I use a ZFS pool for my data and a combination of LXCs and Docker in an LXC to run my services. Proxmox is flexible enough that you can pretty much get any configuration to work. I even have my Intel iGPU passthrough set up to my Docker LXC to my jellyfin docker container. Caddy and Authelia are on one LXC for reverse proxy and authentication, and I can point it to my Docker LXC by its IP address. I use bind mounts to mount folders on my ZFS pool to the LXCs/Docker.

One advantage of Caddy running in Docker is you can use the caddy-docker-proxy module to automatically generate a Caddyfile from Docker labels of your containers.

I started my ZFS pool from scratch with new hard drives. If you want to reuse your existing ones without wiping your data, you may want to look in to MergerFS.

The best consumer NVIDIA card is the 3090ti because of its 24GB memory, so you can run bigger LLM models. I have a 3060ti 12GB which works pretty well with 7B and 13B LLM models.