After spending way too much time fighting NVIDIA driver issues, DKMS errors, and Proxmox kernel issues, I finally got a clean, reliable setup working for older NVIDIA GPUs with LXC passthrough.
This setup works well for:
- GTX 1060 / 1070 / 1080
- Quadro P4000 / P5000
- Tesla V100
- Most Pascal / Volta cards
Use cases:
- Ollama
- CUDA
- Docker
- Plex transcoding
- Local LLMs
- AI inference
- GPU acceleration inside LXC containers
IMPORTANT NOTES
You do NOT need:
- NVIDIA vGPU
- enterprise licensing
- pve-nvidia-vgpu-helper
- patched kernels
- weird passthrough scripts
Also:
- Avoid Linux Kernel 7.x right now with older Pascal/Volta GPUs
- Many users are hitting DKMS compile failures on kernel 7
- Kernel 6.x is MUCH more stable right now
Recommended:
- Proxmox kernel 6.8 or 6.14
- NVIDIA 550 or 575 drivers
HOST INSTALLATION
Install required packages:
Code:
apt update
apt install -y build-essential dkms pve-headers-$(uname -r)Download NVIDIA driver (550 stable branch):
Code:
wget https://us.download.nvidia.com/XFree86/Linux-x86_64/550.163.01/NVIDIA-Linux-x86_64-550.163.01.runAlternative newer driver (575 branch):
Code:
wget https://us.download.nvidia.com/tesla/575.57.08/NVIDIA-Linux-x86_64-575.57.08.runMake executable:
Code:
chmod +x NVIDIA-Linux-x86_64-550.163.01.runOR for 575:
Code:
chmod +x NVIDIA-Linux-x86_64-575.57.08.runInstall driver:
Code:
./NVIDIA-Linux-x86_64-550.163.01.runOR:
Code:
./NVIDIA-Linux-x86_64-575.57.08.runInstaller recommendations:
- YES to DKMS
- NO to OpenGL libraries if headless
Reboot:
Verify:
You should now see the GPU on the Proxmox host.
LXC GPU PASSTHROUGH
Edit your container config:
Basic passthrough config:
Code:
lxc.cgroup2.devices.allow: c 195:* rwm
lxc.cgroup2.devices.allow: c 509:* 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-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=fileRestart container:
MULTIPLE GPU SETUPS
If your server has multiple GPUs, things work a little differently.
Example:
- GPU0 = GTX 1060
- GPU1 = Tesla V100
- GPU2 = Quadro P4000
Each GPU gets its own device node:
Code:
/dev/nvidia0
/dev/nvidia1
/dev/nvidia2You can verify GPU numbering with:
or:
Passing Through ONLY One Specific GPU
Example: pass ONLY GPU1 (Tesla V100) into container.
Use:
Code:
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=fileThis isolates the container to ONLY that GPU.
Passing Through Multiple GPUs
Example: expose GPU0 and GPU2 to the same container:
Code:
lxc.mount.entry: /dev/nvidia0 dev/nvidia0 none bind,optional,create=file
lxc.mount.entry: /dev/nvidia2 dev/nvidia2 none bind,optional,create=fileKeep the shared NVIDIA devices:
Code:
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=fileRestricting CUDA to Specific GPUs
Inside the container, you can limit visible GPUs using:
Code:
export CUDA_VISIBLE_DEVICES=0or multiple GPUs:
Code:
export CUDA_VISIBLE_DEVICES=0,1VERY useful for:
- Ollama
- Docker
- AI workloads
- Multi-user servers
Docker + NVIDIA Toolkit Inside LXC
If running Docker inside the container:
Code:
apt install -y nvidia-container-toolkitTest with:
Code:
docker run --rm --gpus all nvidia/cuda:12.3.2-base-ubuntu22.04 nvidia-smiINSIDE THE CONTAINER
Download the same NVIDIA driver version used on the host.
550 branch:
Code:
wget https://us.download.nvidia.com/XFree86/Linux-x86_64/550.163.01/NVIDIA-Linux-x86_64-550.163.01.run575 branch:
Code:
wget https://us.download.nvidia.com/tesla/575.57.08/NVIDIA-Linux-x86_64-575.57.08.runInstall userspace libraries ONLY:
550 example:
Code:
chmod +x NVIDIA-Linux-x86_64-550.163.01.run
./NVIDIA-Linux-x86_64-550.163.01.run --no-kernel-module575 example:
Code:
chmod +x NVIDIA-Linux-x86_64-575.57.08.run
./NVIDIA-Linux-x86_64-575.57.08.run --no-kernel-moduleTest GPU access:
If everything is correct, the container should now see the GPU.
OPTIONAL: CUDA INSTALLATION INSIDE LXC
Download CUDA installer matching the NVIDIA driver version.
CUDA 12.9.1 + NVIDIA 575.57.08:
Code:
wget https://developer.download.nvidia.com/compute/cuda/12.9.1/local_installers/cuda_12.9.1_575.57.08_linux.runMake executable:
Code:
chmod +x cuda_12.9.1_575.57.08_linux.runRun installer:
Code:
./cuda_12.9.1_575.57.08_linux.runIMPORTANT:
- Skip driver installation during CUDA setup
- The NVIDIA driver is already installed on the Proxmox host
- Inside LXC you only need CUDA userspace libraries/tools
Recommended CUDA components:
- CUDA Toolkit
- CUDA Runtime
- nvcc
- CUDA libraries
You can usually skip:
- Driver
- DKMS
- Kernel modules
VERIFY CUDA
Check GPU visibility:
Check CUDA compiler:
You should now have:
- CUDA working inside LXC
- GPU passthrough working
- Docker/Ollama/AI workloads accelerated
EXAMPLE NVIDIA-SMI OUTPUT
Example nvidia-smi output after successful passthrough:
Code:
+-----------------------------------------------------------------------------------------+
| NVIDIA-SMI 575.57.08 Driver Version: 575.57.08 CUDA Version: 12.9 |
|-----------------------------------------+------------------------+----------------------|
| GPU Name Persistence-M | Bus-Id Disp.A | Volatile Uncorr. ECC |
| Fan Temp Perf Pwr:Usage/Cap | Memory-Usage | GPU-Util Compute M. |
| | | MIG M. |
|=========================================+========================+======================|
| 0 Tesla V100-SXM2-16GB Off | 00000000:65:00.0 Off | 0 |
| N/A 36C P0 45W / 300W| 0MiB / 16384MiB | 0% Default |
| | | Disabled |
+-----------------------------------------+------------------------+----------------------+
| 1 Quadro P4000 Off | 00000000:B3:00.0 Off | N/A |
| 34% 42C P8 12W / 105W| 256MiB / 8192MiB | 3% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
| 2 GeForce GTX 1060 6GB Off | 00000000:C4:00.0 Off | N/A |
| 27% 39C P8 8W / 120W| 128MiB / 6144MiB | 1% Default |
| | | N/A |
+-----------------------------------------+------------------------+----------------------+
+-----------------------------------------------------------------------------------------+
| Processes: |
| GPU GI CI PID Type Process name GPU Memory |
| ID ID Usage |
|=========================================================================================|
| 1 N/A N/A 2481 C ollama 220MiB |
| 2 N/A N/A 3154 C python 110MiB |
+-----------------------------------------------------------------------------------------+Useful commands:
Live monitoring:
Show GPU list only:
Check CUDA compiler:
OPTIONAL: ADD CUDA TO PATH
Add to ~/.bashrc:
Code:
export PATH=/usr/local/cuda/bin:$PATH
export LD_LIBRARY_PATH=/usr/local/cuda/lib64:$LD_LIBRARY_PATHReload shell:
FINAL THOUGHTS
This setup ended up being WAY simpler than all the vGPU guides floating around online.
For older NVIDIA cards, standard LXC passthrough works perfectly fine for:
- Ollama
- CUDA
- Docker
- AI workloads
- Plex
without needing enterprise NVIDIA features.
Hopefully this saves someone else a few hours of troubleshooting.