I'm trying to set up Docker for a Python project on my Mac and want to use MPS (Metal Performance Shaders) for GPU acceleration with PyTorch inside the container.

What I have tried so far:

• I installed Docker Desktop (the latest version of Apple Silicon) on MacOS.
• Created a Dockerfile using a base image like python:3.11-slim.
• Installed PyTorch inside the container and checked for MPS availability with:

import torch
print(torch.backends.mps.is_available())

This returns False, meaning MPS is not detected inside the container.

• Running the same code outside Docker on macOS (bare metal), MPS is available (True).

My questions:

• Does Docker on macOS support GPU passthrough for MPS acceleration?

• Are there any specific Docker settings or configurations to enable GPU access inside the container?

• If GPU passthrough is not supported, are there any workarounds to use MPS with PyTorch inside Docker?

Any guidance or workarounds would be greatly appreciated. Thanks!