SolitaryThinker
eb2cd4e1cd
docs: on-tray handoff for the VSA-H3 DMD2 overfit experiment
...
staged state, copy-paste launch steps (container entry, env, config fix,
torchrun), expected signals, parity criteria, and the pre-paid traps.
(--no-verify: known mypy worktree-dirname hook issue only)
2026-08-09 15:21:52 -07:00
SolitaryThinker
d81f1f166b
[train] VSA-H3 DMD2 overfit: per-role backends, VSA metadata wiring, wandb guard
...
- MiniMaxH3Model: allow per-role attention backends (TORCH_SDPA/FLASH_ATTN/
VIDEO_SPARSE_ATTN_H3); predict_noise routes attn_metadata_vsa for
attn_kind='vsa' (dense roles keep None).
- MiniMaxH3DMDModel.prepare_batch builds real VSA-H3 metadata (packed
text|cond|audio prefix segments via _h3_vsa_prefix_segments, sparsity from
training.vsa.sparsity) when the role runs VIDEO_SPARSE_ATTN_H3.
- VSA-H3 tile(): fresh buffer for grad-tracking forwards; shared-buffer
reuse trips autograd's in-place version check at student backward.
- build_tracker: degrade to DummyTracker when wandb is unimportable or has
no credentials (modular stack already logs per-step loss dicts via the
wandb tracker; no new callback needed).
- dmd2_vsa0_overfit.yaml: student VSA-H3 @ sparsity 0.0, teacher/critic
FLASH_ATTN, 3-step DMD schedule, wandb project h3-dmd2-vsa.
- preprocess_minimax_h3_overfit: --video/--prompt/--output-dir/--model-path
CLI; overfit_vsa0.sh: preprocess + 4-GPU torchrun with FASTVIDEO_FA4=1.
- CPU tests: VSA metadata build/routing, per-role backend override,
wandb guard + loss-dict logging, tile-under-grad regression.
2026-08-09 14:46:58 -07:00
SolitaryThinker
7f81df6acf
Merge remote-tracking branch 'origin/vsa-h3' into feat/h3-dmd2-vsa
2026-08-09 14:28:26 -07:00
SolitaryThinker
dc70667483
[train] MiniMax H3 DMD2 distillation via packed dual-modality adapter
...
Add MiniMaxH3DMDModel, a DMD-capable subclass of the H3 training plugin
that presents both modality streams (video + stereo audio) to DMD2Method
as one packed [1, N] latent tensor, keeping dmd2.py untouched:
- pack/unpack adapters over [1,T,24,H,W] video + [1,2,32,Ta] audio
- integer [0,1000] method timesteps map to H3's shared base noise
amount, shifted per modality (video 12.0 / audio 3.0) as in SFT
- add_noise / predict_noise / predict_x0 in the packed convention;
explicit timesteps rewrite batch.(audio_)timesteps for row plans
and backward forward-context coherence
Base wrapper: accept attn_kind='vsa' as dense under TORCH_SDPA (real
VSA-H3 metadata slots in once the VSA-H3 backend lands), and support
conditional=False teacher-CFG forwards via cfg_uncond={text: zero}
(H3 has no negative-prompt encoder at training time).
Also: H3 DMD2 example config + launch script, tiny-arch fixture, and
CPU contract tests covering one full single_train_step (both rollout
modes, both optimizers) plus the packed-adapter units.
Committed with --no-verify: the mypy pre-commit hook rejects this
worktree's dirname ('h3-dmd2 is not a valid Python package name'),
a known worktree-path issue unrelated to the change.
2026-08-09 13:34:46 -07:00
SolitaryThinker
7dece0471f
[misc] VSA-H3: /simplify pass — dedupe with the Wan VSA backend, drop dead state
...
- scatter_into_tile_buf(): one copy of the zero-padded tile-buffer reuse
invariant, shared by both VSA backends (was byte-identical twice)
- the H3 tile buffer is now builder-owned, so one buffer serves the whole
denoising loop instead of a ~1.4 GB alloc+memset per step
- compute_topk(): H3's inline clamp math now reuses the Wan helper whose
semantics are pinned by test_vsa_cur_topk_clamps_to_valid_block_range
- layer_idx_from_prefix() in abstract.py replaces two regex spellings
(vmoba raises, H3 defaults to -1, both unchanged behaviorally)
- token_tile_and_valid(): single encoding of the pad-validity contract for
the probe and the test oracle (was spelled three ways)
- vsa_mode: one knob, one validation point — FASTVIDEO_H3_VSA_MODE (zero
readers) removed; batch.extra path now validates instead of silently
mapping typos to compete
- dropped dead metadata fields (non_pad_index unread; tile_partition_indices
had one test-only reader already covered by the untile-roundtrip check)
- removed cuda.py's redundant backend pre-import; fixed a stale layer.py
shape comment
(--no-verify: mypy hook rejects the checkout dir name; yapf+ruff ran and
their fixes are included)
2026-08-09 12:58:59 -07:00
SolitaryThinker
281d07d978
[feat] VSA for MiniMax H3: packed mixed-modality sparse attention (FA4 CuTe sm10x), exempt/compete mask policies, gate-compress branch, per-request sparsity/schedule knobs, attention-mass probe
2026-08-09 12:39:48 -07:00
Shao Duan
541ef014ee
[perf] MiniMax-H3: rank-reduced AdaLN pruned model option (-39% params, -23 GiB VRAM) ( #1699 )
2026-08-09 12:31:57 -07:00
William Lin
ffc1a7a58b
[refactor] H3 pipeline cleanup: shared helpers, dead machinery, loop-invariant hoists ( #1698 )
2026-08-09 04:51:31 -07:00
William Lin
9028953625
[misc] yapf pass under CI's interpreter (3.12) + pin hook language_version ( #1702 )
2026-08-08 22:24:17 -07:00
KyleNeverGivesUp and Claude Opus 5
6eb95693a1
[misc]: re-run yapf on main so pre-commit passes again ( #1700 )
...
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com >
2026-08-07 16:17:52 -07:00
Junda Su
c3567eb468
[feat] add Minimax H3 sft pipeline ( #1688 )
2026-08-07 16:12:04 -07:00
William Lin
15568f27db
[perf]: H3 torch.compile + CUDA graphs (1.2-1.3x) with denoising step marking ( #1689 )
2026-08-06 16:23:33 -07:00
Kaiqin Kong
126a75ad63
[misc] Support partial Hugging Face model downloads ( #1684 )
2026-08-06 16:09:46 -07:00
Raghav K and SolitaryThinker
a2bfc7cdb2
[docs] DGX Spark (GB10) performance & tuning guide + reproduction examples ( #1631 )
...
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com >
2026-08-06 14:36:38 -07:00
William Lin
b963a24612
[bugfix]: hard-fail when ATTN_QAT_INFER is selected but the kernel is unusable ( #1690 )
2026-08-06 13:12:27 -07:00
William Lin
fb7be2fe2c
[ci]: add golden-gate lane — single-layer bitwise DiT fingerprints for all SSIM-covered families ( #1682 )
2026-08-05 12:47:03 -07:00
William Lin
ab00392664
[bugfix]: add GB200 to the inline SSIM device tables #1676 missed ( #1681 )
2026-08-05 10:33:46 -07:00
Shao Duan
9f1e7c19d2
[bugfix] Wan I2V: CLIP image conditioning silently dropped when passed as a tensor during training ( #1673 )
2026-08-05 01:35:09 -07:00
Kaiqin Kong
e8b0e4c61e
[feat] Add MiniMax H3 ( #1674 )
2026-08-04 13:54:39 -07:00
Haochen Jiang
9145ffdc46
[bugfix]: keep _resolved_attention_backend out of the positional config signature ( #1678 )
2026-08-03 17:09:15 -07:00
William Lin
c3d07c870b
[bugfix]: give GB200 its own SSIM reference folder instead of B200's ( #1676 )
2026-08-03 15:14:25 -07:00
William Lin
e8812bef0b
[docs]: batched docs cleanup (landing page, links, requirements, nav) ( #1644 )
2026-08-02 18:03:26 -07:00
William Lin
b9be2449dc
[refactor]: delete the dead global attention-backend override ( #1672 )
2026-08-02 18:00:42 -07:00
Adhvay Iyer and SolitaryThinker
bc7a804618
[bugfix]: harden FastVideo Studio UI reliability and accessibility ( #1659 )
...
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com >
2026-08-02 15:26:18 -07:00
William Lin
7b094c945b
[refactor]: resolve attention backend once per component at load time ( #1657 )
2026-08-02 15:07:42 -07:00
William Lin
eeb3e8a597
[bugfix]: left-align Gemma connector tokens per batch row ( #1664 )
2026-08-02 15:01:34 -07:00
Suhaan Khurana
05406c5d1b
[misc]: consolidate dataset download scripts under examples/datasets/ ( #1667 )
2026-08-02 14:21:24 -07:00
KyleNeverGivesUp and Claude Opus 5
99d04a7f98
[bugfix] keep loader-populated text encoder configs in the validation pipeline ( #1669 )
...
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com >
2026-08-02 14:10:16 -07:00
William Lin
1b2b2a0161
[bugfix]: copy text encoder outputs out of CUDAGraph static buffers ( #1650 )
2026-07-27 21:52:53 -07:00
William Lin
98d65835b5
[misc]: refresh stale sm_120-only validation notes in QAT recipes ( #1655 )
2026-07-27 19:35:29 -07:00
William Lin
422585d08f
[misc]: add LTX-2.3 fine-tuning example recipes ( #1651 )
2026-07-27 18:42:44 -07:00
ryanM154
e59a1ce16a
[bugfix] Report actual package version in fastvideo --version ( #1652 )
2026-07-27 16:32:01 -07:00
William Lin
d71acc0eb5
[bugfix]: fix stale imports in LTX-2.3 gradio local demo ( #1640 )
2026-07-27 16:31:07 -07:00
William Lin
af2934dd6b
[feat]: FA4-FP4 ATTN_QAT_INFER on sm_100/sm_103 + NVFP4 weight purge ( #1647 )
2026-07-27 15:46:52 -07:00
William Lin
1801512818
[docs]: cover all registered models in the support matrix ( #1641 )
2026-07-27 11:13:53 -07:00
William Lin
5ae05b032e
[misc]: add LTX-2 fine-tuning example recipes ( #1645 )
2026-07-27 11:11:49 -07:00
Mac Lee
7a592ff09a
[ci]: cache FastVideo kernel builds in Modal ( #1562 )
2026-07-26 04:21:09 -07:00
Lev Novitskiy and Claude Sonnet 5
8b23984c79
[feat] Add Kandinsky5 QAD training pipeline: data preprocessing, QAT finetune, QAT-aware DMD distillation ( #1601 )
...
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com >
2026-07-26 02:59:21 -07:00
William Lin
bf18371afe
[feat] Enable LTX-2 NVFP4 linear and attention QAT fine-tuning ( #1626 )
2026-07-26 02:19:34 -07:00
William Lin
69349dd2aa
[docs]: unify community links on the README Slack invite ( #1643 )
2026-07-25 17:26:58 -07:00
Yogya Mehrotra and Claude Sonnet 5
8d89f30d3f
[bugfix] Skip CUDA-only fastvideo-kernel/flashinfer-python deps on non-Linux ( #1574 )
...
Co-authored-by: Claude Sonnet 5 <noreply@anthropic.com >
2026-07-23 14:23:45 -07:00
William Lin
10546353da
[ci]: extend Full Suite training lane timeouts ( #1616 )
2026-07-23 13:00:00 -07:00
pkisfaludi-nv and Claude Opus 4.8
9fb74b9732
Make LTX-2 RMSNorm out-of-place so torch_tensorrt + Ulysses SP compiles ( #1623 )
...
Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com >
2026-07-22 04:32:04 -07:00
Adriel Fung and SolitaryThinker
521dee0e82
[perf]: enable per-block torch.compile for LTX2 with persistent cache ( #1602 )
...
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com >
2026-07-20 15:59:13 -07:00
William Lin
229419208e
[ci]: opt in to fork-PR head checkout after actions/checkout guard change ( #1625 )
2026-07-20 13:12:22 -07:00
William Lin
65f3b946b9
[feat] Add LTX-2 and LTX-2.3 fine-tuning to the modular trainer ( #1624 )
2026-07-20 11:58:39 -07:00
Zhang Peiyuan
191fcbf46c
[feat] add qat docs ( #1621 )
2026-07-19 18:37:47 -07:00
Junda Su
755a4e4470
[new-model] Add LingBot-Video Dense and MoE/refiner T2V inference ( #1595 )
2026-07-18 20:18:16 -07:00
9709b7513b
[feat] Port NVFP4 QAT/QAD to modular train framework ( #1619 )
...
Co-authored-by: Peiyuan Zhang <email>
Co-authored-by: Peiyuan <a>
2026-07-18 15:01:09 -07:00
William Lin
32cd603515
Revert docs trusted-branch-only workflow ( #1618 )
2026-07-17 00:22:56 -07:00
SolitaryThinker