All pull requests

Pull requests

• [bugfix] Keep image edit uploads inside the uploads directory

  scope: inferencescope: infratype: bugfix
  #1955·JuniverseLi opened Oct 13, 2026·2 of 4 tasks·

  ·
  2

• [perf] Pin FSDP CPU-offloaded shards at exact size (up to ~2x less host RAM)

  scope: infrascope: modeltype: perf
  #1954·JuniverseLi opened Oct 13, 2026·3 of 4 tasks·

  ·
  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

  scope: docsscope: 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

  10

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

  ·
  2

• [perf]: FastH3 on one DGX Spark: V2 480p from 140 s to 78 s (cuDNN NVFP4 GEMMs, dense batched VAE decode)

  scope: docsscope: inferencescope: infrascope: modeltype: perf
  #1946·aryan5v opened Oct 9, 2026Collaborator·7 of 8 tasks·

  ·
  4

• [bugfix] Stream checkpoint conversion so the full checkpoint is never resident during load

  scope: infrascope: modeltype: bugfix
  #1945·Godmook opened Oct 9, 2026Contributor·3 of 6 tasks·

  ·
  5

• [feat] Add LTX-2.5 fine-tuning to the modular trainer

  scope: docsscope: inferencescope: infrascope: trainingtype: feat
  #1944·alanhuangyoo opened Oct 9, 2026Contributor·4 of 6 tasks·

  ·
  3

• [perf] Cache device-resident RoPE tables for Wan

  scope: infrascope: modeltype: perf
  #1943·Godmook opened Oct 9, 2026Contributor·3 of 6 tasks·

  ·
  5

• [bugfix] Fix FastH3 regional compile recompiling for every prompt length

  scope: attentionscope: infrascope: modeltype: bugfix
  #1942·aryan5v opened Oct 9, 2026Collaborator·

  ·
  1

  8

• [bugfix] Fix stale and uncollectable tests in the top-level tests/ suite

  type: bugfix
  #1939·zyuanbing opened Oct 8, 2026·3 of 4 tasks·

  ·
  1

  5

• [kernel] GB10: forward-consistent dV for Attn-QAT training

  scope: docsscope: kernel
  #1938·fusheng-ji opened Oct 7, 2026Contributor·4 tasks done·

  ·
  6

• [bugfix] Honor PyTorch deterministic mode in FA2 backward

  scope: attentionscope: docsscope: infratype: bugfix
  #1936·fusheng-ji opened Oct 7, 2026Contributor·4 tasks done·

  ·
  6

• [bugfix] Keep the split P@V path for Attn-QAT training on GB10 (SM121)

  scope: docsscope: kerneltype: bugfix
  #1934·fusheng-ji opened Oct 7, 2026Contributor·4 tasks done·

  ·
  7

• [bugfix] NVFP4 fake-quant: mask padding lanes and guard zero-scale groups

  scope: kerneltype: bugfix
  #1933·fusheng-ji opened Oct 7, 2026Contributor·4 tasks done·

  ·
  4

• [misc] Remove DreamX World, HunyuanGameCraft, GLM-Image, LongCat, TurboDiffusion, and Z-Image

  needs-rebasescope: attentionscope: docsscope: inferencescope: infrascope: modelscope: trainingtype: misc
  #1926·SolitaryThinker opened Oct 6, 2026Collaborator·5 of 6 tasks·

  ·
  3

• [DRAFT] Onboarding Qwen-Img 2.1

  needs-rebasescope: docsscope: inferencescope: infrascope: model
  #1918·loaydatrain opened Oct 5, 2026Collaborator·6 tasks·

  ·
  2

• [refactor]: one typed config system with ordered resolution and a read-only runtime config

  needs-rebasescope: attentionscope: datascope: docsscope: inferencescope: infrascope: modelscope: studioscope: trainingtype: refactor
  #1913·Davids048 opened Oct 2, 2026Collaborator·

  ·
  2

• [feat] Add reusable serving cookbook recipes and configuration API

  needs-rebasescope: docsscope: inferencescope: infratype: feat
  #1912·leo2ultimate opened Oct 2, 2026Contributor·3 of 9 tasks·

  ·
  3

• [feat] Add sequence parallel inference support for Kandinsky 5

  scope: docsscope: infrascope: modeltype: feat
  #1905·MichaelY310 opened Sep 30, 2026Contributor·4 of 6 tasks·

  ·
  6

• [new-model] Add Wan2.1-VACE 1.3B/14B controllable video generation

  scope: distributedscope: docsscope: inferencescope: infrascope: modelscope: trainingtype: new-model
  #1903·fusheng-ji opened Sep 30, 2026Contributor·5 of 6 tasks·

  ·
  1

  14

• [bugfix] MiniMax-H3: accumulate AdaLN table gradients in fp32

  needs-rebasescope: infrascope: modeltype: bugfix
  #1894·shaoxiongduan opened Sep 29, 2026Collaborator·3 of 6 tasks·

  ·
  5

• [feat]: add TDM (trajectory distribution matching) training for Wan and MiniMax H3

  needs-rebasescope: attentionscope: docsscope: inferencescope: infrascope: trainingtype: feat
  #1890·macthecadillac opened Sep 26, 2026Collaborator·5 of 6 tasks·

  ·
  1

  8