33 Pull requests merged by 21 people

• Integration with ONNX 1.19

  #25678 merged Sep 6, 2025

• [QNN EP] fix memory leak in OrtEpFactory::GetSupportedDevices()

  #25968 merged Sep 5, 2025

• Make name an std::string in OrtMemoryInfo

  #25960 merged Sep 5, 2025

• cherry picks for 1.23.0 release

  #25959 merged Sep 5, 2025

• [GH Pages] Add plugin EP library documentation

  #25880 merged Sep 5, 2025

• Upgrade wheel to 0.45.1

  #25957 merged Sep 4, 2025

• Use vmlaq_f32 in MlasMultipleAddFloat32x4 for Android armeabi-v7a.

  #25955 merged Sep 4, 2025

• [NV TensorRT RTX] Handle unsupported data types

  #25953 merged Sep 4, 2025

• [MLAS] Add 8-bit weights ARM64 Gemm implementation

  #25110 merged Sep 4, 2025

• [WebNN] Support IsNaN and IsInf

  #25930 merged Sep 4, 2025

• [TRT RTX EP] Fixing the stream parameter in CopyTensors API and passing cudastream by value in createNotification API

  #25937 merged Sep 4, 2025

• Compile API: output model and initializer stream write functions

  #25455 merged Sep 4, 2025

• Remove std::string::data() non-const usage from public headers

  #25943 merged Sep 4, 2025

• Fix transpose of empty initializers called from constant folding

  #25922 merged Sep 4, 2025

• [QNN-EP] Add negtive padding value support.

  #25928 merged Sep 4, 2025

• Improve packaging pipelines RC version support

  #25938 merged Sep 4, 2025

• [Core] Fix debug node input output compilation after Fp4 support was enabled in ORT

  #25940 merged Sep 4, 2025

• Bump next from 15.2.4 to 15.4.7 in /js/web/test/e2e/exports/testcases/nextjs-default

  #25904 merged Sep 4, 2025

• Migrate model tests to ONNX Model ZOO only

  #25888 merged Sep 3, 2025

• upgrade dawn to 13c1635a14574ebb7116b56a69f5519301417fda

  #25703 merged Sep 3, 2025

• Updated MetaHuman testimonial from Epic Games.

  #23737 merged Sep 3, 2025

• POWER : Implement MlasGemmQuantKernel using VSX builtins for M = 1

  #25490 merged Sep 3, 2025

• Revert "[QNN EP] Remove workaround for syntax error macro (#25923)"

  #25939 merged Sep 3, 2025

• [Core] Support fp4 type in ORT

  #25767 merged Sep 3, 2025

• [CXX] Introduce C++ API for new C entry points

  #25897 merged Sep 3, 2025

• Compile API: disable optimizations by default

  #25474 merged Sep 3, 2025

• [TRT RTX EP] Add support for RTX runtime caches

  #25917 merged Sep 3, 2025

• [TRT RTX EP] Memory map the engine buffer

  #25909 merged Sep 3, 2025

• [TRT RTX EP] Add sync method

  #25898 merged Sep 2, 2025

• [QNN EP] Remove workaround for syntax error macro

  #25923 merged Sep 2, 2025

• Support DynamicQuantizeLinear op

  #25905 merged Sep 2, 2025

• Replace vmlaq_f32 with vfmaq_f32 (fused multiply-add)

  #25669 merged Sep 2, 2025

• [webgpu] Add back missing code comments for flash decoding

  #25879 merged Sep 1, 2025

24 Pull requests opened by 15 people

• [CUDA] Support SwiGlu in MoE and qMoE (#25530)

  #25903 opened Aug 30, 2025

• Bump ruff from 0.12.9 to 0.12.11

  #25906 opened Sep 1, 2025

• Bump clang-format from 20.1.8 to 21.1.0

  #25907 opened Sep 1, 2025

• [webgpu] Manually normalize `Conv-Transpose` dispatch group size for Intel GPUs

  #25908 opened Sep 1, 2025

• Enables Matmul and Gemm for float16 on CPU

  #25913 opened Sep 1, 2025

• [TRT RTX EP] Remove unused legacy code from TRT EP that is no longer supported for TRT RTX

  #25916 opened Sep 2, 2025

• [Web] Magic import comment missing for Vite (/* @vite-ignore */)

  #25919 opened Sep 2, 2025

• Update AppendExecutionProvider docs

  #25924 opened Sep 2, 2025

• Fix Local Attention off by 1 bug

  #25927 opened Sep 2, 2025

• [webgpu] Optimize flash decoding by merging QKT and SplitVx shader

  #25929 opened Sep 3, 2025

• [WebNN] Fix bug in quantizeLinear and dequantizeLinear

  #25931 opened Sep 3, 2025

• [webgpu] Add dispatchWorkgroupsIndirect support

  #25934 opened Sep 3, 2025

• [WebGPU] allow async shader compilation

  #25941 opened Sep 3, 2025

• Add onnxruntime_plugin_ep_onnx_test for plugin EPs to run onnx test

  #25942 opened Sep 4, 2025

• [webgpu] Unify the present_sequence_length in flash attention

  #25945 opened Sep 4, 2025

• Bump electron from 28.3.2 to 35.7.5 in /js/web

  #25950 opened Sep 4, 2025

• [CPU] MoE Kernel

  #25958 opened Sep 4, 2025

• [webgpu] Move copy_kv_cache to template

  #25961 opened Sep 5, 2025

• Split large inputs into smaller buffers to bypass maxStorageBufferBindingSize limit

  #25962 opened Sep 5, 2025

• [TRT RTX EP] Rename TRT RTX libs to not collide with TRT

  #25964 opened Sep 5, 2025

• Fix Attention GQA implementation on CPU

  #25966 opened Sep 5, 2025

• Update React Native iOS CI build.

  #25967 opened Sep 5, 2025

• Clean up CUDA contrib ops folder

  #25969 opened Sep 5, 2025

• [WIP] Fix Mac Catalyst build options.

  #25970 opened Sep 5, 2025

21 Issues closed by 7 people

• GetEpDevices() does not Detect Intel NPU via OpenVINO EP

  #25557 closed Sep 5, 2025

• [Build] compiling errors for armeabi-v7a: vfmaq_f32 undecared

  #25949 closed Sep 4, 2025

• In Dark Mode, "Interested in using other languages? See the many others we support" link is not properly visible: A11y_ONNX Runtime & Ecosystem_Runtime_User Preferences

  #24964 closed Sep 3, 2025

• onnxruntime errors out due to the wrong process of GatherElements operator with the CPUExecutionProvider: Out of range value in index tensor

  #24917 closed Sep 3, 2025

• [BUG] Non-zero status code returned while running Resize node. in Direct ML backend

  #24928 closed Sep 3, 2025

• Incorrect Use of CUDA Constants in MIGraphXExecutionProvider::CreatePreferredAllocators (Should Use HIP)

  #25268 closed Sep 3, 2025

• [Performance] How to used pinned memory in onnxruntime.

  #20947 closed Sep 2, 2025

• [Build] can't build CUDA (+ vino and directML) for latest v1.22 on windows

  #25081 closed Sep 2, 2025

• Initializers use wrong allocator

  #25108 closed Sep 2, 2025

• [Build] CMake configurations files for bin release 1.22.0 are broken

  #25242 closed Sep 2, 2025

• [DirectML EP] Error when validating attributes of `Slice` operator

  #25252 closed Sep 2, 2025

• [Build] CMake configurations files for bin release 1.22.0 are broken for Linux

  #25279 closed Sep 2, 2025

• Invalid MIGraphX EP option: migraphx_load_compiled_path

  #25379 closed Sep 2, 2025

• [Mobile] OnnxTensor FLOAT16 creation works, but runtime fails with "unsupported type FLOAT16

  #25891 closed Sep 1, 2025

• [Feature Request] Unable to run microsoft/Phi-3.5-vision-instruct-onnx on GPU with C#

  #25892 closed Sep 1, 2025

• [Build] CCCL API migration issue.

  #24774 closed Aug 31, 2025

• [Build] How to build ONNX Runtime as a dynamic framework (.dylib/.framework) for iOS？

  #25256 closed Aug 31, 2025

• about infer ocr Memory exception

  #25258 closed Aug 31, 2025

• onnxruntime can execute the model correctly with the CUDAExecutionProvider, while returning error code with the CPUExecutionProvider: NOT_IMPLEMENTED : Could not find an implementation for Resize(13) node with name ''

  #24919 closed Aug 30, 2025

• [Web] Error using opus-mt-mul fp16 models with WebGPU

  #25125 closed Aug 30, 2025

• When cast a float tensor to uint32, different results are produced by the CPUExecutionProvider and CUDAExecutionProvider

  #25207 closed Aug 30, 2025

16 Issues opened by 15 people

• [Performance] Long-context OOM on newer architectures

  #25965 opened Sep 5, 2025

• CPU and CUDA provider outputs differ for 2D UNet CNN with GraphOptimizationLevel.ORT_ENABLE_BASIC; ORT_DISABLE_ALL fixes issue.

  #25963 opened Sep 5, 2025

• [Web] Static WASM library: Actually selecting WebGpuExecutionProvider

  #25952 opened Sep 4, 2025

• Crash in `MlasRotaryEmbedOneRow` when running EmbeddingGemma

  #25951 opened Sep 4, 2025

• [Runtime bug on Windows] resize op exception. Tensor type mismatch.

  #25948 opened Sep 4, 2025

• [Build] Wheel file not present after build from source with `--build_wheel`

  #25947 opened Sep 4, 2025

• [Performance] Performance Regression in ReduceMean Operator on Specific Axes

  #25946 opened Sep 4, 2025

• Latest release v1.22.2 not available on pypi

  #25944 opened Sep 4, 2025

• [Build] CUDA 13 Failed

  #25936 opened Sep 3, 2025

• Could not find an implementation for QuantizeLinear(23)

  #25932 opened Sep 3, 2025

• LayerNormalization - The parameter is incorrect error

  #25926 opened Sep 2, 2025

• [Build] build for catalyst support will generate invalid compile command

  #25920 opened Sep 2, 2025

• [Web] Magic import comment missing for Vite (/* @vite-ignore */)

  #25918 opened Sep 2, 2025

• Deadlock Bug in Win10 + ORT + Specific Execution Mode

  #25915 opened Sep 2, 2025

• No opset import for domain 'com.microsoft.nchwc' error

  #25914 opened Sep 2, 2025

• Windows 7 compatibility for Microsoft.ML.OnnxRuntime NuGet package with C#

  #25910 opened Sep 1, 2025

57 Unresolved conversations

Sometimes conversations happen on old items that aren’t yet closed. Here is a list of all the Issues and Pull Requests with unresolved conversations.

• [AIX] test-suites failure fixes

  #25791 commented on Sep 5, 2025 • 17 new comments

• FP16 inference performance improvement on CPU

  #25680 commented on Sep 2, 2025 • 10 new comments

• NEON kernels for NCHWc Convolution and Pooling

  #25580 commented on Sep 5, 2025 • 9 new comments

• Support int4 and uint4 for reshape on opset 21+.

  #25645 commented on Sep 6, 2025 • 6 new comments

• Extend ORT perf test to report submit vs inference time

  #25895 commented on Sep 5, 2025 • 6 new comments

• Integrate SME1 SGEMM KleidiAI kernels

  #25760 commented on Sep 5, 2025 • 6 new comments

• [Don't review][webgpu] Make graph capture work on LLM

  #25868 commented on Sep 3, 2025 • 6 new comments

• [ARM CPU] SVE support for Elementwise kernels

  #25238 commented on Sep 2, 2025 • 5 new comments

• [CANN] Fix the ACL_ERROR_REPEAT_INITIALIZE error that occurs when coexisting with torch_npu

  #25902 commented on Aug 30, 2025 • 5 new comments

• Fix incorrect symbolic shape inference for ConstantOfShape

  #25857 commented on Sep 1, 2025 • 1 new comment

• Add model compilation in ORT perf test

  #25797 commented on Sep 5, 2025 • 1 new comment

• [Feature Request] Making the cuda version explicit in the python package name

  #19438 commented on Sep 5, 2025 • 0 new comments

• onnxruntime 1.17.0: transformers benchmarking failing for int8 quantized inference.

  #19409 commented on Sep 5, 2025 • 0 new comments

• Compile API: output EPContext binary data to write function

  #25471 commented on Sep 4, 2025 • 0 new comments

• LayerNormalization shape check error

  #25896 commented on Sep 5, 2025 • 0 new comments

• [Web] Avoid unnecessary data copy for pre-allocated tensors

  #25571 commented on Sep 5, 2025 • 0 new comments

• Error messages from QNN are turned into verbose level messages

  #24876 commented on Sep 5, 2025 • 0 new comments

• [CANN] When using onnxruntime-cann for inference, it failed to utilize the NPU for inference

  #22229 commented on Sep 5, 2025 • 0 new comments

• Update ORT to handle explicit OpSchemaRegisterOnce API in ONNX >= 1.18.0 for fluent chaining

  #24561 commented on Aug 30, 2025 • 0 new comments

• Bump transformers from 4.50.0 to 4.53.0 in /onnxruntime/python/tools/transformers/models/stable_diffusion/requirements

  #25672 commented on Sep 6, 2025 • 0 new comments

• Add `onnxruntime_show_ep_devices`

  #25780 commented on Sep 2, 2025 • 0 new comments

• webgpu: support bool type for Flatten and Unsqueeze operators

  #25804 commented on Sep 4, 2025 • 0 new comments

• Fix CPU EP Tile 0D overvalidation

  #25821 commented on Sep 2, 2025 • 0 new comments

• Add ort api for CustomOp compute events.

  #25843 commented on Sep 1, 2025 • 0 new comments

• [webgpu] Optimize MatMul Op

  #25854 commented on Sep 5, 2025 • 0 new comments

• [QNN EP] Add Case-2 LPBQ pattern support for Gemm and Matmul nodes

  #25865 commented on Sep 5, 2025 • 0 new comments

• [TRT-RTX EP] add missing thread header include to TRT-RTX provider

  #25875 commented on Sep 3, 2025 • 0 new comments

• [DRAFT] Attention and MHA CUDA BFloat16 Support - not ready for review

  #25885 commented on Sep 5, 2025 • 0 new comments

• [Documentation] Update ONNX Runtime Documentation to Position WinML as Preferred Windows Path and DirectML in Sustained Engineering

  #25901 commented on Sep 3, 2025 • 0 new comments

• [Build] add presets for CMake configuration.

  #22628 commented on Aug 30, 2025 • 0 new comments

• [Auto EP] How to set graph optimization level specifically for a registered EP?

  #25593 commented on Aug 30, 2025 • 0 new comments

• [Bug] Models that used to run on onnxruntime 1.16.3 causes AccessViolationException when loading on onnxruntime 1.22.1

  #25588 commented on Aug 30, 2025 • 0 new comments

• [OpenVINO] SessionOptionsAppendExecutionProvider_OpenVINO API loads NULL config file

  #23871 commented on Aug 30, 2025 • 0 new comments

• [Bug] [Performance] Cannot write_calibration_table for per channel quantization calibration

  #25621 commented on Aug 31, 2025 • 0 new comments

• Segmentation Fault running model

  #25613 commented on Aug 31, 2025 • 0 new comments

• [Bug] CUDAExecutionProvider fails to load due to missing libcudnn.so.9 in LD_LIBRARY_PATH when using onnxruntime-gpu==1.22.0

  #25609 commented on Aug 31, 2025 • 0 new comments

• Unable to load gpt-oss onnx model with dotnet

  #25793 commented on Aug 31, 2025 • 0 new comments

• quantize_dynamic causes shape mismatch when model has Gemm with transB = 1

  #25890 commented on Sep 1, 2025 • 0 new comments

• what(): Exception during loading: MLDataType for: tensor(int32) is not currently registered or supported

  #25812 commented on Sep 1, 2025 • 0 new comments

• [Mobile] CoreML Execution Provider not listed in supported values

  #25869 commented on Sep 2, 2025 • 0 new comments

• [Build]

  #13606 commented on Sep 3, 2025 • 0 new comments

• Does BatchNormalization support 2D shape of `X` input

  #25230 commented on Sep 3, 2025 • 0 new comments

• [Feature Request] Python support for AddExternalInitializersFromFilesInMemory

  #25873 commented on Sep 3, 2025 • 0 new comments

• IExecutionProvider::FusedNodeAndGraph set intermediate unused results as model outputs

  #25647 commented on Sep 3, 2025 • 0 new comments

• [Build] cmake cannot find KLEIDIAI - Windows 11 ARM

  #24865 commented on Sep 3, 2025 • 0 new comments

• [Mobile] MatMulNbits Q8 Errors out on Android

  #24769 commented on Sep 4, 2025 • 0 new comments

• Cannot run bfloat16 convolution.

  #25740 commented on Sep 4, 2025 • 0 new comments

• [Mobile] libonnxruntime4j_jni.so 16KB page size compatibility on Android ARM64

  #25859 commented on Sep 4, 2025 • 0 new comments

• [Web] WebGPU first‑run model warm‑up causes long GPU‑blocking operations (maxDispatchNumber & synchronous pipeline creation)

  #25882 commented on Sep 4, 2025 • 0 new comments

• Inference fails with 4 bit quantization

  #25631 commented on Sep 4, 2025 • 0 new comments

• Cannot use Microsoft.ML.OnnxRuntime NuGet package 1.22.1 with Microsoft.SemanticKernel.Connectors.Onnx

  #25287 commented on Sep 4, 2025 • 0 new comments

• Onnx Runtime for Java is packaged with 200MB onnxruntime.pdb in the win-x64 native package

  #12084 commented on Sep 4, 2025 • 0 new comments

• [Feature Request] Windows ARM64 native wheels

  #19161 commented on Sep 4, 2025 • 0 new comments

• [Performance] cudaMemcpyAsync dominates runtime for batch inference with FP32 inputs

  #25852 commented on Sep 5, 2025 • 0 new comments

• [Performance] ONNX FP16 model is having performance bottle neck when compared to FP32 variant

  #25824 commented on Sep 5, 2025 • 0 new comments

• Type mismatch error when loading a Float16 model

  #25522 commented on Sep 5, 2025 • 0 new comments

• [WebGPU] Subgroups feature is not enabled for ort-web WebGPU EP

  #25595 commented on Sep 5, 2025 • 0 new comments