All pull requests

Pull requests

• make import torchao fork-safe on cuda builds

  CLA Signed
  #5088·devanshi-code18 opened Nov 10, 2026·

  ·
  1

• make mx_formats import not initialize cuda

  CLA Signed
  #5087·devanshi-code18 opened Nov 10, 2026·

  ·
  1

  1

• Fix out-of-bounds scales stores in the jagged fp8 scales kernels

  CLA Signed
  #5085·mark14wu opened Nov 9, 2026Contributor·

  ·
  1

• Speed up Intx embedding lookups

  CLA Signed
  #5084·Hang-get opened Nov 9, 2026·

  ·
  1

• Fix default float8 dtypes when reusing cast configs

  #5081·bingdongni opened Nov 9, 2026·

  ·
  2

• [XPU][CI]Upgrade python to 3.11 in XPU CI

  ciflow/xpuCLA Signedxpu
  #5080·zxd1997066 opened Nov 9, 2026Contributor·

  ·
  2

• Pack sparse Squared-ReLU activations saved for backward

  CLA Signed
  #5079·Rohit1mag opened Nov 8, 2026·2 of 3 tasks·

  ·
  4

• Fix SYMMETRIC_NO_CLIPPING_ERR scale for unsigned dtypes

  CLA Signed
  #5078·davrens101 opened Nov 8, 2026·

  ·
  1

• [ROCm] Move PT2E ROCm tests to ciflow/rocm workflow

  ciflow/rocmCLA Signedmodule: rocm
  #5077·akashveramd opened Nov 7, 2026Collaborator·

  ·
  2

  

• Remove host syncs from FakeQuantize.forward on CUDA (bitwise identical)

  CLA Signedmeta-exported
  #5075·cgreenberg opened Nov 6, 2026·

  ·
  2

• fix(utils): report accelerators that have no dedicated branch

  CLA Signed
  #5074·li-lizhe opened Nov 6, 2026·

  ·
  3

• fix(pt2e): do not double-fold conv output into multiple batch norms

  CLA Signed
  #5073·davrens101 opened Nov 6, 2026·

  ·
  1

  1

• Fix BlockwiseQuantLinear dtype handling

  CLA Signed
  #5071·gaurav-redhat opened Nov 6, 2026·

  ·
  1

  1

• CI: Parallelize pytest suites in single-H100 jobs

  CLA Signedmodule: not user facing
  #5069·vkuzo opened Nov 6, 2026Contributor·

  ·
  3

• Handle zero_point=None in quantize_affine

  CLA Signed
  #5066·G26karthik opened Nov 5, 2026·

  ·
  6

• [X86] Fuse external bias adds into linear dynamic fp16 patterns

  CLA Signed
  #5061·ZRICHARD9527 opened Nov 5, 2026·

  ·
  4

• [FSDP] Preserve precomputed FP8 scales across tensor flattening

  CLA Signed
  #5058·0z5a opened Nov 4, 2026·

  ·
  1

• [MXFP8] Fix native token padding for float32

  CLA Signed
  #5057·0z5a opened Nov 4, 2026·

  ·
  1

  1

• [NVFP4] Fix native token padding for 128-row alignment

  CLA Signed
  #5056·0z5a opened Nov 4, 2026·

  ·
  1

  1

• [ROCm] Drive CI ROCm/Python versions from test-infra matrix

  ciflow/rocmCLA Signedmodule: rocm
  #5051·akashveramd opened Nov 4, 2026Collaborator·

  ·
  1

• [mxfp8 moe training] Run MXFP8 grouped mm tests on all SM 10.x GPUs, not only 10.0

  CLA Signed
  #5049·0z5a opened Nov 3, 2026·

  ·
  7

• docs: fix spelling typos in docs, README and setup.py

  ciflow/docsCLA Signed
  #5047·Jah-yee opened Nov 2, 2026·

  ·
  2

• [X86] Use VNNI kernels for DA8W4 linear on AVX512-VNNI CPUs without AMX (30-170x)

  CLA Signed
  #5045·TheNetherWatcher opened Oct 31, 2026·

  ·
  4

• Support weight-only QAT in MXFakeQuantizedLinear

  CLA Signed
  #5044·2021peng opened Oct 31, 2026·

  ·
  1

  14

• Skip unneeded gradients and saved tensors in MX/NVFP4 QAT backward

  CLA Signed
  #5043·Levius-Fubuki opened Oct 31, 2026·

  ·
  3