Sync with Microsoft ONNX Runtime - 03/09/2025 #806
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
### Description When using attention bias input for GQA op with FP16, on the platforms that don't natively support FP16 math a cast to fp32 needs to be performed, and thus a temporary buffer needs to be created to store the fp32 values. The issue is that this temporary buffer was being allocated / deallocated inside of a loop for every token being processed. Refactored the implementation so that the allocation takes place only once. Phi model throughput increased by 15%.
### Description This change fixes correctness issues in two areas that were causing failures in onnxruntime_test_all: - DynamicQuantizeMatMul.WithConstantBInputs - AttentionTest.Attention3DDefault - AttentionTest.Attention3DWithPastAndPresentQkMatmul What was wrong and how it’s fixed 1) DynamicQuantizeMatMul.WithConstantBInputs - Root cause: The Kleidi dynamic quantization GEMM path could be selected even when the B scales contained values such as (zero, negative, or non-finite). That violates kernel assumptions and can lead to incorrect results. - Fix: In `onnxruntime/contrib_ops/cpu/quantization/dynamic_quantize_matmul.cc`, we now explicitly validate that all B scales are finite and strictly positive before enabling the Kleidi/MLAS dynamic path. If any scale is invalid, we disable that path. 2) Attention tests (Attention3DDefault, Attention3DWithPastAndPresentQkMatmul) - Root causes in `onnxruntime/core/mlas/lib/kleidiai/sgemm_kleidiai.cpp`: - Incorrect handling of GEMM corner cases for alpha/beta and K==0 (e.g., not respecting C = beta*C when alpha==0 or K==0). - Unnecessary or premature fallbacks for small shapes. - Fixes: - Add early-outs for degenerate sizes: if M==0 or N==0, return handled. - Correctly implement alpha/beta semantics: --------- Signed-off-by: Jonathan Clohessy <[email protected]>
This change adds skip test for QMoE CPU tests when running on TensorRT or CUDA EP. In the QMoE kernel there was a memory overwrite bug in the accumulate part, updated that and this fixed the python tests back
…pi_ (microsoft#25741) ### Description Delay the call to `OrtGetApiBase()` until the first call to `Ort::GetApi()` so that `OrtGetApiBase()` is typically called after dynamic library loading. ### Motivation and Context When ORT_API_MANUAL_INIT is not defined (which is the default), the static `Ort::Global<void>::api_` has a dynamic initializer that calls `OrtGetApiBase()->GetApi(ORT_API_VERSION)` This dynamic initialization can cause problems when it interacts with other global/static initialization. On Windows in particular, it can also cause deadlocks when used in a dynamic library if OrtGetApiBase()->GetApi() attempts to load any other libraries. * Replace the templated `Global<void>::api_` with an inline static initialized to nullptr. * `Ort::GetApi()` now calls `detail::Global::GetApi()` which calls `detail::Global::DefaultInit()` if initialization is needed. * When `ORT_API_MANUAL_INIT` is defined, `DefaultInit()` returns nullptr, which will eventually cause the program to crash. The callers have violated the initialization contract by not calling one of the `Ort::InitApi` overloads. * When `ORT_API_MANUAL_INIT` is not defined, `DefaultInit()` uses a function-level static to compute the result of `OrtGetApiBase()->GetApi(ORT_API_VERSION)` once and return it. * `Ort::Global<void>` has been replaced with a non-templated type and moved inside a `detail` namespace. Since the `Global<void>` object was documented as being used internally, it is believed that these changes here are non-breaking, as they do not impact a public API. The public APIs, `Ort::InitApi()` and `Ort::InitApi(const OrtApi*)` remain unchanged. * Add `#pragma detect_mismatch` to surface issues with compilation units that disagree on how ORT_API_MANUAL_INIT is defined. (MSVC only.) --------- Co-authored-by: Copilot <[email protected]>
### Description <!-- Describe your changes. --> 1. A Small change to use the shared allocator in Python binding. 2. Remove the FP64 support from the EP. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> The Python GPU IO binding is necessary for performance. The change will enable the shared allocator for GPU allocation. The FP64 was using the FP32 inference—aligned WRT TRT RTX support. --------- Co-authored-by: Gaurav Garg <[email protected]>
### Description Add a `enable_cann_subgraph` feature parameter. this parameter controls whether graph splitting is performed and can help quickly identify issues in certain scenarios.
…ting Node_GetTensorAttributeAsOrtValue (microsoft#25886) ### Description Replace `Node_GetTensorAttributeAsOrtValue` with `OpAttr_GetTensorAttributeAsOrtValue`. Change the API signature to make it one of the `OpAttr` interfaces instead of the `OrtNode` interface. The original API was added [here](microsoft#25566).
### Description This change builds on top of microsoft#25841 , and adds the scaffolding necessary to call into this API from C++ / C# / Python. ### Motivation and Context microsoft#25454 talks more about the broader notion of precompiled model compatibility. This change is directed at app developers whose apps may want to determine if a particular precompiled model (e.g. on a server somewhere) is compatible with the device where the application is running. There is functionality in `OrtEpFactory` for making this determination, which was exposed as a C API in microsoft#25841, and this change makes the API more broadly available in other languages. ### Testing and Validation Introduced new unit test cases across each language, and verified that the API was being called and returned the correct result for the default CPU EP. --------- Co-authored-by: Aditya Rastogi <[email protected]>
…25883) ### Description - Introduce Level1 Transformer into qnn.preprocess to support various optimizations. ### Motivation and Context - This change brings in several useful optimizations such as `ConvBnFusion` and `ConstantFolding`, which are part of `TransformerLevel::Level1` and can benefit QNNEP. - The goal is to optimize the ONNX model before quantization by integrating these passes into the Python tooling workflow.
…icrosoft#25887) ### Description Minor fix weight name missing when not valid QDQ node group ### Motivation and Context Some quantized model failed QDQ node group validation, the weights then won't be folded as initializer. QNN EP failed to handle the dynamic weights here due to the transpose op input name look up. This change make sure we process the weights tensor before adding transposes.
## Summary Adds EP metadata library path support to enable custom ops DLL registration with proper path resolution. ## Changes - Added `library_path` metadata key to EP metadata infrastructure - Pass resolved library path directly to `EpLibraryProviderBridge` constructor - Simplified implementation per reviewer feedback (removed virtual method complexity) - Added `#include <utility>` for std::move compliance ## Purpose Enables downstream applications (like onnxruntime-genai) to resolve relative custom ops library paths using EP metadata, improving DLL registration reliability. ## Files Modified - `plugin_ep/ep_factory_provider_bridge.h` - `plugin_ep/ep_library.h` - `plugin_ep/ep_library_plugin.h` - `plugin_ep/ep_library_provider_bridge.cc` - `plugin_ep/ep_library_provider_bridge.h` - `utils.cc`
### Description This update introduces multiple improvements, fixes, and feature enhancements to the OpenVINO Execution Provider (OVEP) and related components in ONNX Runtime: #### Configuration & Properties - Updated load_config mapping to act as a passthrough to OpenVINO properties. - Added support for providing layout information to inputs/outputs in OpenVINO. #### Inference & Tensor Handling - Improved OVInferRequest::SetTensor to correctly handle cached binding shape mismatches. - Added support for self-detecting on-the-fly bfloat16 → float16 conversion. - Fixed issues with input ONNX models when used with shared execution contexts. #### Model Handling & Operator Support - Fixed model copying behavior for QDQ stripping. - Updated operator support status for OpenVINO 2025.2. #### Platform & Integration Fixes - Applied multiple PSU Lora fixes and related updates. - Resolved filename confusion issues with wrapped OVIRs in EPCtx. - Enabled memory-mapped native binaries for OpenVINO 2025.3. #### Quality & Maintenance - Addressed linting issues. - Fixed coverage gaps in OVEP. - Added a new test script for OpenVINO with ORT ABI integration. --------- Co-authored-by: Ankit Maheshkar <[email protected]> Co-authored-by: Ryan Metcalfe <[email protected]> Co-authored-by: Klimenko, Mikhail <[email protected]> Co-authored-by: sfatimar <[email protected]> Co-authored-by: Garth Long <[email protected]> Co-authored-by: Copilot <[email protected]> Co-authored-by: MayureshV1 <[email protected]> Co-authored-by: Eric Crawford <[email protected]> Co-authored-by: jatinwadhwa921 <[email protected]> Co-authored-by: Vishnudas Thaniel S <[email protected]> Co-authored-by: Javier Martinez <[email protected]>
### Description Java API for compile model and EP discovery APIs. Roughly equivalent to the C# version in microsoft#24604. cc: @skottmckay. I haven't quite got the CMake configured so the Java tests for the ep registration only run when the ONNX Runtime shared provider support is built, but everything else works. I expect that to be a quick fix, but I'm not sure in what conditions it should be built and how we should handle it so I don't know where/when to plumb it through. ### Motivation and Context API parity for Java.
### Description 1. Check process exit code when running 7z.exe . Currently the errors were silently ignored. 2. Add snld20 flag to the 7z.exe commands, which is needed to be compatible with the latest 7z release.
…icrosoft#25881) ### Description Fix illegal memory access in GetInputIndices with optional inputs ### Motivation and Context When an input is optional, its ValueInfo may be nullptr. The current implementation directly calls InputValueInfo->GetName(), leading to illegal memory access. Update logic to skip optional inputs when valueInfo is nullptr .
### Description <!-- Describe your changes. --> Re-enable cpuinfo for ARM64EC build and fix `CPUIDINFO_ARCH_ARM` so it is actually used. Patch cpuinfo to support vcpkg ARM64EC build. See pytorch/cpuinfo#324. ### Motivation and Context <!-- - Why is this change required? What problem does it solve? - If it fixes an open issue, please link to the issue here. --> Fix for workaround in microsoft#25831.
…#25879) Restore accidentally removed comments when using WGSL template.
### Description The [vfmaq_f32](https://developer.arm.com/architectures/instruction-sets/intrinsics/vfmaq_f32) intrinsic compiles to the [FMLA](https://developer.arm.com/documentation/ddi0596/2021-03/SIMD-FP-Instructions/FMLA--vector---Floating-point-fused-Multiply-Add-to-accumulator--vector--?lang=en) instruction which is more performant than separate `fmul`+`fadd` instructions that [vmlaq_f32](https://developer.arm.com/architectures/instruction-sets/intrinsics/vmlaq_f32) compiles to on latest GCC versions: https://godbolt.org/z/aYc9as5Wh Note that this is not a breaking change, as vmlaq_f32 compiles to FMLA instructions already on the latest clang compilers (which are the default for MacOS ORT builds already) ### Motivation and Context With this change, the NEON version of `MlasMultiplyAddFloat32x4` achieves parity with the x86 version that uses `_mm_fmadd_ps`. It also achieves up to ~15% speedups compared to the current `vmlaq_f32` implementation when tested on top of microsoft#25580
### Description Until QAIRT 2.37.0, `QNN_IR_GRAPH_SERIALIZATION_OPTION_INIT` was unusable due to a missing semicolon. Now that it's been fixed, revert the workaround.
This PR adds a missing sync method and fixes the linux CI
### Description Change from fread to mmap to save on system memory. This also accelerated the load time of a ~4GB model in my testing by 1.5X.
### Description Runtime caches can accelerate the JIT time when deserializing an engine of TRT RTX. Here we introduce a per engine caching in a user specified folder. The cache file will be named after the fused node name - which will also be the node name of an ep context node. @chilo-ms we would like to pick this to 1.23
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
Description
Synchronizing intel/onnxruntime ovep-develop branch with latest changes from microsoft/onnxruntime master branch.