> GPU in other ways, such as training ML models and then using them via an inference engine all powered by your local GPU?

Have a look at wonnix https://github.com/webonnx/wonnx

A WebGPU-accelerated ONNX inference run-time written 100% in Rust, ready for native and the web