How to use Container Device Interface devices with Podman API?

[suptejas]

The following command works as expected with nvidia-smi accessible inside the podman container for me:

podman run -d --name jupyter-gpu \
  --device nvidia.com/gpu=all \
  --security-opt=label=disable \
  -p 8888:8888 \
  quay.io/dimension/jupyter-server

However, I'm unable to figure out how to get it to work with the Podman API. Would anyone be kind enough to share an example of using a CDI device? My recent attempt resulted in the following error:

container create: stat nvidia.com/gpu=all: no such file or directory

this gives me the feeling that it is not identifying nvidia.com/gpu=all as a cdi device.

Here's the rust code I'm using (keep in mind the bollard crate will use the docker-compatible API for Podman (I think).

pub async fn create(podman: &Podman, request: ProvisionNotebookRequest) -> Result<()> {
    let container_name = "jupyter-gpu";

    // Pull the image if it doesn't already exist
    let mut pull_stream = podman.create_image(
        Some(CreateImageOptions {
            from_image: "quay.io/dimension/jupyter-server".to_string(),
            ..Default::default()
        }),
        None,
        None,
    );

    while let Some(pull_result) = pull_stream.next().await {
        pull_result.into_diagnostic()?;
    }

    // Create container options
    let options = CreateContainerOptions {
        name: container_name.to_string(),
        ..Default::default()
    };

    // Set up port bindings for the container
    let port_bindings: HashMap<String, Option<Vec<PortBinding>>> = vec![(
        "8888/tcp".to_string(),
        Some(vec![PortBinding {
            host_ip: Some("0.0.0.0".to_string()),
            host_port: Some("8888".to_string()),
        }]),
    )]
    .into_iter()
    .collect();

    // Container configuration
    let config = Config {
        image: Some("quay.io/dimension/jupyter-server".to_string()),
        host_config: Some(HostConfig {
            port_bindings: Some(port_bindings),
            security_opt: Some(vec!["label=disable".to_string()]),
            devices: Some(vec![bollard::models::DeviceMapping {
                path_on_host: Some("nvidia.com/gpu=all".to_string()),
                path_in_container: Some("nvidia.com/gpu=all".to_string()),
                cgroup_permissions: Some("rwm".to_string()),
            }]),
            ..Default::default()
        }),
        env: Some(vec![format!("JUPYTER_TOKEN={}", request.auth_password)]),
        ..Default::default()
    };

    // Create the container
    let container = podman
        .create_container(Some(options), config)
        .await
        .into_diagnostic()?;

    // Start the container
    podman
        .start_container(&container.id, None::<StartContainerOptions<String>>)
        .await
        .into_diagnostic()?;

    println!(
        "Container {} started successfully with GPU support",
        container_name
    );

    Ok(())
}

podman --version
podman version 5.1.2

nvidia-ctk cdi list
INFO[0000] Found 2 CDI devices
nvidia.com/gpu=0
nvidia.com/gpu=all

[KevinZonda]

I do not write python, but I have go snippet:

ops.Resources.DeviceRequests = []container.DeviceRequest{
			{
				Driver: "cdi",
				DeviceIDs: []string{
					"nvidia.com/gpu=all", // TODO: Make this dynamic,
					// Related issue: https://github.com/containers/podman/pull/25171
					// Should works in podman v5.4.2+
				},
			},
		}

Resources is at HostConfig