Skip to content

7g.79gb does not work as expected. #51

Description

@houms-sony

using gpu-operator (helm 23.9.1), and nos (helm 0.1.2)

I have an issue with nvidia.com/mig-7g.79gb. when specifying it it causes nos to create the mig configuration as expected, but it seems to be specified as nvidia.com/mig-7g.80gb as shown in log below from nvidia-device-plugin.

I0312 23:04:34.682199       1 server.go:165] Starting GRPC server for 'nvidia.com/mig-7g.80gb'
I0312 23:04:34.682673       1 server.go:117] Starting to serve 'nvidia.com/mig-7g.80gb' on /var/lib/kubelet/device-plugins/nvidia-mig-7g.80gb.sock
I0312 23:04:34.684745       1 server.go:125] Registered device plugin for 'nvidia.com/mig-7g.80gb' with Kubelet

Additionally, the labels created on the node look like this

Screenshot_20240312_162714

But the issue is because we specified nvidia.com/mig-7g.79gb the pod stays in pending. Note the config below (all other nvidia examples commented out below work except 7g.79gb.

---
apiVersion: batch/v1
kind: Job
metadata:
  name: job-test-7g80g
spec:
  template:
    spec:
      runtimeClassName: nvidia
      restartPolicy: Never
      containers:
      - name: nvidia
        image: nvidia/cuda:12.3.2-devel-ubuntu22.04
        command: ["sleep", "12000"]
        resources:
          limits:
            nvidia.com/mig-7g.79gb: 1
            #nvidia.com/mig-1g.10gb: 1
            #nvidia.com/mig-2g.20gb: 1
            #nvidia.com/mig-4g.40gb: 1

I tried adding 7g.80gb to allowedGeometries, but it did not work as expected. Briefly looked at code and see https://github.com/nebuly-ai/nos/blob/main/pkg/gpu/mig/known_configs.go#L93, so not sure if I missed something, or if there is a way to get the desired behavior?

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions