All pull requests

Pull requests

• [XNNPACK] Support Conv/ConvTranspose with symbolic spatial dims

  #34015·Danny On The Air (dannyota) opened Nov 9, 2026Contributor·

  ·
  1

  1

• Add a C API to serialize an OrtGraph to an ONNX model

  #34014·Danny On The Air (dannyota) opened Nov 9, 2026Contributor·

  ·
  1

  1

• Add EP-agnostic callbacks for persisting EP caches

  #34013·Danny On The Air (dannyota) opened Nov 9, 2026Contributor·

  ·
  1

  2

• Preserve FP16 rounding before explicit consumer Casts

  #34012·Asuka (akasakariko) opened Nov 9, 2026Contributor·

  ·
  1

  2

• [WebGPU] Enable SimplifiedLayerNormFusion for the WebGPU EP

  #34011·Danny On The Air (dannyota) opened Nov 9, 2026Contributor·

  ·
  1

  5

• Ship the py.typed marker in the Python wheel

  #34010·Danny On The Air (dannyota) opened Nov 9, 2026Contributor·

  ·
  2

• Improve session creation time further

  #34009·Christian Bourjau (cbourjau) opened Nov 9, 2026Contributor·

  ·
  2

• Discover GPUs through D3DKMT on WSL2

  #34007·mingyue (mingyueliuh) opened Nov 9, 2026Contributor·

  ·
  4

• Register the existing CUDA BF16 Less kernel

  #34005·Baiju Meswani (baijumeswani) opened Nov 9, 2026Contributor·

  ·

• Fix the CPU Pad slowdown for tall inputs with a small inner extent

  #34004·Chinmay Shrivastava (JonSnow1807) opened Nov 9, 2026·

  ·
  1

  2

• Reduce native packed decode tile and FC2 loader overhead

  #34003·Rui Ren (rui-ren) opened Nov 9, 2026Contributor·7/7·

  ·
  1

• Speed up native Bolt packed projections and IQ2 experts

  #34002·Rui Ren (rui-ren) opened Nov 9, 2026Contributor·6/7·

  ·

• Fix MIGraphX model cache keys for dynamic shapes

  #34000·自分 (2Zest) opened Nov 9, 2026·

  ·
  1

  2

• WebGPU: inherit Dawn cache configuration when omitted

  #33999·Sushanth Rajasankar (sushraja-msft) opened Nov 9, 2026Contributor·

  ·
  1

• [CUDA] Support FP16 KV cache and FP16 query/key in DynamicSparseAttention<BFloat16> and FP32 rotary cache / FP16 output in RMSNormRotary

  #33997·Tianlei Wu (tianleiwu) opened Nov 9, 2026Contributor·

  ·
  7

• Restore vectorized PRelu for per-channel slopes on CPU

  #33996·Chinmay Shrivastava (JonSnow1807) opened Nov 9, 2026·

  ·
  1

  5

• [CUDA] Extract Bolt projected-sigmoid routing into MoERouterPacked

  #33994·Rui Ren (rui-ren) opened Nov 8, 2026Contributor·5/7·

  ·

• Keep 1DS SDK version and auth mutex alive for late exit workers

  #33993·bmehta001 opened Nov 8, 2026Contributor·

  ·
  3

  

• [TEST ONLY] Consolidated Bolt stack for Microsoft benchmarking

  #33991·Rui Ren (rui-ren) opened Nov 8, 2026Contributor·

  ·

• docs: fix duplicated word in build info comment

  #33990·ddlw opened Nov 8, 2026·

  ·
  1

  4

• fix(optimizer): guard Relu quant fusion for non-positive scales

  #33988·自分 (2Zest) opened Nov 8, 2026·

  ·
  4

• fix: Resolve FTBFS with protobuf 36

  #33987·Airat Makhmutov (rautyrauty) opened Nov 8, 2026·

  ·
  4

• Fix SoftmaxCrossEntropyLoss shape with reduction none in symbolic shape inference

  #33986·Lanre Shittu (Shizoqua) opened Nov 8, 2026·

  ·
  1

  2

• [js/web] Rebind external tensors on every run when graph capture is enabled

  #33985·naft-san (Nominy) opened Nov 8, 2026·

  ·
  9

• Fix LayerNormalization stats dtype in symbolic shape inference

  #33983·Lanre Shittu (Shizoqua) opened Nov 8, 2026·

  ·
  1

  2