What must be on the host for Docker containers to use GPUs?
Why C is the answer
The NVIDIA driver and Container Toolkit
The toolkit connects containers to the host driver.
In detail
- The toolkit connects containers to the host driver.
- The toolkit connects containers to the driver.
Use case
An illustrative scenario, not a real incident.
Containers cannot see GPUs until the host has the driver and Container Toolkit.
See it in different ways
flowchart TD A["NVIDIA driver on the host"] --> B["NVIDIA Container Toolkit"] B --> C["Docker runtime configured"] C --> D["docker run with --gpus"]
sequenceDiagram participant D as Host driver participant T as Container Toolkit participant C as Docker D->>T: Exposes the GPUs T->>C: Configures the runtime C->>C: docker run with --gpus all
mindmap
root(("Docker host"))
n1["Needs"]
n2["NVIDIA driver"]
n3["Container Toolkit"]Key point: Containers need host driver and toolkit.
Why the other options are wrong
- A. Slurm and Pyxis on every storage node
- Those are for Slurm.
- B. A vGPU license on each container image
- That is about virtual machines.
- D. A full CUDA install inside the BMC
- The BMC does not run CUDA.