Comment-only pass over the Python this PR touches: drop what the code already
says, collapse multi-line explanations that still read on one line, and keep
the reasoning that is not recoverable from the code. No code, docstring
semantics or behaviour changes; verified with an AST comparison against the
previous revision, and the backend suite is unchanged (same 37 environment
failures as before: the API integration tests that need a live keyed server,
the flash-attn install hooks, and the GPU memory fields).
Reports the repos and exact files a pick needs so the download manager can stage
them with the loader's own file scope. A plain snapshot would add the packaged
root single, transformer shards and fp16 twins the loader never opens.
Four bugs in the batched inference path, all found by review:
- A mixed-prompt batch sent a scalar negative prompt against a prompt
list. Z-Image asserts on the length, and Qwen-Image, Krea 2 and FLUX
true-CFG encode a batch-1 negative against batch-N latents and fail in
the transformer's text/image concat. Broadcast it to match the batch.
- The FBCache step-cache reset sat above the chunk loop. diffusers only
resets that state at the end of a successful call, so a forward that
raised (the OOM the backoff is meant to recover) left its own residual
behind and the halved retry died on a shape mismatch. Reset before
every forward instead.
- The conditioning cache keyed on the checkpoint alone, but a GGUF or
single-file load takes its text encoders from the companion base, so
the same checkpoint against a different base reused the previous
base's embeddings. Key the base too.
- Gallery records stored the base seed and the requested batch size even
when a prompts/seeds list drove the run, so restoring the second image
of seeds=[5, 99] replayed seed 5. List-driven outputs now record as
single-image recipes on their own seed.
Also bound strength above 0: every img2img pipeline derives its step
count from it, so 0 leaves zero denoising steps and either raises or, on
SDXL, crashes on empty latents.
Resolve the drift between PR #6763 and current main:
- deletion: main moved cached-model deletion into hub/services/models/deletion.py,
so the PR's Images/Video in-use guards move there too as _diffusion_blocks_delete
and _video_blocks_delete, keeping main's fail-closed 503 contract.
- llama_keepwarm: take main's rewrite, re-apply the PR's image/video inference
suffixes so a generation in flight blocks an idle unload.
- routes/training: keep main's sidecar-swap 409 and resume_source_run_id, run
start_training in the worker thread the PR's unload hook needs.
- model picker: main renamed components/assistant-ui/model-selector ->
features/model-picker/... and rewrote pickers.tsx, so the PR's picker work is
ported onto main's version (task/catalog props, task gating of hub + cached +
local rows, single-device expanderGpuGb, fine-tuned section hidden when scoped)
rather than reverting main's pinned-models and per-model-config work.
- images/video pages: imports repointed at the new model-selector path.
- tests: delete-guard tests retargeted at the deletion service.
Typecheck, i18n parity and model-catalog checks pass.
Fixes#7244
The Studio per-model config dropdown only surfaced bf16, q8_0, q5_1,
and q4_1 even though llama.cpp already accepts q4_0, q5_0, iq4_nl, and
f32. Add the missing options to KV_CACHE_DTYPES and align API field
descriptions with the backend _valid_cache_types set.
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Roland Tannous <115670425+rolandtannous@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio: add Voice settings tab (dictation, dictionary, read aloud)
New Voice tab in Settings, placed just before About:
- Dictation: microphone picker, browser STT engine, recognition language,
and an inline mic test with a live transcript
- Dictation dictionary: entries rewrite matching speech to their exact
spelling and casing, applied in both dictation paths
- Recent dictations: last 20 final transcripts with copy and clear, so
text can be recovered if it lands in the wrong place
- Read aloud: optional button on assistant responses with two engines,
curated system voices (novelty and legacy voices filtered, quality
ranked, capped at 20) or the TTS audio model loaded in Unsloth via
/audio/generate (e.g. Orpheus), plus speed, pitch, volume and preview
Settings persist in localStorage (unsloth_voice_settings) and are read
at call time so changes apply without reloading the runtime. Adds en
keys plus the tab label for ja, zh-CN and pt-BR.
* Studio: drop the single option STT engine select, rename TTS option
The STT engine dropdown only had one entry, so it added noise without
giving a real choice. The engine row can come back once local STT
models land. Also renames the TTS engine option Unsloth TTS model to
Load TTS model to make the action clearer.
* Studio: harden Voice settings against edge cases found in simulation
Simulated the feature across Chromium, Firefox and WebKit plus node
level unit runs and backend contract checks. Fixes from the findings:
- Dictionary rewrite used a replacement string, so entries containing
dollar patterns corrupted transcripts (A$$AP became A$AP, $& injected
the match). Switched to the callback form of String.replace
- Persisted voice settings now validate types on hydration: non string
micDeviceId, dictationLanguage and ttsVoiceURI, and non boolean
ttsEnabled fall back to defaults instead of flowing into the UI
- Dictionary entries are trimmed, capped at 120 chars and re-sanitized
on hydration
- The Test dictation panel now falls back to the default microphone
when the saved device is unplugged, matching the composer adapter
Test coverage: 46 unit assertions (dictionary regex edge cases across
unicode, word boundaries and injection, voice curation for simulated
macOS, Windows and Linux voice inventories, corrupt storage merge),
13 backend contract checks against /audio/generate on an isolated
instance, and 60 browser assertions across the three engines covering
rendering, degradation without SpeechRecognition, curation in a real
DOM, dictionary persistence with unicode and dollar entries, the
no-model preview error path and corrupt localStorage recovery.
* Studio: address Voice settings review feedback
Verified each review comment before acting. Confirmed and fixed:
- Editing a dictionary entry was broken in two ways: the store trimmed
on every keystroke so spaces could not be typed, and clearing the
field deleted the entry and unmounted the input mid edit. Updates now
keep the raw value and a blur commit trims or removes the entry
- The unplugged mic fallback checked instanceof DOMException, but a
cross browser probe showed Firefox and WebKit throw
OverconstrainedError objects that are not DOMExceptions, so the
fallback never fired there. Matching on the error name now
- When the browser ended a dictation test on its own (silence timeout),
the mic stream stayed open. All recognition end paths now stop the
tracks and save the transcript through a single finalize path
- The studio TTS audio element now releases its WAV data URL as soon as
playback ends, fails or is cancelled
- Allow microphone now reports insecure contexts (no mediaDevices)
accurately instead of claiming access was blocked
- Voice tab copy moved into i18n keys per src/i18n/AGENTS.md, so locale
overlays can translate it; en is the baseline and parity passes
- unsloth_voice_settings added to the Reset all local preferences key
list so voice preferences obey the reset
- Non default microphones note that the system default is used when the
browser speech engine cannot bind a specific device, since browsers
without the start(track) overload ignore the argument silently
Re-ran the full simulation set after the changes: 46 unit assertions,
13 backend contract checks and 60 browser assertions across Chromium,
Firefox and WebKit all pass, plus a dedicated browser probe for the
dictionary editing behavior.
* Studio: use the chat mic icon in Voice settings for consistency
The Voice tab and its buttons used the hugeicons Mic02 glyph while the
chat composer uses a custom filled mic. Extract that composer icon into
a shared lib/mic-icon component, drop the duplicate inline copies in
thread.tsx and shared-composer.tsx, and use it for the Voice tab icon
and the tab's mic buttons so the microphone looks the same everywhere.
* Studio: address second round of Voice settings review feedback
Verified each new comment against the current code first. One item was
already fixed in the previous round (recording transcripts when the
browser ends a dictation test on its own). Confirmed and fixed:
- The microphone row showed a picker with generic names when browsers
enumerate unlabeled devices before permission, leaving no way to
grant access from the row. It now branches on whether labels are
visible and shows Allow microphone otherwise
- Compare chat dictation ignored the selected microphone. It now opens
the chosen device with the same fallback rules as the main adapter,
passes the track to recognition where supported and releases the
stream when recognition ends
- Closing the Voice tab cancelled the shared speechSynthesis even when
read aloud was playing a chat message. Cleanup now only cancels when
the tab owns an active preview
- Double clicking Start test could race two recognizers and leak the
first stream. A starting flag set before the getUserMedia await makes
start reentrancy safe
- Turning off the read aloud setting mid playback removed the only stop
control. The stop button now renders whenever a message is speaking
- When an engine lacks the start(track) overload, both dictation paths
now release the selected device stream before retrying with the
default microphone instead of holding it open
- Read aloud support no longer requires Web Speech synthesis: the
Unsloth TTS engine only needs audio playback, so it stays available
in WebViews without speechSynthesis, with a clear error if the system
engine is chosen there
Not addressed here: cancelling in flight backend TTS generation on
stop. The route runs generation in a worker thread without a
cancellation path, which is shared pre existing behavior with audio
chat generation and belongs in a backend change.
All suites re-run green: 46 unit, 13 backend contract and 60 browser
matrix assertions across Chromium, Firefox and WebKit, plus probes for
the unlabeled device branch and the double click race.
* Studio: drop empty and duplicate voiceURIs so the Voice tab never renders a crashing Select item
* Studio: guard dictation mic lifecycle in Voice test and Compare composer
Release a microphone opened after the component unmounts, and stop Compare
dictation on a permission or security failure instead of silently recording
from the default device, matching the main chat adapter.
* Studio: fix dictation and read-aloud lifecycle edge cases in Voice settings
- Join final dictation chunks with a space so recorded transcripts do not merge words
- Ignore a stale recognizer onend so a quick stop then restart is not torn down
- Use previewingRef so a double click on TTS preview does not orphan the first request
- Keep the read-aloud stop control visible when a new run starts while a message is spoken
- Stop the dictionary remove button from deleting an adjacent entry on a blur then click race
* Studio: trim redundant Voice settings comments
* Studio: fix Voice preview and Compare dictation edge cases
- Only cancel the shared speechSynthesis for a system-voice preview, so stopping
a Studio preview no longer stops an unrelated chat read-aloud
- Release the Studio preview audio and its WAV data URL on normal completion
- Iterate every finalized result in Compare dictation so batched phrases are kept
- Cap persisted recent dictations to the last 20 on hydration
* Studio: use clipboard fallback for recents and release failed preview audio
- Copy recent dictations via the copyToClipboard helper so the execCommand
fallback works in Safari and insecure http LAN contexts
- Release the Studio preview audio when play() rejects, not just on ended/error
* Studio: add local speech-to-text dictation engine
Add an offline dictation engine that transcribes with a local faster-whisper
model, alongside the existing browser (Web Speech) engine. The browser engine
streams audio to Apple or Google speech services and needs internet; the new
engine runs on the server, works offline, and drives any chat model without
evicting it (it loads in the backend process, separate from the model
subprocess). It also gives Firefox dictation, which has no Web Speech support.
Backend: a lazily-loaded, kept-warm faster-whisper sidecar and three routes
under /api/inference/audio (stt/status, stt/load, transcribe). faster-whisper
is torch-free, so this does not disturb the existing model stack.
Frontend: a Dictation engine setting (browser or local model), a curated model
picker with sizes, and MediaRecorder capture posted to the transcribe route.
The model warms automatically when the engine is selected, with live status.
* Studio: stream local STT transcription as you speak
Local dictation showed nothing until you stopped, because the whole clip was
transcribed once on stop. Now the growing recording is re-transcribed on a
fast pass every second and emitted as live interim text, with an accurate
final pass on stop. Partial recordings decode fine, and the model refines
earlier words as more audio arrives.
Adds an interim flag to the transcribe route (beam 1, no VAD) for the fast
preview pass; the final stop uses the accurate path.
* Studio: make local dictation stop instant and reliable
Stopping local dictation waited for a final network transcription before the
session ended, so the stop button did not flip and a second click ended the
session early and dropped the text. Now stop commits the live transcript
immediately, releases the mic at once, and ignores a second stop while
finalizing. Previews run more often so the committed text is current.
* Studio: record local dictation in short clips for reliable streaming
Re-transcribing a growing buffer every second got slower as it grew, flooded
the backend, showed stale words, and could leave the stop button stuck waiting
on a backlog. Record short independent clips instead and transcribe each once,
appending the text as you speak. Work per clip is bounded, so stopping is
prompt (with a hard timeout as a safety net) and long dictations stay smooth.
* Studio: dictate then transcribe once on stop, ChatGPT style
Local STT dictation streamed by re-transcribing the growing clip, which
was quadratic and saturated the backend (multi-second lag), and stop only
halted the recorder without releasing the mic, so it kept recording. Record
the microphone continuously, release it the instant the user stops, and
transcribe the whole clip once. Stopping is immediate and the transcript
lands in about a second. Also add the tiny model for the fastest option.
* Studio: surface dictation and read-aloud failures instead of failing silently
- Compare dictation reports microphone and speech-recognition errors via toast,
reusing the main chat adapter's describeMediaError and describeSpeechError
- Read-aloud toasts genuine model or synthesis failures while ignoring cancellations
* Studio: ChatGPT-style recording bar for dictation
Clicking the mic now drops the composer into a dedicated recording bar
with a live waveform, a discard (X) and a confirm (tick), instead of a
plain stop button. The tick stops recording and transcribes the clip;
the X throws the recording away and keeps whatever text was already in
the composer. The model adapter taps the mic with an analyser to drive
the waveform, and the router tracks the live session so the X can cancel
it without transcribing.
* Studio: transcribe dictation while speaking, ChatGPT layout
Match ChatGPT's recording layout: the bar now renders in place of the
input with the left plus button kept, the waveform in the middle, and
the discard and confirm buttons together on the right.
Cut the post-confirm delay by transcribing in the background as the user
talks. The audio is split at natural pauses (voice-activity detection off
the same analyser that drives the waveform) and each clip is transcribed
as it is cut, so confirming only has to finish the short final tail. The
model is also warmed when recording starts so the first run never pays a
cold load.
* Studio: ChatGPT waveform, hide tools while dictating, faster STT
Make the recording UI read like ChatGPT: the waveform is now a dense row
of round dots that rise into thin centered bars, and while dictating only
the plus button shows, with the mode badge and tool toggles hidden so the
bar is just the waveform and controls.
Speed up transcription: decode greedily (beam_size=1), which is several
times faster on CPU with negligible accuracy loss on short dictation
clips, and cap background segments at 6s so the final tail after confirm
stays short.
* Studio: finish ChatGPT voice bar and low-latency STT
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: full-width waveform with a timer that freezes on stop
Use the full-width waveform for the recording bar: brighter, bigger bars
that advance on a fixed cadence (keeping peaks between advances) so they
glide instead of racing by, inset from the composer edges. Keep a visible
timer and the green confirm button, matching the ChatGPT reference, and
freeze the timer and waveform the moment the user confirms.
* Studio: fix multilingual local dictation
* Studio: speed up dictation and release local STT
* Studio: harden dictation finalization and STT decoding
* Studio: restore Firefox dictation fallback
* Studio: add dictation history manager
* Studio: manage speech model downloads
* Studio: remove em dash from voice model label
* Studio: move dictation history into Voice
* Studio: source local STT from Unsloth Whisper models
Point the dictation STT sidecar and its Model Hub download entries at
Unsloth's Hugging Face Whisper repos (small, large-v3-turbo, large-v3)
and run them through Transformers, so Studio only ever downloads
Unsloth-uploaded weights. Drop faster-whisper and the Systran/mobiuslabs
repos; keep the Model Hub as the only download path via local_files_only,
and keep PyAV for audio decoding.
Device selection uses float16 on CUDA and float32 on MPS and CPU, since
Whisper's decoder is unstable in float16 on MPS and repeats tokens.
Shorten the model picker labels to name plus download size and update the
STT tests for the new backend.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: smooth dictation waveform and keep pill height
* Studio: align STT model dropdown width and tidy voice copy
* Studio: guide to local engine when browser dictation is offline
* Studio: clarify voice section and STT model copy
* Studio: keep STT warm with training-aware eviction
* Harden STT lifecycle and browser compatibility
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model discovery test lint
* Harden cross-browser microphone errors
* Harden cross-browser microphone errors
* Surface voice test recognition errors and fall back to Studio TTS
- Voice test now toasts non-abort speech-recognition failures instead of
ending silently, matching the main and Compare dictation paths.
- Read-aloud routes to the backend model when the runtime lacks Web Speech
synthesis (audio-only WebView), so it no longer errors immediately.
* Fix reviewed STT lifecycle races
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix read-aloud fallback controls
* Guard read-aloud stop when deleting a non-speaking message
aui.message().stopSpeaking() throws unless this message is the one being
read aloud, so calling it unconditionally rejected the delete handler before
the message was removed. Only stop speech when this message is speaking.
* Cap recent dictation transcript length before persisting
Recent dictations only limited entry count, so a long transcript stored the
full text in the persisted voice settings and a few could exceed the
localStorage quota, throwing synchronously from the uncaught dictation cleanup
path. Truncate each entry on save and on hydration, matching the dictionary cap.
* Studio: keep dictation mic clickable and guide to local model
Register the dictation adapter unconditionally so the mic stays enabled
for any engine and starts working right after switching to the local
model on an already-open thread.
When the browser engine cannot run (Firefox, Brave, non-secure origins),
clicking the mic shows a toast that points to the local speech-to-text
model instead of leaving a disabled button. The toast stacks its action
below the text with a fully rounded button.
* Studio: add bottom padding below the dictation guidance toast button
* Studio: increase bottom padding under the dictation toast button
* Studio: add bottom padding inside the dictation toast button
* Studio: add five Whisper defaults and custom model search
Add private UnslothAI Tiny and Base mirrors to the curated local STT choices while keeping Small as the default. Let users search or paste a Transformers-compatible Whisper repository and validate it end to end.
Keep short dictations in one clip to avoid repeated padded encoder work, then split longer recordings near Whisper's 30-second boundary.
Update hidden model filters and tests, including the CPU-only CI runtime stub for PyAV.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: use public Unsloth Whisper repositories
Point the Tiny and Base dictation defaults to the public unsloth repositories and remove the private mirror references from model filtering and tests.
* Studio: update Whisper download sizes
Reflect the cleaned public Tiny and Base repositories in the curated model labels.
* Studio: right-align STT model size, fix dropdown wheel scroll, refresh sizes
- Show the download size on the right of each model row so long names
like Whisper Large v3 Turbo no longer hide it
- Update curated Whisper sizes to the safetensors weights actually
downloaded: Tiny 151 MB, Base 290 MB, Small 967 MB
- Drive the model list scroll from a wheel handler so the mouse wheel
scrolls it inside the Settings dialog, not just the scrollbar
- Add a search icon and shorten the placeholder to Search model
* Studio: do not search when a dictation model is picked, shrink repo label
- Treat the filled-in model text as a selection, not a query, so choosing
a model no longer kicks off a Hugging Face search
- Make the repository line under each model name smaller
* Studio: tighten dictation model and local engine descriptions
* Studio: keep model display on pick instead of the query, shrink row text
- Guard the combobox input so selecting a model shows its name and does
not echo the typed query back or start a search
- Map the item label to the friendly display so picks fill the field
- Reduce the model name and size text in each row
* Studio: show only the model name in the dictation field, shrink size label
- Drop the download size from the search field; the name alone is shown
once a model is selected, with sizes kept in the dropdown list
- Reduce the size label text in each row
* Studio: clarify the dictation model description
* Studio: drop Hugging Face from the dictation model description
* Studio: move the dictation dictionary to its own Manage subpage
- Replace the inline entry list with a Manage row, matching Dictation
history, so a long dictionary no longer crowds Voice settings
- Add a DictationDictionaryView subpage that holds the entry editor
* Studio: match STT field font, use best voice for System default
- Bump the dictation model field text to text-sm so it matches the
engine dropdown next to it
- Resolve the System default read-aloud voice to the top curated voice
instead of the browser default, which is a robotic legacy voice on macOS
* Studio: rerank read-aloud voices and drop duplicate voice entries
- Rank by vendor quality, then the user's locale, then a preferred list of
natural voices, so the best voice leads instead of the first alphabetically
- Collapse voices that macOS reports twice under one name and language
* Studio: fold dictionary and recents into the dictation section
- Drop the separate Dictation dictionary and Recent dictations headings;
their Manage rows now sit under Dictation, split by the row divider
- Shorten the custom spellings description
* Studio: add search and sort to dictation history
- Filter saved dictations by text with a search field
- Sort by newest, oldest, or A to Z; show a no-matches message
- Keep Clear all available regardless of the current filter
* Studio: settle cancelled STT loads before training and fix dictation review items
Wait for a cancelled STT load to exit and release its memory before
reporting it freed for training, so the loader cannot still be inside
from_pretrained()/.to(device) holding VRAM when the training subprocess
starts. A load that finishes before observing the cancel now gets
unloaded so the memory is actually reclaimed.
Clear the accelerator cache before the CPU fallback in load() so a failed
CUDA/MPS load does not strand reserved VRAM once the sidecar is marked
CPU-resident.
Send the saved Hugging Face token when polling STT download progress so a
gated or private repo resolves and shows the correct Load/Downloaded
state instead of reporting missing.
Mark the composer Dictate button as type="button" so clicking it does not
also submit the draft when the composer already has text or attachments.
* Studio: pin dictation settings per session and close STT startup races
Capture the STT model and language when a dictation session starts and
pass them to every queued segment and the warm-up load, so changing the
model or language mid-recording no longer transcribes the same clip with
the wrong model or a model that is not downloaded.
Check the local runtime at the top of transcribe(), before the model
cache lookup and the bounded audio decode, so a server missing PyTorch or
Transformers returns 501 up front instead of decoding a long clip first.
Treat the training startup window as active for STT device selection.
start_training frees VRAM in before_spawn but only assigns _proc later, so
a concurrent STT load could take the GPU that was just cleared. A startup
flag now reports training active from the free until the process is live,
forcing those loads to CPU; a finally clears it on every exit.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stub the STT runtime check in transcribe orchestration tests
transcribe() now verifies the local runtime up front, so the unit tests
that exercise transcription orchestration must treat the runtime as
present to keep passing where PyTorch, Transformers, and PyAV are not
installed. Stub ensure_stt_available in the shared fixture and restore
the real check in the availability and load-rejection tests.
* Harden custom Whisper dictation models
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add whisper.cpp dictation engine with per-engine downloads and history rework
Engines
- New GGML STT sidecar that runs a managed whisper-server subprocess with
idle unload, plus a pinned static build script (scripts/build_whisper_cpp.sh)
- Dictation engine picker now offers Browser, Local transcription
(whisper.cpp), and Local transcription (Transformers)
- Both local engines serve the same five curated Whisper models and download
them directly with byte-level progress reported by /audio/stt/status
- Models auto load on selection and when their download finishes
- Unload and training admission account for both engines
Benchmarks (Apple Silicon, greedy, warm, same checkpoints)
- whisper.cpp transcribes 2.4x to 5x faster than Transformers and loads in
about 0.45s vs 0.86s for Whisper Small
- whisper.cpp GGUF path is unchanged by the Transformers addition
(load 0.445s -> 0.444s, short clip 0.391s -> 0.347s, long 1.197s -> 1.129s)
Voice settings UI
- Plain curated model select replaces the searchable combobox
- Single download progress bar with transfer rate for both engines
- Dictation history now stores every dictation with Show more pagination,
a top Clear history action, and links back to the chat it was spoken into
- Archived chats dialog gets the same pagination
- Delete dialog offers deleting a dictation together with its chat
Tests: 88 backend STT tests pass, including new snapshot download coverage.
Frontend typecheck, lint, i18n parity, and production build pass.
* Merge local engines into one option and source GGML models from unslothai
Engine selection
- The dictation engine dropdown is back to two choices: Browser and Local
transcription. The selected model decides the backend: curated ids run
GGML checkpoints through whisper.cpp, searched Hugging Face repositories
run safetensors through Transformers
- Model picker lists the curated models and searches Hugging Face for other
Whisper repositories, validating them before selection. The trigger is a
plain button so the selection never renders inside a text input
- /audio/stt/status accepts a model query param so downloaded state works
for custom repositories; the engine param on load, transcribe, and
download routes is derived from the model everywhere
Model source
- Curated GGML checkpoints now download from the Unsloth-hosted
unslothai/whisper-*-GGUF repositories (one repo per model) instead of
ggerganov/whisper.cpp; cache lookups, progress totals, and in-flight blob
tracking are per-model
Fixes
- Voice settings and dictation history were not persisting: the quota-safe
localStorage wrapper was declared after the store that uses it, so the
persist storage factory failed silently. Every settings write also threw
mid-click, which kept the model picker popover from closing on selection
- is_model_downloaded now verifies config, preprocessor config, and real
weight files instead of trusting an offline snapshot lookup, so a partial
download left by an aborted fetch shows the Download button instead of
failing to load
- Removed whisper.cpp mentions from user-facing text: the ready status
shows Loaded instead of the runtime name, picker rows show the source
repository, and runtime error messages say local transcription runtime
Verified with automated browser sessions and live API checks: selection
closes the picker with no page errors, persisted settings hydrate on
reload, a stale partial snapshot triggers download then loads on MPS and
transcribes, and curated models download from the unslothai repos. 88
backend STT tests, typecheck, lint, i18n parity, and build pass.
* Skip the duplicate source line for custom models in the STT picker
A custom repository's display name is its id, so search results and the
appended current selection rendered the same string twice. The source
line now only renders when it differs from the name; curated rows keep
their name, unslothai source repository, and download size.
* Verify every shard of a sharded checkpoint in the downloaded check
A snapshot holding one of N shards (or a corrupt shard index) passed the
downloaded check and then failed at load. When model.safetensors.index.json
exists, every shard in its weight map must now be present. Found by
simulation; covered by a regression test.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Rename stale _starting references in the pump resilience tests
The startup flag on TrainingBackend was renamed to _spawn_in_progress but
two tests added alongside it still asserted on the old name, failing the
Python 3.11 to 3.13 CI jobs.
* Make the selected model row clearly highlighted in the STT picker
The current selection was a faint background tint. It now uses the accent
background with a medium weight name. Two line rows use a small corner
radius; single line custom repo rows keep the pill shape.
* Address review feedback on STT snapshot checks, VRAM release, and dictation UX
Verify snapshot completeness in the load preflight so a partial download
fails before the audio is decoded, for curated and custom repos alike.
Drop the failed accelerator traceback before the CPU retry so the cache
clear can actually release that memory. Keep unloading the GGUF sidecar
after cancelling an in-flight Transformers load; both engines can hold
memory at once. Allow Auto language with English-only .en checkpoints,
matching the backend which sends no forced language. Keep the discard
button usable while a transcription is pending so a slow or hung request
cannot trap the composer in dictation mode. Stop linking Compare and
settings test dictations to the unrelated active single chat thread.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Move the CPU retry out of the exception handler
On Python 3.10 the interpreter exception state keeps its own reference
to the traceback, so dropping it from the caught exception was not
enough to release the failed accelerator load during the retry. Leaving
the handler before clearing the cache works on every supported version.
* Address review feedback on session handoff, chat pinning, and server lifetime
Starting a dictation from a second entry point now cancels the session
it replaces, so the old recording cannot keep the microphone open or
save a transcript with no discard button pointing at it. The linked
chat is pinned when recording starts, so switching threads while a
transcription finalizes cannot relink the transcript to the newly
opened chat. whisper-server is now bound to Studio's lifetime like the
other long-lived children: PDEATHSIG on Linux, the parent job object on
Windows, and pid adoption so the shutdown sweep reaps it; before this
it survived a Ctrl+C exit as an orphan still holding the model.
* Remove the dictation mic test from Voice settings
The composer dictate button covers the same check, so the test row, its
transcript panel, the unsupported fallback row, and their strings and
search entry are gone.
* Studio STT: gate GGUF whisper-server on training and fix dictation retry and dictionary edits
GGUF (whisper.cpp) sidecar:
- Launch whisper-server with --no-gpu while training is active, mirroring the Transformers sidecar's CPU device choice, so a mid-training dictation cannot reclaim the VRAM training just freed.
- Report is_loading() during whisper-server startup so training VRAM admission accounts for the accelerator memory it is about to bind.
- Require PyAV in is_available() so /audio/stt/status reports the engine unavailable when uploads cannot be decoded, instead of loading fine and then 501ing at transcription.
- Reject a missing model before decoding audio, matching the Transformers download preflight.
Voice settings:
- The download Retry button now restarts the download; the sidecar error is sticky until a new start(), so re-polling alone never cleared it.
Dictation dictionary:
- Tabbing from an emptied entry to its remove button no longer commit-splices the row first, which shifted indices and deleted the wrong entry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio STT: fix curated GGUF whisper filenames to match hosted repos
The unslothai/whisper-*-GGUF repos host the checkpoint as whisper-<id>.bin,
not ggml-<id>.bin, so every curated dictation download and cached-path
lookup 404'd and the whisper.cpp engine could never load a model. Point
GGML_STT_MODELS at the real filenames and guard the naming with a test.
* Studio STT: validate a custom dictation repo before downloading it
The Transformers STT engine accepts an arbitrary owner/model repo, but the
download route handed it straight to snapshot_download, pulling a possibly large
non-Whisper repository into the shared HF cache. Confirm the repo is a Whisper
checkpoint first with the existing metadata-only validate_remote_model (no
weights); curated ids short-circuit and the GGUF engine (curated-only) is
unaffected. A non-Whisper repo now 422s before any download.
* Studio STT: preempt a still-loading GGUF server for training admission
A whisper-server still in its startup window binds accelerator memory but has no
loaded_model yet, so training admission could miss it and launch into an OOM.
Make the GGUF startup cancellable (cancel_pending_load signals an abort event and
terminates the starting process without the load lock; _wait_for_server observes
it and raises SttLoadCancelledError; wait_for_load_to_settle blocks on the lock
until the killed server is reaped), and always fold the GGUF sidecar into the
resident-STT summary so a resident Transformers model cannot mask a loading GGUF
server. free_stt_model_for_training now cancels an in-flight load and waits for it
to settle before training claims the memory.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio STT: fall back to Transformers when whisper-server is absent
A curated dictation model (including the default small) hard-pinned the GGUF
engine, but standard installs do not ship whisper-server, so every recording
501'd instead of using the Transformers engine that serves the same checkpoint
-- the GGUF sidecar's own documented contract. Add _resolve_serving_stt_engine:
a GGUF request for a curated id (the only ids GGUF accepts, all Transformers-
servable) downgrades to Transformers when whisper-server is unavailable, applied
consistently to download, load and transcribe (not unload, which targets a
specific engine). The Voice tab likewise falls back to the Transformers status so
the model is not shown unavailable and download is not blocked.
* Studio STT: hide custom Whisper caches from the legacy model pickers
The legacy /cached-models (and /cached-gguf) routes called is_hidden_model with
only the owner/model id, which cannot reach the config-based Whisper check, so a
downloaded custom (non-curated) Whisper checkpoint was still offered as a chat
model. Pass the cached snapshot path so _path_is_whisper_model inspects the repo
config and hides it, matching the discovery route.
* Studio STT: hide GGUF dictation repos, lock-free status, unload fallback, split training eviction
- Hide the curated GGUF dictation repos (unslothai/whisper-*-GGUF) from the chat
model inventory and pickers, backend and frontend. Only their Transformers
safetensors companions were hidden; the GGUF repos use a different org and a
-GGUF suffix and carry a raw .bin with no whisper config.json, so they leaked
into chat pickers.
- Make the GGUF sidecar loaded_model/device accessors lock-free, mirroring the
Transformers sidecar. transcribe() holds self._lock across the whole inference
call, so /audio/stt status polls and training admission previously blocked
behind an in-flight transcription.
- stt_unload resolves through the serving resolver: a "gguf" pick on a host
without whisper-server is served by the Transformers fallback, so unload must
target that engine or the resident model is never freed. Unload also attempts
every engine even if one raises, so a failure freeing one backend no longer
skips the other.
- free_stt_model_for_training frees the Transformers and GGUF sidecars under
independent exception boundaries so a failure unloading one no longer skips
the other before training claims the memory.
Adds tests/test_stt_review_fixes.py covering all four.
* Studio STT: resolve Auto dictation language for the model engine + snapshot process liveness
- The model dictation adapter sent the raw setting (the literal "auto") to the
backend, while the browser engine resolves Auto via resolveDictationLanguage.
A batch of non-English voice notes came back mostly English on Auto. Add
resolveModelDictationLanguage: only the literal "auto" is resolved to a
concrete locale, gated so it becomes a language the model AND Whisper can
honor (mirroring the backend's known-whisper-languages set); an explicit
language, or a locale Whisper cannot honor, stays unchanged/auto-detect. Wire
it into both adapter call sites.
- GgmlSttSidecar._process_alive() read self._process twice; a concurrent
unload() nulls it under the lock while loaded_model/device read lock-free, so
a null between the two reads called None.poll(). Snapshot once. Adds a
deterministic regression test.
* studio: tighten comments and docstrings in the dictation modules
* studio: harden dictation model downloads, GGML readiness, and recording paths
Address review findings on the STT dictation feature:
- build_whisper_cpp.sh refuses to delete a whisper.cpp tree under a custom
Studio home unless it carries the Studio ownership marker, matching the
setup.sh policy, and marks trees it creates
- _snapshot_is_complete validates every shard of a sharded PyTorch
(pytorch_model.bin.index.json) checkpoint like the safetensors path, and
requires tokenizer assets (tokenizer.json or vocab.json + merges.txt)
- custom-repo downloads pin the revision resolved at validation time and
restrict snapshot_download to the model/tokenizer/config/preprocessor file
classes Studio loads
- the GGML sidecar holds its port reservation until just before spawning
whisper-server and only accepts readiness from a responder that both looks
like whisper.cpp's server and belongs to the still-running managed child,
probing twice, so mic audio cannot be posted to a foreign local process
- the recording adapter transcribes every non-empty segment; the RMS meter
only shapes segment boundaries and can no longer discard quiet speech
- Compare-pane dictation can cancel a pending transcription on second click,
with the button relabeled while finalizing
- localStorage quota recovery halves the dictation history until the save
fits, so small histories shrink too
- the System default TTS voice resolves to the platform default voice
- new dictation UI imports go through the chat and hub feature barrels
Regression tests cover the build-script gate, sharded PyTorch and tokenizer
completeness, revision pinning and allow patterns, and the whisper-server
readiness probe.
* Fix STT download and voice picker follow-ups
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Add dictation button regression coverage
* Studio: prebuilt whisper.cpp via the shared llama.cpp install core, slim bundles paired to the llama prebuilt (#7294)
* Studio STT: add prebuilt whisper.cpp (whisper-server) installer
New install_whisper_prebuilt.py downloads a per-platform whisper-server
bundle published by the unslothai/whisper.cpp prebuilt CI into the managed
whisper.cpp dir (build/bin/whisper-server) so local dictation needs no
compiler. Mirrors install_node_prebuilt.py / install_llama_prebuilt.py:
host + backend detection, sha256 pins (whisper_prebuilt_pins.json) as the
trust anchor, staging + install lock + atomic swap, traversal-safe extract,
co-located shared libs (RUNPATH=$ORIGIN), an UNSLOTH_WHISPER_PREBUILT_INFO.json
marker with idempotent "already matches", and exit codes 0/1/2/3. Not wired
into setup yet; the pins ship empty so every asset fails closed until the
first fork release is published and its digests are reviewed in.
* Studio STT: install prebuilt whisper.cpp during setup and update
Add a fail-open whisper.cpp block to setup.sh after the llama.cpp section so
`unsloth studio update` (and a fresh install) fetch the prebuilt whisper-server
into the managed whisper.cpp dir the sidecar discovers. It skips a user-set
WHISPER_SERVER_PATH/UNSLOTH_WHISPER_CPP_PATH, honors UNSLOTH_SKIP_WHISPER_INSTALL,
forwards the resolved ROCm gfx, and never aborts setup: a busy install keeps the
existing runtime, and an unavailable prebuilt stays quiet (source build is opt-in
via UNSLOTH_WHISPER_FORCE_COMPILE) since Transformers STT and browser dictation
remain. Register UNSLOTH_WHISPER_PREBUILT_INFO.json as Studio-owned evidence.
* Studio STT: harden whisper-server child env + WSL ROCm detection
- Sidecar spawns whisper-server with a scrubbed child env that prepends the
binary dir (co-located GPU libs) to the loader path, and on WSL2 ROCm loads
the system HIP first (HSA_ENABLE_DXG_DETECTION=1) so a bundle's bare-metal HIP
does not segfault on /dev/dxg. Secret-bearing vars are dropped from the child.
- find_whisper_server_binary now requires an executable, not just a file.
- Installer rocm probe passes HSA_ENABLE_DXG_DETECTION and falls back to
/opt/rocm/bin/rocminfo so a WSL ROCm host is not misdetected as CPU-only;
gfx parsing skips the gfx000 CPU agent and generic ISA lines.
- Tests for the child env (secret scrub, lib dir, WSL HIP precedence), the
executable check, and the WSL rocm detection.
* Studio STT: in-app whisper.cpp prebuilt update stack + ship pins in the wheel
Mirror the llama.cpp update stack for the whisper.cpp prebuilt so Studio can
detect and install a newer whisper-server release from inside the app:
- backend/utils/whisper_cpp_freshness.py: read UNSLOTH_WHISPER_PREBUILT_INFO.json
and compare the installed release against the newest unslothai/whisper.cpp
release. Whisper tags are v<upstream>-unsloth.<N>, so is_behind compares a
(major, minor, patch, serial) key with a strict downgrade guard; 24h cache;
fail-open.
- backend/utils/whisper_cpp_update.py: run install_whisper_prebuilt.py to fetch
and atomically swap the newest bundle, unloading the warm GGUF sidecar first.
- backend/routes/whisper.py mounted at /api/whisper (update-status + update).
- pyproject: add whisper_prebuilt_pins.json to studio package-data so the
installer's trust anchor ships in the wheel (it is a data file, not a .py
module, so package discovery alone does not include it; node_prebuilt_pins.json
is listed for the same reason). Without this a pip-installed wheel had no pins
and the prebuilt install aborted to Transformers STT.
Adds test_whisper_cpp_freshness.py (version parser, is_behind matrix + downgrade
guard, marker layouts, stale decision, fail-open).
* Studio STT: verify whisper prebuilts via the release checksum index, like llama.cpp
Re-align the whisper.cpp prebuilt installer to install_llama_prebuilt.py's trust
model: instead of a committed whisper_prebuilt_pins.json, verify every download
against the release's own whisper-prebuilt-sha256.json checksum index, fetched
from the same GitHub release.
- parse_release_checksums / fetch_release_checksums / expected_sha256_for replace
the pins layer. The index is validated for schema/component and that its
release_tag matches the resolved release; an asset absent from it, a release
that does not publish it, or a manifest sha256 that disagrees with it all fail
closed to a source build.
- resolve_release_tag now resolves the newest published release at runtime (or an
explicit --published-release-tag), matching llama and the freshness check;
removed the pinned-default and the UNSLOTH_WHISPER_ALLOW_UNVERIFIED opt-in.
- Delete studio/whisper_prebuilt_pins.json and drop its pyproject package-data
entry (nothing to ship now, same as llama which has no committed pins).
- Adds test_install_whisper_prebuilt_checksums.py (index parser, fail-closed on
uncovered asset, tampered-manifest guard, newest-release resolution).
This is a same-origin checksum (integrity, not authenticity), identical to the
llama.cpp installer; pair releases with GitHub artifact attestations for provenance.
* Resolve whisper prebuilt release via the download host (no GitHub API)
Mirror install_llama_prebuilt.py's fast path: resolve the release tag from
the releases/latest redirect and fetch the manifest + checksum index from
constructed releases/download URLs, so the common install path makes zero
api.github.com calls (unauthenticated api.github.com is capped at 60 req/hour
per IP; the download host is not). Fall back to the GitHub API only on a 404,
malformed asset, or tag mismatch.
* Studio STT: coverage-aware whisper prebuilt selection via a shared core
whisper's select_artifact returned the first os/arch/backend manifest match and
ignored the SM-coverage fields the release manifest already carries, so a
Blackwell B200 (sm_100) was served cuda12-legacy (sms 50-61) -- runnable only via
forward PTX JIT. install_llama_prebuilt.py on the same host correctly picks
cuda13-newer.
Extract the coverage-aware selection into a shared, component-agnostic core under
studio/backend/utils/prebuilt/ (selection + GPU host-capability detection), lifted
from llama's linux_cuda_choice_from_release / _artifact_covers_sms / _sm_range and
generalised over a normalised artifact. whisper's HostInfo now records the GPU
compute caps + driver CUDA version (honoring CUDA_VISIBLE_DEVICES), and
select_artifact routes CUDA/ROCm through the shared selector: every visible SM
must be covered, the tightest-covering profile wins (Blackwell-aware runtime-line
ordering), ROCm matches the gfx target exactly, and an uncovered GPU falls back to
the CPU bundle. CPU/Metal/Vulkan keep first-match. The resolver JSON, exit codes,
and "already matches" contract are unchanged.
On the B200 the installer now resolves cuda13-newer, matching llama.
* Studio STT: gate whisper CUDA selection on the on-disk runtime, like llama
The prebuilt CUDA bundles are dynamically linked and intentionally do NOT ship
libcudart/libcublas -- they load the same runtime the host already has. So the
driver's advertised CUDA version is only an upper bound: a cuda13 bundle still
needs cuda13 runtime libraries present on disk. Port llama's on-disk runtime
scan (detected_linux_runtime_lines / detected_windows_runtime_lines) into the
shared core and intersect it with the driver-compatible lines in
select_cuda_attempts. A host with a cuda13 driver but only cuda12 runtime (e.g.
torch-cuda12) now correctly gets a cuda12 bundle instead of an unloadable cuda13
one; a host with no CUDA runtime at all falls back to CPU.
Fixes a glob bug in the port (any(Path(d).glob(p) for d in dirs) tests generator
truthiness, not a match) that made every major report present; add a real
filesystem test that exercises the scan.
* studio: harden shared prebuilt core to full llama parity
Apply the review findings on the shared coverage-aware prebuilt-consumer
core so whisper.cpp selection is exactly equivalent to the llama.cpp path.
hosts.py: port llama's CUDA_VISIBLE_DEVICES handling. A GPU hidden by an
index/UUID selector now reports has_usable_nvidia False instead of staying
usable, via supports_explicit_visible_device_matching plus the physical /
explicit-match branches, and _select_visible_rows now matches rows the way
llama does (index or UUID, gpu- prefix optional) and skips unmatched tokens
rather than keeping all rows. Adds the Linux /proc/driver/nvidia/gpus
fallback and has_physical_nvidia. Adds parse_macos_version.
runtime_libs.py: the Linux on-disk scan now requires the exact libcudart /
libcublas SONAME (libcudart.so.13), not a libcudart.so.13* glob, so a bare
versioned file without the SONAME symlink no longer counts as loadable.
Hardens the ldconfig parse against an empty left-hand side.
selection.py: fix the Blackwell/torch reordering so it keys on the covering
runtime lines (falls through to the torch preference when the covering lines
were filtered out), matching linux_cuda_choice_from_release. Corrects the
compatible_runtime_lines_for_driver docstring: the bundles do not ship the
CUDA runtime, so the driver version is only an upper bound and the caller
must intersect with the on-disk scan.
install_whisper_prebuilt.py: enforce a macOS artifact's min_os (new
HostInfo.macos_version) so a bundle that cannot load on the host OS version
is dropped. Keep resolver stdout to only the JSON line by leaving logs on
stderr in --resolve-prebuilt mode, and map an unexpected probe failure to
prebuilt_available False instead of a traceback.
Tests: new host-probe suite for the visible-device logic, exact-SONAME
runtime-scan cases, macOS min_os filtering, resolver stdout-only-JSON,
exit-code mapping, and the repo key.
* studio: fix whisper prebuilt selection + launch parity gaps from review
A parallel review surfaced integration defects where the whisper path could
select or launch a bundle that cannot run on a concrete host. Each is fixed to
match install_llama_prebuilt.py.
macOS min_os: the manifest labels macOS requirements as macos-<version>
(e.g. macos-14.0), which the version parser could not read, so the guard was a
no-op and a macOS-13 host would install the macos-14 Metal bundle. Strip the
platform prefix before parsing.
ROCm gfx detection: _detect_rocm_gfx returned the first gfx token and ignored
HIP_VISIBLE_DEVICES / ROCR_VISIBLE_DEVICES / CUDA_VISIBLE_DEVICES. Since exact
ROCm matching treats that token as the active GPU, a mixed APU + dGPU host
(gfx1151 + gfx1100) with HIP_VISIBLE_DEVICES=1 installed the wrong archive. Route
through a shared pick_rocm_gfx_target (lifted from llama) that parses per-GPU
sections and honors the visibility vars (empty / -1 -> no AMD GPU).
--rocm-gfx override: recording the arch without setting has_rocm left the host on
its CUDA/CPU path so the ROCm bundle was never picked. --rocm-gfx now implies
has_rocm and clears NVIDIA state, like llama's _apply_host_overrides.
CUDA launch env: a CUDA bundle ships the ggml CUDA backend but not
libcudart/libcublas, and the sidecar launch env exposed only the bundle dir, so
on a host whose CUDA runtime lives only in the PyTorch wheels the selection would
gate cuda usable but the server could not load it. Add the CUDA-from-PyTorch
runtime dirs to the child loader path for CUDA bundles (bundle dir still first),
mirroring binary_env.
Also normalize a manifest artifact's supported_sms defensively (parity with
llama's parser) and document that blackwell_min_toolkit_for_caps is retained for
the Phase B llama Windows path.
Not changed (verified parity, not defects): Linux/Windows min_os is enforced
nowhere in llama (macOS only); the resolver is optimistic about the checksum
index and the install path verifies.
* studio: tighten prebuilt-core code comments
* studio: lift shared prebuilt installer core out of the whisper installer
* studio: reuse the llama.cpp prebuilt installer machinery for whisper
* studio: unify llama and whisper prebuilt installers on a shared descriptor core
* studio: consolidate prebuilt installer tests into the shared core suite
Grow tests/studio/install/test_prebuilt_core.py from 62 to 164 tests so every
component-agnostic behavior runs against both descriptors: the full seven
profile CUDA release matrix (multi-GPU, on-disk runtime gating, shuffle
stability, missing SM metadata, dotted SM normalization, no-driver fallback
policy), the ROCm gfx family matrix, macOS min_os gating and its helper,
backend resolution incl. cpu-fallback precedence and Intel-mac auto detect,
checksum-index non-object and plain-lookup cases, the tar symlink/hardlink
extraction guards moved from the llama suite, and the compute-cap, visible
device, runtime-line and Blackwell helper value tables moved verbatim from
the llama characterization suites.
Delete only tests whose exact behavior the master now asserts for the same
component: 40 pure-alias helper cases in test_selection_logic.py (replaced by
value-identical master tables plus an alias-identity pin), 6 extraction moves
and the master-absorbed zip-symlink case in the llama logic suite, 3 routing
twins in test_rocm_support.py already pinned byte-for-byte in
test_selection_logic.py, the 2 Blackwell helper tables in the backend resolve
suite, 28 whisper logic tests and 10 whisper checksum tests re-asserted by
the master whisper parameterization. Wrapper wiring pins, the llama release
plan dialect, fingerprints and every llama-only behavior stay untouched.
* studio: dedupe sidecar and update helpers into the backend prebuilt package
* studio: chain whisper.cpp prebuilt updates onto the llama.cpp update flow
* studio: consume paired slim whisper prebuilts via the llama ggml runtime
* studio: serve every whisper backend from slim prebuilts
* studio: drop the whisper fat per-accelerator selection chain
unslothai/whisper.cpp releases are slim-only from v1.9.1-unsloth.2: one
ggml-less bundle per os/arch, paired to the llama.cpp prebuilt that provides
every ggml backend. Delete the whisper-side fat CUDA/ROCm/metal/vulkan
selection glue; keep slim selection + pairing, link_ggml_runtime, and one
legacy shape, the published fat CPU bundle of an explicitly pinned pre-slim
release. Exit 2 now reads as prebuilt unavailable (whisper never source
builds); setup already treats it that way.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Wire libomp runtime DLL alongside ggml in slim whisper installs
llama's clang-built windows-arm64 ggml-base.dll imports
libomp140.aarch64.dll, shipped in the llama bundle but not a system DLL.
Without it next to whisper-server.exe the loader fails with
STATUS_DLL_NOT_FOUND before main. MSVC x64 links vcomp140.dll from
System32 and Linux ggml uses system libgomp.so.1, so only windows-arm64
was affected. The empty-runtime guard still requires a real ggml
library; libomp alone is not a pairing.
* studio: drop whisper-side fat-selection support structure
Slim whisper bundles are selected per os/arch only; all accelerator
capability comes from the installed llama.cpp prebuilt, whose installer
already did the coverage-aware selection. Remove the machinery that only
existed to pick among fat per-accelerator whisper bundles:
- prebuilt_core: delete the generic CUDA/ROCm coverage selection
(select_cuda_artifact, select_rocm_artifact, ArtifactView adapters,
detected_cuda_runtime_lines, the exact-SONAME linux probe) that no
shipped component routes through; llama keeps its own selection chain
and whisper shadows select_artifact with the slim-only version.
select_artifact is now a plain os/arch/backend first-match.
- install_whisper_prebuilt: drop the HostInfo CUDA fields
(compute_caps, driver_cuda_version, torch_runtime_line) and the torch
runtime probe that populated them; nothing reachable reads them, and
the resolver payload sources runtime_line from the artifact.
- whisper_cpp_update: delete the standalone start_update job worker;
whisper applies only run as the chained phase of the combined
llama+whisper update. The status payload keeps its job field (idle).
- routes/whisper: drop the progress logger that could never fire.
- tests: remove tests of the deleted paths and tests duplicating the
descriptor-parameterized core suite or the llama freshness suite.
Contracts unchanged: resolver JSON keys, exit codes, marker fields,
pairing logs, and the pinned pre-slim fat CPU escape hatch.
* Address review feedback on the whisper prebuilt update and install paths
- Pin the chained whisper phase to the release the freshness check
offered, so the download-host latest pointer cannot reinstall an
older build in a loop
- Wire the whisper prebuilt install into setup.ps1 (Windows setup
previously skipped it entirely)
- Treat a non-executable server or missing wired ggml libraries as a
broken install instead of reporting already matches
- Keep whisper sidecar reloads out of the job-level reload flag and
resync chat state after a partial chained update that unloaded llama
- Repoint home and profile vars for the whisper-server subprocess at a
managed scratch dir and drop credential-store pointers
- Clear the prebuilt marker before the opt-in source build overwrite
- Write the prebuilt marker with explicit utf-8 encoding
* Tighten comments in the whisper prebuilt consumer
* Harden the Windows whisper setup phase and the chained update edges
- setup.ps1: honor WHISPER_SERVER_PATH / UNSLOTH_WHISPER_CPP_PATH /
UNSLOTH_SKIP_WHISPER_INSTALL, run the custom-home ownership guard
before the atomic install, and forward the release-tag pin and ROCm
hints like setup.sh
- sidecar: a cpu-selected install launches whisper-server with --no-gpu
(slim wiring links every llama backend, so the flag is what keeps a
deliberate CPU choice off the GPU)
- chained update: leave whisper unpinned on macOS (the llama phase can
walk back there, and a newest-tag pin could be an impossible pairing
on every retry) and treat installer exit 2 as kept-existing-runtime
instead of failing the combined job
- job.to_tag now comes only from the llama phase, so a whisper-only
round cannot report a llama update that never ran
* Fix slim whisper runtime follow-ups
* Address remaining whisper update reviews
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address remaining prebuilt update reviews
* Fix remaining chained update reviews
* Fix remaining whisper runtime review edges
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: danielhanchen <unslothai@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
---------
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Batched generation: /images/generate takes a prompts list (one image per
prompt, txt2img only) or a seeds list (one prompt, one image per seed);
the legacy batch_size path derives per-image seeds base..base+n-1 like
the native engine. Every image gets its own torch.Generator so any batch
member replays alone from its gallery recipe; the whole list runs as one
forward by default with OOM backoff that halves a failed chunk, and an
explicit batch_size caps images per forward. Validated 10-22x over
serial engines on 32-image suites with LPIPS deltas within 0.002.
Conditioning cache on the inference path: UNSLOTH_DIFFUSION_COND_CACHE_DIR
(the inference sibling of the trainers' cond_cache_dir, same persistent
store) wraps encode_prompt so repeated prompts skip the text-encoder
forward entirely; verified bit-identical outputs. Bypassed while LoRA
adapters are attached; tensor-argument calls pass through uncached.
Compile cache: GGUF loads fingerprint their own bundles (quant=gguf, a
different compiled graph than the dense family) and batched calls
register every distinct (w, h, batch) chunk shape they ran, so the heavy
GGUF batched warmups (~159 s at batch 32 on 12B-class, ~655 s on 20B
CFG-batched) are paid once ever.
GGUF loader: strip the sd.cpp model.diffusion_model. container prefix in
the single-file converter; diffusers' FLUX.2 converter KeyErrors on it
and the Qwen-Image identity mapping strands the model on meta.
* refactor(studio): move chat model picker into features/model-picker
Relocate model-selector + its support files from components/assistant-ui
into a self-contained features/model-picker feature (own barrel), mirroring
the modular Hub layout. Pure move + import repoint; no behaviour change.
* feat(model-picker): add per-model config persistence layer
Superset PerModelConfig (customContextLength, kvCacheDtype, speculativeType,
specDraftNMax, tensorParallel, chatTemplateOverride, trustRemoteCode) persisted
to localStorage (unsloth_model_configs) with schema versioning + LRU budget.
KV-dtype and speculative value sets match main's sidebar (no q4_0/ngram-simple).
Reuses features/hub/lib/model-identity for normalization; adds storage-key layer
and applyPerModelConfigToRuntime (sets tensorParallel, which the old PR omitted).
* feat(picker): modular backend for chat-template validate + default fetch
New studio/backend/picker package (schemas/service/routes) mounted at /api/picker:
- POST /api/picker/validate-chat-template (Jinja syntax validation, no false positives)
- GET /api/picker/chat-template/{model_name} (default template from tokenizer_config.json,
reusing get_cache_path/resolve_cached_repo_id_case; graceful null, no model-code exec)
Frontend api/templates.ts client + hooks/use-model-defaults lazy cache. No backend
changes to the existing inference load route (per-model load fields already supported).
* feat(model-picker): bind picker on-device list to shared hub inventory
Picker now sources cached + local models from useHubInventory (the Hub's shared
store) via a thin adapter, replacing its own /api/models/* fetchers + module
caches. Hub, download manager, and picker now share one source of truth, so
completed downloads reflect in the picker automatically. Partial/live-download
rows are filtered from the cached lists (unchanged rendering). Local naming/search
preserved via additive LocalInventoryRow modelId/displayName. Variant expander,
scan-folder management, recommended-fit, search, external providers untouched.
Known minor: cached 'Downloaded date' sort tiebreak degrades to alphabetical
(hub cached rows carry no mtime); default 'recent' (load-time) sort preserved.
* feat(model-picker): per-model config step inside the picker
Picking a (non-external) model now opens an in-picker config view built from
main's current load controls (context length, KV cache dtype, speculative
decoding, draft tokens, tensor parallel) plus a chat-template editor backed by
the picker validate/default endpoints. 'Remember for this model' persists the
config per model+variant; Run forwards the config to the existing load flow via
meta.config. External models bypass the step. Two-view orchestration lives in
model-selector (single interception point); pickers.tsx call sites untouched.
trustRemoteCode dropped from PerModelConfig to preserve main's per-load consent.
* feat(chat): apply/persist per-model config through the load flow
handleCheckpointChange threads meta.config into the selection; stageOrLoad and
the autoload/Hub-run paths now apply the picker config (explicit pick or saved
remembered config) via applyPerModelConfigToRuntime before staging/loading, with
keepSpeculative set so a remembered speculative mode survives the model switch.
Replaces the old remembered-load-settings seeding (resolveInitialConfig now the
single source). SelectedModelInput carries config.
* refactor(chat): remove per-model load config from the right sidebar
The load knobs (context, KV cache, speculative, draft tokens, tensor parallel)
and the chat-template editor now live only in the picker config step. The sheet's
Model section keeps the staged Load/Cancel flow (config is applied at pick time);
sampling params, system prompt, and RAG are unchanged. Deletes the superseded
remembered-load-settings module + the store's applyRememberedLoadSettings action,
removes the now-dead sheet state/imports, and points the settings reset at
unsloth_model_configs. Delete-cleanup deferred (stale config is LRU-capped).
* fix(model-picker): remove leftover sidebar-staging cogwheel + empty Model section
The downloaded-variant gear (ModelLoadSettingsAction) staged a model straight
into the right-sidebar Run-settings flow -- the old 'configure before load' path
now fully replaced by the in-picker config step. Removed the gear + its component.
Also gate the sheet's 'Model' section to staged picks only (pendingSelection):
after the load-knob strip its content is staged-only, so it was rendering an
empty section header whenever a model was merely loaded.
* chore(chat): remove dead per-model-config setters + modelControlsDisabled
After the load-config UI moved into the picker, the store's per-model setters
(setKvCacheDtype/setSpeculativeType/setSpecDraftNMax/setTensorParallel/
setCustomContextLength/setChatTemplateOverride) had zero callers
(applyPerModelConfigToRuntime writes via setState), and the sheet's
modelControlsDisabled was unreferenced. Verified dead across the whole tree.
* fix(chat): config-step Load actually loads (ignore Load-on-selection)
Root cause: with Settings > Chat > 'Load on selection' turned OFF, the config
step's load went down the deferred-staging path -- opening the right sidebar with
'<model> is staged, not loaded yet / Choose Load model'. The in-picker config step
IS the deliberate load action, so its Load now loads immediately (or downloads +
auto-loads when not cached) regardless of the toggle. Renamed the button
'Run model' -> 'Load model' to match. Native/dropped picks still honor the toggle.
* refactor(chat,hub): retire 'Load on selection' — config step is the only load flow
The in-picker config step (and the Hub Run button) now fully supersede the old
stage-to-sidebar flow, so the Load-on-selection toggle is removed everywhere:
- chat stageOrLoad: every pick loads immediately, or downloads + auto-loads when
not cached (the previous default behaviour, now universal).
- hub Run: drops the stage branch; downloaded GGUFs load directly with their saved
per-model config (no collision with the chat config step — both end at selectModel).
- store: removed loadOnSelection field/setter/key/default; Settings>Chat toggle and
its settings-reset entry removed.
- staged sidebar section is now a download-progress view (auto-loads on completion).
No manual staging remains; stageModel is used only for background auto-load downloads.
* feat(model-picker): default chat template from GGUF + thread variant through config flow
Read the embedded tokenizer.chat_template from GGUF files (read_gguf_chat_template
in gguf_metadata) and use it as the per-model default. Plumb gguf_variant through
the picker service, /api/picker/chat-template route, frontend templates API, and
use-model-defaults so the right variant's template is fetched.
Also refine the picker config-page/model-selector wiring, drop the dead
ggufNativeContextLength runtime path, and add the per-model-config storage keys to
the settings prefs export.
* feat(model-picker): read safetensors chat template + hide editor where it has no effect
Resolve the default chat template for safetensors models: prefer the modern
chat_template.jinja, fall back to the tokenizer_config.json chat_template field,
then chat_template.json (multimodal processor), then the GGUF embedded template.
Applied to local dirs, the HF cache snapshot scan, and the HF remote fetch.
Hide the chat-template editor in the picker for safetensors models — the override
is only applied at load by the GGUF/llama.cpp backend, so editing it on safetensors
currently has no effect. GGUF keeps the editor. Nothing removed; the dialog stays
for when the safetensors apply path is wired up in a later branch.
* fix(model-picker): set legacy-migration flag only after the write succeeds
Set unsloth_model_configs_migrated only once writeMap confirms the migrated
map persisted, so a quota/storage failure no longer marks migration done and
silently drops the user's pre-existing remembered settings — the next load retries.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* MVP model picker fixes
* MVP picker config fix
* MVP safetensors config
* MVP max seq config
* MVP max seq fix
* Fix static max tokens cap ignoring model context
* Fix picker GGUF scan parity
* fix(studio): harden model picker config loading
Apply remembered per-model configs consistently from picker and Hub loads, keep default configs from overriding standing speculative settings, add config access for direct local GGUF files, and support saving or forgetting active model settings without a reload.
* Fix model picker config flow
* Fix model picker config loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Avoid recursive per-model config migration reads
* Apply the displayed context length when loading a GGUF
* Fix template validation, cached template lookup, and failed load rollback
- Validate chat templates with the loopcontrols extension so templates
that use break or continue tags pass the picker validator, matching the
inference renderer that already accepts them.
- Read the default chat template from the newest cache snapshot rather than
an arbitrary iterdir order, so an older cached revision no longer prefills
a stale template.
- Capture the runtime per-model config before a load and reapply it when the
load fails, so a failed switch leaves the active model context, KV cache,
template, and speculative settings as they were.
* Make chat template view only for safetensors models
Custom chat template overrides are applied at inference only for GGUF
models, which pass the template to llama-server. The safetensors backend
renders with the model built-in template and ignores the override, so
editing it would save a value that never loads. For safetensors the
config page now opens the template as a read-only preview with a note
that editing is not available yet. This can become editable once
inference support for custom safetensors templates lands in main.
* Fix model picker config edge cases
- Restore prior runtime config when a load no-ops for the active model
- Cap the picker validator request body via the protected prefixes
- Keep the GGUF context slider max above the loaded context
- Fetch subfolder chat templates for uncached Hub repos
- Show the compare side config when reopening the picker
* Keep saved GGUF context above the fallback ceiling
* Show the model config in the run settings sidebar
* Fix model config sidebar reset and context slider
- Stack the remember toggle and action buttons in the sidebar
- Reset the config to defaults instead of the loaded values
- Fetch the native context so the slider max is not the loaded value
* Fix model picker config and download regressions
- Run picker chat template routes off the event loop
- Depth and root guard local template directory scans
- Restore download manager flow for uncached hub picks
- Apply per model context length on reload
- Import model picker symbols from the feature barrel
* Fix model picker config and cached download sorting
- Restore load settings when a Hub run is rejected mid load
- Reuse one NumericValueInput instead of a duplicate copy
- Fix double decode of the model name in the template route
- Remove the unused reset-to-loaded settings action
- Fix cached model download sorting
* Fix model picker per-model config edge cases
Honor a saved or typed max seq length above the model's native context so
RoPE extended values are no longer clamped and silently overwritten. Allow
typing past native while the slider keeps native as a soft ceiling.
Guard the fetch success paths in use-model-defaults against an aborted
signal, and refetch when the HF token changes.
Hash the chat template content in the sidebar remount key instead of its
length. Enable reset for a GGUF whose native context is unknown, and floor
the context slider max so it can never fall below the min.
* Fix GGUF context auto-fit and gated model config token
Stop forcing a 32768 context when a GGUF native context is unknown so the backend auto-fits to VRAM again, while still honoring an explicit context edit.
Send the HF token as a query param so gated safetensors models resolve their max position embeddings.
Derive model default state during render to drop the set-state-in-effect calls.
* Fix native GGUF context ceiling and guard picker template reads
Restore the native context store field so the sidebar slider keeps the
full ceiling for drag and drop GGUFs. Limit local chat template reads to
the browse allowlist, skip malformed repo ids, and drop unused model
picker exports.
* Fix model picker lint boundaries
* Fix model picker review findings
Chat template editor never seeded its draft. Radix only calls onOpenChange
from internal events, so the seed in the nextOpen branch was dead and a model
with a saved override opened empty. Saving then cleared the override. Drop the
dead branch, treat draft as an untouched sentinel, and reset it on every close.
Uncached Hub picks could auto load a model after the user left the chat. Main
detached the staged pick on route exit and on chat context change. Carry the
context key on the pending pick and skip the load when it no longer matches.
Also clear configTarget when the picker closes, restore the onUpdated ref so
variant rows stop resubscribing on every parent render, skip the LRU write when
the entry is already most recent, import NumericValueInput relatively, and drop
the unused ModelUpdateAction barrel export.
* Preserve GGUF context on active reload
* Fix model picker per-model config regressions
- Stop reloading the already loaded model on re-pick
- Hide infra models from the chat picker
- Detect vision support on cached GGUF repos
- Honor saved maxSeqLength on auto load
- Restore default chat template for local GGUFs
- Warn on save failure and revert config on cancel
- Refetch picker inventory on open
- Persist read only per model config safely
* Fix stale model auto load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model picker numeric input sizing and constraints
Size value inputs to their content so long context lengths are not clipped,
restrict them to numeric characters, and stop the speculative decoding label
from truncating in the sidebar.
* Fix picker CI tests and harden chat template resolution for PR #6647
- tests: point the descender guard at the moved model-selector.tsx path
- tests: exclude the disabled Reload model button from the regenerate locator so .first targets the real Regenerate
- picker/service.py: reject symlinked template/gguf leaves that resolve outside the browse allowlist (HF cache reads unchanged)
- compare mode: resolve each pane's own remembered chat template instead of inheriting the other pane's from the store
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Protect future-schema per-model configs from deletion for PR #6647
savePerModelConfig already refuses to overwrite a stored config whose schema version is newer than this client understands, but deletePerModelConfig did not. Unchecking Remember on an older client therefore silently destroyed a newer client's saved config. Apply the same guard on delete and surface the blocked case through the existing saveFailed toast.
* Protect future-schema per-model configs from quota eviction for PR #6647
The save and delete guards already refuse to touch a stored config whose schema version is newer than this client understands, but the quota-eviction path did not, so a full store on an older client could still evict a newer client's config. Skip future-schema entries when evicting and fail the save if the budget cannot be met without them.
* Fix GGUF context persistence, compare context, and rollback settings for PR #6647
Persist a GGUF context override from the user's intent instead of collapsing it against the loaded context, which reintroduced the context-reset (f4838782cb reverted the native-baseline fix). model-config-page now collapses the saved value against native, and use-chat-model-runtime and chat-adapter retain the requested context on load so re-saving another setting keeps the override; a null request stays null so a VRAM auto-fit never becomes a stored override.
shared-composer: a compare pane with no explicit GGUF context now loads at native (0) like single-view, not the session maxSeqLength that silently shrank the shown context.
use-chat-model-runtime: restore the previous model's KV cache dtype and chat template on a failed-load rollback so it runs as it was, not with backend defaults.
* Preserve native path token when reloading the active model for PR #6647
handleReloadActiveModel rebuilt the selection without the store's activeNativePathToken, so reloading a file-picked GGUF after a settings change validated the display label as a repo/path and failed. Thread the active native token through the reload selection so native-loaded models reopen correctly.
* Make picker template validation resilient and accept HF generation tags for PR #6647
Import Jinja lazily inside validate_chat_template so a backend without the optional jinja2 package (GGUF-only installs) still starts instead of raising ModuleNotFoundError at import time. Register a no-op extension for the Transformers {% generation %} assistant-mask tag so pasting a valid HF chat template validates, matching the renderer, rather than being rejected as an unknown tag.
* Honor remembered compare config and parse processor chat_template.json for PR #6647
* Fix failed-load rollback context and processor template map fallback for PR #6647
* Restore speculative decoding config on failed-switch rollback
When a model switch fails after the previous model was unloaded, the
rollback reload restored tensor_parallel, KV cache dtype and the chat
template override, but omitted speculative_type and spec_draft_n_max and
cleared their loaded shadows to null. The previous model therefore came
back running at backend defaults (speculation off) while the UI still
showed it enabled, and the status resync confirmed the off state. Resend
the previous model's speculative settings in the rollback load and keep
the store's active and loaded speculative fields in sync with them.
* Reset max sequence length when a model has no saved config
applyPerModelConfigToRuntime reset every per-model field except
maxSeqLength, which it only wrote when the incoming config had one.
maxSeqLength is the sole field carried on store.params, so selecting a
model with no remembered config left the previous model's value in place
and later loaded the new model at that leaked length. Fall back to the
standing default so an unremembered model loads at its own default.
* Surface a message when a variant update cannot start
startManagedUpdate handled the conflict and error start outcomes but let
busy fall through as if the update began, so the confirm dialog closed
with no job created and the cached variant stayed stale. Show an info
message when the repo is busy with a sibling transfer so the click is
not silently dropped.
* Keep per-model speculative choices out of the global default
A staged load with a per-model or one-off config sets keepSpeculative,
which already skips reading the global speculative preference. The
matching save still ran unconditionally, so the model-specific choice was
written to the global unsloth_chat_speculative_type and a later model with
no saved config started from it instead of Auto. Skip saveSpeculativeType
when keepSpeculative so the per-model choice stays isolated.
* Seed non-active model settings from the app default max length
The Run settings page captured initialMaxSeqLength from the loaded
model's runtime params and fell back to it for a model with no saved
config. Opening settings for a different, unloaded model and clicking
Load then sent the active model's context (for example 64k) instead of
the 4096 default, risking validation failures or OOMs. Seed the default
for non-active models and keep the runtime value only for the active one.
* Prefer sidecar tokenizer chat template over the GGUF copy for variants
_chat_template_from_dir returned the embedded GGUF template first when a
variant was selected, reversing the tokenizer-first precedence of the
no-variant path. A model whose chat_template.jinja or tokenizer_config.json
supersedes a stale embedded template then got the wrong template on
variant selection. Keep tokenizer files first regardless of variant; the
variant only picks which GGUF is the fallback. Adds regression tests for
both the tokenizer-wins and gguf-fallback cases.
* Keep per-model speculative choices load-local in autoload and compare
The interactive load path treats a per-model speculative choice as
load-local and skips writing it to the global default. Autoload and
generalized compare still called saveSpeculativeType unconditionally, so a
remembered off or ngram setting leaked into unsloth_chat_speculative_type
and later models with no saved config inherited it. Persist the global
preference only when the value came from the global settings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: record the compare pane's loaded context in runtime state so the active model's settings and any reload or save use it, not the previous context
* Studio: notify the user when a Hub autoload can't start because another download for the model is already running, instead of silently dropping it
* Studio: drop the merge's orphaned staged-model store helpers and unused alert imports
The main merge left isPendingGguf and pendingSelectionMatches referencing the
removed PendingModelSelection type, and the alert-dialog/alert imports unused
after the permission-mode dropdown replaced the bypass dialog, so tsc -b failed.
* Studio: cache a null default chat template so the viewer stops re-fetching it
A model with no sidecar or embedded template resolves to a terminal null, but
that result was never cached, so reopening the template viewer re-ran the
backend and Hugging Face lookup every time.
* Studio: detect direct-file GGUFs in run settings so Max Tokens uses their context
A GGUF loaded from a local file or custom folder has no variant label, so the
run-settings panel treated it as non-GGUF and clamped Max Tokens to the session
max_seq_length instead of the loaded GGUF context. Detect it via the reported
GGUF context and the .gguf checkpoint suffix, matching the chat page.
* Studio: prompt to re-select a local model file when its lease expired before reload
A file-picked GGUF is reachable only through a native path token that the
desktop host prunes after a TTL. Reloading reused that token blindly, so a
reload long after the initial load failed with an opaque error. Track the
token's expiry and, when it has passed, ask the user to re-select the file
instead of attempting a doomed reload.
* Fix descender-clipping test to tolerate sidebar layout utilities
The sidebar account-block div carries layout utilities (min-w-0, flex-1)
between 'flex' and 'flex-col', so the descender-clipping guard's regex,
which required 'flex' immediately followed by 'flex-col', no longer matched
and the test failed to locate the account-block div. Generalize the prefix
to allow intervening flex utilities while still capturing the leading-*
class before the collapsible visibility utility and asserting leading-tight,
so the guard against clipped glyph descenders is fully preserved.
* Harden picker chat-template resolution
Enforce the 64 KiB chat-template contract at the validate endpoint's request
model so a direct caller cannot submit a template far larger than the frontend
allows (MaxBodyMiddleware only bounds the whole request body, not this field);
oversized templates now return a clean 422.
Apply sidecar-over-GGUF template precedence globally across cached snapshots
instead of per snapshot. A repo with multiple cached revisions previously
returned the first snapshot's template, so a newer GGUF-only revision could
win over an older revision's maintained chat_template.jinja sidecar, which
contradicted the documented intent that sidecars supersede the embedded copy.
* Guard per-model config against future-schema and lossy migration
Two forward-compatibility gaps in the versioned per-model config store:
- The load/apply path returned and normalized a stored record without checking
its schema version, so a record written by a newer client was reinterpreted
under the current schema and applied to a live model load, even though save,
delete and eviction all refuse to touch future-schema records. Reject
future-schema records on load too.
- The one-time legacy migration enforced the storage budget without protecting
the entries it had just migrated and set the completion flag unconditionally.
When storage was already full of future-schema records (which are unevictable
by an older client), the migrated entries were the only evictable ones and
could be dropped while migration was still marked complete. Protect the
migrated keys during eviction and only mark migration complete when they
survive, so it retries once space frees up.
* Discard chat-template validation results after the dialog closes
Server-side template validation is async, but closing or cancelling the editor
did not abort it, so a late-arriving valid response still called onSave and
applied a template the user had already dismissed. Track a validation token
that is bumped on close and ignore any validation result whose token is stale.
* Record native lease expiry when loading a picked GGUF from the chip
The pending-native-model chip loaded via stageOrLoad directly, bypassing
loadNativeModelIntent, so activeNativePathExpiresAtMs was never recorded for a
chip-loaded file. A later reload then either skipped the lease-expiry guard
entirely (expiry left null) or compared against a previously loaded file's
stale expiry, so reload could reuse an already-pruned token or wrongly block a
still-valid one. Route the chip through loadNativeModelIntent, which builds the
same selection and records the expiry.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Prefer sidecar template for a directly selected local GGUF file
A direct .gguf file path read its embedded chat template without checking the
parent directory for a maintained sidecar (chat_template.jinja /
tokenizer_config.json), while directory and variant selections already prefer
the sidecar. That let the config editor preview or save a stale embedded
template for the same model depending on how it was selected. Check the parent
directory sidecars first, then fall back to the embedded copy, and cover both
paths with tests.
* Resolve cached chat template per revision, newest first
The earlier change searched every cached snapshot for a sidecar before
considering any snapshot's embedded GGUF template, which let an obsolete sidecar
from an older revision override the newest revision's template. Restore
per-snapshot resolution (newest first): a revision's sidecar still supersedes
its own embedded GGUF copy, but a newer revision is no longer overridden by an
older revision's sidecar.
* Preserve autoload transport conflicts and surface background busy downloads
- When a Hub autoload hits a transport conflict, keep pendingHubAutoLoad bound
instead of clearing it. Clearing it re-keyed the download surface and its
cleanup cancelled the conflict the toast tells the user to resolve, so the
Hub resume affordance was gone the moment it appeared. Return early on
conflict, mirroring the started branch, so resolving it from the Hub still
auto-loads on completion.
- The background-download branch handled started and conflict but silently
dropped a busy outcome, leaving the user with no feedback when a peer variant
of the same repo was already downloading. Surface the same busy toast the
autoload path uses.
* Fix context length, GGUF template, fetch state and lease expiry bugs
Keep explicit context length values instead of collapsing to null at
native. The collapse made the slider jump back at the native maximum
and made Reload load the previous context instead of the chosen one.
Prefer the first split when resolving a GGUF without a variant. Later
splits carry no chat template metadata, so picking the largest file
could return no template for a sharded model.
Clear stale fetch state when template and metadata lookups retry, so
a previous terminal error is not shown while a new fetch is running.
Record native path lease expiry together with the token when a load
commits. The expiry was written by only one load path and even when
the load did not start, so a reload could be blocked with an expired
file message for a still valid token.
* fix(model-picker): resolve review findings across config, inventory, and templates
- Apply remembered per-model config in the training-compare chat handoff so a
prior model's customContextLength no longer leaks into the next load
- Match GGUF variant labels with the inventory extractor too, so cached
no-quant-token files resolve their default chat template
- Show "Auto" instead of a fabricated 32768 when native context is unknown
- Reuse the identical staged auto-load object on same-pick so a re-pick during
download pre-flight no longer disarms auto-load via "busy"
- Union supports_vision when deduping cross-cache inventory rows
- Serve hidden-model needles from a new GET /api/hub/hidden-models endpoint and
merge them client-side, covering runtime-configured RAG embedders
- Clamp GET chat templates to MAX_CHAT_TEMPLATE_BYTES (route + jinja sidecar),
matching the validate endpoint's contract
- Lower-clamp stored customContextLength to shared CONTEXT_LENGTH_MIN
- Wipe unsloth_chat_load_on_selection in Settings "Reset all"
- Drop stale pendingHasContext comment describing deleted staging machinery
* Fix stale defaults cache, token in query string and rounded up context ceiling
Refresh cached chat template and max position data when a model update
completes. Send the HF token for model config requests in the dedicated
header instead of the URL. Snap the native sequence length ceiling down
to the nearest step so the slider cannot exceed the declared maximum.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix compare pane reverting active checkpoint on non-GGUF load
Re-read runtime params after setCheckpoint so the fresh checkpoint is
kept instead of being overwritten by the pre-setCheckpoint snapshot.
* Send the HF token via header for the vision and embedding checks
checkVisionModel and checkEmbeddingModel still passed the HuggingFace
token as a ?hf_token= query parameter, so it landed in server access
logs, proxy logs, and browser history. Move them to the
X-Unsloth-HF-Token header like getModelConfig already does, and accept
the header on the check-vision and check-embedding routes with the
existing query parameter kept as a fallback for older clients.
* Cap the chat template on the model load path
The load endpoint accepted an unbounded chat_template_override, so a
direct caller could hand llama.cpp an arbitrarily large Jinja template
even though the frontend, the validate endpoint, and the read paths all
enforce the 64 KiB limit. Reuse MAX_CHAT_TEMPLATE_BYTES in the
LoadRequest validator, rejecting oversized templates with a fast
character-count check before the exact UTF-8 byte check.
* Protect existing per-model configs during legacy migration
When the one-time legacy import pushes the store over budget, eviction
now protects the entries the user already has and drops only the
just-migrated legacy entries, so importing old load settings can never
discard a newer per-model config.
* Reset clears the context override instead of pinning the native value
Reset wrote the discovered native context into customContextLength for
GGUF models, but isDefaultConfig treats any non-null customContextLength
as an explicit pin, so Reset with Remember enabled persisted a fixed
context and future loads stopped using the native auto context. Reset
now restores the full default (customContextLength null); the native
value is still shown through the existing display fallback.
* Bound chat-template sidecar reads to a size limit
The chat_template.json, tokenizer_config.json, and Hub-downloaded sidecar
readers decoded and json-parsed the whole file before the extracted
template hit the 64 KiB response cap, so an oversized metadata file could
exhaust memory. Read them through a bounded reader (4 MiB envelope) that
returns None when the file is larger, matching the existing chat_template.jinja
size guard. Adds tests for oversized tokenizer_config.json and chat_template.json.
* Keep the native-path token and lease expiry in sync
Rollback after a failed reload restored the previous token but left the
failed load's expiry in the store, so a later reload could be falsely
blocked as expired (token A paired with load B's lease). Restore the
previous lease alongside the token, and clear the expiry wherever the
token is cleared on a non-GGUF transition, so the two never diverge.
* Clear the native file lease on compare-pane loads
* Studio: add regression tests for the model-picker per-model-config
Guard the specific regressions that reverted the predecessor change:
- backend pytest (studio/backend/tests/test_model_picker_regression.py):
infra-model hiding, HF token via header with query fallback, and the
chat-template byte caps.
- source contracts (tests/studio/test_model_picker_contracts.py): the token
stays out of the URL, the context ceiling is floored, the native lease is
cleared on compare-load and restored on rollback, the default caches key on
the inventory version, and the hidden needles stay present.
- Playwright E2E (tests/studio/playwright_model_config.py) wired into
studio-ui-smoke.yml on port 18898: Context Length persists across a reload,
Reset clears the stored override, and infra models are absent from the picker.
- optional GPU-gated inference smoke (tests/studio/test_gpu_inference_smoke.py)
that auto-skips on GPU-less CI and stays short on a GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: model pinning, row menus, hub inference settings, and inventory filters
Pinning
- Add a pinned models store (localStorage) with repo and per-quant pins
- Pinned section in the model selector's On Device list and the hub inventory,
with newest pins first so Pin to top lands on top
- Deleting a repo drops its pins
Row menus
- Replace loose row icons with a shared 3-dots menu (pin, reveal in file
manager, copy identifier, copy path, delete) on picker rows, hub quant rows,
the hub run bar, and on-device inventory rows
- Menus only render for models actually on disk; platform-aware reveal labels
- Backend: cached-model-path and reveal-cached-model endpoints resolving
managed HF-cache repos only
Hub inference settings
- Gear in the GGUF run bar opens an Inference settings dialog reusing the chat
page's controls: model config (context length, KV cache, speculative
decoding, chat template), system prompt, reasoning, sampling, tools and
retrieval
Inventory
- Model-type filter (text, vision, embedding, STT, TTS, diffusion) beside the
sort pill, both with a sort icon, capped widths and truncation so the
On device heading never wraps
- Unsloth-owned repos without an upstream provider logo fall back to the
Unsloth mascot avatar
- Discover / On Device tabs widened; hub search bar narrowed to match
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: revert the Unsloth mascot avatar fallback
Unsloth-owned repos without an upstream provider match go back to the
colored-initial tile, and unslothai is no longer a relabeled owner.
* Studio: run-bar options on single models, and aligned type/capability filters
- Give single-model (non-GGUF) run bars the same 3-dots options menu and
settings gear as GGUF, at repo level
- Drop Pin to top from the run-bar menus; pinning stays in the On Device list
- Add an Image to text (diffusion) capability with detection, and surface it
in both the hub Discover capability filter and the On Device type filter
- Align the On Device type filter with the Discover capability options and
share the same detection so both dropdowns match
* Studio: apply hub inference config on reload, eject action, and run-bar polish
- Fix inference settings not applying: the hub dialog now writes the config to
the runtime before reload, matching the chat page (selectModel reads runtime
state, not the selection)
- Order the settings gear before the 3-dots menu in the run bars
- Replace the loaded-model run-bar action (New Chat) with Eject, wired through
the inspector to the hub's ejectModel
- Truncate the results heading so a long search query clips instead of
overlapping the header pills in split view
- Use a plain magnifying-glass icon for the no-results empty state
* Studio: fix GPU settings loss, load guards, pins, filters, and cached paths
Reloading a model from the chat sidebar or the hub gear dialog rebuilt the
per-model config without the GPU memory fields, so manual GPU layers, MoE
placement, and the GPU pick were reset on every reload and could be saved
over a remembered config. The active config now comes from a shared
useActiveModelConfig hook that carries the GPU fields for GGUF models, and
the sidebar remount signature tracks them through a shared gpuFieldsSignature
helper.
The in-flight load guard lived in a ref inside each useChatModelRuntime
instance, so the chat page, hub page, and gear dialog could not see each
other's loads. A load started from the gear dialog left the hub page free to
eject the model mid-reload or start a second concurrent load. The runtime
store now records the loading pick, selectModel checks it across instances,
and ejectModel refuses with a toast while any load is in flight.
The cached-model-path endpoint matched GGUF files by basename and excluded
only mmproj, so Copy path and Reveal could return an MTP drafter for a quant
and returned 404 for directory layouts like BF16/model-00001.gguf. Variant
files are now resolved from snapshot-relative paths with the same drafter,
mmproj, and big-endian exclusions as the load path, shared through a new
_main_variant_gguf_label helper.
Hub and picker fixes:
- rename the diffusion capability label from "Image to text" to
"Image generation", since it detects image generators
- validate pinned quants through the cached variant listing, keep the last
verified set while revalidating, and drop deleted quants immediately
- pass a measured scroll margin to the on-device virtual list so rows past
the overscan stay visible below the pinned block
- keep the delete menu for stopped partial safetensors downloads
- give the inventory type filter a reset in Clear filters, a truthful empty
state with a Show all types action, and hide it on the datasets view
- order picker pinned rows by pin recency, include pinned matches in the
empty-state check, and sync pins across browser tabs
- count only the visible rows in the On device list header
Tests: contract checks for each fix in test_model_picker_contracts.py and a
backend test for the variant label selection.
* Studio: reveal cached models in Windows Explorer under WSL
The reveal endpoint only branched on macOS, Windows, and generic Linux.
Under WSL the Linux branch spawned xdg-open, which is missing on a stock
distro without a Linux desktop, so the request failed with a 500 and the
UI showed a failed to open file manager error.
WSL is now detected with the existing helper and the path is converted
with wslpath before opening explorer.exe, selecting the file the same
way native Windows does. Directories open directly. When interop is
unavailable the old xdg-open fallback still runs. The macOS, native
Windows, and native Linux branches are unchanged, and the Tauri app is
covered since its hub reveal calls this same local endpoint.
Tests: platform guards for the WSL reveal, the interop fallback, and
the unchanged native Linux behavior in tests/studio/test_reveal_file_manager.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Adjust model picker row spacing and cogwheel hover consistency
* Studio: exact hidden model ids and newest revision cached paths
A custom RAG embedder repo was published to the frontend as a basename
substring needle, so a generic name like org/model could hide unrelated
models in the pickers. The hidden-models endpoint now sends full repo ids
that are matched exactly.
Copy path and Reveal picked a GGUF variant from an arbitrary cache
revision when the same file existed in more than one. The newest revision
now wins, matching the whole repo lookup.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Fix model picker GPU config, metadata, and cache selection
Load each compare model with its saved GPU memory mode, GPU layers, CPU MoE layers, and selected GPU IDs. Reconcile saved GPU IDs with the current hardware. Include the active native GGUF path token in metadata checks. Search all Hugging Face cache roots when resolving cached models and select the largest visible cache entry. Remove obsolete barrel exports and the staging-only GPU memory helper.
* Studio: hide hub inference settings gear for now
The cogwheel in the hub download cards is out of scope for this PR. The
dialog component stays in place and a TODO marks where the button
returns in a future PR.
* Refresh hidden model matchers
* Fix GGUF detection, compare context pin, and picker delete staleness
Treat any pick with a GGUF variant as GGUF in selectModel so the first
load after downloading an uncached quant validates and sizes with the
right GPU settings instead of unloading the current model on a wrong
preflight. Variant picks now also set isGguf on their selection meta.
Stop compare panes from inheriting the active model's context pin when
their own saved config says Auto. Null context in a remembered config
now means no pin, matching how the pane settings are shown.
Route picker deletes through the hub inventory client, which
invalidates the HF cache scan and the variants cache. The legacy
delete route left the scan cache warm, so deleted models reappeared
in the picker until the TTL expired. Removed the now unused legacy
delete client and updated the contract test to match.
* Studio: fix stale GGUF load-marker ordering test
The load-in-flight marker still precedes the hub-download guard and the
unload, but the llama_extra_args inheritance that used to sit between the
marker and the guard now runs ahead of the GGUF branch, so it is no
longer a landmark inside the sliced source. Drop it from the ordering
assertion and keep the marker -> guard -> unload invariant.
* Studio: fix per-model config edge cases in compare loads and saved defaults
- chat-settings-sheet: gate the MTP fallback note and context/VRAM warning on
the broader isGguf (variant, loaded gguf context, or .gguf suffix) instead of
isLoadedGguf, so direct-file and custom-folder GGUF loads still surface
those diagnostics.
- shared-composer: a compare pane's context now comes from its own config only
(a saved pin, else null for Auto/native). It no longer inherits the active
model's shared snapshot, which resolveFitMaxSeqLength treated as an explicit
pin and could load a pane at another model's context (VRAM/OOM), matching the
single-model load path.
- model-config-page: when an auto-fit GGUF is saved with fixed GPU layers
(Manual) and Remember, pin the displayed fitted context so a later fresh load
keeps the placement instead of sending native/0 and recreating the OOM.
- per-model-config: treat Auto GPU memory mode and Auto/default speculative type
as follow-global defaults; do not persist them as per-model overrides so later
global preference changes keep applying.
* Studio: gate vision capability on GGUF projectors and bound remote template downloads
- cache_inventory: only mark a cached repo vision-capable when it holds an actual
GGUF mmproj projector, not any file whose name merely contains "mmproj" (e.g.
mmproj_config.json), matching the runtime's GGUF-only projector detection.
- picker/service: pre-check the remote file size before downloading an uncached
repo's chat template / tokenizer config, so a maliciously large sidecar is
skipped instead of fetched and retained in full, mirroring the size gate the
local-file path already applies.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add source-contract guards for the per-model-config edge-case fixes
Guard the four per-model-config fixes against silent regression in CI:
- local GGUF diagnostics gate on the broad isGguf, not the variant-only isLoadedGguf
- fixed-layer GGUF saves pin the displayed context
- Auto GPU mode and Auto/default speculative are not persisted as per-model overrides
- a compare pane's context comes from its own config, not the active model's snapshot
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: clear manual GPU knobs on Default and resolve local embedders before repo-id
- model-config-page: switching GPU Memory back to Default now clears the Manual-only
knobs (gpuLayers/nCpuMoe/selectedGpuIds); otherwise a remembered config kept stale
pins that a later load re-applied when the global GPU preference was Manual, despite
the page showing Default.
- routes/models hidden_model_matchers: resolve an existing local path before the repo-id
regex, mirroring is_hidden_model, so a local embedder shaped like "models/embedder" is
hidden by exact path instead of leaking as a chat model.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: add _is_mtp_drafter to the model_config stub in the export-paths test
routes/models.py imports _is_mtp_drafter from utils.models.model_config at module
load, but the lightweight stub in test_export_absolute_paths.py did not provide it,
so loading the module under the stub raised ImportError on Backend CI. Add the stub.
* Studio: read a picked GGUF's chat template through the native path lease
The picker chat-template GET has no native-path-lease plumbing, so a
desktop-picked (drag-drop) GGUF could not show its default chat template
in Run Settings until the model was loaded: the endpoint only receives
the display label, not the leased file path.
Read the embedded template through the existing lease-aware
/api/inference/validate probe instead. A new include_chat_template flag
resolves the granted canonical path and returns the GGUF's own embedded
template, never a sibling sidecar (the grant authorizes just that one
file); it skips the training guard like include_context_length and is
bounded by MAX_CHAT_TEMPLATE_BYTES. The frontend fetch mints a one-shot
validate-model lease when a native token is present and keeps the plain
GET path for HF and allowlisted local models.
Adds backend and source-contract regression tests.
* Studio: call worker.direct_wheel_url in the ROCm wheel-url test
The ROCm Mamba/SSM test referenced worker.py's private _direct_wheel_url,
but the worker imports the wheel helper under its public name
direct_wheel_url (utils.wheel_utils). When the worker module loads (its
imports resolve in CI), worker_mod._direct_wheel_url raised AttributeError;
the test only masked it by skipping when the worker could not be imported.
Call the name that actually exists so the assertion runs; it still returns
None for an empty cuda_major (ROCm).
* Studio: reset max sequence length to the app default, not the loaded value
For a non-GGUF active model, the per-model config seeds maxSeqLength from
the loaded runtime value so the panel opens showing the running context.
Reset set config.maxSeqLength to null, but the null fallback resolved back
to that captured runtime value, so the field kept showing the old custom
length and the config saved/reloaded it again. A remembered or active
max-length override therefore could not be cleared from Run settings.
Fall the null/default case back to the app default (clamped to the model's
native ceiling) instead of the active runtime snapshot, so Reset actually
clears the override. The initial view is unaffected: an active model's
config.maxSeqLength is already non-null, so it still shows the loaded value.
Adds a source-contract regression guard.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: persist default max length, refresh deleted quants, hide non-chat locals
Three follow-up fixes from review of the per-model-config picker:
- Max sequence length: the persisted per-model record now keeps config's
maxSeqLength (null after Reset) so isDefaultConfig can clear a remembered
override; the resolved app-default is substituted only into the load
request, never the saved record. Previously Reset saved the concrete
default and left the model pinned/remembered.
- GGUF variant expander: deleting a downloaded quant from a repo that still
has other cached quants now bumps the expander refresh key, so the removed
quant stops showing as downloaded and clickable (which would try to reload
the deleted file) until the repo is collapsed and reopened.
- Local picker rows: require capabilities.canChat before listing a local
models-folder / LM Studio row. A weightless folder (only config.json) is
classified non-chat, and toLocalModelInfo drops capabilities, so selecting
such a row would try to load a path the inventory already marked non-chat.
Adds source-contract regression guards for all three.
* Fix compare-pane and Reset context defaults in model picker
Two related per-model-config default regressions:
- A non-GGUF compare pane with no saved maxSeqLength fell back to the
active model's shared runtime snapshot, so comparing a saved 128K model
against an unconfigured pane loaded the latter at 128K and could OOM. It
now falls back to the shared app default (DEFAULT_MAX_SEQ_LENGTH), the
same fallback the single-model config path uses.
- contextAtDefault treated an explicit customContextLength equal to the
native ceiling as a default, which wedged the Reset button disabled for
a deliberate pin-to-native. It now counts as default only when there is
no override at all.
DEFAULT_MAX_SEQ_LENGTH becomes a single exported constant in
per-model-config.ts so the single-model config and the compare path share
one source of truth. Adds source-contract guards for both fixes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Skip over-cap remote Jinja templates so the tokenizer template wins
The remote chat-template resolver bounded raw chat_template.jinja downloads
only by MAX_TEMPLATE_METADATA_BYTES (4 MiB), then returned the first
non-empty Jinja unconditionally. The picker route drops any template larger
than MAX_CHAT_TEMPLATE_BYTES (64 KiB), so an uncached repo whose
chat_template.jinja sits between 64 KiB and 4 MiB returned no template at
all, even when a valid smaller tokenizer_config.json template existed. The
local path already skips oversized .jinja files and falls through.
Gate the extracted Jinja on MAX_CHAT_TEMPLATE_BYTES and continue searching
when it exceeds the cap, matching _chat_template_from_jinja_file. The 4 MiB
download bound stays for JSON files that merely embed a small template. Adds
a regression test that a big Jinja plus a valid tokenizer config resolves to
the tokenizer template.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard legacy per-model-config migration idempotency
The v1->v2 localStorage migration (unsloth_load_settings ->
unsloth_model_configs) runs on every store read, so it must migrate exactly
once and never re-run, duplicate, or clobber a newer per-model config on a
reload or restart. That was covered only by a manual proof, so add durable
guards:
- Source-contract test pinning the three idempotency layers (the in-memory
legacyMigrationChecked guard, the persistent unsloth_model_configs_migrated
flag set in every terminal branch, and the non-overwriting Object.hasOwn
merge-skip) plus the readMap invocation. Reddens if any layer is dropped.
- Playwright model-config E2E: promote the legacy-migration step to a gating
check (soft_fail, which gates under the CI STUDIO_UI_STRICT=1) that the
migrated value is preserved and the flag is set, then reload again with a
fresh legacy seed present and assert the stored key set is unchanged, so a
second reload cannot re-migrate, duplicate, or clobber.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Note the migration E2E now gates idempotency under STUDIO_UI_STRICT
* Tighten model-picker per-model-config code comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: shimmyshimmer <shimmyshimmer@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio: GPU memory dropdown — llama.cpp --fit on and manual gpu-layers/cpu-moe
* Studio: simplify GPU memory changes (reuse ParamSlider, GPU_LAYERS_ALL, loadedGpuMemoryFields helper)
* Studio: GPU picker — choose which GPUs a GGUF model loads on (gpu_ids)
* Studio: simplify GPU picker (share /api/system fetch, validate gpu_ids)
* Studio: GPU picker review fixes (gate relative indices, no cross-model leak, validate, types)
* Studio: group GPU controls under a collapsible GPU section
* Studio: GPU feature review fixes (fix fit-ctx test, behavior-test the floor, comment accuracy)
* Studio: make GPU a top-level settings section (not nested under Model)
* Studio: flatten GPU controls into the Model section, group by GPU/context/generation
* Studio: move GPU Memory to the bottom of Model with its dependent controls beneath it
* Studio: move GPU Memory below Tensor Parallelism and GPUs below GPU Memory
* Studio: tighten GPU Memory and GPU Layers tooltip copy
* Studio: fix fit-mode context slider track-click, restore GPU Memory tooltip, shorten fit dropdown label
* Studio: GPU Memory tooltip one mode per line, briefer
* Studio: note HIP_VISIBLE_DEVICES (ROCm) in the GPUs picker tooltip
* Studio: narrow the GPU Memory dropdown to fit the shortened label
* Studio: use 'llama.cpp --fit' in the GPU Memory tooltip for consistency
* Studio: allow Tensor Parallelism in Manual GPU mode
* Studio: graduated MoE-on-CPU offload (--n-cpu-moe) replacing the all-or-nothing toggle
* Studio: size the MoE-offload slider for staged (deferred-load) models
* Studio: share one GGUF header walk for the context-length and MoE-count readers
* Studio: size the GPU Layers slider for staged models (one staged-header read)
* Studio: move Tensor Parallelism below the GPUs picker
* Studio: GPU split (--tensor-split) per-GPU model share in Manual mode
* Studio: tolerate whitespace in GPU split input, move it below GPU Layers
* Studio: rename the GPU split control to "Split ratio"
* Studio: Split ratio sends explicit even input; fix blank=free-VRAM (not even) copy
* Studio: tighten llama.cpp --fit VRAM margin with --fit-target 512
* Studio: GPU memory review fixes (rollback re-baseline, single-GPU TP gate, accurate copy)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: move Split ratio below MoE Layers on CPU
* Studio: address PR review (fix GPU-info hydration race, share fit context-length across load paths)
* Studio: address codex review (manual single-GPU TP guard, GPU-aware spec defaults in fit/manual, GGUF-only context/preference)
* Studio: address codex review round 2 (gpu_present seed, single-GPU tensor-split guard, staged manual-knob reset, strip inherited offload flags)
* Studio: address codex review round 3 (strip inherited --n-cpu-moe, CPU-fallback warning in Manual mode)
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: address codex review round 4 (preserve pinned fit context across a later Apply)
* Studio: address codex review round 5 (honor GPU picker for diffusion GGUFs, clear fit pin on cross-model switch)
* Studio: preserve the pending GPU Memory mode when staging a model
* Studio: pin diffusion GPU device order and reset GPU-memory state for diffusion loads
* Studio: address codex review round 6 (fit-Auto rollback context, preserve manual non-tensor split modes, persist GPU mode on load not select)
* Studio: persist the applied GPU Memory mode, not the requested one (skip diffusion loads)
* Studio: replace Manual-mode split-ratio field with per-GPU layer sliders
* Studio: clarify per-GPU layer split hint for tensor-parallel mode
* Studio: address codex review round 7 (allow GGUF gpu_ids past the legacy guard, replay GPU-memory fields on respawn)
* Studio: address codex review round 8 (size the validate preflight like the load in fit mode, across both load paths)
* Studio: skip the training-OOM guard for llama.cpp --fit GGUF loads (they spill to RAM)
* Studio: drop the now-redundant compare-path validate sizing (the --fit guard skip makes it moot)
* Studio: address codex review round 9 (keep the training guard for fit loads, forward gpu_ids to validate, strip inherited manual tensor-split)
* Studio: address codex review round 10 (gate GPU-memory adoption on is_gguf, record manual knobs only in Manual mode)
* Studio: handle diffusion GGUFs symmetrically in the GPU Memory controls (preserve the standing mode preference, hide the inapplicable mode/TP controls)
* Studio: remember the GPU Memory settings per model
* Studio: consolidate --fit mode and Manual mode into a single Manual mode
* Studio: preserve the per-GPU layer split across GPU Layers changes
* Studio: trim overly long GPU Memory comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address GPU memory config review comments
* trim redundant GPU memory tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reconcile manual-mode TP drops with the #6659 drop-site invariants
* Preserve quantized KV in manual --fit, charge GGUF companions in full, reconcile GPU pick on load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Clear stale GPU baseline on non-GGUF loads so it can't read as dirty
* Fix no-context-shift test for the conditional -c flag
* Credit manual GPU-layer offload for cached HF GGUFs
* Reset per-model load knobs on GGUF quant switch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Strip inherited tensor-split when manual ratio is cleared
* Match auto-load validation to safetensors placement
* Reset editable manual knobs after Auto GGUF loads
* Record a single device for diffusion GPU picks
* Reset per-model GPU knobs before applying saved settings
* Address review comments
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Guard manual tensor splits and keep remembered context on auto-load
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Snapshot compare knobs, seed splits from free VRAM, flag zero-offload loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Exempt CPU-only loads from the guard floor and harden compare and reseed paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reach full offload from the layers slider and charge extras drafters in the guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Warm the GPU device cache before pick reconciles and disable staged GPU controls
* Align the training guard with inherited extras, spec mode, and compare targets
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Hide GPUs from companion-less zero-offload loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size diffusion picks per device, own manual offload flags, reject XPU picks
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop tensor flags at zero layers and exempt CPU-pinned drafters
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Allowlist the zero-layer tensor parallel drop site
* Keep validate and load guards on the same extras and refresh stale baselines
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop mismatched manual tensor splits before launch
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate XPU picks on the real backend field and harden split and hydration paths
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Weight full GPUs as zero, clamp split shares, and refine the zero-layer mask gate
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Carry fit context across mode changes and align drafter and picker gates
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Catch variant switches, uncached diffusion repos, and text-only mmproj skips
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check companions on the first device and size native and remote zero-layer loads
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Replace the training guard's precise VRAM modeling with a conservative bound
* Baseline context pins on non-GGUF hydration and reprobe list-seeded staged GGUFs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size manual splits by their largest share and preserve resolved context from Default
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Default-deny unsized required companions and price KV at the effective cache dtype
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve MTP draft KV and MLA target-copy in the training guard
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Size tensor-parallel loads per device and show GPU controls for native GGUFs
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve MTP overhead for uncached remote GGUFs and the mmproj runtime factor
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Drop the training-coexistence VRAM estimation this PR added
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate remembered load settings to GGUF picks
* Lock the remaining load-time controls during a staged load
* Clear the stale native-path token on compare loads
* Drop a stale guard reference from the zero-offload masking comment
* Seed GPU baselines from the rollback response and drop never-emitted offload flags
* Match validate's training guard to load and keep the native reload token
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim verbose GPU-memory comments
* Thread the variants header walk off the event loop, honor device pins on zero-offload, and hold staged GPU edits
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor manual placement and classify pinned zero-offload loads
* Close diffusion admission and status hydration gaps
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check the actual diffusion GPU during training
* Align staged baselines and manual reload dedupe
* Fix GGUF placement and rollback state
* Harden manual GGUF placement boundaries
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove unused resolve_tensor_parallel import in llama_cpp.py
The name is used only in llama_server_args.py, routes/inference.py, and tests,
not in llama_cpp.py; the unused hoisted import trips the import-hoist verifier
in the source-lint CI job.
* Fix diffusion GPU dedup and training guard for non-numeric device tokens
The diffusion runner drives only its single lowest device and the backend
records that one device (self._gpu_ids = [sorted(gpu_ids)[0]]), but the reload
dedupe compared it against the full requested list, so a multi-GPU pick that
resolves to the same device forced a needless reload. Normalize the request the
same way for a loaded diffusion model in both _already_in_target_state and the
route _request_matches_loaded_settings.
The chat-during-training coexistence guard called int() on the single-device
token and hard-rejected when it could not parse. A non-numeric token (a CUDA
UUID / MIG handle) now sizes against the whole visible pool like the GGUF guard
instead of falsely blocking the load, and an empty token (a CPU-only runner such
as a CPU diffusion GGUF) is allowed outright since it uses no GPU VRAM.
* Tighten comments added by the GPU memory config changes
* Harden GGUF placement from independent review: VRAM sizing, diffusion TP reset, tensor_split validation
- Training coexistence guard: a single-device runner pinned through an
unresolvable UUID/MIG token was sized against the aggregate visible-VRAM pool,
so a load could pass on capacity it cannot use and then OOM active training.
Size against the worst-case visible device (min free) instead, keeping the
guard's documented default-deny contract. The empty-token (CPU-only runner)
allow path is unchanged.
- Diffusion startup: _start_diffusion_server now resets self._tensor_parallel to
False alongside the other placement resets. A prior tensor-parallel chat load
(process killed but not fully unload-reset) otherwise left /status misreporting
tensor parallelism and made an identical diffusion re-Apply reload against the
stale state.
- tensor_split: reject negative / non-finite / all-zero splits up front. They
were dropped at launch but still compared raw in the reload dedupe, so an
identical Apply reloaded indefinitely.
- Tests: the shared httpx stub was incomplete and, installed via setdefault
before real httpx loaded, broke a combined pytest run (collection errors on
httpx.Response). Import the real installed httpx instead.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <unslothshared@gmail.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
* Replace standalone Studio wording with Unsloth
Replace the single word Studio with Unsloth wherever it is used as
shorthand for Unsloth Studio in docs, CLI output, UI strings, i18n
locales, workflow display names, comments and docstrings.
Kept unchanged: the full name Unsloth Studio, third party product
names (LM Studio, Visual Studio, Mac Studio), feature names
(Recipe Studio, Fine-tuning Studio and its translations), and all
identifiers such as env vars, commands, paths and filenames.
* Address review feedback on the Studio wording rename
Use "an" before Unsloth where the rename left the article as "a".
Restore the split brand where Unsloth and Studio render as two halves
of the full product name: the onboarding sidebar subtitle and the
IPv6 localhost warning. Scope two messages to the full name Unsloth
Studio where plain Unsloth was misleading: the AMD README bullet and
the CLI studio setup error.
Two live-test findings on the video progress endpoints:
- load-progress downloaded_bytes froze mid-download: the counter used
scan_cache_dir, which skips in-flight *.incomplete blobs, so it sat at the
last completed blob for the whole multi-GB shard pull while the disk kept
filling. Count the repo's cache directory directly (completed plus incomplete
blobs, snapshot symlinks skipped so nothing is double-counted).
- generate-progress reported total_steps=null / fraction=0 while step advanced:
the video API only carried the native total field while the image API exposes
total_steps and fraction, so one poller could not work against both. Derive
the image-compatible aliases in generate_progress and declare them on the
response model; the native total stays for back-compat.
Adapters are baked at load time: they attach to the dense transformer,
then quantize_ converts only the frozen base linears (the lora_ side
path is excluded by name), then the loader compiles. Post-quant PEFT
injection is not possible on a manually quantized module, so the
prequant shortcut is skipped for a baked load and the memory plan is
sized for the dense build (force_dense on the quant candidate).
At generation time the baked topology is frozen: weight tweaks and
disabling (scale 0 reproduces the quantized base exactly) go through
set_adapters, while adding or removing adapters returns a clean 400
telling the client to reload with the new selection.
supports_lora now returns True for int8/fp8 diffusers loads (checked
before the gguf-kind early return, since the quant fast path keeps the
picker kind); nvfp4/mxfp8 and GGUF-via-diffusers stay blocked. The
load request model takes an optional loras list, threaded through
begin_load on both engines (native ignores it and keeps applying LoRA
at generation).
Verified end to end on GPU: Z-Image GGUF picker + int8 + trained
adapter loads through the API, bake marker logged, weight 1.0 vs 0
renders differ visibly, weight 0.5 accepted live, unknown adapter
rejected as 400. Affected suites: 296 passed.
* Studio: permission levels for chat tool calls (Ask, Approve for me, Off, Full access)
Replace the Bypass permissions on/off toggle with a four level permission
selector, available in Settings > General (new Permissions section above
Notifications), the chat settings panel, the composer plus menu, and a new
always visible composer pill.
Levels:
- Ask for approval: every local tool call pauses for allow/deny.
- Approve for me: only calls detected as potentially unsafe pause; the
python/terminal sandbox stays on.
- Off: never pauses; sandbox stays on (previous default behavior).
- Full access: never pauses and the sandbox is disabled. Still requires
the danger confirmation and is never restored across reloads.
Backend adds permission_mode to the OpenAI compatible and Anthropic
passthrough payloads and threads it through both tool loops. Auto mode
uses a fail closed classifier in tools.py: terminal commands must be on
a read only allowlist with no redirection or substitution, python code
is AST scanned for writes, exec, process and network use, MCP tools
auto run only with read only style names. Unknown tools always ask.
Legacy bypass_permissions and confirm_tool_calls keep their exact
behavior for existing API callers.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio permissions: Off is a plain toggle below Full access
Off moves to the bottom of the level menu with a short description and
acts as the feature-off state: the composer pill is hidden entirely
while Off, and reselecting the active level toggles back to Off.
* Studio permissions: higher contrast composer pill text
The permission pill uses a foreground based grey instead of the shared
muted pill color, so it reads darker in light mode and lighter in dark
mode. Full access keeps the danger yellow.
* Studio permissions: panel dropdown layout and shorter tooltip
Chat settings panel: the Bypass permissions label sits on one line with
a full width dropdown underneath, styled like the other panel selects.
Tooltip shortened and wording uses Unsloth instead of Studio.
* Studio permissions: harden auto-mode unsafe detection
Extend the Approve for me classifier to catch write and exec paths that
slipped through:
- terminal: sort -o, tree -o, xxd -r, find -exec/-execdir/-ok/-delete
and find -fprint/-fprintf/-fls now ask; plain read-only forms still
auto-run. awk is no longer allowlisted since its program can write and
call system().
- python: from-imports of mutating names (from os import remove [as rm])
and star imports now ask.
Found by a fuzz and edge-case simulation matrix; pinned in
test_permission_mode.py.
* Studio permissions: split multi-line terminal commands in auto detection
A shell runs each line as its own command, but shlex reads newlines as
whitespace, so "ls\nrm -rf x" demoted rm to argument position and
auto-ran. Normalize newlines and CR to separators, and treat any all
separator token as a command boundary so runs of blank lines still
split. Found by the simulation matrix; pinned in tests.
* Studio permissions: address review feedback on auto-mode detection
Auto-mode (Approve for me) safety classifier hardening:
- Python: flag any reference to a mutating attribute, not only direct
calls, so indirect refs (f = os.remove; f(x)) and aliases ask. Detect
Path.open(mode) write modes and wrap the AST walk to fail closed.
- Terminal: match attached short output flags (sort -o/tmp/out) and keep
find context across grouping parens so find ( -delete ) asks.
- Both: ask before reads that escape the sandbox workdir via parent
traversal or hit credential paths (.ssh, .aws, id_rsa, .pem, etc.).
permission_mode plumbing:
- Fold permission_mode=full into bypass_permissions at the request model
so route-level confirm-gate guards see it as bypass.
- Reject ask/auto on the Anthropic Messages server-tools path, which has
no confirmation channel (mirrors the confirm_tool_calls rejection).
- Keep forced RAG autoinject in auto mode: the safe search_knowledge_base
retrieval never gates, so derive the skip from the real confirm need.
- Reset all local preferences now also clears the legacy confirm key so a
reset restores the fresh default instead of the old level.
Regression tests added for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio permissions: close auto-mode classifier gaps from review round 2
Auto mode ("Approve for me") let a few mutating calls through as safe:
- os.open(...) always creates/writes a descriptor, so treat it as unsafe
even though builtin open in read mode stays safe.
- fd -x/--exec/-X/--exec-batch runs a command per match; scan for these
alongside find's -exec/-delete.
- tempfile writes artefacts and hands back writable handles, so importing
it now asks.
- Calling the result of a call (getattr(os, "remove")("x"), partials) is a
dynamic target the AST can't vet, so fail closed.
- An MCP tool whose name pairs a read verb with a mutating one
(get_or_create_issue, read_and_delete_file) no longer auto-runs on the
read prefix alone.
Also fold permission_mode="off" into confirm_tool_calls=False on both
request models so the non-stream route guard sees the disabled gate, and
drive the Confirm tool calls toggle off permission_mode="ask" so auto no
longer shows it on.
* Harden auto-mode classifier and normalize bypass to full for PR #7079
Approve for me now asks for a few cases it previously auto-ran:
- os.open via an os alias (import os as o; o.open(path, O_CREAT))
- pathlib symlink_to / hardlink_to / link_to
- importlib.import_module dynamic imports
- os.mkfifo / os.mknod / os.utime
Also fold bypass_permissions into full when a stale ask/auto permission_mode
is sent alongside it, so the Anthropic route guard no longer 400s those legacy
callers. Adds classifier and request-model regression tests.
* Close more auto-mode classifier gaps for PR #7079
Approve for me now asks for cases the review surfaced:
- builtin open aliased to a name (f = open; from builtins import open as w)
or looked up dynamically (globals()['open'])
- pickle / marshal / shelve / dill deserialization
- io.FileIO write handles
- sort --compress-program (runs an external program)
- MCP names carrying save/archive/submit/commit/push/sync/register verbs
Also refine the attribute open() write check so an explicit read mode
(ZipFile.open(name, "r")) stays auto while os.open flags still ask. Adds
test coverage for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close three more auto-mode gaps for PR #7079
- rg runs an arbitrary program per file via --pre / --hostname-bin, so
"Approve for me" now asks for those flags (rg is on the read-only
allowlist).
- A path-qualified command token (./ls, /tmp/cat) is an arbitrary
executable, not the trusted utility its basename matches, so it asks
before running.
- A direct /chat/completions caller that sets permission_mode ask/auto
but omits the legacy confirm_tool_calls flag now self-enables the
confirmation gate, so tools can no longer run ungated on that path.
Adds classifier and request-model tests for each case.
* Close auto-mode classifier gaps from review round 3 for PR #7079
Approve for me now asks for cases the latest pass surfaced:
- short-option clusters bundling a write flag (sort -uo out => -u -o)
- procfs reads that leak a process env/args/memory
(cat /proc/self/environ, /proc/PID/cmdline, maps)
- env-assignment prefixes that change command lookup/loading
(LD_PRELOAD=x ls, PATH=. ls, IFS=x ls); benign FOO=1 cmd stays auto
- os.open imported as a bare callable (from os import open as o)
Also drops ps from the safe terminal allowlist: its BSD environment
flags (ps auxe, ps eww) dump a parent process's unscrubbed env and
cannot be flag-parsed reliably, so ps always asks now. Adds classifier
tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 4 for PR #7079
Terminal (Approve for me now asks for these):
- cd dropped from the safe allowlist: cd /; cat etc/passwd moves the
shell out of the session workdir so a later relative read escapes it
- env -C/--chdir (workdir escape) and -S/--split-string (builds a fresh
command line); wrapper flags are now checked
- /etc//passwd and /etc/./passwd normalize to /etc/passwd before the
sensitive-path scan
- a sensitive path split across an assignment and an argument
(p=/etc; cat $p/passwd) via best-effort NAME=value expansion
Python:
- builtins.exec / builtins.eval attribute calls (dynamic code execution)
- destructured open aliases (f, _ = (open, print); f('out', 'w'))
- a sensitive path composed from literals (os.path.join('/etc','passwd'),
'/etc' + '/passwd')
- ZipFile/TarFile write modes (ZipFile(name, 'w')); the reader stays auto
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 5 for PR #7079
Terminal (Approve for me now asks for these):
- procfs reads hidden by shell quotes (cat /proc/$PPID/enviro''n) or
quoted/nested-variable assignments (p="/proc/$PPID"; cat $p/environ):
quotes are stripped and NAME=value prefixes expanded before the scan
- LESSOPEN/LESSCLOSE, which make less run an input preprocessor command
Python:
- os.chdir / os.fchdir, which move the cwd so a later relative read
escapes the sandbox workdir
- sensitive paths composed via a pathlib / chain (Path('/etc') / 'passwd')
or an f-string of literals (f'/proc/{pid}/environ')
- runpy (import) and runpy.run_path / run_module, which run arbitrary code
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 6 for PR #7079
Approve for me now asks for these:
- a mutating callable reached through a getattr alias
(rm = getattr(os, "remove"); rm("f")): calls through a getattr-bound
name fail closed
- compound MCP tool names carrying clone/checkout/comment/fork/tag/
invite/share, which start with a read verb but still mutate
- a sensitive path hidden behind a glob (cat /e??/passwd,
cat /e[t]c/passwd): a ? / * / [..] token is matched against the
sensitive-file set and bracket classes are de-obfuscated; benign
globs (ls *.py) stay auto
Also run first-pass RAG retrieval in off mode: like auto, off never
prompts, so a direct caller passing a stale confirm flag should not lose
document retrieval (both tool loops).
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 7 for PR #7079
Approve for me now asks for these:
- __builtins__.exec / __builtins__.eval (dynamic code via the dunder)
- terminal reads that hide a credential path behind a backslash escape
(cat /et\c/passwd)
- read-named MCP filesystem calls pointed at a credential path
(mcp__fs__read_file {"path": "/etc/passwd"})
- compound MCP names carrying append / prepend
- open aliased through a subscript or builtins attribute
(f = globals()["open"]; f = builtins.open) then called to write
- open(..., **{"mode": "w"}) where a kwargs splat hides the write mode
- a sensitive path with a dynamic segment (open(f"/etc/{name}"),
os.path.join("/etc", name)); /tmp/{name} stays auto
- urllib3 networking
Also stop folding permission_mode ask/auto into confirm_tool_calls for
external-provider requests: that branch rejects confirm_tool_calls with
tools, and the mode only governs local tool calls. Local requests still
self-gate. Adds tests for each case.
* Close auto-mode classifier gaps from review round 8 for PR #7079
Approve for me now asks for these:
- dbm on the unsafe-module list: dbm.open(file, "c"/"n") creates files,
and importing the family signals a persistence writer
- reads of ~/.azure and ~/.config/gh credential stores (Azure/GitHub
tokens), in terminal, MCP arguments, and Python literals
- compound MCP names carrying upsert / assign
Adds classifier tests for each case.
* Gate secret mounts and fix the composer pill count for PR #7079
- Add Docker/Kubernetes secret mount dirs (/run/secrets,
/var/run/secrets) to the sensitive-path checks, so Approve for me asks
before reading injected credentials (terminal, MCP args, Python).
- Count the always-visible permission pill in the composer's compact
threshold so labels collapse at the intended width instead of
overflowing by one pill.
Adds classifier tests for the secret mount paths.
* Close auto-mode classifier gaps from review round 10 for PR #7079
Approve for me now asks for these:
- qualified pathlib constructors (pathlib.Path('/etc') / name), folded
the same as bare Path(...), so a dynamic sensitive path is detected
- open aliased through an annotated assignment (f: object = open;
f('out', 'w')), tracked like a plain assignment
- recursive searches rooted at an absolute path (grep -R TOKEN /home,
rg TOKEN /, fd pattern /etc), which read host files outside the
sandbox tree; sandbox-relative searches stay auto
Adds classifier tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 11 for PR #7079
Approve for me now asks for these terminal reads, which bash would
expand into a sensitive path only after the classifier had approved:
- a glob that resolves into a secret mount or credential dir
(cat /r?n/secrets/hf_token, cat /root/.s??/id_rsa)
- a recursive search rooted at a tilde home (grep -R TOKEN ~root,
grep -R TOKEN ~/logs)
- a brace expansion that builds a credential path (cat /etc/pass{w,}d)
- a default/alternate parameter expansion that builds one
(cat /etc/pass${x:-wd})
- an input redirection that hides a glob (cat </e??/passwd)
And these python calls:
- a str.format-built sensitive path (open('/etc/{}'.format('passwd')))
- writer methods that persist to disk without open() (numpy.save,
Image.save, plt.savefig, DataFrame.to_csv, json.dump)
Segment-wise directory matching keeps benign globs (ls /home/*/projects)
auto. Adds regression tests for each case and its safe counterpart.
* Close auto-mode classifier gaps from review round 12 for PR #7079
Approve for me now asks for these too:
- a terminal read whose parent traversal hides behind a redirection with
no following space (cat <../../notes)
- a python read whose path is built with str.join
(open(''.join(['/etc', '/passwd']))), told apart from os.path.join
- a dynamic-code builtin reached through an alias
(from builtins import eval as e; e(...); x = builtins.exec; x(...))
Adds regression tests for each case and its safe counterpart.
* Close auto-mode classifier gaps from review round 13 for PR #7079
Approve for me now asks for these too:
- a recursive search whose root is hidden behind an assignment
(p=/; grep -R TOKEN $p): the recursive-root test now runs on the
assignment-expanded tokens as well
- a python read whose sensitive path is split through a literal variable
(base = '/etc'; open(base + '/passwd')), including via an f-string
- numpy ndarray.tofile, which persists without open()
- a sequence brace read (cat /etc/pass{w..w}d), expanded alongside the
comma brace form before the sensitive-path scan
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 14 for PR #7079
Approve for me now asks for these python reads that assemble a sensitive
path in a form the fold did not yet recognize:
- a pathlib object reused through a name (p = Path('/etc'); p / 'passwd')
- old-style percent formatting ('%s/%s' % ('/etc', 'passwd'))
- Path.joinpath ('/etc'.joinpath('passwd'))
- a bytes path literal (open(b'/etc/passwd'))
And these terminal reads, which bash expands into a sensitive path only
after the classifier had approved:
- a substring parameter expansion off an assignment
(p=passwd; cat /etc/${p:0:6})
- an ANSI-C quoted path (cat $'/etc/pass\x77d')
- a glob into an Azure or GitHub CLI config dir
(cat /home/*/.az?re/..., cat /home/*/.config/g?/...)
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 15 for PR #7079
Approve for me now asks for these terminal reads, which bash expands into
a sensitive path only after the classifier had approved:
- a per-thread procfs env alias (cat /proc/$PPID/task/$PPID/environ)
- a recursive root behind a default parameter (grep -R TOKEN ${root:-/home})
- a path built by pattern replacement (p=passXd; cat /etc/${p/X/w})
And these python reads:
- a pathlib .parent/.parents chain that escapes the session workdir
((Path.cwd().parent / 'other' / 'notes').read_text())
- a sensitive path resolved through glob (glob.glob('/e??/passwd')[0])
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 16 for PR #7079
Approve for me now asks for these terminal reads, which bash expands into
a sensitive path only after the classifier had approved:
- a case-modifying parameter expansion (p=PASSWD; cat /etc/${p,,})
- a mutating find action hidden behind an assignment (f=-delete; find . $f)
- a glob assembled through an assignment (g=e??; cat /$g/passwd)
- a POSIX bracket class glob (cat /etc/pass[[:lower:]]d)
And these python reads/writes:
- a glob pattern folded from a literal variable
(base='/e??'; glob.glob(base + '/passwd'))
- a directly imported os.path.join (from os.path import join; join('/etc', 'passwd'))
- a directly imported writer (from numpy import save; save(...))
- an aliased pathlib constructor (from pathlib import Path as P; P('/etc') / 'passwd')
The find/fd and glob scans now run on the assignment/parameter-expanded
command, and pathlib/join/writer import aliases are tracked. Adds
regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode gaps from review round 17 for PR #7079
Two fixes:
- Gate sqlite3 in auto mode. sqlite3.connect(path) creates or mutates a
database file (and runs DDL/DML) with no open()/writer attribute for
the AST checks to catch, so treat the module like dbm and ask.
- Only self-enable confirm_tool_calls for Studio's own tool loop. The
ask/auto fold previously set confirm on every non-provider request,
including a plain client-tool passthrough (client-supplied tools that
Studio does not execute), which then tripped the local-tool
streaming-confirm route guard and rejected the passthrough. Restrict
the fold to requests that actually ask Studio to run tools
(enable_tools / enabled_tools / mcp_enabled).
Adds regression tests for the sqlite3 write and for the passthrough vs
tool-loop confirm behavior.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode gaps from review round 18 for PR #7079
Classifier (auto mode asks for these):
- os.open through a module alias (import os as o; o.open(...)); os/posix
aliases are tracked like the literal module name.
- less/more pagers, whose escapes (+cmd, !shell, -o/--log-file, LESSOPEN)
can run a command or write a file the command-name allowlist cannot
see, so they are no longer auto-approved.
- a read-named MCP tool carrying a mutating query
(query_database {"query": "DELETE FROM runs"}); DML/DDL statements are
matched as whole statements so a natural-language query that merely
contains "delete" stays safe.
- ML persistence helpers (save_pretrained / save_file / save_model /
save_weights / save_lora / save_checkpoint) that export weights to disk.
Route:
- Honor CLI-forced tools when deriving the confirm gate. When a process
policy (unsloth run --enable-tools) opens the local tool loop without a
request-level tool signal, a permission_mode ask/auto request now
derives confirm at the route (GGUF and safetensors paths) so the mode
still gates the call, and a non-streaming ask/auto request is rejected
rather than running unprompted. A plain client-tool passthrough (no
local loop) is unaffected.
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 19 for PR #7079
Approve for me now asks for these too:
- a terminal read whose path is built by indirect parameter expansion
(x=passwd; p=x; cat /etc/${!p})
- a bash /dev/tcp or /dev/udp redirection, which opens a network socket
(cat </dev/tcp/host/port)
- a python read via pathlib's receiver-plus-pattern glob
(Path('/etc').glob('passw?'))
- a python read whose sensitive root passes through a normalizer
(os.path.abspath('/etc'), Path('/etc').resolve())
- a pickle-backed loader that can execute code on load
(torch.load, joblib.load, pandas.read_pickle), tracked through module
import aliases
- compiled code wrapped into a callable (compile(...) + types.FunctionType)
Adds regression tests for each case and its safe counterpart.
* Honor unset permission_mode as ask across the local tool loop for PR #7079
Three gaps where an omitted permission_mode did not behave as the
documented default ("ask"):
- The frontend only sent permission_mode / confirm_tool_calls /
bypass_permissions when a tool pill was on. A process policy
(unsloth run --enable-tools) can open the tool loop with no pill, so
the backend never saw the selected gate. Send the three permission
fields at the top level of every local chat payload instead.
- The backend read payload.confirm_tool_calls directly at the
pre-switch guard and both late per-backend derivations, so an unset
mode fell through as no-gate even for an explicit ask/auto. Add
_permission_mode_confirm(payload): explicit confirm_tool_calls wins,
explicit ask/auto engage the gate, off/full never prompt, and an
unset mode defaults to ask only where realizable (streaming), keeping
the legacy no-gate run for non-streaming unset requests.
- A forced ask/auto tool loop (CLI --enable-tools) with no stream now
400s at the pre-switch guard before evicting the resident model,
matching the existing confirm-without-stream rejection.
Adds test_permission_mode_confirm_derivation covering the derivation
truth table.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Declare permission_mode and bypass_permissions on the local chat request type
The previous change moved permission_mode, confirm_tool_calls and
bypass_permissions to the top level of the local chat payload. They had
lived inside a conditional spread, which is not subject to excess
property checking, so the fields were never declared on
OpenAIChatCompletionsRequest. At the top level tsc flagged
permission_mode as unknown (TS2322), failing the frontend build and
every job whose Studio install builds the frontend.
Add permission_mode and bypass_permissions to the request interface
(confirm_tool_calls was already present).
* Close auto-mode classifier gaps from review round 21 for PR #7079
Auto mode ("Approve for me") now asks for these too:
- a pathlib read built from a concrete constructor (PosixPath, WindowsPath
and their Pure* forms), which the folder previously ignored so
PosixPath('/etc') / 'passwd' lost its /etc root and ran unprompted
- a terminal or python read of the ssh host keys under /etc/ssh, which
the sensitive-path regex only covered for passwd/shadow/sudoers
- a read whose path variable is reassigned: the whole-tree pre-scan kept
the last binding, so base = '/etc'; open(base + '/passwd'); base = 'data'
folded to data/passwd and ran even though execution reads /etc/passwd;
any multiply-bound name now folds to the escape sentinel and asks
Also stop the pre-switch guard from rejecting a plain client-tool
passthrough. permission_mode only implies the confirm gate for Studio's
own local tool loop (enable_tools / enabled_tools / mcp_enabled); a
non-streaming client-tool passthrough that carries permission_mode
ask/auto (confirm_tool_calls left unset by the validator) must forward to
the provider branch. Only an explicit confirm_tool_calls=True still forces
the local-confirm rejection there.
Adds regression tests for each case and its safe counterpart.
* Fix permission-pill compaction count and Full-access confirm sync for PR #7079
Two frontend consistency issues in the permission-level UI:
- The composer collapses tool pills to icons above four, but the count
left out the permission pill, which renders in every mode except off.
With one optional pill also shown the row reached five pills without
collapsing and could overflow. Count the pill when it is visible
(permission_mode != off).
- Entering Full access via setPermissionMode('full') or
setBypassPermissions(true) left confirmToolCalls at its previous value,
so a Full-access run (which sends confirm_tool_calls=false) could still
report confirmations as enabled in response metadata. Set
confirmToolCalls false at both entry points.
* Close auto-mode classifier gaps from review round 23 for PR #7079
Auto mode ("Approve for me") now asks for these too:
- a command using an abbreviated GNU long option that reaches a
write/exec action (sort --out= for --output, env --ch= for --chdir,
fd --base-dir= for --base-directory); a prefix of an unsafe long flag
now fails closed
- printf -v NAME, which assigns to a shell variable, so
printf -v PATH %s .; ls can rewrite PATH and run ./ls unprompted
- fd --base-directory / --search-path, which move the search root
outside the session workdir without any positional slash token
- an MCP tool whose compound read name carries a copy-style mutator
(read_and_copy_file, get_and_snapshot_volume): copy, duplicate,
import, export, download, backup, restore, snapshot, mirror
Also treat an omitted permission_mode as its documented default ("ask")
on the Anthropic Messages server-tool path. That branch has no
confirmation channel and already rejects explicit ask/auto, so an
omitted mode now falls into the same rejection instead of silently
running server tools unprompted, unless the caller opted out with
confirm_tool_calls=false (the legacy equivalent of "off"). off/full and
that opt-out still run; the two routing tests that relied on the old
implicit run now set permission_mode="off".
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Refine permission gating from review round 24 for PR #7079
Four fixes from the latest review:
- Anthropic Messages server tools: an omitted permission_mode no longer
rejects a request that only runs safe server tools (web_search), so
existing Anthropic callers keep working. It still rejects an omitted
mode when a local tool (terminal/python) is selected, and an explicit
ask/auto is still rejected outright. off/full and a
confirm_tool_calls=false opt-out always run.
- Pre-switch confirm-without-stream guard: use
_explicit_studio_tool_loop_requested (the same predicate the
passthrough router uses) instead of the policy-inclusive
_effective_enable_tools, so a process --enable-tools policy no longer
turns a client-tool passthrough into a local-loop rejection.
- Auto mode now asks for `uniq INPUT OUTPUT`: uniq writes its second
file positional, so a second positional (numeric flag values skipped)
is treated like `sort -o`. A lone `uniq file` or piped `... | uniq`
stays safe.
- MCP mutation check now strips SQL comments before matching, so
DELETE/**/FROM and UPDATE/**/users (comment-as-whitespace) no longer
slip past the DML/DDL denylist.
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode gaps from review round 25 for PR #7079
Auto mode ("Approve for me") now asks for these Python cases too:
- a bare archive constructor with a write mode (from zipfile import
ZipFile; ZipFile('out.zip', 'w')), tracked through import aliases like
the zipfile.ZipFile attribute call already was
- a dynamic lookup aliased through getattr (g = getattr;
rm = g(os, 'remove'); rm('file')), not just direct getattr(...) calls
- a callable that wraps open or a writer via functools.partial
(w = partial(open, mode='w'); w('out.txt')), which hides the write mode
Also:
- Always-safe tools (render_html) stream their early provisional canvas
card in auto mode again. The provisional-card guard mirrored the raw
confirm flag, which suppressed the early card under Approve-for-me; it
now reuses the auto-mode safety decision (is_always_safe_tool).
- The assistant-ui composer no longer counts the permission pill toward
its collapse threshold when the level is Off (the pill renders null
there), matching the other composer.
Adds regression tests for each case and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align permission-mode confirm guards with the router (review round 26)
Three pre-switch confirm-gate checks disagreed with how the tool
loop actually enters, so a valid request could 400 (or an invalid
one could evict the resident model) at the wrong point:
- The /chat/completions pre-switch guard only looked at explicit
request fields, so a process --enable-tools policy that forces the
loop on (request omits enable_tools, no client tools) slipped past
it and only 400ed after _maybe_auto_switch_model had swapped the
model. It now mirrors the router's own loop-entry gate
(_effective_enable_tools or mcp, tool_choice="none" disabling it
unless explicitly asked) while still deferring to client-tool
passthrough, so the policy-forced case is caught before the switch.
- The ChatCompletionRequest full/off fold treated enabled_tools by
itself as a local-loop request and set confirm_tool_calls=True.
The router never starts the loop on enabled_tools alone (it only
filters which tools run), so a non-streaming passthrough carrying
client tools plus enabled_tools 400ed instead of routing verbatim.
The fold now keys off the same enable_tools / mcp_enabled signals.
- The Anthropic /v1/messages unsupported-mode rejection (ask/auto,
or an omitted mode selecting terminal/python) ran inside the
post-switch server-tools block, so an invalid request evicted the
resident model before the 400. It now runs before the auto-switch,
determined from the requested server tools, like the neighboring
malformed- and mixed-tool guards.
Adds regressions for each: a policy-forced non-streaming ask/auto
guard rejection that never reaches the switch, an enabled_tools-only
passthrough that keeps confirm unset, and an Anthropic rejection that
precedes _maybe_auto_switch_model.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close auto-mode classifier gaps from review round 27 for PR #7079
Auto mode ("Approve for me") now asks for these host-mutating or
host-reading cases it previously ran unprompted (the sandbox does not
jail filesystem reads, and terminal commands can change host state):
- Destructured string literals fold into the scanned path now, so
base, leaf = ('/etc', 'passwd'); open(base + '/' + leaf).read()
resolves to /etc/passwd and asks, like the single-assignment form
already did. The tuple/list unpacking branch tracked only aliases to
open; it now also binds literal and folded-path elements.
- pathlib name rewrites fold to the rewritten path:
Path('/etc/x').with_name('passwd').read_text() (and with_stem /
with_suffix) spell no literal /etc/passwd but resolve to it, so they
are folded and caught. Benign in-sandbox rewrites stay safe.
- hostname NAME (or -F/--file, -b/--boot) sets the hostname, so a
positional or a set flag asks; bare hostname and the display flags
(-f/-i/-I/...) stay read-only.
- date -s/--set STRING and the bare MMDDhhmm... positional set the
system clock and now ask; the display forms stay read-only (+FORMAT,
-u/-R, and -d/-r/-f whose following value is skipped so date -d
tomorrow is not mistaken for a clock-setting positional).
Adds regression rows for each gap and its safe counterpart.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close more auto-mode classifier gaps from review round 28 for PR #7079
Auto mode ("Approve for me") now asks for these cases too:
- Mapping-style %-formatted paths. '/etc/%(f)s' % {'f': 'passwd'} folds
to /etc/passwd and asks; a dynamic value or a non-literal mapping
leaves the NUL marker so /etc/<dynamic> still fails closed. The path
folder previously handled only tuple/scalar % right-hand sides and
returned None for a dict, hiding the sensitive segment.
- A read-named MCP database tool carrying PostgreSQL COPY. COPY ... FROM
bulk-loads a table and COPY ... TO writes a server-side file, so both
are matched as mutating queries like DELETE/UPDATE already were. A
'copy' substring in a column name stays safe (word boundary).
- logging file handlers. logging.FileHandler('out.log', mode='w') (and
the default append mode, RotatingFileHandler/TimedRotatingFileHandler/
WatchedFileHandler, and the bare from-import form) create or truncate
a file like open(..., 'w'), so they are classified as writer calls.
StreamHandler / NullHandler and logging reads stay safe.
Adds regression rows for each gap and its safe counterpart.
* Fix writer aliases, GraphQL mutations, and auto server tools (review round 29)
- Auto-mode Python: an aliased writer or archive constructor is tracked
like the existing open alias, so from numpy import save; s = save;
s('out.npy', arr) (and z = ZipFile; z('a.zip', 'w'), incl. the
destructured forms) ask instead of running the write unprompted. A
benign builtin alias (x = len) stays safe.
- Auto-mode MCP: a read-named tool carrying a GraphQL mutation now asks.
query_graphql {"query": "mutation { deleteIssue(id: 1) }"} matches a
leading mutation keyword (GraphQL uses # comments, so it scans the raw
payload); GraphQL read queries stay safe.
- Anthropic /v1/messages: permission_mode "auto" no longer 400s a
safe-only server-tool selection. auto only needs a confirmation
channel for an unsafe call, so like the omitted default it runs for
web_search / RAG / render and rejects only when a gate-needing local
terminal/python tool is selected. ask still always rejects (it asks
per call, which this passthrough cannot honor). The rejection stays
ahead of the model auto-switch.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate asyncio spawn, net clients, default-captured open; allow safe-only auto (round 30)
Auto-mode Python now asks for more process/network/write vectors:
- asyncio process spawners (asyncio.create_subprocess_exec/shell and a
loop's subprocess_exec/shell) run an arbitrary program without the
terminal blocklist, so they gate like os.system/subprocess.
- stdlib network clients imaplib / poplib / nntplib / xmlrpc(.client) /
webbrowser open outbound connections the sandbox does not namespace
off, so their import asks like the other network modules.
- a callable captured as a function or lambda parameter default
(def f(o=open): o('out', 'w')) now binds that parameter into the same
alias set, so the later write through it is gated. A benign default
(o=len) stays safe.
Also, permission_mode "auto" no longer 400s a non-streaming local tool
request whose selection is always-safe-only (web_search / RAG / render).
auto only prompts for a classifier-flagged call, so a safe-only auto
request needs no stream, while ask, an explicit confirm_tool_calls=true,
MCP, and an unrestricted or unsafe selection still require it. Applied
via a shared _confirm_gate_needs_stream helper at the pre-switch, GGUF,
and safetensors confirm-stream guards; the loop's per-call confirm flag
is unchanged.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Catch brace-glob paths and attribute writer aliases; unfold auto (round 31)
- Terminal auto mode now runs the glob-sensitive scan over every
expansion candidate, so a brace-expanded glob (cat /e{t,}c/pass?d,
which bash expands to /etc/pass?d and then globs to /etc/passwd) asks.
Brace expansion alone spells no literal /etc/passwd and the glob only
resolves once the brace group is expanded, so scanning both together
is required. A benign brace + glob stays safe.
- Python auto mode now tracks a mutating attribute captured as a plain
name: s = np.save; s('out.npy', arr) binds a writer alias, a captured
.open bound method (p = Path('out').open; p('w')) fails closed on any
call since its mode position varies, and z = zipfile.ZipFile is gated
like the bare import. A benign attribute alias (x = np.mean) stays safe.
- permission_mode "auto" is no longer folded to confirm_tool_calls=true
on the request model. Folding it defeated the safe-only-selection
exception in _confirm_gate_needs_stream (an explicit confirm forces
stream=true), so a non-streaming safe-only auto request was rejected.
Leaving it unset lets the route apply the exception; the mode still
drives the loop's per-call gate. "ask" still folds (it gates every
call).
Adds regression rows/cases for each.
* Harden SQL/GraphQL/writer classification and passthrough guards (round 32)
MCP argument mutation detection (read-named query tools):
- CREATE DDL now matches modifiers and the broader object set, so
CREATE OR REPLACE VIEW, CREATE UNIQUE INDEX, CREATE TEMP TABLE,
CREATE MATERIALIZED VIEW and CREATE FUNCTION ask.
- Stored-procedure invocation (CALL proc(...), EXEC/EXECUTE) and VACUUM
ask; a natural-language "call me back" stays safe via the trailing
"(" / ";" / end lookahead.
- GraphQL # comments are stripped before the mutation match, so
mutation # note\n { deleteIssue(id: 1) } no longer hides the mutation.
Python auto-mode classification:
- numpy.memmap / open_memmap and pandas ExcelWriter / HDFStore create or
truncate a file on construction, so they gate like open(..., "w").
- asyncio networking (asyncio.open_connection, loop.create_connection /
create_server and unix variants) opens outbound connections/listeners
the sandbox does not isolate, so it gates like socket.connect.
Terminal auto-mode: file -C / --compile writes a compiled magic database.
Routing:
- A JSON-schema response_format is guided-decoding passthrough, not a
local tool loop, so a --enable-tools policy no longer 400s a
non-streaming ask/auto structured-output request at the confirm guard.
- An explicit confirm_tool_calls=False opts out of the Anthropic Messages
server-tool gate entirely (it wins over the mode, mirroring
_permission_mode_confirm and the GGUF path), so it runs even under ask.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Track path-ctor aliases, exempt empty selection and safe safetensors card (round 33)
- Python auto mode now propagates path constructor / join aliases, so
assigning Path or os.path.join to another local name is still folded:
P = Path; (P('/etc') / 'passwd').read_text() and j = os.path.join;
open(j('/etc', 'passwd')) ask, while a benign /tmp alias stays safe.
- _confirm_gate_needs_stream now distinguishes an omitted enabled_tools
(None, all tools) from an explicit empty list ([], no tools). An empty
selection runs no built-in tool and cannot prompt, so a non-streaming
auto request with enable_tools=true, enabled_tools=[] is no longer
400ed under a --enable-tools policy.
- The safetensors provisional render_html card now uses permission_mode:
render_html is always safe and never prompts, so its early canvas card
streams under auto (which ships confirm_tool_calls=true) instead of
being suppressed, matching the GGUF path's is_always_safe_tool exemption.
Adds regression rows/cases for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Extend auto-mode classifier: SQLite mutations, more net/xattr/compressed writers
Additional fail-closed gaps found by a fresh adversarial pass, each with a
reproduction and a benign control:
- MCP read-named tools now ask on SQLite-flavored writes the base DML/DDL regex
missed: ATTACH / DETACH DATABASE, a write-form PRAGMA (PRAGMA journal_mode=WAL
/ user_version=42 / foreign_keys(0), while the read-form PRAGMA journal_mode
stays safe), and load_extension() which loads and runs an arbitrary shared
library.
- Python auto mode now gates the remaining asyncio network entry points
(start_server, open_unix_connection, loop.create_datagram_endpoint,
sock_connect), os.setxattr / os.removexattr metadata writes, the gzip / bz2 /
lzma single-stream writers (GzipFile / BZ2File / LZMAFile, mode-gated like
ZipFile so a read stays safe), pandas to_xml, and the websockets client.
Benign controls (SELECT 1, read-form PRAGMA, asyncio.sleep, gzip read, numpy
read, natural-language "attach"/"analyze") stay safe. Regression rows added to
test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close follow-up auto-mode gaps: SQLite/GraphQL variants, more writers and net
A fresh adversarial pass on the previous round found consistent extensions of
the same fail-closed rules, each reproduced with a benign control:
- MCP read-named tools: DROP / ALTER now cover the same broad object set as
CREATE (DROP FUNCTION, ALTER INDEX, DROP MATERIALIZED VIEW); ATTACH is caught
without the optional DATABASE keyword via its quoted-path form; a
schema-qualified write PRAGMA (PRAGMA main.user_version=1) is matched; and a
GraphQL mutation carrying directives (mutation M @audit { ... }) is treated as
a mutation.
- Python auto mode: os.startfile (Windows program launch), asyncio
start_unix_server, and the socketserver framework now ask; a gzip/bz2/lzma
open imported under an alias (from gzip import open as gopen) is gated like
builtin open; and a dynamic path prefix that can form a sensitive absolute
root (open(chr(47) + "etc/passwd"), open(os.sep + "etc/passwd")) is treated as
sensitive, while a dynamic prefix with a benign suffix stays safe.
Benign controls (read-form PRAGMA, natural-language "attach ... as", "drop the
idea", SELECT dropped_at, query @cached, gzip read alias, dynamic prefix +
data/file suffix) stay safe. Regression rows added to test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate GNU time -o, basicConfig/methodcaller/fileinput, and more SQL mutations
Another adversarial pass surfaced further consistent fail-closed gaps, each
reproduced with a benign control:
- Terminal: GNU time -o/--output/-a/--append truncate or append to a file with
timing output; time is a wrapper, so the flag is checked before the wrapped
command like env -C.
- Python auto mode: logging.basicConfig(filename=...) opens a log file for
write; operator.methodcaller("write_text"/...) hides a writer method behind a
string and is now treated as dynamic dispatch (like getattr/partial);
fileinput.input(..., inplace=True) rewrites a file in place (the default read
form stays safe).
- MCP read-named tools: UPDATE now matches quoted, bracketed, and
schema-qualified targets (UPDATE "users" / public.users / ONLY public.users /
[users] / `users` SET); SELECT ... INTO OUTFILE/DUMPFILE writes a server file;
and state-changing SQL functions inside a SELECT (pg_terminate_backend,
setval, pg_write_file, lo_export, ...) ask.
Benign controls (time ls / time -p, basicConfig(level=), methodcaller("upper"),
fileinput read, NL "update ... set", setval_col column, PL/pgSQL SELECT INTO
var) stay safe. Regression rows added to test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten auto-mode classifier comments
Collapse the multi-line rationale blocks in the permission classifier to one or
two lines each without dropping the exploit each branch closes. Comments and
whitespace only (no code change); the classifier tests are unchanged and pass.
* Retry transient SSE stalls in the tool-calling smoke probes
The tool-calling job flaked with a bare "TimeoutError: timed out": the
server-side python/bash probes stream over post_sse(), which (unlike
post()) had no transport-level retry, so a single stalled stream on a
shared CI runner hard-failed the whole step even though function calling
had already passed.
post_sse() now mirrors post(): a transport-level stall (stream open or a
mid-stream read timing out) is retried once with a fresh request capped
at 300s, while HTTP status errors still surface immediately. The
Linux _run_tool_probe caps each attempt at 360s and treats a stall that
outlives the retry as a failed attempt (rotate to the next seed) instead
of raising, and the web_search probe uses the same 360s cap. A genuine
server wedge still fails (the retry also times out), so real regressions
are not masked. Applied to the Linux, macOS, and Windows inference-smoke
workflows, which share the probe.
* Close five more auto-mode classifier gaps from review
Each reproduces with a benign control:
- Path constructor aliased through an attribute (P = pathlib.Path) now folds
like the bare-name alias, so (P('/etc') / 'passwd').read_text() asks while a
/tmp alias stays safe.
- Callable defaults that are not plain names now bind the parameter: an
attribute writer (def f(s=np.save)), an archive constructor, a captured .open,
and partial(open, mode='w') fold like the equivalent assignment; a benign
default (np.mean) does not.
- A dynamic piece inside a sensitive name (open('/et' + chr(99) + '/passwd'),
which folds to '/et\x00/passwd') now asks: the literals around each dynamic
segment are matched against a credential target with the segment as any run of
non-separator chars, so an all-dynamic ('1 + 1') or segment-spanning
(a + '/' + b) path stays safe.
- MCP read-named tools now ask on REFRESH MATERIALIZED VIEW and REINDEX; a
'refresh' column or natural-language 'refresh' stays safe.
- A writer/open alias handed to a higher-order invoker (map(open, names, modes),
starmap(np.save, ...)) is gated even without a direct call site; a benign
map(len, ...) is unaffected.
Regression rows added to test_permission_mode.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Default tool pills off on model load so tool execution is opt-in
resolveToolsEnabledOnLoad turned the web-search and code pills on for
any tool-capable model when the user had expressed no preference. Default
them off instead, so tool execution is enabled only when the person
clicks the pill to turn it on; a saved preference (on or off) is still
honoured, so a user who already enabled tools keeps them on.
* Gate mark/subscribe MCP verbs and qualified higher-order writer invokers
- A read-prefixed MCP tool name carrying mark / subscribe / unsubscribe
(get_and_mark_read, get_and_subscribe) now asks; a 'mark' substring inside
one token (list_bookmarks) stays safe.
- The higher-order writer check now also fires for a qualified invoker
(itertools.starmap(open, ...), functools.reduce(open, ...)), matching the
bare-name map/filter form; the writer-check on the first arg keeps a benign
itertools.starmap(len, ...) or itertools.chain(...) safe.
Regression rows added to test_permission_mode.py.
* Close more auto-mode gaps and align the ask confirm fold across paths
Each classifier change reproduces with a benign control:
- MCP read-named tools now ask on reply / notify verbs (get_and_reply_email,
list_and_notify_users), on catalog writes COMMENT ON / SECURITY LABEL / LOCK
TABLE and CREATE|DROP|ALTER POLICY, and on state-changing PostgreSQL functions
inside a read-shaped SELECT (nextval, set_config, pg_notify, the advisory-lock
family). A 'comment' column, a 'locks' table, and a 'nextval' column prefix
stay safe; the natural-language NOTIFY/SET ROLE statement forms are left out
because SET/NOTIFY overlap ordinary prose.
- Python auto mode now gates loader.exec_module (runs a module's code), archive
extractall (zip-slip file writes), the ensurepip / venv modules (install pip /
build an environment), and pydoc.writedoc. The Hugging Face login token
(~/.cache/huggingface/token and stored_tokens) is now a sensitive path, while
the rest of that cache (model data) stays readable.
- ChatCompletionRequest no longer overwrites an explicit confirm_tool_calls=false
when permission_mode='ask': the fold only self-enables the gate when the flag
is unset, so an explicit opt-out wins on the chat path exactly as it already
does via _permission_mode_confirm and the Anthropic pre-switch guard.
Regression rows added to test_permission_mode.py.
* Gate sort -T, xxd outfile positional, and the legacy HF token path
- sort -T / --temporary-directory writes spill files to a caller-chosen dir,
so it joins -o / --output in sort's unsafe-flag set.
- xxd [infile [outfile]] writes its second positional, like uniq; xxd now uses
the same second-positional-write handling (xxd in.bin out.hex asks, xxd
in.bin and xxd -c 16 in.bin stay read-only).
- The sensitive-path regex now also covers the legacy ~/.huggingface/token
location (optional leading dot), not just ~/.cache/huggingface/token; an
unrelated dir like myhuggingface/token stays safe.
Regression rows added to test_permission_mode.py.
* Catch multi-char SQL mutation targets, globbed credential names, digit outfiles
Three fail-open gaps in the auto-mode classifier, each with a benign control:
- SQL: the trailing word boundary on the MCP mutation regex meant a bare \w
stopped at the first character, so TRUNCATE users, GRANT SELECT ON t, and
REVOKE ALL ON t (multi-character names) slipped through while single-letter
targets matched. Match the whole identifier instead, and accept an explicit
AS alias on UPDATE (UPDATE users AS u SET). The implicit-alias form is left
out because it is indistinguishable from the prose "update <noun> <noun> set".
A truncate_log column and a grants table stay safe.
- A glob that resolves to a credential basename anywhere (cat ~/.huggingface/tok?n
-> token, cat proj/.netr? -> .netrc, cat repo/.aws/cred*) now asks; the fixed
target list only covered a handful of home paths. notes/dra?t.txt and
token_counts.tx? stay safe.
- uniq / xxd counted file positionals but skipped every numeric token to ignore
a flag value, so a file literally named with digits (uniq 123 out) hid the
output positional. Track each command's value-taking flags and consume only
the value, so uniq -f 2 in stays safe while uniq 123 out asks.
Regression rows added to test_permission_mode.py.
* Isolate the permission-mode loop tests from process-global state
The loop-driving tests (auto/off/full/bypass) drove run_safetensors_tool_loop
against a process-global approval registry (state.tool_approvals._pending)
keyed by a single shared session id, and read os.environ. Other backend test
modules mutate both, some at import time, so in the full-suite ordering a stale
pending approval or a leaked env var could make the loop deny or skip a call
these tests expect to run. It passed when the file ran alone but failed only in
the complete tests/ run on CI.
Add an autouse fixture that snapshots and restores os.environ and the approval
registry around each test, and give every _drive call a unique session id so a
leaked approval can never collide. Attach a compact event-stream dump to the
loop assertions so any residual full-suite-only failure reports what the loop
actually did instead of a bare diff.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden auto-mode classifier for recursive listers, sort file lists, aliased invokers, single-member extract
Close four fail-open gaps in is_potentially_unsafe_tool_call:
- terminal: tree/du (always recursive) and ls -R rooted at an absolute or
tilde path now ask, matching the existing grep/rg/find recursive-read gate;
relative walks stay safe.
- terminal: sort --files0-from=F reads the file list named in F, so it can
read arbitrary host files indirectly; added to sort's unsafe flags.
- python: track aliases of the higher-order invokers (m = map;
from itertools import starmap as sm) so an aliased invoker handed open/a
writer is still gated; a benign callable (map(len, ...)) stays safe.
- python: single-member archive extract (ZipFile/TarFile.extract) writes to
disk like extractall and is vulnerable to a crafted member path, so gate it.
Also update the stale _FakeExecuteTool in test_permission_mode.py to accept
the thread_id keyword that run_safetensors_tool_loop now forwards to
execute_tool after the main merge, which had broken the five tool-loop tests.
Adds regression rows covering each gap plus benign controls.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: normalize unknown permission_mode to 'ask' instead of a 422
The request models validated permission_mode with Literal[ask, auto, off,
full], so an unrecognized value from a newer UI/client was rejected with a 422
before the tool loops could apply their unknown -> ask fallback
(safetensors_agentic.py:464, llama_cpp.py:9001). That made the intended
forward-compat degradation unreachable at the API boundary for both Chat
Completions and the analogous Anthropic field.
Accept a plain string on both ChatCompletionRequest and AnthropicMessagesRequest
and normalize in a before-validator: None stays unset, the four known modes pass
through, and any other value degrades to the safest gate ('ask'), matching the
loops. Adds a regression test covering unknown/None/known across both models.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: close five more auto-mode classifier gaps
- terminal: xargs is no longer a safe wrapper. It appends arguments read from
stdin that the scan never sees, so `echo -o out /etc/passwd | xargs sort`
forwards to `sort -o out /etc/passwd` (a write + sensitive read) while only
the allow-listed literals are visible. Any xargs command now asks.
- terminal: ionice -p/-P/-u change the I/O priority of an already running
process / group / user instead of forwarding to a wrapped read-only command,
so `ionice -c 3 -p <pid>` now asks. ionice -c 3 <cmd> stays safe.
- MCP: gate ALTER SYSTEM, which persists PostgreSQL server configuration and was
not one of the DDL objects the mutation detector matched.
- MCP: a credential noun in a read-named tool (read_secret, list_tokens,
get_credentials, fetch_api_key) is a sensitive disclosure, so it asks even
without a mutating verb or a path/SQL argument. Scoped *_key nouns keep a
primary_key / keyboard lookup safe.
- render_html: no longer unconditionally safe. A static canvas still auto-runs,
but one whose HTML/JS reaches the network (fetch/WebSocket/remote script) asks,
since it can egress under the canvas CSP when artifact network access is on.
Its early provisional card is suppressed under the auto confirm gate, and the
confirm-without-stream guard now requires a stream when render_html is
selectable.
Adds regression rows and benign controls for each, and updates the render_html
provisional-card and confirm-gate tests to the new behavior.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: extend auto-mode gates for indirect file lists, dynamic lookups, HTML network loads, and Anthropic render_html
Follow-ups on the previous classifier round:
- terminal: wc/du/find --files0-from (and find's -files0-from primary) read a
NUL-separated list of input paths from a file, the same indirect mechanism as
sort --files0-from, so a crafted list reads arbitrary host files past the
literal path/root checks. Gate them like sort.
- python: a namespace lookup through a dict-style call (f =
__builtins__.__dict__.get('open'), globals().get('open'), vars(x).get(...))
can return open/eval/a mutator, so poison the bound name like getattr/subscript
lookups already are. An ordinary dict .get or os.environ.get stays safe.
- render_html: broaden the network detector so a canvas that loads a resource
via CSS url()/@import, srcset, or a root-relative (/path) or protocol-relative
(//host) src/href is treated as networked, not just fetch/WebSocket/remote
script. Relative ./x and url(#id)/data: refs stay static/safe.
- Anthropic /v1/messages: drop render_html from the unprompted-safe server-tool
set. Since it can prompt (networked canvas) and this channel invokes the loop
without confirm, selecting it under ask/auto/omitted now rejects like
terminal/python; off/full (or an explicit confirm opt-out) run it.
Adds regression rows and benign controls for each, plus an Anthropic route test.
* Studio: close six more auto-mode classifier gaps
- terminal: a glob that expands to a project .env (cat .e?v) now asks; .env
joins the sensitive glob-basename set, matching the literal-path gate.
- python: an open bound onto an attribute (box.f = open; box.f('out','w'))
is tracked by attribute name, and open invoked via .__call__
(open.__call__('out','w'), unwrapped to the underlying callable) is gated,
so neither slips past the name-based open-alias checks. Benign attribute
callables and .__call__ on non-writers stay safe.
- python: a namespace lookup via .get/.pop/.setdefault already covered the
builtins case; unchanged here.
- MCP: a mutating HTTP verb in a method/verb argument (get_url
{"method": "DELETE"|"POST"|"PUT"|"PATCH"}) now asks, so a generic HTTP
tool cannot mutate an external service unprompted; GET/HEAD stay safe.
- MCP: a credential/secret environment-variable value (get_env
{"name": "OPENAI_API_KEY"}) is treated as a sensitive read via the same
credential-noun match used for tool names; PATH/HOME stay safe.
- render_html: self-navigation sinks (location.assign/replace, window.open,
assigning a URL to (window.)location(.href)) join the network detector, so a
canvas that navigates itself to an external URL asks; location.reload() /
history.back() stay static.
Adds regression rows and benign controls for each.
* Studio: gate obfuscated canvas egress, sensitive-dir iteration, and MCP metadata-host reads
- render_html: strip block comments before the network scan so fetch/*x*/(...)
cannot hide egress, and match bracket-access forms (window['fetch'](...),
self['open'](...)). Line // comments are left alone so the // in an https URL
is not eaten. A comment-only canvas stays static.
- python: enumerating a directory outside the sandbox (Path('/etc').iterdir(),
os.scandir('/etc'), os.listdir('/home'), os.walk('/')) reads host filenames
the direct /etc/passwd checks would prompt for, so gate it when the target dir
folds to an absolute/tilde/sensitive path; a relative dir stays safe and an
unresolved dynamic dir is left to other checks.
- MCP: a read-named HTTP tool pointed at a cloud-metadata / link-local host
(fetch_url {"url": "http://169.254.169.254/..."}, metadata.google.internal)
reads instance credentials, so classify those URL arguments as sensitive,
mirroring the sandbox SSRF blocklist; ordinary and localhost URLs stay safe.
Adds regression rows and benign controls for each.
* Studio: gate meta-refresh navigation, pandas HTML/markdown exporters, absolute glob roots, and checksum verify mode
* Studio: gate starred open writes, builtins.__import__, computed render_html sinks, and procfs fd reads in auto mode
* Studio: gate remote worker canvases, huggingface_hub downloads, and write callables passed to user helpers in auto mode
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Unsloth <michaelhan@Michaels-MacBook-Pro.local>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
* Studio: offer the latest transformers release for brand-new architectures
When a model's config.json model_type is absent from every installed
transformers overlay (base 4.57.x and the .venv_t5_530/550/510 sidecars),
Studio now checks, unauthenticated and cached, whether the newest
transformers ships it:
- utils/transformers_latest.py fetches the latest release version from
https://pypi.org/pypi/transformers/json and the CONFIG_MAPPING_NAMES
sources for that tag and for main from raw.githubusercontent.com
(never api.github.com), parsing them with the same AST extractor the
static router uses (no code execution, no trust_remote_code). Results
are cached in memory and in a JSON snapshot under studio_root()/cache
with a one day ttl; fetches are bounded to 5s with one retry and a
failure backoff, and offline mode or the new kill switch
UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS=1 short-circuits to None.
- POST /api/inference/validate gains requires_transformers_upgrade plus
a transformers_upgrade payload (model_type, pypi_version,
supported_in_pypi, supported_in_main) so the frontend can raise the
install consent dialog before /load, mirroring the existing
remote-code consent flow. The check fires only when the model_type is
unknown to all installed overlays and the hardcoded tier tables.
- POST /api/inference/install-latest-transformers provisions a new
persistent .venv_t5_latest sidecar after user consent, pinned to the
exact PyPI version (re-verified server-side) with the same
--target/--no-deps recipe as the fixed sidecars. A JSON pin marker
inside the dir records the installed package set, so restarts
revalidate it and routing resolves the new highest-ranked tier
automatically. A dependency preflight (compat_plan) compares the
release's requires_dist against the running env: unsatisfied
tokenizers/safetensors floors are shadow-installed as exact pins into
the sidecar, anything else unsatisfied blocks the install with a
clear message.
Routing for every already-supported model_type is unchanged: the
hardcoded lists and the 530/550/510 static resolver run first, the new
tier only participates once its venv exists, and the probe order gains
the latest sidecar only when provisioned. Verified against live PyPI
and GitHub (transformers 5.13.0: 674 model_types, 26 absent from all
installed overlays, e.g. cosmos3_omni; 4 dev-only on main) and with a
real sidecar install plus restart persistence. 64 new tests; the
existing 200-test transformers_version suite passes unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest-transformers check: fetch outside the lock, serialize installs
Release the module lock during the network refresh so a slow fetch cannot
stall other threads in the ASGI pool; concurrent callers during a fetch get
None (the graceful fallthrough) via an in-flight flag instead of stacking
fetches. Serialize install_latest_transformers with an in-progress flag so
concurrent consents cannot race the sidecar delete and recreate; the loser
gets a structured already-in-progress refusal.
* Latest-transformers check: LoRA bases, pin-gated mapping, live reverify
Run the upgrade check over the [adapter, base] target set so a LoRA whose
base model is a brand-new architecture surfaces the prompt (the worker
activates transformers for the base, not the adapter).
Gate the latest overlay's mapping lookup on a valid pin marker, matching
activation and the probe order, so a partial or manual .venv_t5_latest dir
cannot be routed to and then refused at activation.
Re-verify the requested version against a live PyPI snapshot at install
time, falling back to the cached one on fetch failure, so a release
published inside the cache TTL is not silently missed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest-transformers check: nested config types and latest-tier vision probe
Collect every model_type in the config (top level plus each nested
sub-config) and signal on the first one missing from all installed
overlays, so a supported wrapper carrying a brand-new backbone still
surfaces the upgrade prompt; wrappers instantiate sub-configs through
CONFIG_MAPPING and would fail on the nested type.
Route the vision capability subprocess through the pinned latest sidecar
when the model resolves to the latest tier, so latest-only VLMs are not
misclassified as text-only; every other tier keeps the 5.5 sidecar used
today.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest tier: nested routing, vision probe after raw miss, safe upgrades
Route by every model_type in the config: a nested sub-config type can raise
the tier (wrappers instantiate sub-configs through CONFIG_MAPPING), so a
supported wrapper with a latest-only backbone routes to latest once
installed instead of staying on default. An unknown nested type never
vetoes; the primary type keeps its previous semantics. The collector is
shared with the upgrade checker.
Vision detection: when the raw heuristics say False for a model that routes
to the latest tier, run the AutoConfig subprocess under the pinned latest
sidecar instead of trusting heuristics built from older transformers.
Provisioning: stage-and-swap. Build the new sidecar in .venv_t5_latest.staging
and swap it in only when the install and pin marker are complete, so a failed
upgrade never destroys a previously working sidecar; restore the old dir if
the final swap fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Latest-transformers checker, vision subprocess, and cache fixes
Require the latest release to support every missing model_type (the
primary included) before prompting; a nested-only match cannot make the
model loadable, so no install is offered for it.
The vision-check subprocess now unions the active sidecar's own
registry mappings into the inlined parent-process detection sets, so
architectures only the sidecar knows classify correctly.
A successful sidecar install clears the tier probe cache, the latest
tier's model_type mapping, and the vision-detection cache so the new
venv takes effect without a restart. Tests for all three.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Aggregate upgrade support flags and keep install off /v1
The upgrade signal now reports supported_in_pypi only when the latest
release covers every missing model_type; a mix with a main-only nested
type surfaces as dev-only so no PyPI install is offered that would
still fail at load. The consented install endpoint moves to
studio_router so it is not reachable through the OpenAI-compatible /v1
mount. Tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor the latest-transformers kill switch in routing
With UNSLOTH_STUDIO_NO_LATEST_TRANSFORMERS set after the sidecar was
provisioned, the latest tier still joined mapping and probe routing
because only the pin was checked. Both admission points now also check
the kill switch, so operators can roll back a problematic sidecar
without deleting files.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Repair the latest sidecar through stage-and-swap
The lazy repair path installed into the live .venv_t5_latest, which
_ensure_venv_dir wipes first, so a failed repair deleted the pinned
sidecar and its marker. Both the consented install and the repair now
share one stage-and-swap helper: the incomplete-but-pinned dir survives
any failure and a later attempt can still repair it.
* Tighten comments
* Remove the staging dir when a latest-sidecar install fails
A pip failure inside _ensure_venv_dir returns False without raising, so
the except cleanup never ran and the partial .venv_t5_latest.staging
leaked until a later attempt. Also note on the validate response fields
that frontend consumption ships in the follow-up PR.
* Add the transformers-upgrade consent dialog to the frontend
When /validate reports requires_transformers_upgrade, every explicit load
path (chat runtime and the compare composer) now pauses on a consent
dialog modeled on the remote-code one: it names the model_type and the
latest PyPI transformers version, and on Accept calls
/api/inference/install-latest-transformers itself, shows an installing
state, and resumes the original load automatically on success. Errors
surface in the dialog with a retry; Cancel aborts the load like the
trust dialog's deny path. Architectures shipped only on transformers
main get a dev-only notice with no install button. Background auto-load
skips upgrade-requiring candidates instead of prompting, mirroring the
trust_remote_code rule. The dialog mounts once in the root layout and
runs before the security dialogs, since no load can proceed without the
runtime.
* Route a non-installable new architecture to the custom-code consent as a last resort
When the upgrade dialog has no installable PyPI release (the architecture
is only on transformers main, which Studio never installs), the dialog now
says so explicitly, and when the model also declares custom (auto_map)
code it offers Continue with custom code: resolving the paused load into
the existing trust_remote_code consent gate instead of hard-aborting.
Models with no custom code keep the Cancel-only notice. The backend
returns no upgrade signal at all for architectures unknown to both PyPI
and main, so those still route straight to the unchanged security gate.
* Force a 16-bit load for models on the latest-transformers sidecar
Live validation with Zyphra/ZAYA1-8B (model_type zaya, shipped by
transformers 5.13.1 but unknown to every installed tier) surfaced a
generation crash when the consented sidecar load kept the default bnb
4-bit quantization: transformers' grouped-MoE kernels feed the packed
uint8 expert weights straight into torch._grouped_mm, and generation
dies (plain 16-bit works). New latest_tier_active_for() mirrors the
sidecar activation's tier resolution and never raises; the inference
worker flips load_in_4bit off when it reports true, and the load route
applies the same flip so the pre-load VRAM guard and the worker command
agree. Fixed tiers are untouched. With the guard, ZAYA1-8B loads and
generates correctly in Studio chat.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Offer the custom-code fallback when a latest-sidecar install fails
* Fail remote mapping fetches wholesale and mirror the 16-bit flip in validate
A transient fetch or parse failure of one auto-mapping file no longer caches
a partial latest-release map for the TTL (a real 404 on pre-5.10 tags is
still tolerated), and validate_model now applies the same latest-sidecar
16-bit sizing flip as /load before the training guard so the two agree.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the latest-transformers changes
* Resolve remote LoRA bases, fold nested tiers, and guard the sidecar swap
latest_tier_active_for now resolves a remote adapter's base model the same
way worker pre-activation does (and returns early without a sidecar pin), a
hardcoded fast-path tier is raised when a nested sub-config's model_type
needs a higher sidecar, and the install route refuses to swap .venv_t5_latest
while training runs on it and unloads a latest-tier chat model first.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gate the sidecar install on worker liveness and size installable upgrades 16-bit
The install route now refuses while any training or export runs (tier
re-resolution without the load token is unreliable for gated repos), holds
the inference lifecycle gate across the unload and the swap so no load can
interleave, and passes the model name to unload_model. validate_model runs
the upgrade check before the training guard and sizes an installable
upgrade as 16-bit, matching what /load and the worker will force after the
consented install.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close the sidecar install races and honor the kill switch over cached mappings
Training starts and mutating export routes now refuse while a transformers
install is in progress (shared is_install_in_progress flag), the chat unload
and idle export-worker teardown moved into a before_swap hook that runs only
once the staged install succeeded, and _config_model_types checks the kill
switch before returning a cached latest mapping.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Reserve the sidecar swap before the gate wait and abort it on failed teardown
The install-in-progress flag moved into a shared sidecar swap reservation in
transformers_version, taken by the install route before awaiting the
inference lifecycle gate (so training and export starts see it for the whole
window) and by the lazy .venv_t5_latest repair path. The before_swap hook
now raises when the chat unload or export teardown reports failure, leaving
the previous sidecar untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Back the sidecar swap reservation with a cross-process lock file
The lazy repair runs inside worker subprocesses, where a module-level flag
is invisible to the parent's route checks. The reservation now also creates
a lock file next to .venv_t5_latest (O_EXCL, owner-only removal, stale after
two hours for crashed owners), so is_install_in_progress sees a repair from
any Studio process.
* Hand the swap reservation to the installer thread and harden pre-swap teardown
A cancelled install request no longer releases the reservation while the
installer thread is still staging (the thread owns and releases it, shielded
from cancellation). The route refuses while another inference request is
generating, export teardown runs before the chat unload and is judged by
worker liveness rather than the cleanup return value, and a live inference
worker with no active model (failed load residue) is shut down before the
swap.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the lifecycle gate with the installer and recheck the swap at spawn time
The gate moved into the shielded install task so a cancelled POST cannot
release the guard /load honors while the installer still runs, cached latest
probe results are ignored while the kill switch is set, and the training and
export subprocess spawns recheck the sidecar swap reservation right before
spawning (the route-level guards are one-shot and validation can outlast an
install's start).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Close the spawn-registration windows against the sidecar install
Training marks the spawn in progress before its reservation recheck and
is_training_active honors the flag, so the install route sees a start that
has passed proc.start() but not yet recorded _proc. Export load-checkpoint
rechecks the reservation after setting _export_active and before tearing
down the old worker, so losing the race keeps the loaded checkpoint instead
of surfacing a 500.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Refine the install-window interleavings around worker teardown
The inference busy count is rechecked under the lifecycle gate (streams
start by taking that gate, so nothing slips past a held gate), the training
handshake moved ahead of the VRAM-freeing before_spawn hook so a lost race
leaves chat/export intact, the export spawn-time check is op-aware (inside
an active op the install is the side that aborts), and the Xet-stall respawn
waits out a transient reservation instead of stranding the run.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Track the install's server-side unload and guard export ops against the swap
The upgrade dialog store records when its install actually ran (the server
unloads the active chat model before swapping), and the load flow then marks
the previous model as unloaded so a later cancelled gate still triggers
rollback; the custom-code fallback leaves the flag unset. _run_export gained
the same reservation handshake as load_checkpoint so an install cannot block
behind an hours-long export op instead of returning 409.
* Tighten comments in the install-guard and upgrade-consent changes
* Surface install-race refusals cleanly and roll back after a failed swap unload
/load refuses while the sidecar swap is reserved so a load cannot succeed
and immediately be unloaded by the pre-swap teardown, worker starts that
lose the install race raise a typed SidecarSwapInProgress mapped to 409
instead of a 500, the install response reports model_unloaded even on a
structured failure so the client can restore its state, and the compare
flow tracks the server-side unload like the primary load path and clears a
stale checkpoint on abort.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Type the export install races, scope the lock release, and keep the unload signal
Export load-checkpoint and export ops raise SidecarSwapInProgress (mapped to
409 in every export route) instead of a 400-shaped failure, the export spawn
check distinguishes repair reservations (always refused) from install ones
(op-aware), the swap lock release only unlinks a lock this process wrote so
a stale-superseded owner cannot drop the new owner's live lock, and the
frontend unload signal survives a superseding consent via read-and-clear
consumption instead of a reset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Finalize a stalled run when the respawn loses the install race and latch the unload signal
The Xet-stall respawn timeout now finalizes the run as a failure instead of
raising into the pump's broad finalization catch (which stranded it in a
training state with no worker), and a successful install retry ORs the
model_unloaded signal with the latched value so a failed-after-unload first
attempt still triggers rollback.
* Recheck the swap under the load gate and latch the unload before resolver checks
/load rechecks the sidecar reservation after acquiring the lifecycle gate
(an install can reserve while the load queues on it), and the dialog store
latches model_unloaded as soon as the install response arrives, before any
resolver-identity guard, so a superseded consent's unload still reaches
whichever load consumes the signal next.
* Report cleared-state unload failures, guard queued installs, and fold name tiers
A failed chat unload that still cleared the orchestrator's model state now
reports model_unloaded so the client rolls back, the installer aborts with
a 409 when a model load completed while it waited on the lifecycle gate,
and the fixed-tier name fast path consults the config mapping when a latest
sidecar is pinned so an accepted upgrade routes to the sidecar it installed
(no I/O added to the unpinned path).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Report cleared-state unload failures and harden the spawn handshake flag
The failed-unload branch in before_swap now detects that the orchestrator
cleared its model state and reports model_unloaded before aborting (the
earlier commit claimed this fix but a scripting error dropped the edit),
the installer's queued-load check compares a load generation counter so a
same-model reload is caught, and both training spawn sites wrap everything
after the handshake in a guard that resets _spawn_in_progress on any
exception so a failed start cannot wedge is_training_active.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Bump the load generation when the load is published, not at load start
A start-time bump is already visible when the installer snapshots mid-load,
so a same-model reload completing after the snapshot looked unchanged and
could be unloaded by the swap. The counter now increments alongside the
active_model_name publish.
* Self-heal a broken pinned sidecar, guard lazy repairs, and refresh stale retries
A valid pin whose transformers source dir vanished now triggers the repair
from the routing path (with a five minute backoff after failures) instead of
silently routing latest-only models to older tiers, the lazy repair refuses
while parent-visible chat/training/export workers are active since it has no
teardown of its own, and a version-mismatch install failure carries the
superseding release so the dialog's Retry re-requests a version that can
succeed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Flip latest-tier loads to 16-bit outside chat and protect export state
Training and export workers now apply the same latest-sidecar 16-bit flip
as the chat worker so a brand-new grouped-MoE architecture cannot reach bnb
4-bit through those paths, the latest-tier vision override returns None on
an inconclusive probe so a transient failure is not cached as not-vision,
and the install route refuses while an idle export checkpoint is loaded
rather than discard it with no rollback signal on a failed swap.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address parallel-review findings on the sidecar guards and install checks
The training route sizes latest-tier jobs 16-bit before GPU selection, the
inference subprocess spawn rechecks the swap reservation like training and
export (covering the OpenAI auto-switch path) with the typed error mapped
to a retryable 409, compat_plan blocks the install when dependency metadata
cannot be fetched instead of proceeding unverified, snapshot model-type
lists must contain only strings, and pin-marker package specs are validated
against the sidecar's own package set before ever reaching pip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Parent-only repairs, live-owner locks, remote-base activation, pre-teardown recheck
Lazy sidecar repairs now refuse inside worker children (whose empty backend
singletons cannot see live siblings) and run only in the parent where the
active-worker guard is real, swap-lock staleness requires the owner pid to
be dead so a slow live install is never superseded, both activation entry
points resolve a remote adapter's base model like the inference worker and
latest_tier_active_for already do, and load_model rechecks the reservation
before tearing down the old worker so losing the race keeps the current
model loaded.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Check workers under the repair reservation and keep state on refused swaps
The lazy repair now reserves first and checks workers under the reservation
(worker starts set their active markers before rechecking, so every
interleaving aborts one side), with export ops and in-flight inference loads
counted as active. The inference pre-teardown and spawn guards refuse only
repair reservations since an install shares the load's lifecycle gate and
aborts via its queued-load snapshot, a SidecarSwapInProgress raised before
teardown no longer clears the live model mirrors, and an export spawn abort
after teardown clears current_checkpoint so the page cannot claim a loaded
checkpoint with no worker.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Repair a present-but-incomplete latest sidecar from routing
The routing self-heal only fired when the pinned sidecar's transformers/
dir was missing. A sidecar that kept transformers/ but lost another pinned
package still routed models to the latest tier, and workers refuse
parent-only repairs, so every load failed until a manual reinstall. Routing
now validates the full pin (via _venv_dir_is_valid) and repairs any
incomplete sidecar under the same swap reservation and 5-minute backoff.
* Treat an unrepaired latest sidecar as unavailable in routing
When the pinned sidecar is incomplete and the lazy repair fails (offline,
pip failure, workers active) or is inside the backoff window, routing
returned the source dir anyway, sending models to a tier whose worker
activation is known to fail. Return None instead so models an older tier
supports keep loading there until a repair succeeds, matching the behavior
when the sidecar dir is missing entirely.
* Harden sidecar swap and repair against crash, survivor, and 16-bit paths
Reclaim a swap lock as soon as its recorded owner PID is dead instead of
waiting out the two-hour cutoff, so a crash mid-install no longer wedges
/load, training, export, and repair for hours. A lock whose PID cannot be
read yet still uses the long cutoff so the create-before-write window is
never mistaken for dead.
Probe process liveness with OpenProcess on Windows: os.kill(pid, 0) there
is CTRL_C_EVENT (a real Ctrl+C via GenerateConsoleCtrlEvent), not a
harmless check, and psutil is not always present.
Return whether _shutdown_subprocess actually killed the worker and keep the
live handle when it survives terminate/kill (an uninterruptible CUDA
syscall can outlive SIGKILL). The pre-swap liveness guard now trusts that
result, so the destructive .venv_t5_latest rename cannot proceed while a
live worker still holds sidecar modules.
Recover a sidecar stranded at .old when a swap's activation rename and its
rollback both fail: reading the pin restores it when no swap holds the
reservation, so latest-tier models are not permanently broken.
Resolve the latest tier in the parent for export loads and for explicitly
16-bit training runs, not only 4-bit ones: tier resolution self-heals an
incomplete sidecar, and repairs are parent-only, so those paths could not
recover before. Sidecar integrity and quantization are independent.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Revert the parent-side latest-tier repair probe on training and export loads
The probe ran before the route freed VRAM, so a resident chat or export worker
made _workers_active_for_repair() refuse the parent-only repair; the route then
tore that worker down and spawned a child that also cannot repair, so an
incomplete sidecar still failed to load. Repairing correctly requires running the
repair between the worker teardown and the child spawn, decoupled from VRAM
sizing, which is a larger change tracked separately. Restore the prior behavior
so these paths match the reviewed form and do not partially attempt a repair that
cannot complete while workers are resident.
* Honor failed worker shutdowns on load and revalidate the cached latest mapping
The fresh-load paths spawned a new worker straight after _shutdown_subprocess
without checking its result, so a worker that outlived terminate/kill (a wedged
CUDA syscall) had its handle overwritten by the replacement while it still held
GPU memory, and is_worker_alive/the pre-swap guard could no longer see it. Both
the inference load and the export checkpoint load now abort when the old worker
did not exit, so the load can be retried once it does.
_config_model_types returned a cached latest mapping without re-checking the
sidecar, so a sidecar deleted or broken in-process after its first parse was
never re-validated: routing kept sending latest-only models to the stale latest
tier while activation failed. The cached latest mapping is now dropped and
re-resolved (self-healing) when the sidecar is no longer intact.
* Drop cached latest mapping when the pin is gone; keep 4-bit for custom-code fallback
_latest_sidecar_intact now returns False when the pin marker itself is gone, not
just when a pinned package is missing. Otherwise a cached latest mapping outlived
a deleted pin: _config_model_types kept returning it, so routing sent latest-only
models to a tier whose worker activation then failed (no pinned version) until
restart. It now drops the cache and re-resolves to no latest tier. The
_overlay_transformers_dir caller already gates on a present pin, so it is
unaffected.
validate_model forced 16-bit sizing whenever a PyPI upgrade was merely offered,
even for a model that can fall back to its own auto_map code. /load loads such a
model 4-bit without the install, and the install route refuses while training is
active, so 16-bit sizing here returned a VRAM 409 for the only viable 4-bit path.
The offered-upgrade flip is now gated on the absence of a custom-code fallback;
an already-active latest sidecar still always sizes 16-bit.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: persistent stdio MCP sessions so server state survives across tool calls
call_tool_sync spawned a fresh stdio subprocess per tool call
(keep_alive=False) and tore it down when the call returned, so any stateful
MCP server lost its state between calls: with @playwright/mcp,
browser_navigate opened the page in one subprocess and
browser_take_screenshot ran in a brand-new one, screenshotting about:blank.
Keep one connected client per (command, env) on a dedicated event-loop
thread and reuse it across calls:
- idle sessions are reaped after 5 minutes (in-flight calls excluded) and
everything closes at exit, preserving the old design's no-orphans property
- a dead subprocess is detected via is_connected() and retried once on a
fresh session; tool-level errors leave the session alone
- cancel and timeout semantics are unchanged, and a timed-out call does not
tear the session down
- updating a server's endpoint/env/enabled state or deleting it closes its
live session
- HTTP/SSE servers stay one-shot per call
* address review feedback
* fix stdio session cleanup
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review: per-thread MCP scope, close-during-connect and abort races
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* address review: unblock no-limit calls on close, drain borrowers before close, scope closes to url+env
* don't retry sessions closed by config changes, re-verify server row before caching, keep env secrets out of generation keys
* fail fast on connect errors and make the stdio key-lock wait cancellable
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* quote MCP scope parts so IDs with colons can't collide
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* serialize per-session stdio calls, span one timeout budget across connect and call, hash urls in generation keys
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Harden persistent stdio MCP sessions: crash recovery, concurrency, scoping
- Evict a stdio session on any transport-level (non-ToolError) call failure and
do not replay it, so a mid-call subprocess crash can no longer poison the scope.
Never gate liveness on Client.is_connected() (it only reports that a session
object exists, not that the subprocess is alive); add a version-adaptive
dead-transport probe that works on fastmcp 3.0.2 and newer.
- Re-check closed/defunct/config and transport liveness after acquiring the call
lock, and retire a session before releasing the lock, so a queued same-scope
caller never reuses a session that another caller's timeout already retired.
- Force a ProactorEventLoop on Windows so the stdio transport can always spawn
subprocesses regardless of the active event-loop policy.
- Scope stdio sessions per conversation: require thread_id to persist, and tag
the fields so a session_id and a thread_id with the same value cannot collide.
A session_id alone is project-wide, so it now falls back to a safe one-shot
session instead of sharing browser/DB/REPL state across conversations.
- Forward thread_id on the Anthropic Messages path.
- Treat timeout=None as unlimited on connect and the key lock (was capped at 60s).
- Bound the session cache (default 32, override via
UNSLOTH_STUDIO_MAX_STDIO_MCP_SESSIONS) with LRU eviction of idle sessions.
- Run config_check on cache hits, and log a redacted exe#digest label instead of
the raw command so credentials in argv never reach the logs.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim the stdio MCP session cache on release and skip close-generation for HTTP servers
Two fixes from review of the persistent stdio session lifecycle:
- Re-enforce the session cap when a session goes idle. A concurrent burst of
distinct-scope calls can overshoot the cap while every cached session is busy
(insert-time eviction only reclaims idle sessions), and the overshoot used to
persist until the 5-minute idle reaper. _release_stdio_session now trims the
idle overshoot back within the cap, without ever evicting an in-flight call.
- close_stdio_sessions() now no-ops for a specific non-stdio (HTTP/SSE) url.
Those transports are never cached as stdio sessions, so calling it on every
HTTP server update or delete used to accrue an unbounded close-generation entry.
Both are covered by regression tests that fail before the change and pass after.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the live stdio MCP session across a display-name rename
The edit dialog resends url, headers, and use_oauth unchanged whenever a
server is saved, so gating the tool-cache invalidation and stdio session
close on field presence dropped the persistent process on a plain rename
or any no-op edit. Gate on a real value change against the stored row so
only a genuine endpoint, auth, or enable change closes the session.
Regression tests: a rename that resends unchanged url/headers/oauth keeps
the session; a real command change still closes it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments in the stdio MCP session lifecycle
Collapse a few verbose comments to fewer lines with the wording preserved,
and drop one that restated the clear_oauth_tokens_async docstring. Comments
only; no code change.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <danielhanchen@gmail.com>
Gallery clear/delete now scope to Studio-owned files: image_gallery and
video_gallery skip PNGs / MP4s without a readable recipe (a hand-dropped or
orphan file the listing already hides), so clear() and a guessed-id delete no
longer destroy files the gallery never surfaced.
Remote ControlNets now force use_safetensors: a bare owner/name reaches
from_pretrained without the base trust gate, and the Hub scan fails open when
unavailable, so requiring safetensors closes the pickle deserialization vector.
POSIX uninstall now stops resident sd-server / sd-cli under an owned sd.cpp root
before removing the tree (marker-gated), mirroring the Windows stop-before-delete
scan; a live native server no longer survives unlinking its binary.
Diffusion dataset containment: the training-start read path and the discovery
picker route bare names through the protected resolver, so a symlinked dataset
is rejected / not advertised like the caption/delete routes already do. Uploads
gain the inference decode guard (oversized real images 400 before OOMing the
trainer) and dataset upload/caption/delete/import are blocked with 409 while a
diffusion run is active.
JSONL readers (trainer + routes) tolerate non-object JSON and invalid UTF-8
instead of raising AttributeError / 500.
LoRA family compatibility is enforced in the shared resolver, not only the
picker, so a direct API client cannot apply a mismatched-family adapter.
GPU arbiter gains release_if so the image/video unload idle-check and release
are atomic against a concurrent same-owner load's registration. Native batch
recipes persist the base batch_seed and restore replays from it, so a native
batch_index>0 image no longer advances its seed twice.
FLUX.2-klein selects its sd.cpp text encoder by variant (4B -> Qwen3-4B,
9B -> Qwen3-8B) instead of the single family default.
POST /video/generate previously held the response open for the whole
generation (multi-minute for 720p), so in --secure mode the Cloudflare
quick tunnel's ~100s origin-response cap returned a 524 while the server
kept generating, and the frontend treated the run as failed.
Generation now follows the same return-at-once pattern as /video/load:
begin_generate validates synchronously (409 on no model or on a second
concurrent generate via a new busy sentinel) and runs the existing
generate + gallery-persist pipeline, with the route's exact error
mapping, on a daemon thread. GET /video/generate-progress gains optional
terminal fields: phase completed carries the saved gallery record, phase
failed a client-safe error; active only drops together with a terminal
phase. The cancel event is registered before the worker starts so
/video/generate/cancel keeps working across the whole job.
VideoGenerateResponse becomes an accepted acknowledgement (status
started, video kept as an always-null compat field). The video page
fires the POST, then drives completion off the progress poll it already
runs (completed prepends the clip, failed surfaces the error, the
cancelled sentinel stays toast-free). The API-key training-start guards
now also probe the video backend for an in-flight background clip, since
it is no longer visible as an in-flight HTTP request to the keep-warm
counter.
Route tests keep the fake backend for load/generate/status but inherit
the real job machinery, covering immediate accept, concurrent 409, the
terminal completed record, sanitized/ValueError/cancelled failures, and
cancel of a running job.
* Studio /v1/messages: accept thinking and unknown content blocks
The Anthropic-compatible /v1/messages endpoint modeled a message's content as
Union[str, list[{text|image|tool_use|tool_result}]], so any other block type
made Pydantic reject the whole request with
`messages.N.content.str: Input should be a valid string`. Resuming a Claude
session commonly replays assistant turns that carry `thinking` (extended
thinking) blocks, and sometimes a null content for a tool-only turn, both of
which tripped this and returned a 400.
Accept them:
- Add a permissive AnthropicUnknownBlock fallback (any block whose type is not
one of the four known ones), so thinking/redacted_thinking/provider-specific/
future blocks validate. A validator keeps known types on their typed models,
so a malformed known block (e.g. a tool_use without id) still fails cleanly.
- Coerce a null message (and tool_result) content to "" so the converter's
`for block in content` stays safe.
The converter already drops block types it does not translate, so a thinking
block is not forwarded to the model.
* Studio /v1/messages: keep user content validation strict
Make the thinking/null leniency role-aware so it never silently drops real
user input. Assistant turns (replayed history) still accept unknown/thinking
blocks and coerce a null tool-only turn to empty. User turns keep the strict
boundary: a null user content is rejected, and a content block the converter
cannot translate is rejected instead of being dropped into an empty prompt.
Also remove an empty file committed by accident.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio /v1/messages: coalesce resumed user turns and tighten content checks
- The /v1/messages count and generation paths now coalesce the adjacent user
turns that dropping an empty or null assistant turn can leave behind, so a
strict GGUF chat template no longer 400s on non-alternating roles.
- A user content block with a non-string type (list / dict) is rejected as a
clean 400 instead of raising TypeError and escaping as a 500.
- The assistant null-to-empty coercion only applies to an explicit null; an
assistant turn that omits content entirely still fails required-field
validation instead of being silently coerced to an empty string.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio /v1/messages: tighten comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: apply presence_penalty on the safetensors and MLX inference paths
The safetensors and MLX generate paths resolved the inference config and
then dropped presence_penalty before generation, so the same model applied
the configured value under GGUF and 0 under safetensors/MLX. Thread the
already-resolved presence_penalty through the orchestrator command, worker
gen_kwargs, and the safetensors/MLX generate calls, and apply it with a
small logits processor (subtract once per distinct completion token,
prompt excluded, presence not frequency, zero is a no-op, negatives raise).
Backwards compatible: presence_penalty defaults to 0.0 (byte-identical
output when unset) and the GGUF path is unchanged. Also forward min_p on
the legacy /generate/stream route and add the missing min_p field to
GenerateRequest.
* Studio: bound presence_penalty generated ids to valid vocab range on both paths
The presence-penalty logits processors index by generated token ids. The
torch path filtered only the upper bound (seen < vocab_size), so a negative
id would silently wrap to the wrong row; the MLX path had no bound at all,
and MLX out-of-bounds indexing is documented undefined behavior (crash or
memory corruption on Apple Silicon), unlike torch's harmless negative wrap.
Bound generated ids to [0, vocab) consistently on both paths:
- torch: seen[(seen >= 0) & (seen < vocab_size)] (zero-regression safety net;
real completion tokens are always in range).
- MLX: route out-of-range/negative ids to a discarded scratch slot via
mx.where and a (vocab + 1)-wide scatter-assign mask, then subtract. MLX has
no boolean-mask filtering (data-dependent output shape), so this keeps a
fixed shape, stays on-device, and preserves once-per-distinct-token
semantics without any torch/numpy dependency.
Add torch tests for out-of-range and negative ids (only in-range distinct
ids penalized, stray ids ignored, no wrong-index wrap) and a bound-documenting
MLX test that runs on the arm64 macOS CI.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Studio: client-tool passthrough healing for safetensors and MLX
PR 6801 made response-side tool-call healing default-on for the client-tool
passthrough, but only on the GGUF path: the passthrough branch in
/v1/chat/completions is gated on using_gguf, and the safetensors section never
reads payload.tools, so a client-tools request against a safetensors or MLX
model silently dropped the tool schemas and returned prose with no tool_calls.
Add the missing leg. When a non-GGUF model is loaded, the request declares
client tools (or carries tool-role history), server-side tools are off, and the
template supports tools, the route now:
- renders the tools into the chat template for a single turn via the existing
backend.generate_chat_response(..., tools=...) seam (worker templating
already accepts role=tool and assistant.tool_calls messages, normalized with
_openai_messages_for_passthrough);
- non-streaming: promotes text-form calls with heal_openai_message, honors the
opt-in nudge single retry (nudge_should_retry / nudge_messages), caps healed
calls when parallel_tool_calls=false (covers the nudge retry too), and sets
finish_reason=tool_calls with content null on a pure tool-call turn;
- streaming: derives deltas from the worker's cumulative snapshots and feeds
StreamToolCallHealer, emitting healed tool-call deltas and the correct
finish chunk, guarded against repeated or shrinking snapshots.
heal_gate semantics are identical to the GGUF passthrough: default on,
auto_heal_tool_calls=false or UNSLOTH_DISABLE_TOOL_CALL_HEALING=1 relays
verbatim, tool_choice narrows promotion, undeclared names stay text. MLX rides
the same orchestrator seam, so both local backends gain the behavior.
CompletionMessage.content becomes Optional so a promoted pure tool-call turn
matches the OpenAI contract (content null when only tool_calls return).
Adds tests/test_sf_client_tools_passthrough.py (22 cases: healing, gating,
opt-outs, streaming deltas, tool-role history, dict-arguments history, forced
tool_choice, parallel cap, usage, nudge on/off/double-failure, generator error
hygiene, disconnect reset, empty output, MLX path).
* Address review: tool_choice none, developer folding, retry fallback, monitor reply
Four review follow-ups on the safetensors/MLX client-tool passthrough leg:
- tool_choice="none" keeps the tool-history templating but no longer
advertises the tools, so a forced final-answer turn is not prompted into
emitting markup that the (correctly disabled) healer would relay as prose.
Mirrors the GGUF passthrough where llama-server honors tool_choice itself.
- OpenAI "developer" messages fold into a single leading system message via
_set_or_prepend_system_message before templating; local templates reject the
role and the fallback formatter drops it.
- A nudge retry that fails or is cancelled after the original answer exists
falls back to the first response instead of surfacing a 500, matching the
GGUF nudge path.
- The API monitor records the healed tool call summary instead of the raw
markup on a promoted turn.
Adds four regression tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: forced tool_choice templating, content-part flattening, stream monitor parity
- A forced tool_choice function is now the only schema rendered into the
local template, so the advertised tools and the healer allowlist can no
longer disagree (llama-server enforces tool_choice itself on the GGUF path).
- Content-part lists are flattened to their text parts before templating.
Remote image URLs are not decodable locally, so such requests reached this
path with part lists that raise inside apply_chat_template on text-only
templates; the plain non-GGUF path has always flattened them.
- The streaming monitor entry is now fed from the healed events the client
actually receives, recording promoted calls as the [tool_calls] summary
the non-streaming path records.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: gate passthrough on the engaged server path, deserialize templated arguments
- The client-tools gate now keys on _sf_use_tools (whether the server-side
tool path actually claimed the request) instead of the raw mcp_enabled
flag: with an empty MCP registry or a CLI --disable-tools policy, a client
that sets mcp_enabled while declaring its own tools fell through to plain
generation with the tools silently dropped. The GGUF passthrough gate has
no mcp_enabled clause either.
- New _structured_tool_history_for_local_template deserializes assistant
tool_calls[].function.arguments JSON strings into mappings for the
templated copy only: spec-compliant clients send strings, but local chat
templates iterate arguments as a mapping or raise on strings, which
crashed or misrendered multi-turn tool history. The HTTP response and the
GGUF wire shape keep strings.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tighten comments and docstrings in the client-tools passthrough
* Report first-attempt usage when a nudge retry is discarded
When nudge_should_retry fires but the retry produces no healable tool call
(or raises), the first response is still delivered to the client. The retry's
generate() had already overwritten stats_holder, so _monitor_usage recorded
the unseen retry's token counts against the request instead of the first
attempt that was actually returned. Capture the first attempt's stats before
the retry and restore them on both the no-heal and exception paths so the
monitor reports the usage of the response the caller received.
* Do not promote buffered tool markup when a stream is cancelled
The streaming client-tool heal path breaks out of the token loop when
cancel_event is set (the registry "Stop" path), but then still fell through to
healer.finalize(), which heals incomplete tool markup at EOF (allow_incomplete)
and emits a tool_calls delta plus finish_reason=tool_calls. Because the Stop
request only sets the event and leaves the SSE socket open, the client received
that promoted call and executed a tool the user had just cancelled. The disconnect
path already returns before finalize; guard finalize and the finish_reason on
cancel_event too, so a cancelled stream ends with finish_reason=stop and no tool
call. Adds a regression test driving a Stop mid-emission with buffered markup.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim comments in the client-tools passthrough
* Trim client-tools passthrough comments further
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Auto policies: deferred dense compile, video compile default, step cache and precision auto
Image dense loads with speed unset no longer sit at plain off: the load stays
bit-identical eager, and the 3rd generation in a session engages the default
compile profile plus the cuDNN attention upgrade mid-session (a one-off image
never pays the warmup, repeated use amortises it). Video dense loads resolve
straight to the default profile since a clip denoise amortises the compile
within a single run, and never to max.
Video also gains the image backend's tri-state auto policies: unset step cache
now decides from the default schedule and re-checks the actual step count per
generation, and unset precision (transformer_quant) hands the decision to the
hardware ladder instead of staying off. Memory badge reason now says plainly
that everything fits when no offload is planned.
* Rename Dtype to Precision, add the video Precision control, step cache Auto option
The images Advanced panel's Dtype row is now Precision (same control, clearer
name), and the video Advanced panel gains the matching Precision select wired
to the load route's existing transformer_quant field, gated to full-pipeline
loads the way the image control gates to GGUF. Step cache selects on both
pages gain an explicit Auto option as the default (the previous Off default
silently behaved as auto and never let anyone pin off), and the Speed and
Attention tooltips now state the deferred dense compile and the SageAttention
black-frame caveat.
* Model catalog: canonical diffusion model groups with device-aware routing
One canonical name per image/video model, its published artifacts (GGUF, FP8,
bnb-4bit, official BF16) as data, and pure routing helpers: suffix-stripped
canonical keys (owner-preserving; cross-owner merges only via explicit
aliases), group/artifact lookups, a flat back-compat options shim, load-spec
resolution replacing the pages' lookup tables, search matching over old ids
and format tokens, the GGUF fit ladder extracted from the variant expander,
and pickDefaultArtifact/pickDefaultQuant deciding what a bare group click
loads (downloaded first, then the best quality that fits 70 percent of VRAM,
GGUF as the safe fallback). Checked by npm run catalog:check, following the
i18n:check pattern.
* Picker: one canonical row per diffusion model with a format second level
The Images and Video pickers now render the curated catalog as one row per
model in Recommended: clicking loads the best artifact for the device (the
routed GGUF quant, a prequant FP8/bnb-4bit that fits, or the official BF16),
and a chevron opens the per-format list, with the GGUF row nesting the usual
quant expander. Live HF listing rows that belong to a group are deduplicated,
search collapses member repos into their group (old ids and format tokens
still match), and the On Device sections group cached member repos under the
same canonical name with the per-repo rows inside. Curated groups render from
the catalog rather than the HF listing, which finally surfaces LTX-2.3 in the
video Recommended list (its hub pipeline_tag is image-to-video, so the
text-to-video listing always missed it) and exposes the HunyuanVideo 720p
repack next to 480p.
Backend: /cached-models now tags trusted video-family repos text-to-video
instead of blanket text-to-image, and the pickers admit catalog-known
non-unsloth repos On Device, so cached Lightricks/Wan/Hunyuan pipelines
finally appear in the Video picker. Chat pickers pass no catalog and are
unchanged.
* Download formats, tab icons, plain-language train tips, 3-loop autoplay
The image Download button becomes a menu: PNG saves the original bytes with
the embedded recipe, JPEG and WebP re-encode client-side from the fetched
blob (JPEG flattened onto white). The video Download button gains MP4
(original, keeps audio), WebM and GIF; the latter two transcode server-side
from the stored MP4 via PyAV (VP9 realtime profile for WebM, ~12 fps adaptive
palette for GIF) behind a new gallery export route that 501s with a readable
message when a codec is missing.
Generated clips no longer loop forever: the player replays a clip three times
per selection, then pauses with controls up; a new generation or a refresh
gets its own three plays. The Create/Train tabs reuse the sidebar's New Chat
and Train icons (TestTubeOutlineIcon moved to a shared lib module), and every
Train tab helper text is now one plain sentence.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Keep the Create/Train tab icon and label on one line
TabsTrigger renders its children inside a plain inline span and the
Tailwind preflight gives svg display:block, so the HugeiconsIcon forced
the label onto a second line. Wrap icon plus label in their own
inline flex row inside each trigger.
* Strip -int8 and -nvfp4 prequant suffixes in the model catalog key
canonicalKeyFor already lowercases before matching, so -GGUF/-FP8 in any
case were covered; -int8 and -nvfp4 were not in the suffix table, so
such repos rendered as standalone rows in Recommended and On Device
instead of standardizing into their base-name group and routing through
pickDefaultArtifact. Added both suffixes plus case-insensitivity and
routing assertions to the catalog check.
* Standardize non-catalog picker rows to their base model name
The curated catalog already collapses its own groups, but hub listing
rows and cached repos outside the catalog (ERNIE-Image, FLUX.2-klein,
Qwen-Image-Edit-2509, FLUX.2-dev) still rendered raw ids with -GGUF /
-FP8 style suffixes in Recommended and On Device.
- model-catalog.ts: new stripArtifactSuffixesForDisplay, a
case-preserving twin of canonicalKeyFor's stripping that keeps the
owner prefix and original casing for display.
- pickers.tsx: recommended hub rows and the downloaded GGUF/model rows
pass their labels through it when a catalog is present, so only the
diffusion pickers change; chat rows keep raw ids. Click targets keep
the full repo id, and the format badge still shows the artifact kind.
- Catalog check covers the new helper across GGUF/FP8/int8/nvfp4 in
both cases plus no-op and suffix-only names.
* Offer official BF16/FP8 artifacts per model group and fix gallery label clipping
Model picker changes so groups are not limited to unsloth quant repos:
- model-catalog.ts: each image group that has an official vendor pipeline
now carries its BF16 (official) artifact as the top (highest quality)
entry - Tongyi-MAI/Z-Image-Turbo, Qwen/Qwen-Image, Qwen/Qwen-Image-2512,
Qwen/Qwen-Image-Edit-2511, black-forest-labs/FLUX.1-dev, FLUX.1-schnell
and FLUX.1-Kontext-dev. The LTX-2.3 video group now lists Lightricks'
own bf16 and fp8 distilled single-file checkpoints alongside the GGUF.
Resident sizes are set from the actual weight totals (FLUX ships a
duplicate single-file that from_pretrained ignores, so FLUX bf16 is ~32
GB not 54). The repos that used to be aliases are now real artifacts.
- The router already prefers the highest-quality artifact that fits the
0.7 x GPU budget, so a datacenter GPU now defaults to official BF16
while consumer GPUs still route to the fitting quant or GGUF. That is
why bnb-4bit was the Z-Image-Turbo default before: it was the only
non-GGUF artifact and it was already downloaded.
- diffusion.py: allowlist the four official image repos not previously
trusted (qwen/qwen-image-2512, qwen/qwen-image-edit-2511,
black-forest-labs/flux.1-schnell, flux.1-kontext-dev). All verified as
safetensors-only diffusers model_index pipelines. The LTX-2.3
checkpoints are already on the video trust list.
- catalog check: BF16-wins-on-datacenter, quant-wins-on-consumer, and the
single-file load specs for the LTX-2.3 checkpoints.
Also fixes the video gallery thumbnail caption: the leading duration was
clipped by the rounded corner and selection border, so the strip now has
enough left/bottom padding to clear the curve.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* video gallery: guard export transcode against a stream-less clip
_transcode_webm and _transcode_gif indexed src.streams.video[0] before
checking the stream list, so a container with no video stream raised a bare
IndexError that the broad handlers then re-labeled as a missing libvpx or
decoder. Raise an explicit RuntimeError naming the real cause in both the
WebM and GIF paths.
* Studio: honor explicit attention/format choices, fix distilled-LTX defaults and On Device catalog routing
* Remove stray planning notes accidentally committed to the branch
* video: add transformerQuant to the load callback deps
handleLoad reads transformerQuant but omitted it from the useCallback dep array,
so after the user changes only Precision and then selects a model or clicks
Reapply, the memoized callback keeps the stale closure and loads the previous
precision. The image page's equivalent callback already lists it.
* model picker: honor the format filter when routing catalog clicks; add catalog rows to the roving list
- routedArtifactFor now scopes a group's artifacts to the active format filter
(the same matchesFormatFilter predicate the visibility check uses) before
pickDefaultArtifact, so a group shown only because it owns a GGUF no longer
routes a click to a large non-GGUF download. Covers both the Recommended and
On Device grouped paths.
- hubOptionKeys now includes the catalog-group, search-catalog-group, and grouped
On Device row keys in exact render order, so arrow/Home/End roving reaches the
catalog rows instead of giving them a duplicate missing id and skipping them.
* model picker: don't treat a partial base cache as downloaded
A partially-cached base repo (a cancelled download that left only some weights)
was counted as downloaded, so an On Device click routed to a fresh multi-GB
re-download instead of the complete GGUF. The picker's endpoint (/api/models/
cached-models) did not carry a partial flag at all, so a frontend-only guard
could not see it. Surface partial from that endpoint by reusing the hub inventory
scan's snapshot-partial detector, plumb it through CachedModelRepo (backend +
frontend types), and skip partial base repos when building the downloaded set.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* model picker + diffusion: drop partial/unloadable cached rows, skip defer-compile before a LoRA gen
- On Device (cached non-GGUF) rows filtered partial-download snapshots back in: sortedCachedModels
gated on passesTaskGate + a groupForRepoId key match but, unlike downloadedSet, never checked
c.partial, so an incomplete unsloth snapshot showed as a loadable On Device row (click errors or
silently re-fetches multi-GB). It also admitted repos that only match the catalog by group KEY
(a base / uncurated-quant sibling like Qwen/Qwen-Image-2512) which have no loadable artifact and
dead-end at the trust gate. Add !c.partial and gate on artifactForRepoId (what loadSpecFor
resolves) instead of groupForRepoId, so a cached row shows only when the backend can load it.
- Deferred speed-auto engaged the compile profile on the 3rd generation BEFORE _apply_loras. A
compiled transformer rejects LoRA (supports_lora is False) and _apply_loras raises before its
unchanged-selection no-op, so once compile engaged every LoRA generation on that load failed
permanently. Skip the deferral when a LoRA is requested (compile and LoRA are mutually exclusive)
and let it engage on a later LoRA-free generation.
* Scope the cached-model partial probe to the listed snapshot dir
list_cached_models builds each row from the largest/complete copy across HF cache
roots, but _cached_repo_partial probed is_snapshot_partial with no repo_cache_dir,
so the scan spanned every root: a stale .incomplete copy in one root would flag a
complete copy in another as partial and hide the usable model from the picker (the
click then routes to a re-download). Forward the winning snapshot's repo_path so all
three partial signals are scoped to that copy, matching the sibling inventory paths
(models/dataset cache_inventory, local_inventory).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Do not auto-route to gated repos, prefer complete cached copies, defer compile past attached LoRA, scope group expand keys
Four fixes:
- pickDefaultArtifact's not-downloaded ladder returned the gated BF16 FLUX.1-dev / Kontext-dev
before the open GGUF on a large GPU, so a bare group click routed to a repo the user may lack
license/token access to. Add a gated flag and skip gated artifacts in the not-downloaded ladder
(an already-downloaded gated artifact is still returned).
- list_cached_models picked the largest duplicate cache copy and computed partial only on it, so a
larger partial copy shadowed a smaller complete one; since partial rows are dropped from the
picker the usable model vanished. Prefer completeness, then size.
- the deferred-speed compile engaged on a no-LoRA generation while an adapter from a prior
generation was still attached, baking it into the compiled graph (the later unload is swallowed
on a compiled pipe); also defer while adapters remain attached.
- routeGroupClick's GGUF fallback toggled the context-free canonicalId while the chevron toggles
the context-scoped expandKey, leaving the format list un-collapsible in one context, dead in the
other, and risking cross-context expansion; thread expandKey through.
* Guard video pipeline repos from deletion, drop the always-failing LTX FP8 artifact, prefer 720p Hunyuan
Three round-6 fixes:
- cached non-GGUF video repos now surface in the Video On-Device picker with the normal delete
action, but /delete-cached only guarded chat + the Images engine, so a loaded/loading Wan / LTX /
Hunyuan pipeline could have its HF snapshot removed from under it. Add a VideoBackend
loading_repo_ids accessor and a video loaded/loading guard mirroring the Images one.
- the catalog advertised Lightricks/LTX-2.3-fp8 as loadable, but the LTX-2.3 loader refuses the
official scaled-FP8 single file (.weight_scale/.input_scale) and points to GGUF/BF16, so a pick
routed to a ~76 GB download that always fails on load. Remove the FP8 artifact.
- pickDefaultArtifact only sorts by format, so the HunyuanVideo group's 480p (listed first) beat
the 720p even on GPUs where 720p fits the budget. List 720p first so the fit loop prefers it and
falls back to 480p only on smaller cards.
* diffusion: add compute int8/fp8_dynamic text-encoder quant, wire into video
Add two torchao compute text-encoder quant modes to the diffusion precision
engine, alongside the existing layerwise fp8 and weight-only nvfp4:
- int8: per-token activation + per-channel weight (torch._int_mm), with per-layer
keep-bf16 selection. int8 degrades on large encoders unless the most
quant-sensitive decoder blocks stay bf16, so it engages only for families with
a measured keep-bf16 schedule (qwen-image / qwen-image-edit keep first+last 6,
flux.2-dev keeps first 3); a family without one falls back to fp8.
- fp8_dynamic: per-row fp8 compute (torch._scaled_mm), keeping the matmul in fp8
on the tensor cores instead of upcasting each forward like the layerwise fp8.
The selective int8 caster reuses the committed transformer-quant factory
(_make_quant_config / make_filter_fn / exclude_tokens_for_scheme) plus a small
structural first/last-N block skip, so it depends only on committed APIs.
Wire text-encoder quant into the video backend, which previously loaded the
companion encoder (Gemma3 / UMT5 / Qwen2.5-VL) dense bf16 while quantising only
the DiT. text_encoder_quant is plumbed through the load request, validation, the
load chain, the resolved record, and status, mirroring the image backend; it
applies for every load kind (the encoder is dense regardless of how the DiT was
sourced). Widen the image and video load request Literals and add the video
status field.
Tests: int8 family-schedule routing and fp8 fallback, fp8_dynamic routing,
hardware gates (int8 sm_80+, fp8_dynamic sm_89+), the structural block selection,
the real int8 filter closure (keeps the first blocks plus the vision tower /
lm_head / T5 wo dense), and the video route threading and 422 validation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* text-encoder quant: skip the torchao modes under offload (both backends)
quantize_text_encoders applied int8-with-schedule / fp8_dynamic / nvfp4 (all torchao) to the
text encoder regardless of the offload policy. An offload placement then moves the quantized encoder
with Module.to(), which torchao tensor subclasses reject (aten._has_compatible_shallow_copy_type is
unimplemented) -- a hard crash, the same one the DiT path already skips torchao quant under offload to
avoid. Add offload_active to quantize_text_encoders and skip the torchao modes when set; layerwise fp8
is not torchao and still streams under offload. Both the video and image loaders pass
offload_active = (offload policy != none).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* diffusion: skip non-bf16 linears for scaled_mm quant schemes
The fp8 / mxfp8 / nvfp4 schemes run on torch._scaled_mm and the fp4 / mx GEMMs,
which assert a bfloat16 input weight. On a mixed-precision DiT that keeps some
linears in fp32 for numerical stability (the Wan and Hunyuan video transformers
do this), quantize_ hits the first fp32 linear, raises, and the best-effort
wrapper swallows it to None, so the whole transformer stays dense with no error
and no speedup or memory saving.
Add a require_bf16 gate to make_filter_fn and pass it for the scaled_mm schemes
in quantize_transformer (and the fp8_dynamic text-encoder caster). The gate
skips non-bf16 linears so the scheme engages on the bf16 ones. int8 uses
torch._int_mm, which quantizes fp32/fp16 weights fine, so it leaves the gate off
and keeps its current coverage.
Verified on Wan2.2-TI2V-5B: fp8 and mxfp8 now quantize 303 linears via the
committed quantize_transformer path where they previously engaged 0.
* prequant builder: mirror the scaled-mm bf16 gate offline
The runtime DiT quantizer skips non-bf16 Linears for the scaled_mm schemes (fp8,
nvfp4, mxfp8) so the scheme engages on a mixed-precision transformer instead of
aborting on the first fp32 Linear. The offline prequant builder reused make_filter_fn
without that gate, so building an fp8/nvfp4/mxfp8 checkpoint for a mixed-precision DiT
(Wan, Hunyuan keep _keep_in_fp32_modules in fp32 even under torch_dtype=bf16) would hit
the same fp32 Linear and abort, breaking the builder's stated offline == runtime,
LPIPS-0 invariant. Thread require_bf16 = scheme in _SCALED_MM_SCHEMES through the builder,
record it in the checkpoint metadata, and verify it on load (mirrors the existing
exclude_name_tokens guard) so a future _SCALED_MM_SCHEMES change cannot silently load a
checkpoint built under the old filter.
* Keep nvfp4 fp32 linears quantised (bf16 gate is fp8/mxfp8 only)
Verified on torchao 0.17 / B200: fp8 per-row asserts 'PerRow quantization only
works for bfloat16 precision input weight' and mxfp8 asserts 'Only supporting bf16
out dtype', but NVFP4's high-precision conversion quantises an fp32 weight fine
(forward included). So the bf16 skip-gate must be fp8/mxfp8 only, not all scaled_mm
schemes -- otherwise nvfp4 leaves large fp32 projections dense, losing the intended
memory/speed gain. Rename _SCALED_MM_SCHEMES -> _REQUIRE_BF16_SCHEMES = (fp8, mxfp8)
and thread it through the runtime filter, the offline builder, and the loader
require_bf16 verification (offline == runtime preserved).
* [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>
- _dense_quant_prefetch_needed widened the transformer/ prefetch to pull the base
repo's full dense bf16 shards even when a prequant checkpoint is configured
(candidate.prequant), contradicting its own docstring. That both defeats the
prequant download savings and can hard-fail begin_load on a disk-full (no GGUF
fallback there). Only widen for a real dense build (candidate is not None and
not candidate.prequant).
- DiffusionStatusResponse declared no 'resolved' field, so Pydantic's default
extra='ignore' silently dropped the per-control auto-policy provenance the
backend records (build_resolved_record / state.resolved) -- the plumbing never
reached any client. Declare the field so it round-trips.
Re-plan memory with the quant steady factor when the bf16 table forces
offload a quantised DiT would not need, mirroring the image dense-quant
path, and fall back to the bf16 plan when quant does not engage. Stream
the second expert under group offload (model and sequential already hook
every module). Fail the load cleanly when quant engages on only one
expert instead of running mixed precision with quant reported off.
Persist guidance_2 in the gallery recipe so A14B clips are reproducible.
Register two new video families and wire them through the backend, routes,
and frontend picker: wan2.2-ti2v-5b (single DiT) and wan2.2-t2v-a14b (the
dual-expert MoE). Both share diffusers' WanPipeline + WanTransformer3DModel
+ AutoencoderKLWan, which the VideoFamily dataclass already reserved fields
for (transformer2_class, is_moe, cfg2_kwarg).
Verified against the installed diffusers 0.39.0 before writing code:
- WanPipeline, WanTransformer3DModel, and AutoencoderKLWan are all exported
from top-level diffusers 0.39.0.
- WanPipeline.__call__ (pipeline_wan.py:383) defaults to num_frames=81,
num_inference_steps=50, guidance_scale=5.0. guidance_scale_2 DOES exist
in 0.39 (line 392) and its check_inputs raises if it is passed when the
pipeline's boundary_ratio is None (line 322), so the second guidance is
threaded ONLY for the MoE family and only when inspect.signature accepts
it (the same gate frame_rate already uses).
- The Wan VAE temporal factor is 4 (autoencoder_kl_wan.py scale_factor_temporal),
and the pipeline snaps num_frames to 4k+1 (line 493), so frame_step is 4,
unlike LTX-2's 8k+1. Sizes patchify at spatial 8 * patch 2 = 16, so
resolution_multiple is 16.
- boundary_ratio and transformer_2 come from model_index.json: TI2V-5B ships
boundary_ratio=null and transformer_2=[null,null] (single DiT), while A14B
ships boundary_ratio=0.875 and transformer_2=WanTransformer3DModel (dual
DiT). boundary_ratio lives in the pipeline config, so it needs no per-call
plumbing.
- WanTransformer3DModel declares _repeated_blocks=["WanTransformerBlock"] and
inherits CacheMixin (transformer_wan.py:508/551), so regional compile and
First-Block-Cache both work.
bf16-resident component sizes, measured from each diffusers repo's on-disk
safetensors (all stored bf16), feed the auto memory table:
TI2V-5B: transformer 20.0, UMT5 text encoder 11.4, VAE 2.8 GB.
A14B: two experts 57.2 each (114.3 total), text encoder 11.4, VAE 0.5 GB.
Backend changes make the optimisation layers dual-DiT aware: a small
_SecondDiTView proxy presents transformer_2 as pipe.transformer so the
existing single-DiT helpers (apply_speed_optims, apply_attention_backend,
apply_step_cache, quantize_transformer) cover BOTH experts on an is_moe load
without forking any helper; single-DiT loads are unchanged (views is just
(pipe,)). The two Wan base repos are added to the trusted non-GGUF allowlist.
A transformer_quant option is added to the load path, mirroring the image
backend's dense torchao fast path: on a pipeline-kind load the dense DiT(s)
are quantised in place onto the low-precision tensor cores and the engaged
scheme is surfaced in status. generate() threads guidance_2 through the
family's cfg2_kwarg when the loaded pipeline accepts it.
Routes and Pydantic models gain the optional transformer_quant (load /
status) and guidance_2 (generate) fields. The frontend picker gains the two
Wan models with 50-step / CFG 5.0 defaults; fps is supplied per family by
the backend.
Tests extend the fake runtime with WanPipeline and per-DiT transformer fakes
(single-DiT and dual-DiT), and cover family detection for both repos, 4k+1
frame snapping, default application, dual-DiT speed/cache/attention/quant
coverage on both experts, cfg2 threading gated on the pipeline signature,
trusted-repo validation, and the new route fields. Both the standard and the
diffusers/torchao-blocked CI-sim runs are green.
routes/video.py mirrors the /images/* routes one-for-one: validate-before-evict
load ordering (a bad pick must not evict a working chat model and then 400),
the training-active interlock, the device-gated GPU arbiter handoff with the
new VIDEO owner, the exact-match sentinel mapping (VIDEO_NOT_LOADED_MSG /
VIDEO_CANCELLED_MSG to 409, ValueError/FileNotFoundError to 400 with native
paths redacted, everything else a sanitized 500), and the gallery CRUD shape
with fetch-one-extra has_more paging. Generate persists the encoded MP4 plus
its full recipe through video_gallery.save and returns the gallery record; the
file endpoint serves video/mp4 with an immutable Cache-Control, 404 on any id
that fails the containment check.
models/inference.py gains the video request/response set (VideoLoadRequest,
VideoGenerateRequest/Response, GalleryVideo, gallery list, both progress
shapes, VideoGenerationDefaults nested in VideoStatusResponse), reusing
DiffusionResolvedControl for the resolved-provenance badges. The router is
registered in main.py after the images router under the same /api/inference
prefix and auth dependency.
Tests: 20 route tests on a stubbed backend + real tmp gallery (load happy path
and arbiter acquisition, 400/409 mappings, generate persistence round trip,
mp4 file serving + 404, delete/clear, unload releases ownership); the
diffusion route suite stays green after the shared models edit.
An unset transformer_quant used to mean off (run the GGUF as-is), so the
hardware ladder only engaged when auto was explicitly chosen and the panel
showed Off as the default. Unset (or auto) now hands the decision to the
ladder: a dense-capable GPU gets at least int8, data-center silicon fp8,
falling back to the GGUF when the device, VRAM, family deny table or disk
cannot take it. An explicit none/off pins GGUF-as-is and is now
expressible in the API (previously only omission meant off, so pinned-off
and unset were indistinguishable); an explicit scheme pins that scheme.
The dense candidate also gains a free-disk gate: with auto as the default
the bf16 base download (up to ~40 GB) must never wedge a nearly-full
model-cache disk, so the candidate is dropped (GGUF build kept) when free
space cannot hold it plus a 10 GiB margin. Unprobeable disk passes.
Frontend: the Dtype select defaults to Auto (fastest for GPU), keeps Off
as an explicit choice, and sends none through instead of omitting it.
Suite: 622 diffusion tests green (default-load test rewritten to the new
contract, explicit-off short-circuit covered), CI-sim green.
GET /api/inference/images/info returns each family's bf16 component sizes and
the estimated resident GB under bf16/int8/fp8/mxfp8/nvfp4, computed purely from
the auto-policy tables (no GPU probing, torch-free), so the panel can show the
Dtype tradeoff before anything is loaded.
DiffusionStatusResponse gains an additive resolved field: per-control
{value, source, reason} provenance the loader already records. The Advanced
panel renders a muted Auto: X pill next to Speed / Dtype / Attention / Memory /
Step cache / CPU offload when the backend decided that control (source auto),
with the reason as the tooltip; an explicit user choice renders no badge.
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: eager patches + torch.compile cache speed phase
Adds the opt-in speed path for the GGUF diffusion transformer behind a
selectable speed mode (default off, so output is unchanged until a profile
is chosen):
- diffusion_eager_patches.py: shared eager fast-paths (channels_last,
attention/backend selection, fused norms and QKV) installed at load and
rolled back on unload or failed load.
- diffusion_compile_cache.py / diffusion_gguf_compile.py: a persistent
torch.compile cache and the GGUF-transformer compile wiring.
- diffusion_arch_patches.py: architecture-specific patches.
- diffusion_patch_backend.py: shared install/restore plumbing.
- diffusion_speed.py: speed-profile planning.
Tests for each module plus the benchmarking and probe scripts used to
measure speed, memory, and accuracy of the path.
* Studio diffusion: image workflows (safetensors, image-conditioned, editing) + Images UI
Backend:
- Load non-GGUF safetensors models: full bnb-4bit pipelines and single-file
fp8 transformers, gated to the unsloth org plus a curated allowlist.
- Image-conditioned workflows built with Pipeline.from_pipe so they reuse the
loaded transformer/VAE/text-encoder with no extra VRAM: img2img, inpaint,
outpaint, and a hires-fix upscale pass.
- Instruction editing as its own family kind (Qwen-Image-Edit-2511,
FLUX.1-Kontext-dev) and FLUX.2-klein reference conditioning (single and
multi-reference) plus klein inpaint.
- Auto-resize odd-sized inputs to a multiple of 16 (and resize the matched
mask) so img2img/inpaint/edit no longer reject non-/16 uploads. Bound the
decoded image size and cap upscale output to avoid OOM on large inputs.
- Fixes: from_pipe defaulting to a float32 recast that crashed torchao
quantized transformers; image-conditioned calls forcing the slider size
onto the input image. Native sd.cpp engine rejects image-conditioned and
reference requests it cannot serve.
Frontend:
- Redesigned Images page with capability-gated workflow tabs (Create,
Transform, Inpaint, Extend, Upscale, Reference, Edit), a brush mask editor,
client-side outpaint, and a multi-reference picker.
- Advanced options moved to a right-docked panel mirroring Chat: closed by
default, toggled by a single fixed top-bar button that stays in place.
sd.cpp installer: pin the release, verify each download's sha256, add a
download timeout, and make the source repo configurable for a future mirror.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: correct the Advanced panel comment (closed by default, fixed toggle)
* Studio diffusion: LoRA adapters for the Images workflow
Add community LoRA support across both diffusion backends, the single
biggest step toward broad image-workflow coverage.
Backend
- New shared module core/inference/diffusion_lora.py: adapter discovery
(local scan + curated catalog + owner/name[:file] Hub refs), download
via hf_hub_download_with_xet_fallback, alias sanitization, native
managed-dir materialization with collision-broken aliases, prompt-tag
injection (deduped against user-typed tags), and a supports_lora gate.
- Native sd-cli: resolve + materialize selected LoRAs into a per-run
managed dir, inject <lora:ALIAS:w> tags, pass --lora-model-dir with
--lora-apply-mode auto. The arg builder already emitted these flags.
- Diffusers: non-fused load_lora_weights + set_adapters manager, tracked
on the pipe so an unchanged selection is a no-op and a model swap
resets; cleared on unload. Never fuses (breaks quantized transformers
and blocks live weight tweaks).
- Gated off where unsupported: torchao fp8/int8 dense, GGUF-via-diffusers,
and native Qwen-Image (no LoRA name-conversion branch upstream).
- Request contract: optional loras on DiffusionGenerateRequest; empty or
omitted is identical to today. supports_lora surfaced in status; chosen
LoRAs persisted in gallery recipe metadata.
- New GET /api/models/diffusion-loras for the picker (family-filtered).
Frontend
- Repeatable multi-LoRA picker (adapter select + weight slider 0..2 +
remove), gated by the loaded model's supports_lora and family, max 8.
Tests
- New test_diffusion_lora.py (14): helpers, request validation, native
tag/dir wiring, diffusers set_adapters manager, supports_lora matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: ControlNet for the Images workflow (diffusers)
Add ControlNet conditioning, the #2 most-used diffusion workflow after
LoRA, on the diffusers backend for the families with ControlNet pipelines
(FLUX.1 and Qwen-Image), with Union models as the default picks.
Backend
- New core/inference/diffusion_controlnet.py: family-gated discovery
(curated Union models + local dirs + bare owner/name repos), resolution
to a loadable repo/dir, control-image preprocessing (passthrough +
a dependency-free canny edge map), and a supports_controlnet gate.
- diffusion.py: a ControlNet manager parallel to the LoRA one. Loads the
(small) ControlNet model once via from_pretrained (cached by id) and
builds the family's ControlNet pipeline via Pipeline.from_pipe(base,
controlnet=model), reusing the resident base modules at their loaded
dtype (no reload, no recast). Passes the control image + conditioning
scale + guidance start/end at generate time; cleared on unload.
- Families: FLUX.1 -> FluxControlNetPipeline/Model, Qwen-Image ->
QwenImageControlNetPipeline/Model. Others declare none (gated off).
- Gated off for the native engine, GGUF-via-diffusers, and torchao
fp8/int8 dense (same rule as LoRA). v1 conditions txt2img only.
- Request contract: optional controlnet on DiffusionGenerateRequest;
supports_controlnet in status; the choice persisted in gallery meta.
- New GET /api/models/diffusion-controlnets for the picker.
Frontend
- A ControlNet control in the Images rail (model select + control-image
upload + control-type select + strength slider), gated by the loaded
model's supports_controlnet + family, shown for text-to-image.
Tests
- New test_diffusion_controlnet.py (10): discovery/resolve/preprocess/gate
helpers, request validation, family wiring, and the diffusers pipe
manager (loads once, caches, from_pipe with controlnet, rejects
unsupported families).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio ControlNet: show the picker on the Create tab (workflow id is 'create', not 'txt2img')
The ControlNet control gated on workflow === 'txt2img', but the Images workflow tab ids are create/transform/inpaint/extend/upscale/reference/edit -- there is no 'txt2img'. So the picker never rendered even with a ControlNet-capable model loaded. Gate on 'create' (the text-to-image tab) for both the picker and the request wiring. Found via a live Playwright capture of the running Studio.
* Studio: do not force diffusers pipelines cross-tagged gguf into the GGUF variant expander
Some diffusers image repos (e.g. unsloth/Qwen-Image-2512-unsloth-bnb-4bit) carry a
stray "gguf" tag on the Hub but ship no .gguf files. The model search classified
them as GGUF from the bare tag, so the picker rendered the GGUF variant expander,
which then dead-ended at "No GGUF variants found." Trust the bare gguf tag only when
the repo is not a diffusers pipeline; the -GGUF name suffix and real gguf metadata
(populated via expand=gguf) remain authoritative, so genuine GGUF repos are unaffected.
* Studio Images: load non-curated unsloth/on-device diffusers repos instead of no-op
handleModelSelect only loaded curated safetensors ids and GGUF variant picks; any other
non-GGUF pick (an on-device diffusers folder, or a future unsloth diffusers image repo
surfaced by search) silently did nothing. Treat such a pick as a full diffusers pipeline
load when the id is unsloth-hosted or on-device (the backend infers the family + base repo
and gates loads to unsloth/* or local paths), and show a clear message otherwise instead
of silently ignoring the click. Curated and GGUF paths are unchanged.
* Studio Images: keep curated safetensors models in Recommended after download
The curated bnb-4bit / fp8 diffusion rows were filtered out of the Images picker's
Recommended list once cached (curatedSafetensorsRows dropped anything in downloadedSet),
so they vanished from the picker after the first load and could only be found by typing an
exact search. The row already renders a downloaded badge, matching how GGUF Recommended
rows stay visible when cached. Drop the exclusion so the curated safetensors always list.
* Studio diffusion LoRA: sanitize dots out of adapter aliases
The LoRA alias is used as the diffusers PEFT adapter name, and PEFT rejects names
containing "." (module name can't contain "."). sanitize_alias kept dots, so a LoRA whose
filename carries a version tag (e.g. Qwen-Image-2512-Lightning-8steps-V1.0-bf16) failed to
apply with a 400. Replace dots too; the alias stays a valid native <lora:NAME:w> filename
stem. Adds regression coverage for internal dots.
* Studio Images: clarify the GGUF transformer-quant Advanced control
Renamed the confusing "Transformer quant / GGUF default" control to "GGUF speed mode"
with an "Off (run the GGUF)" default, and reworded the hint to state plainly that FP8/INT8/
FP4 load the FULL base model (larger download + more VRAM) rather than re-packing the GGUF,
falling back to the GGUF if it can't fit. Behavior unchanged; labels/hint only.
* Studio Images: list on-device unsloth diffusion models in the picker
The Images picker's On Device tab hid every non-GGUF cached repo whenever a
task filter was active, so downloaded unsloth diffusion pipelines (bnb-4bit
and FP8 safetensors) never showed up there. List cached repos that pass the
task gate, limited under a filter to unsloth-hosted ones so base repos (which
fail the diffusion load trust gate) don't appear only to dead-end on click.
Chat behavior is unchanged: the task gate still drops image repos there.
* Studio: hide single-file image checkpoints from the chat model picker
The chat picker treats a cached repo as an image model, and hides it, only
when it ships a diffusers model_index.json. Single-file, ComfyUI, and
ControlNet image checkpoints (an FP8 Qwen-Image, a z-image safetensors, a
Qwen-Image ControlNet) carry none, so they surfaced as loadable chat models.
Fall back to resolving the repo id against the known diffusion families, the
same resolver the Images backend loads from, so these checkpoints are tagged
text-to-image and stay in the Images picker only.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: add the FLUX.2-dev model family
Loading unsloth/FLUX.2-dev-GGUF failed because detect_family knew only the
Qwen3-based FLUX.2-klein, so FLUX.2-dev (the full, Mistral-based Flux2Pipeline)
resolved to nothing and the load errored. Add a flux.2-dev family: Flux2Pipeline
+ Flux2Transformer2DModel over the black-forest-labs/FLUX.2-dev base repo (gated,
reachable with an HF token), with its FLUX.2 32-channel VAE and Mistral text
encoder wired for the sd-cli path from the open Comfy-Org/flux2-dev mirror.
text-to-image only: diffusers 0.38 ships no Flux2 img2img / inpaint pipeline for
dev. Frontend gets sensible dev defaults (28 steps, guidance 4), distinct from
klein's turbo defaults. Verified live: GGUF load resolves the family + gated base
repo and generates a real 1024x1024 image on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: clearer error for an unsupported diffusion model
When a repo id resolves to no diffusion family the load raised 'Could not infer a
diffusion family... Pass family_override (z-image)', which points at an unrelated
family and doesn't say what is supported. Replace it with a message that lists the
supported families (from a new supported_family_names helper) and notes that video
models and image models whose diffusers transformer has no single-file loader are
not supported. Applies to both the diffusers and native sd.cpp load paths. Also
refreshes two stale family-registry comments that still called FLUX.2-dev omitted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async task scratch outputs committed by mistake
* ControlNet: reject filesystem-like ids and do not cache a model past an unload race
Two review findings on the ControlNet path:
- resolve_controlnet's bare-repo fallback accepted any id with a slash, so a
path-shaped id (/tmp/x, ../x) reached from_pretrained as a local directory.
Restrict the fallback to a strict owner/name HF repo id shape.
- _controlnet_pipe now re-checks the cancel event after the blocking
from_pretrained: an unload that raced the download had already cleared the
caches, so caching the late module would pin it past the unload.
* ControlNet: address review findings on the diffusers path
- resolve_controlnet enforces catalog family compatibility so a direct API call
cannot load a ControlNet built for another family through the wrong pipeline.
- Unknown ControlNet ids now surface as a 400 (call site maps FileNotFoundError
to ValueError) instead of a generic 500.
- strength 0 disables ControlNet entirely, so a no-op selection never pays the
download / VRAM cost; the control image is decoded and validated BEFORE the
ControlNet is resolved or built, so a malformed image fails fast for the same reason.
- ControlNet loads use the base compute dtype (state.dtype is a display string,
not a torch.dtype, so it silently fell back to float32) and honor the base
offload policy via group offloading instead of forcing the module resident.
- Empty/malformed HF token coerced to anonymous access.
- Flux Union ControlNet control_mode mapped from the selected control type.
- resolve_controlnet drops the unused hf_token/cancel_event params.
- ControlNetSpec validates guidance_start <= guidance_end (clean 422).
- Images UI ControlNet Select shows its placeholder when nothing is selected.
Adds regression tests for family enforcement and the union control-mode map.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Diffusion LoRA: harden resolution, native tag precedence, and diffusers teardown
Address review findings on the LoRA path:
- resolve_one: normalise a blank/whitespace hf_token to None (anonymous access)
and reject a client-supplied weight file with traversal / absolute path.
- resolve_specs: convert FileNotFoundError from an unknown/stale id to ValueError
so the route returns 400 instead of a generic 500.
- _scan_local: disambiguate local adapters that share a stem (foo.safetensors vs
foo.gguf) so each is uniquely addressable.
- inject_prompt_tags: the backend-validated weight now wins over a user-typed
<lora:ALIAS:...> for a selected adapter; unselected user tags are left alone.
- diffusers _apply_loras: reject a .gguf adapter with a clear error before touching
the pipe (diffusers loads safetensors only).
- _unload_locked: drop the explicit unload_lora_weights() on teardown; the pipe is
dropped wholesale (freeing adapters), so the previous call could race an in-flight
denoise on the same pipe.
- Images page: use a stable LoRA key and clear the selection (not just the options)
when the catalog refresh fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Diffusion: guard trust check against OSError and validate conditioning inputs
- _is_trusted_diffusion_repo: wrap Path.exists() so a repo id with invalid
characters (or a bare owner/name id) can't raise OSError; treat any failure as
not-a-local-path and fall through to the unsloth/ allowlist. validate_load_request
still raises the clear FileNotFoundError for a genuinely missing local pick.
- generate(): reject mask_image / upscale / reference_images supplied without an
input image, and reject reference_images on a family that does not support
reference conditioning, instead of silently degrading to txt2img / img2img.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review findings on the image-workflows PR
Keep diffusion.py importable without torch: the compile/arch patch modules
import torch at module level, so import them lazily at their load/unload
call sites instead of at module load. This restores the torchless contract
so get_diffusion_backend() works on a CPU/native sd.cpp install.
Match family reject keywords and aliases as whole path/name segments, not
raw substrings, so an unrelated word like edited, edition, or kontextual no
longer misroutes or hides a valid base image model, while supported edit
families (Qwen-Image-Edit, FLUX Kontext) still resolve. Mirror the same
segment matching in the picker task filter.
Route FLUX.2-dev native guidance through --guidance like the other FLUX
families rather than --cfg-scale. Reject native upscale requests that have
no input image. Read image header dimensions and reject over-limit inputs
before decoding pixels, so a crafted small-payload image cannot spike
memory. Reject an upscale that would shrink the source below its input
size. Validate the model_kind against the filename extension before the
GPU handoff. Estimate a local diffusers pipeline's size from its on-disk
weights so auto memory planning does not skip offload and OOM. Report
workflows: [txt2img] from the native backend status so the Create tab
stays enabled for a loaded native model. Clamp the outpaint canvas to the
backend's 4096px decode limit.
Adds regression tests for segment matching and kind/extension validation.
* Harden diffusion LoRA handling on the diffusers and native paths
Reject LoRA on a torch.compile'd diffusers transformer (Speed=default/max):
diffusers requires the adapter loaded before compilation, so applying one to
the already-compiled module fails with adapter-key mismatches. The status
gate now hides the picker and generate raises a clear message instead.
Convert a cancelled Hub LoRA download (RuntimeError Cancelled) to the
diffusion cancellation sentinel in resolve_specs, so an unload/superseding
load during resolution maps to a 409 instead of a generic server error.
Drop weight-0 LoRA rows before the native support gate so a request carrying
only disabled adapters stays a no-op on families where native LoRA is
unsupported, matching the diffusers path.
Reject duplicate LoRA ids in the request model: both apply paths suffix
colliding names, so a repeated id would stack the same adapter past its
per-adapter weight bound.
Strip all user-typed <lora:...> prompt tags on the native path (only the
selected adapters are materialized in the managed lora-model-dir, so an
unselected tag can never resolve), and restore saved LoRA selections from a
gallery recipe so restore reproduces a LoRA image.
* Harden ControlNet resolve, gallery metadata, and the control-type picker
Check cancellation immediately after a ControlNet from_pretrained and before
any device placement, so an unload/eviction that raced the download does not
allocate several GB onto the GPU after the load was already cleared.
Require a loadable weight or shard index (not just config.json) before a local
ControlNet folder is advertised, so an interrupted copy is hidden instead of
failing deep in from_pretrained as a generic 500.
Do not record a strength-0 ControlNet in the gallery recipe: it is treated as
disabled and skipped, so the image is unconditioned and the metadata must not
claim a ControlNet was applied.
Build the control-type picker from the selected ControlNet's advertised
control_types instead of a hardcoded passthrough/canny pair, so a union model
with a precomputed depth or pose map sends the correct control_mode.
* Address further Codex findings on the image-workflows PR
- Persist the actual output image size in the gallery recipe instead of the
request sliders: Transform/Inpaint/Edit derive the size from the uploaded
image, Extend grows the canvas, and Upscale resizes it, so the sliders
recorded (and later restored) the wrong dimensions for those workflows.
- Reject a remote '*-GGUF' repo loaded as a full pipeline (no single-file
name) in validate_load_request, so the unloadable pick fails before chat is
evicted rather than deep in from_pretrained.
- Only publish an image-conditioned from_pipe wrapper to the shared aux cache
when the load is still current: from_pipe runs under the generate lock but
not the state lock, so an unload racing its construction could otherwise
cache a wrapper over torn-down modules that a later load would reuse.
- Verify the Windows CUDA runtime archive checksum before extracting it, like
the main sd-cli archive, so a corrupt or tampered runtime is rejected rather
than extracted next to the binary.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: eager patches + torch.compile cache speed phase
Adds the opt-in speed path for the GGUF diffusion transformer behind a
selectable speed mode (default off, so output is unchanged until a profile
is chosen):
- diffusion_eager_patches.py: shared eager fast-paths (channels_last,
attention/backend selection, fused norms and QKV) installed at load and
rolled back on unload or failed load.
- diffusion_compile_cache.py / diffusion_gguf_compile.py: a persistent
torch.compile cache and the GGUF-transformer compile wiring.
- diffusion_arch_patches.py: architecture-specific patches.
- diffusion_patch_backend.py: shared install/restore plumbing.
- diffusion_speed.py: speed-profile planning.
Tests for each module plus the benchmarking and probe scripts used to
measure speed, memory, and accuracy of the path.
* Studio diffusion: image workflows (safetensors, image-conditioned, editing) + Images UI
Backend:
- Load non-GGUF safetensors models: full bnb-4bit pipelines and single-file
fp8 transformers, gated to the unsloth org plus a curated allowlist.
- Image-conditioned workflows built with Pipeline.from_pipe so they reuse the
loaded transformer/VAE/text-encoder with no extra VRAM: img2img, inpaint,
outpaint, and a hires-fix upscale pass.
- Instruction editing as its own family kind (Qwen-Image-Edit-2511,
FLUX.1-Kontext-dev) and FLUX.2-klein reference conditioning (single and
multi-reference) plus klein inpaint.
- Auto-resize odd-sized inputs to a multiple of 16 (and resize the matched
mask) so img2img/inpaint/edit no longer reject non-/16 uploads. Bound the
decoded image size and cap upscale output to avoid OOM on large inputs.
- Fixes: from_pipe defaulting to a float32 recast that crashed torchao
quantized transformers; image-conditioned calls forcing the slider size
onto the input image. Native sd.cpp engine rejects image-conditioned and
reference requests it cannot serve.
Frontend:
- Redesigned Images page with capability-gated workflow tabs (Create,
Transform, Inpaint, Extend, Upscale, Reference, Edit), a brush mask editor,
client-side outpaint, and a multi-reference picker.
- Advanced options moved to a right-docked panel mirroring Chat: closed by
default, toggled by a single fixed top-bar button that stays in place.
sd.cpp installer: pin the release, verify each download's sha256, add a
download timeout, and make the source repo configurable for a future mirror.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: correct the Advanced panel comment (closed by default, fixed toggle)
* Studio diffusion: LoRA adapters for the Images workflow
Add community LoRA support across both diffusion backends, the single
biggest step toward broad image-workflow coverage.
Backend
- New shared module core/inference/diffusion_lora.py: adapter discovery
(local scan + curated catalog + owner/name[:file] Hub refs), download
via hf_hub_download_with_xet_fallback, alias sanitization, native
managed-dir materialization with collision-broken aliases, prompt-tag
injection (deduped against user-typed tags), and a supports_lora gate.
- Native sd-cli: resolve + materialize selected LoRAs into a per-run
managed dir, inject <lora:ALIAS:w> tags, pass --lora-model-dir with
--lora-apply-mode auto. The arg builder already emitted these flags.
- Diffusers: non-fused load_lora_weights + set_adapters manager, tracked
on the pipe so an unchanged selection is a no-op and a model swap
resets; cleared on unload. Never fuses (breaks quantized transformers
and blocks live weight tweaks).
- Gated off where unsupported: torchao fp8/int8 dense, GGUF-via-diffusers,
and native Qwen-Image (no LoRA name-conversion branch upstream).
- Request contract: optional loras on DiffusionGenerateRequest; empty or
omitted is identical to today. supports_lora surfaced in status; chosen
LoRAs persisted in gallery recipe metadata.
- New GET /api/models/diffusion-loras for the picker (family-filtered).
Frontend
- Repeatable multi-LoRA picker (adapter select + weight slider 0..2 +
remove), gated by the loaded model's supports_lora and family, max 8.
Tests
- New test_diffusion_lora.py (14): helpers, request validation, native
tag/dir wiring, diffusers set_adapters manager, supports_lora matrix.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: do not force diffusers pipelines cross-tagged gguf into the GGUF variant expander
Some diffusers image repos (e.g. unsloth/Qwen-Image-2512-unsloth-bnb-4bit) carry a
stray "gguf" tag on the Hub but ship no .gguf files. The model search classified
them as GGUF from the bare tag, so the picker rendered the GGUF variant expander,
which then dead-ended at "No GGUF variants found." Trust the bare gguf tag only when
the repo is not a diffusers pipeline; the -GGUF name suffix and real gguf metadata
(populated via expand=gguf) remain authoritative, so genuine GGUF repos are unaffected.
* Studio Images: load non-curated unsloth/on-device diffusers repos instead of no-op
handleModelSelect only loaded curated safetensors ids and GGUF variant picks; any other
non-GGUF pick (an on-device diffusers folder, or a future unsloth diffusers image repo
surfaced by search) silently did nothing. Treat such a pick as a full diffusers pipeline
load when the id is unsloth-hosted or on-device (the backend infers the family + base repo
and gates loads to unsloth/* or local paths), and show a clear message otherwise instead
of silently ignoring the click. Curated and GGUF paths are unchanged.
* Studio Images: keep curated safetensors models in Recommended after download
The curated bnb-4bit / fp8 diffusion rows were filtered out of the Images picker's
Recommended list once cached (curatedSafetensorsRows dropped anything in downloadedSet),
so they vanished from the picker after the first load and could only be found by typing an
exact search. The row already renders a downloaded badge, matching how GGUF Recommended
rows stay visible when cached. Drop the exclusion so the curated safetensors always list.
* Studio diffusion LoRA: sanitize dots out of adapter aliases
The LoRA alias is used as the diffusers PEFT adapter name, and PEFT rejects names
containing "." (module name can't contain "."). sanitize_alias kept dots, so a LoRA whose
filename carries a version tag (e.g. Qwen-Image-2512-Lightning-8steps-V1.0-bf16) failed to
apply with a 400. Replace dots too; the alias stays a valid native <lora:NAME:w> filename
stem. Adds regression coverage for internal dots.
* Studio Images: clarify the GGUF transformer-quant Advanced control
Renamed the confusing "Transformer quant / GGUF default" control to "GGUF speed mode"
with an "Off (run the GGUF)" default, and reworded the hint to state plainly that FP8/INT8/
FP4 load the FULL base model (larger download + more VRAM) rather than re-packing the GGUF,
falling back to the GGUF if it can't fit. Behavior unchanged; labels/hint only.
* Studio Images: list on-device unsloth diffusion models in the picker
The Images picker's On Device tab hid every non-GGUF cached repo whenever a
task filter was active, so downloaded unsloth diffusion pipelines (bnb-4bit
and FP8 safetensors) never showed up there. List cached repos that pass the
task gate, limited under a filter to unsloth-hosted ones so base repos (which
fail the diffusion load trust gate) don't appear only to dead-end on click.
Chat behavior is unchanged: the task gate still drops image repos there.
* Studio: hide single-file image checkpoints from the chat model picker
The chat picker treats a cached repo as an image model, and hides it, only
when it ships a diffusers model_index.json. Single-file, ComfyUI, and
ControlNet image checkpoints (an FP8 Qwen-Image, a z-image safetensors, a
Qwen-Image ControlNet) carry none, so they surfaced as loadable chat models.
Fall back to resolving the repo id against the known diffusion families, the
same resolver the Images backend loads from, so these checkpoints are tagged
text-to-image and stay in the Images picker only.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: add the FLUX.2-dev model family
Loading unsloth/FLUX.2-dev-GGUF failed because detect_family knew only the
Qwen3-based FLUX.2-klein, so FLUX.2-dev (the full, Mistral-based Flux2Pipeline)
resolved to nothing and the load errored. Add a flux.2-dev family: Flux2Pipeline
+ Flux2Transformer2DModel over the black-forest-labs/FLUX.2-dev base repo (gated,
reachable with an HF token), with its FLUX.2 32-channel VAE and Mistral text
encoder wired for the sd-cli path from the open Comfy-Org/flux2-dev mirror.
text-to-image only: diffusers 0.38 ships no Flux2 img2img / inpaint pipeline for
dev. Frontend gets sensible dev defaults (28 steps, guidance 4), distinct from
klein's turbo defaults. Verified live: GGUF load resolves the family + gated base
repo and generates a real 1024x1024 image on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: clearer error for an unsupported diffusion model
When a repo id resolves to no diffusion family the load raised 'Could not infer a
diffusion family... Pass family_override (z-image)', which points at an unrelated
family and doesn't say what is supported. Replace it with a message that lists the
supported families (from a new supported_family_names helper) and notes that video
models and image models whose diffusers transformer has no single-file loader are
not supported. Applies to both the diffusers and native sd.cpp load paths. Also
refreshes two stale family-registry comments that still called FLUX.2-dev omitted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async task scratch outputs committed by mistake
* Diffusion LoRA: harden resolution, native tag precedence, and diffusers teardown
Address review findings on the LoRA path:
- resolve_one: normalise a blank/whitespace hf_token to None (anonymous access)
and reject a client-supplied weight file with traversal / absolute path.
- resolve_specs: convert FileNotFoundError from an unknown/stale id to ValueError
so the route returns 400 instead of a generic 500.
- _scan_local: disambiguate local adapters that share a stem (foo.safetensors vs
foo.gguf) so each is uniquely addressable.
- inject_prompt_tags: the backend-validated weight now wins over a user-typed
<lora:ALIAS:...> for a selected adapter; unselected user tags are left alone.
- diffusers _apply_loras: reject a .gguf adapter with a clear error before touching
the pipe (diffusers loads safetensors only).
- _unload_locked: drop the explicit unload_lora_weights() on teardown; the pipe is
dropped wholesale (freeing adapters), so the previous call could race an in-flight
denoise on the same pipe.
- Images page: use a stable LoRA key and clear the selection (not just the options)
when the catalog refresh fails.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Diffusion: guard trust check against OSError and validate conditioning inputs
- _is_trusted_diffusion_repo: wrap Path.exists() so a repo id with invalid
characters (or a bare owner/name id) can't raise OSError; treat any failure as
not-a-local-path and fall through to the unsloth/ allowlist. validate_load_request
still raises the clear FileNotFoundError for a genuinely missing local pick.
- generate(): reject mask_image / upscale / reference_images supplied without an
input image, and reject reference_images on a family that does not support
reference conditioning, instead of silently degrading to txt2img / img2img.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review findings on the image-workflows PR
Keep diffusion.py importable without torch: the compile/arch patch modules
import torch at module level, so import them lazily at their load/unload
call sites instead of at module load. This restores the torchless contract
so get_diffusion_backend() works on a CPU/native sd.cpp install.
Match family reject keywords and aliases as whole path/name segments, not
raw substrings, so an unrelated word like edited, edition, or kontextual no
longer misroutes or hides a valid base image model, while supported edit
families (Qwen-Image-Edit, FLUX Kontext) still resolve. Mirror the same
segment matching in the picker task filter.
Route FLUX.2-dev native guidance through --guidance like the other FLUX
families rather than --cfg-scale. Reject native upscale requests that have
no input image. Read image header dimensions and reject over-limit inputs
before decoding pixels, so a crafted small-payload image cannot spike
memory. Reject an upscale that would shrink the source below its input
size. Validate the model_kind against the filename extension before the
GPU handoff. Estimate a local diffusers pipeline's size from its on-disk
weights so auto memory planning does not skip offload and OOM. Report
workflows: [txt2img] from the native backend status so the Create tab
stays enabled for a loaded native model. Clamp the outpaint canvas to the
backend's 4096px decode limit.
Adds regression tests for segment matching and kind/extension validation.
* Harden diffusion LoRA handling on the diffusers and native paths
Reject LoRA on a torch.compile'd diffusers transformer (Speed=default/max):
diffusers requires the adapter loaded before compilation, so applying one to
the already-compiled module fails with adapter-key mismatches. The status
gate now hides the picker and generate raises a clear message instead.
Convert a cancelled Hub LoRA download (RuntimeError Cancelled) to the
diffusion cancellation sentinel in resolve_specs, so an unload/superseding
load during resolution maps to a 409 instead of a generic server error.
Drop weight-0 LoRA rows before the native support gate so a request carrying
only disabled adapters stays a no-op on families where native LoRA is
unsupported, matching the diffusers path.
Reject duplicate LoRA ids in the request model: both apply paths suffix
colliding names, so a repeated id would stack the same adapter past its
per-adapter weight bound.
Strip all user-typed <lora:...> prompt tags on the native path (only the
selected adapters are materialized in the managed lora-model-dir, so an
unselected tag can never resolve), and restore saved LoRA selections from a
gallery recipe so restore reproduces a LoRA image.
* Address further Codex findings on the image-workflows PR
- Persist the actual output image size in the gallery recipe instead of the
request sliders: Transform/Inpaint/Edit derive the size from the uploaded
image, Extend grows the canvas, and Upscale resizes it, so the sliders
recorded (and later restored) the wrong dimensions for those workflows.
- Reject a remote '*-GGUF' repo loaded as a full pipeline (no single-file
name) in validate_load_request, so the unloadable pick fails before chat is
evicted rather than deep in from_pretrained.
- Only publish an image-conditioned from_pipe wrapper to the shared aux cache
when the load is still current: from_pipe runs under the generate lock but
not the state lock, so an unload racing its construction could otherwise
cache a wrapper over torn-down modules that a later load would reuse.
- Verify the Windows CUDA runtime archive checksum before extracting it, like
the main sd-cli archive, so a corrupt or tampered runtime is rejected rather
than extracted next to the binary.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: eager patches + torch.compile cache speed phase
Adds the opt-in speed path for the GGUF diffusion transformer behind a
selectable speed mode (default off, so output is unchanged until a profile
is chosen):
- diffusion_eager_patches.py: shared eager fast-paths (channels_last,
attention/backend selection, fused norms and QKV) installed at load and
rolled back on unload or failed load.
- diffusion_compile_cache.py / diffusion_gguf_compile.py: a persistent
torch.compile cache and the GGUF-transformer compile wiring.
- diffusion_arch_patches.py: architecture-specific patches.
- diffusion_patch_backend.py: shared install/restore plumbing.
- diffusion_speed.py: speed-profile planning.
Tests for each module plus the benchmarking and probe scripts used to
measure speed, memory, and accuracy of the path.
* Studio diffusion: image workflows (safetensors, image-conditioned, editing) + Images UI
Backend:
- Load non-GGUF safetensors models: full bnb-4bit pipelines and single-file
fp8 transformers, gated to the unsloth org plus a curated allowlist.
- Image-conditioned workflows built with Pipeline.from_pipe so they reuse the
loaded transformer/VAE/text-encoder with no extra VRAM: img2img, inpaint,
outpaint, and a hires-fix upscale pass.
- Instruction editing as its own family kind (Qwen-Image-Edit-2511,
FLUX.1-Kontext-dev) and FLUX.2-klein reference conditioning (single and
multi-reference) plus klein inpaint.
- Auto-resize odd-sized inputs to a multiple of 16 (and resize the matched
mask) so img2img/inpaint/edit no longer reject non-/16 uploads. Bound the
decoded image size and cap upscale output to avoid OOM on large inputs.
- Fixes: from_pipe defaulting to a float32 recast that crashed torchao
quantized transformers; image-conditioned calls forcing the slider size
onto the input image. Native sd.cpp engine rejects image-conditioned and
reference requests it cannot serve.
Frontend:
- Redesigned Images page with capability-gated workflow tabs (Create,
Transform, Inpaint, Extend, Upscale, Reference, Edit), a brush mask editor,
client-side outpaint, and a multi-reference picker.
- Advanced options moved to a right-docked panel mirroring Chat: closed by
default, toggled by a single fixed top-bar button that stays in place.
sd.cpp installer: pin the release, verify each download's sha256, add a
download timeout, and make the source repo configurable for a future mirror.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: correct the Advanced panel comment (closed by default, fixed toggle)
* Studio: do not force diffusers pipelines cross-tagged gguf into the GGUF variant expander
Some diffusers image repos (e.g. unsloth/Qwen-Image-2512-unsloth-bnb-4bit) carry a
stray "gguf" tag on the Hub but ship no .gguf files. The model search classified
them as GGUF from the bare tag, so the picker rendered the GGUF variant expander,
which then dead-ended at "No GGUF variants found." Trust the bare gguf tag only when
the repo is not a diffusers pipeline; the -GGUF name suffix and real gguf metadata
(populated via expand=gguf) remain authoritative, so genuine GGUF repos are unaffected.
* Studio Images: load non-curated unsloth/on-device diffusers repos instead of no-op
handleModelSelect only loaded curated safetensors ids and GGUF variant picks; any other
non-GGUF pick (an on-device diffusers folder, or a future unsloth diffusers image repo
surfaced by search) silently did nothing. Treat such a pick as a full diffusers pipeline
load when the id is unsloth-hosted or on-device (the backend infers the family + base repo
and gates loads to unsloth/* or local paths), and show a clear message otherwise instead
of silently ignoring the click. Curated and GGUF paths are unchanged.
* Studio Images: keep curated safetensors models in Recommended after download
The curated bnb-4bit / fp8 diffusion rows were filtered out of the Images picker's
Recommended list once cached (curatedSafetensorsRows dropped anything in downloadedSet),
so they vanished from the picker after the first load and could only be found by typing an
exact search. The row already renders a downloaded badge, matching how GGUF Recommended
rows stay visible when cached. Drop the exclusion so the curated safetensors always list.
* Studio Images: clarify the GGUF transformer-quant Advanced control
Renamed the confusing "Transformer quant / GGUF default" control to "GGUF speed mode"
with an "Off (run the GGUF)" default, and reworded the hint to state plainly that FP8/INT8/
FP4 load the FULL base model (larger download + more VRAM) rather than re-packing the GGUF,
falling back to the GGUF if it can't fit. Behavior unchanged; labels/hint only.
* Studio Images: list on-device unsloth diffusion models in the picker
The Images picker's On Device tab hid every non-GGUF cached repo whenever a
task filter was active, so downloaded unsloth diffusion pipelines (bnb-4bit
and FP8 safetensors) never showed up there. List cached repos that pass the
task gate, limited under a filter to unsloth-hosted ones so base repos (which
fail the diffusion load trust gate) don't appear only to dead-end on click.
Chat behavior is unchanged: the task gate still drops image repos there.
* Studio: hide single-file image checkpoints from the chat model picker
The chat picker treats a cached repo as an image model, and hides it, only
when it ships a diffusers model_index.json. Single-file, ComfyUI, and
ControlNet image checkpoints (an FP8 Qwen-Image, a z-image safetensors, a
Qwen-Image ControlNet) carry none, so they surfaced as loadable chat models.
Fall back to resolving the repo id against the known diffusion families, the
same resolver the Images backend loads from, so these checkpoints are tagged
text-to-image and stay in the Images picker only.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: add the FLUX.2-dev model family
Loading unsloth/FLUX.2-dev-GGUF failed because detect_family knew only the
Qwen3-based FLUX.2-klein, so FLUX.2-dev (the full, Mistral-based Flux2Pipeline)
resolved to nothing and the load errored. Add a flux.2-dev family: Flux2Pipeline
+ Flux2Transformer2DModel over the black-forest-labs/FLUX.2-dev base repo (gated,
reachable with an HF token), with its FLUX.2 32-channel VAE and Mistral text
encoder wired for the sd-cli path from the open Comfy-Org/flux2-dev mirror.
text-to-image only: diffusers 0.38 ships no Flux2 img2img / inpaint pipeline for
dev. Frontend gets sensible dev defaults (28 steps, guidance 4), distinct from
klein's turbo defaults. Verified live: GGUF load resolves the family + gated base
repo and generates a real 1024x1024 image on GPU.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio Images: clearer error for an unsupported diffusion model
When a repo id resolves to no diffusion family the load raised 'Could not infer a
diffusion family... Pass family_override (z-image)', which points at an unrelated
family and doesn't say what is supported. Replace it with a message that lists the
supported families (from a new supported_family_names helper) and notes that video
models and image models whose diffusers transformer has no single-file loader are
not supported. Applies to both the diffusers and native sd.cpp load paths. Also
refreshes two stale family-registry comments that still called FLUX.2-dev omitted.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Remove stray async task scratch outputs committed by mistake
* Diffusion: guard trust check against OSError and validate conditioning inputs
- _is_trusted_diffusion_repo: wrap Path.exists() so a repo id with invalid
characters (or a bare owner/name id) can't raise OSError; treat any failure as
not-a-local-path and fall through to the unsloth/ allowlist. validate_load_request
still raises the clear FileNotFoundError for a genuinely missing local pick.
- generate(): reject mask_image / upscale / reference_images supplied without an
input image, and reject reference_images on a family that does not support
reference conditioning, instead of silently degrading to txt2img / img2img.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address Codex review findings on the image-workflows PR
Keep diffusion.py importable without torch: the compile/arch patch modules
import torch at module level, so import them lazily at their load/unload
call sites instead of at module load. This restores the torchless contract
so get_diffusion_backend() works on a CPU/native sd.cpp install.
Match family reject keywords and aliases as whole path/name segments, not
raw substrings, so an unrelated word like edited, edition, or kontextual no
longer misroutes or hides a valid base image model, while supported edit
families (Qwen-Image-Edit, FLUX Kontext) still resolve. Mirror the same
segment matching in the picker task filter.
Route FLUX.2-dev native guidance through --guidance like the other FLUX
families rather than --cfg-scale. Reject native upscale requests that have
no input image. Read image header dimensions and reject over-limit inputs
before decoding pixels, so a crafted small-payload image cannot spike
memory. Reject an upscale that would shrink the source below its input
size. Validate the model_kind against the filename extension before the
GPU handoff. Estimate a local diffusers pipeline's size from its on-disk
weights so auto memory planning does not skip offload and OOM. Report
workflows: [txt2img] from the native backend status so the Create tab
stays enabled for a loaded native model. Clamp the outpaint canvas to the
backend's 4096px decode limit.
Adds regression tests for segment matching and kind/extension validation.
* Address further Codex findings on the image-workflows PR
- Persist the actual output image size in the gallery recipe instead of the
request sliders: Transform/Inpaint/Edit derive the size from the uploaded
image, Extend grows the canvas, and Upscale resizes it, so the sliders
recorded (and later restored) the wrong dimensions for those workflows.
- Reject a remote '*-GGUF' repo loaded as a full pipeline (no single-file
name) in validate_load_request, so the unloadable pick fails before chat is
evicted rather than deep in from_pretrained.
- Only publish an image-conditioned from_pipe wrapper to the shared aux cache
when the load is still current: from_pipe runs under the generate lock but
not the state lock, so an unload racing its construction could otherwise
cache a wrapper over torn-down modules that a later load would reuse.
- Verify the Windows CUDA runtime archive checksum before extracting it, like
the main sd-cli archive, so a corrupt or tampered runtime is rejected rather
than extracted next to the binary.
* [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: oobabooga <112222186+oobabooga@users.noreply.github.com>
* Studio diffusion: cross-platform device policy, fp16 guard, lock split, validate-before-evict
Phase 1 of porting the richer diffusion stack onto the image-generation backend.
- Add a compartmentalized device/dtype policy module (diffusion_device.py)
resolving CUDA/ROCm/XPU/MPS/CPU with capability flags. Keeps the NVIDIA
capability-based bf16 choice; ROCm and XPU are isolated; MPS uses bf16 or
fp32, never a silent fp16 that renders a black image.
- Add a per-family fp16_incompatible flag (Z-Image) and promote a resolved
float16 to float32 for those families so they do not produce black images.
- Split the backend locks: a generation holds only _generate_lock, so status,
unload, and a new load are never blocked by a long denoise. Add per-generation
cancellation via callback_on_step_end so an eviction or a superseding load
preempts a running generation; a replacement load waits for it to stop before
allocating, so two pipelines never sit in VRAM at once.
- Validate a load request before the GPU handoff so an unloadable pick never
evicts a working chat model, and reject missing local paths up front.
- Add CPU-only tests for the device policy, dtype guard, lock split and
cancellation, and validate-before-evict, plus a GPU benchmark/regression
script (scripts/diffusion_bench.py) measuring latency, peak VRAM, and PSNR
against a saved reference.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2A): measured-budget memory planner + offload/VAE policy
Add a lean, backend-agnostic memory policy that picks a CPU-offload policy and
VAE tiling/slicing from measured free device memory vs the model's estimated
resident footprint, then applies it to the built pipeline. auto stays resident
when the model fits (byte-identical to the prior resident path), and falls to
whole-module offload when tight; fast/balanced/low_vram are explicit overrides.
Sequential submodule offload is unreliable for GGUF transformers on diffusers
0.38, so it falls back to whole-module offload and status reports the policy
actually engaged.
Verified on Z-Image-Turbo Q4_K_M (B200): auto reproduces the resident image with
no VRAM/latency regression (PSNR inf); balanced/low_vram cut generation peak VRAM
47.9% (15951 -> 8318 MB) with byte-identical output, at the expected latency cost.
73 prior + 35 new CPU tests pass.
* Studio diffusion (Phase 2D): streamed block-level offload + functional VAE tiling
Add a streamed 'group' offload tier (diffusers apply_group_offloading, block_level,
use_stream) that keeps the transformer flowing through the GPU a few blocks at a
time while the text encoder / VAE stay resident, and fix VAE tiling to drive the
VAE submodule (pipelines like Z-Image expose enable_tiling on pipe.vae, not the
pipeline). apply_memory_plan now returns the (policy, tiling) actually engaged so
status never overstates either, and group falls back to whole-module offload when
the transformer can't be streamed.
Measured on Z-Image (B200), all lossless (PSNR inf vs resident): balanced/group
cuts generation peak VRAM 32% (15951 -> 10840 MB) at near-resident speed (2.07 ->
2.99s); low_vram/model cuts it 48% (-> 8318 MB) but is slower (7.99s). Mode names
now match that tradeoff: balanced = stream the transformer, low_vram = offload
every component. auto picks group when the companions fit resident, else model.
112 CPU tests pass.
* Studio diffusion (Phase 5): image quality-vs-quant accuracy harness
Add scripts/diffusion_quality.py, the accuracy analogue of the KLD workflow: hold
prompt + seed fixed, render a grid with a reference quant (default BF16), then render
each candidate quant and measure drift from the reference. Records mean PSNR + SSIM
(pure-numpy, no skimage/scipy) and optional CLIP text-alignment + image-similarity
(transformers, --clip), plus file size, latency, and peak VRAM, then prints a
quality-vs-cost table and recommends the smallest quant within a quality budget.
--selftest validates the metrics on synthetic images with no GPU or model.
Verified on Z-Image (B200): the table degrades monotonically with quant size
(Q8 -> Q4 -> Q2: PSNR 21.7 -> 15.5, SSIM 0.82 -> 0.61), while CLIP-text stays flat
(~0.34) -- quantization erodes fine detail far more than prompt adherence.
* Studio diffusion (Phase 3): opt-in speed layer (channels_last / compile / TF32)
Add a speed_mode knob (off by default, so the render path stays bit-identical):
default applies channels_last VAE + regional torch.compile of the denoiser's
repeated block where eligible; max also enables TF32 matmul and fused QKV. Regional
compile is gated off for the GGUF transformer (dequantises per-op) and for families
flagged not compile-friendly (a new supports_torch_compile flag, False for Z-Image),
so it activates automatically only once a non-GGUF bf16 transformer is loaded. Speed
optims run before placement/offload, per the diffusers composition order. status now
reports speed_mode + the optims actually engaged.
Verified on Z-Image (B200): default -> ['channels_last'], max -> ['channels_last',
'tf32'], compile correctly skipped for GGUF; generation works in every mode.
121 CPU tests pass.
* Studio diffusion (Phase 2B): opt-in fp8 text-encoder layerwise casting
Add a text_encoder_fp8 knob that casts the companion text encoder(s) to fp8 (e4m3)
storage via diffusers apply_layerwise_casting, upcasting per layer to the bf16
compute dtype while normalisations and embeddings stay full precision. Applied
before placement, gated to CUDA + bf16, best-effort (a failure leaves the encoder
dense). status reports which encoders were cast.
Verified on Z-Image (B200, balanced/group mode where the encoder stays resident):
generation peak VRAM dropped 37% (10840 -> 6791 MB, below the lowest-VRAM offload)
at near-resident speed. It is a memory-vs-quality tradeoff, not free -- ~20 dB PSNR
vs the bf16 encoder, a larger shift than one transformer quant step -- so it is off
by default and documented as such, with the Phase 5 harness to size the cost.
127 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 2C): NVFP4 text-encoder quant (+ generalise fp8 knob)
Generalise the text-encoder precision knob from a fp8 bool to text_encoder_quant
(fp8 | nvfp4). nvfp4 quantises the companion text encoder to 4-bit via torchao
NVFP4 weight-only (two-level microscaling) on Blackwell's FP4 tensor cores; fp8
stays the broader-hardware path (cc>=8.9). Both are gated, best-effort, and run
before placement; status reports the mode actually engaged. This is the lean
realisation of GGUF-native text-encoder quant: 4-bit on the encoder without the
3045-line port.
Verified on Z-Image (B200, balanced/group where the encoder stays resident), vs the
bf16 encoder: nvfp4 cut generation peak VRAM 48% (10840 -> 5593 MB, the lowest TE
option, below whole-model offload) at near-fp8 quality (16.4 vs 17.1 dB PSNR), and
both quants ran faster than bf16. A memory-vs-quality tradeoff (off by default);
size it per model with the Phase 5 quality harness. diffusion_bench gains
--text-encoder-quant.
129 CPU tests pass.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): native stable-diffusion.cpp engine for CPU/Mac
Adds the CPU / Apple-Silicon tier of the two-engine strategy, mirroring the
chat backend's llama.cpp shell-out. Diffusers stays the default on CUDA / ROCm
/ XPU; this covers the hardware diffusers serves poorly, consuming the same
split GGUF assets Studio already curates.
- sd_cpp_args.py: pure sd-cli command builder. Maps the family to its
text-encoder flag (Z-Image Qwen3 to --llm, Qwen-Image to --qwen2vl, FLUX.1
CLIP-L + T5), and the diffusers memory policy (none/group/model/sequential)
to sd.cpp's offload flags (--offload-to-cpu / --clip-on-cpu / --vae-on-cpu /
--vae-tiling / --diffusion-fa), so one user knob drives both engines.
- sd_cpp_engine.py: SdCppEngine over a located sd-cli. find_sd_cpp_binary()
with the same precedence as the llama finder (env override, then the Studio
install root, then in-tree, then PATH), an is_available/version probe, and a
one-shot subprocess generate that streams progress and returns the PNG.
runtime_env() prepends the binary's directory to the platform library path
so a prebuilt's bundled libstable-diffusion.so resolves.
select_diffusion_engine() is the pure routing decision (GPU backends to
diffusers, CPU/MPS to native when present).
- install_sd_cpp_prebuilt.py: resolve + download the per-host prebuilt
(macOS-arm64/Metal, Linux x86_64 CPU, Vulkan/ROCm/Windows variants) into the
Studio install root. resolve_release_asset() is a pure, unit-tested
host-to-asset matrix.
- scripts/sd_cpp_smoke.py: end-to-end native generation harness.
Tests (CPU-only, subprocess/filesystem stubbed): 49 new across args, engine,
routing, runtime env, and the installer resolver. Full diffusion suite 166
passing.
Verified on a B200 box: built sd-cli (CUDA) and the prebuilt (CPU) both
generate Z-Image-Turbo Q4_K end to end through SdCppEngine: balanced (group
offload, 5.0s gen), low_vram (full CPU offload + VAE tiling, 13.4s), and the
dynamically-linked CPU prebuilt (50.4s on CPU), all producing coherent images.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 6): img2img / inpaint / edit / LoRA / upscale on the native engine
Builds on Phase 4's native stable-diffusion.cpp engine, extending it from
text-to-image to the wider feature surface, since sd.cpp supports all of these
through the binary already. Pure command-builder additions plus one engine
method, so the txt2img path is unchanged.
- sd_cpp_args.py: SdCppGenParams gains image-conditioning fields. init_img +
strength make a run img2img, adding mask makes it inpaint, ref_images drives
FLUX-Kontext / Qwen-Image-Edit style editing (repeated --ref-image), and
lora_dir + the <lora:name:weight> prompt syntax select LoRAs. New
SdCppUpscaleParams + build_sd_cpp_upscale_command for the ESRGAN upscale run
mode (input image + esrgan model, no prompt / text encoders).
- sd_cpp_engine.py: the subprocess runner is factored into a shared _run() so
generate() (now carrying the conditioning flags) and a new upscale() reuse
the same streaming / error / output-check path.
- scripts/sd_cpp_smoke.py: --task {txt2img,img2img,upscale} with --init-img /
--strength / --upscale-model / --upscale-repeats.
Tests: 10 new across the img2img / inpaint / edit / LoRA flag construction, the
upscale builder and its validation, and the engine's img2img + upscale paths.
Full diffusion suite 176 passing.
Verified on a B200 box through SdCppEngine: img2img (Z-Image-Turbo Q4_K, the
init image conditioned at strength 0.6, 4.8s) and ESRGAN upscale
(512x512 -> 2048x2048 via RealESRGAN_x4plus_anime_6B, 2.7s), both producing
coherent images. Video and the diffusers-path feature wiring are deferred.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): accuracy-preserving speed pass
Re-review of the diffusion stack (#6675/#6679/#6680) surfaced one real accuracy
bug and a dead-on-arrival speed path; this fixes both and adds the lossless /
near-lossless wins, all measured on a B200.
Correctness:
- TF32 global-state leak (fix). speed_mode=max flipped torch.backends.*.allow_tf32
process-wide and never restored them, so a later `off` load silently inherited
TF32 and was no longer bit-identical. Added snapshot_backend_flags /
restore_backend_flags (TF32 + cudnn.benchmark), captured before the speed layer
runs and restored on unload. Verified: load max -> unload -> load off is now
byte-identical (PSNR inf) to a fresh off.
- sd-cli timeout could hang forever. _run() blocked in `for line in stdout` and
only checked the timeout after EOF, so a child stuck in model load / GPU init
with no output ignored the timeout. Drained stdout on a reader thread with a
wall-clock deadline. Added a silent-hang regression test.
Speed (diffusers path), near-lossless, opt-in tiers:
- Regional torch.compile now runs on the GGUF transformer. The is_gguf gate (and
Z-Image's supports_torch_compile=False) were stale: compile_repeated_blocks
compiles and runs ~2.2x faster on the GGUF Z-Image transformer on
torch 2.9.1 / diffusers 0.38 (the per-op dequant stays eager, the rest of the
block compiles). Measured: off 1.80s -> default 0.82s/gen (+54.7%), PSNR 37.7 dB
vs eager -- far above the Q4 quant noise floor (~21 dB), so it does not move
output quality. Gate relaxed; default tier delivers it.
- cudnn.benchmark added to the default tier (autotunes the fixed-shape VAE convs).
- torch.inference_mode() around the pipeline call (lossless, strictly faster than
the no_grad diffusers uses internally).
Memory path:
- VAE tiling (not bit-identical >1MP) restricted to the model/sequential/CPU tiers;
the balanced (group) tier keeps exact slicing only, so it is now bit-identical to
the resident image (verified PSNR inf) and slightly faster.
- Group offload adds non_blocking + record_stream on the CUDA stream path to
overlap each block's H2D copy with compute (lossless; gated on the installed
diffusers signature so older versions still work).
Native (sd.cpp) path:
- native_speed_flags: a first-class speed knob (default -> --diffusion-fa, a
near-lossless CUDA win that was previously only added on offload tiers; max also
-> --diffusion-conv-direct). conv-direct stays opt-in: measured +45% on CUDA, so
it is never auto-on. Engine generate() merges it, de-duped against offload flags.
Default profile: a GGUF model with no explicit speed_mode now resolves to the
`default` profile (resolve_speed_mode), since compile's perturbation sits below the
quantisation noise floor and so does not reduce quality versus the dense reference;
out of the box a GGUF Z-Image generation drops from 1.80s to 0.81s. Dense models
stay `off` / bit-identical, and an explicit speed_mode -- including "off" -- is
always honored, so the byte-identical path remains one flag away and is the
regression reference.
Tooling: scripts/compile_probe.py (eager vs compiled GGUF probe), scripts/
perf_verify.py (the B200 verification above), and diffusion_bench.py gains
--speed-mode so the speed tiers are benchmarkable.
Tests: 183 passing (was 166); new coverage for the backend-flag snapshot/restore,
GGUF compile eligibility, the balanced tiling/slicing split, native_speed_flags +
the engine de-dup, and the sd-cli silent-hang timeout.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7): max tier uses max-autotune-no-cudagraphs + engine/lever benchmarks
The opt-in `max` speed tier now compiles the repeated block with
mode=max-autotune-no-cudagraphs (dynamic=False) instead of the default mode:
Triton autotuning for GEMM/conv-heavier models, gated to the tier where a longer
cold compile is acceptable. CUDA-graph modes (reduce-overhead / max-autotune) are
deliberately avoided -- both crash on the regionally-compiled block (its static
output buffer is overwritten across denoise steps), measured.
Adds two reproducible benchmarks used to validate the optimization research:
- scripts/compare_engines.py: PyTorch (diffusers GGUF) vs native sd.cpp head-to-head.
- scripts/leverage_probe.py: coordinate_descent_tuning + FirstBlockCache probes.
Measured on B200 (Z-Image Q4_K_M, 1024px, 8 steps): default compile 0.80s/gen;
coordinate_descent_tuning 0.79s (within noise, already covered by max-autotune);
FirstBlockCache does not run on Z-Image (diffusers 0.38 block-detection / Dynamo).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): opt-in fast transformer (torchao int8/fp8/fp4 on a dense source)
Add an opt-in transformer_quant mode that loads the dense bf16 transformer and
torchao-quantises it onto the low-precision tensor cores, instead of the GGUF
transformer (which dequantises to bf16 per matmul and so runs at bf16 rate). On a
B200 (Z-Image-Turbo, 1024px/8 steps): auto picks fp8 at 0.614s vs GGUF+compile's
0.823s (1.34x), int8 0.626s (1.32x), both at lower LPIPS than GGUF's own 4-bit floor.
GGUF+compile stays the low-memory default and the fallback. The mode is gated on
CUDA + bf16 + resident VRAM headroom (the dense load peaks ~21GB vs GGUF's 13GB);
any unsupported arch/scheme, OOM, or quant failure falls back to GGUF with a logged
reason. auto picks the best scheme per GPU via a real quantise+matmul smoke probe
(Blackwell nvfp4/fp8/mxfp8, Ada/Hopper fp8, Ampere int8); a min-features filter skips
the tiny projections that crash int8's torch._int_mm. New module mirrors
diffusion_precision.py; quant runs before compile before placement.
184 -> tests pass; new test_diffusion_transformer_quant.py plus backend/route
coverage. scripts/diffusion_bench.py gains --transformer-quant; scripts/quant_probe.py
is the standalone torchao lever probe.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): consumer-GPU tuning - lock fp8 fast accumulate, prefer fp8 over mxfp8, reject 2:4 sparsity
Consumer Blackwell halves tensor-core throughput on FP32 accumulate (fp8 419 vs 838
TFLOPS with FP16 accumulate; bf16 209), so:
- fp8 config locks use_fast_accum=True (Float8MMConfig). torchao already defaults it on;
pinning it guards consumer cards against a default change. On B200 it is identical
speed and slightly better quality (LPIPS 0.050 vs 0.091).
- the Blackwell auto ladder prefers fp8 over mxfp8 (measured faster + more accurate).
2:4 semi-structured sparsity evaluated and rejected (scripts/sparse_accum_probe.py):
2:4 magnitude-prune + fp8 gives LPIPS 0.858 (broken image) with no fine-tune, the
cuSPARSELt kernel errors on torch 2.9, and it does not compose with torch.compile
(our main ~2x). Documented as a dead end, not shipped.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add fp8 fast-accum overflow verification probe
scripts/fp8_overflow_check.py hooks every quantised linear during a real Z-Image
generation and reports max-abs + non-finite counts for use_fast_accum True vs False.
Confirms fast accumulation is an accumulation-precision knob, not an overflow one:
across 276 linears, including Z-Image's ~1.0e6 activation peaks (which overflow FP16),
0 non-finite elements and identical max-abs for both modes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): detect consumer vs data-center GPU for fp8 accumulate, with user override
Consumer/workstation GPUs (GDDR) halve fp8 FP32-accumulate throughput, so they want
fast (FP16) accumulate; data-center HBM parts (B200/H100/A100/L40) are not nerfed and
prefer the higher-precision FP32 accumulate. Add _is_consumer_gpu() (token-exact match
on the device name per NVIDIA's GPU list, so workstation A4000 != data-center A40;
GeForce/TITAN and unknown default to consumer) and gate the fp8 use_fast_accum on it.
Measured: fast accumulate is ~2x on consumer Blackwell and ~8% on B200 (0.608 vs 0.665s),
no overflow, quality below the quant noise floor. So the default leans to accuracy on
data-center; a new request field transformer_quant_fast_accum (null=auto, true/false=force)
lets the operator override per load (scripts/diffusion_bench.py --fp8-fast-accum auto|on|off).
187 diffusion tests pass (+ consumer detection, _resolve_fast_accum, and the override
threading).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): add NVFP4 probe documenting it is not yet a win on torch 2.9
scripts/nvfp4_probe.py measures NVFP4 via torchao on the real Z-Image transformer.
Finding (B200, 1024px/8 steps): NVFP4 is a torchao feature and DOES run with
use_triton_kernel=False (the default triton path needs the missing MSLK library), but
only at bf16-compile rate (0.667s vs fp8 0.592s) -- it dequantises FP4->bf16 rather than
using the FP4 tensor cores. The real FP4 speedup needs MSLK or torch>=2.11 + torchao's
CUTLASS FP4 GEMM. The smoke probe (default triton=True) already keeps NVFP4 out of auto
on this env, so auto correctly stays on fp8; NVFP4 activates automatically once fast.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 8): prefer fp8 over nvfp4 in Blackwell auto ladder
Validated NVFP4 on torch 2.11 + torchao CUTLASS FP4 in an isolated env. The FP4
tensor-core GEMM is genuinely active there (a 16384^3 GEMM hits ~3826 TFLOPS,
2.52x bf16 and 1.37x fp8), but it only beats fp8 on very large GEMMs. At the
diffusion transformer's shapes (hidden ~3072, MLP ~12288, M~4096) NVFP4 is both
slower (0.81x fp8 end to end on Z-Image 1024px) and less accurate (LPIPS 0.166
vs fp8's 0.044). Reorder the Blackwell auto ladder to fp8 before nvfp4 so auto is
correct even on a future MSLK-equipped box; nvfp4 stays an explicit opt-in. Add
scripts/nvfp4_t211_probe.py (extension diagnostics + GEMM micro + end-to-end).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): pre-quantized transformer loading
The Phase 8 fast transformer_quant path materialises the dense bf16 transformer on
the GPU and torchao-quantises it in place, so its load peak is ~2x GGUF's (~21 vs
13.4 GB) plus a ~12 GB download. Add a pre-quantized branch: quantise once offline
(scripts/build_prequant_checkpoint.py) and at runtime build the transformer skeleton
on the meta device (accelerate.init_empty_weights) and load_state_dict(assign=True)
the quantized weights, so the dense bf16 never touches the GPU.
Measured (B200, Z-Image fp8): full-pipeline GPU load peak 21.2 -> 14.6 GB (matching
GGUF's 13.4), on-disk 12 -> 6.28 GB, output bit-identical (LPIPS 0.0). It is the same
torchao config + min_features filter the runtime path uses, applied ahead of time.
New core/inference/diffusion_prequant.py (resolve_prequant_source +
load_prequantized_transformer, best-effort, lazy imports). diffusion.py
_load_dense_quant_pipeline tries the pre-quant source first and falls back to the
dense materialise+quantise path, then to GGUF, so the default is unchanged.
DiffusionLoadRequest gains transformer_prequant_path; DiffusionFamily gains an empty
prequant_repos map for hosted checkpoints (hosting deferred). Hermetic CPU tests for
the resolver, the meta-init+assign loader, and the backend branch selection +
fallbacks; GPU verification via scripts/verify_prequant_backend.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): attention-backend selection
Add a selectable attention kernel via the diffusers set_attention_backend
dispatcher. Attention is memory-bandwidth bound, so a better kernel is an
end-to-end win orthogonal to the linear-weight quantisation (it speeds the QK/PV
matmuls torchao never touches) and composes with torch.compile.
auto picks the best exact backend for the device: cuDNN fused attention
(_native_cudnn) on NVIDIA when a speed profile is active, measured ~1.18x
end-to-end on a B200 (Z-Image 1024px/8 steps) with LPIPS ~0.004 vs the default
(below the compile/quant noise floor); native SDPA elsewhere and when speed=off
(so off stays bit-identical). Explicit native/cudnn/flash/flash3/flash4/sage/
xformers/aiter are honored, and an unavailable kernel falls back to the default
rather than failing the load.
New core/inference/diffusion_attention.py (normalize + per-device select + apply,
best-effort, lazy imports). Set on pipe.transformer BEFORE compile in load_pipeline;
attention_backend threads through begin_load / load_pipeline / status like the other
load knobs. New request field attention_backend + status field. Hermetic CPU tests
for normalize / select policy / apply fallback, plus route threading + 422. Measured
via scripts/perf_levers_probe.py.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 11): prefer int8 on consumer GPUs in the auto ladder
Consumer / workstation GPUs halve fp8 (and fp16/bf16) FP32-accumulate tensor-core
throughput, while int8 runs at full rate (int32 accumulate is not nerfed). Public
benchmarks (SDNQ across RTX 3090/4090/5090, AMD, Intel) confirm int8 via torch._int_mm
is as fast or faster than fp8 on every consumer part, and the only path on pre-Ada
consumer cards without fp8 tensor cores. So when transformer_quant=auto, reorder the
arch tier to put int8 first on a consumer/workstation GPU (detected by the existing
_is_consumer_gpu name heuristic), while data-center HBM parts keep fp8 first.
Pure ladder reorder via _prefer_consumer_scheme; no new flags. Verified non-regression
on a B200 (still picks fp8). Hermetic tests for consumer Blackwell/Ada/workstation
(-> int8) and data-center Ada/Hopper/Blackwell (-> fp8).
* Studio diffusion (Phase 12): First-Block-Cache step caching for many-step DiT
Add opt-in step caching (First-Block-Cache) for the diffusion transformer. Across
denoise steps a DiT's output settles, so once the first block's residual barely
changes the remaining blocks are skipped and their cached output reused. diffusers
ships it natively (FirstBlockCacheConfig + transformer.enable_cache, with the
standalone apply_first_block_cache hook as a fallback).
Measured on Flux.1-dev (28 steps, 1024px): ~1.4x on top of torch.compile (2.83 ->
2.03s) at LPIPS ~0.08 vs the no-cache output, well inside the quality bar.
OFF by default and a per-load opt-in: the win scales with step count, so it is for
many-step models (Flux / Qwen-Image) and pointless for few-step distilled models
(e.g. Z-Image-Turbo at ~8 steps), where a single skipped step is a large fraction
of the trajectory. It composes with regional compile only with fullgraph=False (the
cache's per-step decision is a torch.compiler.disable graph break), which the speed
layer now switches to automatically when a cache is engaged. Best-effort: a model
whose block signature the hook does not recognise is caught and the load proceeds
uncached.
- new core/inference/diffusion_cache.py: normalize_transformer_cache + apply_step_cache
(enable_cache / apply_first_block_cache fallback; threshold auto-raised for a
quantised transformer per ParaAttention's fp8 guidance; lazy diffusers import).
- diffusion_speed.py: apply_speed_optims takes cache_active; compile drops fullgraph
when a cache is engaged.
- diffusion.py: apply_step_cache before compile; thread transformer_cache /
transformer_cache_threshold through begin_load -> load_pipeline and report the
engaged mode in status().
- models/inference.py + routes/inference.py: transformer_cache (off | fbcache) and
transformer_cache_threshold request fields, engaged mode in the status response.
- hermetic tests for normalisation, the enable_cache / hook-fallback paths, threshold
selection, and best-effort failure handling, plus route threading + validation.
- scripts/fbcache_flux_probe.py: the Flux validation probe (latency / speedup / VRAM /
LPIPS vs the compiled no-cache baseline).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 14): fix int8 dense quant on Flux / Qwen (skip M=1 modulation linears)
The opt-in dense int8 transformer path crashed on Flux.1 and Qwen-Image with
'torch._int_mm: self.size(0) needs to be greater than 16, but got 1'. int8 dynamic quant
goes through torch._int_mm, which requires the activation row count M > 16. A DiT's AdaLN
modulation projections (Flux norm1.linear 3072->18432, Qwen img_mod.1 / txt_mod.1, Flux.2
*_modulation.linear) and its timestep / guidance / pooled-text conditioning embedders are
computed once from the [batch, dim] conditioning vector (M = batch = 1), not per token, so
they hit _int_mm at M=1 and crash. Their feature dims are large, so the existing
min_features filter did not exclude them.
Fix: the int8 filter now also skips any Linear whose fully-qualified name matches a
modulation / conditioning-embedder token (norm, _mod, modulation, timestep_embed,
guidance_embed, time_text_embed, pooled). These layers run at M=1 once per block and are a
negligible share of the FLOPs, so int8 keeps the full speedup on the attention / FFN layers
(M = sequence length). fp8 / nvfp4 / mxfp8 use scaled_mm, which has no M>16 limit and
quantises these layers fine, so the exclusion is int8-only. Sequence embedders
(context_embedder / x_embedder / txt_in, M = seq) are deliberately not excluded -- note
'context_embedder' contains the substring 'text_embed', which is why the token is the
specific 'time_text_embed', not 'text_embed'.
Measured on a B200 (1024px, transformer_quant=int8 + speed=default), int8 now runs on every
supported model and is the fastest dense path on Flux/Qwen (int8 runs full-rate vs fp8's
FP32-accumulate): FLUX.1-dev 9.62s eager -> 1.98s (4.86x, vs fp8 2.15s), Qwen-Image -> 1.87s
(5.57x, vs fp8 2.09s), FLUX.1-schnell -> 0.41s (3.59x). Z-Image and Flux.2-klein (already
working) are unchanged.
- diffusion_transformer_quant.py: add _INT8_EXCLUDE_NAME_TOKENS; make_filter_fn takes
exclude_name_tokens; quantize_transformer passes it for int8 only.
- hermetic test that the int8 filter excludes the modulation / embedder linears (and keeps
attention / FFN / sequence-embedder linears), while fp8 keeps them.
- scripts/int8_linear_probe.py: the meta-device probe used to enumerate each transformer's
Linear layers and derive the exclusion list.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 15): build int8 pre-quantized checkpoints (skip M=1 modulation linears)
The prequant-checkpoint builder applied the dense quant filter without the int8-only
M=1 modulation / conditioning-embedder exclusion the runtime path uses, so a built int8
checkpoint baked those projections as int8 and crashed (torch._int_mm needs M>16) at the
first denoise step on Flux / Qwen. Factor the scheme->exclusion decision into a shared
exclude_tokens_for_scheme() used by both the runtime quantise path and the offline builder
so they can never drift, and apply it in build_prequant_checkpoint.py. int8 prequant now
produces a working checkpoint on every supported model, giving int8 (the consumer-preferred
scheme) the same ~2x load-VRAM and download reduction fp8 already had.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16): route no-GPU loads to the native sd.cpp engine
When no CUDA/ROCm/XPU GPU is available, route diffusion load/generate to the
native stable-diffusion.cpp engine instead of diffusers, with diffusers as the
guaranteed fallback. On CPU sd.cpp is 1.4-2.8x faster and uses 1.5-2.2x less RAM.
- diffusion_engine_router: centralised engine selection (built on the existing
select_diffusion_engine), env opt-outs, MPS gating, recorded fallback reason.
- sd_cpp_backend (SdCppDiffusionBackend): the diffusers backend method surface
backed by sd-cli, with lazy binary install, registry-driven asset fetch,
step-progress parsing, and cancellation.
- diffusion_families: per-family single-file VAE + text-encoder asset mapping.
- sd_cpp_engine: cancellation support (process-group kill + SdCppCancelled).
- routes/inference + gpu_arbiter: drive the active engine via the router; the
API now reports the active engine and any fallback reason.
- tests for the backend, router, route selection, and cancellation.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Phase 16 review fixes: engine-switch unload, sd.cpp error mapping, per-image seeds, Qwen sampler
Address review feedback on #6724:
- engine router: unload the engine being deactivated on a switch, so the old
model is not left resident-but-unreachable (the evictor only targets the active
engine).
- generate route: sd.cpp execution errors (nonzero exit / timeout / missing
output) now map to 500, not 409 (which only means not-loaded / cancelled).
- native batch: return per-image seeds and persist the actual seed for each image
so every batch image is reproducible.
- Qwen-Image native path: apply --sampling-method euler --flow-shift 3 per the
stable-diffusion.cpp docs; other families keep sd-cli defaults.
- honor speed_mode (native --diffusion-fa) and, off-CPU, memory_mode/cpu_offload
offload flags on the native load instead of hardcoding them off.
- fail the load when the sd-cli binary is present but not runnable (version()
now returns None on exec error / nonzero exit).
- size estimate: only treat the transformer asset as a possible local path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 9): gate request-supplied local prequant paths behind operator opt-in
load_prequantized_transformer ends in torch.load(weights_only=False), which executes
arbitrary code from the pickle. The transformer_prequant_path load-request field reached
that unpickle for any local file an authenticated caller named, so a request could trigger
remote code execution. Refuse the source.kind=='path' branch unless the operator sets
UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1; the first-party hosted-repo checkpoint stays trusted
and unaffected. Document the requirement on the API field and add gate tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 10): reset the global attention backend on native, gate arch-specific kernels, accept sdpa
- apply_attention_backend now restores the native default when no backend is requested or a
kernel fails. diffusers keeps a process-wide active attention backend that
set_attention_backend updates, and a fresh transformer's processors follow it, so a load
that wanted native could silently inherit a backend (e.g. cuDNN) an earlier speed-profile
load pinned, breaking the bit-identical/off guarantee.
- select_attention_backend drops flash3/flash4 up front when the CUDA capability is below
Hopper/Blackwell. diffusers only checks the kernels package at set time, so an explicit
request on the wrong card set fine then crashed mid-generation; it now falls back to native.
- Add the sdpa alias to the attention_backend Literal so an API request with sdpa (already a
valid alias of native) is accepted instead of 422-rejected by Pydantic.
- Drop the dead replace('-','_') normalization (no alias uses dashes/underscores).
- perf_levers_probe.py output dir is now relative to the script, not a hardcoded path.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 12): only engage FBCache on context-aware transformers; quantized threshold for GGUF
- apply_step_cache now engages only via the transformer's native enable_cache (the diffusers
CacheMixin path), which exists exactly when the pipeline wraps the transformer call in a
cache_context. The standalone apply_first_block_cache fallback installed on non-CacheMixin
transformers too (e.g. Z-Image), whose pipeline opens no cache_context, so the load reported
transformer_cache=fbcache and then the first generation crashed inside the hook. Such a model
now runs uncached per the best-effort contract.
- GGUF transformers are quantized (the default Studio load path), so they now use the higher
quantized FBCache threshold when the caller leaves it unset, instead of the dense default
that could keep the cache from triggering.
- fbcache_flux_probe.py: compile cached runs with fullgraph=False (FBCache is a graph break, so
fullgraph=True failed warmup and silently measured an eager cached run); output dir is now
relative to the script, not a hardcoded path.
* Studio diffusion (Phase 11): keep professional RTX cards on the fp8 ladder
_is_consumer_gpu treated professional parts (RTX PRO 6000 Blackwell, RTX 6000 Ada) as
consumer because their names carry no datacenter token, so the auto ladder moved int8 ahead
of fp8 and the fp8 path chose fast accumulate for them. The rest of the backend already
classifies these as datacenter/professional (llama_cpp.py _DATACENTER_GPU_RE), so detect the
same RTX PRO 6000 / RTX 6000 Ada markers here and keep fp8 first with precise accumulate.
Also fix the consumer-Blackwell test to use compute capability (10, 0) instead of (12, 0).
* Studio diffusion (Phase 8): tolerate missing torch.float8_e4m3fn in the mxfp8 config
Accessing torch.float8_e4m3fn raises AttributeError on a torch build without it (not just
TypeError on older torchao), which would break the mxfp8 config helper instead of falling
back to the default. Catch both so the fallback is robust.
quant_probe.py: same AttributeError fallback; run LPIPS on CPU so the scorer never holds
CUDA memory during the per-row VRAM probe; output dir relative to the script.
* Studio diffusion (Phase 7): robust backend-flag snapshot/restore and restore on failed speeded load
- snapshot_backend_flags reads each flag defensively (getattr + hasattr), so a build/platform
missing one (no cuda.matmul on CPU/MPS) still captures the rest instead of skipping the
whole snapshot. restore_backend_flags restores each flag independently so one failure can't
leave the others leaked process-wide.
- load_pipeline restores the flags (and clears the GPU cache) when the build fails after
apply_speed_optims mutated the process-wide flags but before _state captured them for unload
to restore -- otherwise a failed default/max load left cudnn.benchmark/TF32 on and
contaminated later off generations.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4): enforce the sd-cli timeout while reading output
Iterating proc.stdout directly blocks until the stream closes, so a sd-cli that hangs
without producing output (or without closing stdout) would never reach proc.wait and the
wall-clock timeout was silently bypassed. Drain stdout on a daemon thread and wait on the
PROCESS, so the main thread always enforces the timeout and kills a hung process (which
closes the pipe and ends the reader). Add a test that times out even when stdout blocks,
and make the no-binary test hermetic so a host-installed sd-cli can't leak in.
* Studio diffusion (Phase 14): guard the int8 exclusion filter against a None fqn
The filter callback can be invoked without a module name, so fqn.lower() would raise
AttributeError on None. Fall back to an empty name (nothing matches the exclusion tokens,
so the linear is kept) instead of crashing the quantise pass.
* Studio diffusion (Phase 16) review fixes: native engine robustness
- sd_cpp_backend: stop truncating explicit seeds to 53 bits (mask to int64);
a large requested seed was silently collapsed (2**53 -> 0) and distinct seeds
aliased to the same image. Random seeds stay 53-bit (JS-safe).
- sd_cpp_backend: sanitize empty/whitespace hf_token to None so HfApi/hf_hub
fall back to anonymous instead of failing auth on a blank token.
- sd_cpp_backend: a superseding load now cancels the in-flight generation, so the
old sd-cli can no longer return/persist an image from the previous model.
- diffusion_engine_router: run the previous engine's unload() OUTSIDE the lock so a
slow 10+ GB free / CUDA sync does not block engine selection.
- diffusion_engine_router: probe sd-cli runnability (version()) before committing to
native, so a present-but-unrunnable binary falls back to diffusers at selection.
- diffusion_device: resolve a torch-free CPU target when torch is unavailable, so a
CPU-only install can still reach the native sd.cpp engine instead of failing load.
- tests updated for the runnability probe + a not-runnable fallback case.
* Studio diffusion (Phase 9) review fixes: prequant safety + validation
- SECURITY: a request-supplied local pre-quant path is now unpickled only when it
resolves inside an operator-configured ALLOWLIST of directories
(UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH = dir[:dir...]). The previous boolean opt-in,
once enabled for one trusted checkpoint, allowed torch.load(weights_only=False) on
any path a load request named (arbitrary code execution). realpath() blocks symlink
escapes; a bare on/off toggle is no longer a wildcard.
- Validate the checkpoint's min_features against the runtime Linear filter, so a
checkpoint that quantised a different layer set is rejected instead of silently
loading a model that mismatches the dense path while reporting the same scheme.
- Tolerant base_model_id compare (exact or same final path/repo segment), so a local
path or fork of the canonical base is accepted instead of falling back to dense.
- _has_meta_tensors uses any(chain(...)) (no intermediate lists).
- prequant verify/probe scripts use repo-relative paths (+ env overrides), not the
author's absolute /mnt paths.
- tests: allowlist-dir opt-in, outside-allowlist refusal, min_features mismatch, fork tail.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 7) review fixes: offload fallback + bench scripts
- diffusion_memory: when group offload is unavailable and the plan falls back to
whole-module offload, enable VAE tiling (the group plan left it off, but the fallback
is the low-VRAM path where the decode spike can OOM). Covers both the group and
sequential fallback branches.
- perf_verify: include the balanced-vs-off PSNR in the pass/fail condition, so a
balanced bit-identity regression actually fails the check instead of exiting 0.
- compare_engines: --vae/--llm default to None (were author-absolute /mnt paths), and
the load-progress poll has a 30 min deadline instead of looping forever on a hang.
- test for the group->model fallback enabling VAE tiling.
* Studio diffusion (Phase 8) review fixes: quant compile + nvfp4 path
- diffusion: a torchao-quantized transformer is committed only compiled. A dense model
resolves to speed_mode=off, which would run the quant eager (~30x slower than the GGUF
it replaced), so when transformer_quant engaged and speed resolved to off, promote to
default (regional compile); warn loudly if compile still does not engage.
- diffusion_transformer_quant: build the nvfp4 config with use_triton_kernel=False so the
CUTLASS FP4 path is used (torchao defaults to the Triton kernel, which needs MSLK);
otherwise the smoke probe fails on CUTLASS-only Blackwell and silently drops to GGUF.
- nvfp4_probe: repo-relative output dir + --out-dir (was an author-absolute /mnt path).
- test asserts the eager-quant -> default-compile promotion.
* Studio diffusion (Phase 10) review fixes: attention gating + probe isolation
- diffusion_attention: gate the auto cuDNN-attention upgrade on SM80+; on pre-Ampere
NVIDIA (T4/V100) cuDNN fused SDPA is accepted at set time but fails at first generation,
so auto now stays on native SDPA there.
- diffusion_attention: _active_attention_backend handles get_active_backend() returning an
enum/None (not a tuple); the old unpack always raised and was swallowed, so
the native-restore short-circuit never fired.
- perf_levers_probe: free the resident pipe on a skipped (attn/fbcache) variant; run LPIPS
on CPU so it isn't charged to every variant's peak VRAM; reset force_fuse_int_mm_with_mul
so the inductor_flags variant doesn't leak into later compiled rows.
- tests for the SM80 cuDNN gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review fixes: sd.cpp installer + engine hardening
- install_sd_cpp_prebuilt: download the release archive with urlopen + an explicit
timeout + copyfileobj (urlretrieve has no timeout and hangs on a stalled socket);
extract through a per-member containment check (Zip-Slip guard); expanduser the
--install-dir so a tilde path is not taken literally; and on Windows CUDA also fetch
the separately-published cudart runtime DLL archive so sd-cli.exe can start.
- sd_cpp_engine: find_sd_cpp_binary honors UNSLOTH_STUDIO_HOME / STUDIO_HOME like the
installer, so a custom-root install is discovered without UNSLOTH_SD_CPP_PATH; start
sd-cli with the parent-death child_popen_kwargs so it is not orphaned on a backend
crash; reap the SIGKILLed child (proc.wait) so a cancel/timeout does not leave a zombie.
- tests: Zip-Slip rejection, normal extraction, studio-home discovery.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 4) review round 2: collect sd-cli batch outputs
Codex review: when batch_count > 1, stable-diffusion.cpp's save_results() writes
the numbered files <stem>_<idx><suffix> (base_0.png, base_1.png, ...) instead of
the literal --output path. SdCppEngine.generate checked only the literal path, so
a batch generation would exit 0 and then raise 'no image' (or return a stale
file). generate now returns the literal path when present and otherwise falls
back to the numbered siblings; single-image behavior is unchanged.
Test: a fake sd-cli that writes img_0.png/img_1.png (not img.png) is collected
without error.
* Studio diffusion (Phase 6) review round 2: img2img source dims + upscale repeats
Codex review on the native engine arg builder:
- build_sd_cpp_command emitted --width/--height unconditionally, so an
img2img/inpaint/edit run that left dims unset forced a 1024x1024 resize/crop of
the input. width/height are now Optional (None = unset): an image-conditioned
run (init_img or ref_images) with unset dims omits the flags so sd.cpp derives
the size from the input image (set_width_and_height_if_unset); a plain txt2img
run with unset dims keeps the prior 1024x1024 default; explicit dims are always
honored. width/height are read only by the builder, so the type change is local.
- build_sd_cpp_upscale_command used a truthiness guard (params.repeats and ...)
that silently swallowed repeats=0 into sd-cli's default of one pass, turning an
explicit no-op into a real upscale. It now rejects repeats < 1 with ValueError
and emits the flag for any explicit value != 1.
Tests: img2img unset dims omit width/height (init_img and ref_images), explicit
dims emitted, txt2img keeps 1024; upscale rejects repeats=0 and omits the flag at
the default. (Two pre-existing binary-discovery tests fail only because a real
sd-cli is installed in this dev environment; unrelated to this change.)
* Studio diffusion (Phase 9) review round 2: correct prequant allowlist doc
Codex review: the transformer_prequant_path field description still told operators
to enable local checkpoints with UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH=1, but the
prior security fix made that variable a directory allowlist -- _allowed_prequant_roots
deliberately drops bare on/off toggle tokens (1/true/yes/...). An operator
following the documented =1 would have every transformer_prequant_path request
silently refused. The description now states it must name one or more allowlisted
directories and that a bare on/off value is not accepted.
Test: asserts the field help references UNSLOTH_ALLOW_LOCAL_PREQUANT_PATH, does
not say =1, and describes an allowlist/directory (guards against doc drift).
* Studio diffusion (Phase 10) review round 2: cudnn/flash3 gating + registry reset
Codex review on attention-backend selection:
- Explicit attention_backend=cudnn skipped the SM80 gate that auto applies, so on
pre-Ampere NVIDIA (T4 SM75 / V100 SM70) it set fine then crashed at the first
generation with no fallback. select_attention_backend now applies
_cudnn_attention_supported() to an explicit cuDNN request too.
- flash3 used a minimum-only capability gate (>= SM90), so an explicit flash3 on a
Blackwell B200 (SM100) passed and then failed at generation -- FlashAttention 3
is a Hopper-SM90 rewrite with no Blackwell kernel. The arch gate is now a
(min, max-exclusive) range: flash3 is SM9x-only, flash4 stays SM100+.
- apply_attention_backend's success path left diffusers' process-wide active
backend pinned to the kernel it set; a later component whose processors are
unconfigured (backend None) would inherit it. It now resets the global registry
to native after a successful per-transformer set (the transformer keeps its own
backend), best-effort. Also fixed _active_attention_backend: get_active_backend()
returns a (name, fn) tuple, so the prior code stringified the tuple and never
matched a name, defeating the native-restore short-circuit.
Tests: explicit cudnn dropped below SM80; flash3 dropped on SM100 and allowed on
SM90; global registry reset after a successful set; _active_attention_backend
reads the tuple return.
* Studio diffusion (Phase 11) review round 2: keep GH200/B300 on the fp8 ladder
Codex review: _DATACENTER_GPU_TOKENS omitted GH200 (Grace-Hopper) and B300
(Blackwell Ultra), though it has the distinct GB200/GB300 superchip tokens. So
_is_consumer_gpu returned True for 'NVIDIA GH200 480GB' / 'NVIDIA B300', and the
auto ladder moved int8 ahead of fp8 on those data-center parts -- contradicting
llama_cpp.py's datacenter regex, which lists both. Added GH200 and B300 so they
are treated as data-center class and keep the intended fp8-first behavior.
Test: extends the datacenter parametrize with 'NVIDIA B300' and
'NVIDIA GH200 480GB' (now _is_consumer_gpu False).
* Studio diffusion (Phase 14) review round 2: apply int8 M=1 exclusion in the builder
Codex review: the M=1 modulation/embedder exclusion was wired only into the dense
runtime quantiser; the offline builder scripts/build_prequant_checkpoint.py called
make_filter_fn(min_features) with no exclusion. So an int8 prequant checkpoint
quantised the AdaLN modulation and conditioning-embedder linears, and loading it
via transformer_prequant_path (the load path only loads already-quantised tensors,
it can't re-skip them) reintroduced the torch._int_mm M=1 crash this phase fixes
for the runtime path.
Extracted int8_exclude_name_tokens(scheme) as the single source of truth (int8 ->
the M=1 exclusion, every other scheme -> none) and use it in both the runtime
quantiser and the builder, so a prequant artifact's quantised-layer set always
matches the runtime. fp8/fp4/mx artifacts are byte-identical (empty exclusion).
Test: int8_exclude_name_tokens returns the exclusion for int8 and () for
fp8/nvfp4/mxfp8.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion (Phase 16) review round 2: native CPU arbiter, status offload, load race
Codex review on the native-engine routing:
- The /images/load route took the GPU arbiter (acquire_for(DIFFUSION) -> evict chat)
unconditionally after engine selection. A native sd.cpp load on a pure-CPU host
never touches the GPU, so that needlessly tore down the resident chat model. The
handoff is now gated: diffusers always takes it, a force-native sd.cpp load on a
CUDA/XPU/MPS box still takes it, but a native sd.cpp load on a CPU host skips it.
- sd_cpp status() hardcoded offload_policy 'none' / cpu_offload False even when
_run_load computed real offload flags (balanced/low_vram/cpu_offload off-CPU), so
the setting was unverifiable. status now derives them from state.offload_flags
(still 'none' on CPU, where the flags are empty).
- _run_load committed the new state without cancelling/waiting on a generation that
started during the (slow) asset download, so a stale sd-cli run against the OLD
model could finish afterward and persist an image from the previous model once the
new load reported ready. The commit now signals the in-flight cancel and waits on
_generate_lock before swapping _state (taken only at commit, so the download never
serialises against generation), mirroring the diffusers load path.
Tests: CPU native load skips the arbiter while a GPU native load takes it; status
reports offload active when flags are set; _run_load cancels and waits for an
in-flight generation before committing.
* Studio diffusion (Phase 14) review round 2: align helper name with the stack
Rename the int8 exclusion helper to exclude_tokens_for_scheme, matching the
identical helper already present higher in the diffusion stack (Phase 16). The
helper definition, the runtime quantiser call, and the offline builder are now
byte-identical to that version, so the two branches no longer introduce a
divergent name for the same single-source-of-truth and the stack merges without
a conflict on this fix. No behavior change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio diffusion: persistent sd-server for the native engine (load once, serve many)
The native sd.cpp tier ran sd-cli one-shot per image, so begin_load only resolved
asset paths and every generation re-spawned sd-cli and reloaded the multi-GB GGUF
from disk (a batch of N = N full reloads). This makes it a resident backend backed by
stable-diffusion.cpp's persistent sd-server, mirroring the chat backend's llama-server
lifecycle:
- begin_load spawns sd-server once (the model loads there) and polls /v1/models until
ready; unload kills it.
- generate submits ONE async /sdcpp/v1/img_gen job for the whole batch (no reload),
polls it to completion, and decodes the returned images. Step progress and ETA come
from the server's stdout (the job JSON has no per-step field).
- The one-shot sd-cli path is kept as an automatic fallback: it is used when sd-server
is absent, and also when a present sd-server fails to start, so behavior is never
worse than before. The public backend surface is unchanged, so routes/router need no
change.
New: sd_cpp_server.py (SdCppServer manager: spawn/readiness/job-submit-poll/cancel/stop,
process spawned inside the drain thread so PR_SET_PDEATHSIG binds to the interpreter, not
a transient thread; empty scratch dir for the server's per-request LoRA/upscaler/embd
scans). Extended: sd_cpp_engine.py (find_sd_server_binary), sd_cpp_args.py
(build_sd_cpp_server_command + build_img_gen_request), sd_cpp_backend.py (server/one-shot
modes, ensure_sd_server_binary upgrades existing sd-cli-only installs), and the prebuilt
installer (locate + chmod sd-server, which ships in the same archive as sd-cli).
Verified on a B200 (Z-Image-Turbo-GGUF, CUDA sd-server): one model load across multiple
generations (server pid stable, a single 'listening on:'), a batch served from one job
with distinct per-image seeds, the second generation faster than the first, and
unload/reload spawning a fresh process. 105 sd.cpp + 81 diffusion tests pass.
Addresses the review of the Phase 16 native-engine PR.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio native diffusion: harden the persistent sd-server path
Addresses review findings on the sd-server backend:
- Router: treat a runnable sd-server as native availability, so an
sd-server-only install (no sd-cli) still routes to the native engine
instead of silently falling back to diffusers.
- Backend: probe the sd-server binary before the multi-GB asset download,
falling back to one-shot sd-cli up front when it cannot run.
- Backend: a lazily cached one-shot fallback engine no longer pins the
backend to one-shot; only an explicitly injected engine does, so a
now-available server can be used on the next load.
- Backend: mask explicit seeds to sd.cpp's signed int64 range before
submitting a server job (large seeds were rejected/wrapped in server
mode only), and split batches above the server's per-job limit into
chunks, each with a timeout proportional to its image count.
- Backend/server: make server startup cancellable. stop() signals an abort
event before taking the lifecycle lock so a blocking readiness wait bails
promptly; unload() stops a not-yet-committed pending server.
- Backend: status() clears stale loaded state when the resident server has
exited, so clients reload instead of hammering a dead process with 500s.
- Server: abandon a poll whose best-effort cancel is not honored within a
grace window (releasing the generate lock), report a pre-submit
stop/cancel as cancellation (409, not 500), and verify JSON responses are
the expected type before indexing.
- Server: use a bounded deque for the stdout tail buffer.
- Add native_mode to DiffusionStatusResponse so the field is not dropped by
the response model.
Adds regression tests for each behavioral change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* sd-server: harden native lifecycle and GPU install path
- Treat a crashed sd-server probe (signal death / non-127 nonzero) as
unavailable so a broken prebuilt falls back to diffusers instead of
routing to a server that dies on startup.
- Drop stale loaded state when a resident server has exited before a
generate, returning the recoverable not-loaded path rather than a 500.
- Reject incomplete server batches (fewer blobs than requested) like the
one-shot path instead of silently dropping images.
- Bound the server log tail in place (keep the deque(maxlen)) and bypass
HTTP(S) proxies for the loopback client (trust_env=False).
- Honor a stop() that arrives after the server is published but before
start() takes the lock, so a cancelled load cannot leak a spawned model
process.
- Map a closed-client RuntimeError during poll to a cancellation when the
generation is being cancelled, so unload races surface as 409 not 500.
- Stop a timed-out server job (best-effort cancel then teardown) so an
abandoned generation cannot keep denoising and block later loads.
- Install the accelerator-matched sd-server build (ROCm/Vulkan/CUDA) and
probe the resident server before auto-installing sd-cli, so a server-only
or GPU host does not fetch the wrong or an unused binary.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
Daniel Han