Running Modular MAX on a Canadian GPU VPS

I’m planning to serve an open model with MAX for a client whose data has to stay in Canada. PIPEDA, BC public body rules and Quebec Law 25 all push us toward keeping prompts, outputs and logs on Canadian servers, so we’re looking at a GPU VPS here instead of a hosted inference API in a US region. I’d like a sanity check from people already running MAX themselves.

1. GPU choice. For an 8B model with moderate traffic, is a single 24 GB card enough, or should I plan for 48 GB?

2. Install. Docker container vs installing MAX with pixi or pip on the host: which has been easier to upgrade?

3. Drivers. Any NVIDIA driver or CUDA version gotchas on Ubuntu 24.04?

4. Serving. Do you put the OpenAI-compatible endpoint behind Nginx with TLS and an API key, or something else?

5. Quantisation. Have you seen a real quality drop with the quantised weights for chat use?

6. CPU fallback. Is CPU-only serving usable for a small internal tool while we wait on a GPU node?

7. Monitoring. What metrics do you watch for latency and GPU memory?

What would you change in this setup, and what surprised you the first time you ran MAX on your own server?