Release ONNX Runtime QNN Execution Provider v2.2.0 · onnxruntime/onnxruntime-qnn

This release delivers operator coverage improvements, multi-NPU device selection, and build fixes.

ONNX Runtime Compatibility: >= 1.24.1 (compiled with v1.24.4)

QAIRT SDK Compatibility: 2.46.0

pip install onnxruntime==1.24.4
pip install onnxruntime-qnn==2.2.0

Bug Fixes

QNN EP: Fixed GlobalMaxPool/GlobalAveragePool falsely claiming rank-3 support; unified the 3D→4D reshape path with windowed pool ops. (#201)
QNN EP: Restored Genie builds against QAIRT SDKs older than 2.45.0 by keying conditional compilation off the Genie API version (GenieDlc.h breaking change). (#225)
QNN EP: Fixed GCC 13 build failures: corrected memory_order_acq_rel on std::atomic::store() to memory_order_release, and suppressed a false-positive -Wmaybe-uninitialized in TestInputDef. (#228)
QNN EP: Fixed HNRD Model Compatibility checks incorrectly running on x86 platforms where they don't apply. (#319)

QNN EP: Relaxed QDQ BatchNormalization selector to accept BN nodes with 2 dequantized inputs (instead of requiring 3), matching the common pattern where bias/mean/variance stay as float initializers. Reduces CPU fallback and graph fragmentation. (#209)
QNN EP: NPU device selection now supports HTP cores with non-zero device IDs. (#215)

WoS AMD64 — Python 3.11 installer issue causes inference failure — On Windows on Snapdragon, ep.get_library_path() returns the amd64 folder path instead of arm64ec, causing inference to fail in the AMD64 Python 3.11 environment, due to a known issue with the installer. As a workaround, manually construct the path to the arm64ec library. This issue affects Python 3.11 only.

Package	Windows ARM64	Windows x64	Linux ARM64
Python Wheel	Inference	AOT compilation + Inference	Inference
NuGet	Inference	—	—
ZIP	Inference	—	—
tgz	—	—	Inference

Full Changelog: rel-2.1.0...rel/ort-qnn-ep/2.2.0

This release includes contributions from: