-
fa2d1324c6
[feat]: FastH3 Omni (Ref2VA) on Apple Silicon MLX (#2000)
main
Aryan KumarandAryan Kumar
2026-10-21 12:01:54 -07:00
-
57662a9ea5
[fix] Sync PAN2 padding parity with latest upstream (#1961)
feat/pan2
SolitaryThinker
2026-10-20 22:28:39 +00:00
-
f80184af75
[misc] Merge main into PAN2 port (#1961)
SolitaryThinker
2026-10-20 22:18:40 +00:00
-
-
-
-
cdda1b7535
[feat] Add H3 PDD/DMPDD training and packed Ref2VA preprocessing (#1984)
William Lin
2026-10-20 11:15:57 -07:00
-
089a5a55a0
[feat] Cookbook serving, hardware catalog, and landing page (#1999)
Aryan KumarandAryan Kumar
2026-10-20 09:10:29 -07:00
-
137d9110ca
[docs] Make cookbook copy actions explicit and hardware opt-in (#1992)
Aryan Kumar
2026-10-20 09:00:56 -07:00
-
8503617df3
[kernel] GB10 VSA tile-128: second review pass (device check, NaN rows, routing, linkage)
kernel/gb10-vsa128-tile128
SolitaryThinker
2026-10-20 15:21:51 +00:00
-
d4f144581d
[feat] Add capacity metadata to existing cookbook model manifests (#1983)
2026-10-20 01:24:19 -07:00
-
5f96664abf
[feat] Added a single-GPU VRAM calculator for cookbook deployments (#1978)
Ishan
2026-10-20 01:12:06 -07:00
-
3f0d72f22f
[kernel] GB10 VSA tile-128: review fixes for routing, limits and padding
SolitaryThinker
2026-10-20 07:33:13 +00:00
-
32d8a8765b
[kernel] GB10 (DGX Spark) CUDA VSA forward for 128-token tiles
SolitaryThinker
2026-10-20 05:39:54 +00:00
-
-
-
376e445f0e
[feat]: FastH3 OmniRef inference on consumer devices (RTX 5090 / PRO 6000 / DGX Spark / 4090) (#1980)
2026-10-19 14:57:36 -07:00
-
2099dd5533
[feat]: Community MiniMax-H3 LoRA loading with quantized side branch (#1973)
2026-10-19 14:29:23 -07:00
-
2e4878c5e6
[infra]: leave FlashAttention out of the GB10 image
infra/gb10-image-sm121a
SolitaryThinker
2026-10-19 10:09:54 +00:00
-
fba50356b1
[infra]: skip the FA4 cute overlay in the GB10 image
SolitaryThinker
2026-10-19 08:13:04 +00:00
-
b4cd6ea00a
[feat] Add release T2AV DMD training and runtime optimizations
feat/h3-t2av-dmd-training
SolitaryThinker
2026-10-19 04:05:25 +00:00
-
-
-
431231a285
[infra]: build the GB10 image kernel for sm_121a
SolitaryThinker
2026-10-19 03:50:29 +00:00
-
-
-
0033a9b11d
[bugfix] Fix FastH3 regional compile recompiling for every prompt length (#1942)
2026-10-18 18:22:06 -07:00
-
91a1c50909
[fix] Fix latent handoff and FastWan-QAD example path (#1557)
2026-10-18 17:36:40 -07:00
-
3658b120da
[bugfix] Fix tests that fail on main: FastH3 Trim Spark audio, test isolation (#1957)
2026-10-18 16:07:59 -07:00
-
81b73f6b88
[bugfix] Fix cookbook clipboard test blocking docs deployment (#1979)
William Lin
2026-10-17 23:08:49 -07:00
-
d4ed61cc7c
[refactor] VSA: rename the sm100a block-sparse kernel to ptx_blackwell (#1956)
lpc0220androot
2026-10-17 21:10:19 -07:00
-
377c79fa7e
[docs] Add William Lin to the contact page (#1977)
William Lin
2026-10-17 20:52:55 -07:00
-
25940a7e86
[docs] Unify cookbook model editors and launch commands (#1970)
Aryan KumarandAryan Kumar
2026-10-16 17:26:51 -07:00
-
3f45f2da21
[bugfix] Repair DreamVerse and failing CI gates (#1958)
2026-10-16 17:15:38 -07:00
-
6212e7262e
[docs] Migrate documentation and cookbook to Blume (#1969)
Aryan KumarandAryan Kumar
2026-10-16 16:09:04 -07:00
-
39a8db2f5d
[feat] Reuse Wan inference weights through normal loaders
feat/pipeline-startup-profiling
Junda Chen
2026-10-15 11:20:05 -07:00
-
042e8ee580
[feat] Retain native model weights in a CUDA IPC donor
Junda Chen
2026-10-15 01:35:13 -07:00
-
994be72623
[infra] Profile native encoder and DiT startup phases
Junda Chen
2026-10-15 01:14:06 -07:00
-
ba371e33ed
[ci] Stop automatically marking issues and PRs stale (#1964)
William Lin
2026-10-15 01:05:48 -07:00
-
fc1b93f15c
[docs] Clarify CUDA startup timing boundaries
feat/startup-loader-foundation
Junda Chen
2026-10-15 00:50:58 -07:00
-
0ee43c01f9
[refactor] Separate native VAE loading phases and measure startup
Junda Chen
2026-10-15 00:49:49 -07:00
-
-
-
61ebcbb635
[bugfix] Reset VAE tiling for each request
fix/vae-tiling-reset
Junda Chen
2026-10-15 00:31:25 -07:00
-
06c8c5f4b8
[docs] Add public contact page (#1959)
William Lin
2026-10-15 00:26:44 -07:00
-
-
-
c535797e3d
[skill] Remove the add-model and decompose-pipeline-pr agent skills (#1963)
William Lin
2026-10-15 00:10:49 -07:00
-
59fddb00f2
[ci] Remove automated community comments (#1962)
William Lin
2026-10-14 23:40:45 -07:00
-
9f59519959
[misc] Remove DreamX World, HunyuanGameCraft, GLM-Image, LongCat, TurboDiffusion, and Z-Image (#1926)
William Lin
2026-10-14 23:16:32 -07:00
-
26719dcfe7
[new-model] Add PAN2 text-to-video and image-to-video
SolitaryThinker
2026-10-15 03:44:35 +00:00
-
-
b7d4b37efd
Add jobs timers
klin/h3-studio-editor
kevin314
2026-10-14 01:05:13 +00:00
-
4d0796b281
[bugfix]: Hide only the VSA-H3 block mask from regional compile
fix-h3-regional-compile-recompiles
aryan5v
2026-10-09 23:37:59 -07:00
-
d5253fd710
[feat] Add serving cookbook recipes and configuration API (#1941)
2026-10-09 15:49:08 -07:00
-
08fee053f0
[bugfix]: Drop the eager recompile-limit fallback; copy shards only after VSA prep succeeds
aryan5v
2026-10-09 15:46:50 -07:00
-
8bd5e88f8c
[perf]: reuse a pre-touched pixel buffer in the H3 video decode stage
aryan/h3-spark-gb10
Aryan KumarandClaude Opus 5.5
2026-10-07 20:24:18 -07:00
-
1c1aa75f33
[perf]: faster FastH3 decode and NVFP4 GEMMs in the DGX Spark recipes
Aryan KumarandClaude Opus 5.5
2026-10-06 23:28:42 -07:00
-
d0b91810b2
[perf]: skip the NVFP4 activation pad copy for 128x4 scale layouts
Aryan KumarandClaude Opus 5.5
2026-10-06 23:08:46 -07:00
-
-
-
3c58fd7af4
[bugfix]: FastH3 regional compile reuses one graph across prompt lengths (#1940)
aryan5v
2026-10-08 23:09:13 -07:00
-
-
-
2b164405c3
[bugfix] Accept log_queue in StreamingVideoGenerator.from_fastvideo_args (#1937)
William Lin
2026-10-07 15:58:03 -07:00
-
1033fa8e1c
[tests] Use the env helpers in the HunyuanVideo batch test (#1935)
William Lin
2026-10-07 15:39:39 -07:00
-
e1e2559359
[perf] FastH3 OmniRef: lossless speed-ups behind default-off switches (#1930)
Davide Locatelli
2026-10-07 15:35:37 -07:00
-
8d2e1cf328
[bugfix] Skip streaming IPC queues for standard VideoGenerator (#1902)
Wenbo Ji 嵇文博
2026-10-08 00:33:11 +02:00
-
4ebf11f7d9
[bugfix] Fix test env defaults and server dependencies (#1901)
Wenbo Ji 嵇文博
2026-10-08 00:32:22 +02:00
-
ef63de1f60
[perf] Fuse VSA64 tile layout for SM100 inference (#1900)
Wenbo Ji 嵇文博
2026-10-08 00:31:58 +02:00
-
e5f31c1b4c
[bugfix] Fix HunyuanVideo forward pass with batch size greater than one (#1921)
Michael Yang
2026-10-07 15:15:12 -07:00
-
b7bd7775ed
[perf] Default SM100 long-sequence Attn-QAT tiling (#1899)
Wenbo Ji 嵇文博
2026-10-08 00:14:38 +02:00
-
dc72a35f4e
[bugfix] Return individual images for batched image API requests (#1917)
Boyang ZHONG
2026-10-08 00:06:02 +02:00
-
7d61590213
[bugfix] Fix the CUDA device of Ray workers on multi-GPU nodes (#1932)
William Lin
2026-10-07 15:00:23 -07:00
-
3c713d808a
[bugfix] Prevent ParquetDatasetWriter from overwriting shards across flushes (#1922)
Chanu Ollala
2026-10-07 14:56:53 -07:00
-
6ef5b78020
[bugfix] LTX-2: train temporal RoPE at the clip's fps (#1892)
alanhuangyoo
2026-10-08 05:54:34 +08:00
-
6275ac2c8e
[bugfix] Validate image parameters before generation and uploads (#1916)
Boyang ZHONG
2026-10-07 23:52:15 +02:00
-
73e878a01c
[bugfix]: fix the MLX server and a GPU-only test leak after the 2026-10-07 merges (#1931)
William Lin
2026-10-07 14:35:05 -07:00
-
3907a69727
[bugfix] Preserve image output formats (#1915)
Boyang ZHONG
2026-10-07 23:23:15 +02:00
-
d1416b5930
[misc]: re-run pre-commit after the 2026-10-07 merges (#1929)
William LinandClaude Opus 5.5
2026-10-07 06:13:43 -07:00
-
c520878fd2
[perf]: QAT/QAD add safe regional compile for modular training (#1718)
Zhang PeiyuanandClaude Opus 5.5
2026-10-07 05:24:24 -07:00
-
99f8f3e7a7
[feat] Add streaming and GPU-accelerated LoRA extraction (#1784)
2026-10-07 04:48:35 -07:00
-
14e4b7189b
[feat] Add Cosmos Predict2.5 DFD continuation to DreamVerse (#1768)
Raghav KandClaude Opus 5.5
2026-10-07 04:46:54 -07:00
-
8530fdc550
[feat] Add external-launcher executor for multi-node inference (#1746)
2026-10-07 04:46:26 -07:00
-
315b25d26d
[new-model] Add Cosmos Predict2.5 distilled T2W and DFD V2W inference (#1767)
Raghav KandClaude Opus 5.5
2026-10-07 04:42:44 -07:00
-
b8920ba50a
[misc]: add env var to control process-aware logging (#1680)
2026-10-07 04:31:52 -07:00
-
3467f8befb
[feat] Add end-to-end MMAudio V2A training, inference, and evaluation (#1648)
2026-10-07 12:25:00 +01:00
-
7320d9d350
[feat] Add HunyuanVideo 1.5 embedding preprocessing pipeline (#1663)
Kun LinandClaude Opus 5.5
2026-10-07 07:20:58 -04:00
-
af78c759f1
[new-model] Add Wan2.2-Animate-14B character animation/replacement inference (#1765)
2026-10-07 04:19:37 -07:00
-
f82c190cbb
[kernel] Build + allow attn_qat_infer FP4 attention on sm_121a (DGX Spark) (#1598)
2026-10-07 04:09:39 -07:00
-
e722023bd6
[perf] Wan kernel fusion: Triton-fused residual+LayerNorm+modulate inference path (#1708)
2026-10-07 04:08:07 -07:00
-
bc0191d775
[perf] Harden Ulysses agreement and tune bounded H3 transfers (#1807)
2026-10-07 04:01:53 -07:00
-
2d6e7d9b6b
[feat]: MiniMax-H3 encoder split: dedicated text-encoder node group for 720p on DGX Spark (#1877)
2026-10-07 18:46:33 +08:00
-
5e61accc37
[misc] Clean up QAD 5090 example inference scripts (#1496)
Kevin LinandClaude Opus 5.5
2026-10-07 05:44:49 -05:00
-
f60514183c
[ci] Separate performance behavior tests (#1620)
2026-10-07 03:43:12 -07:00
-
156611bab3
[ci]: run the full attention test directory in the unit lane (#1658)
William LinandClaude Opus 5.5
2026-10-07 03:37:46 -07:00
-
6d6195a9f1
[feat]: fp8 PV mode for the FA4-FP4 attention path (#1654)
William LinandClaude Opus 5.5
2026-10-07 03:36:54 -07:00
-
8348e83e80
[feat] Import MiniMax H3 Comfy NVFP4-AWQ text encoder (#1865)
2026-10-07 18:34:40 +08:00
-
2452022f15
[new-model] Add LTX-2.5 inference support (#1704)
2026-10-07 03:34:18 -07:00
-
4b3f99b224
[new-model] Added LingBot-World-Fast image-to-video support (#1665)
2026-10-07 03:32:38 -07:00
-
d967b99928
[bugfix]: registry ambiguity warning by specificity + QAD weights file path (#1649)
William LinandClaude Opus 5.5
2026-10-07 03:30:55 -07:00
-
8f1eb2992c
[bugfix]: retain NVFP4 weights for LTX2 refinement LoRA (#1705)
2026-10-07 03:29:46 -07:00
-
683689de4f
[bugfix] Fix Ulysses sequence-parallel crash in self-forcing causal Wan DiT (#1627)
pkisfaludi-nvandClaude Opus 5.5
2026-10-07 03:29:22 -07:00
-
d5287f8352
[feat]: add MiniMax H3 Ref2VA and LoRA training support (#1757)
2026-10-07 18:21:45 +08:00
-
cdfbd64b04
[perf] FP8: fuse the dynamic activation quantization into Triton kernels (#1861)
2026-10-07 18:20:04 +08:00
-
299d0f6838
[bugfix] wrap the generic DenoisingStage in the inference_denoising profiler region (#1707)
Haochen JiangandClaude Opus 5.5
2026-10-07 03:19:42 -07:00
-
c18436125a
[misc]: Refactor separate training pipeline runner (#1696)
2026-10-07 03:19:21 -07:00
-
5d515ee617
[perf]: remove per-step host syncs and repeated work from causal inference (#1687)
2026-10-07 03:18:56 -07:00
-
37b9dc1836
[new-model] Add native Helios-Distilled T2V pipeline (#1670)
2026-10-07 03:18:24 -07:00
-
a0291a57c1
[ci] Add Wan2.2-TI2V-5B SSIM regression test (#1772)
2026-10-07 02:44:24 -07:00
-
6102fac00d
[bugfix]: fix modular ops activation checkpoint recomputation (#1716)
2026-10-07 02:41:12 -07:00
-
1ad33c1832
[feat] Support training_cfg_rate > 0 for LTX-2 (#1752)
2026-10-07 17:38:42 +08:00
-
da87973412
[ci] Clarify dashboard cohort grouping for legacy v1 records (#1613)
2026-10-07 02:33:18 -07:00
-
a3fd1d0bab
[ci]: prevent direct tests from minting suite gates (#1615)
William LinandClaude Opus 5.5
2026-10-07 02:32:48 -07:00
-
3e87d6f4ae
[feat] ROCm: route the video sparse attention backends to their Triton kernels (#1851)
2026-10-07 17:32:27 +08:00
-
4bb5627fa8
[feat] MiniMax H3: block-FP8 text encoder on Hopper, its converter, and a 4xH100 FastH3 profile (#1829)
2026-10-07 17:31:52 +08:00
-
08ab8b5c00
[bugfix] ROCm: detect the platform from a HIP torch build when amdsmi is unavailable (#1849)
2026-10-07 17:20:47 +08:00
-
fb5efebdab
[feat] Add HunyuanVideo 1.5 T2V training support (#1662)
2026-10-07 02:19:22 -07:00