All pull requests

Pull requests

• [CUDA] Add KDA per-key-dim decay support to GatedDeltaNet across CPU, WebGPU, and CUDA

  #33817·Tianlei Wu (tianleiwu) opened Nov 2, 2026Contributor·3 tasks done·

  ·
  8

• fix(QNN): dont reject cumsum with int64 axis

  #33816·Nicolas Gry (grybouilli) opened Nov 2, 2026Contributor·

  ·
  1

  3

• [Aarch64] Add SGEMV ASIMD kernel

  #33814·Aditya Tewari (aditew01) opened Nov 2, 2026·

  ·
  4

• Use local Detox CLI in React Native CI

  #33813·eserscor opened Nov 2, 2026Collaborator·

  ·
  1

• Fix PEP 668 failures in Linux host CI jobs

  #33812·eserscor opened Nov 2, 2026Collaborator·

  ·
  1

• Enable macOS compiler caching with shared sccache actions

  #33811·eserscor opened Nov 2, 2026Collaborator·

  ·
  1

• [🤗 WebGPU] TopK optimizations when k=1

  #33810·Joshua Lochner (xenova) opened Nov 2, 2026Contributor·

  ·
  2

• webgpu: Apply legacy provider configuration before registration

  #33809·Jiajia Qin (qjia7) opened Nov 2, 2026Contributor·

  ·

• FlashAttention: use subgroup size 32 on Intel

  #33808·Jie Chen (jchen10) opened Nov 2, 2026Contributor·

  ·
  5

• [CUDA] Optimize Qwen3.8-Flash-Next DSA, Indexer, HC, and QMoE kernels for SM12x and 256K context

  model:qwen
  #33802·Tianlei Wu (tianleiwu) opened Nov 1, 2026Contributor·

  ·

• [CUDA] Add opt-in lossless BF16 norm/cast fusion

  #33801·David Fan (jiafatom) opened Nov 1, 2026Contributor·

  ·
  9

• Fuse DQ->MatMulNBits for shared initializers and skip outer-scope weights

  #33800·Vijay Kapse (vijay-kapse) opened Nov 1, 2026Contributor·

  ·
  1

  3

• Bump joi from 17.13.7 to 17.13.9 in /js/react_native/e2e

  dependenciesjavascript
  #33797·dependabot[bot] opened Nov 1, 2026Bot·

  ·
  2

• Fix typos in comments and doc comments

  #33794·RoomWithOutRoof (Jah-yee) opened Nov 1, 2026·

  ·
  4

• [WebGPU] Add int4 and uint4 QuantizeLinear and Reshape

  #33792·adrastogi opened Oct 31, 2026Contributor·

  ·
  1

• [WebGPU] Support int8 data in Gather

  #33790·adrastogi opened Oct 31, 2026Contributor·

  ·
  1

• [WebGPU] Add int32 support to Relu opset 14

  #33789·adrastogi opened Oct 31, 2026Contributor·

  ·
  1

• [CUDA] Add INT2 QMoE CPU/GPU offload support and optimize spin-pool wakeup

  qmoe_offload
  #33787·Tianlei Wu (tianleiwu) opened Oct 31, 2026Contributor·8/8·3 tasks done·

  ·
  24

• [CUDA] Add Tensor-Core BF16 INT2 QMoE GEMV kernel and 3-mask BF16 converter

  model:qwen
  #33786·Tianlei Wu (tianleiwu) opened Oct 31, 2026Contributor·3 tasks done·

  ·
  8

• [CPU/MLAS] Add QMoE shared layout INT2 AVX-512/AVX2 kernels and direct CPU QMoE

  qmoe_offload
  #33785·Tianlei Wu (tianleiwu) opened Oct 31, 2026Contributor·

  ·
  24

• [CPU] Speed up sub-byte DequantizeLinear (int4/uint4/int2/uint2)

  #33784·Bharat Kaurav (TheNetherWatcher) opened Oct 31, 2026·

  ·
  6

• Update Kusto dashboard SDK dependencies to 6.0.0

  #33783·Mustapha Jaber (mustjab) opened Oct 31, 2026Contributor·

  ·
  1

• fix(web): handle opaque origins for bundled file URLs

  #33782·Lenin Mookiah (leninworld) opened Oct 31, 2026Contributor·

  ·
  1

  2

• Align WebGPU CPU/GPU profiling and add decoder optimization tools

  #33781·Yang Gu (gyagp) opened Oct 31, 2026Contributor·

  ·
  1

• Fix spelling typos flagged by misspell linter

  #33780·adrastogi opened Oct 31, 2026Contributor·

  ·
  1