All pull requests

Pull requests

• [moe_training] Fused dual-axis MXFP8 quantization with correct metadata

  CLA Signedmodule: training
  #4995·sanketpurandare opened Oct 12, 2026·

  ·
  1

• Back out D123598050 to fix qnn_test/test_backward_compatibility numerics (#4994)

  CLA Signedmeta-exportedmodule: pt2e_quant
  #4994·pquochuy opened Oct 12, 2026Contributor·

  ·
  2

• Fix Float8Tensor scale reshaping for 3D-to-2D views

  CLA Signed
  #4992·Pratyush439 opened Oct 11, 2026·

  ·
  1

  6

• Remove redundant weight_dtype validation in MXDynamicActivationMXWeightConfig

  CLA Signed
  #4991·bogoconic1 opened Oct 11, 2026·

  ·
  3

• Support root parameters in safetensors state dictionaries

  CLA Signed
  #4990·AaravMehta-07 opened Oct 10, 2026·

  ·
  3

• Make _triton_permute_bwd handle non-contiguous grads and bad indices

  CLA Signed
  #4989·indianspeedster opened Oct 8, 2026·

  ·
  1

• Use int64 addressing in _triton_permute_bwd_kernel

  CLA Signed
  #4988·indianspeedster opened Oct 8, 2026·

  ·
  1

• Run EP permute tests on ROCm gfx950

  ciflow/rocmCLA Signedmodule: rocm
  #4987·indianspeedster opened Oct 8, 2026·

  ·
  2

• [pat] (4/5) Accelerate coupled pruning for sharded DTensors

  CLA Signedmodule: not user facing
  #4986·lisjin opened Oct 8, 2026Contributor·

  ·
  3

• [pat] (5/5) Add coupled shared-axis pruning

  CLA Signedmodule: not user facing
  #4985·lisjin opened Oct 8, 2026Contributor·

  ·
  3

• [pat] (3/5) Add direct projected-weight optimization mode

  CLA Signedmodule: not user facing
  #4984·lisjin opened Oct 8, 2026Contributor·

  ·
  3

• [pat] (2/5) Add opt-in global group-count score correction

  CLA Signedmodule: not user facing
  #4983·lisjin opened Oct 8, 2026Contributor·

  ·
  3

• [pat] (1/5) Support fixed head_dim in attention groupers

  CLA Signedmodule: not user facing
  #4982·lisjin opened Oct 8, 2026Contributor·

  ·
  3

• Honor PT2E functional-equivalent partition opt-out

  CLA Signed
  #4980·adenzhou1350 opened Oct 8, 2026·

  ·
  1

• Fix MXFP8 AUTO GEMM scale layout in forward and backward

  CLA Signed
  #4979·wsj040303 opened Oct 8, 2026·

  ·
  1

  3

• Clarify FP8 backward operand names

  CLA Signed
  #4978·yncxcw opened Oct 7, 2026·

  ·
  1

• Enable INT8 quantization numerics tests on XPU

  CLA Signed
  #4977·eryk-roch opened Oct 7, 2026·

  ·
  7

• Fix misaligned positional args in _addmm_mx_dispatch activation quantization

  CLA Signedmeta-exportedmodule: not user facing
  #4974·AlbertDachiChen opened Oct 6, 2026·

  ·
  3

• fix: experttensorparallel dynamo tracing crash

  CLA Signed
  #4973·devtyagi3909 opened Oct 5, 2026·

  ·
  1

  1

• fix: benchmark_model CPU timing units

  CLA Signed
  #4972·AdudodlaVarish opened Oct 4, 2026·

  ·
  3

• Skip register_constant for Enum subclasses in register_as_pytree_constant

  CLA Signed
  #4971·vijay-kapse opened Oct 3, 2026·

  ·
  1

  1

• Reject complex input in the pt2e min/max observers instead of silently dropping the imaginary part

  CLA Signed
  #4970·vijay-kapse opened Oct 3, 2026·

  ·
  1

  1

• Fix Int4OpaqueTensor A16W4 crash on mismatched activation dtype

  CLA Signed
  #4969·tritsystem opened Oct 3, 2026·

  ·
  1

• Guard triton imports in prototype.attention.quantization

  CLA Signed
  #4967·JoshBong opened Oct 1, 2026·

  ·
  2

• Fix MaskedLayerNorm with bias=False

  CLA Signed
  #4966·adenzhou1350 opened Sep 30, 2026·

  ·
  1