[Bug]: DinD doesn't allow passing `--gpus` flag

[mtaran]

Steps to reproduce

1. make a repro.dstack.yml with:

type: task
name: my-repro-task
image: dstackai/dind:latest
privileged: true
commands:
  - start-dockerd
  - sleep infinity
resources:
  cpu: 4..
  memory: 6GB..
  gpu:
    count: 1

2. dstack apply -f repro.dstack.yml -y
3. in another terminal: dstack attach my-repro-task
4. in yet another terminal: ssh my-repro-task
5. in the ssh session, try to run docker run --rm --gpus=all hello-world

Actual behaviour

docker: Error response from daemon: failed to create task for container: failed to create shim task: OCI runtime create failed: runc create failed: unable to start container process: error during container init: error running hook #0: error running hook: exit status 1, stdout: , stderr: Auto-detected mode as 'legacy'
nvidia-container-cli: mount error: mount operation failed: /usr/bin/nvidia-smi: no such file or directory: unknown.

Expected behaviour

The container should run, with access to all the GPUs of the host.

dstack version

0.18.18

Server logs

No response

Additional information

No response

[peterschmidt85]

@mtaran Is the issue still relevant? You mentioned you made it work.

[un-def]

As TensorDock is “a marketplace of independent hosts”, the setups are not consistent at all.

In addition to

nvidia-container-cli: mount error: mount operation failed: /usr/bin/nvidia-smi: no such file or directory: unknown.

i've got

nvidia-container-cli: mount error: mount operation failed: /usr/bin/nv-fabricmanager: no such file or directory: unknown

and

nvidia-container-cli: mount error: mount operation failed: /usr/bin/nvidia-persistenced: no such file or directory: unknown

when requesting instances with the same resources configuration.

NVIDIA/CUDA driver versions also vary:

NVIDIA-SMI 535.161.07             Driver Version: 535.161.07   CUDA Version: 12.2

and

NVIDIA-SMI 550.54.15              Driver Version: 550.54.15      CUDA Version: 12.4   

on two NVIDIA RTX A4000 instances in the same region.

[github-actions[bot]]

This issue is stale because it has been open for 30 days with no activity.

[github-actions[bot]]

This issue was closed because it has been inactive for 14 days since being marked as stale. Please reopen the issue if it is still relevant.