April 2026 (2026-04) monthly summary for microsoft/onnxruntime. Key stabilization effort centered on the VitisAI Execution Provider: corrected a typo in the provider factory from 'external_ep_libray' to 'external_ep_library', ensuring proper execution provider identification and utilization. This fix reduces runtime misconfigurations and enhances reliability of the VitisAI integration for hardware-accelerated inference. No new features were shipped this month; the primary impact comes from a targeted bug fix that improves robustness and downstream deployment stability for VitisAI workloads. Technologies demonstrated include provider integration debugging, code hygiene in C++/Python components, and cross-repo collaboration with the VitisAI integration effort.

March 2026: Delivered Windows ML compatibility fix for microsoft/onnxruntime-genai by correctly initializing ep_path_, ensuring reliable startup and runtime behavior with Windows ML. The change enables GenAI workloads to run more reliably on Windows platforms. Commit 639dcc06fbffff14c06612b64d3f7773a7b98a2f as part of the RyzenAI WinML compatibility fix (#2026).

February 2026 monthly update for microsoft/onnxruntime-genai: Implemented enabling the VitisAI external execution provider in Windows Machine Learning (WinML), restoring hardware-accelerated inference path and improving Windows compatibility for GenAI workloads.

January 2026: Delivered major hardware acceleration and external-provider capabilities for ONNX Runtime via RyzenAI and VitisAI integrations. Implemented RyzenAI Execution Provider across platforms with Linux compatibility improvements and ensured backward compatibility for older RyzenAI models. Added VitisAI external provider loader to enable external provider usage, and enhanced benchmarking capabilities with dynamic sequence length configuration, improved prompt generation, and CLI controls. These developments expand hardware options, improve performance and reliability, and empower customers with greater control over model execution and benchmarking.