* 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>
* 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>
* 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>
#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>
* 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>
* 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>
* 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>
* 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>
* 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>
* 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>
* 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>
* 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>
* 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>
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.
* 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.
* 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 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>
* 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>
* 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>
* 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>
* 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>
* 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>
* 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>
* 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>
Suchitra Malimbada