Added support for MPS on Apple silicon

[rampadc]

Tested on Macbook Pro M1 Max.

• Added new virtual environment setup for MPS (start-mps_mac.sh)
• Updated device detection logic to properly handle MPS devices
• Enabled MPS fallback for unsupported operations

Notes

• Only the aten::angle operation falls back to CPU, all other operations utilize MPS
• Environment variables set automatically in startup script:
  • DEVICE_TYPE=mps
  • PYTORCH_ENABLE_MPS_FALLBACK=1

[fireblade2534]

would it be posible to also add support in the form of a docker container?

[rampadc]

@fireblade2534 I'm not sure, maybe, but that will require a separate PR.

Docker on Mac doesn't support GPU, but Podman does, through Vulkan GPU APIs.

The only Pytorch reference for Vulkan I can find is https://pytorch.org/executorch/stable/native-delegates-executorch-vulkan-delegate.html

Or maybe Vulkan on onnx: microsoft/onnxruntime#21917

[fireblade2534]

ah ok

[fireblade2534]

I don't have a mac so I can't test this but it looks good

[remsky]

Looks great, will run a few checks when I'm at my Mac too, and see if I can bring a docker container around

[remsky]

Script looks good