All pull requests

Pull requests

• [CUDA] Add state_status output to PackedSparseAttentionIndexer

  #33383·Tianlei Wu (tianleiwu) opened Oct 15, 2026Contributor·3 tasks done·

  ·

• Fix ConvMulFusion with non-finite scales

  #33382·King Star (jstar0) opened Oct 14, 2026·

  ·
  1

  3

• Fix arena configuration handling and optimize CUDA GEMV and prefill

  #33381·Akshay Sonawane (apsonawane) opened Oct 14, 2026Contributor·

  ·
  3

• Skip QLinearWhere fusion for per-axis quantization parameters

  #33380·Samridh (samridh10exe) opened Oct 14, 2026·

  ·
  1

  3

• Fix GatherToSliceFusion with negative Range limit

  #33377·Sahil Choudhary (Choudhary-Sahil099) opened Oct 14, 2026·

  ·
  1

  4

• Fix/gate cuda ci registry auth

  #33376·Akshay Sonawane (apsonawane) opened Oct 14, 2026Contributor·

  ·
  2

• Validate fp16 MatMulNBits prepack inputs

  #33375·Akshay Sonawane (apsonawane) opened Oct 14, 2026Contributor·

  ·
  1

• Experiment with INT8 activations for INT2 QMoE CUDA decode

  #33374·David Fan (jiafatom) opened Oct 14, 2026Contributor·

  ·
  12

• [CUDA] Fix CUDA architecture gating for SM12x, family/arch-specific targets and -virtual entries

  #33373·Connor Baker (ConnorBaker) opened Oct 14, 2026·

  ·
  2

• [docs] Design region-aware CPU/GPU placement under a memory budget

  #33372·Chi Lo (chilo-ms) opened Oct 14, 2026Contributor·

  ·
  1

• [WebGPU] Fuse rank-32 LoRA and QLoRA projections with MatMulNBits

  #33370·Jiajia Qin (qjia7) opened Oct 14, 2026Contributor·

  ·
  1

• Add sm_87 for Jetson Orin CUDA 13 packaging

  #33369·Alexander Kluge (alexk201) opened Oct 14, 2026·

  ·
  1

  3

• Auto-generated baselines by 1ES Pipeline Templates

  #33368·microsoft-github-policy-service[bot] opened Oct 14, 2026Bot·

  ·
  2

• [CUDA] Optimize dense prefill for NVFP4 hybrid QSA/GDN models

  #33367·Tianlei Wu (tianleiwu) opened Oct 14, 2026Contributor·

  ·
  2

• [CUDA] Add fused decode routing, single-pass finalize, and PDL overlap for NVFP4 QMoE

  #33366·Tianlei Wu (tianleiwu) opened Oct 14, 2026Contributor·

  ·
  2

• [WebGPU] Stabilize LayerNormalization for constant rows

  #33365·Yang Gu (gyagp) opened Oct 14, 2026Contributor·

  ·
  1

  3

• Preserve distinct epsilons in GQA pre-norm fusion

  #33363·Yang Gu (gyagp) opened Oct 14, 2026Contributor·

  ·
  1

  4

• [WebGPU] Capture GPU buffer copies for graph replay

  #33362·Xiaofei Han (xiaofeihan1) opened Oct 14, 2026Contributor·

  ·
  1

  1

• Add Python 3.15 to Python packaging

  #33359·Danny On The Air (dannyota) opened Oct 14, 2026Contributor·

  ·
  6

• [CUDA] Cosmos3 - Fix variance cancellation in fused BERT layer normalization

  #33357·Rui Ren (rui-ren) opened Oct 14, 2026Contributor·

  ·
  3

• [WebGPU] Refactor FlashAttention into eligibility and execution helpers

  #33356·Jianhui Dai (daijh) opened Oct 14, 2026Contributor·

  ·
  5

• [WebGPU] Fix GQA local-window test parallelism assertion on hybrid CPUs

  #33355·Jianhui Dai (daijh) opened Oct 14, 2026Contributor·

  ·
  2

• [CUDA] Add decode GEMV kernel for K-packed NVFP4 QMoE

  #33354·Tianlei Wu (tianleiwu) opened Oct 14, 2026Contributor·3 tasks done·

  ·
  2

• Fix CPU Shrink precision loss and integer boundary handling

  #33353·Mohit Nirmal (Mohit-25-tech) opened Oct 13, 2026Contributor·

  ·
  1

  6

• [WebGPU] Extend FP16 MatMul GEMV fast path to M=2–8

  #33351·aciddelgado opened Oct 13, 2026Contributor·

  ·
  8