All pull requests

Pull requests

• [perf]: training-free token reduction for MiniMax H3

  scope: inferencescope: infrascope: modeltype: perf
  #2004·KyleNeverGivesUp opened Oct 22, 2026Collaborator·2 of 5 tasks·

  ·
  1

• [feat] MiniMax-H3 Ref2VA SFT: train each clip at its own canvas and frame count

  scope: datascope: docsscope: infrascope: trainingtype: feat
  #2003·tianwei-mo opened Oct 22, 2026·

  ·
  4

• [perf] Cache MLX H3 text refinement and speed up attention

  scope: docsscope: inferencescope: infratype: perf
  #2002·posiposi443-ops opened Oct 22, 2026·4 of 5 tasks·

  ·
  6

• [feat] MiniMax-H3 Ref2VA: continue an earlier video from its last frames

  scope: docsscope: inferencescope: infratype: feat
  #2001·tianwei-mo opened Oct 21, 2026·4 tasks done·

  ·
  7

• [kernel] GB10 (DGX Spark) CUDA VSA forward for 128-token tiles, replacing the split-Triton route

  needs-rebasescope: attentionscope: docsscope: infrascope: kernel
  #1998·SolitaryThinker opened Oct 20, 2026Collaborator·4 tasks done·

  ·
  5

• [bugfix] Normalize local and HF-cache model paths before registry lookup

  scope: infratype: bugfix
  #1997·Jackmeson1 opened Oct 20, 2026·3 of 4 tasks·

  ·
  1

  3

• [feat]: shared FastH3 NVFP4 FP4-attention QAD recipe (OmniRef + V2 T2AV)

  needs-rebasescope: attentionscope: docsscope: infrascope: kernelscope: modelscope: trainingtype: feat
  #1989·aryan5v opened Oct 19, 2026Collaborator·4 of 6 tasks·

  ·
  24

• [bugfix] SP warmup: use the rank's own device instead of hardcoded CUDA

  scope: distributedscope: infratype: bugfix
  #1988·li-lizhe opened Oct 19, 2026Contributor·4 tasks done·

  ·
  1

  4

• [infra]: fix the DGX Spark (GB10) Docker image: importable flashinfer, sm_121a kernels, native-matching output

  scope: docsscope: infratype: infra
  #1986·SolitaryThinker opened Oct 19, 2026Collaborator·

  ·
  3

• [feat] Add release T2AV DMD training and runtime optimizations

  needs-rebasescope: attentionscope: datascope: docsscope: inferencescope: infrascope: modelscope: trainingtype: feat
  #1985·SolitaryThinker opened Oct 19, 2026Collaborator·3 of 5 tasks·

  ·
  1

• [kernel] sm_103a VSA: take the softmax row max from tcgen05.ld.red

  scope: kernel
  #1982·lpc0220 opened Oct 18, 2026Contributor·3 of 4 tasks·

  ·
  2

• [kernel] VSA backward (64-token blocks): programmatic dependent launch optimization to improve perf

  scope: docsscope: inferencescope: kernel
  #1981·lpc0220 opened Oct 18, 2026Contributor·3 of 4 tasks·

  ·
  2

• [bugfix] End partial causal Wan mask blocks at the real token boundary

  scope: infrascope: modeltype: bugfix
  #1976·zmy1260222041 opened Oct 17, 2026·3 of 4 tasks·

  ·
  2

• [feat] MiniMax-H3 Ref2VA: selectable reference-image resize mode

  needs-rebasescope: docsscope: inferencescope: infratype: feat
  #1975·tianwei-mo opened Oct 17, 2026·4 tasks done·

  ·
  3

• [feat] VSA-H3 tile 128: run two 64-token Triton tiles where the sm_100a kernel is unavailable

  needs-rebasescope: attentionscope: docsscope: infratype: feat
  #1974·tianwei-mo opened Oct 17, 2026·4 tasks done·

  ·
  4

• [docs] Propose inference VRAM estimation and hardware recommendations

  scope: docstype: docs
  #1972·leo2ultimate opened Oct 17, 2026Contributor·3 tasks done·

  ·
  1

• [feat] Reuse Wan inference weights across process restarts

  needs-rebasescope: docsscope: inferencescope: infrascope: modeltype: feattype: infra
  #1967·GindaChen opened Oct 15, 2026Contributor·

  ·
  1

• [bugfix] Reset VAE tiling for each request

  scope: inferencescope: infratype: bugfix
  #1965·GindaChen opened Oct 15, 2026Contributor·

  ·
  1

• [new-model] Add PAN2 text-to-video and image-to-video

  scope: docsscope: inferencescope: modeltype: new-model
  #1961·SolitaryThinker opened Oct 15, 2026Collaborator·4 of 6 tasks·

  ·
  5

• [bugfix] Count the wait for a concurrency slot in bench latency

  scope: inferencescope: infratype: bugfix
  #1960·konstde00 opened Oct 14, 2026·4 tasks done·

  ·
  5

• [bugfix] Aggregate peak memory across multiprocessing workers

  scope: inferencescope: infratype: bugfix
  #1953·Sl123117 opened Oct 12, 2026·4 tasks done·

  ·
  4

• [bugfix] Resolve Studio attention settings and isolate job execution

  needs-rebasescope: docsscope: infrascope: studiotype: bugfix
  #1952·FlappyBob opened Oct 11, 2026Contributor·3 of 4 tasks·

  ·
  4

• [perf] Shard tokens across SP ranks in causal Wan training (opt-in)

  scope: docsscope: infrascope: modeltype: perf
  #1951·zmy1260222041 opened Oct 11, 2026·4 tasks done·

  ·
  1

  12

• [perf] Enable VSA64 fused layout on DGX Spark GB10

  scope: attentionscope: infrascope: kerneltype: perf
  #1948·fusheng-ji opened Oct 9, 2026Contributor·4 tasks done·

  ·
  6

• [bugfix] Fix TurboDiffusion INT8 quantization tail strides

  scope: infrascope: kerneltype: bugfix
  #1947·fusheng-ji opened Oct 9, 2026Contributor·4 tasks done·

  ·
  2