What would you like to be added:

Serving workloads are different than training - they can be easily trimmed - a Deployment can run at 70% or 50% of Pods. This is different to most AI training workloads, where all Pods need to run. We want to leverage this fact and optimize preemptions.

In particular, when a new high priority workload comes in and we have multiple serving workloads, we want to distribute the preemptions across the serving workloads, rather than preempting one completely.

Note that this is also related to the partial preemption for batch workloads: #975. We may consider having a solution which solves both problems, but for now it seems reasonable to have this dedicated issue, emphasizing that serving workloads are special in this regard.

Why is this needed:

To improve experience of hosting mix of training and inference workloads. When the high-priority workload comes, we can make room for it by trimming multiple serving workloads, rather than preempting completely one.

Completion requirements:

This enhancement requires the following artifacts:

• Design doc
• API change
• Docs update

The artifacts should be linked in subsequent comments.