Address the Codex review round on the video/quant work:
- Companion auto-quant now honors an explicit Speed=off. Both loaders already pin the DiT dense
under an explicit off (bit-exact reference), but the unset text-encoder / VAE quant still promoted
to auto and silently fp8/int8'd the companions, breaking the bit-exact request. An UNSET speed
still auto-quantises; an explicit companion scheme still forces it.
- The HunyuanVideo joint-attention trim is a speed lever (it swaps to the fused SDPA kernel), so gate
it on a non-off speed tier exactly like the adjacent attention-backend selection -- the off path
keeps the stock dense-mask attention.
- Explicit torchao text-encoder modes (int8 / fp8_dynamic / nvfp4) now run the same kernel smoke
test the auto ladder uses. They could clear the capability gate yet fail the real GEMM on a build
where quantize_ wraps the encoder but the kernel is broken; the caster's try/except only covers the
cast, not the first forward, so the load would report engaged then crash at generation. Now it
falls back to dense. Layerwise fp8 has no torchao GEMM, so the probe is a no-op for it.
- The trim pre-hook's fallback restores the caller's original kwargs (it may have emptied the image
stream / trimmed a text stream before failing), so the stock dense-mask path runs on exactly what
it expects, matching the empty-prompt guard.
- video_speedmem_bench mirrors the loader: installs the Hunyuan trim before the backend set (gated on
an active tier) and skips the auto int8 quant when it is the fp8-denied memory fallback and dense
fits resident, so the shipped/auto rows measure what the loader actually runs.
Tests: TE explicit-mode kernel probe (+ layerwise-fp8 bypass), trim mid-trim restore, and loader-level
speed=off companion suppression + trim skip for both backends. 262 backend tests pass; ruff clean.
Two fidelity fixes to the video speed/mem bench so its numbers match production:
- _build_pipe loaded the pipeline with a scalar bf16 torch_dtype and then upcast the
VAE, which truncates the fp32-stored Wan VAE at load (a later .to(float32) only widens
the lossy values). Pin the VAE fp32 per-component like the production loader
({"vae": fp32, "default": bf16}) so the bench decodes the same weights production does.
- The persisted reference frames were written/read as a single unkeyed ref_frames.npz in
the fixed default --out dir, so a reference-less run of a different family/seed/steps/
frames/resolution scored LPIPS against a stale baseline. Key the cache by those
parameters so a run only reuses a reference computed for the same parameters.
The image load route calls engine.begin_load(..., vae_quant=request.vae_quant, ...)
uniformly for both engines, but the native SdCppDiffusionBackend.begin_load accepted
every other diffusers-only knob except vae_quant and had no **kwargs, so a native
(CPU-only / MPS / forced-native) image GGUF load raised TypeError on every request
(vae_quant is always passed, defaulting to None). Accept and ignore it like the other
diffusers-only knobs; sd.cpp has no torchao VAE quant.
test_diffusion_attention.py documents itself as hermetic with no torch/diffusers
needed, but the HunyuanVideo trim tests added a module-level 'import torch' that
aborted collection of the whole file when torch is absent. Move those 15 tests to
test_diffusion_attention_trim.py (which declares the torch dependency) so the
attention-backend policy tests stay collectable and runnable without torch.
HunyuanVideo-1.5's DiT runs a joint [video; text] self-attention and, on every
block and step, builds a dense [B,1,N,N] boolean mask so the video never attends
to the padded text. A dense bool attn_mask disables every fused SDPA kernel
(flash rejects it; cuDNN and memory-efficient fall back), so the attention runs
the slow math-style path: at the production shape (121 frames, 480p, N about 50k)
one attention call is ~421ms with the mask vs ~19ms with attn_mask=None. The text
is ~99.5% padding (a t2v prompt fills ~9 of ~1985 slots), so nearly all of that
cost is spent masking padding.
install_hunyuan_attention_trim installs an eager forward pre-hook that drops the
all-zero image stream (t2v) and trims the mllm/byt5 text streams to their
globally-valid columns, plus a null-mask attention processor that runs
attn_mask=None once no partially-padded column remains (the batch-1 /
per-guidance-branch case) and otherwise delegates to the stock dense-mask
processor. The model already zeroes and masks the padded text and discards its
attention output (only the video split feeds proj_out), so removing it is exact
for the video; the only numeric change is the SDPA kernel (masked fallback to
fused). Measured on a B200: 23.3s to 1.3s per DiT forward at 121 frames (~18x with
regional compile, 0 graph breaks); per-forward cosine 0.99998 vs stock; equal
distance to an fp32 reference (LPIPS fp32-vs-stock 0.292, fp32-vs-trim 0.307), so
it is not less accurate than the current bf16 default.
Wired auto-on for HunyuanVideo-1.5 in the video loader, before the attention
backend set so the requested kernel pins onto the new processors; a no-op for
every other family and reversible (stock dense-mask path on any anomaly). Adds
hermetic tests and the diagnostic/validation scripts.
Root-caused "HunyuanVideo-1.5 int8 is slower than dense" with a per-forward profiler
(scripts/hunyuan_int8_profile.py, dynamo-reset, back-to-back on a clean B200): int8 compiles
cleanly (0 recompiles, 0 graph breaks, steady 268.3 ms/forward) and is only ~7% slower than dense
+ regional compile (250.5 ms/forward), not the 38% a contended-GPU bench run suggested. int8 is
also less accurate (LPIPS 0.085 vs dense+compile 0.037). So for a family where fp8 is denied
(Hunyuan black-frames on per-row fp8), int8 is a MEMORY lever, not a speed win, yet the auto-quant
default quantised it even when the dense DiT already fit resident.
Fix: is_int8_memory_fallback(target, family) is True only when AUTO quant lands on int8 as a
denied/black-frame fallback on a data-center, fp8-capable GPU (fp8 would be the arch pick but is
denied for the family). The video loader now skips the auto-quant and runs dense+compile when that
holds AND the bf16 memory plan already fits resident (offload_policy == none), so there is no new
OOM risk. Scoped tightly: only an AUTO request (explicit int8/fp8 honored), only int8-fallback
families (Wan / LTX resolve to fp8 -> keep quantising), only data-center fp8-capable parts (consumer
GPUs and pre-Ada, where int8 is a genuine accelerator, keep int8), and only when dense provably
fits; a memory-constrained plan still quantises. Result: Hunyuan on a resident-fit B200 now runs
faster AND more accurate, quantising only when memory is the constraint.
Also resets dynamo per config in the video bench (so compiled graphs cannot leak across configs in
one process) and adds the per-forward profiler used for the diagnosis.
Adds wan2.2-t2v-a14b and hunyuanvideo-1.5-720p to the bench family table and makes _apply_levers
quantize / compile BOTH experts of a dual-expert MoE (Wan2.2-A14B) via a _SecondExpertView proxy
that mirrors the loader's _SecondDiTView, so A14B latency and accuracy are measured on the real
two-DiT path instead of only the first expert.
Used to validate that every video family is on the fastest DiT quant scheme that is not less
accurate than its alternative (B200, 512x320, 25 frames, 30 steps, no cache, LPIPS vs dense bf16):
- Wan2.2-TI2V-5B fp8 49.9 ms/step vs int8 64.6 vs dense 59.8; LPIPS 0.129 vs 0.180
- Wan2.2-T2V-A14B fp8 195.8 ms/step vs int8 193.5 vs dense 369.1; LPIPS 0.288 vs 0.349 (both experts)
- LTX-2 / LTX-2.3 fp8 133.5 ms/step vs int8 138.4 vs dense 204.7; LPIPS 0.026 vs 0.027
- HunyuanVideo-1.5 fp8 is black (denied, not localizable); int8 21.2 s vs dense+compile 15.4 s -- a
memory saving at a speed cost, so int8 stays only because fp8 is impossible there.
fp8 is faster than dense AND more accurate than int8 on all three Wan/LTX families (the two Wan ones
via the condition_embedder exclude; LTX-2 needs none -- its conditioning has no zero-amax padding
rows). The auto-ladder + deny + exclude already select exactly these schemes, so no scheme-selection
change was needed; this commit is the bench faithfulness improvement that let the campaign confirm it.
The Wan fp8 black frame was root-caused (scripts/fp8_layer_ablation.py,
measured on B200 with the production torch._scaled_mm path): per-row fp8
scales each activation row by row_amax/448, and the text prompt is padded to
512 tokens (~all padding for a short prompt), so condition_embedder's text
embedder divides a zero padding row by a zero scale, which infs and renders
every frame black. That embedder's bias makes every downstream row non-zero,
so the whole 30-block attn1/attn2/ffn stack is fp8-clean (fp8-except-
condition_embedder measured cosine 0.9998 vs bf16, 0 non-finite; fp8-
everywhere is 100% non-finite).
So the blanket fp8 deny was heavier than needed for Wan. Remove fp8 from the
Wan deny and keep only condition_embedder in bf16 via a new
_FP8_FAMILY_EXCLUDE_NAME_TOKENS; auto now restores fp8 (the Blackwell ladder
head) for Wan2.2-TI2V-5B and -T2V-A14B (shared DiT class and padded-text
conditioning). Full-generation check (512x320, 25 frames, 30 steps, cache on
and off): mixed-fp8 is non-black (mean luma 182.6 vs dense 181.2), more
accurate than int8 (LPIPS 0.129 vs 0.180 no-cache, 0.224 vs 0.251 with
FBCache), faster (49.9 vs 64.6 ms/step; int8 was a per-step regression vs the
59.8 ms/step dense), at the same memory (19.34 GB, both -20% vs dense).
HunyuanVideo-1.5 keeps the fp8 deny: its MMDiT masks the padding text tokens
to zero inside every block, so the per-block context stream (add_*_proj /
to_add_out / ff_context) regenerates zero rows layer after layer (fp8 on only
the main blocks is 100% non-finite) so no small exclude set exists and int8
stays. mxfp8 / nvfp4 remain denied for Wan (same per-row scaled_mm family, not
separately validated).
exclude_tokens_for_scheme now takes an optional family, threaded through the
runtime quantiser and the offline prequant builder + validator so offline ==
runtime (a stale Wan fp8 checkpoint baked without the exclude is rejected and
re-quantised rather than loaded). Adds scripts/fp8_layer_ablation.py (the
per-layer ablation probe) and a mean-luma black-frame metric plus mixed-fp8
vs int8 configs to the video bench.
Measured the fp8 DiT auto-quant path across the remaining dense-pipeline video families on
B200 (production torch._scaled_mm per-row fp8, no MSLK):
- HunyuanVideo-1.5 (480p + 720p repacks): every frame black (mean luma 0.0, LPIPS 0.82);
int8 is clean (mean 102.7 vs dense 99.9). Same failure as Wan / qwen-image.
- LTX-2: fp8 renders clean (mean 153.7, matches int8's 157.7) -- NOT a black-frame family.
So deny fp8/mxfp8/nvfp4 for hunyuanvideo-1.5 and hunyuanvideo-1.5-720p (fall to int8), and
deliberately leave LTX-2 on fp8. The deny stays measured per family, not a blanket video rule:
a blanket deny would have wrongly forced LTX-2 off fp8. Adds a Hunyuan deny test that also
asserts LTX-2 keeps fp8; 49/49 transformer-quant tests pass.
video_speedmem_bench.py gains guidance_via_guider support (HunyuanVideo-1.5 sets CFG on a
guider component and its __call__ takes no guidance_scale / callback_on_step_end), so the
harness can drive Hunyuan the same way the loader does.
video_speedmem_bench.py drives the real video-loader lever functions
(quantize_transformer / quantize_text_encoders / quantize_vae / apply_speed_optims
/ apply_attention_backend / apply_step_cache) with the loader's own defaults, so each
measured config reflects a real load. It decomposes the video speed/memory stack
(compile, cuDNN attention, First-Block-Cache, DiT/TE/VAE quant) with per-step latency,
peak resident GB, and per-frame LPIPS vs a bit-exact reference. This is the harness that
surfaced and validated the Wan fp8 black-frame fix.
quant_speedmem_bench.py gains the DiT-quant mode (dense vs fp8/int8/mxfp8 speed, peak
memory, and LPIPS vs the dense render) plus the shared LPIPS(AlexNet) helper.
The dense video default engages transformer auto-quant, and on Blackwell the
auto ladder leads with fp8. On the Wan DiT the production per-row fp8 path
(torch._scaled_mm) renders every frame black (mean luma 0.0 at 512x320 and
704x480, LPIPS ~0.80 vs bf16): Wan's activation outliers exceed per-row fp8's
range, the same failure already denied for qwen-image. First-Block-Cache then
over-caches the degenerate activations (per-step collapses to ~10ms),
compounding it.
Add the Wan families (wan2.2-ti2v-5b, wan2.2-t2v-a14b, same WanTransformer3DModel)
to _FAMILY_SCHEME_DENY for fp8/mxfp8/nvfp4 so auto falls through to int8, which is
clean on Wan (per-token, outlier-robust), saves the same weight memory on the DiT,
and lets First-Block-Cache engage normally instead of over-caching. mxfp8/nvfp4 are
denied alongside fp8 conservatively so auto lands on the battle-tested int8; they
can be re-enabled per family once validated in-bar, like the nvfp4 auto-ladder TODO.
Validated on B200: the shipped video default now selects int8 for the Wan DiT and
renders clean frames (mean 172.6) at 15.6 GB resident (down from 24.2 GB dense),
with First-Block-Cache engaged. Adds two deny tests; 48/48 transformer-quant tests pass.
A B200 speed/memory sweep (new scripts/quant_speedmem_bench.py) shows the VAE quant
win is a video story. Image AutoencoderKLs are ~0.15-0.26 GB, so fp8 saves ~0.1 GB
and only slows their tiny decode (+6-16%); the video Conv3d VAEs are ~2.5 GB and
halve to ~1.2 GB at ~2% decode cost. So VAE auto now only engages above a ~1 GB size
floor: small image VAEs stay dense (faster decode, no quant quality risk), video VAEs
still quantize. An explicit fp8 / fp8_dynamic request skips the gate (opted in).
The same sweep confirmed the text-encoder default is already right: fp8_dynamic is
E2E-neutral (denoise per-step unchanged; +2% one-time encode) and, by hidden-state
cosine vs bf16, marginally more accurate than layerwise fp8 -- so that default is left
as is. Tests cover the gate (small skipped, large quantized, explicit bypasses).
Five diffusion source and test files landed without the standard two-line SPDX
header the rest of studio/backend carries. Prepend it (matching the sibling
convention) so the whole backend is uniformly licensed. Header-only, no code
change.
A B200 decoded-image LPIPS/SSIM sweep vs the dense bf16 VAE (new
scripts/quant_accuracy_sweep.py) settles the two VAE schemes:
- Layerwise fp8 (storage-only) holds across families (SSIM >= 0.977 on all but
SDXL), so auto now engages layerwise fp8 ONLY. For a VAE decode (a few percent
of end to end) fp8_dynamic's fp8-matmul speedup over storage fp8 is negligible,
so auto never takes the accuracy risk.
- fp8_dynamic (torchao PerTensor conv compute) is in-bar on only FLUX.2 and
Hunyuan and out-of-bar or catastrophic elsewhere (Qwen-Image SSIM 0.46), so it
is now an explicit opt-in, re-gated by a per-family deny list derived from the
sweep. SDXL denies both schemes (its small VAE stays dense).
Also fixes a real decode-time crash: torchao 0.17's fp8 conv kernel rejects
pointwise (1x1 / 1x1x1) convs ("Activation and filter channels must match"), so
an explicit fp8_dynamic request cast fine then threw at the first decode on most
families. The conv filter now keeps 1x1 convs dense, the smoke probe uses a
spatial 3x3 conv (so it exercises the path that actually runs), and an explicit
fp8_dynamic request runs that probe before casting.
Tests updated for the fp8-only auto ladder, the 1x1 exclusion, the explicit
probe gate, and the shipped deny list.
The transformer and text encoder auto-quantize; the VAE stayed dense. VAEs are
convolutional, so torchao int8 (Linear/2D-only) does not apply, but
Float8DynamicActivationFloat8WeightConfig quantizes Conv2d/Conv3d weights with
PerTensor granularity (auto-skipping convs whose channels are not a multiple of 16,
so the 3-channel RGB head stays dense). New diffusion_vae_quant.py offers two
schemes: fp8_dynamic (torchao conv compute fp8, cc>=8.9, resident) and fp8
(diffusers layerwise storage cast, any conv, survives offload); no int8 (no Conv3d
int8 kernel). select_vae_quant_scheme walks (fp8_dynamic, fp8) with a live conv
smoke probe, an offload gate, a per-family deny list, and a force_fp32 gate; the
image + video loaders map unset vae_quant to auto, skip the vae_force_fp32 Wan
families, and record the engaged scheme. Guards _align_vae_dtype to skip the
img2img/inpaint re-cast when the VAE is quantized (its fp8 tensor subclasses reject
.to(dtype=)). Verified on a B200: %16 Conv2d/Conv3d/Linear -> Float8Tensor, conv_out
dense, forward runs.
The transformer already defaults to auto-quant (fp8/int8); the companion text
encoder was opt-in and stayed dense bf16 unless a scheme was named, even though it
is often the largest resident component. Add an auto policy mirroring the
transformer's ladder: select_te_quant_scheme walks a per-capability ladder
(data-center fp8-GEMM: fp8_dynamic -> int8 -> layerwise fp8; Ampere: int8 -> fp8),
reorders int8 first on consumer GDDR parts, falls to layerwise fp8 under group
offload (the only offload-safe cast), only picks int8 for a family with a measured
keep-bf16 schedule, honors a per-family deny list, and smoke-probes the torchao
kernel so a missing build degrades gracefully. The image + video loaders now map an
unset text_encoder_quant to auto (explicit none/off stays dense; a named scheme is
forced), so the shipped default quantizes the encoder to the fastest accurate
scheme for the GPU. Records the engaged scheme in the image resolved-record too.
Verified on a B200: auto -> fp8_dynamic, offload -> layerwise fp8.
The auto ladder still listed nvfp4 in the Blackwell tier even though the comment
above it says nvfp4 must never be the auto pick for diffusion: at the DiT's real
shapes it is slower (0.81x end-to-end on Z-Image 1024px) and less accurate (LPIPS
0.166 vs fp8's 0.044) than fp8. It was unreachable in practice (fp8 precedes it
and the only fp8-denied families also deny nvfp4), but the entry contradicted the
stated intent and left a latent path to the worse scheme. Drop nvfp4 from the
ladder so Blackwell auto is fp8 -> mxfp8 -> int8, keeping the previous line
commented with a TODO to restore it once the FP4 GEMM wins at these shapes. nvfp4
stays fully available as an explicit transformer_quant="nvfp4" request.
The LoRA row div was keyed on sel.id, the editable repo-id field. Typing the
first character changed the key from the fallback index to that character, which
remounted the row and dropped focus on the input, so only one character was ever
captured and manual entry of a repo id was impossible (paste still worked). The
list is index-addressed (every mutation uses j === i) and rows are removed
explicitly via the delete button, so the index is the stable key, matching the
reference-image rows above.
torch.cuda.is_bf16_supported() defaults to counting pre-Ampere bf16 EMULATION as
supported, so on a T4/V100/RTX 20xx the DiT-training bf16 gates all passed even
though the trainer requires native Ampere-or-newer bf16: /diffusion/info advertised
the DiT precision modes, /diffusion/start's preflight let the run through and freed
resident GPU models, then the trainer child hit the real unsupported bf16 path. The
inference device resolver already fixed this (issue #6658) by gating NVIDIA on
capability major >= 8; the training path never got it. Add a shared
native_bf16_supported() helper (NVIDIA cap major >= 8; ROCm keeps the trustworthy
is_bf16_supported()) and use it in the three DiT bf16 sites -- train_precision_modes,
bf16_unsupported_reason, and the trainer guard -- so a pre-Ampere card is offered
nf4 only and never advertises/evicts-then-fails. Tests now exercise the emulation
case (is_bf16_supported True but capability < 8).
img2img and inpaint take their output size from the uploaded image and only
snap it to a multiple of 16, so an ordinary phone photo (up to the 4096/side
decode cap, 4x the txt2img 2048 ceiling and ~16x the area) drove an OOM-scale
latent and an opaque 500 on a normal card, while txt2img, upscale, edit, and
FLUX.2-klein inpaint are all already megapixel-bounded. Clamp the init longest
side to 2048 (the txt2img ceiling) before deriving width/height; edit is exempt
since its pipeline resizes to ~1MP internally.
_cast_nvfp4 quantized every nn.Linear with no filter, unlike the int8 and fp8
torchao text-encoder modes which exclude the VLM vision tower / lm_head / T5 wo.
On qwen-image / qwen-image-edit that 4-bit quantized the Qwen2.5-VL image tower,
degrading the edit/image conditioning the sibling schemes protect. Apply the
same make_filter_fn exclusion (require_bf16, mirroring _cast_fp8_dynamic).
union_control_mode() only matched the short catalog id, so a client naming a
curated union ControlNet by its bare HF repo id (the form resolve_controlnet
documents and accepts) got None and the generate path omitted control_mode,
which makes FluxControlNetModel.forward raise controlnet_mode cannot be None.
Fall back to matching repo_id against the curated entries so the bare repo id
resolves to the same union mode; non-union bare repos still return None.
The images page also wired Reapply for a resident single_file model with no
checkpoint filename (status carries none), but the backend rejects a
single_file/gguf load without a filename, so clicking Reapply 400'd. Narrow the
resident-Reapply wiring to pipeline (the one kind that needs no filename),
matching the existing GGUF handling, so single_file stays a no-op instead of
erroring.
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.
The SDXL trainer drew min(train_batch_size, len(pairs)) indices, so a dataset with
fewer images than the batch trained at a smaller effective batch than configured
while the scheduler and samples-per-second still assumed the full batch. The shared
PermutationBatchSampler already refills across permutation cycles to return exactly k
indices, and the DiT trainer calls it with the full batch size, so drop the clamp and
pass train_batch_size through for parity and to honor the configured batch.
* Speed up Studio startup path
* Studio: recheck managed binary executability on preflight cache hit and ignore stale unauthenticated platform fetches
Preflight: a matching capability cache fingerprint no longer skips the
runnability check when the managed binary's executable bit was cleared
(size and mtime unchanged, since chmod bumps ctime not mtime). The cache
fast path now confirms the binary is still executable, otherwise it falls
back to the CLI help probe so preflight reports Stale and can repair,
instead of returning Ready and failing later at backend start. Adds a
regression test.
Frontend: now that first render is no longer gated on fetchDeviceType,
the initial unauthenticated health call can resolve after an
authenticated platform fetch. Guard the store so a late unauthenticated
or failed non-forced response cannot overwrite an already authoritative
device type, tunnel URL, or secure flag. Forced refreshes and the first
unauthenticated load are unaffected.
* Studio: use access(X_OK) for the preflight cache executability guard
A mode bitmask treats any execute bit as launchable, but the executable
bits can be set only for another owner or group, or be denied by an ACL,
so the current user could still hit PermissionDenied at launch and the
cached fast path would wrongly return Ready. access(X_OK) checks real
executability for the calling user, so an ownership or permission change
correctly falls back to the CLI help probe and the Stale repair path.
* Studio: ignore any stale non-forced platform fetch once authoritative
Extend the platform store guard so a non-forced health response never
overwrites an already authoritative result, not only unauthenticated
ones. With a saved token the post-render non-forced request can be
authenticated but older than a later forced refresh that already picked
up the tunnel URL and secure flag; if that earlier request resolves last
it would null those fields. Now any non-forced response is dropped once
the store holds a server-reported platform. Forced refreshes and the
first authoritative write are unaffected.
* Studio: run the managed CLI help probe before trusting the preflight cache
Restore running the managed CLI help probe before returning Ready from
the desktop capability cache, so a managed install whose venv interpreter
or a runtime dependency is broken (while path, size, mtime, and markers
are unchanged) is reported Stale for repair rather than proceeding to a
backend start that cannot spawn. The capability cache still skips the
heavier desktop-capabilities probe on a hit, so a warm cache runs one
probe instead of two. Removes the executable-access shortcut, which the
help probe now subsumes.
---------
Co-authored-by: Daniel Han <danielhanchen@gmail.com>
A scanned On-Device model's id is its full filesystem path, and the family-token
matcher treats any path segment as a hint, so a folder under a family-named
parent (e.g. /models/qwen-image/misc) tagged an unrelated single-file as
text-to-image. That surfaced it in the Images picker, and a load would evict the
GPU owner before from_single_file failed on the unrelated weights. Match on
Path(model.id).name so only the leaf name counts, for both the image and video
task-tag loops (the video loop is gated on the same detection).
The native sd.cpp one-shot engine re-reads its companion VAE and text-encoder
files from the HF cache on every generation, but the delete-cached guard only
compared the loaded main diffusion repo_id. Deleting a companion repo such as
comfyanonymous/flux_text_encoders while a FLUX GGUF was loaded therefore
succeeded and broke the next generation. Add a loaded_repo_ids() accessor to the
native backend (mirroring loading_repo_ids(), reconstructed from the committed
family's VAE + text-encoder repos) and have the guard refuse those companions
too. Server mode and the diffusers engine hold companions in-process/VRAM, so
they are unaffected.
Video: a local FILE picked for a gguf/single_file load is handed straight to the
loader (the resolver returns the file itself, ignoring gguf_filename), so its own
suffix must match the kind. A .gguf picked as single_file (or a .safetensors picked
as gguf) slipped past the gguf_filename checks, evicting the resident GPU owner
before failing in from_single_file. Reject the mismatch in validate, before the handoff.
Engine router: publish the newly selected engine only after the old one finishes
unloading. The arbiter's diffusion evictor unloads the active engine, so flipping the
active name to the new (empty) engine first let a concurrent chat/video acquire evict
that empty engine and take the GPU while the old model was still freeing VRAM.
The two candidate-resolution tests exercised the real cache-disk gate, so a
small CI disk (< transient + 10 GiB free) dropped the candidate and the tests
failed non-deterministically across runners. Neutralize the disk probe in the
shared selector helper; the two disk-gate tests re-patch it to cover the gate.
Two evict/OOM fixes on the diffusion load paths:
- The video load moved a pipeline onto the GPU (apply_memory_plan) and
committed it while holding no lock, so an unload / GPU-arbiter eviction --
which bumps the load token and then barriers on _generate_lock before
freeing -- could hand VIDEO to chat/images and let the new owner allocate
concurrently with the in-flight placement, OOMing. Hold _generate_lock
across placement + the locked commit, mirroring the image backend, so an
evicting owner waits until this worker's placement is torn down or
committed. Lock order stays _generate_lock -> _lock (unload takes _lock
then releases it before the barrier), so there is no deadlock.
- resolve_local_single_file reinterpreted an On-Device folder as a base
single_file load whenever it held exactly one .safetensors, so a PEFT LoRA
adapter folder (adapter_config.json + adapter_model.safetensors) with a
family-token name was picked as a base checkpoint, evicting the resident
model before from_single_file failed on the adapter weights. Skip adapter
folders (adapter_config.json) and the adapter_model basename so the pick
stays a pipeline load and 400s in validation, before the GPU handoff.
Adds regression tests for both.
The video load-request preflight gated its local-pipeline check on
root.is_dir(), so a bare local file (e.g. /models/ltx-2.safetensors) sent
with model_kind=pipeline skipped it, passed validation, and the route then
evicted the resident GPU model before from_pretrained failed on the
non-directory path. Gate on root.exists() instead (mirroring the image
loader's diffusion.validate_load_request), so a local file is rejected up
front. Add a regression test.
* fix: match qwen3-thinking chat template double-newline in response pattern
The Qwen3-thinking chat template generates `<think>\n\n` (double newline)
after the think tag, but `train_on_responses_only` was looking for
`<think>\n` (single newline).
`\n\n` is token 271 while `\n` is token 198 -- different tokens, so the
pattern match in `train_on_responses_only` fails, masking ALL tokens and
dropping 100% of training samples.
Update the response pattern from `<think>\n` to `<think>\n\n` to match
what the actual qwen3-thinking template generates.
Fixes#6919
* fix qwen3 thinking response marker
---------
Co-authored-by: Ayushman Paul <ayushman@HP>
Co-authored-by: Etherll <61019402+Etherll@users.noreply.github.com>
Two evict/corruption fixes surfaced by review of the diffusion training path:
- krea-2 sets force_bf16 in the DiT trainer spec, but the route-level
_FORCE_BF16_FAMILIES preflight listed only qwen-image and z-image, so a
krea-2 start with mixed_precision=fp16 passed the route check, reserved
training and evicted resident GPU models, and only the child trainer then
raised. Add krea-2 to the set and a drift-guard test asserting it equals
the trainer specs whose force_bf16 is set.
- The dataset-upload same-stem duplicate check compared stems
case-sensitively, so on case-insensitive filesystems (Windows / default
macOS) sample.png and Sample.jpg both passed even though their caption
sidecars sample.txt / Sample.txt resolve to the same file, silently
sharing and corrupting one caption. Compare stems and the same-name guard
with casefold at both the on-disk and in-batch sites.
* show chat by by last activity
* Update chat thread updated_at logic and enhance sidebar chat item handling
* [pre-commit.ci] auto fixes from pre-commit.com hooks
for more information, see https://pre-commit.ci
---------
Co-authored-by: Lee Jackson <130007945+Imagineer99@users.noreply.github.com>
Co-authored-by: pre-commit-ci[bot] <66853113+pre-commit-ci[bot]@users.noreply.github.com>
The /delete-cached guard refused deleting a cached repo whose id is a bare
prefix of the loaded one: with Qwen/Qwen-Image-2512 loaded, deleting the
separate cached Qwen/Qwen-Image returned 400 "Unload the model before
deleting" because loaded_id.startswith(repo_id) matched across the repo
boundary. Both are real, independently cached catalog artifacts, so a
supported delete was wrongly blocked.
Add a /-boundary aware _loaded_id_matches_repo helper (mirroring
hub/services/models/deletion.py) and use it at all six guard sites (chat,
Images, Video, and their in-flight loading-id loops), so only the loaded
repo itself or a file within it blocks deletion. Add a regression test
covering the sibling-prefix pair.
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).
test_delete_keeps_sidecar_listable_when_mp4_unlink_fails patched os.unlink, but delete()
calls path.unlink(). On Python 3.10 Path.unlink dispatches through a cached _accessor bound
to os.unlink at import, so patching os.unlink had no effect there and the simulated mp4
unlink wrongly succeeded (the assert delete() is False failed on 3.10 while passing on
3.11+, where Path.unlink looks up os.unlink dynamically). Patch Path.unlink itself, which
delete() invokes directly on every supported Python version. Test-only; no behavior change.
union_control_mode fell back to control_mode=0 (the canny head) for ANY unmapped control
type, so a typo'd or unsupported value like 'detph' silently conditioned the map as canny
instead of failing. preprocess_control passes non-canny maps through unchanged, so that map
would be interpreted under the wrong mode with no error. Now only 'passthrough' (or an empty
type) keeps the deliberate mode-0 default; any other unknown type raises ValueError, which
the generate route maps to a 400. Known modes are unchanged.
Follow-up to removing piecewise_constant from the trainable scheduler allow-list: the
DiffusionTrainingStartRequest.lr_scheduler Literal still advertised it, so a client that
picked it straight from the schema passed request validation and then hit the 400 from
normalized(). Remove it from the enum too so the API only offers schedulers the trainers
can actually run, and add a test asserting the enum never advertises a scheduler outside
the validation allow-list.
* Studio: account for DeepSeek-V4 compute buffer in context auto-fit
DeepSeek-V4-Flash's lightning indexer plus compressed sparse attention reserve a
large context-scaling compute buffer that _compute_buffer_ctx_bytes did not model
(the KQ-mask and dequant-scratch rates both miss it, even with an f16 cache).
Measured on UD-Q4_K_XL at ub 512 it is about 65.5 GiB at 1M context, which the
mask estimate puts near 1.5 GiB, so the auto-fit kept the full 1M train context
and llama-server OOM'd allocating the ~70 GB buffer, then spilled to CPU (~4
tok/s). Add a deepseek4-gated flat plus per-token term so the fit caps the context
(about 256k on a B200) and the model stays fully on GPU.
* [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>
piecewise_constant is the only diffusers scheduler that needs a step_rules string, and
neither diffusion trainer passes one (get_scheduler is called with only warmup/training
steps, and there is no config field for it). Accepting it let /diffusion/start pass
normalized(), free the resident GPU workloads, spawn the trainer, and only then crash in
the subprocess (get_piecewise_constant_schedule does step_rules.split(",") on None) -- the
exact evict-then-fail the up-front validation exists to prevent. Reject it now with a clear
400. The remaining six schedulers all run with only warmup/training steps.
