LiteRT: Maximum performance, simplified- Google Developers Blog

Introduces LiteRT's MLDrift for enhanced GPU acceleration with optimized tensor layouts, workgroup optimization, and improved data handling. Adds NPU support for MediaTek and Qualcomm, abstracting vendor SDKs and providing unified model/runtime distribution via AI Packs and Feature Delivery. Simplifies hardware acceleration setup with a new CompiledModel API. Enhances performance with seamless buffer interoperability (TensorBuffer) for zero-copy data transfers between hardware memory types and asynchronous execution using OS-level mechanisms (e.g., EGL sync fences) to minimize latency and power consumption by overlapping computations across CPU, GPU, and NPUs.