All pull requests

Pull requests

• Place pt2e input observers after in-place writes through views

  CLA Signed
  #5046·msluszniak opened Oct 31, 2026·

  ·
  1

  1

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

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

  ·
  4

• feat(qat): support weight-only MX QAT in MXFakeQuantizedLinear

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

  ·
  1

  1

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

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

  ·
  3

• fix(pt2e): insert activation input observers before consumer node

  CLA Signed
  #5041·devtyagi3909 opened Oct 30, 2026·

  ·
  1

  1

• Fix Triton 3.9 compile errors in FP8 scale indexing

  ciflow/rocmciflow/rocm-mi300CLA Signed
  #5036·jagadish-amd opened Oct 30, 2026Contributor·

  ·
  6

• Keep asymmetric int8 correction in higher precision

  CLA Signed
  #5035·adityuhkapoor opened Oct 29, 2026·

  ·
  1

  4

• switch CI to pytorch 2.15

  ciflow/rocmciflow/rocm-mi300ciflow/xpuciflow/4xh100CLA Signedmodule: not user facing
  #5033·vkuzo opened Oct 29, 2026Contributor·

  ·
  3

• [ROCm][CI] Update ROCm version to 10.1

  ciflow/rocmCLA Signedmodule: rocm
  #5032·akashveramd opened Oct 28, 2026Collaborator·

  ·
  7

• Add jeffdaily as ROCm co-owner of low-precision training paths

  ciflow/rocmCLA Signedmodule: rocm
  #5030·jeffdaily opened Oct 28, 2026Contributor·

  ·
  4

• Raise a clear error when Int8StaticActivationInt8WeightConfig has no act_quant_scale

  CLA Signed
  #5025·Sonica-B opened Oct 25, 2026·

  ·
  1

• [pat] Add make_m_by_n_mask helper for explicit N:M masks

  CLA Signed
  #5024·VikingDeng opened Oct 23, 2026·

  ·
  6

• Allowlist QAT modules for config_from_dict round-trip

  CLA Signedmodule: qat
  #5023·Cross2pro opened Oct 23, 2026·

  ·
  1

  8

• [INT4][WOQ] Make int4 weight-only F.linear dispatch handle non-contiguous activations on Int4PlainInt32Tensor and Int4TilePackedTo4dTensor

  CLA Signed
  #5018·hoshibara opened Oct 22, 2026·

  ·
  1

  1

• feat(safetensors): add safetensors serialization and loading for static INT8 + SmoothQuant

  CLA Signed
  #5016·piyushumate opened Oct 20, 2026·

  ·
  1

• feat(smoothquant): support single-pass static smoothquant via analytical scale

  CLA Signed
  #5015·piyushumate opened Oct 20, 2026·

  ·
  1

• feat(smoothquant): default use_running_absmax to True and warn on False

  CLA Signed
  #5014·piyushumate opened Oct 20, 2026·

  ·
  1

• fix(smoothquant): fix static convert crash and 0-d scale rank mismatch

  CLA Signed
  #5013·piyushumate opened Oct 20, 2026·

  ·
  3

• Fix INT8 GPTQ scales when rows span multiple blocks

  CLA Signed
  #5012·Lorenzo-Mazza opened Oct 17, 2026·

  ·
  3

• [CI] Use Python 3.11 for nightly regression and wheel builds

  CLA Signed
  #5011·jontb opened Oct 17, 2026·

  ·
  5

• Honor keyword arguments in Int8Tensor linear

  CLA Signed
  #5010·mottopanikeiku opened Oct 17, 2026·

  ·
  1

• Fix Int8Tensor scale indexing when selecting rows

  CLA Signed
  #5009·mottopanikeiku opened Oct 17, 2026·

  ·
  1

• Honor exact parameter configs before parameter regexes

  CLA Signed
  #5008·mottopanikeiku opened Oct 17, 2026·

  ·
  1

• Fix bitpacking default unpack dimension

  CLA Signed
  #5004·Cross2pro opened Oct 16, 2026·

  ·
  1

  3

• Fix NaN from all-zero input in two-level scaling (#5002)

  CLA Signedmeta-exportedmodule: inference
  #5002·markivfb opened Oct 15, 2026·

  ·
  7