January 2026 monthly summary for CodeLinaro/onnxruntime: Delivered targeted quantization and translation work in the QNN execution provider to improve model efficiency and GPU compatibility for large models. Key features: Case-2 LPBQ support for Gemm and Matmul fusion with optional QuantizeLinear nodes to reduce model size while preserving performance; translation of MatMulNBits contrib op to QNN FullyConnected with INT4 BlockQuantized weights to broaden GPU support for LLM workloads. No major bugs fixed were documented for this period.

November 2025 monthly summary for ROCm/onnxruntime focusing on performance optimization through backend-aware graph fusion. Delivered a QNN Gelu fusion which collapses the Gelu pattern into a single QNN Gelu node, eliminating the need to decompose Gelu into Div, Erf, Add, and Mul across EP boundaries. This change reduces graph partitioning and cross-engine data movement, leading to faster inference for Gelu-heavy models and better hardware utilization on the QNN backend.

Month: 2025-09. Concise monthly summary focusing on ONNX Runtime performance enhancements and reliability. This period centers on delivering a tangible runtime tuning capability for the QNN Execution Provider, alongside clear documentation of impact and capabilities.

Concise monthly summary for 2025-08 focusing on ONNX Runtime work in the QNN Execution Provider. Delivered feature: ONNX ScatterElements support in the QNN EP, including handling of various reduction types and integration with existing optimization and testing frameworks. Commit included: f755b8a8f4e225a09c2c4076f217e8c62bcbe895 ("[QNN EP] Add ONNX ScatterElements support (#24811)").

July 2025: Delivered Low Power Block Quantization (LPBQ) pathway for Gemm on the NPU backend in ONNX Runtime, enabling more energy-efficient, accuracy-sensitive inference. Added LPBQ encoding support for the MatMul operator in the QNN EP, broadening quantization coverage for NPU-backed models. Implemented unit tests to guard LPBQ fusions and prevent regressions. These changes are captured in commits 91e91186aa8ab67da4785e24c69e11303ddaa25d, ecc358f069488a79c5abc16c5ddfbc4bd5b3c771, and 5c0a7d81c0b812e7209e7555246aafa9aaaf433c.

June 2025 monthly summary for microsoft/onnxruntime focusing on QNN EP milestones and business impact.

In May 2025, focused on strengthening the QNN Execution Provider in mozilla/onnxruntime by expanding operator support, fixing critical quantization edge-cases, and broadening support for scatter operations. These changes improve inference performance, compatibility with QDQ ONNX models, and overall reliability for production workloads.

April 2025 – mozilla/onnxruntime (QNN EP). Key feature delivered: Expand Op now accepts INT64 shape inputs by converting to INT32, with unit tests validating the behavior. The change was implemented under PR #24389 and committed as f7028a3a087bef85daf204fa65b53d714011ad0b. Impact: extends operator compatibility for models using 64-bit shapes, reduces shape-related runtime errors, and improves reliability of QNN EP workloads. This work contributes to broader deployment readiness and model interoperability across ONNX Runtime.