All pull requests

Pull requests

• Remove thread-pool parallelism assertion in GQA CPU test

  #33342·n1harika opened Oct 13, 2026Contributor·

  ·
  2

• Fix CPU MeanVarianceNormalization epsilon

  #33336·Silu Panda (SiluPanda) opened Oct 13, 2026Contributor·

  ·
  1

  2

• Gate self-hosted Linux builds for fork PRs

  #33332·Akshay Sonawane (apsonawane) opened Oct 13, 2026Contributor·

  ·
  3

• Validate PackedAttention token offsets

  #33330·Akshay Sonawane (apsonawane) opened Oct 13, 2026Contributor·

  ·
  3

• [CUDA] Optimize packed QSA scoring and multi-token verification

  #33328·kunal-vaishnavi opened Oct 13, 2026Contributor·2/2·

  ·

• Fix OpenVINO provider info namespace for Python builds

  #33327·Mustapha Jaber (mustjab) opened Oct 12, 2026Contributor·

  ·
  1

  3

  

• [CUDA] Add opt-in native E4M3 weight decoding to block-FP8 QMoE GEMV

  #33324·Sayan Shaw (sayanshaw24) opened Oct 12, 2026Contributor·

  ·

• [CUDA] Add opt-in native E4M3 block-scale conversion to NVFP4 GEMV

  #33323·Sayan Shaw (sayanshaw24) opened Oct 12, 2026Contributor·

  ·

• [CUDA] Bind fpA/intB profiling to allocator device

  #33322·Sayan Shaw (sayanshaw24) opened Oct 12, 2026Contributor·

  ·
  2

• Simplify 1DS shutdown ordering around the Env singleton

  #33321·bmehta001 opened Oct 12, 2026Contributor·

  ·

• [CUDA] Commit and enforce phase-aware GQA dispatch policies

  #33316·Chi Lo (chilo-ms) opened Oct 12, 2026Contributor·

  ·

• [CUDA] Make MatMulNBits and packed attention L1 estimates route-complete

  #33313·Chi Lo (chilo-ms) opened Oct 12, 2026Contributor·

  ·

• CUDA: fix Nemotron-H relu2, RMSNorm, and activation precision

  #33312·Rui Ren (rui-ren) opened Oct 12, 2026Contributor·

  ·
  12

• [Framework] Honor Level-1 estimates for nodes missing from memory profiles

  #33311·Chi Lo (chilo-ms) opened Oct 12, 2026Contributor·

  ·
  1

• [WebGPU EP] 0.5.0 cherry-picks round 1

  #33308·Edward Chen (edgchen1) opened Oct 12, 2026Contributor·

  ·
  8

• [CoreML EP] Fix "Failed to create model URL from path" for model paths with spaces

  #33307·Alessandro Valli (naturalinteraction) opened Oct 12, 2026·

  ·
  1

  3

• Fix brittle GQA thread pool test assertion

  #33305·Sahil Choudhary (Choudhary-Sahil099) opened Oct 12, 2026·

  ·
  2

• Don't report the Jetson Thor display controller as a GPU in Linux DRM discovery

  #33303·LinCanNerd (LinCanNerd) opened Oct 12, 2026Contributor·

  ·
  4

• Accept an installed Microsoft.GSL 5

  #33302·Joel Kiptoo (Kiptoo-Deus) opened Oct 12, 2026·

  ·
  6

• Add sm_110 (Jetson Thor) to the Linux aarch64 CUDA 13 packaging architectures

  #33301·LinCanNerd (LinCanNerd) opened Oct 12, 2026Contributor·

  ·
  4

• [CUDA] Add opt-in tail-safe INT4 MatMulNBits dispatch for N % 8 == 4, M <= 4

  #33299·Tianlei Wu (tianleiwu) opened Oct 12, 2026Contributor·3 tasks done·

  ·

• Fix session creation with a shared prepacked weights container for fp16 LayerNorm and SkipLayerNorm

  #33295·Yoonseok Kim (yoonseok-kim) opened Oct 12, 2026·

  ·
  7

• Plan If-branch initializer outputs on the If node's device

  #33293·Zhenwei Jin (kylo5aby) opened Oct 12, 2026Contributor·

  ·
  2

• Fix 1DS telemetry deadlock during Windows process termination

  #33291·Jiajia Qin (qjia7) opened Oct 12, 2026Contributor·6 tasks done·

  ·
  4

• [CUDA] Fix false allocation warning during CUDA graph capture in plugin EP

  #33289·Tianlei Wu (tianleiwu) opened Oct 12, 2026Contributor·2 of 3 tasks·

  ·
  2