Issues

• [Performance] ReduceLogSumExp (CPU EP, float, empty axes with noop_with_empty_axes=1) is ~36x slower since 1.24.1 (#26436)

  performance
  microsoft/onnxruntime#33910 · Filter by author junghyunpark2001 opened on Nov 5, 2026

• CUDA provider has no bool ReduceMax kernel

  ep:CUDA
  microsoft/onnxruntime#33876 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026

• CUDA graph replay of a graph with GatherND fails with an illegal memory access

  ep:CUDA
  microsoft/onnxruntime#33875 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  

• Pad has no bfloat16 kernel (CPU or CUDA); session fails to initialize with an unclear error

  ep:CUDA
  microsoft/onnxruntime#33874 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026

• [Performance] CPU MultiHeadAttention with attention_bias is ~4x slower than the unfused ops

  model:transformer
  microsoft/onnxruntime#33873 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  

• CUDA GroupQueryAttention with an attention_bias runs an unfused kernel, 4x slower than MultiHeadAttention with the same bias

  ep:CUDA
  microsoft/onnxruntime#33869 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026

• CUDA GroupQueryAttention reads a total_sequence_length computed on the GPU as 0 and rejects the attention_bias

  ep:CUDA
  microsoft/onnxruntime#33867 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  1

• CPU MoE gives wrong results when its input is computed in the graph and its output feeds another node

  microsoft/onnxruntime#33866 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  1

  

• GroupQueryAttention fails to load when total_sequence_length is computed from static shapes (ParseData type mismatch)

  microsoft/onnxruntime#33865 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  

• CUDA MoE is up to 3x slower in 1.29 than in 1.23

  ep:CUDA
  microsoft/onnxruntime#33864 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  

• CUDA MoE with activation_type="gelu" and fc3 (gated GELU) gives wrong results since 1.28

  ep:CUDA
  microsoft/onnxruntime#33863 · Filter by author IlyasMoutawwakil opened on Nov 4, 2026
  

• MLAS LinearAttention test exceeds tolerance on Windows ARM64

  platform:windows
  microsoft/onnxruntime#33862 · Filter by author patryk-kaiser-ARM opened on Nov 4, 2026

• [Performance] CPU Pad on tall 2-D inputs is 22–44% slower since 1.24.1 (#26864)

  performance
  microsoft/onnxruntime#33861 · Filter by author junghyunpark2001 opened on Nov 4, 2026

• [Performance] WebGPU: pointwise Conv ~8-15x slower on Intel Arc A750 since #32304 (Conv now dispatches MatMulSubgroup); a 70 s Conformer encoder takes 2.1x as long (onnxruntime-node 1.30.0+)

  ep:WebGPUmodel:transformerplatform:web
  microsoft/onnxruntime#33856 · Filter by author qwertz92 opened on Nov 4, 2026
  3

• [Performance] WebGPU LLM decode ~14-17 % slower since 1.30.0 (first slow build 1.30.0-dev.20260819, which adds #32063): same GPU programs, longer CPU-side step. Intel Arc A750, Windows 11, onnxruntime-node

  ep:WebGPUmodel:transformerplatform:webplatform:windows
  microsoft/onnxruntime#33855 · Filter by author qwertz92 opened on Nov 4, 2026

• [Performance] PRelu (CPU EP, float, per-channel slope) is ~21x slower since 1.27.0 (#28750)

  performance
  microsoft/onnxruntime#33831 · Filter by author junghyunpark2001 opened on Nov 3, 2026

• [Build] Default build fails on musl (Alpine): bundled cpp_client_telemetry includes <execinfo.h>

  buildplatform:mobile
  microsoft/onnxruntime#33827 · Filter by author mievst opened on Nov 2, 2026
  1

• [Performance] [ONNX Runtime] ConvAddFusion misses channel-bias fusion when Add inputs are reversed

  performance
  microsoft/onnxruntime#33818 · Filter by author miyorena opened on Nov 2, 2026

• [Mobile] [QNN] CumSum node gets rejected on HTP if axis type is int64

  ep:QNNplatform:mobile
  microsoft/onnxruntime#33815 · Filter by author grybouilli opened on Nov 2, 2026

• [CUDA EP] Expand receives a wrong shape: CPU buffer of a live int64 shape tensor is reused (1.29.0, 1.31.1)

  ep:CUDA
  microsoft/onnxruntime#33807 · Filter by author jrosskopf opened on Nov 2, 2026
  

• [WebGPU EP] Failed to compile graph with error "Alloc WebGPU EP device-free context must not allocate device memory"

  ep:WebGPUep:WebNNplatform:web
  microsoft/onnxruntime#33806 · Filter by author mingmingtasd opened on Nov 2, 2026
  2

• [Feature Request] [WebGPU] per-channel quantization MatMulNBits

  ep:WebGPUfeature requestmodel:transformerplatform:webquantization
  microsoft/onnxruntime#33805 · Filter by author kylo5aby opened on Nov 2, 2026
  3

• [WebGPU EP] Sessions configured with different OrtEpDevices use the same physical GPU

  api:CSharpep:WebGPU.NETplatform:web
  microsoft/onnxruntime#33803 · Filter by author xiaofeihan1 opened on Nov 2, 2026
  1

• [Performance] [ONNX Runtime] GeluFusion misses fusion when GELU input is produced by Mul

  performance
  microsoft/onnxruntime#33799 · Filter by author miyorena opened on Nov 1, 2026

• FP16 MatMul on Linux arm64 accumulates in float16 from 1.29.0: Qwen3-Embedding-0.6B drifts from float32 (1 − cosine 1.5e-6 → 2.7e-4)

  microsoft/onnxruntime#33798 · Filter by author Laurianti opened on Nov 1, 2026