The --local / GGUF-only install resolves its Python deps from
no-torch-runtime.txt, installed with --no-deps. That file was missing the
RAG group that studio.txt declares (sqlite-vec, pymupdf, python-docx), so a
fresh `unsloth studio` came up with RAG disabled: rag_db.py cannot import
sqlite_vec and logs "RAG unavailable: sqlite-vec extension could not be
loaded", and the knowledge-base routes return 503. python-docx was also
absent, so DOCX ingestion failed.
Add the three RAG store and document-parsing deps with the same pins as
studio.txt so knowledge bases work out of the box on the no-torch path.
sentence-transformers (dense embeddings) was already present.
The post-install path cleared only the in-memory freshness caches and then
re-primed the 24h disk cache with a forced GitHub refresh. When that refresh
cannot reach GitHub, latest_published_release falls back to the last-good disk
value, so a still-fresh same-base mix tag cached before the swap (b9596-mix-aaa
vs the just-installed b9596-mix-bbb) is replayed and the prebuilt reads as
behind, surfacing a false update banner that points back at the build that was
just replaced.
Give reset_caches a drop_disk option and use it on the update path: with the
disk cache gone, an offline post-install refresh leaves latest as None and the
banner fails open (off) instead of lingering on the stale same-base value. The
no-arg form stays in-memory only. Adds regression coverage for the drop, the
default no-op, and the fail-open vs stale-replay contrast.
* Studio: tune llama.cpp env for data-center GPUs
Detect datacenter/professional NVIDIA GPUs at llama-server launch and set
the llama.cpp env flags that help them, gated so consumer GeForce, AMD/ROCm,
CPU and macOS are never touched.
- GGML_CUDA_FORCE_CUBLAS_COMPUTE_32F=1 for any DC GPU (FP32 cuBLAS
accumulation). On a B200 this is ~0% throughput cost with identical
perplexity (7.3230 wikitext-2-raw, baseline and on), where on GeForce the
same flag costs real throughput, hence the gate.
- GGML_CUDA_P2P=1 and CUDA_SCALE_LAUNCH_QUEUES=4x for multi-GPU DC boxes.
Benchmarked on 6x B200: +33-51% prompt processing on tensor (row) split and
+8-16% on the default pipeline (layer) split, with no regression on the
other split or on token generation.
Detection uses torch device names (A100/A30/H100/H200/H800/GH200/B200/GB200/
GB300/L40/L4/RTX PRO 6000/RTX 6000 Ada). A mixed box with one consumer GPU in
the selection is treated as non-DC. All writes are setdefault so a user value
always wins, and UNSLOTH_DISABLE_DC_TUNING=1 turns the whole thing off.
37 unit tests cover detection, multi-GPU gating, user-override precedence, the
disable flag and fail-open on error.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix data-center GPU detection false positives and physical-id mapping
Two issues in the data-center llama.cpp env tuning gate:
- _is_datacenter_gpu matched the marker allowlist as unbounded substrings, so
workstation/laptop parts "NVIDIA RTX A1000" and "NVIDIA RTX A3000" matched
"a100"/"a30" and were wrongly tuned as data-center GPUs (forcing FP32 cuBLAS
accumulation and the multi-GPU env, which carry a real cost on those cards).
Switch to a word-boundary regex.
- gpu_indices carries physical GPU ids (translated from torch ordinals by
_get_gpu_free_memory via CUDA_VISIBLE_DEVICES), but they were passed straight
into torch.cuda.get_device_properties, which expects mask-relative ordinals.
On a masked host (e.g. CUDA_VISIBLE_DEVICES=4,5,6,7) a selection like [4,5]
fell out of range and silently dropped the tuning, and on a mixed mask it could
probe the wrong GPU class. Build a physical-id to device-name map mirroring
_get_gpu_free_memory, then look up the selection by physical id.
Add regression tests for the A1000/A3000 false positives and for masked-host
physical-id selection (reordered and mixed-class masks included).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten data-center GPU tuning comments
Comment-only pass over the DC tuning block and its tests: shorten verbose
docstrings/comments, drop ones that restate the code, collapse multi-line
blocks. Keep the load-bearing rationale (physical-id vs ordinal mapping, the
word-boundary reason, the B200 benchmark numbers). No code change: verified
with comment_tools.py check --strip-docstrings (code unchanged, comments only).
---------
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix/uv-bytecode-timeout
* make sure that win installer upgrades uv for bytecode timeout
* Clarify uv bytecode timeout comment in install.sh and install.ps1
* Read installer scripts as UTF-8 in parity test so it runs on Windows
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Prefer freshly installed uv when an older one shadows it on PATH
---------
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 step count mismatch when sequence packing is enabled
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Emit a single step-0 progress event and guard applyStatus totalSteps
Merge the two consecutive _update_progress calls before train() so the
step-0 gate in _on_progress fires once instead of twice, avoiding a
duplicate startup event and a null-metric step-0 row in training_metrics.
Apply the same positive-number guard to applyStatus that applyProgress
uses, so a stale or startup status poll can no longer overwrite the
packed step count with 0 or replace it with a stale total.
* Log debug message when train_dataset length is unavailable
The TypeError fallback for length-less datasets (e.g. streaming
IterableDataset) was silent, leaving no trace that the step estimate
came from the raw dataset rather than the packed 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: Etherll <61019402+Etherll@users.noreply.github.com>
IOReport energy counters can reset (sleep/wake, power gating), making a poll
delta negative. Return None for a negative total so the monitor shows -- for
that poll instead of a bogus negative wattage; it self-corrects next poll.
* Studio: Add inline confirmation (Allow/Always allow/Deny) for tool calls
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix race in tool-call confirmation gate
* Studio: gate built-in tool calls and harden the confirmation handshake
The Allow / Always allow / Deny controls only lived in the fallback tool
card, but the built-in tools (web search, python, terminal, code
execution, image generation) render with their own components and so
never showed the buttons. Those calls paused after tool_start with no way
to approve them, hanging until the 1 hour timeout. Only MCP tools, which
use the fallback renderer, actually worked.
Render the controls for every tool card by wrapping each registered tool
component (and the fallback) in thread.tsx with a shared
ToolConfirmationControls, so the gate applies uniformly.
Also make the handshake robust:
- The gate keys on a per-call approval_id minted by the backend and
echoed in tool_start, instead of session_id alone, so a stale or
concurrent confirmation can no longer resolve the wrong call.
- The approval slot is registered before tool_start is yielded, closing
the race where a fast click or an auto "Always allow" could reach the
backend before the waiter existed.
- The frontend resolves with the same session id the request was sent
with (plus the approval_id), fixing the new-thread mismatch where the
confirmation targeted a different session than the blocked stream.
- The confirm endpoint returns {resolved}; the UI keeps the buttons and
shows a retry hint until the backend confirms a match, instead of
hiding them on a failed or mistargeted post.
- The gate runs after the disabled-tool and duplicate-call checks, so a
call that will not execute is not put up for approval. A denied call is
still excluded from duplicate detection, so re-issuing and approving it
works.
- "Always allow" is scoped per session to match the backend gate.
Add backend tests for the approval registry, the SSE no-deadlock
handshake, and the loop integration (allow, deny, disabled, duplicate,
re-issue after deny).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move "Confirm tool calls" to the Tools section
* Studio: Keep tool group open while a tool call awaits confirmation
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix tool confirmation session scope for PR #5869
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix confirmation follow-ups for PR #5869
* Apply pre-commit formatting for PR #5869
* Fix confirmation cleanup for PR #5869
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden confirmation lookups for PR #5869
* Studio: make the tool-call confirmation decision immutable
resolve_tool_decision accepted a second confirmation for the same approval_id
and overwrote slot["decision"] in the window before the waiter reads it and
pops the slot, so a duplicate or out-of-order POST could flip an Allow to Deny
(and returned a misleading resolved:true). Reject once the slot's event is
already set so the first decision wins. Adds a regression test.
* Fix/adjust tool confirmations for PR #5869
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* Studio: show Apple GPU temperature and power in the GPU monitor (macOS)
The GPU monitor on Apple Silicon always showed -- for Temperature and
Power: the MLX branch of get_gpu_utilization() hardcoded None because
ioreg's AGXAccelerator PerformanceStatistics carries neither metric.
Add utils/hardware/apple.py, mirroring macmon's no-sudo approach:
- Temperature: average of the AppleSMC "Tg*" float keys via the
AppleSMCKeysEndpoint user client (ctypes/IOKit, macOS 14+).
- Power: IOReport "Energy Model" group, "GPU Energy" channels; each
poll diffs the energy counter against the previous poll's sample, so
the value is the average wattage over the polling window. The first
poll only sets the baseline and returns None.
Both readers latch to None on first failure and never raise, so
non-Mac platforms and locked-down hosts keep the previous behavior.
* Sample IOReport with the subscribed channels descriptor for PR #6187
IOReportCreateSubscription writes the channel descriptor that later samples
must use; sampling with the original requested group can return no Energy
Model entries on hosts that normalize the channel set, leaving power_draw_w
null after the baseline. Use the subscribed descriptor (matching macmon) and
fall back to the requested channels if the OS leaves it unset.
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <23090290+danielhanchen@users.noreply.github.com>
* Fix EmptyLogits gathering in nested structure and patch recursively_apply on accelerator module
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Wire EmptyLogits Accelerate patch into startup and fix find_device, pickling, tests for PR #6092
- Call patch_accelerate_recursively_apply() in _gpu_init.py so real imports
install it; previously it was only invoked by the tests
- Make both wrappers idempotent so repeated calls do not stack
- Rework find_device: skip EmptyLogits while still finding real tensors in any
order, keep returning None for tensor-free payloads (AlignDevicesHook relies
on None), fall back to PartialState().device only for sentinel-only payloads
- Give EmptyLogits stateless __reduce__ and drop the stomped pickle stubs on
EMPTY_LOGITS so debug mode gather_object works in real distributed runs
- Put test tensors on PartialState().device so the debug mode test also passes
on GPU machines, and add drift tests for startup wiring, idempotency and
find_device ordering
Verified on 2x B200: ACCELERATE_DEBUG_MODE=1 torchrun gather/broadcast/pad of
sentinel and mixed payloads all pass, training losses unchanged, full drift
suite 25/25.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Define EmptyLogits equality on the class for PR #6092
Gathered sentinel copies must compare equal in accelerate debug mode
regardless of whether the patched recursively_apply saw the sentinel first
in that process. Class body __eq__ requires restoring __hash__ explicitly.
Verified: 123 case simulation battery on accelerate 0.34.2 through latest,
2 process gloo CPU and NCCL GPU debug mode runs, drift suite 25/25.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Fix UnicodeDecodeError on Windows reading subprocess output in save path
On Windows the default text encoding is the locale code page (cp1252), not
UTF-8. The text-mode subprocess calls in save.py (text=True /
universal_newlines=True) set no explicit encoding, so they decode
llama.cpp / Ollama output with cp1252. When a child process emits a byte
undefined in cp1252 -- e.g. 0x9d, which appears inside the UTF-8 encoding
of common punctuation / box-drawing glyphs and in non-ASCII file paths --
the read raises UnicodeDecodeError and aborts GGUF export.
Add encoding="utf-8", errors="replace" to all 8 text-mode subprocess calls.
errors="replace" also avoids silent mojibake for inputs whose bytes happen
to be valid-but-wrong in cp1252.
Add tests/saving/test_save_subprocess_utf8_encoding.py:
- an AST drift detector asserting every text-mode subprocess call in
save.py pins encoding="utf-8" (runs without importing torch/unsloth_zoo)
- a behavioural test reproducing the cp1252 failure and the utf-8 fix
Relates-to: #2660
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* add FastModel config passthrough
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix fastmodel config passthrough for task configs
* fix config-driven FastModel task model selection
* fix text only fastmodel task config selection
* fix fastmodel task config inference from user configs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix fastmodel problem_type config passthrough
* fix fastlanguagemodel config passthrough: FastLlamaModel owns user config
* fix fastlanguagemodel config passthrough: forward user config to causal loads and keep checkpoint quantization_config
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
The post-download llama-quantize / llama-server smoke test JIT-compiles CUDA kernels on the first GPU forward pass and stalls every install and update by minutes on Blackwell (sm_100). Gate it behind _RUN_STAGED_PREBUILT_VALIDATION, disabled for now, keeping the smoke test and the source-build fallback it triggers fully intact so it can be restored by flipping the flag to True.
Hashless external prebuilts (e.g. lemonade) are not in the approved-sha256 manifest and rely on the functional smoke test as their only integrity gate, so they are always validated regardless of the flag; only approved bundles, already proven by the sha256 manifest, skip it.
The sha256 archive verification and the static Linux/macOS preflights are unchanged and still run for every install.
test_sidebar_account_block_uses_leading_tight hardcoded gap-0.5 in its selector, but the sidebar account-block div moved to gap-px during UI polish (#6196), so the regex stopped matching and the test failed across every studio PR's Repo tests (CPU). Match the gap utility loosely (gap-\S+) since this guard is about the leading-* class for descender clipping, not the spacing.
* fix(studio): adopt server-loaded model before chat auto-load
When the user starts Studio via `studio run -m`, the web UI could still
auto-load a different cached GGUF on the first message because the chat
checkpoint was empty. Sync from /api/inference/status before falling back
to autoLoadSmallestModel so CLI-loaded models are not replaced.
Co-authored-by: Cursor <cursoragent@cursor.com>
* fix(studio): hydrate adopted CLI model and harden auto-load errors
Extract shared inference-status hydration for refresh() and CLI adopt
paths so the first chat turn gets reasoning/tools flags. Wrap auto-load
(including adopt) in try/catch for image-edit cleanup, and drop the
redundant adopt call in run().
Co-authored-by: Cursor <cursoragent@cursor.com>
* Guard model adoption against status failures and mid-flight selection for PR #5900
* ci: trigger pre-commit.ci after main merge
Co-authored-by: Cursor <cursoragent@cursor.com>
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): inherit llama_extra_args and honor --no-mmproj
Reloading the same GGUF from the UI without gguf_variant no longer drops
CLI pass-through args like --no-mmproj. Skip mmproj download and launch
when --no-mmproj is present in llama_extra_args.
Co-authored-by: Cursor <cursoragent@cursor.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix(studio): tighten GGUF llama_extra_args variant inheritance guard
Reject inherited CLI args when the request changes gguf_variant or when
omitted variant resolves differently from the stored extra_args source.
Co-authored-by: Cursor <cursoragent@cursor.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Treat --no-mmproj-auto and --mmproj-auto with last-wins parsing for PR #5902
---------
Co-authored-by: Cursor <cursoragent@cursor.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): load run.py by path for editable installs
`studio update` can leave a partial site-packages/studio/backend/ tree
(plugin build artefacts only). That shadowed tree wins over an editable
install and breaks `from studio.backend.run import ...`. Loading run.py
by file path via importlib sidesteps the conflict.
The module is cached in _RUN_MODULE so repeated calls are cheap.
If exec_module fails, the module is removed from sys.modules before
re-raising so a subsequent retry starts clean.
Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Handle None __file__ when checking cached run module for PR #5909
* Harden _load_backend_auth_storage against None __file__ and resolve cache-key path (PR #5909)
* Adapt studio run/cloudflare in-venv tests to _load_run_module loader (PR #5909)
---------
Co-authored-by: Jim Dawdy <jimdawdy@Jims-MacBook-Pro.local>
Co-authored-by: Claude Sonnet 4.6 <noreply@anthropic.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Handle rope_type 'default' on transformers 5 to stop false RoPE warning
transformers 5 reports rope_type="default" for every plain (unscaled) config
and dropped "default" from ROPE_INIT_FUNCTIONS. _compute_config_rope_inv_freq
then did ROPE_INIT_FUNCTIONS["default"], hit KeyError, returned None and logged
"Could not apply RoPE scaling 'default'; long-context generation may degrade"
on every model load. The inv_freq was still correct (the constructor recomputes
vanilla on None), but the warning is a false alarm for unscaled models.
Compute the unscaled inv_freq directly for rope_type "default"/None instead of
going through ROPE_INIT_FUNCTIONS, so plain configs return the right value with
no warning. Scaled types (llama3/linear/yarn/...) are unchanged.
Also skip test_object_style_rope_scaling_on_config_delegates_correctly when
transformers strict-validates rope_scaling (5.x): it rejects a non-dict object
on config.rope_scaling, so the object-style delegation path cannot be set up
there. The test still runs and asserts on transformers <5.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* fix: deduplicate lemonade ROCm prebuilt selection log
resolve_lemonade_rocm_choice() is called twice per install (direct
planner + resolve_upstream_asset_choice). The API fetch is already
memoised via _fetch_lemonade_release_cached but the selection log
lines were still emitted on both calls, printing the 'trying
lemonade-sdk ROCm prebuilt' banner and hash-manifest NOTE twice.
Add _lemonade_selection_logged set keyed on (gfx_target, asset_name)
and guard the two log() calls behind a membership check so they print
exactly once per process regardless of call count.
Also extend the _clear_lemonade_release_cache test fixture to clear
the new set between tests to prevent cross-test state bleed.
Fixes#6020
* fix: write log() output to stdout to avoid PowerShell NativeCommandError
On Windows, PowerShell treats any stderr output from a native process as
an error record and prefixes it with 'python.exe :' and sets the
ErrorId to NativeCommandError. Since log() wrote to sys.stderr, every
[llama-prebuilt] status line triggered this, making normal progress
output look like errors in the installer console.
Switch log() to sys.stdout. The download progress bar (DownloadProgress)
retains its stderr/tty logic unchanged -- that path is for interactive
terminal rendering, not status logging.
* fix: remove redundant 'or ""' in lemonade log_key
host.rocm_gfx_target is already guaranteed truthy by the early
return at the top of resolve_lemonade_rocm_choice. The fallback
was dead code.
* Keep resolver stdout machine-readable, route install logs to stdout
log() sending everything to stdout breaks the resolver modes: setup.sh
json.load()s the whole stdout, so one helper log line (network retry,
release-tag scan) corrupts the parse and silently drops back to building
"latest". Default log() to stderr and flip to stdout only on the install
path, where PowerShell otherwise renders stderr as NativeCommandError
noise. Also tighten the lemonade dedup comments.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
On a 0.0.0.0 bind whose public ip:port is not reachable (cloud firewall),
the banner still printed "Secure link access via Cloudflare: <url>" right
after "is NOT reachable from the public internet", which reads as if the
tunnel might also be blocked. The Cloudflare quick-tunnel works regardless.
Thread the reachability probe result through a module-level _public_reachable
tri-state and, when the public probe definitively failed but the tunnel is
up, print "Also, the secure link access via Cloudflare works: <url>".
Reachable or undecided cases keep the existing wording.
PyPI release unsloth 2026.6.3 is now live. Bump the pinned floor in
install.sh and install.ps1 from unsloth>=2026.6.2 to unsloth>=2026.6.3
so fresh installs resolve to the new wheel.
* Studio: llama.cpp update banner redesign, About tab license info, inline system prompt editing, naming cleanup
- Redesign the llama.cpp update banner to match the chat composer surface
(borderless rounded card, composer shadow, Hellix Medium title), rename
actions to Update and add a 15 minute Remind me later snooze
- Keep the banner up until the user explicitly acts on it; drop the
outside click dismissal
- Add a Settings > General > Notifications toggle to disable the banner
for training-only setups (on by default)
- Rename the Help settings tab to About and add a License section
(Unsloth Studio AGPL-3.0, Unsloth Core Apache-2.0) linking to the
license files in this repo
- Make the run settings system prompt box an inline editable textarea;
the popup editor opens when the prompt overflows the box
- Pointer cursor on the preset dropdown chevron
- Dark mode toasts use the chat composer surface color
- Replace standalone Studio with Unsloth in user facing strings; keep
Unsloth Studio, LM Studio, Fine-tuning Studio, Recipe Studio and CLI
commands unchanged
* Studio: open the system prompt popup on box click, balance banner padding
- The system prompt box opens the Edit System Prompt dialog on click,
matching the pencil action
- Slightly more bottom padding on the llama.cpp update banner so the
spacing reads even next to the action pills
* Studio: replace unsloth studio update with the installer commands in update guidance
- The unsloth studio update command no longer works, so the About tab
update section now shows the one-line installer (curl or irm) for
PyPI and unknown installs, and git pull plus the local installer for
checkouts
- Add a short note that unsloth studio update is no longer supported
- Link the Installation, Updating and Windows install docs pages
- The package update banner now copies the platform installer command
instead of unsloth studio update
* Studio: rounder account menu, inline system prompt box with popup from the label
- Account menu corners go from 14px to 18px via a specific override,
since list menus pin border-radius globally
- llama.cpp banner bottom padding 22px
- System prompt is an inline editable textarea again; clicking the
System Prompt label opens the popup editor, and an overflowing
prompt opens it on box click
* Studio: show the standard install commands in the About update section
- Both one-line install commands (MacOS/Linux/WSL and Windows
PowerShell) are always shown, labeled like the docs, since running
them again updates an existing install
- Drop the unsloth studio update deprecation note
- Add the Mac install guide to the docs links
* Studio: clearer platform toggle and layout in the About update section
- Section heading is Update
- Platform picker is a pair of pill buttons, MacOS / Linux and Windows,
and only the selected platform's install command is shown
- Intro reads: To install or update Unsloth
- Local update heading separates checkout guidance from the standard
install command
* Studio: report GitHub branch instead of dev for source checkouts
A source checkout not on an exact release tag now shows
GitHub <branch> (e.g. GitHub main) as the Studio version in About.
Detached or unusual HEADs still fall back to dev.
* Studio: tighten the About update section copy and toggle styling
- Platform toggle buttons are borderless pills
- Shorter local update wording and restart note
- Docs links read Mac and Windows
* Studio: tighten line spacing in the sidebar account button
* Studio: fix vanishing compact MCP icon on hover, single line pill tooltips
- Compact caret pills (MCP, RAG) keep their icon on hover for inactive
pills too; the off switch hover rules hid the icon while compact mode
hid the X, leaving an empty slot
- Compact icon tooltips and single line compact tooltips render as full
pills; wrapped tooltips keep the 9px corners. TooltipContent measures
line count in a ref callback since Radix mounts portal content
without re-rendering the wrapper
- 1px gap between the name and Unsloth lines in the sidebar account
button
* Studio: Projects hover plus button, align recents with the label
- Hovering the Projects nav item reveals a plus button that opens the
New project dialog, with the same circular hover treatment as the
chat row actions
- Recent chat titles start at the same x as the Recents label
- The system prompt overflow lock only engages for a non-empty prompt
with a laid-out box, so a mis-measure cannot turn clicks into the
popup
* Clip system prompt overflow inside the rounded box
Wrap the inline system prompt textarea in a rounded overflow-hidden
surface so scrolled text and the scrollbar stay inside the box. The
focus ring moves to the wrapper via focus-within.
* Add updating progress bar to llama banner and shorten settings copy
While an update is applying, the banner action row becomes an
indeterminate progress bar that keeps animating under reduced motion,
matching the other loading indicators. Settings descriptions across
General, Profile, Appearance, Chat, Connections, API, and About are
trimmed without losing meaning.
* Address review: desktop update note, server platform detection, zh-CN keys
The About tab no longer shows terminal install commands in the desktop
app, where the bundled backend updates through the built-in updater;
it shows a short note and the docs links instead.
fetchDeviceType now sends the auth token to /api/health, which only
reports the server platform to authed callers, and caches only a
server-reported value. Copied install commands then match the host
platform rather than the browser when they differ (WSL, SSH).
zh-CN gains translations for the new notification and license keys,
the renamed About tab title, and the desktop update note.
* Real download progress for llama.cpp updates, prompt and sidebar polish
The update worker now streams the installer output and parses its
download percent lines into job progress, exposed via the update-status
API. The installer emits finer non-tty milestones when
UNSLOTH_PROGRESS_PERCENT_STEP is set; the worker requests 5 percent
steps. The banner renders a determinate bar from the reported fraction
and falls back to the sweep until the first percent arrives.
Also removes the focus ring on the inline system prompt box and
slightly shrinks the Projects hover plus icon.
transformers >= 5.x makes _is_package_available always return a
(exists, version) tuple, which is truthy even when the package is
absent. The flash_attn and vLLM availability checks treated the result
as a bool, so they always entered the "package present" branch:
- flash_attn (CUDA + HIP): the inner import raises when flash-attn is
not installed, printing a false "Flash Attention 2 installation seems
to be broken" warning before falling back to xformers.
- is_vLLM_available(): always reported vLLM as installed.
Add a small _package_available() helper that normalises the result to a
bool (handling both the new tuple and the legacy bool return), and route
the three call sites through it.
Fixes#6155
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* fix(rocm): stop overwriting ROCR_VISIBLE_DEVICES in apply_gpu_ids
ROCR_VISIBLE_DEVICES uses HSA agent-level indexing, not physical GPU
indices. Setting it to a bare integer breaks multi-GPU ROCm systems
where the parent already set ROCR_VISIBLE_DEVICES=0,1: narrowing to
1 causes torch.cuda.is_available() to return False in the training
worker, producing a misleading 'no HIP accelerator' error even on a
correctly configured ROCm host.
HIP_VISIBLE_DEVICES is sufficient for GPU selection on ROCm.
Leave ROCR_VISIBLE_DEVICES inherited from the parent environment.
* test(rocm): update apply_gpu_ids test to assert ROCR_VISIBLE_DEVICES is not overwritten
• fix: handle empty responses tool output
Normalize empty Responses `function_call_output.output` values before converting them into Chat Completions `role="tool"` messages. Empty strings, whitespace-only strings, and empty arrays now use the existing no-output sentinel, while non-empty text and content arrays are preserved.
Add regression coverage for empty tool outputs, image payloads outside `output`, content-array serialization, validator round trips, and preserving non-empty text.
---------
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
Co-authored-by: Tai An <antai12232931@outlook.com>
Co-authored-by: Datta Nimmaturi <venkatadattasainimmaturi@gmail.com>
When pinning GPUs for the llama-server child, the ROCm path set both
HIP_VISIBLE_DEVICES and ROCR_VISIBLE_DEVICES to the same physical
indices. These masks filter at different layers and stack:
ROCR_VISIBLE_DEVICES reduces the visible set at the HSA/ROCr layer and
re-indexes from 0, then HIP_VISIBLE_DEVICES indexes into that reduced
set. _select_gpus ranks by free VRAM and picks the most-free card, so a
single non-zero pin (e.g. "1") becomes out of range at the HIP layer,
HIP enumerates 0 devices, and the model silently runs on CPU
("ggml_cuda_init: failed to initialize ROCm: no ROCm-capable device is
detected").
Set only HIP_VISIBLE_DEVICES (which narrows correctly on its own) and
clear any inherited ROCR mask so it can't double up.
Verified on a 2x Radeon AI PRO R9700 (gfx1201) host, ROCm 7.1.1: the
same selected=[1] load that fell back to CPU (~7.7 tok/s) now runs on
the GPU (~78 tok/s).
Fixes#6175
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* studio: import MCP servers from a config file
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* import config' on the add-server form
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix: defensively handle MCP config imports
* fix: address MCP import review follow-ups
* fix: preserve apostrophes in Windows MCP commands
* fix: preserve apostrophe-wrapped Windows MCP args
* fix: align Windows MCP parsing with list2cmdline
* fix: preserve explicit MCP remote transport intent
* fix: trim MCP remote URLs before transport checks
---------
Co-authored-by: Roland Tannous <rolandtannous@gravityq.ai>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* Studio: add `unsloth chat` CLI command
Interactive chat REPL on the shared Studio backend: trained-model picker
when no model is given, /think and /compare toggles (adapter toggle on
CUDA, side-by-side base-model load on MLX), markdown streaming, and
connect-if-running Studio server mode so models stay warm across
sessions and are shared with the UI.
* fix settings
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* fix error handling and compare base precision
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix chat CLI backend imports and GGUF drafter loading
* Hide split thinking tags in chat CLI streams
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: imagineer99 <samleejackson0@gmail.com>
* fix(studio): surface live step with null loss through the SSE progress stream
The metric histories skip non-finite steps, so during a NaN stretch the
SSE live loop and final complete event replayed the last finite
step/loss pair. Follow the live progress step when it is ahead of the
history tail and report its loss honestly (null until recovery).
Completes the NaN honesty fix for the SSE consumer flagged in review.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Apply live-step handling to inactive streams and clear the UI loss on null for PR #6206
Fresh /progress connections after a finished run took the inactive branch
which still replayed the last finite step and loss pair; apply the same
live-step correction there. On the frontend, applyProgress kept the stale
currentLoss when a payload advanced the step with a null loss; clear it so
the display shows -- until the loss recovers. Widen the runtime state type
to number | null, which the view layer already handles.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: report the real llama-server context window and add an opt-in overflow policy for OpenAI-compatible serving
A community report showed OpenCode failing tool calls every few minutes
against Studio's OpenAI-compatible API while the same GGUF was stable on
LM Studio. Root cause: Studio advertises the requested context length, but
llama-server can allocate less (memory-fit step on small GPUs, --parallel
slot split), so clients budget against a window that does not exist. Their
generations truncate mid tool call at the real wall (finish_reason=length
with cut JSON arguments) and eventually the prompt itself exceeds the real
window, returning a 400 that agentic clients treat as non-retryable.
Changes:
- After llama-server health, read default_generation_settings.n_ctx from
/props and adopt it whenever it is below Studio's computed context, with
a warning. The load response, status route, UI value, and the passthrough
max_tokens ceiling all become honest automatically.
- Expose context_length and max_context_length on /v1/models so clients can
budget against the enforced window.
- Accept empty role=tool content (commands with no output are routine in
agentic loops; OpenAI and llama-server both accept it) instead of a 400.
- Add context_overflow=truncate_middle (per request, or server-wide via
UNSLOTH_CONTEXT_OVERFLOW=truncate_middle): on exceed_context_size_error
the passthrough drops whole middle turn-groups (system prompt, first turn,
and recent turns kept; tool calls stay paired with their results), clips
oversized contents middle-out when group-dropping is not enough, clamps
max_tokens to the generation headroom, and retries. Default stays 'error'
with code=context_length_exceeded so clients running their own compaction
keep full control.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: allocate the requested context for real (kv-unified, fit-ctx floor)
Two launch-flag gaps caused the advertised vs allocated divergence at the
source:
- llama-server enables --kv-unified only when the slot count is auto; Studio
always passes --parallel N, which silently splits -c into per-slot windows
of -c/N. Pass --kv-unified when N > 1 so a single request can use the full
advertised window (same total KV memory, shared pool).
- with --fit on the fit step may set ctx as low as 4096; pass
--fit-ctx <requested> for explicit requests so fit offloads or fails into
the existing --fit off retry instead of silently shrinking the window.
Both flags are gated on --help capability probing so older builds keep the
current behavior, where the /props readback remains the backstop. Verified
live: -c 98304 --parallel 4 now serves per-slot n_ctx 98304 (was 24576),
48k-token requests pass through the passthrough, and the readback warning no
longer fires.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix config.rope_scaling being dropped by the replaced rotary embedding (#2405)
On modern transformers, LlamaModel builds its rotary embedding from config
using unsloth's replacement LlamaRotaryEmbedding class, whose config path
computed vanilla inv_freq and ignored config.rope_scaling entirely. The
llama3/linear/longrope dispatch in patch_llama_rope_scaling rewrites
LlamaAttention.__init__, which no longer constructs rotary embeddings, so it
never fires; the model-level rotary is then copied onto every attention
layer. Result: Llama-3.1/3.2/3.3 ran with unscaled RoPE on the
FastLanguageModel path and collapsed into repetition loops past roughly 29K
tokens (PASS at 28867, FAIL at 31767 in needle retrieval). FastModel was
unaffected because vision.py keeps transformers' own rotary. qwen2, qwen3,
qwen3_moe, mistral and cohere assign the same base class, so any rope-scaled
config of those families was equally exposed.
The fix makes the base class config path compute inv_freq and
attention_scaling via transformers' ROPE_INIT_FUNCTIONS (covers llama3,
linear, dynamic, yarn, longrope), with an inline llama3 fallback reading
factors from config for older transformers, degrading to prior behavior on
any failure. attention_scaling is applied in _set_cos_sin_cache (1.0 default,
exact no-op for unscaled paths) and persists across extend_rope_embedding.
A type(self) guard prevents double-scaling via the legacy scaled subclasses.
Adds tests/utils/test_rope_scaling_drift.py (AST tripwire + behavioral
inv_freq/cos-cache/extension checks, validated to fail 4 of 5 on the unfixed
code) and wires it into the existing consolidated CI HARD GATE step.
Verified on GPU: 48K-token needle retrieval flips FAIL to PASS for
FastLanguageModel in bf16 and 4bit, 20K stays PASS, scaled inv_freq matches
transformers exactly, and the left-padded batch generation guard still gets
exact solo-vs-batched token matches.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: normalize object-style rope_scaling, vectorize llama3 fallback
config.rope_scaling can be a config object rather than a dict on newer
transformers; _rope_scaling_as_dict normalizes it (to_dict/dict/vars
fallbacks) before any .get() access, with a regression test using a
dataclass stand-in. The inline llama3 fallback now uses torch.where instead
of a per-frequency Python loop; verified bit-for-bit equal to transformers
ROPE_INIT_FUNCTIONS for factor 8 (Llama-3.1) and factor 32 (Llama-3.2).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: CPU-safe rope guard tests, normalized config for delegation
The rotary constructor builds per-device CUDA caches, so the behavioral tests
that instantiate it cannot run on GPU-less CI. Restructured into three layers:
the AST tripwire now also asserts the constructor stays wired to
_compute_config_rope_inv_freq; the CPU layer tests that pure helper directly
(llama3 dict, llama3 object, linear object, default type) with no
instantiation; the instantiation and cache tests are gated behind a real CUDA
probe (actual tensor allocation, so import-time CUDA spoofs cannot fool the
gate). Verified: 9 passed with GPU; 5 passed 4 skipped with CUDA hidden; 5
failed 4 skipped on the unfixed code in CPU mode.
Delegation to ROPE_INIT_FUNCTIONS now retries with a shallow config copy
carrying the normalized rope_scaling dict when the original was an object the
installed transformers cannot read; covered by a linear-object test, which has
no inline fallback and passes only through that retry path.
* Tighten comments in rope scaling fix and guard test
Comment and docstring reduction only; verified code-identical with
scripts/comment_tools.py check --strip-docstrings (AST signature match on
both Python files). All guard tests unchanged: 20 passed with GPU, 5 passed
4 skipped with CUDA hidden.
* Apply repo kwarg-spacing format
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Studio's frontend exposes a Resume action and submits requests with
resume_from_checkpoint set to a previous run's output_dir. The CUDA
training paths in worker.py read this field from config and pass it to
trainer.train() (see lines 2729-2787 and 3108-3229). The MLX path
_run_mlx_training did neither: it never read config['resume_from_checkpoint']
and called trainer.train() with no args. The MLX trainer also did not
accept the kwarg, so even threading it through would have been a no-op.
With this PR + the unsloth-zoo companion PR adding the trainer-side
support (saves optimizer_state + trainer_state, accepts and applies
resume_from_checkpoint in MLXTrainer.train()), MLX Resume now works
end-to-end. Verified on M2 16GB with Qwen3-0.6B + unsloth/LaTeX_OCR:
loss at every post-resume step matches a fresh run bit for bit
(2.168627977371216 == 2.168627977371216 at step 6, etc).
Two lines: read the field near the other config.get() extractions in
_run_mlx_training, pass it as a kwarg at the trainer.train() call site.
Companion PR: unslothai/unsloth-zoo#751
When training produced a NaN or Inf loss event, the handler filtered the
value to None but never updated progress.loss — clients kept seeing the
last finite value as if everything were fine.
Now: on non-finite loss, clear progress.loss to None and log a one-shot
warning. Training continues (no phase=error, no _should_stop), matching
the expected behavior for a non-fatal numerical event.
Test: tests/test_training_nan_loss_handling.py with 6 cases covering
finite, NaN, +/-Inf, idempotency of the one-shot warning, and recovery
when a finite step follows a non-finite one.
* Studio: auto Cloudflare tunnel for 0.0.0.0 launches
Binding Studio to 0.0.0.0 for remote access often leaves the raw
http://<ip>:<port> URL unreachable (https-vs-http, blocked high ports,
closed cloud security groups). On a wildcard bind, auto-start a free
cloudflared quick tunnel and show its https://*.trycloudflare.com URL in
the startup banner:
Secure link access via Cloudflare: https://<random>.trycloudflare.com
- new studio/backend/cloudflare_tunnel.py: find or download+cache the
cloudflared binary (per-OS/arch GitHub release, safe .tgz extract),
start the tunnel, parse the URL, tear it down. Stdlib only; best-effort
and non-fatal throughout (a missing binary or offline box never blocks
or slows startup).
- run_server starts the tunnel for 0.0.0.0 only (skips loopback, api-only
and Colab), prints the line in the banner, and _graceful_shutdown stops
the child so it never orphans.
- --cloudflare/--no-cloudflare flag (default on) on `unsloth studio` and
`unsloth studio run`, forwarded through the re-exec into run_server.
- tests for the helper, the CLI flag forwarding, and the run.py defaults.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio cloudflare: send a User-Agent on the cloudflared download
GitHub's CDN can 403 the default Python-urllib User-Agent on release asset
downloads. Set an explicit UA and pin it with a test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio cloudflare: address review (opt-out for subcommands, tunnel teardown)
- reject --no-cloudflare placed before a subcommand (it would not reach the
subcommand), mirroring the --parallel guard
- register the tunnel before waiting for its URL so a shutdown during the wait
stops cloudflared instead of orphaning it
- tear the server + children down if `unsloth studio run` startup aborts
(health timeout, model-load error, Ctrl+C) before the wait loop
* [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>
* Strip trailing whitespace in construct_chat_template (fixes#992)
construct_chat_template() only lstrip()s the template, so a template ending
in a newline (e.g. the default Llama-3 template + a trailing '\n') keeps the
trailing whitespace. That breaks the rfind-based repeated-example detection
and, on current main, also the regex fallback, which then raises
"Could not recover a two-example structure from chat_template".
Surrounding whitespace is not significant to the template here (the left side
is already stripped), so strip() both ends. Verified that the valid no-newline
template still parses via the primary path and that the meaningful trailing
token (e.g. <|eot_id|>) is preserved.
Co-Authored-By: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
* Fix trailing whitespace handling via parse retry for PR #6008
---------
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* fix(studio): reuse venv Python in setup instead of re-probing system
* Reuse venv Python for studio setup
Pass the venv interpreter from install.ps1 to studio/setup.ps1 via UNSLOTH_SETUP_PYTHON and prefer it over probing the system. Added Resolve-ReusedSetupPython to accept the handed-off path (or derive the venv python when setup runs standalone), validate it (Python 3.11–3.13 and non-conda), and inject its Scripts dir onto PATH. When a reused interpreter is accepted, py.exe enumeration and further system probing are skipped. install.ps1 also sets the env var before running setup and removes it on cleanup to avoid leaving state behind. This prevents setup from being tripped by unsupported Python 3.14 or Windows Store stubs on PATH.
* Harden setup Python detection for PR #6033: py -All, shared conda check, bare ~ guard
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
* center the search dialog and change the wrong borders
* fix the mistake of 1 to l
* Fix/adjust search dialog radius for PR #6184
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: wasimysaid <wasimysdev@gmail.com>
* Stop HF 429 rate limits from sinking the llama.cpp prebuilt path in Studio CI
The Windows Studio API smoke job failed when anonymous huggingface.co
fetches of the tiny GGUF validation model (stories260K.gguf) hit HTTP 429
on the shared runner IP. The installer correctly refused the unvalidated
prebuilt and fell back to a source build, which the prebuilt assert then
flags. Three layers fix this:
1. Installer: auth_headers sends HF_TOKEN (or HUGGING_FACE_HUB_TOKEN) to
huggingface.co hosts, mirroring the existing GH_TOKEN handling for the
GitHub API rate limit. A redirect handler strips Authorization when a
download is redirected off-host (CDN signed URLs reject foreign auth;
urllib forwards headers on redirect, unlike requests/huggingface_hub).
2. Workflows: the HF_HOME prime steps also prefetch the validation model
so the install's hf_hub_download resolves from the local cache even
when the Hub is rate limiting; cache keys bumped v1 to v2 to repopulate.
This also covers fork PRs, which cannot see secrets.
3. Workflows: every Install Studio / update step that already passes
GH_TOKEN now also passes HF_TOKEN, so both the huggingface_hub path and
the direct URL fallback are authenticated.
Tests: tests/studio/install/test_hf_auth.py covers token-to-host routing,
the cross-host redirect strip, and the download_bytes wiring (offline).
Verified live: authenticated download of the validation model through the
new opener (CDN redirect exercised, pinned sha matches) and an offline
hf_hub_download cache hit against an HF_HOME primed by the new step.
* [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>
* Require a found ROCm DLL before forcing BNB_ROCM_VERSION in Studio paths
main.py previously set BNB_ROCM_VERSION=72 whenever HIP_PATH or ROCM_PATH
was set, and the training worker fell back to a blind 72 when DLL
detection found nothing. On a Windows machine with the AMD HIP SDK
installed but CUDA or CPU torch, that forces a ROCm backend onto a
non-ROCm bitsandbytes wheel, which raises at import. Both paths now only
write the override when a libbitsandbytes_rocm DLL actually exists (or a
seeded value is already present), matching the strict gates in
unsloth/import_fixes.py.
Also removes four redundant local import shutil statements in
unsloth/save.py that shadow the module-level import, the same pattern
that caused the UnboundLocalError fixed in #6149.
* Worker: gate the BNB override on a found ROCm DLL, preserving seeded marker
Review follow-ups: track _found_rocm_bnb in the worker like main.py so a
ROCm DLL with an unparsable name still gets the seeded or 72 fallback,
and skip the env write entirely when no DLL exists so a seeded value
keeps its sitecustomize marker and stays redetectable by later import
fixes.
* Fix bitsandbytes ROCm GPU arch and warp size detection on Windows
bitsandbytes resolves the ROCm GPU architecture (and warp size on
0.49.x) by shelling out to rocminfo / hipinfo.exe via PATH at import
time. On Windows neither tool is normally on PATH (AMD torch wheels
ship hipInfo.exe into the venv Scripts dir, only on PATH while
activated), so every `import bitsandbytes` logs an ERROR and WARNING,
ROCM_GPU_ARCH degrades to unknown, and the 0.49.x warp size defaults
to 64, which is wrong on RDNA (wave 32) and silently disables
pre-quantized 4-bit models via ALLOW_PREQUANTIZED_MODELS.
Install a one-shot MetaPathFinder before unsloth_zoo is imported (the
first bitsandbytes import on ROCm) that swaps get_rocm_gpu_arch and
get_rocm_warpsize for torch-device-properties-first implementations
right after bitsandbytes.cuda_specs executes, before cextension reads
them. Falls back to running hipInfo.exe by absolute path (venv
Scripts, conda Scripts, HIP SDK / AMD installer dirs). Repairs the
constants in place when bitsandbytes was imported first. Strict no-op
on non-Windows, non-ROCm builds, missing bitsandbytes, and versions
that fix this upstream. Opt out with UNSLOTH_DISABLE_BNB_ROCM_FIX=1.
Proposed upstream in bitsandbytes-foundation/bitsandbytes#1969;
shipped here so all bitsandbytes versions are covered. Verified on
gfx1151 Strix Halo, Windows 11, torch 2.11.0+rocm7.13.0 against
bitsandbytes main, 0.49.2, and a torch-props-fixed variant.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the bitsandbytes ROCm detection fix
Comment and docstring pass only. AST comparison with docstrings
stripped confirms every definition is identical to the version the
12 scenario suite ran against, and the suite plus the drift test
pass unchanged on the edited files.
* Keep the bitsandbytes cuda_specs finder installed for reload support
Simulation testing caught a regression in the one-shot design:
importlib.reload(bitsandbytes.cuda_specs) re-resolves the spec through
sys.meta_path, so with the finder already removed the reload reinstalled
the unpatched upstream detector and the Windows ROCm noise returned.
Keep the finder on sys.meta_path permanently, matching the lifecycle of
the existing causal_conv1d and vllm import blockers. The finder matches
a single module name and patching stays idempotent via the sentinel
flags, so repeat hits are no-ops.
Validated on gfx1151 Windows 11: 22 simulation scenarios (conda and
embedded layouts, Program Files scan ordering, paths with spaces and
unicode, hanging probe timeout, lru-wrapped and C-function helper
shapes, reload, failed-import retry, threads, spawn, dormant finder,
Studio PATH coexistence, early fix-block ordering, bnb 0.45.5 / 0.47.0
/ 0.49.2 / main / upstream-fixed) plus the original 12 scenario suite,
CPU-torch and stale-HIP_PATH sandboxes, Python 3.10 to 3.13 gates, and
a WSL Linux leg proving byte-identical Linux behavior with and without
the fix, with and without rocminfo on PATH.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: surface the llama.cpp update affordance when MTP is disabled
When a model asks for MTP (auto on an MTP model, or forced mtp / mtp+ngram)
but it gets disabled, the load already degrades gracefully and serves without
speculative decoding. Until now the UI gave no hint why, or that an update
would fix it.
Record why MTP was dropped on the backend (spec_fallback_reason): the probe
found no mtp token (binary_no_mtp), the spawn aborted with an outdated-arch /
context-build error such as a prebuilt that predates the Gemma drafter
(binary_outdated), or the current build could not run it, e.g. a CUDA kernel
limit (runtime_error). Expose it in the inference status. In the chat
Speculative Decoding section, show a short note and, for the two update-fixable
reasons, an inline Update llama.cpp button that reuses the existing update flow.
A runtime_error gets the note without an update push, since a newer build may
not fix it.
Backend tests cover the reason being set / cleared. Frontend typechecks.
* Address review: tighten the update hint to genuinely outdated binaries
Reserve binary_outdated (which surfaces the Update llama.cpp affordance) for an
unknown-architecture abort, which proves the prebuilt predates the model;
classify the generic memory/context build failures as runtime_error, where an
update may not help. Frontend: only append the "Update llama.cpp to enable it"
sentence when an update is actually available, so the text never points at an
action the UI is not offering.
* fix(studio): prevent UI freeze when switching tabs from heavy pages
Change AnimatePresence mode from "wait" to "popLayout" to fix issue
where switching tabs from Export (or other heavy pages) would cause
the URL to update but the UI to freeze.
With mode="wait", the exit animation must complete before the new
component mounts. If the exiting page has expensive computations,
this blocks the UI. mode="popLayout" allows the new route to mount
immediately while the old one animates out.
Fixes#5850
* fix: add relative positioning for popLayout mode
AnimatePresence mode='popLayout' applies position: absolute to the
exiting element, so the parent container needs position: relative
to prevent layout jumps during transitions.
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
The export route's load_checkpoint waits for the training subprocess to
exit by calling time.sleep(0.5) in a loop (up to 30s) inside an async
function. time.sleep blocks the whole event loop, so every other request
to the server stalls for that duration. Use await asyncio.sleep(0.5),
matching the async pattern already used elsewhere in this file
(asyncio.to_thread, await asyncio.sleep).
Co-authored-by: Wasim Yousef Said <wasimysdev@gmail.com>
The kwarg-dispatch in unsloth_base_fast_generate recognized input_ids,
input, input_features and the misspelled input_embeds, but not HF's
canonical inputs_embeds. So generate(inputs_embeds=...) fell through to the
'first kwarg' fallback, which picks whatever kwarg happens to come first
(e.g. attention_mask) and uses it as input_ids -- giving the wrong tensor /
batch size, or the KeyError reported in #3082 on older versions.
Add an explicit inputs_embeds branch so embedding inputs (e.g. multimodal
audio+text) are routed correctly regardless of kwarg order.
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
* fix(studio): fall back to copy when os.replace is blocked during install activation
On Windows ARM64 the antivirus scanner can transiently hold a freshly
extracted DLL open while MoveFileEx runs, so activating the staged
llama.cpp prebuilt fails with [WinError 5] Access is denied. Attempt
os.replace first, then fall back to a file-by-file copytree which
bypasses the rename.
* address review: keep os.replace for rollback, scope copy-fallback to staging
The copy + rmtree fallback could silently corrupt a live install if the
existing directory is busy. Restrict it to freshly extracted staging
trees (renamed activate_staged_dir) and keep strict os.replace for the
rollback move so a busy active install raises immediately.
* fix(studio): scope copy-fallback to busy-lock errors, log it, and add tests
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
