All pull requests

Pull requests

• [Common] Add benchmark scripts for CuTeDSL

  #3688·kainzhong opened Oct 15, 2026Collaborator·13 tasks·

  ·

• [PyTorch] Reduce THD CP P2P input preparation overhead

  community-contribution
  #3687·yashaswikarnati opened Oct 14, 2026Contributor·

  ·

• [Common] Initialize unused THD half-buffer rows

  community-contribution
  #3686·yashaswikarnati opened Oct 14, 2026Contributor·

  ·
  1

• [PyTorch] Fix full-graph compilation of linear attention

  #3685·pggPL opened Oct 14, 2026Collaborator·6 tasks done·

  ·
  3

• Optimize context parallel attention layouts and gradient accumulation

  community-contribution
  #3683·YangXu1990uiuc opened Oct 14, 2026·4 of 7 tasks·

  ·

• [Common] Optimized kernels for swizzling and remove obsolete narrow scaled kernels

  #3682·vthumbe1503 opened Oct 13, 2026Collaborator·13 tasks·

  ·
  4

• [JAX] Disable NCCL communicator splitting for attention CI

  attentionjax
  #3681·KshitijLakhani opened Oct 13, 2026Collaborator·13 tasks·

  ·
  1

  

• Integrating TE Triton mHC kernels into JAX

  community-contribution
  #3680·dinodeep opened Oct 13, 2026Contributor·8 of 13 tasks·

  ·
  24

• Fuse segmented MXFP8 quantization and projection gradient scaling

  community-contribution
  #3679·layalir opened Oct 13, 2026Contributor·

  ·

• [PyTorch] Skip torch.compile tests when required PyTorch APIs are unavailable

  #3678·pggPL opened Oct 13, 2026Collaborator·4 of 13 tasks·

  ·
  1

• Fused QKV ROPE - CS FP8 quantization

  community-contribution
  #3675·vasunvidia opened Oct 12, 2026Collaborator·13 tasks·

  ·
  7

• [JAX] EP with Fused Prepare and Dispatch (NVL Only)

  #3674·phu0ngng opened Oct 12, 2026Collaborator·8 of 13 tasks·

  ·
  15

• [Common][PyTorch] Fuse the per-head Q/K RMSNorm into the single-QKV RoPE kernel

  attentioncommunity-contribution
  #3673·filaretov opened Oct 12, 2026·

  ·
  2

• [Common] Forward the RHT matrix in NVFP4 quantize dispatch

  community-contribution
  #3669·janbernloehr opened Oct 12, 2026Contributor·4 of 13 tasks·

  ·
  1

• Fix archive links across documentation publishing locations

  community-contribution
  #3665·emcallister20 opened Oct 9, 2026·4 of 7 tasks·

  ·
  2

• [Common][PyTorch] Add row-scaled NVFP4 support to grouped-tensor GEMM

  community-contribution
  #3661·cael-ling opened Oct 9, 2026Contributor·2 of 13 tasks·

  ·
  10

• GroupedTensor Machinery for MXFP8 Dispatch input passed from Megatron

  #3660·vthumbe1503 opened Oct 9, 2026Collaborator·13 tasks·

  ·
  3

• Preserve router probability precision in fused grouped MLP

  community-contribution
  #3653·mkhona-nvidia opened Oct 8, 2026·13 tasks·

  ·
  2

• [PyTorch] Support FP64 second moment in FusedAdam

  community-contribution
  #3652·deepakn94 opened Oct 8, 2026Contributor·5 of 13 tasks·

  ·
  2

• Guard zero-gradient blocks in the default FP8 block-scaling recipe

  community-contribution
  #3651·HollowMan6 opened Oct 8, 2026Contributor·8 of 13 tasks·

  ·
  6

  

• Fuse the residual and bias adds into their neighbours: proj bias in the MXFP8 GEMM epilogue (with BGRAD in the dgrad epi

  community-contribution
  #3650·denys-fridman opened Oct 8, 2026·

  ·

• [Common][PyTorch] Add NVFP4 row-scaled grouped cast with graph-safe path

  community-contribution
  #3649·cael-ling opened Oct 8, 2026Contributor·2 of 13 tasks·

  ·
  5

• [PyTorch] Add BF16 fused grouped MLP support

  community-contribution
  #3646·tingyangk opened Oct 7, 2026Contributor·7 of 12 tasks·

  ·

• perf: Improvements to nvfp4 4over6 cuda kernel

  community-contribution
  #3643·CarlosGomes98 opened Oct 7, 2026Contributor·13 tasks·

  ·
  2

• [Common] More CuTeDSL MXFP8 specialized kernels

  #3642·kainzhong opened Oct 7, 2026Collaborator·8 of 13 tasks·

  ·
  8