All pull requests

Pull requests

• GQA: fall back from auto-enabled cuDNN SDPA when the graph can't be built

  #33424·Dante (KafKafrnZ) opened Oct 17, 2026·

  ·
  1

  7

• Preserve score and layout semantics in Python attention fusion

  #33423·Sherman Gaucia Laing (0-Actual) opened Oct 17, 2026·

  ·
  2

• Remove Rust subtree

  #33422·Carson M. (decahedron1) opened Oct 17, 2026Contributor·

  ·
  3

• Preserve reduction and constant semantics in Python LayerNormalization fusion

  #33421·Sherman Gaucia Laing (0-Actual) opened Oct 17, 2026·

  ·
  3

• [CUDA] Tie GQA phase admission to runtime classification

  #33420·Chi Lo (chilo-ms) opened Oct 17, 2026Contributor·

  ·

• Avoid module-load dynamic initializer for OrtEnv::m_

  #33419·Mustapha Jaber (mustjab) opened Oct 16, 2026Contributor·

  ·
  4

  

• [CUDA] QMoE: honour router_weights and allow top-k 9 single-token decode GEMV

  model:nemotron
  #33418·Tianlei Wu (tianleiwu) opened Oct 16, 2026Contributor·2 of 3 tasks·

  ·
  6

• [WebGPU] Fix weight load acceleration so that weight load acceleration off session followed by on works

  #33417·Sushanth Rajasankar (sushraja-msft) opened Oct 16, 2026Contributor·

  ·
  1

• Wasm without Emscripten

  #33416·Mendy Berger (MendyBerger) opened Oct 16, 2026·

  ·
  9

• Fix dynamic Gemm quantization with annotated or shared weights

  #33415·Ge Li (galleonli) opened Oct 16, 2026Contributor·

  ·
  1

  6

• Parallelize gemm operations in xnnpack provider

  #33413·Josh Minor (jishminor) opened Oct 16, 2026·

  ·
  3

• Specify MIT license expression in NuGet packages

  #33411·Federico Fissore (ffissore) opened Oct 16, 2026·

  ·
  3

• [WebGPU] Move copy implementation into data transfer

  #33410·Xiaofei Han (xiaofeihan1) opened Oct 16, 2026Contributor·

  ·
  1

• Avoid quadratic allocation-planner free-buffer scans

  #33407·Rui Ren (rui-ren) opened Oct 16, 2026Contributor·

  ·
  1

• [CUDA plugin EP] Improve NuGet package metadata

  #33405·Tianlei Wu (tianleiwu) opened Oct 16, 2026Contributor·

  ·
  1

• Design post-partition runtime fusion for existing CUDA kernels

  #33404·titaiwangms with CopilotAI opened Oct 15, 2026·

  ·
  1

  3

  

• [float16] Keep float outputs of LayerNormalization, MelWeightMatrix and window ops

  #33402·Kayvan Zahiri (Kayvan-Zahiri) opened Oct 15, 2026·

  ·
  4

• [CUDA] Add source-ordered CUDA IQ2_XXS, Q8_0 and NVFP4 A16 matmul

  #33400·Rui Ren (rui-ren) opened Oct 15, 2026Contributor·

  ·
  4

  

• Fix CPU Range integer boundary counts and final increment overflow

  #33399·Christopher Boseak (cboseak) opened Oct 15, 2026·

  ·
  1

  3

• Validate quantization parameters before QDQ Add, Mul and Concat fusion

  #33397·rookieLiu2018 opened Oct 15, 2026Contributor·

  ·
  1

  2

• [WebGPU] Automate Dawn upgrades and component inventories

  #33396·Yang Gu (gyagp) opened Oct 15, 2026Contributor·

  ·
  1

• [WebGPU] Unroll bounded NVIDIA FlashAttention reductions in WGSL

  #33395·Yang Gu (gyagp) opened Oct 15, 2026Contributor·

  ·
  3

• P0/P1: packed-code foundation and versioned INT3 QMoE contract

  #33394·David Fan (jiafatom) opened Oct 15, 2026Contributor·

  ·
  15

• Optimize Qwen CUDA sparse attention, indexer, and quantized MoE kernels

  #33393·Akshay Sonawane (apsonawane) opened Oct 15, 2026Contributor·

  ·
  9

• Skip MatMul BatchNormalization fusion when the channel axis moves

  #33392·Samridh (samridh10exe) opened Oct 15, 2026·

  ·
  1

  3