All pull requests

Pull requests

• [PyTorch] Restore in-place softmax backward for eager attention

  community-contribution
  #3815·janbernloehr opened Nov 4, 2026Contributor·7 of 13 tasks·

  ·
  1

  2

• Reduce NVFP4 grouped GEMM dispatch and output scaling overhead

  community-contribution
  #3813·milakov opened Nov 4, 2026·6 of 13 tasks·

  ·
  1

• Fix CPU QuantizedTensor dequantize RecursionError

  community-contribution
  #3812·siddhant-rajhans opened Nov 4, 2026·

  ·
  1

  5

• Support different C and D tensors in nvte_cublas_gemm_v2

  #3811·ksivaman opened Nov 4, 2026Member·8 of 13 tasks·

  ·
  1

  5

• [PyTorch] Add MXFP4 weight x MXFP8 activation (W4A8) support for Linear modules

  community-contribution
  #3810·cael-ling opened Nov 4, 2026Contributor·1 of 13 tasks·

  ·
  3

• CuTeDSL NVFP4 1D quantize kernels

  #3809·tdophung opened Nov 4, 2026Collaborator·7 of 13 tasks·

  ·

• Quantize NVFP4 1D from registers instead of through TMA staging

  #3808·tdophung opened Nov 4, 2026Collaborator·7 of 13 tasks·

  ·

• [Common] Trim mhc configs to reduce memory usage and time to autotune

  #3807·kainzhong opened Nov 3, 2026Collaborator·8 of 13 tasks·

  ·
  3

• Add logit-space bias to the fused sigmoid router

  community-contribution
  #3805·philipcmonk opened Nov 3, 2026Contributor·7 of 13 tasks·

  ·
  2

• [JAX] Reduce test runtime of mHC tests by partially disabling autotuning

  #3803·jberchtold-nvidia opened Nov 3, 2026Collaborator·8 of 13 tasks·

  ·
  1

• [PyTorch] Test FusedAdam with an empty param in the last launch slot

  community-contribution
  #3802·andrewliu2001 opened Nov 3, 2026Contributor·6 of 13 tasks·

  ·
  1

  3

• Update TE to support Python 3.14 in CI environments

  #3801·fheinecke opened Nov 3, 2026Collaborator·1 of 13 tasks·

  ·
  2

• Reduce NVFP4 activation quantization overhead

  community-contribution
  #3800·milakov opened Nov 3, 2026·7 of 8 tasks·

  ·
  1

• Reduce NVFP4 GEMM scale metadata overhead

  community-contribution
  #3799·milakov opened Nov 3, 2026·7 of 8 tasks·

  ·
  1

• [PyTorch] Break circular import between quantization and tensor packages

  community-contribution
  #3798·muutot opened Nov 3, 2026Contributor·5 of 8 tasks·

  ·
  1

• [CI] Disable TF32 for the L0 PyTorch numerics tests

  community-contribution
  #3797·bvolpato opened Nov 3, 2026·5 of 13 tasks·

  ·
  1

• [PyTorch] Keep Flex Attention score_mod graphs on the cuDNN backend

  community-contribution
  #3796·bvolpato opened Nov 3, 2026·5 of 13 tasks·

  ·
  1

• [Common] Fix CUDA library loading for pip-installed frameworks

  community-contribution
  #3795·bvolpato opened Nov 3, 2026·7 of 13 tasks·

  ·
  9

• Raise minimum required FA version for FA4

  #3794·ksivaman opened Nov 3, 2026Member·8 of 13 tasks·

  ·
  4

• [JAX] Return EP FFI errors instead of throwing through XLA

  #3792·phu0ngng opened Nov 2, 2026Collaborator·7 of 13 tasks·

  ·
  2

• [PyTorch] Test sequence-parallel Linear with 3D inputs and small micro-batches

  community-contribution
  #3791·0z5a opened Nov 1, 2026Contributor·6 of 13 tasks·

  ·
  1

• [PyTorch] Add export_components to MXFP8 tensors

  community-contribution
  #3790·bhargavchintam opened Oct 31, 2026·8 of 13 tasks·

  ·
  1

  3

• Update minimum framework versions

  #3788·ksivaman opened Oct 30, 2026Member·13 tasks·

  ·
  3

• [PyTorch] All-reduce MXFP8 E8M0 scales instead of amaxes when casting master weights

  community-contribution
  #3787·rapatel opened Oct 30, 2026Contributor·7 of 13 tasks·

  ·
  1

• [Common] MXFP8 partial amax and cast kernels launch only the row tiles of the shard

  community-contribution
  #3786·rapatel opened Oct 30, 2026Contributor·7 of 13 tasks·

  ·
  6