gpu-pass-through/passthrough.md

GPU Passthrough Status for All Your Containers

Already Fully Configured ✅ — No Changes Needed

CT 120 (nvidia-gpu-cuda) and CT 122 (ollama) already have complete GPU passthrough config. Nothing to do there.

---

Containers That Need GPU Added

Based on your workloads, only CT 108 (openclaw) is the AI container that needs GPU. CT 130 (video.takeone.bh) could optionally benefit for hardware video encoding.

---

Exact Lines to Add

CT 108 — openclaw (AI assistant)

Run on the host:

cat >> /etc/pve/lxc/108.conf << 'EOF'

# GPU Passthrough
lxc.cgroup2.devices.allow: c 195:* rwm
lxc.cgroup2.devices.allow: c 510:* rwm
lxc.cgroup2.devices.allow: c 235:* rwm
lxc.mount.entry: /dev/nvidia0 dev/nvidia0 none bind,optional,create=file
lxc.mount.entry: /dev/nvidia1 dev/nvidia1 none bind,optional,create=file
lxc.mount.entry: /dev/nvidiactl dev/nvidiactl none bind,optional,create=file
lxc.mount.entry: /dev/nvidia-modeset dev/nvidia-modeset 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-caps dev/nvidia-caps none bind,optional,create=dir
lxc.environment: NVIDIA_VISIBLE_DEVICES=all
lxc.environment: NVIDIA_DRIVER_CAPABILITIES=compute,utility
lxc.environment: CUDA_DEVICE_ORDER=PCI_BUS_ID
EOF

---

CT 130 — video.takeone.bh (optional, for hardware video encoding)

cat >> /etc/pve/lxc/130.conf << 'EOF'

# GPU Passthrough (hardware encoding)
lxc.cgroup2.devices.allow: c 195:* rwm
lxc.cgroup2.devices.allow: c 510:* rwm
lxc.cgroup2.devices.allow: c 235:* rwm
lxc.mount.entry: /dev/nvidia0 dev/nvidia0 none bind,optional,create=file
lxc.mount.entry: /dev/nvidia1 dev/nvidia1 none bind,optional,create=file
lxc.mount.entry: /dev/nvidiactl dev/nvidiactl none bind,optional,create=file
lxc.mount.entry: /dev/nvidia-modeset dev/nvidia-modeset 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-caps dev/nvidia-caps none bind,optional,create=dir
lxc.environment: NVIDIA_VISIBLE_DEVICES=all
lxc.environment: NVIDIA_DRIVER_CAPABILITIES=compute,utility,video,encode,decode
lxc.environment: CUDA_DEVICE_ORDER=PCI_BUS_ID
EOF

Note: CT 130 has unprivileged: 1 — the bind mounts will still work because of the optional flag.

---

What to Do Inside Each Container After Restart

Once you restart the container, install the matching NVIDIA userspace libs only (no kernel modules) inside it:

# Inside CT 108 or 130 (Debian/Ubuntu)
# Add the same CUDA repo from the host, then:
apt install -y libnvidia-compute-580 nvidia-utils-580
# Verify:
nvidia-smi

Or if the container already has the CUDA local repo mounted/copied from the host, just run:

nvidia-smi   # should show both RTX 3060s

---

Summary Table

CT	Name	GPU Needed	Status	Action
120	nvidia-gpu-cuda	✅ Yes	Already done	Nothing
122	ollama	✅ Yes	Already done	Nothing
108	openclaw	✅ Yes	Missing	Add config above
130	video.takeone.bh	Optional	Missing	Add config above if you want HW encoding
All others	web/DB/network	❌ No	N/A	Skip

The cgroup device majors to remember for this host: 195:* (GPU cores), 510:* (UVM/CUDA), 235:* (nvidia-caps/MIG).