CDI GPU devices not passed through when using `podman compose`

[matteogrechi]

Issue Description

When using podman compose (backed by the docker-compose plugin), CDI GPU devices specified in the devices: field of a compose file are silently ignored. The container starts without GPU access and falls back to CPU.

The same configuration works correctly with:

• podman run --device nvidia.com/gpu=all
• podman-compose (pip package)

Steps to reproduce the issue

1. Generate CDI spec:

sudo nvidia-ctk cdi generate --output=/etc/cdi/nvidia.yaml

2. Create compose.yaml:

services:
  test:
    image: nvcr.io/nvidia/cuda:12.2.0-base-ubuntu22.04
    command: nvidia-smi
    devices:
      - nvidia.com/gpu=all
    security_opt:
      - label=disable

3. Run:

podman compose up

Describe the results you received

CDI device is ignored. Container starts without GPU. No error or warning is shown.

Describe the results you expected

Container starts with GPU access. nvidia-smi runs successfully inside the container.

podman info output

- OS: Ubuntu 24.04
- Podman version: 4.9.3
- Compose backend: `/usr/libexec/docker/cli-plugins/docker-compose` v2.37.1
- nvidia-container-toolkit: 1.19.0
- GPU: NVIDIA Tesla T4
- CDI spec: generated at `/etc/cdi/nvidia.yaml`

Podman in a container

No

Privileged Or Rootless

Rootless

Upstream Latest Release

No

Additional environment details

No response

Additional information

Workaround

Using podman run directly with --device nvidia.com/gpu=all works correctly.
Using podman-compose (pip package) also works correctly.

Additional Notes

The deploy.resources.reservations.devices block with driver: cdi is also ignored by the docker-compose backend.

This appears to be because podman compose delegates to the docker-compose plugin which does not honor CDI device syntax, bypassing podman's native CDI handling entirely.

[matteogrechi]

May be related to --> #28309

[Honny1]

Docker Compose hits Podman’s Docker-compat create path, which always encodes HostConfig.Devices as host:container:perms, so CDI ids like nvidia.com/gpu=all get extra colons, stop parsing as CDI, and fail or silently drop GPU while podman run --device still works because it passes the id unchanged.

[Honny1]

@matteogrechi, Hi, can you check your setup with the latest podman?

Podman 4.9.3 - does not have the DeviceRequests handling

I think it was fixed by #25171

[github-actions]

A reviewer has determined we need more information to understand the reported issue. A comment on what is missing should be provided. Be certain you:

• provide an exact reproducer where possible
• verify you have provided all relevant information - minimum is podman info
• answer any follow up questions

If no response to the needs-info is provided in 30 days, this issue may be closed by our stale bot.

For more information on reporting issues on this repository, consult our issue guide.