Proxmox 9.2 shipped in August 2026 with a kernel bump that broke NVIDIA’s open kernel modules for a lot of home lab users running local LLMs, and the fix turned out to be a patched driver NVIDIA released about a week later. That is the pattern every Proxmox owner running a GPU for inference will get to know: a kernel update lands, the driver rejects the new headers, nvidia-smi returns “No devices were found”, and Ollama or vLLM fails to start until you rebuild.
The best apps for managing NVIDIA drivers on Proxmox for local LLM workloads reduce that pattern to a routine rather than a crisis. Here are seven desktop tools worth having installed before the next kernel update.
What to look for in a Proxmox NVIDIA stack
Start with the driver install path. The vendor-supplied .run installer works and is worth understanding, but for a home lab you want the reproducibility of a package. That means a signed .deb from NVIDIA’s CUDA repository or a scriptable path that pins a known-good driver version.
Look at DKMS. Dynamic Kernel Module Support rebuilds the NVIDIA module when the kernel updates. If you skip DKMS, you rebuild by hand every kernel bump, and that is where breakage lives. Confirm the DKMS module actually rebuilt after the update rather than silently failing.
Watch the GPU passthrough side. If the GPU is passed through to a VM rather than run on the host, the module lives inside the VM and Proxmox needs vfio-pci binding to keep the host from grabbing the card. Broken passthrough is often mistaken for a broken driver.
Finally, plan monitoring and rollback. nvidia-smi in a loop tells you the card is alive; Netdata shows utilization over time; a snapshot of the host taken before an update means the rollback is one Proxmox restore away.
Quick comparison
| App | Role | Free plan | Paid tier | Standout feature |
|---|---|---|---|---|
| Proxmox VE | The hypervisor NVIDIA has to live in | Free, open source | Paid subscription for enterprise repo | Snapshot before every risky driver update |
| NVIDIA driver installer | Ships the CUDA and kernel modules | Free | Free | Signed .deb and network installer that survive kernel updates |
| Aprox Proxmox client | Reach the Proxmox console from phone or desktop | Free | Free | Restart a VM without opening the browser |
| Ollama | Local LLM runner that consumes the GPU | Free, open source | Free | One-line install, one-line model pull |
| Open WebUI | Chat frontend on top of Ollama or vLLM | Free, open source | Free | Multi-model chat, RAG, no cloud dependencies |
| vLLM | High-throughput LLM inference server | Free, open source | Free | The higher-throughput alternative to Ollama for busy setups |
| NVIDIA System Management Interface | The nvidia-smi command line |
Free, ships with driver | Free | The tool you run first when something breaks |
| Netdata | Live GPU and system metrics | Free tier | Paid cloud plans | GPU utilization, temperature, and VRAM over time |
1. Proxmox VE
Proxmox VE is the hypervisor at the base of the stack. For NVIDIA workloads, the setting that matters most is snapshot discipline: snapshot the host and the GPU-bearing VM before every risky driver or kernel change, so a rollback is a two-click operation. The Proxmox web UI’s Backup schedule can automate the LXC or VM snapshot; the host itself needs a separate approach (a stateless install, a ZFS snapshot of the root pool, or a manual note of the installer flags).
Where it falls short: Proxmox itself does not offer a first-party GPU passthrough wizard. You configure it through the CLI and /etc/pve/qemu-server/<vmid>.conf.
Pricing: Free with community repo. Paid subscription for enterprise repo starts at a modest annual fee per socket.
Platforms: Bare metal on the compute box.
Download: Proxmox VE
Bottom line: the base. Snapshot before every driver update.
2. NVIDIA driver installer
NVIDIA’s driver installer ships as either a signed CUDA repository package or a .run installer. For Proxmox home labs the CUDA repository is the more reliable path: add the repo, install nvidia-driver-open (the open kernel modules) with DKMS, reboot, verify nvidia-smi. When a kernel update lands, DKMS rebuilds the module; when it doesn’t, you rebuild by hand once and log the fix.
Where it falls short: the open kernel modules are still the newer path. Some older cards need the proprietary modules, and the version matrix is worth reading before you upgrade.
Pricing: Free.
Platforms: Linux (Debian, Ubuntu, and derivatives, which Proxmox uses).
Download: NVIDIA CUDA repository
Bottom line: the repo install with DKMS is the setup that survives kernel updates most of the time.
3. Aprox Proxmox client
Aprox is a phone client for Proxmox. When a driver update goes sideways in the middle of the workday, Aprox lets you restart the affected VM, roll back to a snapshot, or check task history without opening a browser. It reads the same Proxmox API the web UI uses.
Where it falls short: it is a client, not a replacement. Deep configuration still goes through the web UI.
Pricing: Free.
Platforms: Android; usable from Windows and ChromeOS via Android runtimes.
Download: Aptoide · Google Play
Bottom line: the quick-recovery client for a home lab you do not want to log in to a full desktop to touch.
4. Ollama
Ollama is the easiest way to run a local LLM on a Proxmox GPU. Install it in a VM or LXC container with the GPU passed through, pull a model with a one-line command, and Ollama exposes a local API on the standard OpenAI-shaped shape. Home labs that started with Ollama and never left rarely regret it.
Where it falls short: throughput is lower than vLLM at concurrency. For a household chat setup this doesn’t matter; for a small team it might.
Pricing: Free, open source.
Platforms: Linux, macOS, Windows.
Download: Ollama
Bottom line: the default runner for a home lab GPU that isn’t already serving vLLM.
5. Open WebUI
Open WebUI is the chat frontend that sits on top of Ollama or vLLM. It handles multi-model chat, RAG against local documents, a shared conversation history, and multiple user accounts. Run it in a container on the same VM as Ollama and it appears at a URL you can bookmark.
Where it falls short: it is another moving part. If you only ever chat from a terminal, curl against the Ollama API works fine.
Pricing: Free, open source.
Platforms: Docker (any host).
Download: Open WebUI
Bottom line: the friendly frontend for a home LLM setup.
6. vLLM
vLLM is the higher-throughput inference server for when a household or a small team is generating enough tokens per second that Ollama slows down. It supports paged attention, batched requests, and hosts models with much higher concurrency. Setup is more involved and CUDA-version-sensitive.
Where it falls short: it is not the “install and forget” experience Ollama is. Expect to read release notes and match CUDA and PyTorch versions.
Pricing: Free, open source.
Platforms: Linux with CUDA.
Download: vLLM
Bottom line: the pick when concurrency matters more than setup simplicity.
7. NVIDIA System Management Interface
nvidia-smi is the command every Proxmox-plus-NVIDIA user runs first when something breaks. It shows driver version, CUDA runtime, GPU utilization, VRAM usage, active processes, temperatures, and power. When a kernel update lands, nvidia-smi is the first honest reading of whether the driver survived.
Where it falls short: it is a CLI. There is a graphical nvidia-smi -l 1 loop you can leave in a tmux pane, but it is not a dashboard.
Pricing: Free, ships with the driver.
Platforms: Linux, Windows (via NVIDIA driver).
Download: Bundled with the NVIDIA driver install above.
Bottom line: the first tool you should run after every kernel or driver change.
8. Netdata
Netdata picks up the NVIDIA collector automatically when the driver is installed. It renders GPU utilization, VRAM, power, and temperature at second-level granularity in the browser. Historically this is how you spot a thermal throttle that a benchmark run does not report.
Where it falls short: long retention on the free tier is limited. Store historical metrics in Prometheus if a year of GPU history matters.
Pricing: Free tier for a small number of nodes. Paid cloud plans add retention.
Platforms: Linux (agent), any browser (dashboard).
Download: Netdata
Bottom line: the fastest way to spot GPU misbehavior over time.
How to pick
Every home lab GPU should start with Proxmox VE (snapshot before driver changes), the NVIDIA CUDA repo with DKMS, and nvidia-smi as the first-check tool. Add Ollama plus Open WebUI for a friendly LLM setup, or vLLM if throughput matters. Install Aprox for phone-side control and Netdata for continuous GPU visibility. When a kernel update lands, snapshot first, update second, run nvidia-smi third, and check Netdata’s GPU panel fourth. If any of those fail, the Aprox client is faster than opening a browser to roll back.
FAQ
Why did Proxmox 9.2 break my local LLMs?
The kernel bump was ahead of NVIDIA’s driver support for it. NVIDIA released a patched driver about a week later; installing that driver, then running dkms status and rebooting, restored nvidia-smi and Ollama.
Should I use the open or proprietary NVIDIA modules on Proxmox?
For recent Ampere and newer cards the open kernel modules are the recommended path in 2026 and are what NVIDIA is supporting forward. Older Turing and Pascal cards still lean on the proprietary modules.
Do I need GPU passthrough for Ollama on Proxmox?
If the GPU is on the compute box and you run Ollama on the host, no. If you run Ollama inside a VM, yes. LXC containers can share the host’s GPU without passthrough with the right cgroup and udev rules.
What is the best local LLM runner on Proxmox?
Ollama for ease of setup, vLLM for concurrency. Both consume the NVIDIA driver installed on the guest.
How do I monitor GPU temperature over time on Proxmox?
Install Netdata on the guest that owns the GPU. It picks up the NVIDIA metrics automatically. For long retention, ship the metrics to Prometheus and view them in Grafana.
