fix(hami-scheduler): compatible with app's stop fast behavior

[dkeven]

• Background
  After appservice: stop app fast if pod was hami schudule failed when resume #2699, an Application that's reported unschedulable by hami-scheduler will be stopped immediately by app-service, however, hami-scheduler also reports unschedulable and make kube-sheduler retry scheduling in many retryable cases, such as node locked by another pod. Also, the asynchronous nature of HAMi's informer may lead to device occupation stats not updated immediately, causing pod to be scheduled only in the next retry. Two changes have been made to make HAMi compatible with this new logic:
  1.Add a new event type reasoned as InsufficientGPU that's dedicated to the case when no available GPU resources can be found for the to-be scheduled pod, separating from other normal retryable cases.
  2.When pod is deleted by HAMi-scheduler itself, update the in-memory device usage immediately rather than relying on the pod informer to update the state, to avoid potential race conditions with the deployment controller.

• Target Version for Merge
  1.12.5, 1.12.6

• Related Issues
  none

• PRs Involving Sub-Systems
  fix(scheduler): compatible with app's stop fast behavior HAMi#17

• Other information:
  none

[vercel]

The latest updates on your projects. Learn more about Vercel for GitHub.

1 Skipped Deployment

Project	Deployment	Actions	Updated (UTC)
olares-docs	Ignored		Mar 18, 2026 1:47pm