This website requires JavaScript.
fca45bc8e1
[perf] Quantize frames to uint8 on-device before the post-decode D->H copy (#1362 )
Raghav K
2026-08-20 19:55:55 -07:00
aa95a4c18e
[misc]: add CPU regression tests for the on-device uint8 frame path
Will Lin and Claude Fable 5
2026-08-21 02:39:09 +00:00
942f7db3db
[test] H3 VAE streaming: legacy-decode oracle, batch slicing, pinned-buffer coverage
Will Lin and Claude Fable 5
2026-08-21 02:38:44 +00:00
aadb23f409
[perf] H3 VAE streamed decode: direct per-plane copies, async with pinned buffers
Will Lin and Claude Fable 5
2026-08-21 02:38:35 +00:00
f56f567042
[misc]: correct the post-decode rationale comments
Will Lin and Claude Fable 5
2026-08-21 02:34:43 +00:00
15a164a052
[bugfix]: derive result size from the decoded output when the samples mirror is skipped
Will Lin and Claude Fable 5
2026-08-21 02:34:16 +00:00
a02634fc72
[bugfix]: H3 encoder truncation review fixes: parity gate, override validation, key-filter bound
Will Lin and Claude Fable 5
2026-08-21 02:32:48 +00:00
aaaa7a14a3
Merge remote-tracking branch 'origin/main' into pr1362-fixes
Will Lin
2026-08-21 02:21:44 +00:00
86d639c848
[docs] Announce FastMetal-QAD (#1721 )
Aryan Kumar
2026-08-19 13:45:49 -07:00
3ab4a7f980
[ci] add CI runner
SolitaryThinker
2026-08-19 12:27:50 -07:00
00338aa9ca
[perf] Add FA4 CuTe backward support for VSA-256 (#1639 )
2026-08-19 11:46:49 -07:00
74b409d7cf
[test]: add H3 VAE parity and memory benchmark
H1yori233
2026-08-19 02:45:23 -07:00
23c285bfb2
[perf]: reduce DMD2 rollout forwards safely under FSDP
root
2026-08-12 06:00:01 +00:00
ceec06dc31
[feat]: integrate torch profiler with modular training
root
2026-08-12 04:05:21 +00:00
8537dcd6de
[feat]: Apple Silicon MLX runtime — INT8 Wan2.1 and Wan2.2 inference (#1638 )
Aryan Kumar and Aryan Kumar
2026-08-18 15:08:00 -07:00
a24b8fc02a
[perf] Add fine-grained FA4 VSA kernels
Will Lin
2026-08-17 17:39:20 +00:00
08e18d333a
[misc] correct fp32 checkpoint size to measured 741 GiB (was ~528 estimate)
Will Lin and Claude Fable 5
2026-08-15 07:37:53 +00:00
1ec849a4e3
[perf]: build the MiniMax H3 text encoder only as far as it is read
Kyle and Claude Opus 5
2026-08-15 00:21:46 -07:00
0246a62726
[bugfix] resume latest: skip incomplete checkpoints (require dcp/.metadata)
Will Lin and Claude Fable 5
2026-08-15 06:16:46 +00:00
80adb0fe1b
[misc] lessons: preflight ruleset distilled from H3 DMD2 v1-v5 launches
Will Lin and Claude Fable 5
2026-08-14 11:20:01 +00:00
deaf60edf1
[train] dcp_to_diffusers: --link-base space-lean export + stale shard-index cleanup
Will Lin and Claude Fable 5
2026-08-14 06:09:50 +00:00
dd9276d1c8
[train] H3 v5 config: keep-4 checkpoint retention (fp32 saves are ~528 GiB)
Will Lin and Claude Fable 5
2026-08-14 06:07:13 +00:00
b45233b86b
[train] H3 v5 config: requeue-safe resume_from_checkpoint=latest
Will Lin and Claude Fable 5
2026-08-14 02:11:48 +00:00
1aeb90a019
[train] DMD2: sigma-space score sampling + shift-aware H3 schedule (v5 config)
Will Lin and Claude Fable 5
2026-08-14 02:08:21 +00:00
33e104ee65
[train] enforce fp32 master weights in every training entry point
Will Lin and Claude Fable 5
2026-08-14 02:08:05 +00:00
b65b37f73e
[train] run ops: checkpoint start-step gating, LR reset on resume, W&B fallback, weights-only DCP export
Will Lin and Claude Fable 5
2026-08-14 02:07:51 +00:00
987b5399b7
[train] MiniMax-H3: text-only schema for data-free DMD2 preprocessing
Will Lin and Claude Fable 5
2026-08-14 02:07:36 +00:00
528cef02c4
optimize VAE memory
H1yori233
2026-08-12 02:48:43 -07:00
038384e12e
[train] DMD2: intermediate-latent visualization callback (legacy training/ parity)
Will Lin and Claude Fable 5
2026-08-12 06:46:49 +00:00
6c5bc858c9
[train] DMD2: reuse the generator rollout for the critic loss on student steps
Will Lin and Claude Fable 5
2026-08-12 05:06:49 +00:00
8e06166115
[train] DMD2: skip the unconditional teacher forward at guidance scale 1.0
Will Lin and Claude Fable 5
2026-08-12 05:06:16 +00:00
19fb3ba75c
[train] DMD2: per-modality loss normalization for packed multi-modality models
Will Lin and Claude Fable 5
2026-08-12 05:05:47 +00:00
825c6365e0
bundle variant -> preset resolution, pipeline pin, and LTX-2.5 examples
SolitaryThinker
2026-08-10 13:40:35 -07:00
b2a72b3933
docs: Slurm handoff for the 32x GB200 data-free DMD2 run
Will Lin and Claude Fable 5
2026-08-11 09:53:58 +00:00
a20b69a4a8
[train] MiniMax-H3 config suite: SFT sparsity sweep, data-free DMD2, 32-GPU Slurm launch, README
Will Lin and Claude Fable 5
2026-08-11 04:41:49 +00:00
69fb5af602
[train] H3 VSA/DMD2 training fixes: bf16 boundaries, validation offload, buffer-free AdamW
Will Lin and Claude Fable 5
2026-08-11 04:41:35 +00:00
10dee7b351
[feat] per-rank sharded base-weight cache for fast training relaunches
Will Lin and Claude Fable 5
2026-08-11 04:41:12 +00:00
95f4f54712
[bugfix] Triton block-sparse backward: scale QK^T after the dot, keep K raw
Will Lin and Claude Fable 5
2026-08-11 04:40:58 +00:00
bb1fb7c149
WIP: route single-file bundles through the modular train stack
SolitaryThinker
2026-08-10 13:06:06 -07:00
436c44e583
bundle route through the real pipeline: module map, component paths, loader branches, connector factorization
William Lin
2026-07-31 07:24:23 -07:00
6343eadbec
resolve pipeline config for single-file bundles from declared transformer class
William Lin
2026-07-29 19:03:16 -07:00
3f5614fd03
single-file bundle loader: metadata config, prefix routing, per-stream ff bias, encoder-root resolution
William Lin
2026-07-29 16:04:10 -07:00
0d1fe20fdb
[misc] Fixed yapf formatting on LingBot-World-Fast pipeline config
Ishan Vaish and Claude Sonnet 5
2026-07-29 18:29:35 -07:00
7f7431779e
[new-model] Added LingBot-World-Fast image-to-video support
Ishan Vaish and Claude Sonnet 5
2026-07-29 18:20:37 -07:00
8208536cd1
[bugfix] profiler region system: record + export actually work, usability roll-up (#1691 )
William Lin
2026-08-09 20:35:53 -07:00
3cc2420e84
make the region system actually usable: short names, summaries, nvtx, dead-code removal
SolitaryThinker
2026-08-06 17:03:59 -07:00
ccaafd327f
fix the profiler region system: it recorded nothing with default env
SolitaryThinker
2026-08-06 16:38:45 -07:00
0653f8f3af
[new-model] Add V2A: native MMAudio inference pipeline (#1622 )
Kai
2026-08-09 20:29:27 -04:00
91bf3f2752
[bugfix] restrict setuptools package discovery to fastvideo/dreamverse
Will Lin
2026-08-09 22:44:35 +00:00
eb2cd4e1cd
docs: on-tray handoff for the VSA-H3 DMD2 overfit experiment
SolitaryThinker
2026-08-09 15:21:52 -07:00
a0abcd837f
[fix]: handle MMAudio integer timesteps
SolitaryThinker
2026-08-06 20:00:12 -07:00
cceacec997
[fix]: preserve causal masks in FlashAttention
SolitaryThinker
2026-08-03 18:19:03 -07:00
4cabd0a05d
[docs]: record MMAudio GB200 validation
SolitaryThinker
2026-08-02 18:06:04 -07:00
be61633ac3
[fix]: update MMAudio VAE component
SolitaryThinker
2026-08-02 17:21:41 -07:00
c4200f9ad7
feat: add native MMAudio V2A inference pipeline
Kai
2026-07-20 14:04:32 +00:00
cbc4195c5a
[docs] Document DGX Spark (GB10) local perf benchmarking + baseline follow-up
Raghav
2026-08-05 20:21:21 -07:00
982ddef6da
[ci] Validate run_config.gpu_types at discovery (list of non-empty strings)
Raghav
2026-08-03 17:28:27 -07:00
0d9d4ad132
[ci] Ground GB10 perf thresholds from 5x reproducibility run (39.3s / 8.4GB)
Raghav
2026-08-03 17:19:53 -07:00
7445aeabfb
[ci] Add DGX Spark (GB10) single-GPU perf config + gpu_types device gate
Raghav
2026-08-01 11:36:28 -07:00
78f719bf08
feat: add native MMAudio components for V2A
Kai
2026-07-20 05:09:29 +00:00
d81f1f166b
[train] VSA-H3 DMD2 overfit: per-role backends, VSA metadata wiring, wandb guard
SolitaryThinker
2026-08-09 14:46:58 -07:00
3ff189671e
[docs]: replace config generator with native cookbook
SolitaryThinker
2026-08-06 16:56:55 -07:00
44a85cc180
[docs] refresh consumer guide recipes and deployment
SolitaryThinker
2026-08-06 15:31:25 -07:00
c46601a673
[misc] enable TypeScript type-checking in config generator build
KyleAtASU
2026-05-21 14:20:59 -07:00
2b2a89ad91
[feat] build config generator from source in CI instead of committing static output
KyleAtASU
2026-05-21 14:08:23 -07:00
dd37a13598
remove useless svg
KyleAtASU
2026-05-16 20:50:28 -07:00
af0a910cdb
[feat] add interactive config generator pages to Quick Start and Advanced Tuning Guide
KyleAtASU
2026-05-16 20:16:30 -07:00
2cde8feb91
[feat]: add consumer guide pages
KyleAtASU
2026-05-05 16:41:07 -07:00
7f81df6acf
Merge remote-tracking branch 'origin/vsa-h3' into feat/h3-dmd2-vsa
SolitaryThinker
2026-08-09 14:28:26 -07:00
dc70667483
[train] MiniMax H3 DMD2 distillation via packed dual-modality adapter
SolitaryThinker
2026-08-09 13:34:46 -07:00
e0d702decb
[feat] VSA for MiniMax H3: packed mixed-modality sparse attention (#1695 )
William Lin
2026-08-09 13:10:51 -07:00
7dece0471f
[misc] VSA-H3: /simplify pass — dedupe with the Wan VSA backend, drop dead state
SolitaryThinker
2026-08-09 12:58:59 -07:00
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
SolitaryThinker
2026-08-06 16:01:18 -07:00
541ef014ee
[perf] MiniMax-H3: rank-reduced AdaLN pruned model option (-39% params, -23 GiB VRAM) (#1699 )
Shao Duan
2026-08-10 03:31:57 +08:00
ffc1a7a58b
[refactor] H3 pipeline cleanup: shared helpers, dead machinery, loop-invariant hoists (#1698 )
William Lin
2026-08-09 04:51:31 -07:00
9028953625
[misc] yapf pass under CI's interpreter (3.12) + pin hook language_version (#1702 )
William Lin
2026-08-08 22:24:17 -07:00
be7cc7a1dd
studio: keep frontend polling noise out of the engine console
SolitaryThinker
2026-08-08 18:36:37 -07:00
63a02d55f4
Add clip fix
kevin314
2026-08-08 12:39:46 +00:00
8c5778fa44
studio: create generators on one persistent loader thread
SolitaryThinker
2026-08-07 18:28:36 -07:00
0fbdbe9174
studio: pin dit_layerwise_offload=False at generator creation
SolitaryThinker
2026-08-07 18:20:09 -07:00
659cf4fbfe
studio: attach the worker log queue at generator creation, not via RPC
SolitaryThinker
2026-08-07 18:11:48 -07:00
6eb95693a1
[misc]: re-run yapf on main so pre-commit passes again (#1700 )
KyleNeverGivesUp and Claude Opus 5
2026-08-07 16:17:52 -07:00
c3567eb468
[feat] add Minimax H3 sft pipeline (#1688 )
Junda Su
2026-08-07 16:12:04 -07:00
609648b114
studio ui: resolution presets, inline video, Studio nav label, cluster panel
SolitaryThinker
2026-08-07 00:59:27 -07:00
9a2eda427f
studio: GET /api/cluster — per-node GPU telemetry via ray
SolitaryThinker
2026-08-07 00:51:09 -07:00
cb7e8e960a
studio ui: show the full prompt on job cards (wrap, don't truncate)
SolitaryThinker
2026-08-07 00:41:16 -07:00
4a4e9b3f31
studio ui: compact job cards to a two-line layout
SolitaryThinker
2026-08-07 00:38:58 -07:00
0f2bbcda10
h3: rank-0 denoising progress bar; studio: strip ANSI in the progress bridge
SolitaryThinker
2026-08-07 00:33:03 -07:00
58f7b913b9
studio: drive job progress from ray-relayed worker output
SolitaryThinker
2026-08-07 00:26:52 -07:00
07f1cfc356
ray executor: wait out async GPU release on shutdown; retry node GPU probe on init
SolitaryThinker
2026-08-07 00:05:38 -07:00
3101331020
studio: serialize all generator-slot transitions; wait for ray PG removal
SolitaryThinker
2026-08-07 00:00:50 -07:00
4a63de71d7
studio ui: default to the warm model, populate from model presets
SolitaryThinker
2026-08-06 23:09:14 -07:00
99df34a78d
studio: GET /api/models/presets — per-model sampling defaults
SolitaryThinker
2026-08-06 22:59:03 -07:00
7b94433092
studio ui: single-slot model panel + engine output console
SolitaryThinker
2026-08-06 22:50:26 -07:00
3a7e7fc69f
studio: single resident generator + engine log tail; fix ray PG leak
SolitaryThinker
2026-08-06 22:40:39 -07:00
02a10268e9
studio ui: warm models panel on the inference page
SolitaryThinker
2026-08-06 22:29:03 -07:00
9c9dd70214
studio: preload/list/unload resident generators
SolitaryThinker
2026-08-06 22:20:28 -07:00
9f94673c34
fix
H1yori233
2026-08-06 21:13:14 -07:00
c150889330
update server
H1yori233
2026-08-06 20:47:35 -07:00
dc13c59dca
fix ray executor: implement the log-queue abstract methods
SolitaryThinker
2026-08-06 19:16:33 -07:00
Raghav K