Junda Su
02f1ce11ae
[misc] Move environment var reads into env registry and rename unprefixed variables ( #1897 )
2026-09-29 14:36:16 -07:00
Junda Su
e3b88bb12a
[feat] Add a typed environment-variable registry, policy doc, and contract test ( #1896 )
2026-09-29 12:50:45 -07:00
Junda Su
cb66acd400
[bugfix] Fix environment-variable bugs in LTX-2 debug logging, HF token lookup, and attention backend reads ( #1895 )
2026-09-29 11:48:31 -07:00
alanhuangyoo and SolitaryThinker
9dd2a837f4
[bugfix] encoders: keep Qwen2.5-VL importable on transformers 5 ( #1791 )
...
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com >
2026-09-24 17:13:23 -07:00
Hexu Zhao and Claude Opus 5
10bdf5e076
[bugfix] frozen offload: no host copy for an already resident module
...
`load` took the host copies before checking which tensors were missing from
the device, so with `vae_cpu_offload=False` — where the loader builds the VAEs
on CUDA and `unload` is never called — it pulled about 11 GB per rank back over
PCIe once and kept a pinned mirror of it for the life of the process, for a
module that never leaves the device.
It now asks what is absent first and returns early when nothing is. The
offloaded path is unchanged: weights still round-trip and come back bit
identical.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com >
2026-09-20 22:46:13 -07:00
Hexu Zhao and Claude Opus 5
3e26db40b0
[perf] H3 decode: pin the frozen VAEs' host copies
...
With the device-to-host copy gone, the remaining host-to-device copy is the
stage's cost, and from pageable memory it is staged through a bounce buffer at
roughly 2.6 GB/s instead of full PCIe speed. The host copies are made once, so
pinning them costs nothing per request.
It does cost the module's size in non-pageable host memory (10.4 GB for the
H3 video VAE, 0.6 GB for the audio VAE, per rank) for as long as the pipeline
is alive. It follows --pin-cpu-memory, which is on by default; --no-pin-cpu-
memory keeps the copies pageable and only loses the bandwidth. Note that this
flag did not already reach the VAE: its other use is FSDP's own CPUOffloadPolicy
and the VAE does not go through FSDP.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com >
Claude-Session: https://claude.ai/code/session_01ATw7g6cxq9N2LtnMratsru
2026-09-20 22:46:13 -07:00
Hexu Zhao and Claude Opus 5
c73dd0ab55
[perf] H3 decode: offload the frozen VAEs without the device-to-host copy
...
The decode stages move the 10.4 GB fp32 video VAE (and the 0.6 GB audio VAE)
with module.to(device) / module.to("cpu") on every request. The weights are
frozen -- the whole stage runs under no_grad -- so the copy back to the host
returns bytes the host already has.
Keep one host copy per tensor, made the first time the module is loaded, and
on unload point .data back at it instead of copying. One H2D per request, no
D2H at all; the module still lives on the host between requests. Where the old
path allocated a fresh host tensor per cycle and freed the previous one, this
reuses one buffer, so peak host memory can only go down.
Co-Authored-By: Claude Opus 5 (1M context) <noreply@anthropic.com >
Claude-Session: https://claude.ai/code/session_01ATw7g6cxq9N2LtnMratsru
2026-09-20 22:46:13 -07:00
430e52154e
[bugfix]: Restore exact Qwen3-VL vision interpolation ( #1737 )
...
Co-authored-by: William Lin <8941107+SolitaryThinker@users.noreply.github.com >
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com >
2026-09-18 05:00:05 -07:00
li-lizhe
39718cd54d
[bugfix] fix(cosmos): make AdaLayerNorm autocast device-agnostic ( #1818 )
2026-09-15 07:44:51 -07:00
William Lin
1c14afd559
[docs]: refresh AGENTS.md maps and add Wan SP/I2V and CI test notes ( #1845 )
2026-09-14 16:06:53 -07:00
Kyle Hu
3a3ad3d209
[feat]: NVFP4 text encoder for MiniMax H3 ( #1837 )
2026-09-12 18:31:59 -07:00
William Lin
556ac7088e
[refactor] Simplify Wan sampling and tests ( #1825 )
2026-09-07 15:57:19 -07:00
Junda Su
e7456f1b75
Add H3 support into Dreamverse ( #1800 )
2026-09-07 13:09:44 -07:00
William Lin
7f83164233
[refactor] Move Wan VAE into the Wan package ( #1824 )
2026-09-05 18:43:55 -07:00
Junda Su
4e52f47d1e
[feat] add MXFP8 support on H3 ( #1796 )
2026-09-05 17:08:58 -07:00
William Lin
e19913f6e9
[refactor] Group Wan transformer and config ( #1823 )
2026-09-05 17:05:32 -07:00
aaef49bfc6
[feat]: run FastH3 across two DGX Sparks with Ray sequence parallel ( #1803 )
...
Co-authored-by: Kyle <shh075@ucsd.edu >
Co-authored-by: Claude Opus 5 (1M context) <noreply@anthropic.com >
Co-authored-by: Satyam Srivastava <srivastavasatyam53@gmail.com >
Co-authored-by: Aryan Kumar <aryan5v@users.noreply.github.com >
2026-09-01 01:37:23 -07:00
Aryan Kumar and Aryan Kumar
cf6a00b9be
[feat]: add opt-in CUDA TAEH3 preview decode for FastH3 ( #1795 )
...
Co-authored-by: Aryan Kumar <aryan5v@users.noreply.github.com >
2026-08-31 23:34:56 -07:00
Aryan Kumar and Aryan Kumar
f82d8be4bf
[perf] Sequential MiniMax H3 start with GPU-direct DiT load ( #1793 )
...
Co-authored-by: Aryan Kumar <aryan5v@users.noreply.github.com >
2026-08-31 04:20:21 -07:00
William Lin and shaoxiongduan
a534ba20a0
[feat] Add MiniMax H3 LoRA inference and preview launchers ( #1771 )
...
Co-authored-by: shaoxiongduan <shaoxiongduan@gmail.com >
2026-08-27 14:43:39 -07:00
KyleNeverGivesUp
9bfa585448
[perf]: stop holding the whole checkpoint during DiT load, unblocking MiniMax H3 on one GB10 ( #1714 )
2026-08-26 15:23:51 -07:00
KyleNeverGivesUp
c9c5585758
[perf]: MiniMax H3 on GB10 - skip text encoder CPU offload on unified memory (5m49s to 30ms) ( #1710 )
2026-08-26 12:03:14 -07:00
William Lin
f9e3680f11
[perf] Align FastH3 optimized inference profile ( #1748 )
2026-08-23 02:01:03 -07:00
William Lin
13213395b4
[perf] Parallelize MiniMax-H3 VAE over sequence ranks ( #1744 )
2026-08-22 18:05:31 -07:00
William Lin
c488fa1211
[perf] Add opt-in packed-varlen FA4 for MiniMax-H3 ( #1742 )
2026-08-22 17:16:47 -07:00
William Lin
d3cff517cd
[perf] Add opt-in regional fullgraph compile for DiT inference ( #1741 )
2026-08-22 12:14:39 -07:00
Junda Su
2f3d407406
[perf] Optimize MiniMax H3 VAE decoding ( #1734 )
2026-08-21 14:57:32 -07:00
Kaiqin Kong
bcffa4026e
[perf] Optimize MiniMax-H3 text encoder memory ( #1732 )
2026-08-21 14:57:06 -07:00
Kaiqin Kong
73dd105f3d
[perf] Add opt-in MiniMax-H3 Sol-Engine fusions ( #1735 )
2026-08-21 12:39:28 -05:00
Shao Duan
c4ad4227c0
[misc] MiniMax-H3: move the AdaLN converter into scripts/checkpoint_conversion ( #1712 )
2026-08-21 02:15:00 -05:00
KyleNeverGivesUp
0462e1b0e7
[perf]: MiniMax H3 - build the Qwen3-VL encoder only as far as it is read (-13.7 GB) ( #1711 )
2026-08-20 23:19:25 -05:00
Kaiqin Kong
e0a3db5651
[perf] Reduce MiniMax-H3 VAE peak memory ( #1703 )
2026-08-20 23:13:23 -05:00
Kai
0653f8f3af
[new-model] Add V2A: native MMAudio inference pipeline ( #1622 )
2026-08-09 17:29:27 -07:00
William Lin
e0d702decb
[feat] VSA for MiniMax H3: packed mixed-modality sparse attention ( #1695 )
2026-08-09 13:10:51 -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
Junda Su
c3567eb468
[feat] add Minimax H3 sft pipeline ( #1688 )
2026-08-07 16:12:04 -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
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
William Lin
422585d08f
[misc]: add LTX-2.3 fine-tuning example recipes ( #1651 )
2026-07-27 18:42:44 -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
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
Junda Su
755a4e4470
[new-model] Add LingBot-Video Dense and MoE/refiner T2V inference ( #1595 )
2026-07-18 20:18:16 -07:00
Junda Su
d4bdd3621a
[new-model] Port LingBot-World-v2 ( #1579 )
2026-07-16 19:51:53 -07:00
6966f9e0bc
[fix] Z-Image ( #1236 ) draft port: rebase + strict-load contract + bf16 encoder parity + PORT_STATUS ( #1339 )
...
Co-authored-by: Mrinaal Dogra <mdogra@ucsd.edu >
Co-authored-by: SolitaryThinker <wlsaidhi@gmail.com >
2026-07-16 18:18:25 -07:00
Junda Su