onyxia icon indicating copy to clipboard operation
onyxia copied to clipboard

[feature request] Multiple heterogeneous GPU ressources

Open Archimonde666 opened this issue 2 weeks ago • 0 comments

As of now, Onyxia allow only for specification of the numeric value of the nvidia.com/gpu key.

For intermediate infrastructure and heterogeneous clusters, many different kind/subdivision of gpu might cohexist and it would be convenient if users could target more specific GPU ressources.

I tried different approaches to do it, first by inserting direct key:values entries in the schemas files (ressources-gpu.json) but it did not work due to the json nature of the schema that parse the ":" character crashing the parser.

The best way i could work around was to define manually new keys for every ressource present in the cluster :

            "nvidia.com/gpu": {
              "type": "string",
              "default": "0"
            },
            "nvidia.com/mig-3g.47gb": {
              "type": "string",
              "default": "0"
            },
            "nvidia.com/mig-4g.47gb": {
              "type": "string",
              "default": "0"
            },
            "nvidia.com/mig-7g.80gb": {
              "type": "string",
              "default": "0"
            },
            "nvidia.com/mig-7g.94gb": {
              "type": "string",
              "default": "0"
            }

That doesn't render catastrophically but it would be more convenient to be able to use the list object, maybe by creating a magic character or a new mechanism.

Image

Archimonde666 avatar Dec 12 '25 18:12 Archimonde666