All pull requests

Pull requests

• Raise ir_version to 10 when MatMulNBitsQuantizer emits INT4/UINT4 tensors

  #33068·Andreas Fehlner (andife) opened Oct 2, 2026Contributor·

  ·
  2

• Align C API tensor element type values with ONNX TensorProto

  #33066·tianleiwu with CopilotAI opened Oct 2, 2026·

  ·
  3

  

• Handle zero-channel ConvTranspose inputs

  #33064·Akshay Sonawane (apsonawane) opened Oct 1, 2026Contributor·

  ·
  3

• Bind external data validation to opened files

  #33063·Akshay Sonawane (apsonawane) opened Oct 1, 2026Contributor·

  ·
  9

• Retain external-data file handles and original model directories

  #33062·Gopalakrishnan Nallasamy (GopalakrishnanN) opened Oct 1, 2026Contributor·

  ·
  1

• Validate Attention prepack dimensions

  #33061·Akshay Sonawane (apsonawane) opened Oct 1, 2026Contributor·

  ·
  2

• Validate NCHWc binary input ranks

  #33060·Akshay Sonawane (apsonawane) opened Oct 1, 2026Contributor·

  ·
  3

• Reject package-controlled optimized model paths

  #33059·Akshay Sonawane (apsonawane) opened Oct 1, 2026Contributor·

  ·
  2

• [WebGPU] Reduce QMoE single-token gate sorting to stable TopK

  #33058·Sayan Shaw (sayanshaw24) opened Oct 1, 2026Contributor·

  ·
  2

• [WebGPU EP] fixes fp16 accumulation bug in gemm kernel

  #33057·Prathik Rao (prathikr) opened Oct 1, 2026Contributor·

  ·
  10

• Handle out-of-range antialiased Resize coordinates

  #33056·Akshay Sonawane (apsonawane) opened Oct 1, 2026Contributor·

  ·
  5

• [WebGPU] Accelerated model loading using Win32 and D3D12

  #33055·Sushanth Rajasankar (sushraja-msft) opened Oct 1, 2026Contributor·

  ·
  5

• Reuse device streams across threads when graph capture is disabled

  #33053·tianleiwu with CopilotAI opened Oct 1, 2026·

  ·
  1

  1

  

• Fix provider bridge initialization failures without exceptions

  #33051·Mustapha Jaber (mustjab) opened Oct 1, 2026Contributor·

  ·
  4

• [MLAS] Cover non-finite and lane-position inputs in the fp16 exp test

  #33050·Vijay Kapse (vijay-kapse) opened Oct 1, 2026·

  ·
  5

• [MLAS] Build the SVE asm generator's input TUs as C++20

  #33047·Vijay Kapse (vijay-kapse) opened Oct 1, 2026·

  ·
  2

• Fix LayerNorm fusion exponent and epsilon validation

  #33046·Raunak (raun6k) opened Oct 1, 2026·

  ·
  1

  2

• [CUDA] Plumb the attention dispatch policy session option into the L1 estimator (reader-only)

  #33042·Chi Lo (chilo-ms) opened Oct 1, 2026Contributor·

  ·

• [MLAS] Exercise ARM64 SymmQgemm Lit (Ld64) and Big kernels in unit tests

  #33041·Yair Ben Avraham (yairbenavraham) opened Oct 1, 2026·

  ·
  3

  

• [MLAS] Fix fp16 overflow/NaN bugs in ARM64 SVE Gelu/Erf and add test coverage

  #33040·Vijay Kapse (vijay-kapse) opened Oct 1, 2026·

  ·
  5

• Enable static CPU offload for FP16 CUDA MoE

  #33039·Xavier Dupré (xadupre) opened Oct 1, 2026Member·

  ·
  30

• fix(qnn/cpu): enable stft on cpu backend of QNN EP

  #33035·Nicolas Gry (grybouilli) opened Oct 1, 2026Contributor·

  ·
  17

• [Python] Add an experimental CUDA 13 unified-wheel packaging pipeline

  #33034·Tianlei Wu (tianleiwu) opened Oct 1, 2026Contributor·

  ·
  12

• Preserve shared half-x nodes during FastGeluFusion

  #33033·Raunak (raun6k) opened Oct 1, 2026·

  ·
  1

  3

• Fuse CPU linear Resize and ArgMax

  #33032· Stern (Hosi121) opened Oct 1, 2026Contributor·

  ·
  5