Can't access to GPU using podman compose

[joaopalma5]

Issue Description

Hello,

as we know the deploy resources (deploy.resources.reservations.devices) not are working with podman compose as described in the next issue:
Describe your issue

so this is the solution until now:
ollama:
image: ollama/ollama:latest
container_name: ollama
restart: unless-stopped
ports:
- "11434:11434"
volumes:
- ollama_data:/root/.ollama
# ── GPU passthrough (NVIDIA) — comment out for CPU-only ──
devices:
- nvidia.com/gpu=all
healthcheck:
<<: *hc-defaults
test: ["CMD", "curl", "-f", "http://localhost:11434/api/tags"]

but now is giving this error when we are creating the container

• Container ollama Creating 0.2s
  Error response from daemon: container create: stat nvidia.com/gpu=all: no such file or directory

best regards

[mheon]

I was not aware that deploy.resources.reservations.devices was not working properly; the issue you're linking to is closed. Can you provide more details?

[joaopalma5]

With this code

  ollama-gpu:
    image: ollama/ollama:latest
    container_name: ollama
    restart: unless-stopped
    ports:
      - "11434:11434"
    volumes:
      - ollama_data:/root/.ollama
    devices:
      - "nvidia.com/gpu=all"

we get

[+] Running 3/4
 ✔ Network dlmcp_default          Created                                                                                                                              0.0s 
 ✔ Volume "dlmcp_openwebui_data"  Created                                                                                                                              0.0s 
 ✔ Volume "dlmcp_ollama_data"     Created                                                                                                                              0.0s 
 - Container ollama               Creating                                                                                                                             0.2s 
Error response from daemon: container create: stat nvidia.com/gpu=all: no such file or directory

but with this, works normally
podman run -d --pod ai-stack --name ollama --security-opt label=disable --device nvidia.com/gpu=all -v ollama_data:/root/.ollama ollama/ollama:latest

The issue linked before is closed but as you can see without proper solution, until now we can made a work around using devices - "nvidia.com/gpu=all" but not works anymore...

podman version 5.7.1
Versão WSL: 2.7.0.0
Versão do kernel: 6.6.114.1-1
Versão WSLg: 1.0.71
Versão MSRDC: 1.2.6676
Versão direct3D: 1.611.1-81528511
Versão DXCore: 10.0.26100.1-240331-1435.ge-release
Versão do Windows: 10.0.26200.8037

@mheon can you try on your side to see the results?

[matteogrechi]

Sorry, I just created this issue --> #28436

I think the two issues are related

[Honny1]

@joaopalma5 Could you share your docker-compose version?

I have suspision that stat nvidia.com/gpu=all error means the CDI name is arriving as a DeviceMapping on HostConfig.Devices, older docker-compose (< v2.30, before docker/compose#12184) does this instead of routing CDI names to DeviceRequests.

PR #28495 fixes the HostConfig.Devices path to preserve CDI qualified names. Could you test with the patched binary?