Issues

• [Build] OneCore CRT override is skipped for x64-hosted ARM64/ARM64EC OneCore builds

  api:Javabuild.NET
  microsoft/onnxruntime#34153 · Filter by author adrastogi opened on Nov 12, 2026
  

• [CPU EP] MatMulNBits 8-bit with accuracy_level <= 3 is 2x-40x slower than fp32 MatMul on AVX2 without VNNI when B <= 16 MiB

  platform:web
  microsoft/onnxruntime#34150 · Filter by author n24q02m opened on Nov 12, 2026
  1

• Symbolic shape inference rejects Concat axis=-1 for known shape values

  microsoft/onnxruntime#34143 · Filter by author itsrayeedscodespace opened on Nov 12, 2026

• [Performance] [ONNX Runtime] ConstantFolding misses sliced Shape folding when an unselected dimension is symbolic

  performance
  microsoft/onnxruntime#34140 · Filter by author nikokurumi opened on Nov 12, 2026

• [Performance] [ONNX Runtime] ConvBNFusion misses Conv–BatchNormalization fusion with an explicitly empty bias input

  performance
  microsoft/onnxruntime#34139 · Filter by author nikokurumi opened on Nov 12, 2026

• Symbolic shape inference ignores Resize keep_aspect_ratio_policy

  microsoft/onnxruntime#34123 · Filter by author itsrayeedscodespace opened on Nov 11, 2026

• Symbolic shape inference fails when Gather reads Constant values

  microsoft/onnxruntime#34122 · Filter by author itsrayeedscodespace opened on Nov 11, 2026

• [OpenVINO EP] "Output names mismatch between OpenVINO and ONNX" on dynamically quantized (int8) NeMo FastConformer encoder

  ep:OpenVINOquantization
  microsoft/onnxruntime#34108 · Filter by author m-umair-khan-dev opened on Nov 11, 2026
  1

• [Bug][DirectML] Valid Reshape with allowzero=1 and -1 fails during session initialization

  api:CSharpep:DML
  microsoft/onnxruntime#34107 · Filter by author araujo2kg opened on Nov 11, 2026

• [Mobile] WebGPU EP: VK_ERROR_DEVICE_LOST on Android Arm Mali GPUs since 1.29.0 (Dawn SPIR-V maximal reconvergence)

  ep:WebGPUplatform:mobileplatform:web
  microsoft/onnxruntime#34104 · Filter by author laurens-pilot opened on Nov 11, 2026

• [Performance] WebGPU EP: Conv3D uses a naive per-element shader and runs ~20x slower than the equivalent MatMul (0.08 vs 1.5 TFLOP/s)

  api:Javascriptep:WebGPUperformanceplatform:web
  microsoft/onnxruntime#34087 · Filter by author schlegelp opened on Nov 10, 2026

• [Web] WebGPU DP4A MatMulNBits gives wrong FP16 results on AMD RDNA1 (D3D12): subgroupMax on vec4<f16> in DP4AMatMulQuantize

  ep:WebGPUplatform:webquantization
  microsoft/onnxruntime#34078 · Filter by author julianmesa-gitkraken opened on Nov 10, 2026

• [Performance] GridSample (CPU EP, float, many channels, few output points) is 13–33% slower since 1.26.0 (#28060, #28302)

  microsoft/onnxruntime#34072 · Filter by author junghyunpark2001 opened on Nov 10, 2026

• [Performance] GridSample (CPU EP, float, 3-D, many channels) is ~12% slower since 1.25.0 (#27178)

  microsoft/onnxruntime#34071 · Filter by author junghyunpark2001 opened on Nov 10, 2026

• [Performance] QLinearConv (CPU EP, u8 x u8, AVX-VNNI-INT8) is 10–15% slower since 1.25.0 (#27671)

  microsoft/onnxruntime#34070 · Filter by author junghyunpark2001 opened on Nov 10, 2026

• [Performance] DFT (CPU EP, float, many short transforms) is ~11% slower since 1.27.0 (#27028)

  microsoft/onnxruntime#34069 · Filter by author junghyunpark2001 opened on Nov 10, 2026

• [Performance] ~2× slower quantized MatMul on Intel N100 starting with ORT 1.25.0 (1.24.4 is fast)

  performancequantization
  microsoft/onnxruntime#34035 · Filter by author farooqu opened on Nov 9, 2026
  1

• [QNN EP] Dynamic uint16 × uint16 MatMul is always lowered to 16 × 8 (Convert to uint8), large accuracy loss in attention on HTP v73

  ep:QNNplatform:mobile
  microsoft/onnxruntime#34033 · Filter by author chakib-belgaid opened on Nov 9, 2026
  1

• [Performance] TreeNodeElementId hash function is horrendously inefficient on Windows

  performanceplatform:windows
  microsoft/onnxruntime#34027 · Filter by author GerritNG opened on Nov 9, 2026
  1

• [Performance] [ONNX Runtime] TransposeOptimizer misses inverse transposes when perm is omitted

  performance
  microsoft/onnxruntime#34008 · Filter by author nikokurumi opened on Nov 9, 2026

• [Performance] [ONNX Runtime] LayerNormFusion misses rank-one normalization with axis 0

  performance
  microsoft/onnxruntime#34006 · Filter by author miyorena opened on Nov 9, 2026

• [CPU] ReduceMax, ReduceMin, ArgMax and ArgMin ignore NaN unless it is the first element

  microsoft/onnxruntime#34001 · Filter by author 14H034160212 opened on Nov 9, 2026
  1

• CPU: FP16 MatMul followed by Cast to FP32 bypasses FP16 rounding

  microsoft/onnxruntime#33998 · Filter by author tangyixuan01 opened on Nov 9, 2026
  1

• [Performance] CUDA GroupQueryAttention: each new prompt length waits for a cuDNN graph build (default sdpa_kernel, SM 12.0)

  ep:CUDA
  microsoft/onnxruntime#33995 · Filter by author john-rocky opened on Nov 9, 2026
  4

• [MIGraphX EP] Model cache reuses a compiled program for a different input shape

  ep:MIGraphX
  microsoft/onnxruntime#33992 · Filter by author JustinTArthur opened on Nov 8, 2026