All pull requests

Pull requests

• Fix MatMulNBits CompInt8 on x64 returning zeros with runtime scales or zero points

  #34224·Qiming Bao (14H034160212) opened Nov 14, 2026·

  ·
  1

  3

• [CUDA] Add a session option for the QMoE GEMV row limit

  #34223·BBaditi opened Nov 14, 2026Contributor·

  ·

• onnxruntime-web - Add WebNN browser CI coverage

  #34222·Edward Chen (edgchen1) opened Nov 14, 2026Contributor·

  ·

• Fix PropagateCastOps when a removed Cast feeds the same node input twice

  #34221·Qiming Bao (14H034160212) opened Nov 14, 2026·

  ·
  1

  2

• [WebGPU] Fix GatedDeltaNet shader key collision

  #34220·Edward Chen (edgchen1) opened Nov 14, 2026Contributor·

  ·
  1

• [WebGPU] 0.6.0 cherry-picks round 1

  #34219·Edward Chen (edgchen1) opened Nov 14, 2026Contributor·

  ·

• [CUDA] Support multi-sequence Mamba2Verify and batched MatMulPacked decode for Nemotron Bolt

  model: nemotron-bolt
  #34218·Tianlei Wu (tianleiwu) opened Nov 14, 2026Contributor·

  ·
  15

• Fix QSA IndexShare reuse of held compressed-state prefixes

  #34217·bmehta001 opened Nov 14, 2026Contributor·

  ·
  1

• Add opt-in chunked LinearAttention prefill using GatedDeltaNet

  #34215·Selena Yang (selenayang888) opened Nov 13, 2026Contributor·

  ·

• Skip inactive XQA split-KV CTAs using runtime sequence length

  #34214·Selena Yang (selenayang888) opened Nov 13, 2026Contributor·

  ·

• Guard cuDNN 9.8 single-query grouped SDPA with large heads

  #34213·Selena Yang (selenayang888) opened Nov 13, 2026Contributor·

  ·

• Add opt-in wide INT4 M1 GEMV with uint8 zero points

  #34212·Selena Yang (selenayang888) opened Nov 13, 2026Contributor·

  ·
  4

• [Draft] Add MatMulNBits estimation test matrices and profiling validation

  #34211·Tommaso Adani (tadani3) opened Nov 13, 2026·

  ·
  

• Fix QMoE compatibility report capture on Windows

  #34210·eserscor opened Nov 13, 2026Contributor·

  ·
  4

• [CUDA] Add opt-in Mamba2Verify compact speculative replay

  #34209·BBaditi opened Nov 13, 2026Contributor·

  ·
  2

• [WebGPU] Add Apple INT4 register and FP32 subgroup matrix tiles

  #34208·Nikhil Jain (nikhilJain17) opened Nov 13, 2026·

  ·
  2

• Require Soup Sieve 4.0 or newer in Python documentation dependencies

  #34207·eserscor opened Nov 13, 2026Contributor·

  ·
  1

• [CPU] Fix Pad 'wrap' mode on inputs with a zero-size dimension

  #34206·0z5a (0z5a) opened Nov 13, 2026·

  ·
  1

  2

• Quantize N-D MatMul weights per-tensor in MatMulInteger/QLinearMatMul when per_channel is set

  #34204·Mohammad Sohail Shaikh (sohailshk) opened Nov 13, 2026·

  ·
  4

• Fix duplicate DequantizeLinear output in MatMulNBitsQuantizer QDQ for shared weights

  #34203·Mohammad Sohail Shaikh (sohailshk) opened Nov 13, 2026·

  ·
  4

• Try four opt-in GoogleTest provider shards in Linux CPU Debug CI

  #34201·eserscor opened Nov 13, 2026Contributor·

  ·
  1

• [CUDA] Route short shared-buffer GQA verify steps to speculative-decoding XQA

  #34198·Rui Ren (rui-ren) opened Nov 13, 2026Contributor·

  ·
  2

• fix(dnnl): Remove unused zero-point size calculation

  #34196·Airat Makhmutov (rautyrauty) opened Nov 13, 2026·

  ·
  2

• [CoreML] Fall back for ML Program MatMul with a dynamic K on only one input

  #34195·Vijay Kapse (vijay-kapse) opened Nov 13, 2026Contributor·

  ·
  1

  3

• [CoreML] Saturate out-of-range int64 constants to int32 instead of throwing

  #34193·Vijay Kapse (vijay-kapse) opened Nov 13, 2026Contributor·

  ·
  3