All pull requests

Pull requests

• feat: support qwen image pipefusion

  #926·pds-amd opened Oct 14, 2026Contributor·4 of 7 tasks·

  ·

• feat: support wan 2.2 pipefusion

  #924·pds-amd opened Oct 14, 2026Contributor·4 of 7 tasks·

  ·

• fix(cosmos3): reject more than one prompt per pipeline call

  #923·fusheng-ji opened Oct 13, 2026Contributor·4 of 7 tasks·

  ·

• fix(z-image): reject CFG parallelism when the batch is not doubled

  #922·fusheng-ji opened Oct 13, 2026Contributor·4 of 7 tasks·

  ·

• fix(runtime): skip row-patch metadata when latents are not split by rows

  #921·fusheng-ji opened Oct 13, 2026Contributor·4 of 7 tasks·

  ·

• fix(hybrid-schedule): restart the step schedule on every run and drive it from Krea-2

  #920·fusheng-ji opened Oct 13, 2026Contributor·4 of 7 tasks·

  ·

• fix(causal-wan): reject requests longer than the KV cache

  #919·fusheng-ji opened Oct 13, 2026Contributor·4 of 7 tasks·

  ·

• feat: step caching for pipefusion

  #918·pds-amd opened Oct 13, 2026Contributor·4 of 7 tasks·

  ·

• fix(config): require a warmup step for PipeFusion

  #916·fusheng-ji opened Oct 12, 2026Contributor·4 of 7 tasks·

  ·

• fix(runner): forward negative prompt to Z-Image and guidance to HunyuanVideo-1.5

  #915·fusheng-ji opened Oct 12, 2026Contributor·4 of 7 tasks·

  ·

• refactor: PipeFusion runtime modularisation

  #911·pds-amd opened Oct 12, 2026Contributor·4 of 7 tasks·

  ·

• Quantized Fused A2A Attention

  #907·ksikiric opened Oct 8, 2026Contributor·

  ·

• fix(runner): let one layer own DP, output rank and parallel VAE for legacy pipelines

  #902·fusheng-ji opened Oct 7, 2026Contributor·3 of 7 tasks·

  ·
  4

• fix(hunyuan-video): pad sequence-parallel shards when tokens do not divide evenly

  #901·fusheng-ji opened Oct 7, 2026Contributor·4 of 7 tasks·

  ·

• fix(ideogram4): keep sequence-parallel padding out of attention

  #900·fusheng-ji opened Oct 7, 2026Contributor·4 of 7 tasks·

  ·

• fix(cache): store the Ulysses PipeFusion KV cache patch-major

  #899·fusheng-ji opened Oct 7, 2026Contributor·4 of 7 tasks·

  ·
  2

• MiniMax-H3: fuse QK-norm and rotate_half RoPE into one FlyDSL kernel

  #895·avjves opened Oct 7, 2026Collaborator·4 of 5 tasks·

  ·

• fix(hunyuan15): apply the text mask in attention

  #862·fusheng-ji opened Oct 4, 2026Contributor·4 of 7 tasks·

  ·
  1

• fix(sp): honour text masks for batched prompts of different lengths in Z-Image and Qwen-Image

  #856·fusheng-ji opened Oct 4, 2026Contributor·2 of 7 tasks·

  ·
  2

• fix(sp): handle padding across the remaining transformer wrappers

  #851·fusheng-ji opened Oct 4, 2026Contributor·3 of 7 tasks·

  ·
  2

• fix(sp): keep sequence-parallel padding out of attention keys

  #837·fusheng-ji opened Oct 4, 2026Contributor·3 of 7 tasks·

  ·
  4

• fix(flux): pad image tokens correctly under sequence parallelism

  #827·fusheng-ji opened Oct 4, 2026Contributor·3 of 7 tasks·

  ·
  6

• Refactor GEMM quantization support to be more generic

  #807·avjves opened Oct 1, 2026Collaborator·7 tasks·

  ·
  13

• Add FLUX - AMD Quark quantitation example

  #742·amd-xiaoyu12 opened Jul 17, 2026·

  ·
  6