Address three issues in the Gemma-native tool-call path:
- _quote_gemma_object_keys stopped a bare (unquoted) string value at the
first comma, so an argument like `location:New York, NY` was split
mid-value and the synthesized JSON failed to parse, dropping the whole
tool call. A bare value now ends only at `}` or a comma that begins the
next `key:` pair.
- parse_tool_calls_from_text scanned the entire response for Gemma markers
even inside a tool call already parsed from a `<tool_call>{...}` JSON
block, so a marker-like string inside an argument (data) was promoted to
a second, unintended tool call. Matches inside an already-consumed call
span are now skipped.
- _openai_passthrough_stream relied on _monitor_openai_sse_line to flag a
stream error, which returns early when monitor_id is None
(skip_api_monitor), so an upstream error chunk left saw_stream_error
unset and the synthetic-finish guard emitted a successful finish_reason
after a failed stream. Error chunks are now detected independently of API
monitoring.
Adds tests/test_gemma_tool_parse_edge_cases.py covering the comma and
marker-injection cases.
Resolve studio/backend/routes/inference.py:
- keep the reasoning_content split (the #6471 fix) on the GGUF chat streams
- keep each streaming path's direct _SameTaskStreamingResponse return (the
cancel-scope fix) and route main's new _sse_streaming_response helper through
_SameTaskStreamingResponse too, so every SSE path stays in the request task
- keep both the new _SameTaskStreamingResponse class and main's
_aclose_stream_resources teardown helper
All 466 tests across the 9 changed test files pass.
* CLI: stop `unsloth connect` from leaking Studio credentials to unverified servers
`unsloth connect` (and `unsloth chat`) discovered a Studio base URL from
UNSLOTH_STUDIO_URL or the default localhost port after only an unauthenticated
/api/health probe, then sent credentials to it:
- keyless connect iterated every cached API key and sent each as a bearer token
to {base}/v1/models, so a malicious or port-preempting endpoint could harvest
all of them;
- with no cached key it self-issued a Studio JWT and POSTed it to
{base}/api/auth/api-keys;
- unsloth chat sent the same self-issued JWT to the discovered base.
The key cache was a flat, global list with no binding to a server identity, so a
key minted for one Studio could be replayed to any other.
Changes:
- Scope the agent key cache per base URL so a key is only ever replayed to the
exact server it was minted for. Pre-scoping flat caches are ignored rather
than replayed (at most one extra local mint on the next launch).
- Gate every automatic credential flow to loopback bases. A non-loopback
UNSLOTH_STUDIO_URL now requires an explicit --api-key and nothing is sent
automatically. SSH-tunnelled Studios that land on 127.0.0.1 keep working.
- Mint the API key locally against the Studio auth DB instead of POSTing a
self-issued JWT over the network, so no bearer token leaves the process on the
local path.
- Apply the same loopback gate to connect_studio_server (used by unsloth chat).
Fully closing same-host loopback-port preemption needs a signed /api/health
handshake so the client can verify the server identity before sending anything;
that is tracked as a server-side follow-up.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* CLI: verify Studio server identity before auto-sending credentials
Adds a challenge-response so `unsloth connect` and `unsloth chat` can confirm a
discovered loopback endpoint is really this install's Studio (not a process
that preempted the port) before sending it a cached or freshly minted
credential. This closes the same-host loopback-preemption gap left open by the
previous commit, which could only limit the blast radius.
Server:
- storage.get_or_create_identity_secret(): a dedicated server-wide secret in
app_secrets (kept separate from the per-user JWT secret), readable only by
the same OS user.
- storage.compute_identity_proof(nonce) = HMAC-SHA256(identity secret, nonce).
- GET /api/auth/identity?nonce=<base64url>: unauthenticated, returns the proof.
The nonce is opaque to the server and the proof reveals nothing about the
secret, so answering is safe.
Client:
- verify_studio_identity(base): sends a fresh 32-byte nonce, recomputes the
expected HMAC from the local same-user secret, and constant-time compares.
Fails closed on any error.
- connect._agent_api_key gates the loopback cached-key replay and the local
mint on it; connect_studio_server (used by unsloth chat) gates the
self-issued JWT on it.
A server that cannot read this install's secret (a different OS user, or a
remote/fake endpoint) cannot produce a matching proof, so the client refuses
and falls back to an explicit --api-key.
Tests: studio/backend/tests/test_identity.py (proof determinism, secret
persistence and caching, route response and nonce validation) and additions to
test_connect.py (the verify gate refuses when unverified, an explicit key skips
the check, and an end-to-end client plus server handshake against a stub HTTP
server).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* CLI: mint through the verified server instead of the local auth DB
CodeQL (py/clear-text-storage-sensitive-data) flagged the API key written to
the per-server cache and the agent config files once it was sourced from
storage.create_api_key(); the original HTTP-minted key did not trip the query.
Now that the identity handshake cryptographically confirms the loopback
responder really is this Studio before anything is sent, minting through the
server's /api/auth/api-keys endpoint with a self-issued JWT is safe again and
restores the original, CodeQL-clean data flow. The local-DB mint path is
removed.
The security properties are unchanged: discovery is still loopback-gated and
identity-verified, the key cache is still scoped per server, and a credential
reaches the server only after the handshake has proven its identity. The only
difference from the previous commit is that the self-issued JWT is sent to the
already-verified loopback server rather than the key being minted in-process.
Tests updated to mint through the fake server again.
* CLI: address review feedback on connect credential handling
- Reuse a saved per-server key before the loopback/identity gate. Keys are
scoped per base URL, so a key the user saved with --api-key for a remote or
SSH-tunnelled Studio (whose identity secret the local handshake can't match)
is replayed only to that exact server. The loopback + identity-handshake gate
now guards just auto-minting (self-issuing a JWT and creating a new key),
which is the path that needs a cryptographically verified local Studio. Fixes
keyless reuse being impossible for remote/tunnelled Studios the user had
saved a key for.
- connect_studio_server (unsloth chat / inference): when the user explicitly set
UNSLOTH_STUDIO_URL but the server can't be safely attached (non-loopback, or
identity unverifiable), fail with a clear message instead of silently loading
the model locally. Opportunistic discovery of the local default still falls
back to a local load.
- Harden cache parsing: tolerate a corrupt or hand-edited cache where a base
maps to a non-list (which would otherwise iterate a string into
single-character "keys"), and read the cache as UTF-8.
Tests updated and added: saved-key replay without the handshake for both local
and remote bases, keyless mint still refused when the loopback server is
unverified, and connect_studio_server erroring on an explicit remote while
falling back locally on default discovery.
* CLI: harden connect handshake against relay and gate cached minted keys
Addresses review feedback on the credential handshake:
- Refuse HTTP redirects on credential-bearing requests (the identity handshake,
/v1/models, key minting, and the chat HTTP backend). A process squatting the
discovered port could 302 /api/auth/identity to the real Studio and relay its
valid proof, or bounce a bearer-token request to another base, and urllib
follows redirects by default. A shared no-redirect opener now treats any 3xx
as an error.
- Give cached keys provenance. Keys the user supplied with --api-key are "saved"
and replay without the handshake (needed for remote or SSH-tunnelled Studios
whose secret the local handshake can't match). Keys we auto-mint are "minted"
and replay only after the identity handshake, so a port squatter can't collect
a previously minted localhost key just by answering the health check. New cache
shape: servers[base] = {"saved": [...], "minted": [...]}.
Known residual: a different-OS-user process that squats the port and can also
reach a genuine same-secret Studio elsewhere on loopback can still manually relay
the identity challenge. Fully closing that needs the proof bound to the server's
real listening port, or OS-level peer-credential checks; tracked as follow-up. A
same-user attacker is out of scope, since it can already read the 0600 key cache.
Tests: redirect rejection in the handshake, minted-cache requiring the handshake
while saved-cache bypasses it, and the existing suites updated for the new cache
shape. unsloth_cli (206) and test_identity.py (5) pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* CLI: keep urllib imports function-local in the no-redirect opener
The repo's import-hoist safety linter (scripts/verify_import_hoist.py) flags
hoisting urllib to module level because it re-points the 'urllib' name in the
pre-existing HttpChatBackend._request scope. Build the no-redirect opener lazily
with function-local urllib imports instead, matching this module's convention,
and restore the local 'import urllib.request' in _request and
verify_studio_identity.
* test(identity): skip route tests when routes.auth import chain is unavailable
The identity route tests build a TestClient from routes.auth, which pulls the
whole routes package (routes/__init__ -> inference -> llama_cpp, ...). In a
minimal test matrix without the heavy backend deps, or when another test in the
same process has already broken that import chain, importing it raised and the
two route tests hard-failed. Skip in that case instead: the proof crypto is
covered by the storage-level tests, and the full backend CI still exercises the
route. No behaviour change where the deps are present (5 passed in isolation).
* test(connect): make connect tests pass on native Windows
unsloth connect supports Windows: --no-launch prints PowerShell ($env:X =
"v" / Remove-Item Env:X) instead of POSIX (export/unset), and the launch
path bridges env into a Windows agent .exe over WSLENV. The tests hardcoded
the POSIX shell forms, so on a real windows-latest runner 12 of them failed on
the assertion string even though every command exited 0.
Add OS-aware assertion helpers (_assert_env_set / _assert_env_unset) that check
the right shell syntax for the host OS, and skip the two WSL-from-Linux shim
tests on native Windows (os.name is 'posix' inside WSL, so that path can't run
there). No change on Linux/macOS (57 passed); the connect command's behaviour
is untouched. Validated on a windows-latest staging runner.
* style(connect): tighten comments in the credential-leak fix
Condense the verbose explanatory comments and multi-line docstrings added by
this PR to one or two lines each, drop a few that just restated the code, and
keep the security rationale where it is load-bearing. Comment/whitespace only;
verified with unslothai/scripts comment_tools.py (check --strip-docstrings:
6/6 'code unchanged'). Tests unchanged: connect 57 passed, identity 5 passed.
* CLI/Studio: harden the identity handshake (review round)
Addresses the latest Codex/Gemini review of the handshake:
- Store the identity secret privately. sqlite3.connect created the auth DB
world-readable under a 022 umask, so another OS user could read app_secrets
and forge proofs, defeating the same-user assumption the handshake rests on.
The auth dir and DB are now restricted to owner-only (0700/0600); the JWT
secret and password hashes there get the same protection.
- Bind the proof to the server's listening port. The stateless HMAC(secret,
nonce) was relayable: a process squatting the discovered port could proxy the
challenge to the real Studio on another port and pass it back. The proof now
covers the port the server actually listens on (from the socket, never the
Host header) and the client checks it against the port it connected to, so a
relayed proof from a different port no longer matches. Closes the manual-relay
residual left after the redirect fix.
- Cap the identity response read (the server is still unverified at that point)
and serve the identity route from a sync def so its first-call SQLite read
runs in the threadpool instead of the event loop.
Tests: port-bound proof + relayed-proof rejection added; identity (5) and
unsloth_cli (58) suites pass. Verified end to end against a real backend
(auth DB owner-only, handshake + mint still succeed).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* CLI/Studio: bind the identity proof to the connection address, not just port
Follow-up to the port binding from the last review. Binding only to the port
left a cross-address relay: a squatter on a different loopback address but the
same port (for example localhost resolving to a squatter on ::1 while the real
Studio is on 127.0.0.1) could proxy the nonce to the real Studio and pass back
a proof that still matched, since both share the port.
The proof now covers the address and the port the connection landed on:
- Server: takes the address+port from request.scope, which uvicorn populates
from getsockname, so it is the real local address the client reached even
when Studio is bound to 0.0.0.0 (verified empirically), never the
client-controlled Host header.
- Client: resolves the base host to one concrete IP, talks to exactly that IP,
and binds the proof to (IP, port). A proof relayed from a Studio on a
different address or port was computed for that other endpoint and no longer
matches the one the client dialed.
Both sides normalise the address through ipaddress so equivalent forms compare
equal. Together with the private-secret and redirect fixes, this closes the
cross-user loopback relay an attacker can mount without reading the secret.
Tests: proof now bound to host+port; relayed-proof rejection retained; identity
(5) and unsloth_cli (58) suites pass. Verified end to end against a real backend
(handshake + mint still succeed).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* CLI: pick the loopback address at discovery so localhost does not regress
find_studio_server now resolves a bare localhost base to its concrete loopback
addresses and returns the first that answers /api/health, IPv4 127.0.0.1 first
(where unsloth studio binds by default). The whole flow (health probe, identity
check, credential send) then targets that one address instead of racing
IPv4/IPv6 resolution, where localhost could resolve ::1-first and hide a Studio
bound to 127.0.0.1. A literal IP or remote name is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: use an isolated Node.js for the frontend build instead of replacing the system Node/npm
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address Node isolation review (no-Node probe crash, PATH refresh, OXC provisioning, venv python, runtime node resolver)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust Node isolation for PR #6533
* Studio Node: don't cache a negative node resolution; accept Node metadata in setup.sh ownership guard
- node_runtime: memoize only a version-adequate executable so a Node installed
by a separate-process 'studio update' is picked up without a backend restart.
- setup.sh: _studio_owned_adoptable also accepts UNSLOTH_NODE_PREBUILT_INFO.json,
matching the setup.ps1 Node ownership guard (custom-home parity).
* Studio setup.ps1: skip OXC npm install gracefully when npm is absent
Mirror setup.sh's `command -v npm` guard so a pip-installed Studio with no
system Node skips the OXC runtime install (validator degrades at runtime) instead
of exit 1 aborting the whole setup. Tighten test_node_probe_guard.ps1's probe
regex so it only matches the two system-version probes, not this new npm guard.
* Wire test_node_probe_guard.ps1 into Windows CI for PR #6533
* Harden isolated Node install and probes for PR #6533
- install_node_prebuilt.py: keep an existing, still-usable isolated Node
when nodejs.org's dist index is unreachable instead of aborting the
update on a transient outage (existing_install_usable + tolerant fetch).
- install_node_prebuilt.py: pin NPM_CONFIG_PREFIX/npm_config_prefix and
drop NODE_PATH in _run_node so any npm -g stays inside the isolated
prefix; Windows npm otherwise writes to %APPDATA%\npm.
- install_node_prebuilt.py: resolve tar hard-link targets against the
archive root (symlink targets stay link-parent relative).
- setup.ps1: wrap the system node/npm probes in try/catch so a present
but broken shim degrades to the bundled Node instead of aborting setup.
- setup.ps1: run the isolated Node install with the handed-off/venv Python
(ReusedSetupPython); the main resolver runs later and bare python may be
a Store stub this early.
- setup.sh: log when the OXC validator runtime is skipped for missing npm,
matching setup.ps1.
- node_runtime.py: move the version-floor comment onto _version_meets_floor.
- Tests for the offline-reuse and broken-shim paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim verbose comments across the Studio Node installer for PR #6533
Comments-only pass: collapse the multi-line section banners to single lines,
drop comments that restate obvious code, and tighten the remaining docstrings
and "why" notes without losing intent. No code changes (verified with an AST
comment-only check on the Python files and a non-comment-diff scan on setup.sh
and setup.ps1). Net 109 fewer lines; the install, decision, and probe-guard
suites stay green.
* Harden Node install from review: validated Python, version floor, legacy home, lock race
For PR #6533, addressing the latest review pass:
- setup.ps1: run the isolated Node install with the validated reused/venv Python.
An incompatible reused interpreter (old venv, conda, stale UNSLOTH_SETUP_PYTHON)
is no longer used; fall back to the resolved python instead.
- setup.ps1: a STUDIO_HOME/UNSLOTH_STUDIO_HOME override equal to the legacy default
now uses the legacy sibling node dir (~/.unsloth/node), matching the runtime
resolver and setup.sh, so OXC can find the Node it installed.
- install_node_prebuilt.py: reject an explicit --node-version below the floor
(^20.19 || >=22.12 || >=23) instead of installing a Node the build cannot use.
- install_node_prebuilt.py: atomically rename a stale install lock before unlinking
so two concurrent runs without filelock cannot both acquire it.
Tests added for the version floor (parametrized + explicit-below-floor rejection).
Full install suite: 937 passed, 1 skipped; setup.ps1 parses; decision tests green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address latest review: armv7l + later-fetch offline reuse for PR #6533
- install_node_prebuilt.py: reject 32-bit ARM (armv7l) up front. Node 24 LTS
ships no linux-armv7l build, so the old path failed late with a confusing
"no sha256"; it now fails fast with a clear unsupported-architecture error.
- install_node_prebuilt.py: extend the offline-reuse fallback to the SHASUMS and
archive fetches. If index.json resolves a newer Node but a later download fails
and a usable isolated Node is already on disk, keep it instead of aborting a
non-force update.
Tests added: armv7l/armhf are unsupported; a SHASUMS failure keeps an existing
usable Node and re-raises when none is present. Full install suite: 941 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add UNSLOTH_STUDIO_HOME node-dir tests (install side + resolver) for PR #6533
* Add regression tests pinning the reuse path read-only and isolating installer writes
Lock in the two invariants behind the isolated-Node design: reusing a good
system Node never mutates the user's Node/npm, and the installer's own npm
calls only ever write inside its install_dir.
- tests/studio/install/test_install_node_prebuilt_logic.py: assert _run_node
redirects NPM_CONFIG_PREFIX/npm_config_prefix into install_dir and drops an
inherited NODE_PATH; assert _ensure_npm_floor scopes the npm self-upgrade to
install_dir (never -g against the system) and is a no-op once npm meets the floor.
- tests/sh/test_system_node_readonly.sh (new, wired into studio-backend-ci.yml):
the setup.sh NODE_SOURCE=system arm runs no global install and sets no
NPM_CONFIG_PREFIX, with a positive control that the bundled arm does.
- tests/studio/test_node_decision.ps1: symmetric structural guard that the prefix
pin and the only global install (bun) live in the bundled branch, not the system arm.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* Studio: fix llama.cpp update toast tag and reload hint
The post-update toast used the job's to_tag, which is the bare bNNNN build
number (same as installed_tag), so it showed e.g. "b9726" instead of the full
release tag. Use status.latest_tag (e.g. b9726-mix-<sha>) to match the tag the
banner already shows, falling back to to_tag and then a generic label.
Also drop "Reload your model to use it." when there is nothing to reload: only
append it when a local model is loaded, since external-provider models do not
use llama.cpp.
* Fix/adjust llama update toast for PR #6493
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Recognise the Gemma 4 separate-drafter MTP family, auto-download the drafter with retry, fall back to n-gram with a clear reason when it cannot be resolved, and retry the download on reload. Gemma 3n (ships no drafter) and embedded-MTP models (Qwen) are unaffected.
Fixes#6406
- Quote bare unquoted string values in Gemma native tool-call args (e.g.
{location:Tokyo,unit:celsius}) so they parse; JSON scalars stay typed.
- Stop _detect_safetensors_features from suppressing supports_tools for
templates that emit Gemma native <|tool_call>, which the shared parser
now reads.
- Add tests for both.
* Studio: label Apple Silicon as Metal/unified memory instead of CPU-only in installers
* Studio: drop redundant aarch64 check from the macOS GPU label detection
* studio: run /generate/stream's sync generator off the event loop to avoid blocking it
* fix: close generator in finally on client disconnect in generate_stream
* Fix/adjust generate stream test for PR #6466
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix/adjust generate stream cancellation for PR #6466
* Fix/adjust generate stream cleanup for PR #6466
* fix: cancel incomplete generate stream cleanup
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
_estimate_mtp_overhead_bytes is also reached for the separate-drafter spec modes
(draft-simple / draft-eagle3) through _user_draft_via_extras. Those modes load a
small distinct drafter with its own KV -- already counted in the draft KV +
weights -- and keep no duplicated full target context; only MTP runs a second
context over the target model's own KV geometry (llama.cpp ctx_tgt). Charging the
~main-KV-sized f16 copy there over-reserved by tens of GiB on an MLA model and
needlessly shrank the advertised context, the same under-advertising #6312 set
out to fix.
Thread mtp_keeps_target_ctx through _estimate_mtp_overhead_bytes (True for MTP,
False for separate-drafter modes) and derive _engaged_is_mtp at the fit call site
so the target copy is added only when the engaged mode is actually MTP. MLA + MTP
(GLM-5.2 / DeepSeek / Kimi) is unchanged, so the GLM-5.2 OOM fix is preserved;
non-MLA and the draft-simple / draft-eagle3 paths no longer pay the copy.
test_mtp_mla_target_ctx.py adds a case asserting the separate-drafter reserve
collapses to the draft KV (no target copy) while the default MTP path keeps it.
* Studio: Auto disables MTP for MLA models (GLM-5.2 et al.); UNSLOTH_MLA_MTP_ENABLED to re-enable
Studio's Auto speculative mode promotes any embedded-MTP model >=3B to
--spec-type draft-mtp. For MLA models (GLM-5.2/DeepSeek/Kimi) that is a
regression: llama.cpp's MLA/DSA MTP path keeps a duplicated full target-KV
context and recomputes the sparse-attention indexer every draft step, so it
runs ~2x slower than no speculation (GLM-5.2 UD-IQ1_S bench: 27 vs 45 tok/s,
flat across draft depth 1..6 and 96-100% acceptance, on both prose and code).
vLLM/SGLang get a speedup from the same model, so this is a llama.cpp
implementation gap, not a model property.
Auto now drops embedded MTP for MLA models and falls back to ngram-mod (or
spec-off when the binary lacks ngram-mod), mirroring the existing sub-3B
fallback. The metadata separator is kv_lora_rank: it is present on MLA models
and absent on non-MLA embedded-MTP models (Qwen3.x-MTP), whose MTP module is
structurally identical but fast, so a "full layer" heuristic cannot tell them
apart. Qwen MTP, separate drafters (Gemma, --model-draft), and non-MTP models
are unchanged.
Explicit overrides still engage the slower MTP route: choosing MTP / MTP+Ngram
in Settings, or passing --spec-type in extra args. UNSLOTH_MLA_MTP_ENABLED=1
re-enables Auto promotion for MLA once the upstream path is optimized.
A new spec_fallback_reason value "mla_mtp_disabled" surfaces this as an
Auto-mode policy downgrade (not a binary/update problem), with a settings
banner that points users at the MTP override. It is deliberately kept out of
the "Update llama.cpp" affordance since updating does not help.
Tests: resolver-matrix rows for MLA->ngram-mod / MLA-no-ngram->off /
non-MLA-Qwen->draft-mtp / MLA-separate-drafter->draft-mtp /
non-MTP-MLA->default / forced mtp|mtp+ngram on MLA->draft-mtp / env flag;
kv_lora_rank metadata fixtures; and reload-skip coverage (Auto ngram-mod is
idempotent, forced mtp bounces a reload).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: add an Open button to reveal the models folder in the file manager
* Studio: report models folder creation failures
---------
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
test_safetensors_capability_advertise: detect_reasoning_flags now returns a
reasoning_effort_levels key, so the none-template expectation must include it.
test_tensor_parallel::test_runtime_recovery_reloads_without_mtp: the assertion
raced the recovery thread, which sets _spec_fallback_reason just before its
finally clears _mtp_runtime_fallback_in_progress. Wait for the flag to clear
before asserting.
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Fix clipped card shadows in Hub trending carousel
The carousel scroller only had vertical padding, so with overflow-x set
the first and last cards had their drop shadow clipped on the horizontal
edges. Add px-2 with a matching -mx-2 so the shadow has room while the
cards stay aligned with the section heading, and scroll-px-2 so snap-start
does not scroll the padding away on load.
* Align carousel edge fades with the scroll clip edge
The shadow fix gave the scroller an -mx-2 bleed, but the left/right fade
overlays stayed pinned to the wrapper edges, 8px inside the clip edge. That
left a thin strip where a card showed beside the fade, so the fade read as a
separate block instead of blending into the background. Offset both fades by
the same 8px so their opaque edge sits on the clip edge.
* Add click-and-drag panning to the Hub card carousel
The rows only scrolled by wheel or trackpad, and grabbing a card started
a native drag of its avatar image, so the cards could not be dragged to
move the row. Add mouse drag-to-scroll (touch and pen keep native
scrolling), swallow the click a drag would otherwise fire on a card, keep
plain clicks working, and block the avatar's native drag.
* Trim carousel edge fade width from 56px to 44px
* Smooth out carousel drag panning
Scroll snap was correcting the position on every drag frame, which made
the pan feel sticky. Disable snap while a drag is active and restore it on
release so the row follows the pointer and then settles on a card.
* Drop stale carousel drag when the button is released off-element
If a press ended outside the scroller before the drag threshold was
crossed, no pointerup reached us and the drag stayed armed, so a later
buttonless mousemove would scroll the row. Bail out and clear the drag
whenever the primary button is no longer held.
* Studio: reserve the duplicated MTP target KV context for MLA models
GLM-5.2 UD-IQ1_S advertised its native 1,048,576-token context, loaded, then
crashed cublasCreate on the first generation with "CUDA error: the resource
allocation failed" on a 2x B200 box. The model loaded fine; the decode OOMed.
Cause: when MTP speculative decoding is engaged, llama.cpp keeps a second full
copy of the target model's KV context for draft verification (ctx_tgt=yes in the
spec log), at f16. On an MLA model that copy is ~the main KV again -- for GLM-5.2
at 1M ctx llama.cpp sized it at ~97.5 GiB -- but the auto-fit reserve only
counted the tiny embedded draft head (~2 GiB), 46x too low. So weights (~202 GiB)
+ main KV (~83 GiB) + a 2 GiB reserve looked like it fit in 2x182 GiB, when the
real footprint with the ~97 GiB MTP copy is ~382 GiB and overruns the cards.
Disabling speculative decoding removed the copy and the same context ran fine.
_estimate_mtp_overhead_bytes now adds the duplicated target context (the main KV
re-estimated at f16) for MLA models, so auto-fit backs the context off (or selects
more GPUs) instead of advertising one that OOMs. It is gated strictly on MLA
(kv_lora_rank present), which is exactly the family that keeps the extra copy
(GLM-5.x, DeepSeek, Kimi-K2); non-MLA MTP (Qwen, Gemma) is byte-for-byte
unchanged. The reserve stays deterministic from GGUF dims, matching #6312.
test_mtp_mla_target_ctx.py covers it: the MLA reserve includes the f16 target
copy and dominates the draft head, the copy is f16 regardless of the main cache
type and scales with context, non-MLA embedded heads keep overhead == draft KV,
and _fit_context_to_vram on the GLM-5.2 / 2x B200 budget now returns a context
below the requested 1M where the old draft-only reserve kept the full 1M.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stub structlog/loggers in MTP-MLA test so it is import-order-independent
The new test imports core.inference.llama_cpp, which pulls in orchestrator ->
structlog. In the lightweight test env structlog is absent, so when this file is
collected before test_mtp_vram_budget.py (it sorts first) or run directly,
collection aborted with ModuleNotFoundError. Install the same loggers/structlog
(+ conditional httpx) stubs the sibling MTP tests use before the import, matching
the established per-file convention.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: fix Bypass Permissions menu freeze and show decimal GB for model sizes
Bypass Permissions freeze: the warning dialog lived inside the composer
"+"/More dropdown and kept the menu mounted via onSelect preventDefault,
so confirming or cancelling the dialog left both popovers frozen open.
Lift the dialog out of the menu into a store-driven
BypassPermissionsConfirmDialog mounted at a stable spot in the composer.
The menu item now closes normally on select and just toggles a new
bypassConfirmOpen store flag, so the popovers dismiss as expected.
Model search sizes: formatBytes divided bytes by 1024 but labelled the
result "GB", so unsloth/GLM-5.2-GGUF:UD-IQ1_S showed 201.8 GB where
Hugging Face reports 217 GB. Switch the search display to decimal
(base-1000) units to match what Hugging Face reports. The GPU-fit math
stays base-1024 since VRAM capacity is binary.
* Studio: address review feedback and add GLM-5.2 high/max/disabled thinking
Review feedback on the Bypass Permissions and size-format changes:
- Mount the Bypass Permissions warning dialog once at the chat-page root
instead of inside each Composer. It is driven by global store state, so
the per-composer mount meant Compare mode (multiple composers) rendered
duplicate dialogs and the shared-composer menu had none. A single root
mount fixes both.
- Defer opening the dialog past Radix's menu-close focus restoration with
setTimeout(0), so the dropdown does not steal focus back and break the
dialog's focus trap.
- Clamp the unit index in formatBytes so units[i] cannot go out of bounds
past TB (and to absorb log() float error at exact powers of 1000).
GLM-5.2 reasoning levels:
GLM-5.2's template gates thinking with enable_thinking and also reads a
reasoning_effort level ('high' or 'max'), so it needs high / max /
disabled rather than the binary toggle it got before (its style was
detected as enable_thinking, which made 'high' unreachable). Add a new
reasoning style 'enable_thinking_effort' that reuses the effort dropdown
but, unlike gpt-oss, can be fully disabled:
- detect_reasoning_flags classifies a template that has both
enable_thinking and reasoning_effort, extracting the discrete levels
from the quoted effort literals it branches on. Templates with only one
of the two (gpt-oss, Qwen3, DeepSeek, GLM-4.6) are unchanged.
- _request_reasoning_kwargs maps the new style to enable_thinking plus an
in-range reasoning_effort; disabling sends enable_thinking=false. The
gpt-oss reasoning_effort path is left untouched.
- The backend reports reasoning_effort_levels on the load/status response;
the frontend carries them through to the effort dropdown and sends
enable_thinking + reasoning_effort for this style.
Verified: backend reasoning kwargs render the real GLM-5.2 template to
"Reasoning Effort: High/Max" (thinking) and an empty <think></think>
(disabled); tsc, eslint, i18n parity and the production build all pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address review feedback on reasoning effort and formatBytes
- chat-adapter localReasoningEffort: accept 'minimal' so a template that
branches on it (extracted into reasoning_effort_levels) is sent through
instead of being coerced to 'low' and then dropped by the backend.
- formatBytes: return '0 B' for non-finite / non-positive sizes (missing
metadata -> NaN, Infinity, negatives) and clamp the unit index lower
bound to 0, so sub-1-byte values can't produce a negative index.
* Studio: hybrid reasoning none gate and decimal GB in load progress
- _request_reasoning_kwargs: for enable_thinking_effort models, treat a
raw reasoning_effort='none' (OpenAI 'no reasoning' sentinel) as the
enable_thinking=false off gate, so a direct API caller can disable
thinking even without passing enable_thinking. The frontend already
sends enable_thinking=false; this only affects raw API callers.
- use-chat-model-runtime: the download / 'X of Y GB in memory' load
progress divided bytes by 1024**3 but labelled GB, so it disagreed with
the model picker and Hugging Face. Use decimal GB (1e9) to match.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: carry hybrid reasoning levels on all load paths and harden formatBytes
Review follow-ups on the enable_thinking_effort work:
- Every model-load path now copies reasoning_effort_levels and derives
supportsReasoningOff, via a shared reasoningCapsFromLoad() helper. The
shared/Compare composer load and the three chat-adapter auto-load paths
previously set only reasoningStyle, so a GLM-style hybrid model loaded
through Compare or first-chat auto-load fell back to the default
low|medium|high and lost its Max / Off controls.
- The local send path clamps the effort to the loaded model's advertised
levels (clampReasoningEffortToLevels) instead of a hard-coded list. A
stale "max" carried over from an external provider no longer reaches a
pure reasoning_effort (gpt-oss) model that only accepts none|low|medium|
high, where the backend would have dropped it.
- formatBytes divides iteratively instead of via Math.log, which has float
error at exact powers of 1000 (log(1e12)/log(1000) = 3.9999... would
label 1 TB as "1000 GB"). Keeps the non-finite/non-positive guard.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: restore sidebar roundness to its pre-#6349 state
* Studio: keep train recents as a rounded rectangle, not a pill
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
* Studio: fall back to text-only when a vision projector hard-crashes llama-server
The text-only mmproj fallback (#6075) only fired when llama-server printed a
recognizable projector-format error ("Unknown projector type", exit -6). An
installed llama.cpp that predates a model's projector can instead SIGSEGV
(exit -11) with no parseable output, e.g. unsloth/Qwen3.5-4B-MTP-GGUF +
mmproj-F16 on an older gfx1151 prebuilt: llama-server crashes on load, the
--fit off retry crashes the same way, and load_model gives up with a hard 500
instead of dropping vision.
Generalize the decision: a vision (--mmproj) launch killed by a signal (POSIX
returncode < 0, e.g. -11 SIGSEGV / -6 SIGABRT; Windows 0xC0000000+ access
violation) is treated like a projector incompatibility, so the load retries
once text-only. The retry is skipped if a cancel/unload is pending, mirroring
the MTP guard. Clean non-zero exits (bad GGUF, port bind) and hung processes
keep their own handling; non-vision launches are unaffected.
Reproduced and verified on gfx1151 (Radeon 8060S, ROCm 7.2.1): a current
prebuilt (llama.cpp b9596) loads the exact model + args fine, confirming the
crash is a stale prebuilt. With a wrapper that SIGSEGVs on --mmproj, Studio now
recovers: the load returns 200 (is_vision=false) and serves at ~31 tok/s
text-only instead of failing. New _is_signal_crash helper plus tests pin the
decision.
Also normalize a few em-dashes to ASCII punctuation in existing comments.
* Studio: refine mmproj hard-crash fallback (signal scope + last argv)
- Limit _is_signal_crash to genuine program faults (SIGSEGV, SIGABRT,
SIGILL, SIGFPE, SIGBUS) and Windows 0xC0000000+ statuses. SIGKILL,
SIGTERM and SIGINT no longer count, so an OOM-killer, unload or
supervisor kill is not masked as a projector incompatibility.
- Strip --mmproj from the last attempted argv so the text-only retry
keeps --fit off / --spec-default instead of resurrecting the original
spec flags (matters for MTP vision models on an older llama.cpp).
- Drop stray temp files committed by mistake and gitignore the "~" dir
so they cannot be re-added.
* Studio: tighten comments in mmproj hard-crash fallback
* Studio: retry --flash-attn off before dropping vision on a startup crash
When llama-server hard-crashes at startup, the recovery chain now tries the
least-destructive mitigation first. Flash-attention kernels SIGSEGV at load on
some ROCm/GPU builds (often inside the vision tower's attention); disabling
flash attention keeps BOTH vision and MTP, so a hard program fault with
--flash-attn on now retries once with --flash-attn off before the MTP-drop or
the text-only (mmproj-strip) fallbacks. _is_signal_crash already gates this to
genuine faults (SIGSEGV/SIGABRT/SIGILL/SIGFPE/SIGBUS), so an OOM-kill or unload
(SIGKILL/SIGTERM/SIGINT) does not trigger a retry.
Field context: a gfx1151 user crashes loading a vision GGUF even on the latest
prebuilt, so an update cannot help, and the same model and args load fine on
another gfx1151 box, pointing at a runtime/flash-attn fault. New
_with_flash_attn_off helper plus tests. Verified on hardware with a wrapper
that SIGSEGVs on --flash-attn on: Studio recovers with is_vision=true (vision
and MTP intact) instead of failing or losing vision.
* Studio: name the OOM kill on a too-large model load
When the OS kills llama-server with no diagnostic output (SIGKILL/SIGTERM,
almost always the OOM killer, e.g. a BF16 model too large for the WSL VM's
RAM cap), the recovery ladder correctly does not retry an external kill, so
this is the message the user sees. It fell through to the generic "is the
GGUF valid / out of memory" text. Make it actionable: name the signal and
point at a smaller or more quantized GGUF, a lower context length, or raising
the WSL memory limit. Output-based diagnoses still win and a hard fault keeps
the generic fallback.
* Studio: refuse a model too large for system RAM on a unified-memory APU
On gfx1150/gfx1151 APUs the weights load into shared system RAM (GGML
unified memory). _get_gpu_free_memory reports the full ROCm/APU budget as
free (often ~100 GB), but under WSL the VM's RAM cap is the real ceiling.
Studio trusted the budget, spawned a load larger than RAM, and the OS killed
it mid-flight, taking the Studio process with it (a silent "Terminated" with
no error, the model resident in RAM not VRAM).
Add a pre-flight guard on the APU path: if the weights exceed available
system RAM (psutil, then /proc/meminfo), refuse before spawning with a clear
message (smaller/more-quantized GGUF, lower context, or raise the WSL memory
limit). Weights only so KV/context auto-reduction is not double-counted;
unknown RAM never refuses; non-APU and discrete-GPU paths are untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: refine recovery ladder (keep diagnosed errors, flip all flash-attn)
Two review points on the hard-crash recovery ladder:
1. The signal-only text-only fallback stripped --mmproj on any hard fault,
even when llama-server had already printed a non-projector cause (an OOM
such as "cudaMalloc failed: out of memory", an unsupported architecture, or
a tensor-parallel limit). That masked the real error and told the user to
update llama.cpp for vision. New _output_has_nonprojector_diagnostic gates
the signal path: it fires only when no such marker is present, so a bare
SIGSEGV with no output still retries text-only, but a diagnosed OOM surfaces
the real error instead of silently dropping vision.
2. _with_flash_attn_off only flipped the first --flash-attn. llama.cpp is
last-wins, so a leftover enable from extra_args (--flash-attn on, -fa on, or
the = form) could keep flash attention on and re-crash the retry. It now
flips every occurrence and returns None only when nothing is flippable.
test_llama_cpp_mmproj_fallback.py and the classification/APU suites: 103 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: pass the text-only retry's exit code to the failure classifier
When the text-only fallback retry itself fails, read its exit code before
_kill_process() clears it and forward it to _classify_llama_start_failure, so an
OS-killed retry surfaces the actionable out-of-memory message instead of the
generic one (matching the primary failure path).
* Studio: scope APU RAM guard to selected GPUs, count MTP drafter, neutral SIGTERM
Three refinements to the startup recovery work in this PR:
- The unified-memory APU RAM guard fired whenever any visible GPU was a
gfx1150/gfx1151 APU, so on a mixed APU+dGPU host it could refuse a valid
load placed on the discrete GPU. Scope _amd_apu_wants_unified_memory to the
selected gpu_indices (physical ids, mapped via CUDA_VISIBLE_DEVICES like
_is_datacenter_gpu); None still means every visible GPU. Applied to both the
RAM guard and the GGML_CUDA_ENABLE_UNIFIED_MEMORY env set.
- The RAM guard counted only the main GGUF plus mmproj, so a separate MTP
drafter (also resident in unified system RAM, even when offloaded to CPU)
could push the load past the RAM cap and still get OS-killed mid-load. Add
the drafter weights to the APU RAM total.
- The startup classifier reported SIGTERM (-15) as 'most likely out of memory',
but SIGTERM is also how an unload/cancel or a supervisor stops the server.
Keep the OOM wording for SIGKILL (-9, the OOM killer) and report -15
neutrally.
Tests updated/added accordingly.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address review on the APU guard and decode-probe ladder
- Map APU physical ids via the active ROCm mask (HIP, then ROCR, then CUDA),
mirroring _get_gpu_memory, so a HIP_VISIBLE_DEVICES-selected APU is matched.
- Only add the MTP drafter to the APU RAM total when MTP will actually engage,
so a stale LLAMA_ARG_SPEC_DRAFT_MODEL cannot refuse a non-MTP load.
- After an MTP first-decode hard fault, retry --flash-attn off (keeps MTP)
before dropping speculative decoding, matching the startup rung.
- Fold the --flash-attn= / -fa= rewrite into one branch.
Tests: tensor-parallel decode-probe assertion updated for the FA-off rung.
* Studio: tighten two comments in the APU guard and RAM preflight
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: refine flash-attn retry and APU RAM guard per review
- _with_flash_attn_off now decides on the effective last-wins value: it returns
None when FA is already off (no wasted retry), and neutralizes a bare
--flash-attn / -fa (which llama.cpp reads as on) so the retry cannot re-enable
it. Length is preserved so downstream index slices stay valid.
- _amd_apu_wants_unified_memory uses 'gpu_indices is not None' so an empty
selection is respected (not treated as all-visible).
- The APU RAM refusal now checks the base model only (main + mmproj); an
optional MTP drafter is dropped by the existing MTP-drop fallback rather than
causing a hard pre-spawn refusal of an otherwise loadable model.
Tests: bare-flag / effective-off / empty-selection / HIP-mask cases added.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: free chat model VRAM at training start only when the GPU is tight
The training start route unconditionally tore down the transformers/MLX
inference subprocess before training, and never stopped the llama.cpp GGUF
server at all, so a loaded GGUF chat model kept holding VRAM for the whole
run. Conversely the HF model was always unloaded even when there was plenty
of room to keep it.
Make the unload VRAM aware and cover every inference backend:
- Add routes/training_vram.py with summarize_resident_chat(),
can_keep_chat_during_training() and free_chat_models_for_training(). The
keep/unload decision reuses the same estimator and live per device free
VRAM reader the training GPU selection already uses (auto_select_gpu_ids,
estimate_required_model_memory_gb, get_visible_gpu_utilization), so the
probe agrees with the placement computed later in start_training.
- When a chat model is resident and training fits alongside it with a
conservative margin (required_gb * 1.15 + 4 GB), keep it loaded so the
user can train and chat at the same time; on a multi GPU box training
lands on a different GPU and both coexist. Otherwise unload the HF/MLX
orchestrator and the llama.cpp GGUF server before training starts.
- The export subprocess shutdown stays unconditional and now runs first so
its freed VRAM is reflected in the decision.
Default deny: non CUDA backends, unestimable models, or any probe error
fall back to the previous always unload behavior.
Adds tests/test_training_vram_coexistence.py and updates two existing route
tests in test_gpu_selection.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: per-GPU floor for explicit GPU lists + don't unload chat on invalid gpu_ids
Address review feedback on the chat coexistence probe:
- Explicit gpu_ids mode now enforces a per-GPU floor in addition to the
aggregate free-VRAM check, mirroring auto_select_gpu_ids' min_per_gpu_N.
Without it, an uneven split such as free [45, 10] for a 40 GB job passed
the aggregate threshold and kept chat loaded even though the 10 GB GPU
could not hold its training shard, risking an OOM.
- Invalid explicit gpu_ids (ids outside the visible set, or a UUID/MIG
mask) make resolve_requested_gpu_ids raise. That request is rejected with
a 400 before training starts, so leave the resident chat model untouched
instead of unloading it.
- Tighten the target_modules / gpu_ids type hints to List[str] / List[int].
Adds tests for the per-GPU floor (uneven split unloads, even split keeps)
and for invalid gpu_ids keeping the chat model loaded.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: only free chat VRAM once training will start; handle in-flight and CPU-only chat
Address the second review pass on the chat-coexistence path:
- Run the chat/export VRAM teardown as a before_spawn hook inside
TrainingBackend.start_training, fired only after the start guards pass.
Previously the route freed chat VRAM before calling start_training, so a
refused start (e.g. a lingering pump thread) would tear down the resident
chat model even though no training job began.
- Treat an in-flight HF chat load (loading_models set, no active model yet)
as not safely sizeable: free it rather than risk both OOMing as the load
keeps allocating after training starts.
- Do not count or tear down a GGUF llama-server confirmed to run entirely on
CPU (_gpu_offload_active is False): it holds no VRAM, so killing it cannot
help training fit.
Adds tests for the before_spawn hook (runs on start, skipped when a
subprocess is alive or a pump thread will not die, survives a hook error),
the in-flight load flag, and the CPU-only GGUF exclusion in both the resident
summary and the unload path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: treat any in-flight chat load (HF swap / mid-start GGUF) as unsafe to keep
Tighten the in-flight detection in summarize_resident_chat so the keep check
never sizes a load that is still allocating:
- Flag loading on ANY non-empty loading_models, not only when active_model_name
is empty. load_model adds the new model to loading_models before clearing the
old active_model_name, so a replacement load during a swap was previously
sized as a normal resident and could OOM as the new model finishes loading.
- Flag a GGUF server that is active but not yet healthy (is_loaded False) as
in-flight: it is still mmaping/offloading layers, so its final VRAM footprint
is unknown.
Consolidates the signal into a single resident["loading"] flag; the route frees
the chat model whenever it is set. Adds tests for the replacement HF load and
the mid-start GGUF cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten comments in chat/training VRAM coexistence (comments only)
* Studio: run before_spawn VRAM hook only after GPU-selection validation
Reviewers found the before_spawn hook fired before prepare_gpu_selection
validated gpu_ids (and before config build), so a refused start (invalid
gpu_ids -> 400, or a bad grad-clip value) could still tear down chat/export
VRAM. Move the hook to immediately before proc.start(), once all synchronous
validation and process construction have passed. This also fixes the route's
in-flight-chat loading branch, since that teardown runs inside the same hook.
Add test_hook_skipped_when_gpu_selection_rejects.
* Studio: recompute GPU auto-selection after the before_spawn VRAM hook
Codex P2: with before_spawn moved after prepare_gpu_selection, placement was
frozen against the pre-teardown VRAM state while the hook freed export/chat
afterward. Auto-selection could pin training onto a GPU the hook then cleared
(or onto a kept chat model). Split validation from placement: explicit gpu_ids
are still validated before the hook (raise -> 400, no teardown; explicit
placement is VRAM-independent), but VRAM-dependent auto-selection now runs
after the hook so it sees the freed memory.
Add test_auto_placement_runs_after_hook and test_explicit_placement_validated_before_hook.
* Studio: allow chatting during training (lift sidebar gate + VRAM-aware load guard) (#6335)
* Studio: allow chatting during training (lift sidebar gate + VRAM-aware load guard)
The sidebar disabled New Chat, project, and home navigation while a training
run was active, so users could not chat during training even though the backend
serves inference fine alongside a run. This removes that gate and adds a backend
guard so the one genuinely risky operation, loading a new local chat model
mid-training, is refused with a clear 409 when it would not fit beside the run.
Frontend (app-sidebar.tsx): drop the chatDisabled = isTrainingRunning gate and
its consumers. Navigation triggers no model load on its own, so chat stays
usable during training.
Backend (routes/training_vram.py, routes/inference.py): add
can_load_chat_during_training plus a load/validate guard that sizes the same
effective load the loader performs (LoRA 4-bit to 16-bit resolved first, HF auto
placement via auto_select_gpu_ids, explicit multi-GPU per-GPU floor, GGUF sized
from on-disk shards and companions or the selected remote variant). It is a
no-op when training is inactive, never blocks external providers or
already-resident models, and default-denies only on a CUDA sizing failure so a
load can never OOM the run. Validate refuses early with the real settings so the
frontend does not unload the resident chat model for a load that would be
rejected.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address review feedback for chat-during-training load guard
- Run the load/validate VRAM guard via asyncio.to_thread so the sync
nvidia-smi + HF metadata work never blocks the event loop.
- Size the GGUF KV cache at the requested context (_estimate_gguf_kv_gb)
and add it to the local GGUF estimate so large-context picks are not
under-counted.
- Keep the requested quantization when adapter_config.json is malformed
(not a JSON object) instead of raising in _effective_load_in_4bit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: size the training load guard at the launcher's effective GGUF context
The GGUF KV-cache estimate used max_seq_length only, but the llama.cpp
launcher honors a user --ctx-size/-c in llama_extra_args. A load such as
max_seq_length=4096 with --ctx-size 131072 was sized against a 4k cache
while the server allocates 131k, so the guard could approve a long-context
GGUF load that then OOMs training. Size the guard's KV at the larger of
max_seq_length and the parsed --ctx-size (reusing the launcher's own
parse_ctx_override), keeping the conservative f16 cache so the estimate is
never smaller than what the server allocates.
The chat model picker also validated with the raw max_seq_length while
/load sizes with resolveLoadMaxSeqLength, so validate could pass, unload
the current model, then have /load reject the native-context load. Validate
now uses the same effective context; the load path is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: size the GGUF training guard at the server parallel-slot count
The KV-cache estimate assumed a single slot, but llama-server allocates the
cache across --parallel slots (app.state.llama_parallel_slots). On a Studio
launched with --parallel N>1 the guard under-sized the cache N-fold and could
approve a GGUF chat load that then OOMs training. Thread the same slot count
the loader uses into the guard's KV estimate; default 1 leaves single-slot
setups unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments for chat-during-training guard
* Studio: keep chat generation alive across navigation; Train spinner + Return to Chat
Hoist the base chat runtime above the routed outlet so navigating to Train (or any tab) no longer aborts an in-flight generation; only an explicit Stop cancels. Add a Train sidebar spinner and swap New Chat to Return to Chat while a run is active, with a lightweight completion watch so the spinner clears from any tab. Also respawn a chat llama-server killed mid-session and guard unreadable HF cache dirs that 500'd the hub model list.
* Studio: show Return to Chat on the Train tab whenever a chat is live
Previously the top sidebar item only swapped to Return to Chat while training was running; on the Train tab with an idle/just-finished run it stayed New Chat, which started a fresh thread and cancelled an in-flight generation. Show Return to Chat (and navigate back, preserving the run) whenever a generation is running or its thread is still active, or training is in progress.
* Studio: keep a running chat alive when starting a New Chat
Starting a New Chat (or switching threads) while a generation was in flight
remounted the single-chat runtime provider, which detached the in-flight run
and cut the previous chat off (it showed up frozen / empty when reopened).
Key the single-chat view by project instead of by thread or new-chat nonce so
the provider stays mounted and assistant-ui switches to a fresh thread in place.
The previous generation keeps streaming in the background and autosaves on
completion, and returning to that thread reattaches the live run instead of
reloading a half-saved one.
Also:
- "Return to Chat" now lands on the thread that is still generating rather than
the empty new chat that became active after New Chat.
- Skip the explicit /inference/cancel POST when an abort comes from a runtime
detach (navigation / background switch) rather than an explicit Stop, so a
backgrounded generation is never cancelled behind the scenes.
* Studio: make model export non-blocking and inline
The Export tab opened a full-screen modal that trapped focus, could not be
closed or cancelled while running, and showed no progress. It also stopped
training and unloaded the chat model before loading, so export could not run
alongside them.
Export now mirrors the training runtime pattern:
- Inline panel embedded where the Export Model button was, with no modal or
backdrop, so the rest of the UI stays usable during an export.
- Global export runtime store plus an app-root lifecycle hook, so a run keeps
going and streaming across navigation and is reflected on the Export nav item
from any tab.
- The worker log stream now stays connected across the load to export phase
boundary instead of stranding on "Waiting for worker output".
- Progress bar driven by phase and quant index (quant N of M for GGUF), with
elapsed time and a working Cancel.
- load-checkpoint no longer stops training or unloads inference; export loads in
its own subprocess in parallel and surfaces out-of-memory as a clear error.
- Add POST /api/export/cancel and is_export_active on /api/export/status.
* Studio: show Return to Chat on the Export tab too
Extend the New Chat to Return to Chat swap to the Export route so leaving a
running chat for Export offers a way back to the live generation, matching the
Train tab.
* Studio: smooth out Export animations and polish the panel
- Drop the height-based reveal animations (source switch, run panel, quant
picker, hub fields) that caused flashing and reflow; use instant swaps and
quick opacity fades instead.
- Method and quant cards now transition colors only, with no transition-all or
hover lift, so selecting a method or quant is crisp instead of jumpy.
- Auto-scroll the export panel into view when it opens and add a scroll-to-bottom
button when its output is below the fold, like Chat.
- Show Return to Chat on the Export tab while an export is running, matching how
training drives it on the Train tab.
- Surface the current phase or stage in the live output before the first worker
line arrives so the panel never looks stuck while progress is advancing.
* Studio: show Return to Chat on every non-chat tab
Generalize the Return to Chat swap from just Train/Export to any non-chat route
(Recipes, Projects, Hub, ...) so a running or active chat is always one click
away, instead of showing New Chat there.
* Studio: stream export logs over the Cloudflare tunnel; drop janky export animations
Exporting over a --secure Cloudflare quick tunnel showed "connecting..." with no
logs while the progress bar advanced. Cloudflare buffers text/event-stream and
only flushes when the stream closes, so the SSE log stream never reached the
browser during the run (direct localhost is unaffected, which is why this only
showed up over the tunnel).
Add a tunnel-safe JSON poll fallback (GET /api/export/logs?since=) that the
runtime lifecycle hook polls while a run is active. Short JSON responses are not
buffered by the proxy, so logs show up in near real time over the tunnel. It
shares the orchestrator's monotonic seq cursor with the SSE stream and the store
de-dupes by seq, so the two transports run together (SSE on localhost, poll over
the tunnel) without double-printing. A successful poll marks the panel
"streaming" instead of leaving it stuck on "connecting...".
Also remove the framer-motion AnimatePresence reveals from the export config and
run panel (quant picker, hub fields, the inline run panel, and the live log
section). The expand/slide animations flashed and felt clunky; the sections now
render in place.
* Studio: recover export over the Cloudflare tunnel when the blocking POST times out (524)
A model export over a --secure Cloudflare quick tunnel showed "Request failed
(524)" even though the export succeeded on the backend (the GGUF was written).
Cloudflare returns 524 when a single request takes longer than ~100s to respond,
and a GGUF conversion routinely runs for minutes, so the blocking per-method
export POST is cut off while the backend keeps going.
Confirm completion via short status polls instead of relying on the long POST
response (the same approach that fixed log streaming):
- The orchestrator records each finished op's outcome (status / output_path /
error) with a monotonic seq, exposed on GET /api/export/status.
- parseJson now preserves the HTTP status; a 524/520/522/523/502/503 or a
status-less network drop is classified as a recoverable transport error.
- runExport wraps each phase (load, every export method, each GGUF quant): on a
recoverable failure it keeps the run alive (logs keep streaming, the panel
shows "reconnecting...") and polls status until the still-running op finishes,
then settles from the recorded result, recovering the output path for the
success banner. A real 4xx still fails immediately; localhost still uses the
fast POST response. applyBackendStatus also settles a reloaded run from the
last-op record.
Verified over the tunnel: a 3m14s gemma-4-E4B-it GGUF export now ends on the
success banner with the output path instead of 524.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep the export method + logs visible after navigating away mid-export
While an export was running, navigating to another tab and back to Export
remounted the page and reset the local form state (exportMethod, quant levels),
so the method card showed unselected and the run panel's log area was hidden
until the card was re-clicked. The run itself lives in the global store and was
unaffected.
Seed exportMethod / quantLevels from the active run's summary via lazy useState
initializers on (re)mount, and gate the panel's log area on the live run
(isExporting / logLines / the run's method) rather than only the local form
selection. The card stays selected and the logs/progress stay visible across
navigation; nothing changes when no run is active.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: address export/training review findings
- Export: guard Start against an empty GGUF quant selection so an inline-panel
run with no quant can't settle as success with no file produced.
- Export: thread the source HF token into the background load so gated/private
HF source exports (and gated bases) authenticate, matching the consent path.
- Export: only settle a recovered (non-owned) run as a finished export when the
last backend op was an export, not a standalone load_checkpoint.
- Training: free the export subprocess whenever an export is active, not only
once a checkpoint is loaded, so an in-flight export load can't race training
for VRAM (current_checkpoint is unset during the load phase).
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: polish Bypass permissions toggle and add it to chat menu settings
- Recolor the pill, menu item and settings caption from alarm red to a
bright yellow accent via a new --bypass token (light and dark).
- Use the shield icon on the composer pill (X on hover, like other pills).
- Expand the composer to two rows the moment Bypass permissions is on.
- Keep Bypass leftmost among tool pills; Compare still sits ahead of it.
- Add a Bypass permissions row to Settings > Chat menu so it can be pinned.
- Widen the "More" submenu so the label fits on one line.
* Studio: use shield-ban icon and lighten the Bypass permissions pill
- Swap the lucide shield-off icon for the Hugeicons shield-ban across the
pill, menu item and settings row.
- Lighten the pill background and deepen the yellow text a touch so the
label stays readable on a near-white tint.
* Studio: drop the Bypass permissions resting fill, keep it on hover
Show only the yellow icon and label at rest; the rounded hover pill picks
up the yellow accent like the other toggles.
* Studio: lighten the Bypass permissions hover fill
Drop the hover tint so the yellow label stays readable on hover.
---------
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Switch the sidebar profile button from rounded-full to rounded-[5px] so its hover background is a rounded rectangle instead of a pill. The collapsed icon-only state stays circular. Styling only.
* Studio: fail fast on an invalid first training batch
Training a base vision-language model (e.g. Qwen/Qwen2-VL-7B or
unsloth/Qwen2-VL-7B) on a conversational image dataset crashed on the first
step with 'Expected ... Long, Int; but got torch.cuda.FloatTensor (embedding)'.
Root cause: the base model's chat template is a flat, media-only template that
renders to an empty string for role-based messages, so UnslothVisionDataCollator
hands the processor empty text, the processor returns empty input_ids, torch
defaults the empty tensor to float32, and the embedding lookup rejects it.
Add a preflight that runs one real batch through the trainer's own tokenization
and collation right before train(), and stops the run with an actionable message
when input_ids is empty or non-integer (pointing to the instruction-tuned variant
for the base-model case). Faithful across text, vision and audio-VLM paths, and
never blocks a run whose first batch is valid.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments in the training preflight
* Stub unsloth/trl in preflight test so backend CI collection passes
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
