Issues

[CPU EP] FP16 DQ + MatMul fusion returns zero with MatMulNBits accuracy_level=4

microsoft/onnxruntime#34216 · opened

on Nov 13, 2026

[CPU] Pad with mode="wrap" fails with "Unexpected mode of 3" when the input has a zero-size dimension

microsoft/onnxruntime#34205 · opened

on Nov 13, 2026

[Web] Let `env.wasm` pass a precompiled module to the runtime (`instantiateWasm`)

microsoft/onnxruntime#34202 · opened

on Nov 13, 2026

[CPU EP] DQ-to-MatMulNBits fusion drops FP16 dequantization rounding with accuracy_level=1

microsoft/onnxruntime#34197 · opened

on Nov 13, 2026

[CoreML EP] MLProgram: MatMul with a static K on one input and a dynamic K on the other fails to load (default dynamic dim of 1)

microsoft/onnxruntime#34194 · opened

on Nov 13, 2026

[Optimizer] PropagateCastOps throws when Where consumes the same Cast output twice

microsoft/onnxruntime#34192 · opened

on Nov 13, 2026

[TensorRT EP] `RunAsync` fails on a GPU device other than 0 when `user_compute_stream` is set (regression since #26542)

microsoft/onnxruntime#34191 · opened

on Nov 13, 2026

[CPU EP] FP16 Mul fallback removes an explicit FP32-to-FP16 input Cast

microsoft/onnxruntime#34189 · opened

on Nov 13, 2026

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

microsoft/onnxruntime#34153 · 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

microsoft/onnxruntime#34150 · opened

on Nov 12, 2026

1

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

microsoft/onnxruntime#34143 · opened

on Nov 12, 2026

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

microsoft/onnxruntime#34140 · opened

on Nov 12, 2026

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

microsoft/onnxruntime#34139 · opened

on Nov 12, 2026

Symbolic shape inference ignores Resize keep_aspect_ratio_policy

microsoft/onnxruntime#34123 · opened

on Nov 11, 2026

Symbolic shape inference fails when Gather reads Constant values

microsoft/onnxruntime#34122 · opened

on Nov 11, 2026

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

microsoft/onnxruntime#34108 · opened

on Nov 11, 2026

1

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

microsoft/onnxruntime#34107 · 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)

microsoft/onnxruntime#34104 · 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)

microsoft/onnxruntime#34087 · opened

on Nov 10, 2026

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

microsoft/onnxruntime#34078 · 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 · 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 · 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 · opened

on Nov 10, 2026

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

microsoft/onnxruntime#34069 · opened

on Nov 10, 2026

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

microsoft/onnxruntime#34035 · opened

on Nov 9, 2026

1

Navigation Menu

Repository navigation

Issues

[CPU EP] FP16 DQ + MatMul fusion returns zero with MatMulNBits accuracy_level=4

[CPU] Pad with mode="wrap" fails with "Unexpected mode of 3" when the input has a zero-size dimension

[Web] Let `env.wasm` pass a precompiled module to the runtime (`instantiateWasm`)

[CPU EP] DQ-to-MatMulNBits fusion drops FP16 dequantization rounding with accuracy_level=1

[CoreML EP] MLProgram: MatMul with a static K on one input and a dynamic K on the other fails to load (default dynamic dim of 1)

[Optimizer] PropagateCastOps throws when Where consumes the same Cast output twice

[TensorRT EP] `RunAsync` fails on a GPU device other than 0 when `user_compute_stream` is set (regression since #26542)

[CPU EP] FP16 Mul fallback removes an explicit FP32-to-FP16 input Cast

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

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

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

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

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

Symbolic shape inference ignores Resize keep_aspect_ratio_policy

Symbolic shape inference fails when Gather reads Constant values

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

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

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

[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)

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

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

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

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

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

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