All pull requests

Pull requests

• [bugfix] Make torch.compile reduce-overhead work in the shared denoising loop (Wan T2V)

  scope: docsscope: inferencescope: infratype: bugfix
  #2047·arun-elr opened Nov 1, 2026·4 of 6 tasks·

  ·
  1

• [perf] MiniMax H3: faster FP4 attention padding and VAE pixel path

  scope: attentionscope: inferencescope: infrascope: modeltype: perf
  #2046·aryan5v opened Oct 31, 2026Collaborator·

  ·
  6

• [docs] Fix cookbook component selection and memory reporting

  scope: docstype: docs
  #2045·aryan5v opened Oct 31, 2026Collaborator·

  ·
  1

• [bugfix] Honor output_video_name when selecting output filenames

  scope: docsscope: inferencescope: infratype: bugfix
  #2044·CHYYYYYYYY opened Oct 30, 2026·4 tasks done·

  ·
  1

  1

• [bugfix] FA4: stop forwarding max_seqlen to the varlen kernel

  scope: attentionscope: infratype: bugfix
  #2043·aryan5v opened Oct 30, 2026Collaborator·

  ·
  4

• [perf] MiniMax H3: opt-in faster kernel tier for the DiT fused ops

  scope: attentionscope: docsscope: infrascope: modeltype: perf
  #2042·aryan5v opened Oct 30, 2026Collaborator·

  ·
  12

• [perf] MiniMax H3: fuse VAE decoder QK RMSNorm and partial RoPE

  scope: infrascope: modeltype: perf
  #2041·aryan5v opened Oct 30, 2026Collaborator·

  ·
  8

• [perf] MiniMax H3: decode eight VAE spatial tiles per call by default

  scope: infrascope: modeltype: perf
  #2040·aryan5v opened Oct 30, 2026Collaborator·

  ·
  6

• [kernel] sm_107a CUDA block-sparse VSA forward (Rubin), 128-token blocks

  scope: kernel
  #2039·lpc0220 opened Oct 30, 2026Contributor·4 tasks done·

  ·
  1

• [new-model] Add Pika H3 Flex inference support

  scope: docsscope: inferencescope: infrascope: modeltype: new-model
  #2038·aryan5v opened Oct 30, 2026Collaborator·

  ·
  9

• [feat] Open-H3-Regenerate-2K: regenerate FastH3 QAD, Trim and V2 drafts at 2K (text to 2K 3.1-3.4x faster)

  scope: docsscope: inferencescope: infratype: feat
  #2037·aryan5v opened Oct 29, 2026Collaborator·4 tasks done·

  ·
  23

• [feat] Open-H3-Regenerate-2K: NVFP4 H3 video decoder (decode 23 s to 5 s at 2K)

  scope: docsscope: inferencescope: infratype: feat
  #2036·aryan5v opened Oct 29, 2026Collaborator·

  ·
  7

• [new-model] Open-H3-Regenerate-2K: 768P H3 draft to 2560x1440, served on the OpenAI video routes

  scope: docsscope: inferencescope: infrascope: modeltype: new-model
  #2035·aryan5v opened Oct 29, 2026Collaborator·3 of 4 tasks·

  ·
  20

• [perf] Wan: fused-norm tier at the norm1 site, exact tier on causal Wan

  scope: docsscope: infrascope: modeltype: perf
  #2033·incisors opened Oct 29, 2026Contributor·

  ·
  1

• [feat] Add opt-in BF16 forward/SP4 Wan VSA backend for Ascend NPUs

  scope: attentionscope: distributedscope: docsscope: infratype: feat
  #2032·Gxj230958 opened Oct 29, 2026Contributor·4 of 6 tasks·

  ·
  1

• [feat] OpenAI video API: accept data: URLs for video and audio references

  scope: docsscope: inferencescope: infratype: feat
  #2030·alexzms opened Oct 28, 2026Collaborator·4 of 8 tasks·

  ·
  1

  7

• [new-model] Re-add PAN2 text-to-video and image-to-video, with inference optimizations and guardrails

  needs-rebasescope: docsscope: inferencescope: infrascope: kernelscope: modeltype: new-model
  #2029·SolitaryThinker opened Oct 28, 2026Collaborator·

  ·
  18

• [feat] Per-request community LoRA in the OpenAI video API

  scope: docsscope: inferencescope: infrascope: modeltype: feat
  #2026·aryan5v opened Oct 28, 2026Collaborator·5 tasks done·

  ·
  10

• [feat] Reuse MiniMax-H3 text conditioning across seed sweeps

  scope: docsscope: inferencescope: infrascope: modeltype: feat
  #2022·LeslieWylie opened Oct 27, 2026Contributor·

  ·
  12

• [perf] Kandinsky-5 kernel fusion: Triton-fused residual+LayerNorm+modulate inference path

  scope: docsscope: infrascope: modeltype: perf
  #2021·arcusbuilds opened Oct 26, 2026·

  ·
  1

• [perf] Wan: bit-exact Triton residual/modulate kernels as the default…

  scope: docsscope: infrascope: modeltype: perf
  #2020·incisors opened Oct 26, 2026Contributor·

  ·
  1

• [DoNotMerge] [perf] FastH3 on DGX Spark: 1.8x on one Spark, 3.6x on a pair, 1.27x Omni (same computation)

  scope: distributedscope: docsscope: inferencescope: infrascope: kernelscope: modeltype: perf
  #2015·GindaChen opened Oct 23, 2026Contributor·4 tasks·

  ·
  1

• [new-model] Add native Bernini-R 1.3B video editing

  needs-rebasescope: docsscope: inferencescope: infrascope: modeltype: new-model
  #2010·Nhckdvrl opened Oct 22, 2026·8 tasks done·

  ·
  3

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

  scope: docsscope: inferencescope: infrascope: modeltype: perf
  #2004·KyleNeverGivesUp opened Oct 22, 2026Collaborator·3 of 5 tasks·

  ·
  2

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

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

  ·
  10