Issues

[Performance] CUDA Conv: cuDNN frontend >= 1.24 calls cudaGetDeviceProperties on every plan rebuild (~1 ms per Conv node per input-shape change)

microsoft/onnxruntime#33769 · opened

on Oct 30, 2026

Plugin-EP partition names collide under `ep.share_ep_contexts`: byte-loaded models share one identity, and `ep.context_node_name_prefix` is ignored

microsoft/onnxruntime#33758 · opened

on Oct 30, 2026

[Web] WebGPU SkipLayerNormalization / SkipSimplifiedLayerNormalization return zeros past maxStorageBufferBindingSize

microsoft/onnxruntime#33755 · opened

on Oct 30, 2026

[Web] [WebGPU] Single-pass prefill of ~900 tokens crashes Safari's GPU process on iOS 27 (iPhone) — please split large workloads into smaller submissions / dispatches

microsoft/onnxruntime#33753 · opened

on Oct 29, 2026

[Feature Request] [Web] Expose kvCacheQuantizationBits in the onnxruntime-web WebGPU EP options

microsoft/onnxruntime#33752 · opened

on Oct 29, 2026

[Web] WebGPU bucket buffer cache: 4 MiB → 8 MiB step and low large-bucket limits cause a memory jump and allocation churn in long-context LLM decoding; limits not configurable

microsoft/onnxruntime#33751 · opened

on Oct 29, 2026

1

[Web] WebGPU EP creates a new GPUBindGroup for every dispatch (~317 per decoded token); bind groups are costly in WebKit (memory held until GC)

microsoft/onnxruntime#33750 · opened

on Oct 29, 2026

1

[Feature Request] EP-agnostic callbacks for persisting EP runtime/compilation caches

microsoft/onnxruntime#33748 · opened

on Oct 29, 2026

[Web] Firefox throws URL constructor: null is not a valid URL when opening Vite single-file build via file://

microsoft/onnxruntime#33735 · opened

on Oct 29, 2026

[Bug][DirectML EP] Reshape with allowzero=1 and a -1 dimension misinfers the output shape and crashes

microsoft/onnxruntime#33720 · opened

on Oct 29, 2026

[CPU EP] MatMulNBits (accuracy_level=4) produces wrong results on x86-64 AVX2 when loaded from `session.save_external_prepacked_constant_initializers` output (1.31.1)

microsoft/onnxruntime#33677 · opened

on Oct 28, 2026

[Web/WebGPU] 1.31: a session created on a page-supplied GPUDevice cannot bind gpu-buffer inputs

microsoft/onnxruntime#33642 · opened

on Oct 27, 2026

1

[Build] CUDA EP fails to load in training-enabled builds since #29252: backward cuDNN calls bypass the dynamic loader

microsoft/onnxruntime#33638 · opened

on Oct 27, 2026

1

[Build] Both 1.31.0 and 1.31.1 failed to Build Windows msvc2019 x64 Release with error C2440

microsoft/onnxruntime#33636 · opened

on Oct 27, 2026

[Plugin EP] Provider-bridge probe unloads the EP DLL, so RegisterExecutionProviderLibrary loads and initializes it twice

microsoft/onnxruntime#33603 · opened

on Oct 26, 2026

[Build] cannot build with --disable_contrib_ops on Windows CUDA

microsoft/onnxruntime#33602 · opened

on Oct 26, 2026

[CPU] ReduceSum loses int64 precision even when the result does not overflow

microsoft/onnxruntime#33579 · opened

on Oct 25, 2026

MatMulNBits benchmark generates IR12 models rejected by ORT 1.23.2

microsoft/onnxruntime#33576 · opened

on Oct 24, 2026

CoreML EP (ML Program): Conv without `pads` fails with "Required param 'pad' is missing"

microsoft/onnxruntime#33574 · opened

on Oct 24, 2026

[Web] Next release 1.26 is scheduled for May 2026 has passed

microsoft/onnxruntime#33570 · opened

on Oct 24, 2026

[CoreML EP] Model cache keeps the ML Program package beside the compiled model, doubling its size

microsoft/onnxruntime#33566 · opened

on Oct 23, 2026

[CUDA EP] GroupQueryAttention with past/present bound to the same buffer gives nondeterministic, sometimes wrong results across runs (worse under CUDA graph capture)

microsoft/onnxruntime#33562 · opened

on Oct 23, 2026

[WebNN] [WebGPU EP] [Windows ML] WebGPU EP via ExecutionProviderCatalog: wrong `LibraryPath`, and no way to require/upgrade to a minimum EP version

microsoft/onnxruntime#33558 · opened

on Oct 23, 2026

2

[Python] A requested execution provider that can't be created is silently dropped, e.g. a plugin MIGraphXExecutionProvider

microsoft/onnxruntime#33556 · opened

on Oct 23, 2026

[DirectML EP] MatMul with rank-5 inputs fails the whole session at initialization (80070057) instead of falling back

microsoft/onnxruntime#33549 · opened

on Oct 22, 2026

Navigation Menu

Repository navigation

Issues

[Performance] CUDA Conv: cuDNN frontend >= 1.24 calls cudaGetDeviceProperties on every plan rebuild (~1 ms per Conv node per input-shape change)

Plugin-EP partition names collide under `ep.share_ep_contexts`: byte-loaded models share one identity, and `ep.context_node_name_prefix` is ignored

[Web] WebGPU SkipLayerNormalization / SkipSimplifiedLayerNormalization return zeros past maxStorageBufferBindingSize

[Web] [WebGPU] Single-pass prefill of ~900 tokens crashes Safari's GPU process on iOS 27 (iPhone) — please split large workloads into smaller submissions / dispatches

[Feature Request] [Web] Expose kvCacheQuantizationBits in the onnxruntime-web WebGPU EP options

[Web] WebGPU bucket buffer cache: 4 MiB → 8 MiB step and low large-bucket limits cause a memory jump and allocation churn in long-context LLM decoding; limits not configurable

[Web] WebGPU EP creates a new GPUBindGroup for every dispatch (~317 per decoded token); bind groups are costly in WebKit (memory held until GC)

[Feature Request] EP-agnostic callbacks for persisting EP runtime/compilation caches

[Web] Firefox throws URL constructor: null is not a valid URL when opening Vite single-file build via file://

[Bug][DirectML EP] Reshape with allowzero=1 and a -1 dimension misinfers the output shape and crashes

[CPU EP] MatMulNBits (accuracy_level=4) produces wrong results on x86-64 AVX2 when loaded from `session.save_external_prepacked_constant_initializers` output (1.31.1)

[Web/WebGPU] 1.31: a session created on a page-supplied GPUDevice cannot bind gpu-buffer inputs

[Build] CUDA EP fails to load in training-enabled builds since #29252: backward cuDNN calls bypass the dynamic loader

[Build] Both 1.31.0 and 1.31.1 failed to Build Windows msvc2019 x64 Release with error C2440

[Plugin EP] Provider-bridge probe unloads the EP DLL, so RegisterExecutionProviderLibrary loads and initializes it twice

[Build] cannot build with --disable_contrib_ops on Windows CUDA

[CPU] ReduceSum loses int64 precision even when the result does not overflow

MatMulNBits benchmark generates IR12 models rejected by ORT 1.23.2

CoreML EP (ML Program): Conv without `pads` fails with "Required param 'pad' is missing"

[Web] Next release 1.26 is scheduled for May 2026 has passed

[CoreML EP] Model cache keeps the ML Program package beside the compiled model, doubling its size

[CUDA EP] GroupQueryAttention with past/present bound to the same buffer gives nondeterministic, sometimes wrong results across runs (worse under CUDA graph capture)

[WebNN] [WebGPU EP] [Windows ML] WebGPU EP via ExecutionProviderCatalog: wrong `LibraryPath`, and no way to require/upgrade to a minimum EP version

[Python] A requested execution provider that can't be created is silently dropped, e.g. a plugin MIGraphXExecutionProvider

[DirectML EP] MatMul with rank-5 inputs fails the whole session at initialization (80070057) instead of falling back

📋 Server Request & Proxy Logs

⚙️ Pengaturan Proxy Target

Issues

Issues

Issues