May 2026 — Monthly work summary for Xilinx/onnx-mlir: Strengthened robustness of the shape-inference path for SplitToSlice by adding explicit rank validation. This prevents crashes when shape inference is not possible and improves stability for dynamic models in production pipelines.

Month: 2026-04. Delivered robustness enhancements to MicrosoftGroupQueryAttention in Xilinx/onnx-mlir, focusing on input validation for pastKey and improved handling of rotary embeddings. Implemented static shape enforcement for pastKey before modifications and expanded test coverage for do_rotary=1 under varied inputs to reduce edge-case failures. Reordered match checks and rewritten related modifications to improve maintainability and reliability. This work strengthens model robustness in real-world workloads, reducing risk of shape/misalignment errors in transformer-style attention and improving test coverage for rotary embeddings.