Support native macOS (Apple Silicon) execution without Docker

[davidfowl]

Right now, running Ollama via Docker on macOS doesn’t use the GPU because Docker can’t access Apple Silicon GPUs. That’s a pretty big performance hit, especially when working with larger models.

This blog post lays it out clearly: https://chariotsolutions.com/blog/post/apple-silicon-gpus-docker-and-ollama-pick-two/

When you run Ollama natively, it detects and uses the Metal backend:

level=INFO msg="Dynamic LLM libraries [metal]"

But in Docker:

level=INFO msg="no GPU detected"

It’d be great if Ollama could run natively on macOS without Docker, so we can take advantage of Apple’s GPU acceleration. Performance and dev experience would be much better.

[aaronpowell]

Would you expect it to download/install Ollama if it's not already installed?

Can you specify the port of the Ollama host via the CLI? That'd be important for ensuring you don't have port conflicts on startup.

A current workaround would be to use the ConnectionStringResource and provide the endpoint for Ollama run from outside of Aspire.