3090 Passthrough Runbook

1. Check what the host already has

   where: proxmox host, as root

   If nvidia-smi already works here, skip to step 3. Also confirm neither card is currently passed through to a VM — a GPU bound to vfio-pci for a VM cannot also serve the host.

   # is a driver already loaded?
   nvidia-smi
   
   # which driver owns each card — look for "Kernel driver in use"
   lspci -nnk -d 10de:
   
   # anything bound to vfio-pci is reserved for a VM
   lsmod | grep -E 'nvidia|vfio'

   Copy

2. Install the driver on the host

   where: proxmox host, as root

   Proxmox is Debian, but it uses its own kernel headers package. Take the driver version from NVIDIA's site and keep a note of it — the passthrough route needs the identical version inside the container.

   apt update && apt install -y pve-headers-$(uname -r) build-essential dkms
   
   # stop nouveau claiming the card, then reboot
   echo -e 'blacklist nouveau\noptions nouveau modeset=0' > /etc/modprobe.d/blacklist-nouveau.conf
   update-initramfs -u && reboot
   
   # after the reboot — substitute the version you downloaded
   chmod +x NVIDIA-Linux-x86_64-<VERSION>.run
   ./NVIDIA-Linux-x86_64-<VERSION>.run --dkms
   
   nvidia-smi   # should now list the 3090

   Copy
   Two cards, one decision

   This box has a 3090 and a Quadro RTX 4000. The driver claims both. If the Quadro is doing something else — a VM, display output — check that first, because installing on the host will take it over.

3. Make the device nodes appear on every boot

   where: proxmox host, as root

   The nodes are created lazily by the first CUDA process, which is no good for a container that expects them at start. Load the modules at boot and keep the driver resident.

   echo -e 'nvidia\nnvidia_uvm\nnvidia_modeset' > /etc/modules-load.d/nvidia.conf
   systemctl enable --now nvidia-persistenced
   
   ls -l /dev/nvidia*   # nvidia0, nvidiactl, nvidia-uvm, nvidia-modeset

   Copy

4. Read the real device majors

   where: proxmox host, as root

   Do not copy these numbers from a forum post. nvidia is conventionally 195, but nvidia-uvm is allocated dynamically and differs between machines and reboots-after-upgrade. Read them here and use what you get in the next step.

   grep -E 'nvidia' /proc/devices
   # e.g.  195 nvidia   511 nvidia-uvm   234 nvidia-caps

   Copy

5. Bind the GPU into CT 211

   where: proxmox host, as root · passthrough route only

   Append to /etc/pve/lxc/211.conf, substituting the majors from step 4 for the 195 and 511 below.

   lxc.cgroup2.devices.allow: c 195:* rwm
   lxc.cgroup2.devices.allow: c 511:* rwm
   lxc.mount.entry: /dev/nvidia0 dev/nvidia0 none bind,optional,create=file
   lxc.mount.entry: /dev/nvidiactl dev/nvidiactl none bind,optional,create=file
   lxc.mount.entry: /dev/nvidia-uvm dev/nvidia-uvm none bind,optional,create=file
   lxc.mount.entry: /dev/nvidia-uvm-tools dev/nvidia-uvm-tools none bind,optional,create=file
   lxc.mount.entry: /dev/nvidia-modeset dev/nvidia-modeset none bind,optional,create=file

   Copy

   Then give the container somewhere to put 20 GB of weights, and restart it:

   # either grow the root disk…
   pct resize 211 rootfs +60G
   
   # …or mount host storage in, which keeps models off the container disk
   pct set 211 -mp0 /rpool/models,mp=/opt/models
   
   pct stop 211 && pct start 211

   Copy

6. Install the matching driver inside the container

   where: inside CT 211, as root · passthrough route only

   Same version as the host, and crucially --no-kernel-module: the container shares the host kernel and must not try to build its own.

   ./NVIDIA-Linux-x86_64-<SAME-VERSION>.run --no-kernel-module
   
   nvidia-smi   # the 3090, from inside the container

   Copy
   Version skew is the failure mode

   If the host driver updates and the container's does not, nvidia-smi inside starts failing with a mismatch error and every CUDA call dies. Pin both, or use the serve-over-HTTP route and never think about it again.

7. Run the model

   where: wherever the GPU now works

   Ollama is the least ceremony. Point its model directory at the big disk, expose it on the LAN if this container is going to call it, then pull the model from the next section.

   curl -fsSL https://ollama.com/install.sh | sh
   
   # store weights off the root disk, and listen beyond localhost
   mkdir -p /etc/systemd/system/ollama.service.d
   cat > /etc/systemd/system/ollama.service.d/override.conf <<'EOF'
   [Service]
   Environment="OLLAMA_MODELS=/opt/models"
   Environment="OLLAMA_HOST=0.0.0.0:11434"
   EOF
   systemctl daemon-reload && systemctl restart ollama
   
   ollama pull devstral:24b
   ollama run devstral:24b "write a bash one-liner that lists open ports"

   Copy
   Then tell me the address

   Once it answers, give me the endpoint — http://192.168.40.x:11434 — and I can call it from this session over HTTP without any of the above being true in here.

• My pick · agentic work

  Devstral 24B

  ~14 GB

  58% of 24 GB · ~10 GB left for context

  Built for tool-use and multi-step editing rather than autocomplete, and the headroom is the point: it is the one that still behaves when the context gets long. ollama pull devstral:24b

• Stronger coder, tighter fit

  Qwen3-Coder 30B-A3B

  ~19 GB

  79% of 24 GB · ~5 GB left for context

  A mixture-of-experts model with roughly 3.3B active parameters, so it runs far faster than 30B suggests. Better raw code, but 5 GB of cache runs out quickly in agent loops. ollama pull qwen3-coder:30b

• Verify before trusting

  Laguna XS 2.1

  ~20 GB

  83% of 24 GB · ~4 GB left for context

  Surfaced in current round-ups as a 33B MoE claiming 70.9% on SWE-Bench Verified, which would beat both of the above outright. That is past what I can personally vouch for — check the number holds before building on it.

Claude Code speaks the Anthropic API, and Ollama does not. Bridging them means putting an Anthropic-compatible proxy in front — LiteLLM is the usual choice — and setting ANTHROPIC_BASE_URL at it. It works, it is not an officially supported configuration, and you should expect to debug tool-call formatting rather than have it work first try.