Applies the video round-2 accuracy findings to the image diffusion stack and fixes
two real image-path bugs found while measuring. All numbers B200, production
settings (family default steps/guidance, 1024px, seed 42, 4 fixed prompts), LPIPS
(AlexNet) via the new scripts/image_speedmem_bench.py, which drives the production
lever functions in the loader's own order.
- inductor precision parity: emulate_precision_casts=True on the regional-compile
path (fused pointwise kernels keep fp32 intermediates where eager rounds to bf16
between ops). Pairwise LPIPS of the compiled tier vs the same-stack eager tier:
Qwen-Image 0.019 to 0.006 at identical speed (72.4 vs 72.5 ms/step), FLUX.1-dev
0.046 to 0.029 at +2% step time (69.8 vs 68.3, reproduced), FLUX.2-klein-4B
0.018 to 0.017 at identical speed. Snapshot/restored with the other process-wide
backend flags so an off load never inherits it.
- cache x compile composition: re-point each cache hook's fn_ref.original_forward
at a torch.compile'd wrapper of the same bound method (armed only where the
speed layer compiled the block; restored before every disable_cache and before
the partial-hook cleanup). Qwen-Image FBCache computed steps 91.8 to 71.2 ms
(back at the uncached compiled rate), 1.21x end to end (7.36 to 6.06 s per 4
images); FLUX.1-dev already traced through its FBCache hook and is measured
neutral (same-process armed vs unarmed latents bit-identical). Skip counts
within noise (13 vs 11 of 76; pairwise LPIPS 0.005).
- FBCache mid-session toggle crash: diffusers 0.39 caches the HookRegistry child
list on first cache_context use, so an uncached generation followed by a
20+-step generation (the auto toggle path) enabled hooks the context never
reached and crashed with "No context is set" (reproduced live on FLUX.1-dev).
Invalidate the stale child cache after every enable_cache.
- TE fp8_dynamic zero-row guard: torchao per-row fp8 derives a per-output-channel
scale from the row amax, so an all-zero weight row is 0/0 = NaN. SDXL's
text_encoder_2 (OpenCLIP bigG) ships exactly such a row, and every explicit
fp8_dynamic SDXL render came out black; keep zero-row Linears dense (LPIPS
0.976 black to 0.096 working). Other families' encoders have no such rows and
are byte-identical.
- No AUTO TE quant exists on the image branch (text_encoder_quant defaults dense,
explicit-only), so the video round's auto-dense retune has no image analogue;
the explicit lever's cost is now measured (TE fp8_dynamic alone, LPIPS vs
bit-exact: Qwen-Image 0.038, FLUX.1-dev 0.084, SDXL 0.096; no speed win, VRAM
-6.5 GB on Qwen-Image) for the docs.
Tests: 96 passing across the cache/speed/precision suites (11 new arming, 2
child-registry, 2 zero-row, 4 inductor-flag); ruff clean.
* scripts: refresh scan_packages allowlist baseline
Regenerate scripts/scan_packages_baseline.json against the current
resolved dependency set so the blocking pip scan-packages gate matches
what the scanner now finds. Refreshes evidence hashes for benign
findings whose code shifted lines (unsloth-zoo mlx loader, gguf/mlx
test /tmp fixtures) and adds two mainstream-library entries that were
newly surfaced (torch inductor codecache base64+subprocess compile
cache, torch testing common_utils socket import). Stale entries whose
matching code changed and no longer triggers are dropped.
All entries remain CRITICAL/HIGH findings manually judged benign;
matched on (package, file, check, evidence_hash).
* ci(security-audit): re-run scan when the allowlist baseline changes
The security-audit pull_request trigger listed the scanners but not
their allowlist baselines, so a baseline-only edit never re-ran the
scan that consumes it. A refreshed baseline could therefore merge
without CI confirming its evidence hashes match what the scanner finds.
Add scan_packages_baseline.json and scan_npm_packages_baseline.json to
the paths filter so baseline changes are validated on their own PR.
Resolve the app-sidebar.tsx conflict: main refactored the chat-export dropdown to a
format-based CHAT_EXPORT_OPTIONS + dynamic-import exportConversationByFormat dispatcher,
which the merged body already uses. Keep main's dispatcher and drop the branch's static
export imports; keep TestTubeOutlineIcon imported from the shared @/lib/hugeicons-derived
module (also used by images-page) rather than main's duplicate inline definition. The
branch's Images and Video nav items are preserved.
Fold PR #6872's image-generation fixes into the branch, deduped against the
round-12 dataset-upload and gallery integrity work already on image-generation.
Fixes carried forward from #6872:
- fp8 single-file transformer memory estimate: an fp8 checkpoint loads with no
quantization_config and diffusers upcasts it to bf16 (~2x resident), so budget
it accordingly in _plan_memory and estimate_safetensors_dense_mib.
- dense-quant OOM-evict preflight: when the GGUF fits resident but the dense bf16
transformer this path materializes does not, skip the fast path up front rather
than evict the current pipeline and OOM in finalization. Combined with the
existing offload->resident candidate re-plan so both the family-table estimate
and the on-disk shard measurement gate engagement (unified on the
transformer_resident_override_mib plan override).
- ControlNet: evict the previous module and its from_pipe wrapper before loading a
new one so swapping ControlNets within a base-model load cannot accumulate to OOM.
- ControlNet union_control_mode: raise on an unknown control type instead of
silently defaulting to canny.
- edit-family mask rejection: raise instead of silently dropping a mask on an
image-editing model that has no inpaint pipeline.
- companion cache: walk the snapshot dir and exclude transformer/ so the
dense-quant prefetch's cached shards do not inflate the companion total and
wrongly force offload.
- training: drop piecewise_constant from the LR scheduler enum and force bf16 for
fp16-incompatible families.
- dataset upload: batch-atomic staging with the same-stem duplicate guard.
- images page: guard negative-prompt restore on guidance>0, clear stale ControlNet
selection on restore, and revert an optimistic quant label when a pipeline load
never starts.
- uninstall (sh + ps1): keep the owner-marker guard on sd.cpp removal.
Conflicts resolved in favour of image-generation's evolved memory system,
loadSpecFor catalog, and stop-and-save (lora_path) run detection; #6872's fp8 and
dense-preflight fixes carried forward on top. All affected backend tests pass
(test_diffusion_backend, test_diffusion_training, test_diffusion_lora_trainer,
test_video_gallery, test_diffusion_controlnet).
The custom-root sd.cpp removal requires the .unsloth-studio-owned marker (written by
install_sd_cpp_prebuilt) before deleting, so a user's own stable-diffusion.cpp checkout
beside a custom root is kept. The default-mode removal of ~/.unsloth/stable-diffusion.cpp
was unconditional, so a user who keeps their own checkout at that path (or points
UNSLOTH_SD_CPP_PATH there), or a pre-marker Studio build, would have it deleted on
uninstall. Guard the default-mode removal on the same owner marker in both uninstall.sh
and uninstall.ps1. Extends the sd.cpp uninstall shell test with owned (removed) and
unowned (kept) default-mode cases.
* WSL ROCm: generalize ROCm-on-WSL bootstrap from Strix-only to any RDNA arch
install_rocm_wsl_strixhalo.sh hardcoded gfx1151, so its verify step died on
discrete Radeon cards even though the ROCm + librocdxg setup is arch-agnostic.
Auto-detect the GPU arch from rocminfo (override via UNSLOTH_WSL_GFX), verify any
GPU agent enumerates over DXG, and map the arch to AMD's per-arch wheel family for
the optional smoke test (injecting librocdxg into torch/lib so torch's bundled
ROCr finds the DXG bridge). Verified on gfx1200 (Radeon RX 9060 XT) in WSL2 +
Ubuntu 24.04 -- torch.cuda now enumerates the GPU.
* WSL ROCm: trigger the ROCm-on-WSL bootstrap for discrete Radeon GPUs too
_maybe_bootstrap_rocm_wsl only fired for Strix APUs (matched via /proc/cpuinfo,
which discrete cards don't appear in). Add _wsl_amd_gpu_name() -- queries the
Windows host via WMI -- and broaden the trigger gate plus the 'already-usable
ROCm' rocminfo check from gfx1151-only to any real GPU agent (gfxNNNN, excluding
the gfx11-generic fallback ISA). The generalized bootstrap then auto-detects the
arch. Enables 'curl install.sh | sh' to set up ROCm-on-WSL on discrete Radeon RX
7000/9000 in WSL2 + Ubuntu 24.04, not just Strix Halo/Point.
* WSL ROCm: address review -- filter generic ISA in bootstrap, bound the host GPU query
- install_rocm_wsl_strixhalo.sh: exclude the gfx11-generic fallback ISA in arch
detection (grep -v generic), matching install.sh's rocminfo check, so a generic
agent listed before the real one can't be picked as the arch.
- install.sh: wrap the powershell.exe Win32_VideoController query in _run_bounded
(10s timeout) so an unstable WSL-interop / busy host can't hang the installer.
* WSL ROCm: harden arch-detect + librocdxg copy under set -eo pipefail (review)
- _detected_gfx: append '|| true' so a no-GPU rocminfo (empty pipeline, non-zero
under pipefail) doesn't abort the assignment before the '[ -z ]' branch prints
the diagnostic + die message.
- smoke-test librocdxg copy: gate on '[ -d "$_tlib" ]' instead of '[ -n ]' so a
non-directory value can't make cp rename librocdxg to 'lib'.
* WSL ROCm: address Codex review (gfx000, 24.04 reroute for discrete, test locator)
- Exclude gfx000 (the CPU agent) from the WSL 'usable ROCm' check and the bootstrap
arch-detect: match gfx[1-9] (nonzero arch), so a partial ROCm install that only
reports the CPU ISA no longer short-circuits the librocdxg setup. (P2)
- Reuse the Ubuntu-24.04 reroute for discrete Radeon: broaden
_maybe_reroute_strixhalo_to_2404's gate with the same _wsl_amd_gpu_name (WMI)
fallback, so a discrete card on 26.04 reroutes to a 24.04 distro like Strix does
instead of falling to CPU. Moved _wsl_amd_gpu_name above the reroute and made it
self-contained + 10s-bounded (it runs before _run_bounded is defined). (P2)
- Update TestInstallShDropinPersistence to locate the gate by its unique
'!/generic/' clause now that the gfx1151 literal is gone. (P1)
* Condense ROCm-on-WSL comments in install.sh and bootstrap helper
* Guard WSL reroute from NVIDIA hybrid hosts and fix GFX-override pipefail check
* Honor CUDA_VISIBLE_DEVICES-hidden NVIDIA in the WSL reroute guard
* Reuse _has_usable_nvidia_gpu in the WSL reroute guard
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
- _local_model_task now tags a local diffusers pipeline that resolves to a video
family (LTX / Wan / Hunyuan) as text-to-video, mirroring the cached-repo
_cached_repo_task, so supported local video pipelines surface in the Video
On-Device picker instead of being routed to the Images picker where the image
loader rejects them. Gated on _local_is_diffusers so only a real loadable
pipeline dir reaches the video check.
- Video validate_load_request now rejects a local pipeline pick whose directory has
no model_index.json before the GPU handoff, mirroring the image loader, so a bad
local pipeline can no longer evict the resident model and only then fail deep in
from_pretrained.
- The custom/env-mode uninstall now removes a sibling stable-diffusion.cpp only when
it carries the Studio owner marker. install_sd_cpp_prebuilt writes the canonical
.unsloth-studio-owned marker on install; uninstall.sh and uninstall.ps1 keep any
unowned checkout (a user's own git clone of stable-diffusion.cpp beside a custom
Studio root is no longer deleted). A pre-marker Studio build is left behind rather
than a user file removed.
Adds regression tests: local video pipeline tagged text-to-video (and a video-named
non-pipeline dir stays untagged so it can never trigger a doomed pipeline load), the
video local-pipeline preflight rejection, the install ownership marker, and the
uninstall keeping an unowned sibling while removing an owned one.
* Studio: exclude mlx-lm 0.31.3 (broke gemma4/qwen3_5 QK-norm load)
mlx-lm 0.31.3 regressed the QK-norm archs: its strict load_weights rejects the
q_norm/k_norm tensors with "Received N parameters not in model", so gemma4 and
qwen3_5 checkpoints fail to load. Studio installs the MLX stack unpinned at
latest, which pulls 0.31.3. Verified on a real macos-14 runner: gemma4 fails to
load on 0.31.3 but loads and generates coherently on 0.31.2 and on git-main
(future 0.31.4). See mlx-lm #1242.
Exclude just that release (!=0.31.3) in the installer and the self-heal floor so
--upgrade still resolves to the newest good build, and treat an already-installed
0.31.3 as unsatisfied so the self-heal replaces it.
* Studio MLX: cover fresh-install path + robust bad-version compare
Address PR review:
- Fresh install.sh (Apple Silicon) runs the base 'uv pip install unsloth' with
SKIP_STUDIO_BASE=1, skipping the guarded MLX-stack step, so transitive
resolution could still pull mlx-lm 0.31.3. install.sh already exports
UV_OVERRIDE -> overrides-darwin-arm64.txt before that install, so exclude
mlx-lm 0.31.3 there too; this also strengthens the self-heal (same override).
- Match the known-bad version with parsed packaging.Version so 0.31.3 == 0.31.3.0
(trailing-zero normalization) instead of raw string equality.
* Studio: exclude mlx-lm 0.31.3 on the fresh Apple Silicon install too
The overrides file only applies via UV_OVERRIDE when it exists relative to the
script, which is not true for a curl-piped install, and the guarded MLX step in
install_python_stack.py is skipped there (SKIP_STUDIO_BASE=1). So the base
install could still resolve the transitive mlx-lm to the broken 0.31.3. Append
mlx-lm!=0.31.3 to the base install on Apple Silicon (empty elsewhere), so the
fresh path pins away from 0.31.3 without waiting for the runtime self-heal.
* Studio: exclude mlx-lm 0.31.3 on the migrated install; keep the >=0.22.0 floor
The with-deps migrated install did not append ${_MLX_LM_EXCLUDE_ARG:-}, so a
curl-piped Apple Silicon migration (no repo overrides file, UV_OVERRIDE unset)
could resolve mlx-lm 0.31.3 transitively. Append the exclusion there, matching
the fresh install path. The no-torch migration is left alone since --no-deps
never resolves mlx-lm (same as the fresh no-torch path).
Also restore the >=0.22.0 floor in overrides-darwin-arm64.txt: a uv override
replaces the transitive constraint, so a bare !=0.31.3 could let the resolver
drop below the supported minimum that mlx_repair.py enforces at runtime.
* Triage huggingface_hub 1.22.0 / fastapi / multiprocess scanner false positives
The scan-packages gate red-failed on all three shards after transitive deps
bumped. Every new CRITICAL is a benign false positive, verified against upstream:
- huggingface_hub 1.22.0 added _sandbox.py for the remote HF sandbox feature.
Its job-startup bootstrap string (fetch sbx-server into the container /tmp and
exec it) and the SandboxPool host-reservation loop trip the staged-dropper and
C2-loop heuristics; that script runs inside a remote HF container, not on the
user machine. The bump also re-hashed the already-reviewed benign polling loops
in hf_api.py and utils/_http.py. The PyPI artifact is byte-identical to the
official v1.22.0 tag.
- fastapi 0.139.0 routing.py re-hashed the websocket keepalive while-True loop;
byte-identical to upstream 0.139.0.
- multiprocess 0.70.19 forkserver.py and tests/__init__.py re-hashed the AF_UNIX
fork-server IPC and fd-inheritance tests; genuine uqfoundation release, local
IPC not network.
Added 7 reviewed allowlist entries (no blind regenerate). All three shards
(hf-stack, studio, extras) exit 0 locally.
* Tighten mlx-lm 0.31.3 exclusion comments
* Trim mlx-lm 0.31.3 exclusion comments
- Diffusion dataset upload now streams each file into a sibling temp file and
atomically os.replace()s it into place only after the whole file is written and
within the size cap. A mid-batch 413 (or any abort) removes the temp, never an
example already stored under the same name, so re-uploading a too-large batch can
no longer truncate or delete a previously uploaded image.
- _diffusion_dataset_summary counts an image as captioned only when it resolves to a
non-empty caption via the same sidecar-over-metadata precedence the trainer uses. An
empty (tombstone) sidecar shadows a metadata row and makes the trainer skip the
image, so counting it over-reported caption_count and mislabeled an effectively
uncaptioned dataset as captioned.
- uninstall.sh/.ps1 now remove a custom/env-mode Studio's native diffusion build that
installs beside the root as a stable-diffusion.cpp sibling (find_sd_cpp_binary
resolves it from the Studio home's parent), guarded by the same unsafe-path check,
and stop processes locking the default-mode stable-diffusion.cpp before removing it.
Adds regression tests for the upload data-loss and caption-count paths and a hermetic
shell test for the custom-root stable-diffusion.cpp removal.
The pip scan-packages gate (SCAN_ENFORCE=1) blocks on non-baselined
CRITICAL/HIGH findings. Recent upstream releases of transitive
dependencies added new files/loops that trip the pattern scanner, so all
three shards (extras, hf-stack, studio) red-failed on legitimate library
code. Add the 7 reviewed findings to scripts/scan_packages_baseline.json.
Each entry is genuine upstream code from the official PyPI archive:
- huggingface-hub huggingface_hub/_sandbox.py (staged dropper + C2 loop):
the HF Jobs sandbox bootstrap string and its host-pool reservation
loop. New in huggingface_hub 1.x (pulled via huggingface_hub>=0.34.0).
- huggingface-hub huggingface_hub/hf_api.py, utils/_http.py (C2 loop):
standard polling / retry while True loops.
- fastapi fastapi/routing.py (C2 loop): websocket receive loop.
- fastmcp-slim fastmcp/cli/apps_dev.py (fs enum + network): the FastMCP
dev CLI (PrefectHQ) making httpx/socket calls.
- cffi cffi/_cffi_gen_src.py (compile + exec): cffi generating and
running C extension source, its core purpose.
Additive only: no existing baseline entry is changed or removed. Verified
by re-running the scanner over the full closure on Python 3.12.13 (the CI
interpreter); it now exits 0 with only MEDIUM findings remaining.
* 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>
* studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)
Adds tool calling for Llama-3, Mistral (pre-v11 + v11+ + [ARGS]), and Gemma 4 to the safetensors / transformers and MLX backends. Parser patched against llama.cpp / vLLM / SGLang per-family parsers and normalises to OpenAI shape. 96 targeted unit tests + cross-OS staging CI (ubuntu / macos-14 / windows) green on the multi-format probe.
* studio: tool-call healing parity between safetensors / MLX and GGUF
After the multi-format parser landed in #5615, the safetensors / MLX
agentic loop and the GGUF loop still differed on healing behaviour.
This commit closes the gaps in both directions so the two backends
react the same way to identical model output.
Changes:
1. core/inference/llama_cpp.py -- the GGUF BUFFERING state machine
now wakes on every emission marker the shared parser knows. Was
("<tool_call>", "<function="); is now the five-tuple imported
from core.inference.tool_call_parser (Qwen / Qwen3.5 / Llama-3
<|python_tag|> / Mistral [TOOL_CALLS] / Gemma 4 <|tool_call>).
Stream cleanup is delegated to the same shared strip_tool_markup
so leaked markup from any family is removed from assistant
content.
2. core/inference/llama_cpp.py -- per-tool canonical heal key. When
a tool arguments field is a bare string and JSON parsing fails,
the GGUF path now heals to {"code": raw_args} for python,
{"command": raw_args} for terminal, and {"query": raw_args} for
everything else. Was hard-coded to {"query": raw_args}, which
silently routed every python / terminal emission through
web_search. Mirrors safetensors_agentic._CANONICAL_HEAL_ARG.
3. core/inference/safetensors_agentic.py -- re-prompt on plan-
without-action. When the model emits a short forward-looking
intent ("I'll search for that", "Let me check", "First, I
will...") and no tool call, the loop nudges the model to act
instead of silently returning a plan-only answer. Up to
_MAX_REPROMPTS=3 (matches GGUF). The intent regex, character
cap, and instruction text are byte-identical to the GGUF path.
The buffer-end fall-through is unified so a buffered intent
emission that never exits the BUFFERING state still triggers
the re-prompt.
4. core/inference/safetensors_agentic.py -- extra iteration slots
for re-prompts. The loop now budgets max_tool_iterations +
_MAX_REPROMPTS + 1 total iterations and tracks the tool-call
count separately, so a stalling model can be nudged 3x without
eating the caller's tool-call budget. Mirrors the _extra slot
reservation in the GGUF path.
Tests (14 new safetensors-side units; 5 GGUF parity pins):
TestLoopRePrompt -- intent-trigger, plain-answer,
no-tools, cap-at-three, budget
preserved, buffer-end intent.
TestLoopCanonicalHealKey -- python / terminal / unknown.
TestGGUFSafetensorsHealingParity -- shared markers used, shared
strip used, canonical heal keys
identical, intent regex matches
same phrases, _MAX_REPROMPTS
equal on both backends.
All 110 targeted tests pass locally; the broader tool / inference /
model-config / sandbox / anthropic / mlx suites stay green.
Why this matters
Without this parity, Llama-3.2 / Mistral / Gemma 4 emissions on Mac
(MLX) and Linux-safetensors stop the agentic loop as soon as the
model says "Let me...", because the GGUF re-prompt logic never
existed on these backends. The two-marker GGUF BUFFERING tuple also
let non-Qwen tool emissions stream out as plain prose when
llama-server's structured channel did not pick them up. Both paths
now drain the same way, heal the same way, and re-prompt the same
way -- so a tool call that works on GGUF works identically on
safetensors / MLX.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix tool-call parser bugs from gemini review on #5620
Three high-priority gemini findings on the tool-call parsing additions:
1. unicode_escape on UTF-8 bytes corrupts non-ASCII literals
(e.g. ✨ becomes â\x9c¨). Replace with json.loads on a quoted
string -- preserves emoji / CJK / RTL while still handling
\n \t \uXXXX escapes.
2. Llama-3 sentinel stripping is order-dependent. A leading
`<|eot_id|><|begin_of_text|>` left `<|begin_of_text|>` behind
because the loop had already passed that sentinel. Loop until
no sentinel matches at the start.
3. Mistral v11+ `[TOOL_CALLS] name { json }` regex uses non-greedy
`\{.*?\}` which truncates at the first `}` of a nested JSON
argument, leaking the tail (e.g. `}}`) into user-visible
streamed text. Same problem for the v0.3 array pattern with
nested brackets. Strip those with balanced brace/bracket
scanning via a new `_strip_mistral_closed_calls` helper called
from `strip_tool_markup`.
Also fix the inference routes' parallel `_TOOL_XML_RE`:
- Same nested-JSON truncation in the Mistral patterns; route the
strip through the parser's balanced-scan helper via a thin
`_strip_tool_xml` wrapper that all existing callers now use.
- Llama-3 `<|python_tag|>[^\n<]*` stopped at any `<`, leaking the
tail of any tool call whose argument contained a literal `<`
(queries, code snippets). Relax to `[^\n]*` which keeps the
strip confined to the actual end-of-line.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tool calling for DeepSeek (R1/V3/V3.1), GLM 4.x, Kimi K2
Adds three more emission-family parsers to tool_call_parser.py so the
shared safetensors / MLX / GGUF agentic loop covers the major open-
weight reasoning families. Patterns ported from llama.cpp
(common/chat-parser.cpp legacy pre-PEG branch), vLLM
(tool_parsers/deepseekv3*, glm4_moe, kimi_k2), and SGLang
(function_call/deepseekv31_detector, glm4_moe_detector, kimik2_detector).
All three references are MIT (llama.cpp) or Apache-2.0 (vLLM, SGLang).
Formats covered:
DeepSeek R1 <｜tool▁calls▁begin｜><｜tool▁call▁begin｜>function
<｜tool▁sep｜>NAME\n```json\n{...}\n```<｜tool▁call▁end｜>
<｜tool▁calls▁end｜>
-- args wrapped in a Markdown json fence, ``function``
literal prefix per llama.cpp common_chat_parse_
deepseek_r1 (chat-parser.cpp:801-820)
DeepSeek V3/V3.1
<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>NAME
<｜tool▁sep｜>{json}<｜tool▁call▁end｜><｜tool▁calls▁end｜>
-- bare JSON, no code fence, no ``function`` prefix
per llama.cpp common_chat_parse_deepseek_v3_1
(chat-parser.cpp:822-879)
GLM 4.5/4.6/4.7 <tool_call>NAME\n<arg_key>k1</arg_key>
\n<arg_value>v1</arg_value>...</tool_call>
-- strings raw, non-strings JSON-encoded per
chat_template.jinja; multi-call is back-to-back
blocks. Per llama.cpp common_chat_parse_glm_4_5
(chat-parser.cpp:1040-1052)
Kimi K2 <|tool_calls_section_begin|><|tool_call_begin|>
functions.NAME:IDX<|tool_call_argument_begin|>{json}
<|tool_call_end|><|tool_calls_section_end|>
-- bare name recovered by stripping ``functions.``
prefix and ``:IDX`` suffix; full id preserved as
tool_calls[i].id so the roundtrip replays verbatim.
Per llama.cpp common_chat_parse_kimi_k2
(chat-parser.cpp:896-913)
Marker collisions
GLM uses the same ``<tool_call>`` opener as Qwen but with a bare
function name + ``<arg_key>`` body (Qwen has ``\s*{`` after the tag).
The dispatch keeps Qwen first; Qwen's _TC_JSON_START_RE returns no
matches on a GLM emission, so the fall-through to _parse_glm_tool_
calls handles it correctly. Existing Qwen tests confirm zero
regression.
Streaming buffer
TOOL_XML_SIGNALS extended from 5 markers to 12 so the BUFFERING state
machine wakes on every new family's section opener. Added the
DeepSeek alternative markers (ASCII underscores, short ``<｜tool▁calls｜>``
form) because real checkpoints emit those variants.
Strip patterns
_TOOL_CLOSED_PATS adds DeepSeek envelope (``<｜tool▁calls▁begin｜>...
<｜tool▁calls▁end｜>``) and Kimi section (``<|tool_calls_section_begin|>
...<|tool_calls_section_end|>``). _TOOL_ALL_PATS adds the same plus
the unclosed-tail variants so a truncated stream does not leak
markup.
Route gate
_detect_safetensors_features._PARSER_MARKERS grows to include
DeepSeek and Kimi markers plus ``<arg_key>`` (the unique GLM signal).
_TOOL_XML_RE (the route-layer markup-strip regex) gets DeepSeek and
Kimi closed-pair patterns. _TOOL_TEMPLATE_MARKERS in llama_cpp.py
adds ``message['role'] == 'tool'``, ``message['tool_calls']``, and
``tool_calls is defined`` so the classifier recognises DeepSeek's
subscripted-access template style (it has no top-level
``{% if tools %}`` block).
Tests (39 new):
TestParserDeepSeek (7) -- R1 fence, short-form opener, V3.1 bare,
multi-call, with-reasoning, strip,
signal-wakes-streaming
TestParserGLM (6) -- single, mixed types, multi-call,
unclosed-heal, no-Qwen-regression, strip
TestParserKimi (6) -- single, multi-call, dotted-name, unclosed,
strip, signal-wakes-streaming
TestParserCrossFormatRouting (2) -- dispatch routing, signal coverage
TestLoopBasic loop integration (3) -- DeepSeek / GLM / Kimi end-to-end
Capability advertise (3) -- DeepSeek / GLM / Kimi templates flip
supports_tools=True
All 398 targeted tests pass locally (115 safetensors + 27 capability
+ rest of tool / inference / sandbox / model-config suites). Builds
on PR #5620 (parser + healing parity for Llama-3 / Mistral / Gemma 4);
will rebase cleanly onto main once #5620 lands. PR opened as draft -
do not merge until validated against real models for each family.
Sources
- llama.cpp common/chat-parser.cpp lines 801-913, 1040-1052 (MIT)
- vLLM vllm/tool_parsers/deepseekv31_tool_parser.py (Apache-2.0)
- vLLM vllm/tool_parsers/glm4_moe_tool_parser.py (Apache-2.0)
- vLLM vllm/tool_parsers/kimi_k2_tool_parser.py (Apache-2.0)
- SGLang python/sglang/srt/function_call/{deepseekv31,glm4_moe,kimik2}_
detector.py (Apache-2.0)
- Live chat templates: deepseek-ai/DeepSeek-V3.1, zai-org/GLM-4.6,
moonshotai/Kimi-K2-Instruct, unsloth/DeepSeek-V3-0324,
unsloth/GLM-4.5-Air, unsloth/Kimi-K2-Instruct
* studio/routes: make python_tag strip multi-line aware
Earlier revisions of _TOOL_XML_RE in studio.backend.routes.inference
oscillated between two bug shapes:
5615 r"<\|python_tag\|>[^\n<]*" -- stopped at any literal "<"
so code='if x < 10: pass'
leaked '< 10: pass)' to the
user.
5620.1 r"<\|python_tag\|>[^\n]*" -- single-line only; the second
line of
python.call(code="a\nb")
leaked.
The full parser (_parse_llama3_python_tag) already handles both via
balanced-brace scanning, so the parsing path was fine; the LEAK was
in the streaming strip path that runs on every cumulative emission
while content is still arriving.
Switch to r"<\|python_tag\|>(?:[^<]|<(?!\|))*" so the strip consumes:
* any character that is not a "<" (newlines, JSON, code, ...),
* a "<" only when it is NOT followed by "|" (i.e. NOT a Llama-3
sentinel start like <|eot_id|>, <|eom_id|>, <|begin_of_text|>).
This means:
* code='if x < 10' stays inside the strip (5615 fix preserved),
* multi-line code stays inside the strip (5620 round 2),
* the strip terminates at the next Llama-3 sentinel so trailing
assistant content survives.
Tests: TestRoutesPythonTagStrip (8 cases)
pytest test_safetensors_tool_loop.py test_safetensors_capability_advertise.py
-> 118 passed in 1.81s (was 110).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: review follow-ups for DeepSeek / GLM / Kimi tool calling
Four fixes addressing review of the parent commit:
1. GLM <arg_value> coercion: tighten the
json.loads -> ast.literal_eval -> raw cascade to only deserialize
when the body unambiguously looks like a JSON literal (object,
array, JSON-encoded string, true/false/null, or numeric). Strings
like ``True`` / ``None`` (Python literals, not JSON) and arbitrary
prose now stay raw. The bare-numeric / bare-boolean ambiguity with
string args remains an inherent limitation of the template without
schema access -- documented in the new comment. Drops the ast
import entirely (closes Gemini's :1036 suggestion).
2. Kimi K2 bare-counter ids (e.g. ``<|tool_call_begin|>3``) are now
dropped rather than surfaced as a tool literally named "3". Matches
vLLM behaviour; SGLang's schema-infer fallback is out of scope at
the parse site. Real Kimi K2 emissions use ``functions.NAME:IDX``
so this is the exception path.
3. Restore the elaborate ``<|python_tag|>(?:[^<]|<(?!\|))*`` clause in
routes.inference._TOOL_XML_RE -- the simpler ``[^\n<]*`` form
regressed PR #5620's multi-line / literal-``<`` python_tag fix.
Restore ``TestRoutesPythonTagStrip`` (8 tests) adapted to call
``_TOOL_XML_RE.sub`` directly since the ``_strip_tool_xml`` helper
was inlined this PR.
4. Add the spaced and backslash-escaped DeepSeek opener variants
(``<｜tool calls begin｜>``, ``<｜tool\_calls\_begin｜>``) to
``TOOL_XML_SIGNALS`` for streaming-gate parity with
``_DEEPSEEK_BEGIN_RE``.
Also updates the llama.cpp / vLLM citations in the parser docstrings:
``common/chat-parser.cpp`` was split into ``common/chat.cpp`` +
``common/chat-peg-parser.cpp`` by llama.cpp PR #18675, and vLLM
moved the tool parsers from ``vllm/entrypoints/openai/tool_parsers/``
to ``vllm/tool_parsers/``. Pin to pre-refactor commit ``51fa458a92d6``
where the cited line numbers still resolve.
New regression tests in ``test_pr5624_regressions.py`` cover the GLM
coercion heuristic shapes, GLM literal-``<`` in arg_value, Kimi K2
dotted name, Kimi K2 bare-counter drop, DeepSeek V3.1 truncated
mid-stream, and routes-layer strip across all three new families.
Tests:
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 170 passed in 1.91s
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten verbose comments in tool-call parser sections
Comments were narrating what the code already says. Cut historical
"earlier revisions used X, then Y" narratives down to one-line WHY
notes where the footgun still matters (canonical heal-key parity,
balanced-brace vs non-greedy regex, ``(?:[^<]|<(?!\|))*`` over
``[^\n<]*``/``[^\n]*``). Drop section-header banners.
No behaviour change. Re-ran:
pytest studio/backend/tests/test_safetensors_tool_loop.py \
studio/backend/tests/test_safetensors_capability_advertise.py -q
-> 118 passed.
Regression replay (parser + _coerce_arguments on the 5 #5615 inputs)
-> 21/21.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: GLM 4.7 no-newline emission + Kimi multi-section parity
Two fixes surfaced by triple-confirm verification against the live
HF chat templates and upstream llama.cpp / vLLM / SGLang parsers.
1. GLM 4.7 silent drop
``zai-org/GLM-4.7/chat_template.jinja`` line 65 uses
``{{- '<tool_call>' + tc.name -}}`` which Jinja strips trailing
whitespace from, so the first ``<arg_key>`` follows the function
name with NO ``\n`` between them. Real emissions look like
``<tool_call>get_weather<arg_key>city</arg_key><arg_value>London
</arg_value></tool_call>``. The previous ``_GLM_TC_OPEN_RE`` ended
the name with ``\n`` so GLM-4.7 calls were silently dropped
(parser returned ``[]``).
Fix: relax the name terminator to a lookahead that accepts EITHER
``\n`` OR the next ``<arg_key>``:
_GLM_TC_OPEN_RE = re.compile(
r"<tool_call>\s*([^\n<{][^\n<]*?)\s*(?=\n|<arg_key>)"
)
The first-char restriction ``[^\n<{]`` still excludes Qwen's
``<tool_call>{json}`` form so the Qwen-vs-GLM dispatch remains
mutually exclusive.
2. Kimi multi-section parity with vLLM / SGLang
``vllm/tool_parsers/kimi_k2_tool_parser.py`` and SGLang's
``kimik2_detector.py`` both use ``re.findall`` and so collect every
``<|tool_calls_section_begin|>...<|tool_calls_section_end|>`` block
in a single stream. The previous implementation stopped at the
first ``<|tool_calls_section_end|>``. Kimi K2 doesn't emit
multi-section in practice, but parity is cheap.
Fix: wrap the existing per-call body parser in an outer loop that
advances past each ``<|tool_calls_section_end|>`` and continues to
the next ``<|tool_calls_section_begin|>``. Body parsing extracted
to ``_parse_kimi_section_body`` for clarity. Truncated final
section is still surfaced via the existing in-body balanced-brace
walk.
Verified independently against the live HF templates:
* GLM-4.7 emission constructed from the live template parses to the
expected ``{name, arguments}`` shape.
* GLM-4.5 / 4.6 newline shape continues to parse (the lookahead also
matches ``\n``).
* Qwen ``<tool_call>{json}`` still dispatches to the Qwen path -- the
first-char restriction stops the GLM regex from biting JSON bodies.
* Kimi two-section stream surfaces both calls in order with full ids
preserved.
* Bare-counter Kimi ids still drop.
Tests added in ``test_pr5624_regressions.py``:
* ``test_glm_4_7_no_newlines_between_name_and_arg_key``
* ``test_glm_4_7_no_newlines_multi_call``
* ``test_glm_4_7_does_not_break_qwen_path``
* ``test_kimi_two_sections_in_one_stream_both_parse``
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 174 passed in 1.93s
pytest studio/backend/tests/ -q -k 'not gpu and not llama_cpp_integration'
-> 2038 passed, 15 failed (pre-existing CI gaps).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: parser robustness fixes for PR #5620
Three surgical extensions to the multi-format tool-call parser, each
covering a real fine-tune / template emission shape that the current
parser silently drops. No path narrows; all changes widen what is
accepted.
1. `_parse_tool_call_json` now accepts both `arguments` and
`parameters` keys. A Hermes / Qwen `<tool_call>{json}</tool_call>`
wrapper around a Llama-3.2 fine-tune that emits the `parameters`
key was extracting the tool name and silently discarding the
args, producing a working-shaped call with an empty payload. The
bare-JSON and python_tag paths already accepted both keys; this
path now matches them.
2. `_TC_FUNC_START_RE`, `_TC_PARAM_START_RE`, and `_TC_PARAM_CLOSE_RE`
now also match the attribute form
`<function name="..."><param name="...">v</param></function>` used
by MiniCPM-5 and MiniMax-M2. Names land in either capture group,
and `</param>` is accepted as a short close.
3. `_parse_llama3_bare_json` sentinel-strip now consumes the role
label inserted between `<|start_header_id|>` and
`<|end_header_id|>` by Meta's official Llama-3.x chat template.
Without this, every assistant turn re-fed through the template
prefix `<|start_header_id|>assistant<|end_header_id|>\n\n{json}`
parsed to zero calls, so any history-with-tool-call round-trip
in production silently dropped.
Tests in `studio/backend/tests/test_safetensors_tool_loop.py`:
* `TestParserRobustness::test_tool_call_json_accepts_parameters_key`
* `TestParserRobustness::test_function_xml_attribute_form`
* `TestParserRobustness::test_function_xml_attribute_form_multi_param`
* `TestParserRobustness::test_function_xml_legacy_equals_form_still_works`
(regression guard for the existing `<function=name>` syntax)
* `TestParserRobustness::test_llama3_chat_template_round_trip`
* `TestParserRobustness::test_llama3_round_trip_all_roles`
* `TestParserRobustness::test_llama3_round_trip_with_eot_prefix`
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 118 to 125 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Trim verbose comments in tool-call parser sections for PR #5624
Pure comment / docstring tightening on top of the GLM 4.7 + Kimi
multi-section fixes. No behavioural change.
* Drop multi-paragraph prelude and post-refactor citation chatter in
the DeepSeek, GLM and Kimi parser docstrings; keep the shape and
upstream-commit pin.
* Collapse ``parse_tool_calls_from_text``'s 9 per-family blocks into
a single ordered loop with one combined comment.
* Tighten the GLM coercion, Kimi bare-counter and ``_TOOL_XML_RE``
comments to one or two lines each.
* Same trim pass on ``_PARSER_MARKERS`` and the regression-test
docstrings.
Tests:
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 174 passed in 2.00s
* Fix O(N^2) DeepSeek V3.1 backtracking for PR #5624
Adversarial input ``<｜tool▁calls▁begin｜><｜tool▁call▁begin｜>fn<｜tool▁sep｜>``
followed by a long body that does NOT contain a closing brace caused
the V3 path's ``([^\n<]+?)<｜tool▁sep｜>`` regex to backtrack
quadratically: at each position the lazy quantifier extends one char
at a time looking for a sep that isn't there, taking ~19s on 50k
chars.
Replace the regex search with ``str.find`` on the sep marker plus a
left-walk to recover the name. ``str.find`` is O(N); the walk stops
on ``\n`` (turn boundary), ``<`` (start of a tag), or ``>`` (end of
an optional ``<｜tool▁call▁begin｜>`` prefix). Same observable
behaviour as the regex on every canonical input.
Tests:
test_deepseek_v3_1_huge_truncated_body_is_linear (new) -- 50k chars
must parse in < 1s.
pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py
studio/backend/tests/test_pr5624_regressions.py -q
-> 175 passed in 1.97s
pytest studio/backend/tests/ -q -k 'not gpu and not llama_cpp_integration'
-> 2038 passed, 15 pre-existing failures unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: terminate function-XML body at </function>, not just </tool_call>
`_parse_function_xml` was looking for `</tool_call>` (the Hermes
wrapper) as the body terminator. When a model emits a standalone
`<function=NAME><parameter=K>v</parameter></function>` followed by
explanatory prose (which models routinely do), no `</tool_call>` is
present, so the body extended to end-of-string and the trailing
prose leaked into the LAST parameter value.
Pre-existing on main (the legacy `<function=NAME>` form had this
bug too). Same affects PR #5620's new attribute-form
`<function name="NAME"><param name="K">v</param></function>`
emission used by MiniCPM-5 / MiniMax-M2.
Fix: `_TC_END_TAG_RE` now matches either `</tool_call>` OR
`</function>`. The existing `_TC_FUNC_CLOSE_RE` / `_TC_PARAM_CLOSE_RE`
strips are unchanged. Multi-call inputs still bound each function
at the next `<function=` start, so no over-eager consumption.
New tests:
* `test_function_xml_followed_by_prose` (legacy form + prose)
* `test_function_attribute_xml_followed_by_prose` (attribute form + prose)
Existing `test_code_with_embedded_xml` still passes (a parameter
value containing literal `<a></a>` is preserved because the
embedded close tag is `</a>`, not `</function>`).
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 125 to 127 passed.
* Studio: tighten Llama-3.2 bare-JSON guard
A fuzz pass on PR #5811 turned up that ``_parse_llama3_bare_json``
accepted ``parameters`` as a string, contradicting the docstring's
"parameters or arguments is a dict" guard. Prose JSON like
``{"name":"foo","parameters":"a sentence"}`` would wrongly fire the
parser, which the agentic loop would then heal into a real
``foo(query="a sentence")`` call.
Same code lives on this branch, so the same fix applies here.
Tightened guard:
- ``parameters`` must be a dict (Llama-3 spec).
- ``arguments`` may be a dict, or a JSON-encoded string that
decodes to a dict (OpenAI shape, e.g.
``"arguments":"{\"q\":\"x\"}"``). Plain non-JSON strings or
JSON-strings of lists / scalars / null no longer pass.
Mirrors the fix landed in PR #5811 commit 615b8608. Adds the same
4 regression tests under TestParserMultiFormat.
Existing test suite stays green: 127 -> 131 passing.
* Studio: skip non-scalar args in python_tag JSON form
The JSON sub-path of ``_parse_llama3_python_tag`` was fabricating
``{"value": args}`` when the model emitted a non-dict / non-string
``arguments`` value (e.g. ``42``, ``[1,2,3]``, ``null``, ``true``).
This silently turned a malformed emission into a real tool call,
which the agentic loop would then execute with arguments the model
never intended.
Tightened: skip the call instead of fabricating. The same
behaviour now matches the bare-JSON guard tightened earlier
(strict-guard merge from PR #5620, inherited via merge here).
Added a regression test covering the four non-scalar shapes.
Pass count on this branch: 158 -> 159.
Sites in ``_parse_tool_call_json`` and ``_consume_mistral_call``
keep the existing looser behaviour for now; both are reached
only after explicit ``<tool_call>`` / ``[TOOL_CALLS]`` markers
so the false-positive surface there is much narrower.
* studio: fix safetensors tool-call parser gaps vs llama.cpp (Mistral CALL_ID / THINK, attribute-form signal)
Three GGUF-parity fixes to the safetensors tool-call parser, each matching
llama.cpp's reference behaviour:
- Mistral Small 3.2 emits [TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}. The
parser stopped after the name on seeing [CALL_ID] (neither [ARGS] nor {),
dropping the call. Skip an optional [CALL_ID]<id> segment in both the
parse and strip paths. llama.cpp parses this (test-chat.cpp:4785).
- Magistral wraps reasoning in [THINK]...[/THINK]. A [TOOL_CALLS] inside the
reasoning was parsed as a real call, producing a phantom call. Strip a
leading [THINK] block before scanning so only the post-reasoning call
counts (test-chat.cpp:2285); a literal [THINK] inside a later argument is
left intact.
- The standalone MiniCPM-5 / MiniMax-M2 <function name="..."> attribute form
parsed correctly but was absent from TOOL_XML_SIGNALS and the markup strip
patterns, so the streaming safety-net parse was gated off (dropping the
call) and markup leaked into displayed text. Add the signal and broaden
the strip regexes.
Adds regression tests for all three.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix GLM and Kimi K2 safetensors tool-call parser gaps vs llama.cpp
Four GGUF-parity fixes for the GLM and Kimi K2 families:
- GLM 4.7 zero-argument inline call <tool_call>name</tool_call> was dropped:
the open-tag lookahead only allowed \n or <arg_key> after the name. Allow
</tool_call> too so a no-arg call parses to empty args (vLLM / SGLang /
llama.cpp all parse it).
- GLM string argument values were stripped, losing significant leading /
trailing whitespace in code / diff arguments. Keep the raw value for the
string fallback and only strip the copy used to probe for a JSON literal,
matching vLLM glm4_moe which never strips string args.
- Kimi K2 calls emitted without the <|tool_calls_section_begin|> wrapper
were dropped. llama.cpp makes the section optional (Kimi can call a tool
straight after reasoning without opening a section); parse a bare
<|tool_call_begin|> when no section is present.
- Kimi K2 malformed / truncated JSON in one call dropped every later call in
the section. Skip the bad call and keep parsing so valid subsequent calls
are recovered (vLLM parity).
Adds regression tests for all four.
* studio: fire safetensors tool calls for the bare-JSON (Llama-3.2) form
The agentic loop's streaming safety-net parse was gated on
has_tool_signal(), which is False for the Llama-3.1 / 3.2 bare-JSON tool
form {"name":..,"parameters":..} (no XML marker). Real tool calls were
therefore dropped: the loop logged "model planned without calling tools",
re-prompted three times, then gave up with zero tool calls, while GGUF's
llama-server parses the same emission natively.
Run parse_tool_calls_from_text() unconditionally in the safety net. The
parser is strict (only fires on a valid tool-call shape) so plain answers
are unaffected. Reproduced on a real unsloth/Llama-3.1-8B-Instruct run:
the model emits {"name":"web_search","parameters":{...}} which now
executes the tool instead of being re-prompted into a no-op.
Adds a loop regression test for the bare-JSON form.
* studio: fire safetensors tool calls for Gemma 4 (native template + stripped parser)
Gemma-4 safetensors fired no tools while its GGUF fired reliably. Three gaps:
- The Studio swaps in the Unsloth "gemma-4" chat template, which does not
render the tools schema (the model's native template does), so the model
never saw the tools. Fall back to the model's native template when the
override template renders identically with and without tools. Same fix
helps any family whose override template drops tools.
- skip_special_tokens strips the <|tool_call> wrapper and <|"|> string
markers, so a streamed Gemma-4 call arrives as a bare call:NAME{k:v, ...}
with unquoted values. Parse that form, keeping commas/braces inside a
code or command value, normalising surrounding quotes, and stripping the
leaked markup from the final answer.
- Without a grammar a small model can loop, repeating one call for the whole
tool budget. Collapse exact-duplicate calls within a turn and force a final
answer after a turn that made no new tool progress (llama-server's lazy
grammar prevents this loop on the GGUF side).
Adds parser tests for the bare/stripped Gemma-4 form.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: complete strict-mode contract and fix parser import paths
Address review findings on the multi-format tool-call parser:
- Honor allow_incomplete=False in the remaining sub-parsers. The Llama-3
<|python_tag|>NAME.call(...) parser, the pre-v11 Mistral [TOOL_CALLS] array
parser, and the Gemma 4 <|tool_call> parser ignored strict mode, so a
truncated call (missing closing paren, ], or <tool_call|>) was still healed
and executed with Auto-Heal disabled. Thread strictness through and reject
the unclosed forms, matching the JSON and function-XML paths.
- Drop the duplicate tool_call_parser import block in llama_cpp.py and the
redundant un-aliased TOOL_XML_SIGNALS; only the _SHARED_TOOL_XML_SIGNALS
alias is used as a value.
- Import _strip_mistral_closed_calls from core.inference.tool_call_parser in
routes/inference.py instead of studio.backend.core... The self-contained
run.py launch mode only puts studio/backend on sys.path, so the absolute
package path raised ModuleNotFoundError on the server-tool strip path.
Add strict-mode regression tests for the truncated Llama-3 dot-call and the
unclosed Mistral array.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden DeepSeek/Kimi tool-call parsing and strip
Address review findings on the DeepSeek and Kimi parsers:
- Honor allow_incomplete=False for DeepSeek. An envelope with no closing
<｜tool▁calls▁end｜> is truncated mid-stream; reject it in strict mode
instead of healing the body out to EOF, matching the strict XML and Mistral
paths.
- Do not skip a following tool call when the current call's end marker is
missing. The DeepSeek V3 and Kimi loops advanced by searching forward for the
next <｜tool▁call▁end｜> / <|tool_call_end|>, which could land on a later
call's end marker and drop the call in between. Advance by the JSON end; the
loop re-locates the next call marker from there.
- Strip truncated DeepSeek and Kimi section blocks in the route-level display
regex. The patterns required the closing marker; add the end-of-text
alternative so a block truncated by EOS does not leak raw markup to the UI.
Add regression tests for the truncated DeepSeek envelope, and for DeepSeek and
Kimi multi-call recovery when the first call's end marker is missing.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: preserve XML param indentation and alias Mistral array parameters
Two parser-correctness fixes found by auditing against the model chat templates
and the SGLang / vLLM reference parsers:
- Qwen3.5 XML parameter values lost their leading indentation. The chat template
emits <parameter=k>\nVALUE\n</parameter>, but the parameter-start regex ate the
wrapping newline AND the value's first-line indentation with a trailing \s*,
then str.strip() removed the rest. Narrow the trailing class to horizontal
whitespace only and trim exactly one wrapping newline (via _trim_param_value),
preserving indentation in code/diff arguments. Matches SGLang's qwen3_coder
detector. Applies to both _parse_function_xml (tool_call_parser.py) and the XML
path in tool_healing.py.
- Mistral pre-v11 array objects keyed on parameters dropped their payload.
_consume_mistral_call read only the arguments key; alias parameters the same way
the JSON/XML paths and SGLang's base detector do.
Add regression tests for preserved multi-line indentation and the array
parameters alias.
* Studio: DeepSeek strip sync, Gemma nested args, GLM/Kimi strict mode
Parser-correctness fixes found by auditing DeepSeek/GLM/Kimi against vLLM,
SGLang, and the model chat templates:
- DeepSeek: the short <｜tool▁calls｜> opener (and the space / escaped-underscore
spellings) was parsed but never stripped, so a short-opener envelope leaked raw
markup to the UI. Share one opener alternation between _DEEPSEEK_BEGIN_RE and
the strip patterns (and the route-level display regex) so a signal we parse can
never be left un-stripped.
- Gemma wrapper-less stream: a nested object/array argument (loc:{city:NYC},
labels:[bug,ui]) was kept as a literal string. Parse it recursively when the
bare value is a balanced {} / [], falling back to the raw string for a
truncated value.
- GLM and Kimi ignored allow_incomplete. With Auto-Heal off, a GLM block with no
</tool_call>, a Kimi section with no <|tool_calls_section_end|>, or a Kimi call
with no <|tool_call_end|> are truncated and must be rejected, matching the
strict behavior of the JSON/XML/Mistral/DeepSeek paths and vLLM/SGLang.
Add regression tests for the short-opener strip, the Gemma nested args, and GLM /
Kimi strict-mode rejection.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: tighten tool-call parser comments
Make the comments in the multi-format tool-call parser and its callers succinct:
compress verbose docstrings/blocks to one or two lines, drop ones that restate the
code, and trim the tiny balanced-scanner helpers. Correctness rationale and
upstream provenance (SGLang/llama.cpp parity, the strict-mode / Auto-Heal
contract, whitespace-preservation, and the Unicode / full-width-pipe notes) are
kept in compact form.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).
* Studio: tighten DeepSeek/GLM/Kimi parser comments
Compress the comments added for the DeepSeek/GLM/Kimi parsers and the Gemma
wrapper-less helpers to one or two lines, keeping the upstream provenance
(llama.cpp 51fa458a92d6), the O(N^2) / strict-mode rationale, and the vLLM parity
notes intact.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).
* Studio: make DeepSeek R1 / GLM parsing linear and close routes strip gaps
Review follow-up for the DeepSeek/GLM/Kimi parser:
- DeepSeek R1 detection used a greedy ``([^\n]+)\n```json`` regex that backtracks
O(N^2) on a fence-less truncated body; scan with str.find instead (mirrors the
V3 path).
- GLM arg pairs used a lazy-group finditer that rescanned to EOF from each bare
<arg_key> in an unclosed body (O(N^2)); walk pairs with str.find.
- The route display strip (_TOOL_XML_RE) accepted fewer DeepSeek openers than the
parser (missed the space / escaped-underscore spellings) and missed bare
section-less Kimi calls, so a call we parse could leak raw markup to the UI.
Reuse the parser's shared _DEEPSEEK_OPEN_RE_SRC and add a bare-Kimi arm.
Add ReDoS-linearity regressions for the R1 and GLM paths, a positive R1
fenced-json parse test, and routes-strip tests for the space/escaped DeepSeek
openers and the bare Kimi call.
* Studio: fix test_mcp_servers _TOOL_XML_RE reconstruction after _DS_OPEN_SRC reuse
The routes strip fix made _TOOL_XML_RE reference the module-level
_DS_OPEN_SRC variable. test_mcp_servers reconstructs the regex by exec-ing
the extracted compile() source in a namespace that only defined _re, so it
raised NameError. Inject _DS_OPEN_SRC into that namespace, matching the same
fix already applied in test_tool_xml_strip.
* Studio: make Llama-3 .call and Mistral-array healing parsing linear
Two more O(n^2) ReDoS paths in the multi-format parser, both reachable from
the agentic loop on a long truncated body with no length cap:
- _LLAMA3_KV_RE.finditer over a .call(...) body retried at every offset of a
long word run / unterminated quote (40K -> 14s). Replace with a hand-scan
that reuses the same key/number/literal sub-regexes via anchored match and
walks the string body by hand, so an unterminated quote is O(n). Verified
byte-identical to the old regex over 200K fuzzed inputs.
- _parse_mistral_array healing ran _balanced_brace_end from every { in the
body (20K -> 17s). Walk top-level objects, advancing past each balanced
{...}; this also drops the phantom call the old scan emitted from a nested
argument object.
Add adversarial-length linearity regressions plus positive .call kwargs and
unclosed-array recovery coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strengthen #5624 regression assertions and strip-test harness guards
- test_strip_tool_markup_handles_deepseek_envelope used `A or B` where B was the
preservation property the next line already asserts, masking the real check.
Replace with an explicit assertion that the call name and args are stripped.
- The test_tool_xml_strip source-extraction harness reconstructs _TOOL_XML_RE and
_strip_tool_xml_for_display from routes/inference.py via lazy regexes that could
silently grab a shorter slice. Assert the extracted regex carries the DeepSeek /
bare-Kimi arms and the helper body reached the _TOOL_XML_RE.sub call.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor strict mode in safety-net, keep empty Gemma args, strip attribute-form function XML
- safetensors safety-net parser now forwards allow_incomplete=auto_heal_tool_calls,
matching the draining path, so a late incomplete tool call is not healed and
executed when Auto-Heal is off.
- Gemma empty bare value ({k:}) now serialises as "" instead of invalid {"k":},
which previously dropped the whole call.
- Route _TOOL_XML_RE also strips the <function name="..."> attribute form
(MiniCPM-5 / MiniMax-M2) so it no longer leaks to the UI.
* Studio: linearize wrapper-less Gemma nested-arg parsing and correct parser provenance
- _gemma_parse_value/_gemma_parse_mapping/_gemma_parse_array now parse nested
{}/[] in a single forward pass instead of pre-scanning each subtree with a
balanced-brace walk and re-parsing it. Deeply nested wrapper-less Gemma args
were O(n^2); they are now ~linear (and ~40x faster at depth 400).
- Correct the DeepSeek/GLM/Kimi provenance comments: the cited commit
51fa458a92d6 is unrelated, and GLM/Kimi were never standalone
common_chat_parse_* functions (llama.cpp uses common_chat_params_init_glm_4_5
plus a generalized XML parser, PRs #15904 / #16932).
- Add tests: Gemma deep-nesting linearity, nested object/array preservation,
same-turn distinct-call cap, and the native-template tool-render fallback.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: guard Gemma value parser against non-advancement and missing tokenizer
Addresses Gemini review:
- _gemma_parse_value now consumes one character when a stray }/]/, sits where a
value is expected, so _gemma_parse_array can never stall at the same index on
malformed input (a latent infinite loop).
- _render_with_native_template returns None when neither a tokenizer nor a
processor is present instead of raising AttributeError.
- Tests for both.
* Studio: fix attribute-form function-XML literal close tag and zero-arg strict call
Addresses Codex review of the <function name="..."> attribute form in
_parse_function_xml (MiniCPM-5 / MiniMax-M2):
- End the call body at the LAST </function> / </tool_call> within the call's
window, so a literal close tag inside a code/search argument (e.g.
print("</function>")) is preserved instead of truncating the call.
- Accept a closed call with no parameters as a valid zero-argument call in strict
mode (the function close is already required), instead of rejecting it as a
truncated call.
- Tests for both, mirroring the legacy <function=...> coverage.
* Studio: drop scratch review/planning artifacts from the branch
* Studio: fix tool-call parser/loop review findings on the multi-format path
Address the live code-review findings on the safetensors/MLX + GGUF tool path:
- routes: include the attribute form <function name="..."> in the safetensors
capability whitelist so MiniCPM-5 / MiniMax-M2 templates keep the tool pill
(parser already handles the form; the post-filter wrongly suppressed it).
- safetensors loop: build the plan-without-action re-prompt from the active
tools instead of a hardcoded web_search/python string, and gate it on
auto_heal_tool_calls, matching the GGUF loop.
- safetensors loop: hold a leading bare-JSON object ({"name":..,"parameters":..})
during BUFFERING until it closes, then drain it as a tool call instead of
streaming the raw JSON to clients. The DRAINING/STREAMING resolvers still
recover a plain JSON answer, so this can never drop content.
- parser: anchor the Llama-3 <|python_tag|>NAME.call(...) scan to the tag and
chain ; -separated calls, so all semicolon-separated built-ins parse and a
literal <|python_tag|>x.call(...) inside a JSON string argument no longer
fires the wrong tool.
- parser: consume the optional trailing </s> after a named Mistral
[TOOL_CALLS]name{json} call, mirroring the array shape.
- GGUF streaming strip: use the shared parser patterns (which know
[TOOL_CALLS] and <|python_tag|>) so a textual tool call entering DRAINING is
stripped instead of leaking the marker to streaming clients.
- routes: hoist the _strip_mistral_closed_calls import to module level.
Adds regression tests covering each fix; existing parser suite stays green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: fix DeepSeek/GLM/Gemma tool-call review findings
Address the live code-review findings specific to the DeepSeek / GLM / Kimi
and native-template additions:
- parser: in strict mode (Auto-Heal off) require the per-call
<｜tool▁call｜end｜> terminator for DeepSeek V3 calls instead of executing on
a bare balanced object closed only by the envelope end.
- parser: keep GLM string arguments that begin with a quote verbatim (drop
the leading-quote case from the JSON-decode probe) so a quoted search query
is not decoded down to its inner text.
- parser: reject a GLM call with an unclosed <arg_value> in strict mode, and
under Auto-Heal keep the partial value rather than dropping it to a no-arg
call.
- parser: add a balanced wrapper-less Gemma strip (call:NAME{...}) so a nested
object/array argument is removed whole instead of leaving a trailing brace;
run the balanced Mistral and Gemma strips on the streaming display paths too.
- safetensors loop: buffer a leading wrapper-less Gemma call:NAME{...} so it
drains and executes instead of streaming the raw call text.
- inference: render the native-template fallback on a shallow tokenizer copy
instead of mutating the shared tokenizer outside the generation lock, and
load the native template from base_model for LoRA adapters.
Adds regression tests for each; existing parser suite stays green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden multi-format tool-call detection from review findings
Apply five targeted fixes from the review pass over the multi-format tool
path:
- routes: route display strip delegates to _strip_tool_xml so Mistral
[TOOL_CALLS] blocks with nested JSON are removed from streamed display
text, not just the XML forms.
- tool_call_parser: skip function/parameter starts that fall inside an
already-open parameter block (_inside_open_parameter) so nested example
payloads are not mis-parsed as new calls; extract
strip_llama3_leading_sentinels so the bare-JSON guard is shared.
- safetensors_agentic: probe bare JSON through strip_llama3_leading_sentinels
before the balanced-brace check so a leaked header sentinel does not defeat
the guard.
- tool_healing: allow dotted tool names in the Gemma wrapped start pattern.
- llama_cpp (GGUF): buffer wrapper-less Llama-3.2 {"name":..} calls that carry
no XML signal, drain a complete object silently and hold an incomplete one,
and run the end-of-stream safety net unconditionally so markerless calls are
detected and never leak the raw JSON (including truncated fragments).
Adds regression tests for the GGUF bare-JSON streaming path and the Mistral
display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop bare-JSON tool calls leaking at EOF, oversized, and into history
The second review pass flagged that the Llama-3.2 bare-JSON tool-call handling
still leaked raw JSON in several spots; ``strip_tool_markup`` only knows
XML/bracket markup, so the bare-JSON form survived it. Fix them symmetrically
across the safetensors and GGUF loops:
- Safetensors stream-end resolver now routes a held bare-JSON fragment to
DRAINING (mirroring GGUF) so a truncated ``{"name":..`` cut off by the end of
the stream is dropped instead of flushed as assistant content. The 7/10
reviewer finding.
- Both loops now drain (suppress) an oversized still-open bare-JSON call once it
passes ``_MAX_BARE_JSON_BUFFER`` instead of streaming the raw prefix, gated on
a ``"name"`` key so a giant plain JSON answer still streams; a complete
oversized call still executes via the safety net.
- Add a shared ``strip_leading_bare_json_call`` helper and apply it to the
content kept for the assistant turn in both loops, so an executed bare-JSON
call is not replayed as visible text or fed back as next-turn history.
Plain JSON answers without a ``"name"`` key are untouched throughout. Adds
regression tests for the EOF, oversized, and next-turn cases on both backends
plus unit tests for the helper.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: bound the Llama-3 python_tag strip on real control sentinels
The route display strip's <|python_tag|> arm ran to the next <| of any kind.
A tool-call argument carrying a literal <|...|> token (for example <|cite|>
inside a string value) truncated the strip early and leaked the call tail into
the visible response. Narrow the stop condition to the genuine Llama control
sentinels (eot_id, eom_id, python_tag, start/end_header_id, begin_of_text,
finetune_right_pad_id) so embedded markup and JSON are consumed while real
header/turn boundaries still bound the strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden GLM/Gemma parsing, cap GGUF textual calls, share native-template fallback
GLM 4.x parser walked a body pre-bounded by the first </tool_call>, so a string
argument containing a literal </tool_call> (e.g. code that prints it) was
truncated. Walk arg_key/arg_value pairs against the full content instead, since
each <arg_value> is delimited by its own </arg_value> and the call's real close
is the </tool_call> that precedes the next <arg_key>.
Add a truncated wrapper-less Gemma pattern (call:NAME{... with no closing brace)
to the markup strip so a call cut off mid-arguments does not leak raw into the
visible stream. It runs after the closed form, so a complete call keeps trailing
prose.
Cap and dedup tool calls parsed from the GGUF TEXTUAL fallback at
_MAX_TOOL_CALLS_PER_TURN, mirroring the safetensors loop. Structured
delta.tool_calls are grammar-bounded by llama-server, but text parsed straight
from content is not, so one runaway turn could fan out into dozens of
executions.
Extract the native-chat-template fallback into chat_template_helpers
(render_native_template / render_with_native_template_fallback) so the
transformers and MLX text backends share one implementation. The MLX text path
now applies it too, so an Unsloth override template that drops the tools schema
no longer silently stops MLX from advertising tools. The MLX VLM path renders
via the processor for image tokens and is intentionally left on its own render.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate markerless bare JSON on enabled tools and close parser/strip asymmetries
The Llama-3.2 custom_tools bare-JSON form has no marker, so any JSON object with a
name key was read as a tool call. An ordinary JSON answer like
{"name":"Alice","parameters":{"age":30}} was misclassified as a call to a
disabled tool and dropped from the visible response. Gate the markerless form on
the enabled tool names (threaded through parse_tool_calls_from_text and
strip_leading_bare_json_call, supplied by both streaming loops): an object whose
name is not an enabled tool is ordinary content. The marker-based forms keep
their name-agnostic behaviour (an explicit signal is a real call attempt), and
unrestricted mode stays ungated.
Also fix two parser/strip asymmetries the parser already tolerated:
- A literal </function> inside a parameter value (print("</function>")) truncated
both the core and route strips at the first close, leaking the tail. Extend the
strip to the call's real close (last </function> before the next opener),
mirroring the parser, without merging separate calls.
- The single-object Mistral [TOOL_CALLS]{...} shape parsed but _strip_mistral_closed_calls
left it, leaking the raw object into display. Strip the balanced object while
keeping trailing prose, matching the array and name shapes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: fix strip/parse symmetry and native-template token for DeepSeek/GLM/Kimi
Pass-3 review follow-ups on the multi-format tool parser:
- Bare Kimi call (<|tool_call_begin|>...<|tool_call_end|> with no section
wrapper) is accepted by the parser, so add it to the closed strip patterns
so the streaming (non-final) display strip removes it instead of leaking the
markup mid-generation.
- Route display strip now also runs the wrapper-less Gemma cleanup, so a
Gemma 4 call:NAME{..} no longer leaks into the visible answer.
- MLX model record carries base_model for a LoRA adapter so the native-template
fallback loads the base repo template rather than the adapter's
(often template-less) tokenizer.
- Native-template reload forwards the load-time HF token so a gated/private
model's repo template can still be fetched (transformers and MLX text paths).
- GGUF end-of-stream bare-call heuristic is gated on the enabled tool names so a
truncated ordinary JSON object ({"name":"Alice","age":) streams as the answer
instead of being dropped as a tool call.
Adds regression tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: gate GGUF bare-JSON suppression on enabled tools and fix python-tag exponent parsing
Pass-4 review follow-ups on the GGUF tool loop and Llama-3 parser:
- The GGUF bare-JSON suppression sites still keyed off a raw "name" substring,
so an ordinary JSON answer whose name is not an enabled tool was dropped when
it was truncated, oversized, or reached the no-tool DRAINING fallback (the
parser, helper, and safetensors paths were already gated). All three sites now
use the shared enabled-name gate, and a held bare-JSON buffer that turns out not
to be an enabled call is shown as the answer instead of dropped at stream end.
- The Llama-3 python-tag numeric kwarg regex matched only the mantissa, so
scientific notation was truncated to its leading digits (1e-3 parsed as 1) and a
tool executed with the wrong value. The regex now accepts exponent and decimal
forms, and the int/float classification keys off the exponent too.
Adds regression tests for the truncated / oversized disabled-name JSON cases (and
a counterpart that a truncated enabled call still does not leak) plus the
scientific-notation kwargs.
* Studio: drop accidentally committed async worker transcripts
Eight generated reviewer / async-worker transcripts were committed under
studio/backend/async_task_outputs/. They are not imported or referenced by any
code and carry only internal task state, so they should never ship in the repo.
Remove them and gitignore the directory so they cannot be re-added.
* Studio tools: gate safetensors bare-JSON drain, fix nested-name gate and function-XML strip
Pass-4 review follow-ups on the shared parser / safetensors loop:
- The safetensors oversized and end-of-stream bare-JSON drain branches keyed off
a raw "name" substring, so a large or truncated ordinary JSON answer whose name
is not an enabled tool was drained instead of streamed. Both now use the shared
enabled-tool-name gate, matching the GGUF path.
- strip_leading_bare_json_call matched the first "name" anywhere, so a plain JSON
answer with a nested name equal to an enabled tool ({"result":{"name":"web_search"}})
was wrongly suppressed. It now extracts the TOP-LEVEL name only, walking past
nested objects/arrays and keeping the text when a top-level value is truncated.
- The function-XML display strip used a regex negative-lookahead that stopped at a
literal <function=...> opener inside a parameter value and then dropped the rest
of the answer to EOF. A scan-based strip mirrors the parser (ignores openers
inside an open <parameter> via _inside_open_parameter) and closes each call at its
real </function>, so trailing assistant text after such a call survives.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep tools prompt when native-template probe raises; make helper tests hermetic
Pass-4 review follow-ups on the native-template fallback:
- render_with_native_template_fallback re-renders the live template with tools=None
to detect whether it dropped the schema. A template that requires tools can raise
on that probe; that must not discard the already-valid tools prompt. The probe is
now wrapped so any error returns the original formatted_prompt (transformers would
otherwise fall back to manual formatting and lose the schema; MLX would let the
exception escape).
- The native-template helper tests imported InferenceBackend just to reach the
thin wrapper, which pulls in unsloth and its optional vllm package metadata. They
now call the dependency-light render_native_template helper directly so they pass
in a backend/test environment without vllm. Adds a probe-raises regression test.
* Tool parsing: 3.9 import safety, disabled-Auto-Heal contract, capability gate
Round-2 review follow-ups on the multi-format tool-call parser:
- tool_call_parser: add `from __future__ import annotations`. The module
is dependency-light by design (external llama-server wrappers import it
standalone) and the package targets python >=3.9, where its PEP 604
`int | None` return annotations would raise TypeError on import.
- safetensors + GGUF drain fallback: gate the leading bare-JSON strip on
auto_heal_tool_calls. With Auto-Heal off, a truncated enabled-name
fragment that did not parse now stays visible, matching the XML strip
in the same branch and the disabled-Auto-Heal contract. With Auto-Heal
on it is still suppressed.
- safetensors capability gate: match the bare-JSON `{"name":` template
marker with a whitespace/escape-tolerant regex so a pretty-printed
`{ "name" :` or JSON-escaped `{\"name\":` template is not mis-classified
as tool-less. The parser already accepts that whitespace via
raw_decode, so the gate must too.
Regression tests added for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* GLM tool-call display strip: treat literal close tag in arg value as data
Round-2 review follow-up on the GLM 4.x tool-call format.
The GLM call shape is <tool_call>NAME<arg_key>k</arg_key><arg_value>v
</arg_value>...</tool_call>. The parser was hardened to walk arg_key /
arg_value pairs so a literal </tool_call> inside an argument value (e.g.
print("</tool_call>")) is treated as data and the call's real close is the
</tool_call> that precedes the next <arg_key>. The display strips still used a
non-greedy <tool_call>.*?</tool_call> regex, which stopped at the literal and
leaked the call's tail into visible content and stale history.
Add _strip_glm_calls, a scan that mirrors the parser's close detection, and run
it before the regex arms in every strip pipeline: the core strip_tool_markup,
the route _strip_tool_xml display/history cleanup, and the safetensors + GGUF
streaming strips. Qwen / Hermes <tool_call>{json} has no NAME token after the
opener, so it is left to the regex arms unchanged.
Regression tests cover the literal-close-tag leak (core + route), normal GLM
calls, back-to-back GLM calls, zero-arg GLM, truncated GLM, and untouched Qwen.
* Tool parsing: symmetric "function" bare-JSON alias and route strip parity
Round-3 review follow-ups, all parser/strip symmetry fixes.
- Bare-JSON "function" alias: the markerless parser accepts a call name via
obj.get("name") or obj.get("function"), but the strip/gates only knew "name",
so a {"function":<enabled tool>} call executed while its raw JSON leaked. Teach
_top_level_bare_json_name the alias (with "name" precedence and the same nested
and truncated-name guards), and widen the guards in strip_leading_bare_json_call,
the safetensors and GGUF _looks_like_enabled_bare_json gates, and the route
capability marker regex.
- Route display/history cleanup: strip a tail-only </param> alias close (the
parser accepts <param name="...">...</param>), and run the parser's guarded
function-XML scan (_inside_open_parameter) before _TOOL_XML_RE so a literal
nested <function=...></function> inside an argument value does not truncate the
strip and leak the tail.
Regression tests added for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: fix DeepSeek strict recovery, Kimi dotted names, Gemma spaced streaming
Round 3 review fixes for the DeepSeek / GLM / Kimi tool-call parsing path.
- DeepSeek R1 and V3/V3.1 strict parsing (Auto-Heal off): when a call is
truncated (missing closing fence or <tool_call_end> terminator), skip it
and keep scanning for later well-formed calls instead of breaking out and
dropping the rest of the envelope. This matches the Kimi strict parser's
recovery behaviour.
- Kimi dotted tool names: keep the full name after stripping only the
functions. prefix and :idx suffix, e.g. functions.mcp.server-list:0 stays
mcp.server-list. The previous split on "." truncated dotted MCP names to
their last segment. This matches current vLLM
(tool_id.split(":")[0].removeprefix("functions.")) and SGLang
(^(?:functions\.)?(?P<name>[\w.\-]+):(?P<index>\d+)$).
- Gemma wrapper-less call streaming: hold the whitespace-tolerant prefix
(call : NAME) in the streaming suppression buffer, matching the parser's
_GEMMA_BARE_TC_RE, so the spaced spelling split across chunks is buffered
instead of leaking as visible text. Applied to both the safetensors and
llama.cpp streaming paths.
- Remove dead _render_with_native_template method and the now-unused copy
import from inference.py; the live path uses render_with_native_template_fallback.
Adds regression tests for DeepSeek R1/V3 strict recovery, Kimi full dotted
name preservation, and the Gemma spaced-call streaming suppression.
* Studio tools: honor tool budget in GGUF loop and guard function-XML streaming strip
Round 4 review fixes. Both are asymmetric-fix bugs where the final/steady path got a
guard the analogous streaming/loop path did not.
- GGUF tool-call budget: the safetensors loop counts real tool-call turns against
max_tool_iterations (re-prompt stalls excepted), but the GGUF loop only bounded the
turn count by the enlarged range (max_tool_iterations + _MAX_REPROMPTS). Since this
PR raised _MAX_REPROMPTS from 1 to 3, a model that keeps making valid tool calls
could run up to three extra tool rounds (with max_tool_iterations=1, four rounds
instead of one). Add a _tool_iters_done counter that increments only when a tool
actually executed in the turn, and stop once the caller's budget is spent so the
post-loop final-answer nudge fires. A duplicate/disabled no-op turn is a correction
turn (like a plan-without-action re-prompt) and does not consume budget, preserving
the existing "already completed" re-prompt behavior.
- Streaming display strip: the final strip runs the guarded _strip_function_xml_calls
scanner (a literal <function=...> inside a parameter value is data, not a nested
call), but the GGUF and safetensors streaming strips still used only the open-ended
regex arms. When a tool-call argument contained literal function markup, the regex
tail ate everything to end-of-text and dropped the real trailing prose after the
call's true </function>. Run the guarded scanner (and the balanced Mistral strip)
before the regex arms in both streaming paths so streaming and final display agree.
Adds regression tests: GGUF valid tool calls respect max_tool_iterations, and the
streaming strip keeps trailing prose after a function-XML call with a literal marker.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: safetensors tool budget counts only executed turns (GGUF parity)
Follow-up to the GGUF budget fix. The safetensors loop charged max_tool_iterations
per non-re-prompt iteration (iteration + 1 - reprompt_count), so a duplicate/disabled
no-op turn spent a budget slot even though no tool ran. With a small cap this dropped
real work: for max_tool_iterations=2, a model that made a valid call, repeated it (an
internal no-op correction turn), then made a distinct valid call executed only the
first -- the third turn was sent with no tools and the distinct call was ignored.
Track whether a turn actually executed a tool (set on record_result) and count only
those turns against the cap, matching the GGUF loop. A duplicate/disabled no-op is a
correction turn -- like a plan-without-action re-prompt -- and no longer consumes
budget, so the model still gets its "already completed" nudge and another tool-enabled
turn. Adds a regression test for the small-cap duplicate-then-distinct-call flow.
* Studio tools: fix stale Kimi dotted-name regression test
test_pr5624_regressions.py still expected functions.my.tool:0 to resolve to the last
segment (tool). The parser now preserves the full dotted name (my.tool) after removing
only the functions. prefix and :idx suffix, matching current vLLM/SGLang so dotted MCP
names like mcp.server-list survive. Update the assertion, name, and module docstring to
the corrected contract (the raw id is still preserved on the call).
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: don't force a tool re-prompt on a negated intent (safetensors parity)
The safetensors _INTENT_SIGNAL claimed to mirror GGUF but was missing the
negative lookahead, so a refusal like "I will not search the web for that"
matched the "i will" intent and triggered the plan-without-action re-prompt
(STOP... you MUST call a tool), overriding a valid no-tool answer. GGUF already
excludes not/never. Add the same (?!\s+(?:not|never)\b) lookahead so both
backends agree. Extends the intent parity test with negated refusals.
* studio: parse the outer envelope before DeepSeek/Kimi markers embedded in its args
parse_tool_calls_from_text ran the DeepSeek/Kimi marker pre-pass before the shared
<tool_call>/<function=...> parser. When a Qwen/Hermes call's argument contained
literal Kimi/DeepSeek markup (for example a user asking the model to explain that
syntax), the pre-pass matched the embedded marker and returned it, executing the
wrong tool and dropping the real call. Skip the pre-pass when a <tool_call> or
<function=...> envelope opens before the first DeepSeek/Kimi marker, so the shared
parser takes the outer call; a genuine marker-led call (no leading envelope) still
goes through the pre-pass. Tests for the embedded-marker case and the control.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: trim redundant comments (comment-only, AST-verified)
* Studio: trim redundant comments (comment-only, AST-verified)
* Studio: prevent Gemma tool-parser DoS on stray delimiters
_gemma_parse_value returned the input index unchanged when text[i] was a
stray delimiter (,}]), so the list and mapping caller loops that advance
on the returned index spun forever at 100% CPU on malformed input such as
[},]. Advance past the delimiter so parsing always terminates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip Magistral [THINK] reasoning from final display/history
strip_tool_markup removed [TOOL_CALLS] and <function> markup but left a
leading Magistral [THINK]...[/THINK] block intact, so its bracket-form
reasoning (not the <think> the reasoning channel renders) leaked into the
safetensors display and conversation history while GGUF/llama.cpp routes
it natively. Drop the leading reasoning block at end-of-turn (final=True)
via the existing _strip_mistral_reasoning helper; streaming is untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: keep times in wrapper-less Gemma tool arguments
The wrapper-less Gemma value scanner used _GEMMA_KEY_RE = [\w.\-]+ for keys,
which also matches a digit-leading token, so a comma followed by a time or
ratio inside a value (call:web_search{query:meet at 10:00, 11:00 tomorrow})
was misread as a new 11: key, truncating the query and injecting a bogus
argument. Require keys to start with a letter or underscore, matching the
identifier-start rule the wrapped path already uses (_GEMMA_NEXT_KEY_RE).
Add a regression test.
* Studio: treat markers/close-tags inside tool-call arguments as data
Four parser correctness fixes where a valid argument string was mistaken for
structure:
- DeepSeek: find the envelope-end token outside JSON strings, so a query/code
argument containing the literal token no longer truncates the body and drops
the whole call.
- GLM: locate the real </arg_value> as the one whose next token is <arg_key> /
</tool_call> / end, so a value containing a literal </arg_value> (or
</tool_call>) is kept instead of executing the tool with corrupted arguments.
- Attribute-form <function name="..."> envelopes now count in the embedded-marker
guard, so a DeepSeek/Kimi marker inside a parameter value does not hijack the
outer call and run the wrong tool.
- Wrapper-less Gemma call:NAME{...} is gated on the enabled tool names (parse and
display strip), mirroring the Llama bare-JSON gate, so a disabled/example name in
prose is not stolen as a call and the real answer is preserved.
Add regression tests for each.
* Gate route Gemma wrapperless strip by enabled tools; make Kimi section-end search string-aware
Route-level display stripping now threads the enabled tool-name set into the
Gemma wrapperless-call strip, so prose that mentions a disabled tool
(call:foo{...}) is preserved while active tool calls are still stripped. This
mirrors the parser-level gate already used in tool_call_parser.
The Kimi section-end lookup now searches outside JSON string literals, so a
section-end marker appearing inside an argument string no longer triggers a
false truncation that drops a valid tool call.
* Run DeepSeek/Kimi pre-pass when a closed tool-call example precedes a real block
The marker pre-pass was skipped whenever any <tool_call>/<function> opener
appeared before the first DeepSeek/Kimi marker, even when that opener was a
CLOSED syntax example in prose that ends before the real block. In that case
parse_tool_calls_from_text skipped the DeepSeek/Kimi parsers and the genuine
tool call was dropped while a phantom tool named in the example ran instead.
Only treat a marker as embedded in a leading envelope when removing the closed
outer <tool_call>/<function> envelopes also removes every marker (the marker
actually sat inside one). A marker left standing is a real call, so the pre-pass
runs. The legitimate case of a marker inside a closed outer envelope's arguments
is preserved.
* Honor reasoning_effort none in safetensors prefill; strip Magistral reasoning while streaming
Two safetensors/MLX reasoning fixes surfaced in review:
_sf_reasoning_prefill_mode only checked enable_thinking, so an
enable_thinking_effort (GLM-5.2) request that disables thinking via
reasoning_effort=none (without enable_thinking=False) still began in
prefilled-<think> mode. A plain answer with no </think> was then swallowed
whole into reasoning_content and the visible response came back empty. Thread
reasoning_effort into the predicate and treat none as disabled, mirroring
_request_reasoning_kwargs.
strip_tool_markup_streaming stripped tool markup but not the leading Magistral
[THINK]...[/THINK] bracket block, so the raw chain-of-thought leaked into the
streamed safetensors content instead of the reasoning drawer (GGUF routes it
natively). Apply _strip_mistral_reasoning first, matching the final strip; an
unclosed [THINK] is held from the marker on so nothing flickers.
* Heal truncated outer tool envelopes and keep quoted Gemma args intact
Two follow-ups from review of the marker pre-pass and Gemma parsing:
The leading-envelope guard only removed CLOSED outer <tool_call>/<function>
envelopes before deciding whether a DeepSeek/Kimi marker was embedded, so a
truncated outer call missing its close tag (whose argument embeds a marker) was
treated as a standalone marker and the embedded sample ran instead of the
intended outer call being Auto-Healed. Decide on the last outer opener before the
marker and whether it closed before the marker instead, so a closed syntax
example still runs the pre-pass while a real closed-or-truncated outer call keeps
it.
The wrapper-less Gemma argument scan tracked bracket depth but not quotes, so a
quoted value containing a comma followed by a key-like token (a search query such
as "weather, location: Boston") was split mid-string, truncating the value and
fabricating an extra argument. Track quote state (with escapes) so the top-level
comma boundary is only taken outside quoted spans.
* Span outer envelopes to their real close when locating embedded markers
Locating the DeepSeek/Kimi marker relative to a leading outer envelope used the
FIRST close tag after the opener, so a literal </function> or </tool_call> inside
an argument value (for example python code that contains the text) was mistaken
for the envelope boundary. The marker after it was then treated as a standalone
call and the embedded sample ran instead of the intended outer call.
Match the closed outer envelopes with the shared patterns that already extend to
the real final close (a literal close inside a value is data), and treat a marker
that survives their removal as embedded only when a still-open (truncated) outer
opener precedes it, so Auto-Heal still repairs a truncated outer call. A closed
syntax example before a genuine block still runs the pre-pass.
* Span the tool_call outer envelope to its real close in the marker guard
The leading-envelope check reused the lazy <tool_call>.*?</tool_call> strip
pattern, so a Qwen/Hermes JSON argument containing a literal </tool_call> ended
the span early. A DeepSeek/Kimi sample later in that same string then survived
the closed-envelope removal, and the pre-pass executed the embedded call instead
of the outer <tool_call>. The <function> arm already spanned to its real close;
give <tool_call> the same real-close pattern (with the negative lookahead that
keeps back-to-back calls separate) so a literal close inside a value is data.
* Preserve no-tool Gemma prose and keep later R1 calls when healing a close
Two review follow-ups:
_gemma_strip_gate returned None when no tools were enabled, and None means
strip every markerless call:NAME{...} block, so a no-tool answer that documents
the syntax (or the Anthropic display path, which passes an empty tool list as
None) had that prose deleted. It is a display/history gate, so return the
enabled-name set instead -- an empty set when no tool is enabled, which strips
nothing because every call:NAME{...} is then prose.
The DeepSeek R1 heal path located the close fence with an unbounded forward
search, so when a first call had balanced JSON but omitted its fence the search
landed on a LATER call's terminator and pos advanced past that valid call,
dropping it. Match the close immediately after the JSON (whitespace-skipped) like
the strict path, and advance by just the JSON when it is absent, so a multi-call
turn keeps its later well-formed calls (heal is now a superset of strict).
* Resume wrapper-less Gemma scan past a consumed call's balanced body
The markerless call:NAME{...} scan used finditer, which resumes right after the
opening call: token, so a nested call:OTHER{...} mentioned inside the first
call's own quoted string argument (for example a web_search query that quotes the
Gemma tool syntax) was re-matched and returned as a spurious second tool call,
executing an unintended tool. Walk with a manual cursor that resumes after the
outer call's balanced body (brace matching already skips quoted braces), so a
call's arguments are never rescanned. Genuinely separate back-to-back calls and
disabled/example prose are unaffected.
* Mistral outer call wins over XML literals; align healer signals with its parser
Two follow-ups on the shared-parser ordering after the healing-passthrough
merge:
- A well-formed [TOOL_CALLS] call whose JSON arguments quote tool XML parsed
the literal instead of the outer call (executing the wrong tool). When the
first XML signal sits inside a leading balanced Mistral body it is argument
data, so the Mistral parser now runs first; an XML signal before the trigger
keeps the normal order, so a [TOOL_CALLS] literal inside an XML call's
arguments still stays data.
- passthrough_healing buffered streams on the parser module's broadened signal
list (now including <|python_tag|> and [TOOL_CALLS]) but promotes with
core.tool_healing, which does not parse those forms: a streamed Mistral or
Llama text call was held until finalization and flushed as prose. The healer
keeps its own signal list limited to the formats it can promote, restoring
immediate streaming for the rest.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: Gemma wrapper-less marker literals and quotes, GLM embedded close pair
- The Gemma fallback deferral now keys on an actual wrapped opener
(_GEMMA_TC_RE), not the wrapper literal anywhere in content: a wrapper-less
call whose argument merely mentions <|tool_call> has nothing tool_healing
can parse, and deferring it lost the call entirely (not executed and
stripped from display).
- New _gemma_body_brace_end boundary scanner honors single- and double-quoted
strings like _gemma_parse_stripped_body, shared by parse and strip, so a
quoted brace in a code argument (code:print('}')) no longer truncates the
executed arguments or the strip span.
- _glm_value_close now requires a structural </arg_value> to sit at balanced
quote state: the full pair </arg_value></tool_call> embedded inside a string
literal is data, not an early close. When no candidate balances, the first
token-valid close wins as before.
* Address review: leading envelopes win over rehearsed literals
- New _first_foreign_tool_signal shared by the leading-envelope guards adds
<|python_tag|> to the protected signal set: the spelled-out literal inside a
Mistral call's arguments (a query about Llama built-in tool syntax) executed
the inner literal instead of the outer call.
- New _xml_signal_inside_leading_bare_json guard, sibling of the Mistral one:
a leading bare-JSON call whose string argument quotes tool XML (a code value
citing <function=...>) had the literal promoted by the shared XML pass
before the bare-JSON parser ran.
- Magistral [THINK]...[/THINK] is dropped once at parse entry instead of only
inside the Mistral parser, so a call rehearsed in the think block in a
foreign format can no longer be promoted while the real call after the
block is lost. Parse now agrees with the display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: a disabled leading bare-JSON object keeps its literals as data
When the leading bare-JSON object is ordinary content (name not an enabled
tool), the guard proved the first tool signal sits inside it, so falling
through to the XML/python_tag passes promoted quoted string data as a real
call. Drop the object and parse only the tail: a real call after the object
still parses, nothing inside it can be promoted.
* Address review: apostrophes in raw Gemma values, GLM strict key contract, per-model template token
- Quote openers in the wrapper-less Gemma boundary and body scanners now
require value-start context (after : { [ ( , =): an apostrophe inside an
unquoted value (query:what's the weather) opened quote mode, swallowed the
real closing brace, and lost the whole call on common contraction queries.
Quoted values keep hiding delimiters as before.
- A GLM <arg_key> with no <arg_value> tag now rejects the call in strict
mode, matching the unclosed-value contract, instead of executing the tool
with the argument silently dropped; Auto-Heal keeps the lenient skip.
- The native-template fallback reads the hf_token stored on the model record
instead of the instance-wide last-load token, so a later token-less load
cannot break template fetches for a previously loaded gated model (both
the transformers and MLX backends).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: Mistral literals inside leading JSON, whitespace-tolerant wrapped Gemma opener
- The leading bare-JSON guard now treats the [TOOL_CALLS] trigger as a
foreign signal: the Mistral parser runs before the bare-JSON one, so a
literal quoted inside the leading object's strings was promoted over the
outer call (or over ordinary JSON content).
- tool_healing's wrapped Gemma opener tolerates whitespace around call and
the colon: sampling drift emits call: name{ and call : name{, and
rejecting those lost the call entirely because no fallback re-parses the
wrapped form. Strict mode still requires the closing tag.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: DeepSeek/Kimi markers inside leading JSON and Mistral envelopes stay data
The DeepSeek/Kimi pre-pass runs before the outer-call parsers, and
_marker_inside_leading_envelope only protected XML envelopes: a marker
quoted inside a leading bare-JSON or Mistral call's argument strings was
promoted as a separate no-arg call and the real outer call dropped. The
guard now recognizes those two leading envelopes as well; standalone
DeepSeek/Kimi calls keep parsing.
* Address review: accept dotted Gemma argument keys in the key-quoting scanner
The scanner quoted keys of [alnum_-] only, so a dotted key (user.name:...)
was left unquoted, json.loads failed, and the whole wrapped call was lost
(parse empty, strip wipes the markup). Dots now match the parser's own
key/name charset.
* Address review: a real DeepSeek/Kimi call after a disabled leading JSON object still parses
DeepSeek/Kimi markers are foreign signals for the leading bare-JSON guard
too: a marker literal inside a disabled leading object made the envelope
guard skip the pre-pass for the whole message, so a real DeepSeek/Kimi call
after the object was dropped. Routing the case through the guard's
drop-and-parse-the-tail recursion reaches the real call while the literal
inside the object stays data.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: leading Mistral call owns the turn, dotted keys after bare values
- A LEADING parseable [TOOL_CALLS] call now runs the Mistral parser first
unconditionally: literal XML in trailing prose after the call was promoted
by the earlier shared XML pass, executing the quoted example instead of
the real leading call. XML leading keeps the normal order.
- _GEMMA_NEXT_KEY_RE accepts dots so a dotted key after a bare value
(query:foo,user.name:bob) ends the value at the comma instead of being
swallowed into it, matching the round-earlier key-quoting charset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: a leading wrapper-less Gemma call owns the turn
A quoted foreign literal inside a leading wrapper-less Gemma call's
argument (a query citing another tool syntax) was promoted by tool_healing
before the Gemma fallback ran, executing the quoted example and dropping
the outer call. New leading guard, sibling of the Mistral and bare-JSON
ones, gated on an enabled name since the form is markerless. Foreign markup
leading keeps the normal order.
* Fix merge resolution: restore both leading-guard test classes intact
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: markup quoted inside a nameless leading JSON answer stays data
The leading bare-JSON guard required a top-level name, so a structured JSON
answer quoting tool markup in its strings (a response_format turn
documenting a tool's syntax) had the literal promoted by the later passes.
A nameless leading object that parses as real JSON now routes through the
same decline-then-parse-the-tail path; non-JSON braced prose keeps the old
behaviour, and a real call after the answer still parses.
* Address review: JSON answers stay data, nested Gemma quotes, earliest envelope, no failure caching
- A whole-content JSON value is a structured answer: the markerless Gemma
scan and its strip no longer promote or strip a quoted example of an
enabled tool's syntax inside it.
- Nested stripped-stream Gemma values now unquote quoted string leaves
recursively, so {loc:{city:"New York"}} hands the tool New York, matching
the top-level coercion.
- The DeepSeek/Kimi pre-pass dispatches by earliest envelope opener, so a
leading real call wins over a trailing example of the sibling format in
either direction.
- A failed native-template fetch is no longer cached as no-template: the
next call retries after the model record's token is fixed or a transient
Hub error clears; only definitive loads are cached.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: closed calls precede the marker pre-pass, truncated Gemma scan stops, quoted nested delimiters
- A closed non-DeepSeek/Kimi call preceding the first DS/Kimi marker owns
the turn: a trailing syntax example, or one quoted inside a wrapped Gemma
argument, was promoted by the pre-pass and dropped the real leading call.
Wrapped Gemma joins the outer-envelope pattern sets.
- An unbalanced wrapper-less Gemma call now stops the scan (mirroring the
strip contract) instead of resuming inside its own argument text, where a
quoted enabled call would be promoted.
- Raw-quoted strings in nested stripped-stream Gemma values hide delimiters,
so {city:"New, York"} is one value instead of a split pair, returned
unquoted like the top-level coercion.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: string-marker literals in wrapper-less args, mid-value quoted phrases
- The wrapper-less deferral guard no longer keys on the <|"|> literal: a
real call whose argument merely mentions the string marker was deferred to
tool_healing, which has no wrapped opener to parse, losing the call. The
wrapped-opener check alone owns the deferral.
- Double quotes now also open at the start of a word, so a quoted phrase
mid-value (query:find "weather, location: Boston", limit:3) hides its
delimiters instead of splitting the value into garbage keys; apostrophes
keep the value-start-only rule so contractions stay prose.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: strict GLM refuses in-quote close fallback, Gemma guard covers preambles
- _glm_value_close gains a strict flag: a truncated value whose only close
candidates sit inside a string literal rejects the call in strict mode
(Auto-Heal keeps the lenient partial), restoring the strict contract the
quote-aware fallback had weakened.
- The leading wrapper-less Gemma guard no longer requires the call to open
the response: a visible preamble before call:NAME{...} is the normal
shape, and the quoted foreign literal inside the argument was promoted
again in that shape. An enabled balanced call beginning before the first
foreign signal owns it.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: contextual GLM quote openers, disabled Gemma examples stay prose, JSON array answers
- The GLM value-close quote tracker uses the same contextual openers as the
Gemma scanners (single quote after punctuation context, double quote also
at word start), so strict mode accepts a normal apostrophe value again
while still rejecting a truncated value whose only close candidates sit
inside a string literal.
- A disabled wrapper-less Gemma call is prose by design, so a tool literal
quoted inside it no longer promotes: the span is dropped for parsing and
the tail parsed, mirroring the nameless-JSON guard.
- Leading JSON ARRAY answers join the leading-JSON envelope guard, so a
marker quoted inside a structured array response stays data.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Align closed-envelope regression test with the document-order contract
The test asserted the pre-round-13 behavior (trailing DeepSeek/Kimi block
wins over a leading closed envelope) while the shipped rule is document
order: the leading closed call owns the turn. Rename the test and assert
the leading call so the suite matches the contract exercised by
test_leading_xml_call_wins_over_trailing_kimi_example.
* Parse a leading Llama-3.2 bare-JSON call before the markerless Gemma scan
The bare-JSON form only ever matches a leading call object, and document
order says that call owns the turn. Running the Gemma wrapper-less scan
first let an enabled call:NAME{...} snippet quoted inside the leading
call's string arguments steal the turn when the JSON was not the whole
content (trailing prose or a second ;-separated call), executing the
quoted tool instead of the real one. Reordering cannot take a leading
Gemma call's turn since that content never starts with an object brace.
* Leading-call ownership: Mistral trigger in Gemma guards, closed bare JSON before markers, depth-aware nested Gemma values
Three parser gaps against the document-order contract:
The wrapperless Gemma leading guards did not count [TOOL_CALLS] as a
foreign signal, so a leading Gemma call quoting a Mistral snippet in its
argument lost the turn to the quoted literal. Both the enabled-call and
disabled-example guards now include the trigger, matching the bare-JSON
guard's local inclusion.
_marker_inside_leading_envelope required the DeepSeek/Kimi marker to sit
inside the first closed bare-JSON or Mistral call. A marker after that
closed call (a trailing example or data in a later ;-chained call's
strings) now also defers to the leading call, the same inside-or-after
rule the closed XML envelope patterns already applied.
The nested Gemma primitive value scan split on every comma, corrupting
arguments like opts:{code:print(1,2),lang:py}. It now applies the same
paren/brace depth, contextual quote openers, and comma-only-before-a-key
mapping rule as the top-level scan.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gemma leading guard: a closed enabled call preceding the signal owns the turn
The wrapperless Gemma guard only claimed the turn when the first foreign
signal sat inside the first enabled balanced call. When that call closed
before the signal (a second call quoting a Mistral or Kimi literal, or a
trailing prose example), the guard forfeited the turn and the foreign
parser promoted the quoted literal, dropping the real Gemma calls. Apply
the same inside-or-after ownership rule as the closed bare-JSON and
Mistral envelopes, gated on an enabled name so the name-agnostic legacy
path is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Marker guard: only an executable leading bare-JSON call owns the turn
The bare-JSON branch of the leading-envelope marker guard claimed the
turn for any NAMED leading object. A disabled-name object is prose by
design (the bare-JSON parser will not execute it), so deferring the
DeepSeek/Kimi pre-pass to it lost the real later call entirely. Gate the
ownership claim on the enabled set (or the name-agnostic None path). A
marker inside the disabled object's own strings stays data, matching the
tail-exclusion contract; a marker after it now falls through so the
pre-pass parses the real call. The Mistral branch stays ungated since
[TOOL_CALLS] parsing is never name-gated.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Gemma scan skips leading JSON answers; GLM heal bounds values at structural tags
Two fixes to the document-order data contracts:
The markerless Gemma scan only exempted whole-content JSON, so a leading
JSON answer followed by prose had an enabled call:NAME{...} snippet
inside its strings promoted to a real executed call and stripped from
the displayed answer. Both the parse and strip scans now start after a
balanced json-valid leading value span, keeping parse and strip
mirrored. Real calls after the answer still parse; mid-prose JSON gets
no exemption.
The GLM heal fallback for a missing closing arg_value tag took the
entire remainder as the value, executing markup-contaminated arguments
like city="NYC</tool_call>" and swallowing trailing prose. The healed
value now stops at the next arg_key or tool_call close and the pair walk
resumes there. EOF-truncated values keep the partial heal, strict mode
still rejects, and closed values holding a literal close tag in quotes
are untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Compress docstrings in the multi-format tool parser to their contract essence
* Condense parser guard comments and test narration to contract essentials
* verify_import_hoist: exempt __future__ imports and same-diff relocations
Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.
* Leading bare-JSON calls own the turn; function calls end at the first balanced close
The XML-signal guard for a leading bare-JSON call required the signal
strictly inside the object, so a trailing XML example stole the turn
from the leading call; it now applies the same inside-or-after rule as
the Mistral guard. Function-XML calls also ended at the LAST close tag,
which let prose after a closed call that mentions a literal close tag
get swallowed into the final parameter value; calls now end at the
first close tag that is not inside an open parameter, and the strip
mirrors the same rule so parse and strip agree.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Attribute-form calls end at the first balanced close; bare-JSON strip requires the call shape
The attribute form parser still kept the last close tag in the call
window, folding prose after a closed call into the final parameter
value. It now takes the first close not inside an open parameter, the
same rule the equals form and the strip already use.
The leading bare-JSON strip deleted any closed object whose top-level
name matched an enabled tool, including plain JSON answers the parser
correctly rejects as non-calls. The strip (and the drain gate that
delegates to it) now requires the parser's exact call shape, so answers
like {"name":"web_search","result":...} stream and display intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* False-alarm markers keep the answer; the bare-JSON strip consumes the whole chain
The trailing strip arms dropped everything from a bare marker to EOF,
so a normal answer that mentions [TOOL_CALLS] or another marker
literally was truncated (or fully swallowed when it started with the
literal) after the no-call drain fallback. Those arms now require a
call-shaped lookahead or marker-at-EOF before dropping; truncated real
calls still strip.
Chained bare-JSON turns executed both calls but stripped only the first
object, so the second call's raw JSON replayed into the next assistant
history message alongside the structured tool_calls. The strip now
consumes the entire chained run of call-shaped enabled objects while
non-call answers, disabled names, and trailing prose stay intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* DeepSeek and Kimi trailing strip arms require a call-shaped lookahead
Same false-alarm rule as the bare-word markers: a prose answer that
mentions a DeepSeek or Kimi marker literally keeps its tail, while
truncated real envelopes and bare end-of-text fragments still drop.
* Attribute-form containment, parameter-close-decides rule, preamble-tolerant Mistral guard, strict strip shape
Four document-order and containment fixes. A leading attribute-form
call now parses before the shared XML pass, so markup quoted in its
parameter stays data. The open-parameter scan lets the parameter's own
close tag decide, so any number of literal function closes inside one
value stay data, restoring the pre-close-scan behavior for multi-close
arguments. The leading-Mistral guard tolerates a visible preamble, with
the leading-bare-JSON guard running first so a trigger quoted inside a
leading JSON object stays data. The bare-JSON strip requires the
parser's top-level name in every mode, so nested-name JSON answers
survive name-agnostic stripping.
* Keep buffering long wrapper-less Gemma tool names instead of leaking the prefix
The streaming buffer stopped holding a call:NAME prefix at a fixed
32-char cap, so a Gemma wrapper-less call to a tool whose name exceeds
that (OpenAI allows 64 chars, MCP names run longer) streamed its raw
call:longname text as visible content before the end-of-turn parser
executed it. Hold the variable-length prefix while it still matches the
call: shape, bounded like the bare-JSON path and self-terminating into
prose, draining once the opening brace arrives.
* Keep prose that only mentions DeepSeek/Kimi markers in the route display strip
The route-level _TOOL_XML_RE DeepSeek/Kimi arms consumed from an opener up to
the end of text whenever the marker appeared, so an answer that merely refers
to a marker (for example "See <|tool_call_begin|> in the docs") had the rest
of the reply truncated. The parser-level _TOOL_ALL_PATS already gates these
arms with a call-shaped lookahead. Mirror it here so a marker is only stripped
when a real call follows it or it is a bare fragment at end of text.
* Tighten tool-calling parser and backend comments
* Pass trust_remote_code when reloading native tokenizers
The native-template fallback re-fetches a model's native chat template from
its repo when an Unsloth override template drops the tools schema. The
secondary AutoTokenizer.from_pretrained threaded hf_token but not
trust_remote_code, so for a model loaded with trust_remote_code=True whose
tokenizer repo carries custom code the reload raised, was swallowed, and the
request silently kept the tool-dropping prompt for a model that supports tools.
Store the loaded trust_remote_code on each backend's per-model info dict and
source it in render_native_template, so the reload re-uses exactly the consent
granted at load. For a LoRA adapter the reload targets the base model, whose
remote code was gated and loaded under the same stored flag, so re-passing it
executes no unconsented code. Falsy stored flag preserves the prior behaviour.
Adds a regression test that fails without the flag (custom-code reload raises,
returns None) and passes with it (tools-advertising native prompt returned).
* Treat <|python_tag|> as an outer marker envelope
A Llama-3 <|python_tag|> tool call (built-in NAME.call(...) or custom
{json} form) whose argument quotes a complete DeepSeek/Kimi example was
hijacked by the DeepSeek/Kimi marker pre-pass: the embedded example (for
example delete_all) executed instead of the real outer call. python_tag
is Llama-3's tool-call envelope, so a marker quoted inside its arguments
is data, the same as for <tool_call>, <function=...>, bare JSON, Mistral
and wrapper-less Gemma, which the guard already covers.
Add <|python_tag|> to _OUTER_ENVELOPE_OPEN_RE with a call-shaped
lookahead (mirroring the _TOOL_ALL_PATS python_tag arm) so the marker
pre-pass is suppressed when a python_tag call opens before the first
marker, while a bare prose <|python_tag|> mention is left untouched.
* Tighten tool-call parser comments
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@users.noreply.github.com>
* studio: tool calling for Llama-3, Mistral, Gemma 4 on safetensors + MLX (#5615)
Adds tool calling for Llama-3, Mistral (pre-v11 + v11+ + [ARGS]), and Gemma 4 to the safetensors / transformers and MLX backends. Parser patched against llama.cpp / vLLM / SGLang per-family parsers and normalises to OpenAI shape. 96 targeted unit tests + cross-OS staging CI (ubuntu / macos-14 / windows) green on the multi-format probe.
* studio: tool-call healing parity between safetensors / MLX and GGUF
After the multi-format parser landed in #5615, the safetensors / MLX
agentic loop and the GGUF loop still differed on healing behaviour.
This commit closes the gaps in both directions so the two backends
react the same way to identical model output.
Changes:
1. core/inference/llama_cpp.py -- the GGUF BUFFERING state machine
now wakes on every emission marker the shared parser knows. Was
("<tool_call>", "<function="); is now the five-tuple imported
from core.inference.tool_call_parser (Qwen / Qwen3.5 / Llama-3
<|python_tag|> / Mistral [TOOL_CALLS] / Gemma 4 <|tool_call>).
Stream cleanup is delegated to the same shared strip_tool_markup
so leaked markup from any family is removed from assistant
content.
2. core/inference/llama_cpp.py -- per-tool canonical heal key. When
a tool arguments field is a bare string and JSON parsing fails,
the GGUF path now heals to {"code": raw_args} for python,
{"command": raw_args} for terminal, and {"query": raw_args} for
everything else. Was hard-coded to {"query": raw_args}, which
silently routed every python / terminal emission through
web_search. Mirrors safetensors_agentic._CANONICAL_HEAL_ARG.
3. core/inference/safetensors_agentic.py -- re-prompt on plan-
without-action. When the model emits a short forward-looking
intent ("I'll search for that", "Let me check", "First, I
will...") and no tool call, the loop nudges the model to act
instead of silently returning a plan-only answer. Up to
_MAX_REPROMPTS=3 (matches GGUF). The intent regex, character
cap, and instruction text are byte-identical to the GGUF path.
The buffer-end fall-through is unified so a buffered intent
emission that never exits the BUFFERING state still triggers
the re-prompt.
4. core/inference/safetensors_agentic.py -- extra iteration slots
for re-prompts. The loop now budgets max_tool_iterations +
_MAX_REPROMPTS + 1 total iterations and tracks the tool-call
count separately, so a stalling model can be nudged 3x without
eating the caller's tool-call budget. Mirrors the _extra slot
reservation in the GGUF path.
Tests (14 new safetensors-side units; 5 GGUF parity pins):
TestLoopRePrompt -- intent-trigger, plain-answer,
no-tools, cap-at-three, budget
preserved, buffer-end intent.
TestLoopCanonicalHealKey -- python / terminal / unknown.
TestGGUFSafetensorsHealingParity -- shared markers used, shared
strip used, canonical heal keys
identical, intent regex matches
same phrases, _MAX_REPROMPTS
equal on both backends.
All 110 targeted tests pass locally; the broader tool / inference /
model-config / sandbox / anthropic / mlx suites stay green.
Why this matters
Without this parity, Llama-3.2 / Mistral / Gemma 4 emissions on Mac
(MLX) and Linux-safetensors stop the agentic loop as soon as the
model says "Let me...", because the GGUF re-prompt logic never
existed on these backends. The two-marker GGUF BUFFERING tuple also
let non-Qwen tool emissions stream out as plain prose when
llama-server's structured channel did not pick them up. Both paths
now drain the same way, heal the same way, and re-prompt the same
way -- so a tool call that works on GGUF works identically on
safetensors / MLX.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fix tool-call parser bugs from gemini review on #5620
Three high-priority gemini findings on the tool-call parsing additions:
1. unicode_escape on UTF-8 bytes corrupts non-ASCII literals
(e.g. ✨ becomes â\x9c¨). Replace with json.loads on a quoted
string -- preserves emoji / CJK / RTL while still handling
\n \t \uXXXX escapes.
2. Llama-3 sentinel stripping is order-dependent. A leading
`<|eot_id|><|begin_of_text|>` left `<|begin_of_text|>` behind
because the loop had already passed that sentinel. Loop until
no sentinel matches at the start.
3. Mistral v11+ `[TOOL_CALLS] name { json }` regex uses non-greedy
`\{.*?\}` which truncates at the first `}` of a nested JSON
argument, leaking the tail (e.g. `}}`) into user-visible
streamed text. Same problem for the v0.3 array pattern with
nested brackets. Strip those with balanced brace/bracket
scanning via a new `_strip_mistral_closed_calls` helper called
from `strip_tool_markup`.
Also fix the inference routes' parallel `_TOOL_XML_RE`:
- Same nested-JSON truncation in the Mistral patterns; route the
strip through the parser's balanced-scan helper via a thin
`_strip_tool_xml` wrapper that all existing callers now use.
- Llama-3 `<|python_tag|>[^\n<]*` stopped at any `<`, leaking the
tail of any tool call whose argument contained a literal `<`
(queries, code snippets). Relax to `[^\n]*` which keeps the
strip confined to the actual end-of-line.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio/routes: make python_tag strip multi-line aware
Earlier revisions of _TOOL_XML_RE in studio.backend.routes.inference
oscillated between two bug shapes:
5615 r"<\|python_tag\|>[^\n<]*" -- stopped at any literal "<"
so code='if x < 10: pass'
leaked '< 10: pass)' to the
user.
5620.1 r"<\|python_tag\|>[^\n]*" -- single-line only; the second
line of
python.call(code="a\nb")
leaked.
The full parser (_parse_llama3_python_tag) already handles both via
balanced-brace scanning, so the parsing path was fine; the LEAK was
in the streaming strip path that runs on every cumulative emission
while content is still arriving.
Switch to r"<\|python_tag\|>(?:[^<]|<(?!\|))*" so the strip consumes:
* any character that is not a "<" (newlines, JSON, code, ...),
* a "<" only when it is NOT followed by "|" (i.e. NOT a Llama-3
sentinel start like <|eot_id|>, <|eom_id|>, <|begin_of_text|>).
This means:
* code='if x < 10' stays inside the strip (5615 fix preserved),
* multi-line code stays inside the strip (5620 round 2),
* the strip terminates at the next Llama-3 sentinel so trailing
assistant content survives.
Tests: TestRoutesPythonTagStrip (8 cases)
pytest test_safetensors_tool_loop.py test_safetensors_capability_advertise.py
-> 118 passed in 1.81s (was 110).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: tighten verbose comments in tool-call parser sections
Comments were narrating what the code already says. Cut historical
"earlier revisions used X, then Y" narratives down to one-line WHY
notes where the footgun still matters (canonical heal-key parity,
balanced-brace vs non-greedy regex, ``(?:[^<]|<(?!\|))*`` over
``[^\n<]*``/``[^\n]*``). Drop section-header banners.
No behaviour change. Re-ran:
pytest studio/backend/tests/test_safetensors_tool_loop.py \
studio/backend/tests/test_safetensors_capability_advertise.py -q
-> 118 passed.
Regression replay (parser + _coerce_arguments on the 5 #5615 inputs)
-> 21/21.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: parser robustness fixes for PR #5620
Three surgical extensions to the multi-format tool-call parser, each
covering a real fine-tune / template emission shape that the current
parser silently drops. No path narrows; all changes widen what is
accepted.
1. `_parse_tool_call_json` now accepts both `arguments` and
`parameters` keys. A Hermes / Qwen `<tool_call>{json}</tool_call>`
wrapper around a Llama-3.2 fine-tune that emits the `parameters`
key was extracting the tool name and silently discarding the
args, producing a working-shaped call with an empty payload. The
bare-JSON and python_tag paths already accepted both keys; this
path now matches them.
2. `_TC_FUNC_START_RE`, `_TC_PARAM_START_RE`, and `_TC_PARAM_CLOSE_RE`
now also match the attribute form
`<function name="..."><param name="...">v</param></function>` used
by MiniCPM-5 and MiniMax-M2. Names land in either capture group,
and `</param>` is accepted as a short close.
3. `_parse_llama3_bare_json` sentinel-strip now consumes the role
label inserted between `<|start_header_id|>` and
`<|end_header_id|>` by Meta's official Llama-3.x chat template.
Without this, every assistant turn re-fed through the template
prefix `<|start_header_id|>assistant<|end_header_id|>\n\n{json}`
parsed to zero calls, so any history-with-tool-call round-trip
in production silently dropped.
Tests in `studio/backend/tests/test_safetensors_tool_loop.py`:
* `TestParserRobustness::test_tool_call_json_accepts_parameters_key`
* `TestParserRobustness::test_function_xml_attribute_form`
* `TestParserRobustness::test_function_xml_attribute_form_multi_param`
* `TestParserRobustness::test_function_xml_legacy_equals_form_still_works`
(regression guard for the existing `<function=name>` syntax)
* `TestParserRobustness::test_llama3_chat_template_round_trip`
* `TestParserRobustness::test_llama3_round_trip_all_roles`
* `TestParserRobustness::test_llama3_round_trip_with_eot_prefix`
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 118 to 125 passed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: terminate function-XML body at </function>, not just </tool_call>
`_parse_function_xml` was looking for `</tool_call>` (the Hermes
wrapper) as the body terminator. When a model emits a standalone
`<function=NAME><parameter=K>v</parameter></function>` followed by
explanatory prose (which models routinely do), no `</tool_call>` is
present, so the body extended to end-of-string and the trailing
prose leaked into the LAST parameter value.
Pre-existing on main (the legacy `<function=NAME>` form had this
bug too). Same affects PR #5620's new attribute-form
`<function name="NAME"><param name="K">v</param></function>`
emission used by MiniCPM-5 / MiniMax-M2.
Fix: `_TC_END_TAG_RE` now matches either `</tool_call>` OR
`</function>`. The existing `_TC_FUNC_CLOSE_RE` / `_TC_PARAM_CLOSE_RE`
strips are unchanged. Multi-call inputs still bound each function
at the next `<function=` start, so no over-eager consumption.
New tests:
* `test_function_xml_followed_by_prose` (legacy form + prose)
* `test_function_attribute_xml_followed_by_prose` (attribute form + prose)
Existing `test_code_with_embedded_xml` still passes (a parameter
value containing literal `<a></a>` is preserved because the
embedded close tag is `</a>`, not `</function>`).
`pytest studio/backend/tests/test_safetensors_tool_loop.py
studio/backend/tests/test_safetensors_capability_advertise.py -q`
goes from 125 to 127 passed.
* Studio: tighten Llama-3.2 bare-JSON guard
A fuzz pass on PR #5811 turned up that ``_parse_llama3_bare_json``
accepted ``parameters`` as a string, contradicting the docstring's
"parameters or arguments is a dict" guard. Prose JSON like
``{"name":"foo","parameters":"a sentence"}`` would wrongly fire the
parser, which the agentic loop would then heal into a real
``foo(query="a sentence")`` call.
Same code lives on this branch, so the same fix applies here.
Tightened guard:
- ``parameters`` must be a dict (Llama-3 spec).
- ``arguments`` may be a dict, or a JSON-encoded string that
decodes to a dict (OpenAI shape, e.g.
``"arguments":"{\"q\":\"x\"}"``). Plain non-JSON strings or
JSON-strings of lists / scalars / null no longer pass.
Mirrors the fix landed in PR #5811 commit 615b8608. Adds the same
4 regression tests under TestParserMultiFormat.
Existing test suite stays green: 127 -> 131 passing.
* studio: fix safetensors tool-call parser gaps vs llama.cpp (Mistral CALL_ID / THINK, attribute-form signal)
Three GGUF-parity fixes to the safetensors tool-call parser, each matching
llama.cpp's reference behaviour:
- Mistral Small 3.2 emits [TOOL_CALLS]name[CALL_ID]<id>[ARGS]{json}. The
parser stopped after the name on seeing [CALL_ID] (neither [ARGS] nor {),
dropping the call. Skip an optional [CALL_ID]<id> segment in both the
parse and strip paths. llama.cpp parses this (test-chat.cpp:4785).
- Magistral wraps reasoning in [THINK]...[/THINK]. A [TOOL_CALLS] inside the
reasoning was parsed as a real call, producing a phantom call. Strip a
leading [THINK] block before scanning so only the post-reasoning call
counts (test-chat.cpp:2285); a literal [THINK] inside a later argument is
left intact.
- The standalone MiniCPM-5 / MiniMax-M2 <function name="..."> attribute form
parsed correctly but was absent from TOOL_XML_SIGNALS and the markup strip
patterns, so the streaming safety-net parse was gated off (dropping the
call) and markup leaked into displayed text. Add the signal and broaden
the strip regexes.
Adds regression tests for all three.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: fire safetensors tool calls for the bare-JSON (Llama-3.2) form
The agentic loop's streaming safety-net parse was gated on
has_tool_signal(), which is False for the Llama-3.1 / 3.2 bare-JSON tool
form {"name":..,"parameters":..} (no XML marker). Real tool calls were
therefore dropped: the loop logged "model planned without calling tools",
re-prompted three times, then gave up with zero tool calls, while GGUF's
llama-server parses the same emission natively.
Run parse_tool_calls_from_text() unconditionally in the safety net. The
parser is strict (only fires on a valid tool-call shape) so plain answers
are unaffected. Reproduced on a real unsloth/Llama-3.1-8B-Instruct run:
the model emits {"name":"web_search","parameters":{...}} which now
executes the tool instead of being re-prompted into a no-op.
Adds a loop regression test for the bare-JSON form.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: complete strict-mode contract and fix parser import paths
Address review findings on the multi-format tool-call parser:
- Honor allow_incomplete=False in the remaining sub-parsers. The Llama-3
<|python_tag|>NAME.call(...) parser, the pre-v11 Mistral [TOOL_CALLS] array
parser, and the Gemma 4 <|tool_call> parser ignored strict mode, so a
truncated call (missing closing paren, ], or <tool_call|>) was still healed
and executed with Auto-Heal disabled. Thread strictness through and reject
the unclosed forms, matching the JSON and function-XML paths.
- Drop the duplicate tool_call_parser import block in llama_cpp.py and the
redundant un-aliased TOOL_XML_SIGNALS; only the _SHARED_TOOL_XML_SIGNALS
alias is used as a value.
- Import _strip_mistral_closed_calls from core.inference.tool_call_parser in
routes/inference.py instead of studio.backend.core... The self-contained
run.py launch mode only puts studio/backend on sys.path, so the absolute
package path raised ModuleNotFoundError on the server-tool strip path.
Add strict-mode regression tests for the truncated Llama-3 dot-call and the
unclosed Mistral array.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: preserve XML param indentation and alias Mistral array parameters
Two parser-correctness fixes found by auditing against the model chat templates
and the SGLang / vLLM reference parsers:
- Qwen3.5 XML parameter values lost their leading indentation. The chat template
emits <parameter=k>\nVALUE\n</parameter>, but the parameter-start regex ate the
wrapping newline AND the value's first-line indentation with a trailing \s*,
then str.strip() removed the rest. Narrow the trailing class to horizontal
whitespace only and trim exactly one wrapping newline (via _trim_param_value),
preserving indentation in code/diff arguments. Matches SGLang's qwen3_coder
detector. Applies to both _parse_function_xml (tool_call_parser.py) and the XML
path in tool_healing.py.
- Mistral pre-v11 array objects keyed on parameters dropped their payload.
_consume_mistral_call read only the arguments key; alias parameters the same way
the JSON/XML paths and SGLang's base detector do.
Add regression tests for preserved multi-line indentation and the array
parameters alias.
* Studio: tighten tool-call parser comments
Make the comments in the multi-format tool-call parser and its callers succinct:
compress verbose docstrings/blocks to one or two lines, drop ones that restate the
code, and trim the tiny balanced-scanner helpers. Correctness rationale and
upstream provenance (SGLang/llama.cpp parity, the strict-mode / Auto-Heal
contract, whitespace-preservation, and the Unicode / full-width-pipe notes) are
kept in compact form.
Comment-only: no code or behavior change (verified with comment_tools.py check
--strip-docstrings; parser suite green).
* Studio: make Llama-3 .call and Mistral-array healing parsing linear
Two more O(n^2) ReDoS paths in the multi-format parser, both reachable from
the agentic loop on a long truncated body with no length cap:
- _LLAMA3_KV_RE.finditer over a .call(...) body retried at every offset of a
long word run / unterminated quote (40K -> 14s). Replace with a hand-scan
that reuses the same key/number/literal sub-regexes via anchored match and
walks the string body by hand, so an unterminated quote is O(n). Verified
byte-identical to the old regex over 200K fuzzed inputs.
- _parse_mistral_array healing ran _balanced_brace_end from every { in the
body (20K -> 17s). Walk top-level objects, advancing past each balanced
{...}; this also drops the phantom call the old scan emitted from a nested
argument object.
Add adversarial-length linearity regressions plus positive .call kwargs and
unclosed-array recovery coverage.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: honor strict mode in safety-net, keep empty Gemma args, strip attribute-form function XML
- safetensors safety-net parser now forwards allow_incomplete=auto_heal_tool_calls,
matching the draining path, so a late incomplete tool call is not healed and
executed when Auto-Heal is off.
- Gemma empty bare value ({k:}) now serialises as "" instead of invalid {"k":},
which previously dropped the whole call.
- Route _TOOL_XML_RE also strips the <function name="..."> attribute form
(MiniCPM-5 / MiniMax-M2) so it no longer leaks to the UI.
* Studio: fix attribute-form function-XML literal close tag and zero-arg strict call
Addresses Codex review of the <function name="..."> attribute form in
_parse_function_xml (MiniCPM-5 / MiniMax-M2):
- End the call body at the LAST </function> / </tool_call> within the call's
window, so a literal close tag inside a code/search argument (e.g.
print("</function>")) is preserved instead of truncating the call.
- Accept a closed call with no parameters as a valid zero-argument call in strict
mode (the function close is already required), instead of rejecting it as a
truncated call.
- Tests for both, mirroring the legacy <function=...> coverage.
* Studio: fix tool-call parser/loop review findings on the multi-format path
Address the live code-review findings on the safetensors/MLX + GGUF tool path:
- routes: include the attribute form <function name="..."> in the safetensors
capability whitelist so MiniCPM-5 / MiniMax-M2 templates keep the tool pill
(parser already handles the form; the post-filter wrongly suppressed it).
- safetensors loop: build the plan-without-action re-prompt from the active
tools instead of a hardcoded web_search/python string, and gate it on
auto_heal_tool_calls, matching the GGUF loop.
- safetensors loop: hold a leading bare-JSON object ({"name":..,"parameters":..})
during BUFFERING until it closes, then drain it as a tool call instead of
streaming the raw JSON to clients. The DRAINING/STREAMING resolvers still
recover a plain JSON answer, so this can never drop content.
- parser: anchor the Llama-3 <|python_tag|>NAME.call(...) scan to the tag and
chain ; -separated calls, so all semicolon-separated built-ins parse and a
literal <|python_tag|>x.call(...) inside a JSON string argument no longer
fires the wrong tool.
- parser: consume the optional trailing </s> after a named Mistral
[TOOL_CALLS]name{json} call, mirroring the array shape.
- GGUF streaming strip: use the shared parser patterns (which know
[TOOL_CALLS] and <|python_tag|>) so a textual tool call entering DRAINING is
stripped instead of leaking the marker to streaming clients.
- routes: hoist the _strip_mistral_closed_calls import to module level.
Adds regression tests covering each fix; existing parser suite stays green.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: harden multi-format tool-call detection from review findings
Apply five targeted fixes from the review pass over the multi-format tool
path:
- routes: route display strip delegates to _strip_tool_xml so Mistral
[TOOL_CALLS] blocks with nested JSON are removed from streamed display
text, not just the XML forms.
- tool_call_parser: skip function/parameter starts that fall inside an
already-open parameter block (_inside_open_parameter) so nested example
payloads are not mis-parsed as new calls; extract
strip_llama3_leading_sentinels so the bare-JSON guard is shared.
- safetensors_agentic: probe bare JSON through strip_llama3_leading_sentinels
before the balanced-brace check so a leaked header sentinel does not defeat
the guard.
- tool_healing: allow dotted tool names in the Gemma wrapped start pattern.
- llama_cpp (GGUF): buffer wrapper-less Llama-3.2 {"name":..} calls that carry
no XML signal, drain a complete object silently and hold an incomplete one,
and run the end-of-stream safety net unconditionally so markerless calls are
detected and never leak the raw JSON (including truncated fragments).
Adds regression tests for the GGUF bare-JSON streaming path and the Mistral
display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: stop bare-JSON tool calls leaking at EOF, oversized, and into history
The second review pass flagged that the Llama-3.2 bare-JSON tool-call handling
still leaked raw JSON in several spots; ``strip_tool_markup`` only knows
XML/bracket markup, so the bare-JSON form survived it. Fix them symmetrically
across the safetensors and GGUF loops:
- Safetensors stream-end resolver now routes a held bare-JSON fragment to
DRAINING (mirroring GGUF) so a truncated ``{"name":..`` cut off by the end of
the stream is dropped instead of flushed as assistant content. The 7/10
reviewer finding.
- Both loops now drain (suppress) an oversized still-open bare-JSON call once it
passes ``_MAX_BARE_JSON_BUFFER`` instead of streaming the raw prefix, gated on
a ``"name"`` key so a giant plain JSON answer still streams; a complete
oversized call still executes via the safety net.
- Add a shared ``strip_leading_bare_json_call`` helper and apply it to the
content kept for the assistant turn in both loops, so an executed bare-JSON
call is not replayed as visible text or fed back as next-turn history.
Plain JSON answers without a ``"name"`` key are untouched throughout. Adds
regression tests for the EOF, oversized, and next-turn cases on both backends
plus unit tests for the helper.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: bound the Llama-3 python_tag strip on real control sentinels
The route display strip's <|python_tag|> arm ran to the next <| of any kind.
A tool-call argument carrying a literal <|...|> token (for example <|cite|>
inside a string value) truncated the strip early and leaked the call tail into
the visible response. Narrow the stop condition to the genuine Llama control
sentinels (eot_id, eom_id, python_tag, start/end_header_id, begin_of_text,
finetune_right_pad_id) so embedded markup and JSON are consumed while real
header/turn boundaries still bound the strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: gate markerless bare JSON on enabled tools and close parser/strip asymmetries
The Llama-3.2 custom_tools bare-JSON form has no marker, so any JSON object with a
name key was read as a tool call. An ordinary JSON answer like
{"name":"Alice","parameters":{"age":30}} was misclassified as a call to a
disabled tool and dropped from the visible response. Gate the markerless form on
the enabled tool names (threaded through parse_tool_calls_from_text and
strip_leading_bare_json_call, supplied by both streaming loops): an object whose
name is not an enabled tool is ordinary content. The marker-based forms keep
their name-agnostic behaviour (an explicit signal is a real call attempt), and
unrestricted mode stays ungated.
Also fix two parser/strip asymmetries the parser already tolerated:
- A literal </function> inside a parameter value (print("</function>")) truncated
both the core and route strips at the first close, leaking the tail. Extend the
strip to the call's real close (last </function> before the next opener),
mirroring the parser, without merging separate calls.
- The single-object Mistral [TOOL_CALLS]{...} shape parsed but _strip_mistral_closed_calls
left it, leaking the raw object into display. Strip the balanced object while
keeping trailing prose, matching the array and name shapes.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: gate GGUF bare-JSON suppression on enabled tools and fix python-tag exponent parsing
Pass-4 review follow-ups on the GGUF tool loop and Llama-3 parser:
- The GGUF bare-JSON suppression sites still keyed off a raw "name" substring,
so an ordinary JSON answer whose name is not an enabled tool was dropped when
it was truncated, oversized, or reached the no-tool DRAINING fallback (the
parser, helper, and safetensors paths were already gated). All three sites now
use the shared enabled-name gate, and a held bare-JSON buffer that turns out not
to be an enabled call is shown as the answer instead of dropped at stream end.
- The Llama-3 python-tag numeric kwarg regex matched only the mantissa, so
scientific notation was truncated to its leading digits (1e-3 parsed as 1) and a
tool executed with the wrong value. The regex now accepts exponent and decimal
forms, and the int/float classification keys off the exponent too.
Adds regression tests for the truncated / oversized disabled-name JSON cases (and
a counterpart that a truncated enabled call still does not leak) plus the
scientific-notation kwargs.
* Studio tools: gate safetensors bare-JSON drain, fix nested-name gate and function-XML strip
Pass-4 review follow-ups on the shared parser / safetensors loop:
- The safetensors oversized and end-of-stream bare-JSON drain branches keyed off
a raw "name" substring, so a large or truncated ordinary JSON answer whose name
is not an enabled tool was drained instead of streamed. Both now use the shared
enabled-tool-name gate, matching the GGUF path.
- strip_leading_bare_json_call matched the first "name" anywhere, so a plain JSON
answer with a nested name equal to an enabled tool ({"result":{"name":"web_search"}})
was wrongly suppressed. It now extracts the TOP-LEVEL name only, walking past
nested objects/arrays and keeping the text when a top-level value is truncated.
- The function-XML display strip used a regex negative-lookahead that stopped at a
literal <function=...> opener inside a parameter value and then dropped the rest
of the answer to EOF. A scan-based strip mirrors the parser (ignores openers
inside an open <parameter> via _inside_open_parameter) and closes each call at its
real </function>, so trailing assistant text after such a call survives.
Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tool parsing: 3.9 import safety, disabled-Auto-Heal contract, capability gate
Round-2 review follow-ups on the multi-format tool-call parser:
- tool_call_parser: add `from __future__ import annotations`. The module
is dependency-light by design (external llama-server wrappers import it
standalone) and the package targets python >=3.9, where its PEP 604
`int | None` return annotations would raise TypeError on import.
- safetensors + GGUF drain fallback: gate the leading bare-JSON strip on
auto_heal_tool_calls. With Auto-Heal off, a truncated enabled-name
fragment that did not parse now stays visible, matching the XML strip
in the same branch and the disabled-Auto-Heal contract. With Auto-Heal
on it is still suppressed.
- safetensors capability gate: match the bare-JSON `{"name":` template
marker with a whitespace/escape-tolerant regex so a pretty-printed
`{ "name" :` or JSON-escaped `{\"name\":` template is not mis-classified
as tool-less. The parser already accepts that whitespace via
raw_decode, so the gate must too.
Regression tests added for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Tool parsing: symmetric "function" bare-JSON alias and route strip parity
Round-3 review follow-ups, all parser/strip symmetry fixes.
- Bare-JSON "function" alias: the markerless parser accepts a call name via
obj.get("name") or obj.get("function"), but the strip/gates only knew "name",
so a {"function":<enabled tool>} call executed while its raw JSON leaked. Teach
_top_level_bare_json_name the alias (with "name" precedence and the same nested
and truncated-name guards), and widen the guards in strip_leading_bare_json_call,
the safetensors and GGUF _looks_like_enabled_bare_json gates, and the route
capability marker regex.
- Route display/history cleanup: strip a tail-only </param> alias close (the
parser accepts <param name="...">...</param>), and run the parser's guarded
function-XML scan (_inside_open_parameter) before _TOOL_XML_RE so a literal
nested <function=...></function> inside an argument value does not truncate the
strip and leak the tail.
Regression tests added for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: honor tool budget in GGUF loop and guard function-XML streaming strip
Round 4 review fixes. Both are asymmetric-fix bugs where the final/steady path got a
guard the analogous streaming/loop path did not.
- GGUF tool-call budget: the safetensors loop counts real tool-call turns against
max_tool_iterations (re-prompt stalls excepted), but the GGUF loop only bounded the
turn count by the enlarged range (max_tool_iterations + _MAX_REPROMPTS). Since this
PR raised _MAX_REPROMPTS from 1 to 3, a model that keeps making valid tool calls
could run up to three extra tool rounds (with max_tool_iterations=1, four rounds
instead of one). Add a _tool_iters_done counter that increments only when a tool
actually executed in the turn, and stop once the caller's budget is spent so the
post-loop final-answer nudge fires. A duplicate/disabled no-op turn is a correction
turn (like a plan-without-action re-prompt) and does not consume budget, preserving
the existing "already completed" re-prompt behavior.
- Streaming display strip: the final strip runs the guarded _strip_function_xml_calls
scanner (a literal <function=...> inside a parameter value is data, not a nested
call), but the GGUF and safetensors streaming strips still used only the open-ended
regex arms. When a tool-call argument contained literal function markup, the regex
tail ate everything to end-of-text and dropped the real trailing prose after the
call's true </function>. Run the guarded scanner (and the balanced Mistral strip)
before the regex arms in both streaming paths so streaming and final display agree.
Adds regression tests: GGUF valid tool calls respect max_tool_iterations, and the
streaming strip keeps trailing prose after a function-XML call with a literal marker.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio tools: safetensors tool budget counts only executed turns (GGUF parity)
Follow-up to the GGUF budget fix. The safetensors loop charged max_tool_iterations
per non-re-prompt iteration (iteration + 1 - reprompt_count), so a duplicate/disabled
no-op turn spent a budget slot even though no tool ran. With a small cap this dropped
real work: for max_tool_iterations=2, a model that made a valid call, repeated it (an
internal no-op correction turn), then made a distinct valid call executed only the
first -- the third turn was sent with no tools and the distinct call was ignored.
Track whether a turn actually executed a tool (set on record_result) and count only
those turns against the cap, matching the GGUF loop. A duplicate/disabled no-op is a
correction turn -- like a plan-without-action re-prompt -- and no longer consumes
budget, so the model still gets its "already completed" nudge and another tool-enabled
turn. Adds a regression test for the small-cap duplicate-then-distinct-call flow.
* Studio: render the reasoning block for safetensors and MLX like GGUF
enable_thinking chat templates (Qwen3/Qwen3.5/GLM) prefill an unclosed <think>
into the generation prompt, so the model emits only the closing </think> then
the answer. The safetensors/MLX chat stream emitted that as plain content, so
the reasoning showed inline with no collapsible thinking block, while GGUF
(which surfaces reasoning via reasoning_content) rendered one. This brings
safetensors and MLX to parity.
- _ResponsesReasoningExtractor gains a reasoning_prefilled mode that starts
inside the reasoning block and splits on the first </think>; default False
keeps GGUF and every existing caller byte-identical. It suppresses a stray
re-emitted <think> and holds partial markers back across chunk boundaries.
- _sf_reasoning_prefill_mode gates the mode on reasoning being enabled for the
request, an enable_thinking or enable_thinking_effort style, and the template
actually using the standard <think>/</think> markers. Models with a bespoke
reasoning channel (e.g. gemma's <|think|>/<|channel>) are excluded so their
answer is never swallowed; gpt-oss (Harmony) and thinking-off requests are
excluded too.
- sf_tool_stream and stream_chunks (the latter also serves MLX) feed text
through the extractor, emitting reasoning_content then content deltas, with a
per-turn reset in the tool loop and a flush before each tool_start; only the
visible delta reaches the monitor reply. The two non-streaming drains split
reasoning_content the same way.
- Tests: extractor prefilled mode (streaming and edge cases), the gate matrix
including the gemma-style exclusion, and a route-replay of the tool-loop
reasoning stream.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* studio: don't force a tool re-prompt on a negated intent (safetensors parity)
The safetensors _INTENT_SIGNAL claimed to mirror GGUF but was missing the
negative lookahead, so a refusal like "I will not search the web for that"
matched the "i will" intent and triggered the plan-without-action re-prompt
(STOP... you MUST call a tool), overriding a valid no-tool answer. GGUF already
excludes not/never. Add the same (?!\s+(?:not|never)\b) lookahead so both
backends agree. Extends the intent parity test with negated refusals.
* Studio: trim redundant comments (comment-only, AST-verified)
* Studio: prevent Gemma tool-parser DoS on stray delimiters
_gemma_parse_value returned the input index unchanged when text[i] was a
stray delimiter (,}]), so the list and mapping caller loops that advance
on the returned index spun forever at 100% CPU on malformed input such as
[},]. Advance past the delimiter so parsing always terminates.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: strip Magistral [THINK] reasoning from final display/history
strip_tool_markup removed [TOOL_CALLS] and <function> markup but left a
leading Magistral [THINK]...[/THINK] block intact, so its bracket-form
reasoning (not the <think> the reasoning channel renders) leaked into the
safetensors display and conversation history while GGUF/llama.cpp routes
it natively. Drop the leading reasoning block at end-of-turn (final=True)
via the existing _strip_mistral_reasoning helper; streaming is untouched.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Honor reasoning_effort none in safetensors prefill; strip Magistral reasoning while streaming
Two safetensors/MLX reasoning fixes surfaced in review:
_sf_reasoning_prefill_mode only checked enable_thinking, so an
enable_thinking_effort (GLM-5.2) request that disables thinking via
reasoning_effort=none (without enable_thinking=False) still began in
prefilled-<think> mode. A plain answer with no </think> was then swallowed
whole into reasoning_content and the visible response came back empty. Thread
reasoning_effort into the predicate and treat none as disabled, mirroring
_request_reasoning_kwargs.
strip_tool_markup_streaming stripped tool markup but not the leading Magistral
[THINK]...[/THINK] bracket block, so the raw chain-of-thought leaked into the
streamed safetensors content instead of the reasoning drawer (GGUF routes it
natively). Apply _strip_mistral_reasoning first, matching the final strip; an
unclosed [THINK] is held from the marker on so nothing flickers.
* Mistral outer call wins over XML literals; align healer signals with its parser
Two follow-ups on the shared-parser ordering after the healing-passthrough
merge:
- A well-formed [TOOL_CALLS] call whose JSON arguments quote tool XML parsed
the literal instead of the outer call (executing the wrong tool). When the
first XML signal sits inside a leading balanced Mistral body it is argument
data, so the Mistral parser now runs first; an XML signal before the trigger
keeps the normal order, so a [TOOL_CALLS] literal inside an XML call's
arguments still stays data.
- passthrough_healing buffered streams on the parser module's broadened signal
list (now including <|python_tag|> and [TOOL_CALLS]) but promotes with
core.tool_healing, which does not parse those forms: a streamed Mistral or
Llama text call was held until finalization and flushed as prose. The healer
keeps its own signal list limited to the formats it can promote, restoring
immediate streaming for the rest.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: leading envelopes win over rehearsed literals
- New _first_foreign_tool_signal shared by the leading-envelope guards adds
<|python_tag|> to the protected signal set: the spelled-out literal inside a
Mistral call's arguments (a query about Llama built-in tool syntax) executed
the inner literal instead of the outer call.
- New _xml_signal_inside_leading_bare_json guard, sibling of the Mistral one:
a leading bare-JSON call whose string argument quotes tool XML (a code value
citing <function=...>) had the literal promoted by the shared XML pass
before the bare-JSON parser ran.
- Magistral [THINK]...[/THINK] is dropped once at parse entry instead of only
inside the Mistral parser, so a call rehearsed in the think block in a
foreign format can no longer be promoted while the real call after the
block is lost. Parse now agrees with the display strip.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: a disabled leading bare-JSON object keeps its literals as data
When the leading bare-JSON object is ordinary content (name not an enabled
tool), the guard proved the first tool signal sits inside it, so falling
through to the XML/python_tag passes promoted quoted string data as a real
call. Drop the object and parse only the tail: a real call after the object
still parses, nothing inside it can be promoted.
* Address review: Mistral literals inside leading JSON, whitespace-tolerant wrapped Gemma opener
- The leading bare-JSON guard now treats the [TOOL_CALLS] trigger as a
foreign signal: the Mistral parser runs before the bare-JSON one, so a
literal quoted inside the leading object's strings was promoted over the
outer call (or over ordinary JSON content).
- tool_healing's wrapped Gemma opener tolerates whitespace around call and
the colon: sampling drift emits call: name{ and call : name{, and
rejecting those lost the call entirely because no fallback re-parses the
wrapped form. Strict mode still requires the closing tag.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: accept dotted Gemma argument keys in the key-quoting scanner
The scanner quoted keys of [alnum_-] only, so a dotted key (user.name:...)
was left unquoted, json.loads failed, and the whole wrapped call was lost
(parse empty, strip wipes the markup). Dots now match the parser's own
key/name charset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: leading Mistral call owns the turn, dotted keys after bare values
- A LEADING parseable [TOOL_CALLS] call now runs the Mistral parser first
unconditionally: literal XML in trailing prose after the call was promoted
by the earlier shared XML pass, executing the quoted example instead of
the real leading call. XML leading keeps the normal order.
- _GEMMA_NEXT_KEY_RE accepts dots so a dotted key after a bare value
(query:foo,user.name:bob) ends the value at the comma instead of being
swallowed into it, matching the round-earlier key-quoting charset.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review: markup quoted inside a nameless leading JSON answer stays data
The leading bare-JSON guard required a top-level name, so a structured JSON
answer quoting tool markup in its strings (a response_format turn
documenting a tool's syntax) had the literal promoted by the later passes.
A nameless leading object that parses as real JSON now routes through the
same decline-then-parse-the-tail path; non-JSON braced prose keeps the old
behaviour, and a real call after the answer still parses.
* Compress docstrings in the multi-format tool parser to their contract essence
* verify_import_hoist: exempt __future__ imports and same-diff relocations
Two false positives fired on this PR's refactor. A from __future__ import
is a compiler directive whose name never appears as a runtime load, so
HOISTED-IMPORT-UNUSED can never see it used, yet the file requires it for
PEP 604 annotations on Python 3.9. TARGET-CHANGED flagged the deliberate
move of the strip-pattern constants into core.inference.tool_call_parser
as a silent re-point even though the old module-level target was removed
and the new one added in the same diff. Both get narrow exemptions; a
re-point to a pre-existing target is still caught, and the self-test
negative controls all pass unchanged.
* Leading bare-JSON calls own the turn; function calls end at the first balanced close
The XML-signal guard for a leading bare-JSON call required the signal
strictly inside the object, so a trailing XML example stole the turn
from the leading call; it now applies the same inside-or-after rule as
the Mistral guard. Function-XML calls also ended at the LAST close tag,
which let prose after a closed call that mentions a literal close tag
get swallowed into the final parameter value; calls now end at the
first close tag that is not inside an open parameter, and the strip
mirrors the same rule so parse and strip agree.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Attribute-form calls end at the first balanced close; bare-JSON strip requires the call shape
The attribute form parser still kept the last close tag in the call
window, folding prose after a closed call into the final parameter
value. It now takes the first close not inside an open parameter, the
same rule the equals form and the strip already use.
The leading bare-JSON strip deleted any closed object whose top-level
name matched an enabled tool, including plain JSON answers the parser
correctly rejects as non-calls. The strip (and the drain gate that
delegates to it) now requires the parser's exact call shape, so answers
like {"name":"web_search","result":...} stream and display intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* False-alarm markers keep the answer; the bare-JSON strip consumes the whole chain
The trailing strip arms dropped everything from a bare marker to EOF,
so a normal answer that mentions [TOOL_CALLS] or another marker
literally was truncated (or fully swallowed when it started with the
literal) after the no-call drain fallback. Those arms now require a
call-shaped lookahead or marker-at-EOF before dropping; truncated real
calls still strip.
Chained bare-JSON turns executed both calls but stripped only the first
object, so the second call's raw JSON replayed into the next assistant
history message alongside the structured tool_calls. The strip now
consumes the entire chained run of call-shaped enabled objects while
non-call answers, disabled names, and trailing prose stay intact.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Attribute-form containment, parameter-close-decides rule, preamble-tolerant Mistral guard, strict strip shape
Four document-order and containment fixes. A leading attribute-form
call now parses before the shared XML pass, so markup quoted in its
parameter stays data. The open-parameter scan lets the parameter's own
close tag decide, so any number of literal function closes inside one
value stay data, restoring the pre-close-scan behavior for multi-close
arguments. The leading-Mistral guard tolerates a visible preamble, with
the leading-bare-JSON guard running first so a trigger quoted inside a
leading JSON object stays data. The bare-JSON strip requires the
parser's top-level name in every mode, so nested-name JSON answers
survive name-agnostic stripping.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Let a leading <|python_tag|> call own the turn over quoted XML literals
The leading-call ownership contract (a leading executable call owns the turn;
foreign markup quoted in its string arguments or trailing prose stays data) was
enforced for the bare-JSON, Mistral and attribute-form leading calls but not
for the Llama-3 <|python_tag|> form. The shared tool_healing XML pass runs
before _parse_llama3_python_tag and does not recognise <|python_tag|>, so a
<function=...> / <tool_call> / [TOOL_CALLS] literal quoted inside a
<|python_tag|> .call(...) string argument (or its JSON parameters) was promoted
and the wrong tool executed. Well-formed single-format examples:
<|python_tag|>web_search.call(query="... <function=foo> ...") -> foo
<|python_tag|>python.call(code="<function=render_html>..</function>") -> render_html
both returned the phantom inner tool instead of the real leading call.
Add a leading-<|python_tag|> guard mirroring the other leading-call guards:
when the tag is the first tool signal, parse it before tool_healing so quoted
foreign markup stays data. A foreign signal before the tag keeps normal
document order. Added TestPythonTagOuterOverXmlLiteral (7 cases).
* studio: tighten tool-calling comments to be shorter and clearer
* studio: shorten tool-format comments in changed files
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
Co-authored-by: Daniel Han <info@unsloth.ai>
Co-authored-by: danielhanchen <danielhanchen@users.noreply.github.com>
A candidate clip that decodes to fewer frames than the reference was
compared only over the shared prefix, so good early frames could PASS a
truncated or corrupt render. Record both frame counts and gate a mismatch
as FAIL, matching the harness contract that the requested shape is held
fixed. Extend the selftest with a truncated-clip case.
The latest huggingface-hub release added the Sandboxes feature. Its
bootstrap (_sandbox.py) fetches the static sbx-server binary into /tmp with
an Authorization header and marks it executable, which is exactly the
staged-dropper pattern the scanner hunts, and three while True polling loops
in _sandbox.py / hf_api.py / utils/_http.py match the beaconing heuristic.
All four verified against the official huggingface/huggingface_hub
repository: the snippet is the documented sandbox server injection and the
loops are deadline-style job and sandbox polling. Entries generated with
--write-baseline and reviewed line by line; scan_packages.py huggingface-hub
now exits 0 with the four findings suppressed.
An empty or corrupt clip decode crashed clip_metrics on frame indexing; it now
returns a zero-frame FAIL record. A NaN candidate audio RMS compared False
against every threshold and slipped past the silence trip-wire; NaN now counts
as a collapse.
The 1.22.0 release adds _sandbox.py (the client for HF Jobs sandboxes,
including the bootstrap that downloads HF's own sandbox server binary)
and the scanner flags it plus three long-standing while True retry and
pagination loops as CRITICAL. Reviewed all four against the upstream
repo and the 1.22.0 wheel: legitimate library code. Entries generated
with scan_packages.py --write-baseline and verified to suppress with
exit 0.
HunyuanVideo-1.5 (the 8.3B DiT with Qwen2.5-VL + ByT5 text encoders) loads
through the hunyuanvideo-community Diffusers repacks; tencent's own repo is
the original non-diffusers layout and cannot load as a pipeline, so only the
community 480p/720p t2v repos are trusted. Two pipeline quirks, both verified
against pipeline_hunyuan_video1_5.py in diffusers 0.39, shape the wiring:
- __call__ takes no guidance kwarg: CFG lives on the pipeline's guider
component (ClassifierFreeGuidance, shipped at scale 6.0). The family gains
guidance_via_guider and generate() writes the requested scale onto
pipe.guider instead of passing cfg_kwarg, which the pipeline would reject.
- __call__ has no callback_on_step_end: progress and cancellation fall back
to a scheduler.step wrapper (one call per denoise step), installed for the
duration of the call and always restored. Cancellation unwinds the loop by
raising through the wrapper and surfaces the same cancelled sentinel the
callback path uses.
The VAE compresses 16x spatial / 4x temporal, so sizes snap to /16 and frame
counts to 4k+1. The transformer declares _repeated_blocks and CacheMixin, so
the regional compile profile and the step cache both apply unchanged.
scripts/video_quality.py is the video accuracy gate, the analogue of
scripts/diffusion_quality.py with the same pure-numpy PSNR/SSIM math so image
and video budgets compare: fixed prompt/seed/shape, one short clip per
candidate against a reference clip, per-frame SSIM/PSNR over sampled frames,
a temporal-consistency deviation (motion-energy series error, catching
flicker SSIM alone misses), black-frame/NaN collapse checks, an audio RMS
silence trip-wire for LTX-2, and wall time + peak VRAM per candidate.
Verdicts map the standing budget: ssim >= 0.75 passes, >= 0.50 warns,
anything lower or any collapse fails. --selftest runs the metric path on
synthetic clips with no GPU or model.
* 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: fix FP8 transformer quant producing noise (per-row scaling)
FP8 dense-quant produced pure noise on models with extreme activation
outliers (z-image), while INT8 was fine. Root cause: torchao's default
fp8 granularity is per-TENSOR, so z-image's MLP activation outliers
(~6.6e4) force a tensor-wide scale that pushes every normal value below
fp8 resolution and the denoise collapses to noise. INT8 dynamic is
per-token by default, which is why it was unaffected.
Fix: request PerRow granularity (per-token activation + per-output-channel
weight) for the fp8 config, confining each outlier to its own row. The
per-row scaled_mm is probed by _smoke_probe, so an arch or build without
it falls through the ladder to int8.
Validated on B200 (z-image, 1024px, 8 steps): per-tensor fp8 = noise,
per-row fp8 = matches bf16; int8 unaffected. Adds a regression test
asserting the fp8 config carries PerRow granularity.
* 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
* [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
* fp8 prequant: reject stale per-tensor checkpoints
The per-row fp8 fix only applies when quantising dense on the fly; a supplied
transformer_prequant_path (or a future hosted fp8 repo) bypassed it, so an fp8
checkpoint baked with the old per-tensor layout still loaded and reproduced the
z-image noise failure. Stamp the fp8 granularity into the checkpoint metadata at
build time and require per-row at load time, so a stale artifact is rejected and
the loader falls back to rebuilding/re-quantising. The gate is fp8-only (int8 and
the others are unaffected). Adds regression tests.
* 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>
Memory planning and dense-quant path: size a local diffusers base's
resident companions from its on-disk VAE and text-encoder weights instead
of folding them to zero, feed the distilled variant hint into the runtime
headroom estimate so turbo and schnell models are not over-reserved, place
group-offload companions resident before attaching the transformer hooks
so a failed placement falls back to whole-module offload instead of
crashing, and bail out of the dense transformer download before it starts
when the requested quant scheme is unsupported so the load falls back to
GGUF cleanly.
sd.cpp stack: scrub the native path lease secret from sd-cli child env,
redact native load-progress errors, forward the resolved accelerator when
auto-installing a forced-native binary, release stale diffusion GPU
ownership on CPU-native loads, and remove the sd.cpp install tree on
uninstall.
Prequant and scripts: reject prequant artifacts missing base_model_id
when a base is requested, expanduser before checkpoint existence checks,
record and validate the int8 exclusion filter and fp8 fast-accum in
checkpoint metadata, make verify_prequant_backend allowlist its local
checkpoint and fail on missing or bad LPIPS and on load-peak regressions,
average only finite PSNR values in diffusion_quality, and reset the
process-wide attention backend between perf probe variants.
API and UI: normalize attention_backend casing before Literal validation,
close hidden popovers when leaving the Images page, and clear the stale
quant label when loading a direct local GGUF file.
The pip scan-packages job keys its baseline on the matched-code hash (main's
scanner). This branch's dependency set resolves a newer fastapi (its routing
while-True loop hashes differently than the baselined one) and adds gguf,
whose HF_TOKEN-authenticated download helper trips the env+network check.
Both reviewed benign: the fastapi hit is its own websocket routing loop and
the gguf hit is the official package's Hub download path. Sync the scanner
and baseline from main and add the two reviewed entries.
* 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 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 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 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
* [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
---------
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 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 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).
* [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
---------
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 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 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 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.
* [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
---------
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 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 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 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 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).
* [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
---------
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 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 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 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.)
* [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
---------
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 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 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 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.)
* [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 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 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.)
* [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 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 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.
---------
Co-authored-by: oobabooga <112222186+oobabooga@users.noreply.github.com>
* scan_packages: key baseline on matched-code hash
The baseline matched on (package, package-relative file, check), which
excluded the matched code, so a future finding of the same check in the
same file was suppressed regardless of what the code did. A malicious
future version of an already-baselined package could place a payload in
the same file under the same check and pass the enforcing gate.
Key the baseline on a hash of the matched code too. The hash is over the
deduped, sorted set of matched spans with L<NN>: line markers stripped, so
version bumps, line shifts and match reordering stay stable while new or
changed flagged code reopens the finding. Version is left out of the key so
routine dependency bumps do not reopen every entry. The hash is capped and
recomputable from the stored evidence.
Regenerate scan_packages_baseline.json against the current dependency set;
the hf-stack, studio and extras scan shards pass enforcing (no active
CRITICAL or HIGH).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: refresh baseline for newer unsloth-zoo release
A newer unsloth-zoo published after the first regenerate added
tests/test_mlx_save_export_regressions.py, a benign test fixture
(temporary_location="/tmp/ignored") that trips the /tmp dropper check.
Regenerate the hf-stack shard against the current set so the entry is
allowlisted; studio and extras are unchanged.
* scan_packages: harden baseline loading against malformed JSON
Guard against a non-dict top-level baseline and non-dict entries so a
corrupt or hand-edited allowlist warns and fails closed instead of
crashing with AttributeError, and treat an explicit evidence: null as
empty.
* scan_packages: hash the full match set, keep indentation, strip only the marker
Address the evidence-hash review feedback:
- Capture every matching line, not the first three, so a payload appended
after existing matches in a baselined file and check reopens the finding
instead of riding the sample.
- Preserve leading indentation so a flagged line moved out of a guarded block
reads as changed.
- Strip only each span's prefix up to the first L<NN>: marker, so an L<NN>:
inside the matched code is kept and a change to it reopens the finding.
Evidence and its hash are stored in full and stay recomputable from the stored
field. Regenerate the baseline; hf-stack, studio and extras pass enforcing with
no active CRITICAL or HIGH.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: bind baseline evidence to full matched code
Address review feedback on the evidence-hash baseline key:
- Split evidence only on real span delimiters (" | " before an L<NN>:
marker, or a newline), so a bitwise-or or union type in matched code
is no longer split apart into separate spans.
- Record matched lines in full (drop the 160-char per-line cap) and
record every distinct multiline match, so code appended past the cap
or a second cross-line match reopens the finding instead of riding the
first one.
- Give the large-JS-bundle and .pth base64-blob findings a content
digest instead of empty or prefix-only evidence, and record all .pth
import lines, so a changed bundle, blob or import no longer inherits a
baselined empty or truncated key.
- Warn when a loaded baseline has entries without evidence_hash so a
legacy baseline is regenerated rather than silently degraded.
Regenerate scripts/scan_packages_baseline.json against the current dep
set and add regression tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: harden multiline and duplicate evidence handling
Follow-up hardening so the evidence hash tracks the full matched code:
- For DOTALL patterns that match across lines, record every line the match
spans (not just the start line), so a change on a continuation line (the
URL inside a baselined C2 loop, a swapped credential path) reopens the
finding. A pathological greedy span is bounded to its head line plus a
digest of the rest.
- Keep duplicate spans in the canonical evidence so a second identical
matched line in a new code path changes the key instead of deduping away.
- Anchor the evidence prefix to strip only a genuine leading label or
line-number marker, leaving a marker-like "L<NN>:" inside raw .pth code
intact.
- Make the legacy-baseline warning explicit that entries without an
evidence_hash reopen rather than suppress under a coarse key.
Regenerate scripts/scan_packages_baseline.json (same finding set; entries
for same-file repeated checks are now tracked separately) and add tests.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: bind every combo and large finding to its full content
Close the remaining asymmetric-evidence gaps so a changed payload cannot
ride a reviewed baseline entry:
- Digest a capped multiline span from the code without line markers, so a
pure line shift stays stable while a continuation-line change reopens.
- Give the "Unusually large executable .pth" finding a content digest
instead of keying on byte size and import-line count alone.
- Record both contributing signals for the JS credential+network stealer,
the shell credential+network and persistence-hook combos, and the hidden
network+exec docstring payload, so changing the network/exec side reopens.
- Allow punctuation in an evidence label prefix so a "network+exec:" label
is stripped and line shifts do not change the key.
Regenerate scripts/scan_packages_baseline.json and add tests for each case.
* scan_packages: bind remaining Python combos; key npm baseline on evidence
Python scanner: the openssl+key, anti-analysis, DNS-exfil and base64+exec+blob
combos recorded only one contributing signal, so a changed payload on the other
side could ride a reviewed baseline entry. Each now binds every co-occurring
signal (and the blob is digested, since it can sit on a separate line from the
decode call).
npm scanner: scan_npm_packages.py keyed its allowlist on (package, path,
pattern) only, the same coarse-key bypass the Python scanner just closed. Add an
evidence hash to the key (schema v3, fail-closed on older baselines) and store
full evidence. The committed baseline stays empty by design.
Regenerate scripts/scan_packages_baseline.json and add tests for each case.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_npm_packages: bind full blob evidence and harden baseline loader
Follow-up on the npm evidence-hash key:
- _evidence now records every match and, when a snippet is truncated for
display, appends a digest of the full match. The obfuscated-blob key was
hashing only the truncated first-match snippet, so a changed payload tail or
an appended blob in the same package/file/pattern could ride a reviewed entry.
- _load_baseline guards that the root is an object, entries is a list, and each
entry is a dict before reading it, so a malformed baseline warns and fails
closed instead of raising AttributeError.
Add tests for a changed blob tail reopening the key and for malformed entries.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan_packages: symmetric baseline-loader guards; bind npm outbound host context
- Python _load_baseline now rejects a non-list "entries" with a warning instead
of raising TypeError, matching the npm loader.
- npm cred-surface-host (outbound) records the host with its URL path / fetch
call / host config, so a changed outbound path, headers or body reopens the
key rather than riding the bare host literal.
Add tests for both.
* scan_npm_packages: migrate v2 baselines and bind host-config outbound context
- _load_baseline now migrates schema v2 entries by recomputing the evidence
hash from stored evidence (with a legacy warning), matching the Python
loader, instead of discarding them; only pre-v2 basename schemas are rejected.
- The cred-surface-host (outbound) host-config branch now captures the whole
line (path, headers, body), so a changed outbound payload on the same
hostname line reopens the key instead of riding the bare host snippet.
Add tests for v2 migration and the host-config context binding.
* scan packages: bind PEM key bodies and npm windowed evidence to baseline keys
scan_packages: embedded-key findings now pin the full PEM block (BEGIN..END)
via a content digest, so a key body swapped under the same marker reopens the
finding instead of riding the unchanged BEGIN line. Single-line and DER keys
were already bound by their full matched line; marker-only references with no
END block (validation header lists) are unaffected, so the committed baseline
is unchanged.
scan_npm_packages: _evidence now digests the full containing line whenever the
shown snippet is only a window into it (short match on a long line, or a
truncated payload), so a changed payload tail outside the display window
reopens the key. The npm baseline is empty, so this changes no suppressions.
Adds regression tests for both cases.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan packages: bind multi-line evidence and every blob to baseline keys
_extract_evidence now extends each single-line match over its bracket
continuations, so a multi-line call binds its argument lines and a changed
URL or body on a continuation line reopens the key. After the per-line pass it
also records cross-line matches the scan cannot otherwise see (a DOTALL regex,
or a multi-line construct appended under a check that already had a one-line
match), so an appended multiline payload reopens instead of riding the key.
_blob_digest hashes every large base64 blob (not just the first) for the
base64+exec finding and the .pth large-blob finding, so an appended or swapped
second encoded payload reopens; single-blob files keep the same digest.
scan_npm_packages _evidence digests the full logical line (the matched line
plus its bracket-continuation lines), so a multi-line fetch's option and header
lines bind and a changed payload on a following line reopens the outbound key.
Regenerated the Python baseline: same package/file/check set, 24 entries pick
up the wider multi-line evidence. Adds regression tests for each case.
* scan packages: stop giant greedy spans from binding a whole-file digest
When a greedy DOTALL pattern (reverse shell socket...subprocess, C2 loop) has
its anchor tokens far apart, the match span covers the whole file. Digesting
that span bound thousands of unrelated lines, so the evidence hash drifted on
any edit between the anchors (a dependency bump reshuffling the file), which
made a baselined finding reopen on an upstream release. The multiline pass now
skips an oversized span when the per-line pass already bound the signal lines,
so the evidence is the stable matched lines; a genuinely appended multi-line
construct stays under the cap and is still recorded.
Regenerated the Python baseline against Python 3.12 (the version the scan CI
shards run) so the resolved dependency set matches CI. Same package/file/check
set. Adds a regression test.
* scan packages: tighten evidence binding (order, string brackets, span size)
Address review follow-ups on the evidence extraction:
- _canon_evidence keeps discovery (line) order instead of sorting. Line-shift
stability already comes from stripping the L<NN>: markers, so order stays
significant and reordering matched lines (a multi-line call's arguments)
reopens the finding.
- _logical_line_end (Python) and _logical_line_text (npm) blank string literals
before counting brackets, so a ) inside a string argument does not close the
logical line early and drop later argument lines.
- The oversized-span skip now only drops a giant whole-file bridge (over 60
lines); a genuinely appended multi-line construct is recorded so its payload
reopens, rather than riding an existing one-line match.
- npm _logical_line_text binds the enclosing bracket group, so a host-config
object whose { is on a prior line binds its path/headers/body lines.
Regenerated the Python baseline (Python 3.12, matching the scan CI shards):
same package/file/check set. Adds regression tests for each.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan npm packages: normalize and bound the logical-line digest
- _evidence whitespace-normalizes the logical line before digesting (matching
_evidence_hash), so a formatter-only reindent of the bound continuation lines
does not change the sha256 suffix and reopen an unchanged finding.
- _logical_line_text follows a bracket group to its close up to a hard 200-line
cap (digest input only), so a config object longer than the backward window
still binds its whole tail instead of silently truncating.
Adds regression tests. npm baseline is empty, so no regeneration is needed.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan: cap single-line evidence and widen npm opener window
Cap each rendered evidence line at 200 chars in scan_packages.py: a long
or minified one-line file is shown as a bounded prefix plus a sha256 of the
full line, so a packed payload cannot dump unbounded content into the CI
logs or baseline while a change past the cutoff still changes the digest
and reopens the finding. Mirrors how the npm scanner bounds its snippets.
Widen the npm backward opener window (_MAX_CONT_LINES 12 to 200, symmetric
with the forward cap) so a host deep inside a large options object binds
the whole object, not just its own line; a changed path, header, or body on
any property reopens.
Regenerate the Python baseline with Python 3.12: only the protobuf
nspkg.pth and unsloth-zoo compiler.py evidence change, both from the new
line cap; the package/file/check key set is unchanged.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* scan: bind all host contexts, deep call continuations, far-back npm openers
Three fail-closed evidence gaps surfaced by review of the previous round.
scan_npm_packages.py: measure the forward bracket-group cap from the matched
line (idx + _MAX_GROUP_LINES) instead of the opener, so an opener found near
the widened backward limit no longer consumes the forward budget and drops
the path, headers, or body that follow the host.
scan_npm_packages.py: _outbound_host_evidence now records every outbound
context form for a host (URL, fetch-context, host-config), claiming each
non-overlapping match in form order, so a separate host-config request added
beside an already-baselined URL changes the evidence and reopens the key.
The common single-context case keeps its existing snippet.
scan_packages.py: follow a matched Python call over its continuations up to a
separate _MAX_CALL_LINES (40), decoupled from the 12-line display threshold,
so a multi-line requests.post( binds its whole argument list in the digest
and a changed body deep in the call reopens; bounded so a miscounted bracket
cannot swallow unrelated code. No baseline change: the current dependency set
has no matched call that closes between 13 and 40 lines, confirmed by a
Python 3.12 regenerate that produced a byte-identical baseline.
* scan: clamp npm depth, pin large bundles, follow backslash and bound .pth dump
Four fail-closed evidence gaps surfaced by review of the previous round.
scan_npm_packages.py: clamp the backward opener scan at depth 0 so a leading
unmatched closer (a preceding block whose opener is outside the backward
window) no longer drives depth negative and masks the real enclosing opener
that follows; a host-config object after such a block now binds and a changed
path reopens.
scan_packages.py: a large JS bundle now pins its whole content even when
another JS heuristic already fired. The bundle digest was only added when no
other finding existed; it is now appended to every finding's evidence on a
large bundle, so an unchanged obfuscation signature no longer lets changed
payload elsewhere ride the matched-line key.
scan_packages.py: _logical_line_end follows explicit backslash line
continuations, so a call split with a backslash before its parenthesis binds
the continuation line (URL/body) instead of returning at the zero-depth API
line.
scan_packages.py: the catch-all .pth import evidence is bounded through
_cap_line (prefix plus a digest of every line) so a large .pth of benign
imports cannot dump the whole member into the logs or baseline while an
appended or swapped import still reopens.
Baseline regenerated with Python 3.12: key set unchanged; one entry
(unsloth-zoo compiler.py) gains the backslash-continued banner lines now
bound by the continuation fix.
* scan: handle multi-line strings, lifecycle bodies, and de-quadratic evidence
Addresses a review round plus a performance audit of the evidence extractor.
Correctness (fail-closed):
- Bind the UNION of the single-line-blanked and multi-line-blanked bracket spans
in both scanners. The multi-line view blanks a triple-quoted Python string or a
backtick template literal that spans lines, so a `)` inside such a string no
longer closes the enclosing call early and drop later arguments. The single-line
view still counts a payload embedded INSIDE a string, so a dropper that hides a
call in a string keeps its argument lines bound. Taking the larger span never
shrinks the binding below either view, avoiding a fail-open regression.
- cred-env-in-lifecycle now pins the whole lifecycle script body via a digest, so
a changed non-token line (e.g. adding a curl exfil beside the token reference)
reopens, not just a change on the token line.
Performance / DoS (the scanner runs on attacker-controlled package files up to the
64 MiB / 16 MiB member caps, with no per-file time budget):
- _extract_evidence precomputes newline offsets once and maps match offsets with
bisect, removing the O(matches) whole-file content.count per match that made the
finditer fallback quadratic (a crafted minified file went from ~13 s/MiB and
hours at the cap to linear).
- npm _index_text splits and string-blanks the file once per evidence call instead
of per match (was O(matches x file) time and allocation).
- Bound evidence output: _MAX_EVIDENCE_SPANS (Python) and _MAX_EVIDENCE_MATCHES
(npm) fold the remainder into a digest so a file with thousands of matches cannot
build a multi-megabyte evidence/baseline blob while an added/removed match past
the cap still changes the key.
- _outbound_host_evidence caps matches per form and bounds the overlap claim so a
host repeated many times cannot make it quadratic.
No baseline change: a Python 3.12 regenerate is byte-identical (the union equals the
legacy single-line span for every current dependency file; the cap thresholds sit
above the largest real entry), so these are forward-looking hardening with no drift.
* scan: count all overflow matches, bind their context, blank JS regex literals
Follow-ups on the evidence output caps from the previous commit.
- _outbound_host_evidence no longer truncates each pattern's match iterator with
islice; it iterates every match and runs the overlap dedup only while the
display list is below the cap (so claimed stays bounded and the check is O(cap)
per match, not quadratic), folding every match past the cap into the overflow
digest. A host context beyond the 64th is counted again, so it reopens.
- The overflow digest (both scanners, via a shared _overflow_digest) binds each
overflow match's logical-line context, not just the regex match text, so a
changed payload on an over-cap line reopens even with the matched token
unchanged.
- The multi-line JS blanked view now blanks regex-literal bodies (tracking the
previous significant char for regex-vs-division and char classes for a literal
`/` inside `[...]`), so a `)` inside `/)/` no longer closes an outbound call
early. The bound span is the union of the single-line and multi-line views, so
an imperfect regex decision only ever grows the span, never shrinks it.
- The Python overflow digest canonicalizes spans (strips L<NN>: markers via
_canon_evidence) before hashing, restoring line-shift stability for the
over-cap region.
No baseline change: the overflow branches only trigger above the per-finding caps
(above the largest real entry), and the npm baseline is empty, so a Python 3.12
regenerate is byte-identical.
* scan: refresh baseline for ipython interactiveshell.py span drift
A newer ipython release changed the filesystem-enumeration span in
IPython/core/interactiveshell.py, so its content digest no longer matched the
baselined evidence and the studio scan shard flagged it as a non-baselined
CRITICAL. Regenerated with Python 3.12: only the ipython entry's evidence_hash
changes; the package/file/check key set is unchanged, and a studio enforcing
spot-check exits 0.
* Bound scanner evidence memory: stream overflow spans and cap lifecycle baseline size
scan_packages.py: _extract_evidence no longer materializes a rendered span
per match before slicing at the display cap. Once out holds _MAX_EVIDENCE_SPANS
spans, further spans fold straight into a running digest, so a minified or
padded file with hundreds of thousands of matching lines keeps memory bounded
to the display cap instead of the match count. The fold reproduces
_canon_evidence(" | ".join(overflow)) byte for byte, so the overflow digest and
every baseline key are unchanged.
scan_npm_packages.py: lifecycle-fetch-exec and cred-path-in-lifecycle stored the
entire install script body as evidence, so --write-baseline on a package with a
multi-MiB lifecycle script bloated the baseline JSON. Both now store a bounded
matched snippet plus a body-sha256 digest, matching cred-env-in-lifecycle. The
digest still binds the whole body, so a change to any line reopens the finding.
Adds tests for the streamed overflow bound and the bounded-but-reopens lifecycle
evidence. Baseline unchanged (byte-identical Python evidence; npm baseline empty).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Make npm bracket-group scan order-aware so a same-line close-then-open binds
_scan_group counted brackets with a per-line net (opens minus closes), which
collapses intra-line order: a line that closes a prior block and then opens the
host-config object on the same line, e.g. `}); const opts = {`, nets to <= 0, so
the trailing `{` was dropped and the group started at the hostname line. A
changed path/headers on the following lines then hashed to the same evidence and
could ride an existing baseline key.
Replace the net count with an order-aware (L, R) reduction per line (L closers
needing an opener to the left, R openers needing a closer to the right) and apply
it in order in both the backward and forward scans, clamping stray closers at 0.
The trailing opener now stays visible so the whole object binds and a changed
payload reopens. Per-line cost is unchanged (one C-level bracket findall), so the
existing outbound-host evidence is byte-identical on all prior shapes; only the
previously-dropped same-line case changes. Adds a regression test for it.
* Harden scanner evidence: bound memory and bind Python call tails fail-closed
Five fixes across both scanners, none of which change the committed baseline (a
full regen of all three pip shards produced a byte-identical 185-key set).
scan_npm_packages.py: _evidence and _outbound_host_evidence collected every regex
match into a list before applying the 64-match display cap, so a text file under
the size cap that repeats a cheap signal (such as NPM_TOKEN) millions of times
could allocate a huge list of re.Match objects and stall or OOM before the
overflow digest ran. They now stream from finditer and fold overflow as matches
arrive via a shared _fold_overflow_match helper, byte-identical to the prior
digest.
scan_packages.py:
- _extract_evidence kept inserting every unique over-cap span into the seen set
even after it stopped appending to the display list, so a generated file with
millions of one-line matches still grew that set unbounded. It now tracks spans
only while filling the display list (per-line spans are unique by line number,
so dropping them past the cap cannot miss a dedup).
- _scan_line_end counted brackets with a per-line net, so a continued statement
that closes on the same line it opens a flagged call (a leading "]" before
"requests.post(") had the call's open paren cancelled and bound only the opener
line. It now applies brackets in order via _bracket_lr (leading closers clamp at
0), matching the npm bracket fix.
- a single-quoted string continued by a trailing backslash was not tracked across
lines, so a close paren inside the continued string on the next line closed the
call early; _blank_code_strings now carries the continuation.
- a call with more argument lines than the soft cap was hashed only through the
cap, so a changed data=/headers tail past it stayed suppressed; a closing call
is now followed to its real close under a 200-line hard limit (a never-closing
opener still stops at the 40-line soft cap so it cannot swallow the file).
Adds regression tests for each. npm baseline is empty; the Python baseline is
unchanged (verified byte-identical by regenerating all three shards).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Bind giant DOTALL span anchors and add context to constant IOC evidence
Two fail-closed gaps where a changed payload could keep the same evidence hash
and stay suppressed by the baseline.
scan_packages.py: a giant greedy DOTALL span (a cross-line IOC match bridging
more than 60 lines, e.g. RE_TEMP_EXEC matching a /tmp line and a much-later
subprocess line) was dropped entirely once the per-line pass had any match, so an
appended cross-line payload -- a new /tmp line plus a later subprocess line that
share no single line, so the per-line pass never binds them -- produced the same
evidence and rode the key. The span is no longer dropped: it is bound by its head
and tail anchor lines plus a digest over just those (no line numbers, so a pure
line shift is stable). An added or moved anchor reopens the finding, while churn
in the bridged interior stays stable, so this does not reintroduce whole-file
drift. Two baseline entries (multiprocess test, unsloth-zoo scanner file) carry
such a span and are refreshed; a full three-shard regen confirmed only those two
keys change.
scan_npm_packages.py: known-ioc-string and cred-surface-host (always-bad) recorded
only the bare needle/host as evidence, so a reviewed tarball that kept the IOC
string while altering the adjacent fetch/exfil body produced an identical key.
They now bind matched-line context: known-ioc-string via the matched line and its
bracket-group continuation, cred-surface-host (always-bad) via the outbound call
context (path/headers/body, falling back to the bare host when not in an outbound
call). A changed payload on the same call now reopens.
Adds regression tests for each. npm baseline is empty; the Python baseline updates
only the two giant-span entries.
* Hash giant-span interiors, bind exec/eval trigger, JS content, intra-literal whitespace
Four fail-closed gaps where a changed payload could keep the same evidence hash.
scan_packages.py:
- A giant bridged DOTALL span was bound only by its head and tail anchors, so a
cross-line payload inserted into the bridged interior between unchanged outer
anchors kept the same key. The whole span content is now digested (via _render),
so any interior change reopens; a pure line shift stays stable because the digest
is over the markerless code. Two baseline entries (multiprocess test, unsloth-zoo
scanner file) carry such a span; with full-interior binding, multiprocess
resolved at two versions across shards now yields two distinct entries where the
anchor digest had collapsed them into one.
- The exec/eval-with-hidden-payload findings omitted the visible exec/eval line
that makes the hidden string executable, so flipping a harmless eval("1+1") to
exec(__doc__) kept the same key while arming the payload. The trigger line from
the real-code view is now bound into the evidence.
- check_js_file extracted evidence with the Python-string-aware extractor, which
does not blank JS backtick template literals, so a template containing a close
paren closed a call's bracket span early and omitted later option/body lines. The
full file content digest is now pinned to every JS finding (not just large
bundles), binding the whole call.
scan_npm_packages.py: the evidence canon collapsed all whitespace via split(),
erasing whitespace inside JS string literals along with harmless indentation, so a
changed request body 'a b' -> 'a b' kept the same key. A new _canon_preserve_strings
collapses whitespace only OUTSIDE string literals (reindent-stable) while preserving
it INSIDE single/double/backtick literals (intra-payload edits reopen). Used for the
evidence hash and the logical-line digests.
Adds regression tests for each. npm baseline is empty; the Python baseline updates
the two giant-span entries and adds the second multiprocess version's entry.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
check_js_file put the bundle's KB size inside the finding's check label, which is
part of the baseline match key (package, file, check). When tensorboard's
projector_binary.js grew from 1918 KB to 1933 KB, the reviewed baseline entry stopped
matching and the benign HIGH resurfaced, red-failing the studio and extras
scan-packages shards. Move the size into the evidence field (shown for review, not
matched) and keep the check label constant, then update the one tensorboard baseline
entry to the size-agnostic label. The finding is suppressed again and will not
re-break when the bundle grows by a few KB. Scanner self-tests pass unchanged.
Co-authored-by: danielhanchen <michaelhan2050@gmail.com>
* Studio: drop OpenEnv and unused ExecuTorch/open_spiel install deps
* Studio: drop 8 more unused install deps from extras
* Studio: restore tomli<3.11 for kernels; tidy dep-cleanup comments and tests
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio: refresh scan-packages baseline for scipy _external + unsloth-zoo tests
scipy moved its vendored array_api_compat from scipy/_lib to
scipy/_external, so the four allowlisted array_api_compat __init__.py
entries stopped matching and resurfaced as unsuppressed CRITICAL
"Downloads and executes remote code" findings on all three pip
scan-packages shards (extras, hf-stack, studio). Add the _external
paths next to the existing _lib ones so both scipy layouts stay covered.
Allowlist two unsloth-zoo test-file false positives now present in the
hf-stack shard: tests/test_mlx_save_export_regressions.py (writes to
/tmp dropper) and tests/test_mlx_trainer_internals.py (obfuscation plus
exec/eval).
Drop nine stale entries for packages removed from the Studio
requirements and no longer in any shard closure (evaluate, pytest,
hypothesis, kgb, langid), confirmed absent via with-deps resolution of
all three shards.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
Three independent upstream CI fixes that currently fail on every open PR:
verify_import_hoist.py: TARGET-CHANGED only flags a genuine swap (a BEFORE
target no longer reachable in AFTER). A pure superset growth such as adding
import urllib.error next to import urllib.request binds the same top-level
package and loses nothing, so it is no longer a blocker (transformers_version.py).
test_vision_cache.py: run each test from a fresh empty cwd. is_vision_model
calls is_local_path first, and a relative model id that happens to exist on
disk short-circuits before the mocked detection runs; the CI cwd and HF cache
can contain dirs colliding with the synthetic ids, causing 'called 0 times'.
Production code is correct; only the test needed cwd isolation.
consolidated-tests-ci.yml: the llama.cpp smoke probes the first of
llama-cli / llama-mtmd-cli / llama-server that exists instead of hard-requiring
llama-cli, which upstream no longer always builds. llama-cli stays first so it
is preferred when present. Adds Windows .exe + build/bin/Release handling.
* Windows installer: fix DiskPart UAC, drive-root cache, spurious rename warning, CPU-base messaging
amd-smi gate (DiskPart UAC mid-install): the AMD torch wheel ships hipInfo.exe
inside the venv, and the bitsandbytes fix prepends that venv Scripts dir to PATH.
shutil.which("hipinfo") then found it and flipped _amd_smi_allowed() to True, so
the post-install AMD probe fell through to `amd-smi list` (the venv hipInfo failed
to report gcnArchName, which is why the arch came from the GPU-name table) and
amd-smi elevated, popping the DiskPart UAC. Fix: a hipinfo resolved inside the
active venv (sys.prefix) is the torch-wheel binary, not a HIP SDK, and must not
open the gate. Mirrored in install_python_stack.py, install_llama_prebuilt.py, and
backend utils/hardware/amd.py (the runtime VRAM poller had the same latent prompt).
TORCHINDUCTOR_CACHE_DIR: move from C:\tc to <StudioHome>\TORCHINDUCTOR_CACHE_DIR so
the inductor/Triton cache lives under the user's Studio home, not the system drive
root. Long paths are already enabled above so deep inductor paths still fit.
unsloth.exe rename: skip the rename (and its "pip may fail with WinError 32"
warning) when SKIP_STUDIO_BASE=1. In the install.ps1 flow base packages are not
reinstalled, so unsloth.exe is never rewritten; the self-rename only failed because
setup runs via unsloth.exe (the running launcher holds its own file). The
'studio update' flow still attempts it.
CPU PyTorch messaging: clarify that the CPU base is temporary and setup replaces it
with GPU ROCm wheels, and print an explicit "GPU ROCm PyTorch installed" line after
the AMD wheels land, so the log makes clear the final install is GPU-accelerated.
Adds two regression tests covering the venv-internal vs external hipInfo gate.
Verified end-to-end on a Strix Halo box (Radeon 8060S / gfx1151): install.ps1
--local from this branch completed exit 0 with no DiskPart prompt, no rename
warning, the cache under the Studio home, and "GPU ROCm PyTorch installed
(gfx1151)"; Studio then booted and detected "ROCm (HIP 7.13.99004) -- AMD Radeon
8060S Graphics".
* Windows installer: drop the unreliable unsloth.exe rename and its WinError 32 warning
setup.ps1 used to rename the running unsloth.exe out of the way before the
base-package upgrade so pip could replace it. That rename never actually
worked: setup runs *via* unsloth.exe, so renaming our own running
uv-trampoline launcher failed with a sharing violation (WinError 32) and only
printed a scary 'could not rename unsloth.exe; pip may fail with WinError 32'
warning on every Windows install and update.
It also was not needed. pip tolerates a running/locked console-script .exe: it
moves the old one aside and writes the new one. The base upgrade routes through
pip on Windows, so the upgrade succeeds (or, in the install.ps1 flow with
SKIP_STUDIO_BASE=1, the base is not touched at all) and unsloth.exe is left
intact either way.
Removing the rename block and its failed-install restore block removes the
false warning for all Windows devices in both the install and update flows.
* Windows installer: gate venv-internal hipInfo.exe in PowerShell amd-smi probe; harden venv path checks
Follow-up to PR #6296.
- install.ps1 and setup.ps1: ignore the AMD torch wheel hipInfo.exe that lives
inside the Studio venv when probing for a HIP SDK, so amd-smi no longer reopens
the DiskPart UAC during install/update. Mirrors _path_inside_venv in the Python
installers, which already do this.
- amd.py, install_llama_prebuilt.py, install_python_stack.py: normcase the venv
containment check (Windows paths are case-insensitive) and run the
HIP_PATH/ROCM_PATH candidate through it too.
- setup.ps1: fall back to a short TORCHINDUCTOR cache dir when long paths are
unavailable, and create the dir wildcard-safely.
- tests: isolate sys.prefix in the gate helper, add HIP_PATH/ROCM_PATH cases, and
assert the PowerShell venv exclusion.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Windows installer: install ROCm PyTorch directly for a known AMD arch
When the GPU arch is known (name-inferred from the GPU-name table) but ROCm
could not be probe-verified (no HIP SDK, no amd-smi), the bootstrap installed
a CPU PyTorch base that setup.ps1 then force-reinstalled as ROCm. The
repo.amd.com wheels bundle their own runtime (no HIP SDK required), which
setup.ps1 already relies on, so the CPU base was a pure wasted download/install.
- Gate the ROCm index on a known arch, not only on probe-verified ROCm, so a
mapped arch installs ROCm torch directly. Unmapped arches and no-GPU hosts
still get CPU (unchanged).
- Fall back to a CPU base if the ROCm-index install fails, so a transient
repo.amd.com outage does not abort the install (setup.ps1 retries ROCm).
- Correct the stale comment that claimed ROCm wheels need a confirmed HIP SDK.
- Add a regression test for the arch-based gate.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Windows installer: correct the unsloth.exe rename-removal comment
The comment claimed the base upgrade 'routes through pip on Windows' and that
pip 'moves the old unsloth.exe aside, then writes the new one'. That is not what
the code does. install_python_stack tries uv first; on a locked launcher uv
aborts and falls back to pip, but the pip fallback strips --upgrade-package and
base.txt lists only bare unsloth/unsloth-zoo, so pip finds them already
satisfied and no-ops. The running unsloth.exe is left intact at its current
version either way. Reword the comment to describe the real uv-first /
pip-fallback-no-op behavior. No functional change.
* Windows installer: close two gaps in the venv-internal hipinfo exclusion
Review follow-up. The amd-smi/DiskPart gate could still reopen in two cases:
- setup.ps1 ran the HIP probe long before $VenvDir is assigned, so without
VIRTUAL_ENV (the `unsloth studio update` path) $venvRoots was empty and the
venv-internal hipInfo.exe was not recognized. Seed the venv root from
UNSLOTH_SETUP_PYTHON and the default Studio home too (both installers).
- The HIP_PATH/ROCM_PATH candidate was accepted without the venv filter, so an
env var pointing into the venv (AMD wheel) still set $HipSdkInstalled. Run
Test-HipinfoIsVenvInternal on the candidate as well (both installers).
Extend the PS gate test to assert both. Both .ps1 parse clean; install tests
pass (the venv-internal / HIP probe coverage at 359 passed).
* Windows installer: correct the CPU-base message for arches with no ROCm wheels
After gating the ROCm index on a known arch, a mapped arch sets $ROCmIndexUrl
and installs ROCm directly, so it no longer reaches the "temporary CPU base"
branch. That branch is now reached only by a name-inferred arch with no ROCm
wheels (e.g. RDNA2 gfx103X), where setup.ps1 does NOT install ROCm. The old
text ("setup replaces it with GPU ROCm wheels ... the final install IS
GPU-accelerated") was therefore always wrong there. Say plainly that PyTorch
stays on CPU for this GPU.
* Windows installer: seed the venv-internal hipInfo check from a custom Studio home
Test-HipinfoIsVenvInternal seeded the venv root from VIRTUAL_ENV, VenvDir, the
setup python, and the default %USERPROFILE% path only. A standalone
`unsloth studio update` with a custom UNSLOTH_STUDIO_HOME (or STUDIO_HOME alias)
and none of those set would not recognize the venv hipInfo on PATH, reopening the
amd-smi/DiskPart gate. Seed the custom home too, in both installers, and assert
it in the gate test.
* Studio installer: resolve venv aliases and expand ~ in the hipInfo venv filter
Two review points on the amd-smi/DiskPart UAC gate:
1. _path_inside_venv compared os.path.abspath of sys.prefix and the hipInfo
path, which does not resolve symlinks, junctions, or 8.3 short names. A venv
reached through an aliased path then fails the check, so its bundled
hipInfo.exe is mistaken for an external HIP SDK and amd-smi runs (the
DiskPart prompt this fix exists to suppress). Switch to os.path.realpath in
all three copies (amd.py, install_llama_prebuilt.py, install_python_stack.py).
2. setup.ps1's early venv-internal hipInfo probe seeded the venv root from a
custom Studio home (UNSLOTH_STUDIO_HOME / STUDIO_HOME) without expanding a
leading ~, while the canonical resolver does. With a tilde form,
[IO.Path]::GetFullPath kept the literal ~ relative to cwd, so the custom-home
hipInfo escaped the filter and reopened the gate. Expand ~ in the probe the
same way as the resolver.
tests/studio/install/test_pr5940_followups.py: 30 passed (adds a symlink
realpath case and a setup.ps1 tilde-expansion guard).
* Studio installer: mirror the hipInfo venv filter and ROCm wheel pins into install.ps1
Follow-up review on the same install.ps1 paths:
1. install.ps1's venv-internal hipInfo probe (Test-HipinfoIsVenvInternal)
seeded the venv root from a custom Studio home without expanding a leading
~, unlike the canonical resolver and setup.ps1. A tilde form left
[IO.Path]::GetFullPath with the literal ~ (relative to cwd), so the
custom-home hipInfo escaped the filter and reopened the amd-smi/DiskPart
gate. Expand ~ in the probe, matching the setup.ps1 fix.
2. The AMD ROCm path installed torchvision/torchaudio bare while pinning torch
to below 2.12. AMD's per-arch index publishes the companions independently
and may ship torchvision 0.27 (for torch 2.12) before removing 0.26, so a
bare resolve can pick an ABI-incompatible set and fall back to CPU. Add
torchvision/torchaudio floor maps and pass the pinned specs, mirroring
setup.ps1 and install_python_stack.py.
3. The ROCm-to-CPU fallback torch install used Invoke-InstallCommand (no
retry), the only torch step in the file without it. Switch to
Invoke-InstallCommandRetry so the recovery path survives a transient index
failure.
tests/studio/install/test_pr5940_followups.py: 33 passed (parametrized tilde
check over both installers, a torch/companion floor-map parity test, and a
CPU-fallback retry guard).
* Studio installer: scan all PATH hipinfo so the venv copy can't shadow a real HIP SDK
The amd-smi HIP-SDK probe used shutil.which("hipinfo") / Get-Command hipinfo,
which return only the first hit on PATH. The AMD torch wheel ships hipInfo.exe
inside the venv and the bnb fix (plus the Studio backend) prepend the venv
Scripts dir to PATH, so that venv-internal copy lands first. When a real HIP SDK
hipinfo sits later on PATH with HIP_PATH/ROCM_PATH unset, the first-hit probe
stopped at the venv copy, treated it as "not a HIP SDK", and closed the amd-smi
gate -- AMD users in that PATH-only SDK setup lost amd-smi telemetry and could
fall back to CPU. Scan every PATH entry and keep the first hipinfo that is not
venv-internal; only the venv copy is ignored, so the UAC/DiskPart suppression is
unchanged.
Applied to all three Python copies (install_llama_prebuilt.py,
install_python_stack.py, backend/utils/hardware/amd.py) via a new
_external_hipinfo_on_path helper, and both PowerShell callers (install.ps1,
setup.ps1) now use Get-Command hipinfo -All filtered by Test-HipinfoIsVenvInternal.
tests/studio/install/test_pr5940_followups.py: 36 passed (real-PATH scan tests, a
shadow-regression test for the exact venv-first ordering, and a parity check that
every Python copy uses the scanning helper).
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Studio uninstallers: fix leftovers (false "removed", shared icon, llama lock)
Auditing a dual native+WSL uninstall on a real device surfaced three leftovers:
1. uninstall.ps1 removed the data dir (which holds unsloth.ico) before the
shortcuts that reference that icon, so Explorer's icon cache briefly held it
open. Remove-Item -Recurse reported success yet left the locked file, and the
dir was never re-attempted, so it orphaned with a false "removed" log.
_RemovePath now verifies the path is actually gone (retrying transient locks)
and reports honestly, and the data dir is re-swept after the shortcuts go.
2. install.sh writes a shared unsloth.ico to %LOCALAPPDATA%\Unsloth Studio for
the WSL shortcut, but uninstall.sh never removed it, orphaning the icon (and
dir) after a WSL uninstall. uninstall.sh now drops that icon and the dir when
empty, in both the powershell.exe and drvfs-fallback paths.
3. ~/.unsloth/.llama.cpp.install.lock was never removed, so the rmdir of
~/.unsloth failed and the dir lingered. Both uninstallers now remove the lock.
Verified by running both uninstallers on a real dual install: device fully clean
(no install dirs, shortcuts, PATH/registry entries, shared icon, or lock left).
* install.sh: auto-route Strix Halo WSL to an existing Ubuntu 24.04
ROCm-on-WSL is the GPU runtime for Strix Halo and only targets Ubuntu
24.04. When the installer runs in a newer default distro (e.g. 26.04) it
cannot enable the GPU and silently falls back to CPU. If a 24.04 distro
already exists, re-run the install there and stop in the current one so the
GPU path is taken without the user having to know about the distro
requirement.
Runs before venv creation so the wrong distro is left untouched, guards
against re-route loops via UNSLOTH_WSL_REROUTED, leaves a working ROCm
distro alone (librocdxg present), and skips the GGUF-only / opt-out /
non-Strix cases. When no 24.04 distro exists we keep today's behaviour:
continue to CPU and print the `wsl --install Ubuntu-24.04` guidance, never
auto-downloading a distro.
Adds tests/sh/test_strixhalo_wsl_reroute.sh (hermetic: extracts the
function, rewrites its paths to fixtures, mocks wsl.exe) covering the full
decision matrix, wired into tests/run_all.sh.
* uninstall.ps1: keep shared unsloth.ico for a surviving WSL shortcut
A dual native+WSL install shares %LOCALAPPDATA%\Unsloth Studio\unsloth.ico:
install.sh points the WSL shortcut's icon there while the native install owns the
dir. The native uninstaller removed the whole dir unconditionally, so uninstalling
native while keeping WSL left the WSL shortcut with a blank icon. The old code only
avoided this when Explorer happened to hold the icon open, which is unreliable; on a
real dual install the dir was deleted and the WSL shortcut went blank.
_RemoveDataDirKeepingWslIcon now scans the Start Menu + Desktop for a surviving
"Unsloth Studio (WSL ...).lnk" and, if found, removes everything in the data dir
except unsloth.ico (keeping the dir) instead of deleting it; with no WSL shortcut it
removes the dir as before. uninstall.sh still drops the icon and the empty dir when
WSL itself is uninstalled, so every uninstall order ends clean.
Adds tests/studio/test_uninstall_dual_install_icon.ps1 (AST-extracts the helper and
runs it against a temp dir with controlled shortcut dirs) covering the dual,
native-only, empty, and missing-dir cases, wired into the windows-inference smoke
workflow. Verified on a real dual install: native uninstall now keeps unsloth.ico
and the WSL shortcut's icon stays intact.
* installer: condense AMD/ROCm code comments (no behavior change)
Tighten the comments added for the Strix Halo native+WSL installer work so
they are shorter and clearer without losing intent: the venv-internal hipInfo
amd-smi gate, the ROCm torch/companion floor maps, the WSL 24.04 reroute, and
the dual-install uninstall icon handling. Comment-only; code paths unchanged.
107 insertions, 166 deletions across 11 files.
* install.sh: run the Strix Halo WSL reroute before any STUDIO_HOME write
The reroute fired after mkdir -p "$STUDIO_HOME" and the legacy-venv migration,
so rerouting 26.04 -> 24.04 left an empty ~/.unsloth/studio stub in the origin
distro (and ran venv migration in the distro about to be abandoned). Move the
reroute ahead of the venv section so the origin distro is left untouched, matching
the function's own comment. Behavior is identical on every non-reroute path.
* installer: fix ROCm CPU-fallback, hipinfo gate edge cases, uninstall icon, WSL 22.04
- install.ps1: clear $ROCmIndexUrl/$ROCmTorchFloor after the CPU fallback so the
flavor-repair block does not retry the failed ROCm index and abort the install;
pin the ROCm companion specs ($visionSpec/$audioSpec) in the repair path too.
- install.ps1 + setup.ps1: skip a bare drive root in Test-HipinfoIsVenvInternal so a
non-venv UNSLOTH_SETUP_PYTHON does not match the whole drive; iterate
HIP_PATH/HIP_PATH_57/ROCM_PATH and take the first non-venv hipinfo.
- amd.py, install_llama_prebuilt.py, install_python_stack.py: strip surrounding
quotes from PATH entries before probing for hipinfo.
- install.sh: pipefail the WSL reroute curl|sh; do not reroute supported Ubuntu 22.04.
- uninstall.sh: keep the shared unsloth.ico while any Unsloth shortcut (native or
another WSL distro) still references it, in both the powershell and drvfs paths.
- tests: regression coverage for all of the above.
* installer: forward reroute options, guard ROCm bootstrap, harden hipinfo gate
- install.sh: forward the caller's --package/--python/--verbose/--tauri and a custom
UNSLOTH_STUDIO_HOME into the WSL reroute (was a bare default install); bail on
--local; run the reroute BEFORE dependency/uv install so the origin distro is left
untouched; set UNSLOTH_SKIP_ROCM_WSL_SETUP after a failed reroute so the later
ROCm-on-WSL bootstrap does not install into the unsupported origin distro.
- install.ps1 + setup.ps1: Get-Command hipinfo -CommandType Application so only real
executables match (not an alias/function named hipinfo).
- uninstall.ps1: guard $env:APPDATA when building the default shortcut search dirs.
- tests: cover option forwarding, --local bail, the bootstrap guard, and the gate change.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* installer: guard origin ROCm bootstrap on every CPU-only fallback; harden ~ expansion
WSL reroute: the no-wsl.exe, no-24.04-target and --local fallbacks all tell the
user the install continues CPU-only, but only the failed-reroute branch set
UNSLOTH_SKIP_ROCM_WSL_SETUP=1. The later _maybe_bootstrap_rocm_wsl gate keys off
that flag, so the other three branches could still install ROCm into the
unsupported origin distro (e.g. 26.04). Set the skip guard on all of them.
Forward UNSLOTH_ROCM_WSL_AUTO into the reroute so a Tauri/consented GPU bootstrap
carries through to the rerouted 24.04 child instead of dropping to the prompt path.
install.ps1/setup.ps1: guard the venv-probe ~ expansion on a non-empty
$env:USERPROFILE so Join-Path does not throw on a profile-less service account.
Tests: add no-wsl.exe and UNSLOTH_ROCM_WSL_AUTO reroute cases, the USERPROFILE
guard assertion, and route shell-test fixtures through a single trap-cleaned root.
* installer: pin + soften Windows ROCm Python repair, reroute to 22.04, harden gates
install_python_stack.py: the Windows AMD ROCm repair in _ensure_rocm_torch()
installed bare torch/torchvision/torchaudio via the fatal pip_install -- the same
asymmetry already fixed on the PowerShell side. A transient repo.amd.com failure
could abort the whole install even after install.ps1/setup.ps1 fell back to CPU.
Pin companions per-arch (gfx120X/Strix -> the rocm7.2 trio, mirroring the PS floor
maps) and make the retry nonfatal: keep the existing build and let the user re-run
update to retry ROCm, so the chain install.ps1 -> setup.ps1 -> stack stays CPU-safe.
install.sh: reroute now targets an installed Ubuntu 24.04 OR 22.04 (24.04 preferred);
both are AMD-supported for ROCm-on-WSL, matching the leave-alone set, so a box with
only 22.04 reaches the GPU instead of staying CPU-only.
install.ps1/setup.ps1: a bare ~ for UNSLOTH_STUDIO_HOME left an empty Join-Path child
(PS 5.1 throws); fall back to USERPROFILE directly and only join a real remainder.
_path_inside_venv (amd.py + both installers): guard a root-dir sys.prefix so commonpath
can't classify every path on the drive as venv-internal (defensive; venv never at root).
uninstall.sh: guard an empty LOCALAPPDATA in the PS-interop icon cleanup (mirror APPDATA).
Tests: add 22.04-target reroute cases, Windows ROCm pin+nonfatal coverage (text +
behavioral), root-dir guard coverage, and bare-~/LOCALAPPDATA guard assertions.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* install.sh: match WSL reroute target by exact distro name, not substring
The 24.04/22.04 reroute target was chosen with grep -F (substring), so a custom
distro such as 'Ubuntu-24.04-test' (with no exact Ubuntu-24.04) was picked as the
target; the later 'wsl -d Ubuntu-24.04' then fails and the Strix Halo install stays
CPU-only. Match whole lines (grep -ixF) and reuse the matched name so only a real
Ubuntu-24.04/22.04 is targeted. Adds substring-rejection + exact-vs-custom tests.
* install.sh: keep the WSL reroute target to Ubuntu 24.04 (helper-supported only)
The ROCm-on-WSL bootstrap (scripts/install_rocm_wsl_strixhalo.sh) dies on any
VERSION_ID other than 24.04 and pins the noble repo, so treating 22.04 as
GPU-supported let the parent report a successful reroute while the child fell
back to CPU. Drop 22.04 from the supported set and the reroute target list;
24.04 stays the sole target (keeping the exact whole-line distro match). An
already-working ROCm on any other version is still left alone by the librocdxg
check above.
tests: reroute 22.04 cases updated to the 24.04-only behavior; make the
"no wsl.exe" case hermetic so a real host wsl.exe can't leak in on dev boxes;
stop the tauri exit-order check from mis-flagging the reroute helper's
[ "$TAURI_MODE" = true ] && ... --tauri one-liner.
* installer: tighten comment wording across the Strix Halo install/uninstall paths
Condense the verbose multi-line comment blocks (amd-smi hipinfo gate, ROCm
torch install + CPU fallback, WSL reroute, uninstall icon-keep) into fewer,
clearer lines. Comments and a few docstrings only; no code, logic, or
behavior change. Verified with bash -n, the PowerShell parser, and ast.parse,
and the installer test suite still passes.
* add AGPL-3.0 SPDX headers to the .sh/.ps1 scripts missing them
Every shell and PowerShell script under the Studio/installer surface now
carries the standard SPDX-License-Identifier: AGPL-3.0-only + copyright
header (after the shebang where present): the installer (install.sh,
install.ps1), build.sh, the .github and src-tauri scripts, the installer
test suite, and the moe kernel test. Header-only, line endings preserved;
bash -n, the PowerShell parser, and the installer tests all pass.
* installer: drop the duplicate AGPL header from install.sh and install.ps1
Both already carry an SPDX-License-Identifier: AGPL-3.0-only header below
their usage comment block; the prior header pass added a second one at the
top because it only scanned the first few lines. Remove the duplicate so each
file keeps a single original header.
* installer: force-reinstall CPU fallback torch; propagate Tauri NEED_SUDO from reroute
install.ps1/setup.ps1: when the AMD ROCm wheel install fails and we fall back to a
CPU base, force-reinstall the torch/vision/audio triplet. A failed ROCm install can
leave an unpinned ROCm torch (e.g. 2.10.0+rocm on gfx110X/gfx90a) that still
satisfies the CPU torch>=2.4,<2.11.0 range, so without --force-reinstall uv keeps the
ROCm build and only swaps the companions -- a mismatched venv the flavor-repair block
won't fix. setup.ps1 scopes the forced reinstall to the ROCm-fallback path
() so the genuine CPU-only install stays fast.
install.sh: the Strix Halo WSL reroute treated every nonzero child exit as a reroute
failure and fell back to CPU. In --tauri mode the child uses exit 2 ([TAURI:NEED_SUDO])
to ask the desktop app to elevate for the target distro; capture the child's exit code
and propagate exit 2 in Tauri mode (the child already printed the NEED_SUDO line)
instead of masking it. CLI mode still falls back to CPU on a generic failure.
Tests: reroute Tauri exit-2 propagation (and non-Tauri CPU-fallback) cases;
run_func now preserves the child exit code; force-reinstall assertions for both
PowerShell installers.
Note: codex's _rr_q apostrophe finding is a false positive -- the helper already
emits POSIX-correct 'O'\''Brien' and round-trips under both sh and bash.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* setup.ps1: fix $cpuForce array collapse in the ROCm->CPU torch fallback
An if-expression assignment ($cpuForce = if ($ROCmCpuFallback) { @("--force-reinstall") })
collapses the single-element array to a scalar string, so @cpuForce splatting enumerated
it character-by-character into broken single-letter args (- - f o r c e ...), which made
uv/pip reject the install and aborted the whole Studio setup on the AMD ROCm->CPU fallback
path. Build $cpuForce as a real array assigned outside the if-expression so the splat passes
a single --force-reinstall arg. Genuine CPU-only installs stay fast (empty array, no flag).
Test now asserts the array-build form and rejects the if-expression form.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* uninstall: remove the isolated Node.js runtime (~/.unsloth/node)
The isolated Node.js runtime (install_node_prebuilt.py, added with the managed-Node
change) installs to ~/.unsloth/node in default mode -- a sibling of studio, so deleting
<studio> leaves it behind (~200MB orphaned after uninstall). Both uninstallers already
remove the other default-mode siblings (llama.cpp/.cache/.staging); add node alongside
them. uninstall.ps1 also adds it to the handle-lock sweep so a held node.exe can't block
the delete. Env/custom mode nests node under the custom root, removed with that root.
* [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>
* Package scanners: close fail-open gaps in the sdist fallback and hidden-payload paths
Follow-up hardening on the now-blocking scanners so the enforcing gate cannot
report clean while a malicious artifact goes unscanned.
scan_packages.py
- Hidden payload: also flag a network call AND an os/subprocess exec that live
only in a blanked docstring/string of an exec/eval file (the fetch-then-run
shape of an exec(__doc__) dropper). Either alone in real code was already
covered; hidden together they are the payload.
- Pinned releases fail closed: _release_files no longer falls back to the latest
artifact when a pinned version is missing or empty, so a yanked/bad pin is an
error instead of a different file being scanned in its place.
- requires_dist is read from the pinned release's metadata, not the project-level
(latest) document, so a sdist-only pin follows its own dependency tree.
- Environment markers are evaluated (PEP 508) instead of dropping any marker that
merely contains the word extra, so default-true markers like extra != 'dev' are
kept; conservative fallback keeps a dep on any uncertainty.
- Transitive recovery is a depth-bounded worklist: a wheel dependency whose own
child is sdist-only is fetched (--no-deps) and scanned, then its children are
recovered in turn, rather than being silently skipped.
scan_npm_packages.py
- Baseline keys use the package-relative path instead of the basename, so the
same basename in a different directory is not over-suppressed.
Tests cover each case; full scripts pass AST and ruff checks.
* Address review: tighten marker scope, decoy-proof the dropper check, fail closed on missing pin metadata
- Markers: keep any dep whose marker can hold on another install target
(sys_platform == 'win32', python_version == '3.13'); only drop a marker that
depends solely on extra and is false with no extra. A scanner runs on one
target but must cover code installed on others. Pure-extra markers are
evaluated against default_environment() with extra unset.
- Hidden dropper: the network+exec docstring check now inspects the removed
(blanked) span directly, so a benign visible network or subprocess call cannot
mask a payload that still lives in a docstring. Carrier checks stay
blanked-only (an in-code carrier is already caught by the normal check), so
corpus findings are unchanged.
- requires_dist: a pinned version whose own metadata cannot be fetched recovers
nothing rather than substituting the latest release's dependency tree.
- Transitive recovery: the last-ditch direct-sdist branch also chases the
recovered package's declared deps, matching the other branches.
- npm baseline: schema bumped to v2 (package-relative keys); a pre-v2 baseline
with entries is ignored (fail closed) instead of mis-applying basename keys.
Tests cover each case; scripts pass AST, ruff, and the import-hoist verifier.
* Scanner: exclude comments from hidden-payload check, flag missing pin metadata as incomplete
Hidden network+exec detection now inspects only docstring/string spans (what exec(__doc__)/exec(<str>) can actually run), so a real exec() beside comments that mention a network and a subprocess call no longer false-positives. Missing pinned-release metadata in transitive recovery records a download_error so the --with-deps path fails closed instead of treating it as no dependencies. Adds regression tests for both.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
* Fix scan_packages.py --fix crash on download_packages() tuple return
`download_packages()` returns `(results, download_errors)`, but the two
`--fix`-path call sites still treated the return value as the bare results
list. `find_safe_version` did `downloaded = download_packages(...)` followed
by `if not downloaded:` (always false: a 2-tuple is truthy) and
`for _, archive_path in downloaded:`, which unpacked the results list into
two variables -> ValueError in the normal single-archive `--no-deps` case.
`_run_fix` indexed `downloaded[0][1]`, i.e. the second archive of the results
list instead of the first archive's path -> IndexError. So `--fix` crashed
exactly when a CRITICAL finding needed remediation. The main scan path already
unpacks the tuple; this aligns the two `--fix` sites with it.
Adds CPU-only regression tests for both sites.
Closes#6412
Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
* Update scripts/scan_packages.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Update scripts/scan_packages.py
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
---------
Co-authored-by: Claude Opus 4.8 <noreply@anthropic.com>
Co-authored-by: gemini-code-assist[bot] <176961590+gemini-code-assist[bot]@users.noreply.github.com>
* Package scanners: cut false positives and make the CI gate blocking
scan_packages.py and scan_npm_packages.py red-failed on legitimate
library code, so the security-audit steps were left advisory. Reduce
the false positives at the source and flip both gates to blocking.
scan_packages.py:
- Scan code only: blank comments and bare docstrings/doctests before
matching (line numbers preserved), so prose and >>> examples cannot
trip a finding.
- Drop the platform.system() branch from the anti-analysis regex (under
DOTALL it matched across the whole file, so every cross-platform
library tripped it) and fix the dead /proc/self/status alternative.
- Add a reviewed baseline allowlist (scan_packages_baseline.json) keyed
on (package, basename, check): only non-baselined CRITICAL/HIGH exit
1, and a new kind of finding in a listed file still fails.
- sdist fallback: when --with-deps cannot resolve a shard (a sdist-only
package or a version conflict), drop to per-spec and fetch the raw
sdist from the PyPI JSON API (no pip build, no setup.py), so every
package is still scanned and no shard exits 2.
scan_npm_packages.py:
- Mirror the code-only JS/TS scanning (blank // and /* */ comments,
string/template/regex aware) and the baseline allowlist. The npm
corpus is clean today, so the baseline is empty.
security-audit.yml:
- Flip both scan steps to blocking (SCAN_ENFORCE=1), capturing the
scanner exit via PIPESTATUS so tee does not mask it.
tests/security: add coverage for the strip, baseline and sdist paths.
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
* Address review feedback on the package scanners
- Do not blank f-strings during code-only scanning (they evaluate at
import); and when a file uses exec/eval, rescan the original for
payload carriers hidden in a docstring/string so exec(__doc__) style
payloads stay visible.
- sdist fallback: recover transitive deps with their version specifier
(fetch the pinned version, not latest), and recover deps in the
--no-deps branch too so a sdist-only transitive dependency is still
scanned instead of silently skipped.
- Baseline: key by package-relative path, not basename, so a future
same-named file in another directory is not auto-suppressed.
Regenerated the baseline accordingly.
---------
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
Daniel Han