30 Pull requests merged by 23 people

• fix webgpu linux build on gcc 13.3

  #25099 merged Jun 18, 2025

• [WebGPU EP] fixes bugs in NCHW version of instance norm operator

  #25092 merged Jun 18, 2025

• [webgpu] Bump version of Dawn to 9733be39e

  #25071 merged Jun 17, 2025

• Add version check for IAlloactor::GetStats

  #24989 merged Jun 17, 2025

• [webgpu] Add zero points support for dp4 path

  #24675 merged Jun 17, 2025

• vitisai/imp/graph.cc: fix graph_save() for models with external data

  #25082 merged Jun 16, 2025

• Skip topk uint64 test added from ONNX 1.18

  #25084 merged Jun 16, 2025

• Fix CANN build error

  #25075 merged Jun 16, 2025

• [WebGPU EP] adjusts pow op to use sqrt when exponent is 0.5

  #25026 merged Jun 16, 2025

• Fix some OrtStatus* leaks

  #25036 merged Jun 16, 2025

• Add python bindings to the global thread pool functionality

  #24238 merged Jun 16, 2025

• Disable RMSNorm in onnx_backend_test.py and Topk(uint_64)

  #25056 merged Jun 16, 2025

• [webgpu] support intel subgroup matrix on matmul_nbits

  #24898 merged Jun 14, 2025

• [CUDA] Fix clip operator

  #25057 merged Jun 14, 2025

• [webgpu] make DP4AMatMulNBitsSmallMProgram shader template

  #25025 merged Jun 13, 2025

• Revert "Improve Windows ETW callback registration and fix issues"

  #25055 merged Jun 13, 2025

• Cherry pick telemetry changes from win-onnxruntime

  #24957 merged Jun 13, 2025

• [WebGPU-EP] Disable SubgroupMatrix uniformity checks

  #25028 merged Jun 13, 2025

• Make OrtDevice generic to support plugin EP usage.

  #24985 merged Jun 13, 2025

• Add paralell to win-ci.yml

  #25039 merged Jun 12, 2025

• [WebNN] Expose Module.webnnRegisterMLContext to EM_ASM block

  #25022 merged Jun 12, 2025

• Refactor the cmake code of onnxruntime_shared_lib_dlopen_test

  #25031 merged Jun 12, 2025

• [WebNN] Fix bug in getting the first input name of a WebNN op

  #25027 merged Jun 12, 2025

• Support 8 bit weights "unpacked" compute mode in MatmulNBits kernel

  #24959 merged Jun 12, 2025

• Add Paged Attention Op for CUDA SM80 support

  #24595 merged Jun 12, 2025

• Update OpenVINO-ExecutionProvider.md with missing available options

  #25018 merged Jun 12, 2025

• [test] add retry to onnx node test data download

  #25034 merged Jun 12, 2025

• [QNN EP] Add fixing dynamic input shapes in qnn.preprocess

  #25024 merged Jun 12, 2025

• [CUDA] Suppress MSVC build warnings for sm=90

  #25020 merged Jun 12, 2025

• Fix how Whisper tokenizer is loaded and saved

  #25037 merged Jun 12, 2025

34 Pull requests opened by 19 people

• Fix Int4x2/UInt4x2 to MLFloat16 casting compilation errors on ARM64 Linux

  #25040 opened Jun 12, 2025

• update onnx python package to 1.18.0

  #25041 opened Jun 12, 2025

• Use non-CPU device type and id for host accessible memory

  #25043 opened Jun 13, 2025

• fix: BART attention fusion for key with bias🐛

  #25046 opened Jun 13, 2025

• Fix link to Java API documentation

  #25047 opened Jun 13, 2025

• [wasm] Optimize WASM relaxed simd MlasGemmQuantKernel

  #25048 opened Jun 13, 2025

• [MIGraphX EP][ROCm EP] Update build args to include rocm_gfx_arch input for ROCM/MIGraphX EP builds

  #25059 opened Jun 14, 2025

• [WIP] Feature request: Implement GroupNormalization-21

  #25060 opened Jun 14, 2025

• [WIP] Improve DFT implementation

  #25061 opened Jun 14, 2025

• [Feature] Add numpy array protocol and enhanced dlpack support to OrtValue

  #25062 opened Jun 14, 2025

• [WIP] Squeeze node fails when axes is ""

  #25063 opened Jun 14, 2025

• [WIP] `AddCastNode` in `insert_cast_transformer.cc` sets invalid dtype `-1`

  #25064 opened Jun 14, 2025

• [WIP] ORT aborts with the `linspace` implementation when input is empty

  #25065 opened Jun 14, 2025

• [WIP] [tracking] Improve Sequence operator handling

  #25066 opened Jun 14, 2025

• [WIP] ORT aborts on ConcatFromSequence with empty sequence inputs

  #25067 opened Jun 14, 2025

• Fix outer scope initializer type checking by using IsOuterScopeValue

  #25068 opened Jun 14, 2025

• Fix CUDA GatherND batch dimension validation regression

  #25069 opened Jun 14, 2025

• Allocator and data transfer support for plugin EP API

  #25070 opened Jun 14, 2025

• Bump clang-format from 19.1.7 to 20.1.6

  #25073 opened Jun 16, 2025

• [QNN EP] Add QNN EP UDO support

  #25076 opened Jun 16, 2025

• [WebNN] Handle in-memory external data

  #25079 opened Jun 16, 2025

• Bump protobuf from 3.20.2 to 4.25.8 in /onnxruntime/python/tools/transformers/models/llama

  #25085 opened Jun 17, 2025

• Bump protobuf from 3.20.2 to 4.25.8 in /onnxruntime/python/tools/transformers/models/whisper

  #25086 opened Jun 17, 2025

• Bump protobuf from 3.20.3 to 4.25.8 in /onnxruntime/python/tools/transformers/models/stable_diffusion/requirements

  #25087 opened Jun 17, 2025

• Bump protobuf from 4.21.12 to 4.25.8 in /tools/ci_build/github/linux/docker/inference/aarch64/python/cpu/scripts

  #25088 opened Jun 17, 2025

• [QNN-EP] Fix the input order in gen_qnn_ctx_onn_model.py

  #25090 opened Jun 17, 2025

• [webgpu] Make the GQA's intermediate buffer static

  #25091 opened Jun 17, 2025

• [TRT RTX EP] fix build error

  #25094 opened Jun 17, 2025

• [QNN-EP] Complement PoolOpBuilder to support Pool3d.

  #25100 opened Jun 18, 2025

• [QNN EP] Add remove_initializer_from_input in qnn.preprocess

  #25102 opened Jun 18, 2025

• [TRT-EP] Ignore deprecated warnings for TRT APIs

  #25105 opened Jun 18, 2025

• Skip hash computation for EPContext models

  #25106 opened Jun 18, 2025

• [CUDA] FpA IntB Gemm Kernel Test

  #25109 opened Jun 18, 2025

• [DO NOT REVIEW] [MLAS] 8 bit weights ARM64 Matmul implementation

  #25110 opened Jun 18, 2025

9 Issues closed by 9 people

• Is apple M series GPU supported?

  #25042 closed Jun 17, 2025

• [Build] libc++abi: terminating due to uncaught exception of type std::__1::system_error: mutex lock failed: Invalid argument

  #25038 closed Jun 17, 2025

• [CANN] Failed to compile/build onnxruntime

  #25074 closed Jun 17, 2025

• VSCode (mac) debugger breakpoint on GetTensorDataAsSpan leads to app crash evaluating SpanDebugView

  #25023 closed Jun 16, 2025

• Remove reference to internals in torch.onnx

  #11901 closed Jun 14, 2025

• [Build] Tests fail for debug builds with recent VS versions

  #25013 closed Jun 14, 2025

• Prototipo

  #25058 closed Jun 14, 2025

• InferenceSession - Catastrophic Error or Unspecified Error is thrown

  #22815 closed Jun 13, 2025

• onnxruntime and onnx protobuf version confliction

  #12009 closed Jun 12, 2025

21 Issues opened by 18 people

• Initializers use wrong allocator

  #25108 opened Jun 18, 2025

• Is OPSET=23 compatible with which version of Torch CUDA 12.4?

  #25107 opened Jun 18, 2025

• Multi-GPU inference throws error when thread uses sessions allocated on different device

  #25104 opened Jun 18, 2025

• ORT 1.22.0 fails assertion on python import on aarch64, but not x86_64

  #25103 opened Jun 18, 2025

• [Build] Unable to build ONNX Runtime 1.22 due to dependency update

  #25098 opened Jun 17, 2025

• Persistent Crashes on Android/armeabi-v7a

  #25097 opened Jun 17, 2025

• [Web] Cannot import from web worker

  #25096 opened Jun 17, 2025

• All the reduce ops cause an error with None axis and noop_with_empty_axes=1

  #25095 opened Jun 17, 2025

• [Mobile] Please advise: How to use onnxruntime_c/c++_api to determine whether QNN_EP is effective?

  #25093 opened Jun 17, 2025

• [Feature Request] No Support of Graph optimization for Qwen3 Embedding model

  #25083 opened Jun 16, 2025

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

  #25081 opened Jun 16, 2025

• [WebNN EP] Fail to run some models with in-memory external data

  #25078 opened Jun 16, 2025

• [Documentation] wasmPaths object syntax is wrong

  #25072 opened Jun 15, 2025

• [Feature Request] Fp6 datatype support

  #25054 opened Jun 13, 2025

• onnxruntime with CUDAExecutionProvider crashes: gather_nd.cc:30 CheckBatchDimensionsMatch Batch dimensions differ at index 0: 1 != 3, tensor indices: 0, 1

  #25053 opened Jun 13, 2025

• How can I output every node's output shape when infer with onnx models with lots of if branches

  #25052 opened Jun 13, 2025

• different results between the CPUExecutionProvider and CUDAExecutionProvider when onnxruntime executes a valid model

  #25051 opened Jun 13, 2025

• onnxruntime outputs different results for different opset versions

  #25050 opened Jun 13, 2025

• error C2039: "ComputeV2": is not a member of "FurthestSamplingKernel"

  #25049 opened Jun 13, 2025

• Part of QuantizeLinear results have a bias of 1

  #25045 opened Jun 13, 2025

• Test has construct float input and output nodes

  #25044 opened Jun 13, 2025

44 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.

• Add GetCapability/Compile infrastructure for EP ABI

  #24887 commented on Jun 18, 2025 • 129 new comments

• Update config spec

  #24913 commented on Jun 19, 2025 • 10 new comments

• Add RotaryEmbeddings(23) - CPU

  #24980 commented on Jun 19, 2025 • 6 new comments

• [QNN EP] Add Support for Reciprocal Op in QNN EP

  #25035 commented on Jun 18, 2025 • 4 new comments

• Enable VTCM Back Up Buffer Sharing

  #24962 commented on Jun 18, 2025 • 2 new comments

• Add Int4 and UInt4 support for Cast

  #24973 commented on Jun 17, 2025 • 2 new comments

• [webgpu] Enable graph capture

  #24900 commented on Jun 18, 2025 • 1 new comment

• [QNN EP] Add support for Mean Op in QNN EP

  #25021 commented on Jun 12, 2025 • 1 new comment

• Inter session optimization with early release and without default bucket keys

  #24992 commented on Jun 13, 2025 • 0 new comments

• [QNN EP] Fuse DynamicQuantizeLinear + ConvInteger sequence into Conv

  #25003 commented on Jun 17, 2025 • 0 new comments

• [webgpu] Fix test case 'Asin' on Vulkan backend

  #25008 commented on Jun 16, 2025 • 0 new comments

• [QNN EP] MatMul+Add->Gemm fusion when AttentionFusion isn't enabled

  #25017 commented on Jun 16, 2025 • 0 new comments

• Update abseil

  #25033 commented on Jun 18, 2025 • 0 new comments

• Update xnnpack.cmake for WASM build

  #24860 commented on Jun 18, 2025 • 0 new comments

• use WebGPU EP instead of JSEP in WebAssembly

  #24856 commented on Jun 18, 2025 • 0 new comments

• Fix AutoEpSelection and OrtEpLibrary tests when using AuthenticAMD

  #24754 commented on Jun 18, 2025 • 0 new comments

• Enable SME for sgemm and sbgemm through KleidiAI

  #24346 commented on Jun 17, 2025 • 0 new comments

• Add OpenCL EP

  #23830 commented on Jun 16, 2025 • 0 new comments

• fix: ReduceAggregatorMax fill_for_empty_set

  #22385 commented on Jun 18, 2025 • 0 new comments

• [Feature Request] Native WebGPU Execution Provider

  #22077 commented on Jun 18, 2025 • 0 new comments

• [Web] `Error: [WebGPU] Kernel "[Add] /model/model.5/conv1/ffc/convg2g/Add" failed. Error: Can't perform binary op on the given tensors`

  #24744 commented on Jun 18, 2025 • 0 new comments

• Memory allocation failures due to incorrect requested buffer size

  #18743 commented on Jun 18, 2025 • 0 new comments

• Non-zero status code returned while running LSTM node

  #10768 commented on Jun 12, 2025 • 0 new comments

• [Documentation] "Upcoming Release Roadmap" is outdated

  #24722 commented on Jun 12, 2025 • 0 new comments

• [Build] ORT can't build with cuda 12.9

  #24731 commented on Jun 12, 2025 • 0 new comments

• [Documentation] Attention Contrib Op Bias argument Not optional?

  #25005 commented on Jun 12, 2025 • 0 new comments

• [Build] CCCL API migration issue.

  #24774 commented on Jun 12, 2025 • 0 new comments

• [Documentation] outdated documents about cuda version and onnxruntime

  #24759 commented on Jun 14, 2025 • 0 new comments

• [Build] Fail to pass AutoEpSelection and OrtEpLibrary tests in Windows x64 QNN build

  #24676 commented on Jun 14, 2025 • 0 new comments

• Improve DFT implementation

  #24522 commented on Jun 14, 2025 • 0 new comments

• [Feature Request] Make OrtValue compatible with numpy `__array__` and dlpack protocols

  #24071 commented on Jun 14, 2025 • 0 new comments

• Squeeze node fails when axes is ""

  #21661 commented on Jun 14, 2025 • 0 new comments

• `AddCastNode` in `insert_cast_transformer.cc` sets invalid dtype `-1`

  #16449 commented on Jun 14, 2025 • 0 new comments

• [tracking] Improve Sequence operator handling

  #18355 commented on Jun 14, 2025 • 0 new comments

• ORT aborts on ConcatFromSequence with empty sequence inputs

  #16619 commented on Jun 14, 2025 • 0 new comments

• ORT raises node "does not have type information set by parent node" for initializers declared in outer graph

  #24880 commented on Jun 14, 2025 • 0 new comments

• Feature request: Implement GroupNormalization-21

  #24538 commented on Jun 14, 2025 • 0 new comments

• Execution Provider bridge for TFLite Delegates for Coral Edge TPUs

  #10248 commented on Jun 14, 2025 • 0 new comments

• [Build] Cannot cross platform build for v3.20 alpine arm64

  #24788 commented on Jun 15, 2025 • 0 new comments

• why not support ceil() in version 1.12.1

  #24674 commented on Jun 15, 2025 • 0 new comments

• how to release gpu memory when keep onnxruntime session around.

  #9509 commented on Jun 16, 2025 • 0 new comments

• Static quantize self-attention module not work

  #17278 commented on Jun 16, 2025 • 0 new comments

• [Build] Inconsistent naming of lib directories

  #23642 commented on Jun 17, 2025 • 0 new comments

• [Performance]

  #24787 commented on Jun 18, 2025 • 0 new comments