# Volcano v1.13.4

2026-08-31 · [Original release notes](https://github.com/volcano-sh/volcano/releases/tag/v1.13.4) · [Full summary](https://ratatosk.io/en/releases/volcano/v1.13.4)

Breaking changes 1 · Important fixes 6

## Breaking changes (1)

### HAMi vGPU discovery now uses node allocatable resources

- Affects: NVIDIA/HAMi vGPU scheduling
- Action required: Non-zero node allocatable values for `volcano.sh/vgpu-number`, `volcano.sh/vgpu-cores`, and `volcano.sh/vgpu-memory`

HAMi vGPU scheduling failures occurred in large- and medium-scale clusters. The scheduler no longer uses the `volcano.sh/node-vgpu-handshake` annotation, and discovery returns no devices if allocatable information is missing or any required resource is absent or zero.

Evidence: “Cherry-picks the vGPU scheduling fix from master into `release-1.13` by removing the scheduler/device-plugin handshake path and relying on node allocatable resources instead. This aligns release behavior with the upstream fix for large/medium cluster vGPU scheduling failures.” https://github.com/volcano-sh/volcano/pull/5431

> * Fix HAMi vGPU scheduling failures in large and medium-scale clusters by @linuxfhy in [#5431](https://github.com/volcano-sh/volcano/pull/5431)

## Important fixes (6)

### Preemption reprieves higher-priority victim pods first

- Affects: scheduler preemption (Inferred)

When `SelectVictimsOnNode` drained the victims queue during preemption, its low-to-high priority ordering caused lower-priority pods to be reprieved first.

> * Fix victim reprieve order in preemption by @dengaosong in [#5310](https://github.com/volcano-sh/volcano/pull/5310)

### Succeeded pods counted as ready for job dependencies

- Affects: job task dependencies checked by `isDependsOnPodsReady` (Inferred)

Previously, the check could report a depended-on task as not ready when its pods had succeeded, because those pods were not counted as ready. Now, pods in phase `v1.PodSucceeded` count toward readiness.

> * Treat succeeded pods as ready when checking job dependencies by @avinxshKD in [#5548](https://github.com/volcano-sh/volcano/pull/5548) and [#5549](https://github.com/volcano-sh/volcano/pull/5549)

### PodGroup scheduling counts terminating Pods toward `minMember`

- Affects: PodGroups with gracefully terminating scheduled Pods

When scheduled Pods were gracefully terminating but still occupying their nodes, they were modeled as `Releasing` without counting as scheduled; PodGroups could flip from `Running` to `Pending` and then `Inqueue`, or become `Unknown`, and the gang plugin could report misleading `NotEnoughResources`. The gang plugin now skips that condition when those tasks satisfy `minAvailable` and there is no `Pending` or `Pipelined` work.

> * Prevent PodGroups from falling into `Inqueue` while pods are terminating by @halcyon-r in [#5842](https://github.com/volcano-sh/volcano/pull/5842)

### Unbounded growth of job status conditions fixed

- Affects: Job status conditions updated by the controller (Inferred)

On the changed controller paths, an existing condition with the same phase is removed before the new one is appended; conditions with other phases remain.

> * Fix unbounded `job.Status.Conditions` growth by @avinxshKD in [#5450](https://github.com/volcano-sh/volcano/pull/5450)

### Predicates plugin stops node allocation after `Allocate` errors

- Affects: scheduler setups using the predicates plugin (Inferred)

> * Stop further node allocation after an error occurs in the predicates plugin by @jiahuat in [#5908](https://github.com/volcano-sh/volcano/pull/5908)

### Backfill skips tasks when no best node is selected

- Affects: backfill tasks

When node scoring failed and no best node was selected, the backfill task path could panic.

> * Skip backfill tasks when no best node is selected by @mesutoezdil in [#5922](https://github.com/volcano-sh/volcano/pull/5922)

---

Summaries are produced by AI, so they can be wrong, and they are provided without warranty.
