Open timozerrer opened 3 years ago
So far this is not feasible, since we define a resource called "SharedPod", that is the resource that we use to implement GPU sharing. And in the other case like Kubeflow pipeline, every pipeline stage using Pod as a unit to allocate resource. So .. will it's not possible change that now :)
We are currently working on another branch that will fit your requests by re-implementing our GPU sharing mechanism based on the scheduling framework 2.0. We expect to release the branch by the end of the year, so stay tuned.
Thanks for this solution, succesfully tested this in my cluster.
Is there any way to use the solution without the SharePod custom resources?
E.g. create a Job or Pod directly using annotations / resource requests?
Im looking at the following scenarios:
Thanks