All pull requests

Pull requests

• [PyTorch] Remove EP handle cache and reconstructing handle on the fly

  #3760·phu0ngng opened Oct 27, 2026Collaborator·13 tasks·

  ·
  1

• Remove dummy weight gradients from fused accumulation

  #3759·vthumbe1503 opened Oct 27, 2026Collaborator·5 tasks done·

  ·

• MxFP8 gradient cross entropy option

  #3758·sraman-rgb opened Oct 27, 2026Collaborator·13 tasks·

  ·
  1

  

• [PyTorch] Fuse BHSD permute and GEMM swizzle of MXFP8 attention scales

  #3757·sraman-rgb opened Oct 27, 2026Collaborator·13 tasks·

  ·
  1

  

• [PyTorch] Fuse NVFP4 global scales and GEMM alpha in the grouped MLP

  #3756·sraman-rgb opened Oct 27, 2026Collaborator·13 tasks·

  ·
  1

• [PyTorch] Optionally derive FC1 dy NVFP4 amaxes from the dGLU epilogue

  #3755·sraman-rgb opened Oct 27, 2026Collaborator·13 tasks·

  ·
  2

• [PyTorch] Cache GEMM-swizzled scales of all-gathered NVFP4 weights

  #3754·sraman-rgb opened Oct 27, 2026Collaborator·13 tasks·

  ·
  3

• [Common][PyTorch] Add NVFP4 EP dispatch and combine backward

  #3753·sraman-rgb opened Oct 27, 2026Collaborator·13 tasks·

  ·
  3

• Reduce CPU overhead in grouped-tensor GroupedLinear and MLP

  #3752·vthumbe1503 opened Oct 27, 2026Collaborator·2 tasks done·

  ·
  1

• [Common, PyTorch] Write both NVFP4 amax outputs in the reduction

  community-contribution
  #3750·kimchi-developer opened Oct 27, 2026·7 of 13 tasks·

  ·
  2

• [Common, PyTorch] Fuse NVFP4 scale swizzling on SM120 and SM121

  community-contribution
  #3749·kimchi-developer opened Oct 27, 2026·6 of 13 tasks·

  ·
  1

• Fix LayerNormLinear dummy gradients for sharded weights

  #3748·zhongbozhu opened Oct 27, 2026Collaborator·13 tasks·

  ·
  4

• [JAX] Control MeshResource at API call granularity for Dense APIs

  #3747·jberchtold-nvidia opened Oct 26, 2026Collaborator·8 of 13 tasks·

  ·

• [Common] Make CI fail if CuTeDSL kernels with supported configs are not used

  #3746·kainzhong opened Oct 26, 2026Collaborator·9 of 13 tasks·

  ·
  9

• [PyTorch]Validate grouped split alignment on device

  #3745·vthumbe1503 opened Oct 26, 2026Collaborator·8 of 9 tasks·

  ·
  3

• Don't rely on cuda.core for compute capability check

  community-contribution
  #3743·CarlosGomes98 opened Oct 26, 2026Contributor·13 tasks·

  ·
  2

  

• [Common] Optimize NVFP4 RHT cast-fusion kernels

  community-contribution
  #3740·CalebDu opened Oct 26, 2026·4 of 13 tasks·

  ·
  2

• [PyTorch/JAX] Add NVTE_CUDNN_ATTN as an alias of NVTE_FUSED_ATTN

  #3738·YangXu1990uiuc opened Oct 26, 2026·7 of 13 tasks·

  ·
  2

• [Common] Honor MXFP8 scale strides in specialized casts

  community-contribution
  #3737·bvolpato opened Oct 25, 2026·5 of 13 tasks·

  ·
  5

  

• [JAX] Use local MoE aux-loss computation

  community-contribution
  #3731·dinodeep opened Oct 22, 2026Contributor·8 of 13 tasks·

  ·
  1

• [Common] CGEMM stream ID logging

  #3727·jberchtold-nvidia opened Oct 22, 2026Collaborator·8 of 13 tasks·

  ·
  9

• [PyTorch] Add SiLUGate fusible op

  community-contribution
  #3726·VikingDeng opened Oct 21, 2026·7 of 13 tasks·

  ·
  6

• [JAX] Fix EP=1 case in TE MoEBlock

  #3725·jberchtold-nvidia opened Oct 21, 2026Collaborator·13 tasks·

  ·
  3

• 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

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

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

  ·
  1