August 2025 (2025-08) — intel/onnxruntime: Performance-focused feature delivery on the Phi model via GQA attention bias optimization for FP16. Implemented pre-allocation of a buffer for attention masks to reduce memory allocation overhead, achieving ~15% throughput improvement for Phi model. This work was delivered in the CPU FP16 path and committed under [CPU] Optimize GQA attention bias application for FP16.

Monthly performance summary for 2025-07 focused on delivering a targeted feature for ROCm/onnxruntime with improvements in observability and test coverage, plus alignment with business value.

March 2025 ROCm/onnxruntime monthly performance summary focusing on feature delivery and operational impact. Key accomplishment centers on enhancements to the Group Query Attention (GQA) CPU operator to support custom position IDs and attention bias for speculative decoding, accompanied by a new element-wise addition kernel for applying attention bias and updates to input handling. These changes enable more flexible and accurate speculative decoding workflows in PhiSilica and set the stage for production-grade decoding pipelines. No major bugs were reported in this period for the ROCm/onnxruntime repo; stability and maintainability were maintained.