All pull requests

Pull requests

• [WS1][Qwen-Image] attn_out_bias_gemm (python + triton + cuda)

  #459·yanbao1217 opened Oct 2, 2026·

  ·
  1

• fix(ci): restore lint, docs and CPU tests on main

  #458·Jim-jimu opened Oct 1, 2026·

  ·
  1

  1

• fix(cli): honor --profile when bootstrapping the runtime

  #457·Jim-jimu opened Oct 1, 2026·

  ·
  1

• fix(grpo): stabilize group variance for offset rewards

  #455·MichaelCao0 opened Sep 30, 2026·

  ·
  3

• fix(logp): write indexed results into noncontiguous output views

  #454·MichaelCao0 opened Sep 30, 2026·

  ·
  3

• [WS1][kernels] Add fixed-order joint-attention softmax for Qwen-Image

  #453·BPbruce opened Sep 30, 2026·6 of 8 tasks·

  ·
  6

• fix(triton): bound deterministic GEMM reduction grid on CUDA

  Nemotronplatform: triton
  #450·joielab opened Sep 28, 2026·

  ·
  2

• [DSv4][P5-8] Shared Expert TP column/row, SP reduce-scatter, shared-once

  deepseek-P5DSv4
  #449·jyizheng opened Sep 28, 2026·

  ·
  2

• feat(p6): T01 operator contracts, recorded providers and sealed start kit

  deepseek-P6DSv4
  #448·Flink-ddd opened Sep 27, 2026Collaborator·

  ·
  1

• feat(p2): T01 contract, sealed start kit and H100 conformance

  #447·inaniloquentee opened Sep 27, 2026Collaborator·

  ·
  1

• [DSv4][P1-2] fp32_gemm_rms

  deepseek-P1DSv4platform: cudaplatform: triton
  #443·Junyifannnn opened Sep 24, 2026·

  ·
  2

• feat(debug): add generic portable auto capture

  #442·inaniloquentee opened Sep 23, 2026Collaborator·

  ·
  8

• Sm80 linear locuda: add fused SM80 linear_logp backend for A100gp

  #440·QingYUDC opened Sep 22, 2026·

  ·
  7

• [DSv4][P5-4] MXFP8×MXFP4 Routed Expert grouped GEMM: wgmma triton kernel

  deepseek-P5DSv4platform: triton
  #438·Bignonia7 opened Sep 20, 2026·

  ·
  1

• [DSv4][P1-5] Implement CUDA and Triton rmsnorm_residual

  #414·why1te opened Sep 16, 2026·10 tasks done·

  ·
  1

• feat(ascend): Qwen-Image qk_rmsnorm & multi_axis_rope kernels (issue #386)

  Ascendmultimodal
  #411·erfgss opened Sep 13, 2026Contributor·4 of 7 tasks·

  ·
  1

• Qwen image latent pack unpack

  multimodalplatform: cuda
  #410·nodeeeeee opened Sep 12, 2026·

  ·
  1

• style: format Python sources and benchmark results

  #408·maxiaosong1124 opened Sep 12, 2026Collaborator·

  ·
  1

• feat(musa): add native deterministic gemm kernel

  MUSA
  #395·Arlo-mt opened Sep 9, 2026Collaborator·

  ·
  4

  

• Musa support native fused logp kernel

  MUSA
  #392·Arlo-mt opened Sep 8, 2026Collaborator·

  ·
  10

  

• [CI] Fix black formatting on main and pin line-length in pyproject.toml

  #384·Dnoob opened Sep 3, 2026Contributor·

  ·
  1

  4

• [DSv4][P1-S0] Start kit for the P1 work package (mHC + RMSNorm)

  deepseek-P1DSv4platform: cuda
  #383·zhangj1an opened Sep 3, 2026Collaborator·10 tasks·

  ·
  8

• [DSv4][P5] MXFP4 Routed Expert + LoRA + Shared Expert Main Control thread (CUDA & ROCM)

  deepseek-P5DSv4platform: cuda
  #368·KJLdefeated opened Sep 1, 2026Collaborator·

  ·
  20

  

• [WS1] Add batch-invariant h_aggregate kernel

  deepseek-P1DSv4platform: cuda
  #366·nodeeeeee opened Aug 30, 2026·

  ·
  2

• docs: add DCO 1.1 text and contributor sign-off guide

  type: ci-cd
  #359·Zhifu-Liu opened Aug 29, 2026Contributor·

  ·
  8