November 2025 (2025-11) monthly summary for intel/onnxruntime: API surface expansion focused on custom operator configuration access and extensibility. Highlights: - KernelInfo API: Introduced KernelInfo_GetConfigEntries to expose session configuration entries to custom operators during kernel construction. This enables operators to adapt behavior based on session options without external signaling, improving flexibility and performance tuning. - Cross-language API surface: Added C API addition to OrtApi (KernelInfo_GetConfigEntries) and corresponding C++ wrapper (GetConfigEntries() in KernelInfoImpl), aligning with existing ORT API patterns and maintaining backward compatibility (API Version 1.24). - Core implementation: Implemented in onnxruntime/core/session/custom_ops.cc; iterates config_options_map from OpKernelInfo to produce OrtKeyValuePairs that are returned to the kernel at construction time. - Testing and validation: Updated tests (shape_inference_test.cc) to verify retrieval of config entries within custom kernel constructors, including checks for existing and non-existing keys. - Stability and documentation: All changes are backward compatible and documented within the API; no breaking changes introduced. Impact: - Business value: Enables customization and optimization of custom operators by making session-level configuration accessible at kernel construction time, reducing overhead and enabling behavior that is responsive to deployment-time options. - Technical accomplishments: Clean API design across C and C++ boundaries, robust integration into the OrtKernelInfo flow, and targeted test coverage to prevent regressions. Files touched (high level): includes/onnxruntime/core/session/onnxruntime_c_api.h, include/onnxruntime/core/session/onnxruntime_cxx_api.h, onnxruntime/core/session/custom_ops.cc, and test/framework/shape_inference_test.cc. Commit reference: - 5834bfe8782ccce88558753343216b89a3ec84d7: "Add API to access config entries from KernelInfo (#26589)" with detailed changes and motivation.

October 2025: Delivered a targeted bug fix to OrtShapeInferContext for optional inputs during ONNX Runtime graph optimization, with a focus on VitisAI deployment. The change uses an undefined data type and an empty shape as placeholders to correctly construct OrtShapeInferContext when an optional input is present but lacks a type. This improves graph optimization reliability and may benefit other execution providers by ensuring consistent inference shape handling across providers.