Google has released ML Drift, the GPU acceleration engine used by LiteRT, as an Apache 2.0 open-source project. Developers can use it to build inference and graphics runtimes across different edge devices without writing a separate implementation for every low-level GPU interface.
The engine supports OpenGL ES, OpenCL, Apple’s Metal and WebGPU. It uses tensor virtualization to provide a common shader model and includes extensible custom operations and five-dimensional tensor support. Google has also added optimizations for GPUs from Arm, Intel and Qualcomm.
That portability addresses a persistent problem in on-device AI: an application may encounter many chip architectures, driver versions and APIs after deployment. ML Drift is intended to hide part of that variation while still allowing hardware-aware optimization. Potential uses include computer vision, audio and depth processing as well as generative models that run locally.
Open sourcing the engine gives teams access to the layer beneath LiteRT rather than limiting them to Google’s complete runtime. It does not make performance identical across devices, however. Drivers, memory limits and vendor-specific behavior still affect latency and supported operations. Developers considering ML Drift will need to test on the actual hardware they support, but they can now do that from a shared code base instead of maintaining a separate GPU path for each platform.