All pull requests

Pull requests

• Add snapshot-aware Mamba2Verify CUDA contrib operator

  #33969·Rui Ren (rui-ren) opened Nov 7, 2026Contributor·4/4·

  ·

• Fuse a broadcast Mul and ReduceSum into a MatMul

  #33968·daveycodez (daveycodez) opened Nov 7, 2026·

  ·
  6

• [WebGPU] Leave the bodies of CPU loops to the CPU EP

  #33966·daveycodez (daveycodez) opened Nov 7, 2026·

  ·
  6

• [WebGPU] Fail instead of hanging when the device is lost in the browser

  #33965·daveycodez (daveycodez) opened Nov 7, 2026·

  ·
  5

• [web] Upload external weights in bounded chunks in JSPI WebGPU builds

  #33964·daveycodez (daveycodez) opened Nov 7, 2026·

  ·
  6

• Extend MoEPacked for Bolt ReLU-square routing and BF16 IQ2 precision

  #33963·Rui Ren (rui-ren) opened Nov 7, 2026Contributor·3/4·

  ·

• Extend MatMulPacked for native Bolt scaling and BF16 precision

  #33962·Rui Ren (rui-ren) opened Nov 7, 2026Contributor·2/4·

  ·

• Bound MatMulNBits unpacked-path scratch memory by dequantizing B in blocks

  #33960·daveycodez (daveycodez) opened Nov 7, 2026·

  ·
  1

  6

• [WebGPU] Support signed 8-bit and 64-bit unary math

  #33958·Sushanth Rajasankar (sushraja-msft) opened Nov 7, 2026Contributor·

  ·

• [Framework] Report missing or unavailable execution backends

  #33957·Sushanth Rajasankar (sushraja-msft) opened Nov 7, 2026Contributor·

  ·

• [Quant Tool] Fix symmetric percentile histogram counts, growth and threshold

  #33956·realies opened Nov 7, 2026·

  ·
  4

• Add opt-in post-SiLU clipping to MoE operators

  #33954·Baiju Meswani (baijumeswani) opened Nov 7, 2026Contributor·2/2·

  ·
  1

• [CUDA] Add GatherPacked, MatMulPacked, and MoEPacked contrib ops for block-packed weights

  #33953·Tianlei Wu (tianleiwu) opened Nov 6, 2026Contributor·1/4·3 tasks done·

  ·
  71

• [CUDA] Add opt-in single-warp INT4 GEMV for FP16 decode

  #33952·Sunghoon Choi (hanbitmyths) opened Nov 6, 2026Collaborator·

  ·
  4

• Retire unsupported ONNX Runtime training support

  #33951·eserscor opened Nov 6, 2026Contributor·

  ·

• Use CPU-only preparation build for Linux minimal CI

  #33948·eserscor opened Nov 6, 2026Contributor·

  ·
  2

• Fix DivMulFusion integer division semantics

  #33946·diansen33 opened Nov 6, 2026·

  ·
  1

  3

• [CUDA] Tail expert substitution for QMoE offload decode (stacked on #33935)

  qmoe_offload
  #33944·Tianlei Wu (tianleiwu) opened Nov 6, 2026Contributor·2/2·

  ·
  14

• Pin patched cryptography for dashboard dependencies

  #33942·eserscor opened Nov 6, 2026Contributor·

  ·
  1

• [CUDA] Consolidate source-exact BF16 decoding into existing contrib operators

  #33940·Rui Ren (rui-ren) opened Nov 6, 2026Contributor·

  ·
  7

• Speed up session creation

  #33939·Christian Bourjau (cbourjau) opened Nov 6, 2026Contributor·

  ·
  6

• Kleidiops static int8

  #33938·Damien Dooley (damdoo001-arm) opened Nov 6, 2026Contributor·

  ·
  6

• [WebGPU] Restore direct external initializer loading for the static plugin

  #33937·Xiaofei Han (xiaofeihan1) opened Nov 6, 2026Contributor·

  ·
  7

• [CUDA] Improve QMoE offload decode throughput via AVX2 spill elimination and working-set retention

  qmoe_offload
  #33935·Tianlei Wu (tianleiwu) opened Nov 6, 2026Contributor·1/2·

  ·
  11

• [WebGPU] Add scoped code review skill

  #33933·Yang Gu (gyagp) opened Nov 6, 2026Contributor·

  ·
  7