ML Drift: Next-Gen GPU AI/ML Inference at the Edge- Google Developers Blog

Introduces ML Drift, an open-source, cross-platform, on-device GPU compute engine for AI/ML inference. ML Drift abstracts hardware complexities (OpenGL ES, OpenCL, Metal, WebGPU) to provide a unified foundation for peak performance across diverse edge devices. Key architectural changes include tensor virtualization for unified shaders, an extensible custom op framework with agentic SKILL.md guide, 5D tensor support, performance upgrades for classical models, and stage-aware optimizations for edge LLMs (KV cache prefill and decode stages). It also enables desktop previews via native WebGPU compilation. ML Drift is battle-tested in production across Google products (Chrome, YouTube Shorts, Photos, Meet) and adopted by partners (Adobe Lightroom/Photoshop, Snap), demonstrating significant latency reductions and memory footprint improvements. It is co-optimized with hardware partners (Arm, Intel, Qualcomm).