[feat] Upstream Attn-QAT Video Diffusion Code by RandNMR73 · Pull Request #1225 · hao-ai-lab/FastVideo

RandNMR73 · 2026-04-09T04:14:40Z

Summary

Add Attn-QAT training kernels
Add flashinfer NVFP4 linear layers (currently hardcoded for Wan-2.1 arch)
Add Modified SageAttention3 kernels in the fastvideo-kernel package
Add Attn-QAT video model training scripts

@SolitaryThinker

Per @SolitaryThinker review: the .codex empty marker file was extracted from PR #1225 (sync-branch) but is not wanted on origin/main. Dropping it from Slice 1 of the Attn-QAT decomposition. Attn-QAT-Stack: 1.1/12

@RandNMR73

Adds NVFP4QATConfig — the Attn-QAT flavor of FP4 quantization. Distinct from main's NVFP4Config (NVIDIA Blackwell hardware FP4 inference); this config is for the QAT training side of the Attn-QAT stack. Pure deadcode: no existing code path imports NVFP4QATConfig yet. The `nvfp4_qat` method literal in QuantizationMethods makes it selectable, but the caller lands in slice 3+. Renamed from PR-1225's original `Fp4Config` / `fp4_config.py` / `"fp4"` literal to avoid collision with PR #1334's NVFP4 plumbing already merged to main. Extracted from PR #1225 (#1225) by @RandNMR73. Source SHA: 3f818d0 Attn-QAT-Stack: 2/12 Co-Authored-By: jzhang38 <42993249+jzhang38@users.noreply.github.com> Co-Authored-By: RandNMR73 <99706358+RandNMR73@users.noreply.github.com>

@RandNMR73

Adds the FastVideo-native FP4 linear forward helper module extracted from PR-1225. The source module does not define a public Fp4Linear class, so this slice keeps the original fp4linear.py filename and helper names. Pure deadcode: no existing code path imports this helper yet. Activation lands in Slice-12 of the decomposition. Extracted from PR #1225 (#1225) by @RandNMR73. Source SHA: 3f818d0. Attn-QAT-Stack: 3/12 Co-Authored-By: jzhang38 <42993249+jzhang38@users.noreply.github.com> Co-Authored-By: RandNMR73 <99706358+RandNMR73@users.noreply.github.com>

…/12) Add fastvideo/attention/backends/attn_qat_{infer,train}.py extracted from PR hao-ai-lab#1225. Backends not yet registered in the selector — full deadcode until activation (slice 12). Attn-QAT-Stack: 4/12 Co-Authored-By: jzhang38 <42993249+jzhang38@users.noreply.github.com> Co-Authored-By: RandNMR73 <99706358+RandNMR73@users.noreply.github.com>

… 5/12) Slice 5/12 of the PR hao-ai-lab#1225 decomposition (Attn-QAT stack). Tier-2: backward-compatible additions to the shared attention infra that several backends consume. What this adds (per file): * `fastvideo/attention/backends/abstract.py` (+6/-2) - Moves `VSA_sparsity: float = 0.0` (kw-only) up to the base `AttentionMetadata` dataclass so non-VSA backends can satisfy a uniform metadata interface without re-declaring the field. - Adds `AttentionMetadata.__getattr__` raising `AttributeError` so callers that typo a field get a stable, matchable error message instead of relying on default object behavior. - Loosens `AttentionMetadataBuilder.build` signature from `**kwargs: dict[str, Any]` to `**kwargs: Any` to match how subclass builders are typed throughout the codebase. * `fastvideo/attention/backends/bsa_attn.py` (+3/-5) - Collapses the duplicated double-fallback flash-attn import block to a single import from `fastvideo.attention.utils.flash_attn_no_pad` (which now centralizes the fallback chain — see below). When the centralized impl isn't available, sets the symbol to None and `FLASH_ATTN_AVAILABLE=False`. * `fastvideo/attention/backends/video_sparse_attn.py` (+2/-3) - Removes the redundant `VSA_sparsity: float` field from the `VideoSparseAttentionMetadata` subclass (it is now inherited from the base with a default of 0.0). VSA continues to pass `VSA_sparsity=...` as a kwarg into the metadata constructor. - Adds `-> None` return annotations to `__init__` / `prepare` on the builder. * `fastvideo/attention/backends/vmoba.py` (+8/-4) - Adds `-> None` return annotations to `__init__` / `prepare`. - Tightens the `device` parameter to `torch.device | None = None` (was an unannotated default-`None`). - Narrows the `attn_metadata` parameter on `VMOBAAttentionImpl.forward` from the base `AttentionMetadata` to the concrete `VideoMobaAttentionMetadata`. - Adds an explicit `assert self.layer_idx is not None` guard and an `else: raise ValueError` for the moba_layer chunk-selection switch so a misconfigured config can't silently use an uninitialized `moba_chunk_size`. - Pre-declares `moba_chunk_size` with its union type for mypy. * `fastvideo/attention/layer.py` (+2) - Adds an `attention_mask: torch.Tensor | None = None` parameter to `DistributedAttention_VSA.forward` (downstream backends in the stack consume it). * `fastvideo/attention/utils/flash_attn_no_pad.py` (+47/-33) - Centralizes the varlen-flash-attn fallback chain into a single `_resolve_flash_attn_varlen_func()` helper. The fallback order is: `fastvideo.attention.utils.flash_attn_cute` → `flash_attn_interface` → `flash_attn`. Module-level `flash_attn_varlen_func_impl` is now this helper's output, so consuming backends (bsa, vsa) import one stable symbol. - Adds type annotations to every public function and to the helper. Files in PR hao-ai-lab#1225 considered but NOT applied: * `fastvideo/attention/backends/sage_attn3.py` — the source-SHA diff shrinks `get_supported_head_sizes()` from `[64, 128, 256]` to `[64, 128]`. That removal is unrelated to QAT-compat (head_size=256 has been supported since the original SAGE3 backend landed in hao-ai-lab#815) and is treated as an obsolete edit from the source branch's history. Not applied; current main's behavior is preserved. Test coverage added (`fastvideo/tests/attention/`): * `test_attention_metadata_base.py` — 8 CPU-only smoke tests for the new base-class semantics: `VSA_sparsity` default 0.0, kw-only-ness, subclass inheritance and override, `__getattr__` `AttributeError` message contract, and `asdict_zerocopy` field handling. * `test_flash_attn_no_pad_resolver.py` — 3 tests for the fallback chain (cute → interface → flash_attn). Skips at module level when `flash_attn` is not installed (the file has an unconditional top-level import that pre-exists this slice). Pre-commit gate (yapf + ruff + codespell + mypy) passes on all changed files. Local pytest of the new tests: 8 passed, 1 skipped (resolver suite, no flash_attn installed in this env). Source: extracted from PR hao-ai-lab#1225 (hao-ai-lab#1225, SHA 3f818d0) and 3-way-merged onto current main (slice 4 / PR hao-ai-lab#1358 merged as fda0203). The `bsa_attn.py` and `video_sparse_attn.py` files required 3-way merge because main has diverged from the source SHA's merge-base; both merged cleanly with no manual conflict resolution. Co-Authored-By: jzhang38 <42993249+jzhang38@users.noreply.github.com> Co-Authored-By: RandNMR73 <99706358+RandNMR73@users.noreply.github.com> Attn-QAT-Stack: 5/12

Slice 6 of 12 in the PR hao-ai-lab#1225 decomposition. Tier 2 — backward- compat additions, gated paths only. No activation in this slice. What this adds -------------- * fastvideo/layers/linear.py (+54): adds opt-in shape-tracking instrumentation to ``ReplicatedLinear`` so upcoming QAT-aware backends can discover which GEMM shapes need quantized kernels. Gated by the class attr ``enable_shape_tracking = False``; the default forward path is bit-identical to pre-slice behavior. Adds ``get_shape_mapping``, ``reset_shape_tracking``, ``_track_shape``, ``print_shape_summary``. No new constructor params. * fastvideo/layers/mlp.py (+22): adds an optional ``quant_config: QuantizationConfig | None = None`` kwarg to ``MLP.__init__`` and threads it (plus an explicit ``prefix``) into the two underlying ``ReplicatedLinear`` instances. When ``quant_config is not None``, runs ``process_weights_after_loading`` on each sub-layer's resolved quant method. When ``quant_config is None`` (default), behavior is unchanged: ``ReplicatedLinear`` falls back to ``UnquantizedLinearMethod`` exactly as before. * fastvideo/models/dits/wanvideo.py (+67): wires ``quant_config`` through ``WanSelfAttention``, ``WanI2VCrossAttention``, ``WanTransformerBlock``, ``WanTransformerBlock_VSA``, and ``WanTransformer3DModel`` constructors so a future ``NVFP4QAT``- configured Wan2.1 build can quantize its attention QKV/out projections and FFN. Reads ``config.quant_config`` from ``WanVideoConfig`` (the field is already present on the shared ``DiTBaseConfig``). All new kwargs default to ``None``; default Wan2.1 path stays bit-identical. Files in PR hao-ai-lab#1225 considered but NOT applied -------------------------------------------- The source-SHA ``fastvideo/layers/linear.py`` also contains several edits that pre-date current ``main`` and would silently regress it: * Removal of the ``NVFP4Config``-only-quantizes-a-curated-subset explanatory comments in ``LinearBase.__init__`` and ``ReplicatedLinear.__init__`` (added on main as part of slice 3 / PR hao-ai-lab#1336). * Removal of the ``if self.quant_method is None: self.quant_method = UnquantizedLinearMethod()`` fallback inside ``LinearBase.__init__`` (also part of the slice 3 hardening). * A constructor / ``create_weights`` reformat from multi-line to compact one-line style — pure style noise. * ``assert self.quant_method is not None`` → ``if self.quant_method is None: self.quant_method = UnquantizedLinearMethod()`` in ``ColumnParallelLinear.__init__/forward`` and ``RowParallelLinear.__init__/forward``. ``LinearBase.__init__`` on current ``main`` already guarantees ``quant_method`` is non-None, so the source PR's defensive checks would be no-ops; they pre-date the slice 3 base-class hardening. * The same ``if quant_method is None`` defensive insert in ``ReplicatedLinear.forward`` — also a no-op against current ``main`` for the same reason. None of the skipped edits affect the FP4 path; current ``main``'s behavior on those lines is strictly stronger than the source SHA's. This mirrors slice 5's intentional skip of the ``sage_attn3.py`` head_size removal (see PR hao-ai-lab#1383). Also dropped: an unused ``from contextlib import nullcontext`` import that the source PR staged in ``wanvideo.py`` for a deeper-stack slice (ruff would reject it as unused). Stacking -------- Base: ``main`` (slice 5 / PR hao-ai-lab#1383 merged at ``ba75ad82dbe4a7069412494c051c1c69155fdc9d``). No stack dependency — this is a clean linear PR off ``main``. Provenance ---------- Files extracted from PR hao-ai-lab#1225 (hao-ai-lab#1225) at source SHA ``3f818d0fc532ec6494b465967d5f485150917d0c`` and audited against current ``main``. ``mlp.py`` and ``wanvideo.py`` (modulo the dropped unused import) were applied directly — ``main`` had not diverged from the source's merge-base for those files. ``linear.py`` was hand-merged to preserve current ``main``'s slice-3 hardening (see the ``NOT applied`` list above); only the additive shape-tracking surface was carried over. Pre-commit gate (yapf + ruff + codespell + mypy) passes on all three changed files. Test plan --------- No new tests this slice. The shape-tracking surface is opt-in instrumentation (default disabled) and the ``quant_config`` plumbing is dormant until a future slice sets ``config.quant_config`` to a non-None value. The activation slice (12/12) will carry the contract test for the full FP4 Wan-2.1 path. Sequence -------- Attn-QAT-Stack: 6/12. Earlier merged slices: 4/12 (PR hao-ai-lab#1358), 5/12 (PR hao-ai-lab#1383). Out of scope for this slice: the actual FP4 activation switch, weight-loading conversion, and any cross-cutting config registration (later slices). Co-Authored-By: Peiyuan Zhang <a1286225768@gmail.com> Co-Authored-By: Matthew Noto <notomatthew31@gmail.com>

mergify · 2026-05-29T23:46:13Z

This PR has merge conflicts with the base branch. Please rebase:

git fetch origin main
git rebase origin/main
# Resolve any conflicts, then:
git push --force-with-lease

Slice 6 of 12 in the PR hao-ai-lab#1225 decomposition. Tier 2 — backward- compat additions, gated paths only. No activation in this slice. What this adds -------------- * fastvideo/layers/linear.py (+54): adds opt-in shape-tracking instrumentation to ``ReplicatedLinear`` so upcoming QAT-aware backends can discover which GEMM shapes need quantized kernels. Gated by the class attr ``enable_shape_tracking = False``; the default forward path is bit-identical to pre-slice behavior. Adds ``get_shape_mapping``, ``reset_shape_tracking``, ``_track_shape``, ``print_shape_summary``. No new constructor params. * fastvideo/layers/mlp.py (+22): adds an optional ``quant_config: QuantizationConfig | None = None`` kwarg to ``MLP.__init__`` and threads it (plus an explicit ``prefix``) into the two underlying ``ReplicatedLinear`` instances. When ``quant_config is not None``, runs ``process_weights_after_loading`` on each sub-layer's resolved quant method. When ``quant_config is None`` (default), behavior is unchanged: ``ReplicatedLinear`` falls back to ``UnquantizedLinearMethod`` exactly as before. * fastvideo/models/dits/wanvideo.py (+67): wires ``quant_config`` through ``WanSelfAttention``, ``WanI2VCrossAttention``, ``WanTransformerBlock``, ``WanTransformerBlock_VSA``, and ``WanTransformer3DModel`` constructors so a future ``NVFP4QAT``- configured Wan2.1 build can quantize its attention QKV/out projections and FFN. Reads ``config.quant_config`` from ``WanVideoConfig`` (the field is already present on the shared ``DiTBaseConfig``). All new kwargs default to ``None``; default Wan2.1 path stays bit-identical. Files in PR hao-ai-lab#1225 considered but NOT applied -------------------------------------------- The source-SHA ``fastvideo/layers/linear.py`` also contains several edits that pre-date current ``main`` and would silently regress it: * Removal of the ``NVFP4Config``-only-quantizes-a-curated-subset explanatory comments in ``LinearBase.__init__`` and ``ReplicatedLinear.__init__`` (added on main as part of slice 3 / PR hao-ai-lab#1336). * Removal of the ``if self.quant_method is None: self.quant_method = UnquantizedLinearMethod()`` fallback inside ``LinearBase.__init__`` (also part of the slice 3 hardening). * A constructor / ``create_weights`` reformat from multi-line to compact one-line style — pure style noise. * ``assert self.quant_method is not None`` → ``if self.quant_method is None: self.quant_method = UnquantizedLinearMethod()`` in ``ColumnParallelLinear.__init__/forward`` and ``RowParallelLinear.__init__/forward``. ``LinearBase.__init__`` on current ``main`` already guarantees ``quant_method`` is non-None, so the source PR's defensive checks would be no-ops; they pre-date the slice 3 base-class hardening. * The same ``if quant_method is None`` defensive insert in ``ReplicatedLinear.forward`` — also a no-op against current ``main`` for the same reason. None of the skipped edits affect the FP4 path; current ``main``'s behavior on those lines is strictly stronger than the source SHA's. This mirrors slice 5's intentional skip of the ``sage_attn3.py`` head_size removal (see PR hao-ai-lab#1383). Also dropped: an unused ``from contextlib import nullcontext`` import that the source PR staged in ``wanvideo.py`` for a deeper-stack slice (ruff would reject it as unused). Stacking -------- Base: ``main`` (slice 5 / PR hao-ai-lab#1383 merged at ``ba75ad82dbe4a7069412494c051c1c69155fdc9d``). No stack dependency — this is a clean linear PR off ``main``. Provenance ---------- Files extracted from PR hao-ai-lab#1225 (hao-ai-lab#1225) at source SHA ``3f818d0fc532ec6494b465967d5f485150917d0c`` and audited against current ``main``. ``mlp.py`` and ``wanvideo.py`` (modulo the dropped unused import) were applied directly — ``main`` had not diverged from the source's merge-base for those files. ``linear.py`` was hand-merged to preserve current ``main``'s slice-3 hardening (see the ``NOT applied`` list above); only the additive shape-tracking surface was carried over. Pre-commit gate (yapf + ruff + codespell + mypy) passes on all three changed files. Test plan --------- No new tests this slice. The shape-tracking surface is opt-in instrumentation (default disabled) and the ``quant_config`` plumbing is dormant until a future slice sets ``config.quant_config`` to a non-None value. The activation slice (12/12) will carry the contract test for the full FP4 Wan-2.1 path. Sequence -------- Attn-QAT-Stack: 6/12. Earlier merged slices: 4/12 (PR hao-ai-lab#1358), 5/12 (PR hao-ai-lab#1383). Out of scope for this slice: the actual FP4 activation switch, weight-loading conversion, and any cross-cutting config registration (later slices). Co-Authored-By: Peiyuan Zhang <a1286225768@gmail.com> Co-Authored-By: Matthew Noto <notomatthew31@gmail.com>

Add the attn_qat_infer Blackwell FP4 attention + quantization CUDA kernels (modified SageAttention3) into fastvideo-kernel, gated behind a build flag and landed as deadcode (not yet wired into a backend). - fastvideo-kernel/attn_qat_infer/: Blackwell FP4 attention (blackwell/api.cu) and 4D FP4 quantization (quantization/fp4_quantization_4d.cu) CUDA kernels, plus their Python wrappers (api.py) and microbenchmarks. - CMake: new FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER option (AUTO/ON/OFF). AUTO only builds on CUDA Toolkit 12.8+ with Blackwell sm_120a; otherwise the kernels are skipped, so the change is inert on existing CI/GPUs. - pyproject/MANIFEST: ship the attn_qat_infer module in the wheel. This is the modified-SageAttention3-kernel item from the #1225 tracker. The attn_qat_infer / attn_qat_train attention backends already on main import these modules lazily, so they stay dormant until a follow-up PR wires the backend in. Part of #1225.

Add the attn_qat_infer Blackwell FP4 attention + quantization CUDA kernels (modified SageAttention3) into fastvideo-kernel, gated behind a build flag and landed as deadcode (not yet wired into a backend). - fastvideo-kernel/attn_qat_infer/: Blackwell FP4 attention (blackwell/api.cu) and 4D FP4 quantization (quantization/fp4_quantization_4d.cu) CUDA kernels, plus their Python wrappers (api.py) and microbenchmarks. - CMake: new FASTVIDEO_KERNEL_BUILD_ATTN_QAT_INFER option (AUTO/ON/OFF). AUTO only builds on CUDA Toolkit 12.8+ with Blackwell sm_120a; otherwise the kernels are skipped, so the change is inert on existing CI/GPUs. - pyproject/MANIFEST: ship the attn_qat_infer module in the wheel. This is the modified-SageAttention3-kernel item from the #1225 tracker. The attn_qat_infer / attn_qat_train attention backends already on main import these modules lazily, so they stay dormant until a follow-up PR wires the backend in. Part of #1225. Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com> Co-authored-by: Loay Rashid <42599591+loaydatrain@users.noreply.github.com> Co-authored-by: Kaiqin Kong <k1kong@ucsd.edu>

…ipe finale (12/12) Completes the QAD training recipe: quantization-aware DMD distillation of Wan2.1-T2V-1.3B down to 3 sampling steps, with generator-only Attn-QAT. - component_loader.py: generator-only QAT for DMD distillation. The teacher (real_score) and critic (fake_score) transformers load with the _loading_teacher_critic_model flag; mask the nvfp4_qat quant and the global ATTN_QAT_TRAIN attention env for them so only the generator runs fake-quant attention. Config-driven, no monkey-patching, reuses the existing flag. - distill_dmd_qat.sh: stage-2 DMD distillation script (3-step, generator init from the stage-1 finetune checkpoint). - README: the full two-stage recipe (QAT finetune -> QAT DMD distill to 3 steps). Verified end-to-end on Blackwell (GB200/sm_100): the generator loads with ATTN_QAT_TRAIN while teacher/critic load full precision; the DMD double loop runs (generator updates every generator_update_interval, critic every step, healthy loss), 3-step validation generates videos, checkpoint saved. Part of #1225. Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com> Co-authored-by: Loay Rashid <42599591+loaydatrain@users.noreply.github.com> Co-authored-by: Kaiqin Kong <k1kong@ucsd.edu>

Add FP8 (e4m3) quantization-aware training for the DiT linear layers, mirroring the FP4 linear STE path (#1463). Small follow-on for FP8-class GPUs. - layers/fp8linear.py: _LinearFWD8BWD16Fn — FP8 forward (torch._scaled_mm on sm89+, bf16 fake-quant fallback on older GPUs) + full-precision backward (STE). Absmax tensor/row-wise scaling, FP8_MAX=448. From the fork's FP8 training work (commit 96011d29), adapted to the config-driven path (no monkey-patch flags). - layers/quantization/fp8_qat_train_config.py: a training quant method bridging the STE into quant_config, registered "fp8_qat_train" (Wan to_q/k/v/out + ffn). - register it in layers/quantization/__init__.py. No flashinfer needed; runs on any sm89+ GPU (and older via the bf16 fallback), not just Blackwell. Enable with --transformer-quant fp8_qat_train (the training CLI arg + string->config resolve added in #1463). Verified on Blackwell (GB200/sm_100): unit test = FP8 forward + nonzero weight grad (STE); a finetune smoke with --transformer-quant fp8_qat_train trains 10 steps with healthy, decreasing loss / grad. Part of #1225. Co-authored-by: William Lin <SolitaryThinker@users.noreply.github.com> Co-authored-by: Loay Rashid <42599591+loaydatrain@users.noreply.github.com> Co-authored-by: Kaiqin Kong <k1kong@ucsd.edu>

SolitaryThinker · 2026-07-15T23:06:11Z

closing, as this was upstreamed in other PRs

jzhang38 and others added 30 commits December 20, 2025 11:09

+baseline

da3f43c

+ fp4 linear + fp4 attn, all with 16-bit bwd

85553c2

+ generator sage 3

6a58c3a

update

9161ef6

update

8abfe23

stash

a1ab4a7

update

2a28c08

update

032016f

update

c47be5c

update

fa91001

save

130b466

1005 morning

84ed3ce

update

18aebfa

add real and fake quant precision tests

c137c9f

fake quant done

3a9e4d3

save

b628b5f

checkpoint

ebdb160

nvfp4 utils in progress

9b12552

add inference repo

2fa6fc1

fix DeepGEMM path

fe9e94e

fix DeepGEMM path

6a7f08e

qat attn in progress + refactor nvfp4 utils

b06c038

fix import

d41912c

checkpoint (qat attn in progress)

9bcc411

fix masking + causal and non-causal logic in qat attn

3b70702

5090 testing

7afe836

add SageAttn3 with QAT

8d68a4f

print sageattn file

3641487

adjust sage3 block size to 64x64

460df09

adjus quant kernel block size

52480b2

SolitaryThinker mentioned this pull request May 13, 2026

[feat]: Add NVFP4QAT quantization config (Attn-QAT 2/12) #1348

Merged

SolitaryThinker mentioned this pull request May 14, 2026

[feat]: Add NVFP4QAT linear layer (Attn-QAT 3/12) #1350

Merged

SolitaryThinker mentioned this pull request May 15, 2026

[feat]: Attn-QAT inference + training backends (deadcode) (Attn-QAT 4/12) #1356

Closed

This was referenced May 16, 2026

[feat]: Attn-QAT inference + training backends (deadcode) (Attn-QAT 4/12) #1358

Merged

[refactor]: shared attention infra additions for QAT-compat (Attn-QAT 5/12) #1383

Merged

SolitaryThinker mentioned this pull request May 23, 2026

[refactor]: linear/mlp FP4 path additions for Wan-2.1 (Attn-QAT 6/12) #1390

Merged

mergify Bot added the needs-rebase PR has merge conflicts label May 29, 2026

alexzms mentioned this pull request Jun 12, 2026

[kernel] QAD 5090: Add modified SageAttention3 FP4 inference kernels (7/12) #1455

Merged

alexzms mentioned this pull request Jun 15, 2026

[feat] QAD 5090: QAT training recipe — finetune + DMD distillation (12/12) #1462

Merged

This was referenced Jun 16, 2026

[feat] QAD 5090: FP4 QAT linear STE for training (13/12) #1463

Merged

[feat] QAD 5090: FP8 QAT linear training (14/12) #1464

Merged

loaydatrain mentioned this pull request Jun 16, 2026

[Bugfix] QAD 5090: Torch.compile and other optimizations (15/12) #1466

Merged

SolitaryThinker closed this Jul 15, 2026

Navigation Menu

Uh oh!

[feat] Upstream Attn-QAT Video Diffusion Code - #1225

[feat] Upstream Attn-QAT Video Diffusion Code#1225
RandNMR73 wants to merge 203 commits into
mainfrom
sync-branch

RandNMR73 commented Apr 9, 2026

Uh oh!

mergify Bot commented May 29, 2026

Uh oh!

SolitaryThinker commented Jul 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

15 participants

Uh oh!

Conversation

RandNMR73 commented Apr 9, 2026

Summary

Uh oh!

mergify Bot commented May 29, 2026

Uh oh!

SolitaryThinker commented Jul 15, 2026

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

15 participants