All pull requests

Pull requests

• Enable torch._grouped_mm for Int4PlainInt32Tensor

  ciflow/xpuCLA Signedmodule: inferencexpu
  #4934·xiaowangintel opened Sep 22, 2026·

  ·
  2

• Add generic MX Conv2d QAT support

  CLA Signed
  #4933·Sebastian-Larsson opened Sep 22, 2026·

  ·
  1

• [moe_training] Replace MXFP8 dim-1 (32x1) quantization kernel with a faster one

  CLA Signed
  #4931·protonu opened Sep 21, 2026·

  ·
  1

• [XPU][CI] Add timeout for Intel GPU CI

  ciflow/xpuCLA Signedxpu
  #4928·zxd1997066 opened Sep 20, 2026·

  ·
  1

• [a16-support] Add asymmetric A16 fake quantization

  CLA Signedmeta-exportedmodule: qat
  #4924·insoochung opened Sep 19, 2026·

  ·
  3

• [a16-support] Preserve dynamic signed W4 scale gradients

  CLA Signedmeta-exportedmodule: qat
  #4923·insoochung opened Sep 19, 2026·

  ·
  2

• [prototype][QAT] Add stateless MX fake quantization and grouped MM

  CLA Signed
  #4917·elfiegg opened Sep 18, 2026·

  ·
  1

• [moe_training] Replace MXFP8 dim-0 (1x32) quantization kernel with a faster one

  CLA Signed
  #4916·protonu opened Sep 18, 2026·

  ·
  1

• Fix the Adam step counter saturating under a BF16 default dtype

  CLA Signedmodule: optimizer
  #4915·RachelWanggg opened Sep 18, 2026·

  ·
  6

• Reject complex input in the pt2e min/max observers, scriptably

  CLA Signed
  #4914·rohday opened Sep 18, 2026·

  ·
  1

  6

• Fix AdamW8bit state precision selection for small DTensor shards

  CLA Signed
  #4913·gss10282023 opened Sep 18, 2026·

  ·
  1

  3

• Vectorize complete Arm NEON quantization blocks

  CLA Signed
  #4905·JacobSzwejbka opened Sep 16, 2026·

  ·
  1

• Fix FqnToConfig param regex None overridden by later patterns

  CLA Signedmodule: inference
  #4902·gaurav-redhat opened Sep 15, 2026·

  ·
  2

• fix(quant_api): prevent IndexError in _fqn_to_config_handler when multiple parameter FQNs are None (#4818)

  CLA Signed
  #4901·somuai opened Sep 15, 2026·

  ·
  1

  2

• Fix output addressing race in the jagged rowwise fp8 scales kernel

  CLA Signed
  #4900·mark14wu opened Sep 14, 2026·

  ·
  1

• [moe_training] Add a fused 1x32 + 32x1 MXFP8 quantization kernel

  CLA Signed
  #4895·protonu opened Sep 14, 2026·

  ·
  4

• docs: add Simplified Chinese translation (README.zh.md)

  CLA Signed
  #4892·JasonYeYuhe opened Sep 13, 2026·

  ·
  2

• Make float8 autograd Function allow_in_graph opt-in via config

  CLA Signedmodule: training
  #4879·vkuzo opened Sep 10, 2026·

  ·
  7

• Fix PT2E Quick Start guide example so it runs as documented

  CLA Signed
  #4878·Sanjays2402 opened Sep 10, 2026·

  ·
  1

  1

• Fix scalar collapse for all-ones block_size in tinygemm / dont_preserve_zero qparams

  CLA Signed
  #4871·TusharND12 opened Sep 8, 2026·

  ·
  3

• fix(float8): handle aten.abs and amax-path ops in Float8TrainingTensor dispatch

  CLA Signedmodule: training
  #4870·somuai opened Sep 7, 2026·

  ·
  1

  7

• Make Int8 and Float8 weight-only reapplication safe

  CLA Signed
  #4869·yashb98 opened Sep 6, 2026·

  ·
  1

• Add generalized int1-int8 weight-only MPS quantization with simdgroup GEMM

  CLA Signed
  #4868·mspinelli opened Sep 4, 2026·

  ·
  1

• Warn when Int4WeightOnlyConfig uses the PLAIN packing format

  CLA Signed
  #4867·fletchers16 opened Sep 4, 2026·

  ·
  1

• [nvfp4 training][rl] Add the four-over-six grouped GEMM and dispatcher wiring

  CLA Signed
  #4853·wolfcomos opened Aug 31, 2026·

  ·
  1