The prewarm registered its cancel event in _active_generate_cancel, but a
begin_generate arriving mid-warmup overwrote that slot with its own event
and then queued its worker behind the full warmup on _generate_lock. From
that point unload and cancel_generate signalled the wrong event, so the
warmup could no longer be aborted and the first real request waited out
the 9-54s the prewarm exists to hide.
Track the prewarm's event in a dedicated _prewarm_cancel slot (cleared
identity-checked alongside _active_generate_cancel) and signal it from
begin_generate before registering the real job's event, and from direct
generate() calls that skip begin_generate. The warmup then aborts at its
next step boundary and the real job takes the lock, while unload/cancel
keep working against whichever run is actually active.
vLLM and SGLang finish every compilation at server startup (dummy batches
through each compiled shape) so no request ever pays a compile mid-serving.
The video backend's compiled tier instead paid a first-generation extra after
every restart: ~54 s cold and ~11.3 s even with a warm Mega-cache bundle (the
residual is dynamo tracing plus cudnn.benchmark autotune, which the bundle
cannot carry).
After a compiled DEFAULT-tier resident load commits, a daemon thread now runs
one tiny throwaway generation (192x128 snapped, 4k+1-lattice 9 frames, 2
steps) under the generate lock. The default tier compiles dynamic=True, so
the small trace serves every later resolution. Measured through the real
backend (HunyuanVideo-1.5-480p, B200, 480x288/17f/30 steps):
warm bundle: first-generation extra 11.3 s -> 2.1 s (9.6 s background warmup)
cold start: the full compile moves off the user's first request entirely
(14.5 s background; first generation extra 2.1 s), and the
warmup persists the Mega-cache bundle itself
steady state: unchanged (2.4-2.5 s per 30-step clip in every phase)
Exactly lossless by construction: the warmup only changes when compilation
work happens. It resets its step-cache residuals, the real generation seeds
its own generator, and no process-wide flag is touched.
The warmup registers itself as the active cancellable job, so unload, a new
load, or cancel_generate abort it at a step boundary (verified: unload 2 s
into a running prewarm returns in 6.7 s with the warmup cancelled). It yields
untouched when a real request arrived first and is token-scoped against
superseded loads. Gated per family (supports_compile_prewarm), skipped for
speed=max (static per-shape graphs a warmup shape cannot serve), offload
(every warmup forward would stream the DiT over PCIe), and CFG parallel (its
planner owns compile-sensitive runs); the UNSLOTH_DIFFUSION_COMPILE_PREWARM
kill switch disables it. The decision and reason land in the resolved record.
Tests: +4 hermetic (decision gates, engaged-load spawn with snapped tiny
shape, skip-without-compile, yield to generations / stale tokens); related
backend set 551 passed; ruff clean.
The pre-warmed torch.compile cache (diffusion_compile_cache.py: fingerprinted
bundles over torch.compiler.save/load_cache_artifacts plus a persistent per-key
TORCHINDUCTOR_CACHE_DIR) was wired into the image backend only, so every video
load re-paid the full first-generation compile after every process restart (the
stock inductor dir lives in /tmp).
video.py now mirrors the image backend exactly: compile_cache.begin runs after
the attention-backend set (the fingerprint keys on the engaged kernel) and
before apply_speed_optims on a compile-eligible default/max tier, keyed on the
same fullgraph decision as the compile itself (an engaged or still-toggleable
step cache and a planned offload both drop it); the context is committed to
_VideoLoadState and compile_cache.save persists the bundle after the first
successful generation (env-gated distributor / first-run-warm mode);
_teardown_state restores the inductor dir, and a token-scoped
_rollback_precommit_compile_cache covers loads that die before the state
commit, mirroring the globals and CFG-parallel rollbacks.
Measured on HunyuanVideo-1.5-480p through the real VideoBackend (B200,
480x288/17f/30 steps): the first-generation extra drops from 107.5 s cold to
13.8 s when the 12.8 MB bundle loads into a fresh inductor dir (0.10 s load)
and to 11.7 s from the persistent per-key dir alone; through the wired
production path a restart lands at 10.5-10.8 s (bundle-only included) vs
86.5 s cold. Steady state is unchanged (2.4-2.6 s), and the loaded artifacts
are the same bits a local compile would produce, so numerics are untouched.
Tests: begin/save/restore lifecycle with the fullgraph keying, the Speed=off
and compile-ineligible skips, and the token-scoped pre-commit rollback.
Only settle the CFG-parallel dispatch key after a run that actually routed
the replica: a guidance-near-1 generation disables the overlap without
warming the replica, so its completed key must not unlock thread dispatch
for the next CFG-enabled run at the same shape (that first compile has to
stay serialized).
Restore the process-global thread-safe cuDNN attention patch when the
CFG-parallel install fails after the patch landed: no proxy is committed on
that path, so teardown would never reach it and later single-device
generations would keep running the direct aten replacement.
Tear down a CFG-parallel proxy installed by a load that is cancelled or
fails before the _VideoLoadState commit: the proxy owns a daemon worker,
the DiT replica's VRAM, and possibly the cuDNN patch. The load stashes the
proxy pre-commit and _run_load's error handler rolls it back, token-scoped
exactly like the speed-globals rollback.
Re-engage an EXPLICIT magcache choice when the actual step count differs
from the configured one, so the ratio curve, retention window, and skip
budget are re-interpolated over the real schedule (the on/off choice never
changes); auto already re-engaged via maybe_toggle_step_cache. The step
marker comparison uses endswith so #s5 cannot match inside #s50.
Bench fidelity: the e2e auto row quantizes companions before CUDA placement
(mirroring the loader, so load_peak_gb records the measured configuration),
the video bench clears step-cache residuals before every generation exactly
like VideoBackend.generate, and the image-interface dit/e2e modes reject
video families with a pointer to video_speedmem_bench.py.
Extends the HunyuanVideo-1.5 round-1/2 optimization levers to wan2.2-ti2v-5b,
wan2.2-t2v-a14b (dual-expert MoE) and ltx-2, shipping only what beats the
incumbent on the measured accuracy-speed frontier (B200, LPIPS(AlexNet)
pairwise vs the same uncached compiled stack at identical seed/settings).
- Wan2.2-TI2V-5B auto step cache switches FBCache to calibrated MagCache:
balanced (0.12, 3, 0.2) measures 1.65x at pairwise LPIPS 0.034 vs the
incumbent FBCache 0.08 at 1.49x/0.031, and 1.73x/0.044 vs 1.71x/0.083 at
the fast points (FBCache error grows unboundedly past its threshold while
MagCache's budget caps it). A 50-step calibrated curve ships; cond/uncond
branches agree within 0.0008 so one curve serves both CFG contexts.
- Per-expert MagCache plumbing for dual-expert MoEs: the experts split the
schedule at the boundary timestep (Wan2.2-A14B: 16 + 34 of 50) and the hook
counts each expert's own forwards from 0, so a shared full-schedule curve
would be misaligned for both. apply_step_cache / maybe_toggle_step_cache /
the loader now thread an expert name; a second expert resolves
family::transformer_2 curves and sub-curves scale their configured step
count by steps/50. Single-DiT behaviour unchanged.
- Wan2.2-A14B keeps FBCache: with per-expert curves, FBCache 0.12 at
2.88x/0.128 dominates balanced MagCache (1.80x/0.145) and FBCache 0.08 sits
at 1.28x/0.098; the 16-step high-noise expert starves MagCache's skip
budget. No calibrated curve ships, so an explicit magcache request runs
uncached with a warning instead of engaging a measured-worse mode.
- Wan2.2-A14B TE auto quant resolves dense: TE fp8_dynamic alone costs
pairwise LPIPS 0.1195 for a 1.03x once-per-generation encode (146.7 to
142.7 s e2e). Wan2.2-TI2V-5B shares the UMT5 encoder but stays quantized
(0.0396 pairwise at a real 1.09x on its much faster DiT).
- LTX-2 TE fp8_dynamic family-denied: torchao per-row compute fp8 on the
Gemma3-27B encoder black-frames the whole clip (mean luma 137.9 to 0.0,
LPIPS 0.78; reproduced compiled and eager), while layerwise fp8 is
near-lossless (pairwise 0.0043) at the same shrink, so auto falls through
to it and explicit fp8_dynamic requests are refused.
- LTX-2 step caching deliberately stays unregistered, now documented on
_EXTRA_BLOCK_METADATA: the block returns a joint (video, audio) stream pair
and both cache hook families would substitute text embeddings into the
audio slot on every skipped step; a dual-stream cache is required, and the
distilled checkpoints run below FBCACHE_MIN_STEPS anyway.
- Compile parity (emulate_precision_casts) verified family-neutral and kept
global: wan5b 1.75x/0.0029 on vs 1.54x/0.0082 off; ltx2 1.308x/0.0013 vs
1.307x/0.0025; a14b 2711 vs 2717 ms/step. Cache-hook compile arming
verified to generalize (wan5b fb@0.04 armed 1.216x vs raw 1.048x). Dual-GPU
CFG stays HunyuanVideo-1.5-only: LTX-2 runs batch-CFG in one forward and
the Wan pipelines consume each branch inline with no guider combine hook.
- video_speedmem_bench gains epc_off (compile-parity isolation) and
fbcache_explicit / magcache_explicit configs plus expert-aware cache
application mirroring the loader.
Measured via scripts/video_speedmem_bench.py and the round-3 single-load
probes; full data and per-family decision table in
outputs/video_families_optim_round3.md (workspace). Tests: 235 passing across
the five video inference suite files (9 new: per-expert curve resolution and
step scaling, uncalibrated-expert refusal, toggle expert threading, wan5b
magcache auto load/toggle, ltx2 deny auto+explicit, a14b TE auto-dense);
ruff clean.
Cuts the shipped default's LPIPS vs the bit-exact reference from 0.224 to 0.139
while going faster (24.9 s to 21.2 s at 720p/33f/30 steps, 22.7x vs reference),
and makes the remaining speed/accuracy trade a user knob.
- inductor precision parity: set emulate_precision_casts=True for the regional
compile (fused pointwise kernels kept fp32 intermediates where eager rounds to
bf16 between ops); full-clip LPIPS vs bit-exact 0.221 to 0.052 at zero speed
cost. Snapshot/restored with the other process-wide backend flags.
- cache x compile composition fix: diffusers cache hooks are
torch.compiler.disable'd, so every COMPUTED step ran eager (1.69 vs 1.09
s/step) under MagCache/FBCache in both enable orders. Re-point each 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 so the uncached path stays pristine). Balanced MagCache at 50
steps: 1.48x to 2.17x, identical skip counts, bit-identical uncached rerun
after enable/disable cycles.
- transformer_cache_quality knob (quality|balanced|fast; API + UI + bench)
mapping to (threshold, max_skip_steps, retention_ratio). Auto resolves to the
near-lossless quality preset (0.06, 2, 0.3; 1.63-1.64x at pairwise LPIPS
0.05-0.09) for the HunyuanVideo-1.5 families and to balanced (the pre-knob
values, byte-identical behaviour) everywhere else.
- TE auto-quant resolves dense for HunyuanVideo-1.5: TE fp8_dynamic alone moves
the clip to LPIPS 0.236 vs bit-exact for zero speed win (the quantised encoder
perturbs the conditioning and the trajectory amplifies it chaotically); VAE
fp8 stays in auto (0.053, at the compile floor). Explicit schemes honored.
- dual-GPU CFG branch parallelism (new diffusion_cfg_parallel.py): transformer
proxy + DiT replica on the most-free second CUDA device + worker thread,
branch-routed off the pipeline's own cache_context names. Auto engages only
where measured bit-identical (eager tier: max abs diff 0.0, 1.66x); the
compiled stack is explicit cfg_parallel=on (1.52x over the sequential
default; per-device compiled artifacts differ by 1 bf16 ulp/step, documented
in the resolved record). Fail-soft gates: family allowlist, guider CFG,
pipeline kind, dense DiT, no offload, free-VRAM check; single-GPU loads are
untouched and the memory plan stays single-device.
- video API: the transformer_cache literal now accepts auto/magcache (an
explicit magcache request was rejected at the pydantic layer); the mxfp8
family deny records the round-2 measurement (block-32 MX scaling fixes the
zero-row collapse, no black frames, but is latency-neutral at LPIPS 0.37:
fails both ship bars).
Measured on B200 via the production lever path (video_speedmem_bench.py, which
gained a --cache-quality lever and companion-quant isolation configs). Tests:
441 passing across the video inference suite (32 new for cfg-parallel, 20 for
presets/arming, 3 for the inductor flag, 2 for TE auto-dense); ruff clean.
HunyuanVideo-1.5 loads previously logged 'fbcache unavailable (Model class
HunyuanVideo15TransformerBlock not registered)': diffusers 0.39 ships FBCache
block metadata for HunyuanVideo 1.0 but not 1.5, although the 1.5 DiT is fully
cache-shaped (CacheMixin, homogeneous residual-additive dual-stream blocks,
cache_context per guidance branch). Register the missing metadata at engage
time (deferring to a native registration when a future diffusers ships one).
Measured on a B200 (720p t2v, 1280x720, 33 frames, seed 42), FBCache is fast
but not shippable for this family: 1.44x at 30 steps / 2.41x at 50 steps, at
LPIPS 0.43-0.54 vs the same uncached stack with a +5..8 luma drift (no skip
cap or error budget, so the trajectory derails into a different clip). MagCache
(same registry metadata, also dispatched via enable_cache) is bounded by
design and lands the win: 1.49x end-to-end at 50 steps at LPIPS 0.147 with the
same composition, 1.21x at LPIPS 0.071 on the 480p model. Ship magcache as the
per-family AUTO cache mode for hunyuanvideo-1.5 / -720p with 50-step
calibrated per-family mag_ratios (cond/uncond curves agree within 0.014,
30 vs 50-step calibration within 0.027 after interpolation); every other
family keeps fbcache, and explicit fbcache/magcache requests are honored.
The auto toggle re-engages magcache on a step-count change so the ratio curve
is re-interpolated over the actual schedule.
Two production bugs fixed along the way:
- diffusers' HookRegistry caches its child-registry list, so enabling a cache
AFTER any uncached generation (the auto off-to-on toggle) left the new block
hooks without a context ('No context is set' on the first cached forward).
Invalidate the stale cache after every enable_cache.
- An explicit int8 DiT request crashed under the padded-text trim: torchao's
int8 dynamic path returns a zero-token (M=0) input unprojected (t2v byt5 /
image streams -> cond-type add shape crash) and torch._int_mm requires
M > 16 (an empty negative prompt trims to ~6 tokens -> TokenRefiner crash).
Add per-family int8 excludes for the text-stream linears (context_embedder*,
image_embedder, add_q/k/v_proj, to_add_out, ff_context); they run at tens of
tokens vs the ~32k video stream, so the exclusion costs nothing measurable.
Bench: trim lever key (trim_off / eager_trim isolation configs), int8_cudnn +
shipped_nocache rows, --cache-threshold, warmup timing, per-config frame
persistence for offline LPIPS rescoring, and the loader's per-family auto
cache mode mirrored. Full 720p matrix recorded: reference 481.6s ->
trim+cudnn+compile 35.4s -> shipped default with TE/VAE quant + magcache
24.9s (19.4x, peak VRAM 89.4 -> 81.8 GB), int8 latency-neutral (dense auto
policy confirmed), compile 1.56x per step, trim 13.5x per step at production
shapes.
Validated end to end through the real VideoBackend: load resolves
transformer_cache=magcache with trim + compile + cudnn, generation
re-interpolates 50 -> 30 steps, auto-disengages below 20 steps, re-engages
after an uncached generation, unload restores globals. Hermetic tests cover
the registration, the child-cache invalidation, magcache engage/threshold/
no-curve/no-steps paths, auto-mode routing, toggle re-interpolation, and the
family int8 excludes.
POST /video/generate previously held the response open for the whole
generation (multi-minute for 720p), so in --secure mode the Cloudflare
quick tunnel's ~100s origin-response cap returned a 524 while the server
kept generating, and the frontend treated the run as failed.
Generation now follows the same return-at-once pattern as /video/load:
begin_generate validates synchronously (409 on no model or on a second
concurrent generate via a new busy sentinel) and runs the existing
generate + gallery-persist pipeline, with the route's exact error
mapping, on a daemon thread. GET /video/generate-progress gains optional
terminal fields: phase completed carries the saved gallery record, phase
failed a client-safe error; active only drops together with a terminal
phase. The cancel event is registered before the worker starts so
/video/generate/cancel keeps working across the whole job.
VideoGenerateResponse becomes an accepted acknowledgement (status
started, video kept as an always-null compat field). The video page
fires the POST, then drives completion off the progress poll it already
runs (completed prepends the clip, failed surfaces the error, the
cancelled sentinel stays toast-free). The API-key training-start guards
now also probe the video backend for an in-flight background clip, since
it is no longer visible as an in-flight HTTP request to the keep-warm
counter.
Route tests keep the fake backend for load/generate/status but inherit
the real job machinery, covering immediate accept, concurrent 409, the
terminal completed record, sanitized/ValueError/cancelled failures, and
cancel of a running job.
Speed=off contract: the companion (text encoder / VAE) suppression under
an explicit Speed=off only matched an UNSET request, but auto is
backend-owned like transformer_quant, so an explicit
text_encoder_quant/vae_quant=auto would still engage fp8/int8 and break
the bit-exact request. Match 'auto' as well in both the image and video
loaders (a concrete scheme still forces quant). Covered by new
explicit-auto suppression tests.
Benchmark accuracy:
- quant_speedmem_bench teacc: when quantize_text_encoders returns None
(scheme skipped) the encoder is still dense, so scoring it against the
dense reference falsely certified a scheme that never ran. Record it
NOT engaged instead of collecting accuracy metrics.
- quant_speedmem_bench e2e: report the actual engaged te/vae scheme,
falling back to dense (not the requested auto) when the caster stayed
bf16, so a no-op default is not mislabelled as an auto-quantised run.
- video_speedmem_bench: HunyuanVideo ignores callback_on_step_end, so
step_ts stayed empty and per_step_ms was published as 0.0 for every
row. Time the denoise via a scheduler.step wrapper for that path.
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.
- diffusion_cache: do not engage FBCache when the selected pipeline opens no cache_context.
A CacheMixin transformer is necessary but not sufficient -- Flux Kontext / img2img /
inpaint / controlnet reuse the CacheMixin FluxTransformer2DModel yet their __call__ never
opens a cache_context, so the First-Block-Cache hook raised 'No context is set' on the
first forward, crashing every default FLUX.1-Kontext edit (28 steps, above the FBCache
threshold). Detect it from the pipeline __call__ source, resolved off the instance so the
per-expert proxy view delegates to the real pipe.
- diffusion_attention: honor an explicit aiter backend on ROCm/AMD targets instead of
dropping it via the NVIDIA-only guard (aiter is the AMD ROCm kernel; it only works there).
- video: clear the CUDA cache on a failed load so a partially built pipeline's reserved VRAM
does not OOM the next load (mirrors the image backend), and re-check cancellation after the
export/mux so a clip cancelled during the blocking encode is discarded, not persisted.
- diffusion_auto_policy / diffusion_prequant: validate a request-supplied prequant path
override (present AND allowlisted) before budgeting the small prequant plan, so the loader
does not skip the dense shards and then rebuild dense after evicting the resident pipeline.
- diffusion_controlnet: family-gate a curated ControlNet addressed by its full repo id, not
only its short catalog id, so a cross-family repo id 400s up front instead of downloading
and loading through the wrong ControlNet class.
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.
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.
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.
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.
Three gates the image backend has were missing on the video path:
- kind/extension mismatch: a model_kind 'single_file' with a .gguf name (or 'gguf'
with a non-.gguf) passed the video preflight, so the route acquired VIDEO and
evicted the resident GPU owner before the wrong single-file loader failed in the
background. Reject the mismatch up front, mirroring the image loader.
- Windows-shaped missing local pick: the missing-path check only matched POSIX
prefixes (/ ~ ./ ../), so a missing Windows path (C:\ / C:/ or any backslash path)
was treated as a Hub repo and only failed after the GPU handoff. Use the image
loader's is_absolute()/backslash path-shaped check.
- speed=off auto-quant: a full-pipeline load with Speed=off but Precision=auto still
promoted the unset precision to auto-quant, so on a dense-capable GPU it engaged
torchao quantization and forced the speed back to default -- silently breaking the
user's bit-exact request. Suppress the auto promotion when speed_mode is off, as
the image loader does.
Regression tests for each.
Several image/video/training preflights ran before the route acquires the GPU or
frees resident models, but let a doomed local pick through and only failed deep in
the background load, after the user's chat/Images/Video model was already evicted.
- Local base_repo / base_model: _is_trusted_diffusion_repo accepts any existing
local path, but the base loads via from_pretrained (needs model_index.json). A
local dir that is not a diffusers pipeline passed the trust gate, evicted the
resident model, then failed. Add a shared _assert_local_base_is_pipeline check
and call it in the image, video, and training preflights.
- Dataset images: discover_image_caption_pairs only checked filenames, so a
corrupt or zero-byte upload passed the start-route preflight, freed the GPU, then
crashed the spawned trainer in PIL. Add an opt-in verify_images decode probe
(cheap PIL header check) that the start route enables; the trainers leave it off
since they decode every image anyway.
- Local single-file safetensors: the On-Device scanner advertises a bare
.safetensors directory (no model_index.json) as a text-to-image model, but the
picker starts it as a pipeline with no filename, so every click 400s. Reinterpret
such a pick as a single_file load of the sole checkpoint (resolve_local_single_file)
so the advertised model is actually loadable.
Regression tests for each: local non-pipeline base (image/video/training), the
verify_images decode gate, and resolve_local_single_file.
- _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.
* 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>
default_video_generation_params fell back to a hardcoded LTX 40/4.0 when no identifier matched.
A Wan model loaded from an opaque local path under an explicit family_override (none of
gguf_filename / repo_id / base_repo carrying a wan token) then ran 40/4.0 instead of Wan's 50/5.0.
Add a fallback param and pass the resolved family's own default_steps/default_guidance from both
call sites, so the fallback tracks the actual family.
Two follow-ups to the video-tab review fixes:
- the cached-hub arch fallback only probed the active HF cache, but the cached-gguf picker
scans the active, legacy, and default cache roots; a GGUF cached in a non-active root was
offered yet 400d on load. Probe all three roots.
- on a mount/refresh with a model already resident, status.loaded made the Reapply button render
but lastLoad (set only by our own loads) was null, so clicking it silently did nothing. Track a
session reapply descriptor and hide the button when it is absent rather than offer a dead control.
Three video-tab fixes:
- delete-cached refused a loaded video repo but not one a background load is still downloading
(status().loaded is False in that window); add VideoBackend.loading_repo_ids() mirroring the
image backend and check it in the route so deleting mid-download can no longer yank blobs.
- the cached-gguf picker tags a cached HUB GGUF by its general.architecture, but the loader's
arch fallback only read a LOCAL file, so an opaquely-named cached hub LTX GGUF the picker
offered 400d on load; read the arch from the cached blob (network-free) too.
- on a mount with a model already loaded (refresh / load from another client) steps and guidance
stuck at the pre-load default, so a base checkpoint silently generated a degraded clip; seed
them from the backend-authoritative status.defaults once per newly-loaded model.
Verified against diffusers 0.39 source + the HF configs/safetensors headers:
- The Wan VAE decodes in float32 (WanPipeline loads AutoencoderKLWan at torch.float32
while the pipe runs bf16); the loader cast every component to bf16, degrading every
clip. Add vae_force_fp32 (both Wan families) and pin pipe.vae back to fp32 after build.
- The Wan transformers ship FP32 on disk (safetensors headers are F32; A14B index =
57.15 GB per expert = 14.3B x 4, TI2V = 20.0 GB = 5B x 4), so bf16_components_gb held
the fp32 on-disk sums (114.3 / 20.0) instead of the documented bf16-resident sizes.
Halve to 57.2 (two A14B experts) and 10.0 (TI2V), so the plan no longer over-budgets
the DiTs ~2x and forces needless offload on an 80 GB GPU.
- TI2V-5B's VAE is 16x spatial (vae/config.json), so WanPipeline floors H/W to 16*2 = 32.
Snap TI2V to /32 (was /16) so the recorded size matches the generated clip (a 720
request was recorded but rendered at 704). A14B keeps /16 (Wan2.1 8x VAE).
- delete_cached_model guarded chat (llama.cpp/transformers) and Images but not the
Video backend, so a loaded video GGUF (which shares the On-Device GGUF delete UI)
could be removed from under a live pipeline. Add the mirror guard on
get_video_backend().status().
- The Video picker admits a local GGUF by its general.architecture, but the loader
detected the family only from path/name tokens, so a renamed ltxv file (e.g.
model.gguf) was offered yet failed validate_load_request with a 400. _detect_load_family
now falls back to reading the arch (its string is a family alias) when name detection
misses, so the loader accepts exactly what the picker offered; an unsupported video arch
(wan) still resolves to None and 400s as before.
apply_step_cache on the video load path omitted quant_active, so a quantized video
transformer (an engaged dense transformer_quant, or a GGUF checkpoint) that also
enabled First-Block-Cache without an explicit threshold used the dense bf16 threshold
(0.08) instead of the higher quantized threshold (0.12) the cache helper documents as
needed for quantized transformers to trigger. The advertised quant plus FBCache path
therefore cached far less than intended. Thread quant_active through exactly as the
image path (diffusion.py) does: an engaged transformer_quant or a GGUF transformer both
count as quant-active here.
After the upstream cascade generalised _apply_group_offload to stream every
DiT it finds on the pipe (transformer AND transformer_2), apply_memory_plan's
group tier already block-streams both A14B experts, and model/sequential
offload hook every top-level module. The explicit per-expert pass this branch
added is therefore redundant: on a real group-tier load it re-registers the
group-offload hooks transformer_2 already carries, which diffusers rejects with
a duplicate-hook ValueError. That was swallowed by the helper's try/except and
surfaced as a misleading "second expert stays resident" warning even though the
expert was in fact streamed. Remove the pass and document why MoE needs none.
Review follow-ups on the video inference backend:
- validate_load_request now rejects a -GGUF repo picked as a diffusers
pipeline (no gguf_filename) up front, instead of failing minutes later
in from_pretrained after the GPU owner was already evicted.
- New _detect_load_family helper shared by validate_load_request and
_run_load: when the repo id alone does not carry the family, fall back
to detecting it from the picked GGUF filename, so both paths agree.
- routes/video.py now threads base_repo into validate_load_request so an
untrusted companion repo is refused before the arbiter handoff.
- unload() now drains _generate_lock before _teardown_state so a
cancelled clip actually exits the denoise loop before the VRAM is
reported free.
- load_pipeline re-checks the load token after the generate-lock barrier
and raises if the load was superseded while waiting.
- Pre-commit global mutations (backend flags, gguf compile installs) are
registered per load token and rolled back in _run_load's error path
via _rollback_precommit_globals, so a failed load no longer leaks
process-wide state.
- fp32 memory estimates now apply a 2x dtype scale on non-CPU devices
for pipeline, single-file and companion sizes (bf16 tables assume
2 bytes/param); GGUF quant estimates stay unscaled.
Tests: GGUF-repo-as-pipeline rejection, _detect_load_family fallback and
override semantics; fake route backend accepts base_repo. 66 passed
across test_video_backend, test_video_routes, test_video_families,
test_video_gallery.
The load tail re-ran the already-filtered speed_optims tuple through
.items() as if it were still the raw applied dict from apply_speed_optims.
An empty tuple short-circuited to {} so CPU test runs passed, but on a real
GPU at least channels_last engages, the tuple is truthy, and every load
failed with 'tuple' object has no attribute 'items'. Store the filtered
tuple directly and add a regression test that forces one optimisation to
engage.
