You can use a GPU with docker - at least on some platforms. There's more setup though, nvidia have some details to help https://docs.nvidia.com/datacenter/cloud-native/container-to...
Thank you. I should have specified on MacOS. I ran into this recently trying to setup stable-diffusion-webui/InvokeAI/Foocus and finding it much more complicated to get working for me on my personal laptop than the llms.
Out of curiosity, before you attempted this, what was your impression of the fitness and performance of Macs for generative AI?
Before I attempted, I had no idea. I hadn't ran any AI models locally and I don't follow this stuff too closely, so I wasn't even sure if I could get something usable on my M1 MacBook Air. I went in fairly blind which is why the Ollama Docker installer was so appealing to me–I got to hold off fighting Python and Homebrew until I had a better sense of what the tool could provide.
After my attempt, I think chat is performant enough on my M1. Code gen was too slow for me. Image generation was 1-2 minutes for small pixel art sprites, which for my use case is fine to let churn for a while, but the image generation results were much worse than ChatGPT browser gives me out of the box. I do not know if poor image quality is due to machine constraints or me not understanding how to configure the checkpoint and models.
I would be interested to hear how an M3 or M4 Mini handles these things as those are fair affordable to pick up used.