Model formats and backends for best support.

[rymuelle]

Currently, torch traces are used, as torch offers easy access to CUDA/MPS/CPU/ROCm backends, however, there are two main disadvantages:

1. It requires a large dependency in torch
2. It may limit use on windows AMD devices.

ONNX is a natural alternative format, and is commonly used to distribute weights. However, it has limited support on mac silicone (and AMD). I will use this thread to brainstorm possibilities, and welcome any input.