In March 2026, delivered Grouped Query Attention (GQA) support in FlashAttentionTile for the aobolensk/openvino repo, enabling larger and more configurable attention patterns on NPU-backed inference. Key improvements include shape inference for GQA when numQHeads > numKVHeads and numQHeads % numKVHeads == 0, as well as relaxedRestrictions on the fp32 running sum tensor and added fp16 evaluate function support. The changes are captured in commit f42ebd314ff8741eda1355cc4c5937c2e163703c (NPU) and aligned with ticket #34929, validated on an NPU branch. Business value: enables more flexible transformer attention workloads with improved potential throughput on supported hardware, while preserving FP32/FP16 compatibility and reducing configuration overhead for deployment.

January 2026 performance summary for openvinotoolkit/openvino focused on API cleanup and maintainability for FlashAttentionTile. Delivered API-level simplification by removing the scale input, reducing surface area and potentially improving runtime efficiency. The work was performed with cross-team collaboration (NPU-focused) and aligns with ongoing performance and usability goals. No major bugs fixed this period; emphasis was on design/implementation improvements and code hygiene.

Monthly performance summary for openvinotoolkit/openvino (Dec 2025). Primary focus: feature delivery enabling efficient CPU-based attention computations via a standalone static library, with CPU FP32 support and verifiable reference evaluation. This aligns with performance, scalability, and reusable-build goals for openvino across models and deployments.