All pull requests

Pull requests

• Keep the FP32 MXFP8 RMSNorm intermediate at 2048x7168 on cuDNN 9.26

  community-contribution
  #3723·liayan opened Oct 21, 2026·5 of 13 tasks·

  ·
  1

• [PyTorch] FP8BS grouped GEMM test reference fix

  #3722·denera opened Oct 21, 2026Collaborator·8 of 13 tasks·

  ·
  4

  

• Run minimum-versions build on pushes to main to warm sccache

  community-contribution
  #3721·agarg21 opened Oct 21, 2026·4 of 13 tasks·

  ·
  1

• [Common] Use FP8 fast accumulation in the cuBLASMp comm+GEMM overlap

  community-contribution
  #3719·almogsegal opened Oct 21, 2026Contributor·13 tasks·

  ·
  1

• [PyTorch] Check saved tensors against comm+GEMM overlap buffers

  #3718·denera opened Oct 20, 2026Collaborator·8 of 13 tasks·

  ·
  1

  

• Rely on Recipe object for fusion

  #3717·vthumbe1503 opened Oct 20, 2026Collaborator·13 tasks·

  ·
  10

• [PyTorch] Keep FlashAttention 2 off for head_dim > 192 dropout backward on sm12x

  attentioncommunity-contribution
  #3715·0z5a opened Oct 20, 2026Contributor·8 of 13 tasks·

  ·
  1

  1

• [PyTorch] Allow FlashAttention 2 head_dim > 192 on sm103

  attentioncommunity-contribution
  #3714·0z5a opened Oct 20, 2026Contributor·8 of 13 tasks·

  ·
  1

• [PyTorch] Keep autocast precision state per thread

  community-contribution
  #3713·yashaswikarnati opened Oct 20, 2026Contributor·7 of 13 tasks·

  ·
  2

• [JAX] Exposing cuDNN Indexer, Compressor, and DSA Kernels

  attentioncommunity-contributionjax
  #3712·dinodeep opened Oct 19, 2026Contributor·1 of 13 tasks·

  ·

• [Common] Support compact unswizzled scales in MXFP8 quantization

  community-contribution
  #3710·XFDG opened Oct 19, 2026·7 of 13 tasks·

  ·
  1

  4

• [Common] Use .shared::cta for 2D TMA loads on sm_12x

  community-contribution
  #3708·kimchi-developer opened Oct 18, 2026·6 of 13 tasks·

  ·

• [Docs] Add compute capability 11.0 to the support matrix

  community-contribution
  #3707·0z5a opened Oct 17, 2026Contributor·5 of 13 tasks·

  ·

• [Common] Use software stochastic rounding in the RHT cast fusion kernels

  community-contribution
  #3706·0z5a opened Oct 17, 2026Contributor·7 of 13 tasks·

  ·
  1

• [PyT] [common] Add cuDNN OSS forward dispatch for context parallel attention

  attentioncommunity-contribution
  #3705·YangXu1990uiuc opened Oct 17, 2026·

  ·

• [Common] Give the grouped TMA descriptor updater internal linkage

  community-contribution
  #3704·ravimajeti opened Oct 17, 2026Contributor·6 of 13 tasks·

  ·
  1

  1

• Skip unsupported grouped MLP dprob determinism checks

  community-contribution
  #3702·yeliu-oss opened Oct 17, 2026·5 of 13 tasks·

  ·

• Support GroupedTensor autograd views and stream tracking

  2.21
  #3701·vthumbe1503 opened Oct 16, 2026Collaborator·13 tasks·

  ·
  3

• [Common][Attention][cuDNN] Restore bucketed attention plan reuse

  2.21attention
  #3700·sudhakarsingh27 opened Oct 16, 2026Member·

  ·
  4

  

• Add cuDNN OSS backward dispatch for context parallel attention

  attentioncommunity-contribution
  #3699·YangXu1990uiuc opened Oct 16, 2026·5 of 8 tasks·

  ·

• [PyTorch][CP] Make THD CP helper kernels to be shared memory size aware

  2.21attention
  #3698·sudhakarsingh27 opened Oct 16, 2026Member·7 of 13 tasks·

  ·
  4

  

• Fix EP sharding gather logic so activations stay sharded

  community-contribution
  #3697·dinodeep opened Oct 16, 2026Contributor·8 of 13 tasks·

  ·
  2

• [PyTorch] Add the discovered NCCL and cuDNN include dirs to the PyTorch extension build

  community-contribution
  #3696·0z5a opened Oct 16, 2026Contributor·6 of 13 tasks·

  ·
  1

• [DRAFT] Switch to Python binding for cuDNN frontend

  #3693·vcherepanov-nv opened Oct 16, 2026Collaborator·5 of 13 tasks·

  ·
  9

• [PyTorch] Add torch.compile support for fused MoE routing

  #3691·pggPL opened Oct 15, 2026Collaborator·2 tasks done·

  ·
  2