kubeflow / trainer / 29940707525

68%

Build Type

push

github

Committed by web-flow

Commit Message

fix(runtimes): propagate PodSets.Count into TemplateSpec Parallelism/Completions (#3651)

* fix(runtimes): propagate PodSets.Count into TemplateSpec Parallelism/Completions

RuntimeInfo() helpers now correctly propagate PodSets.Count (set by
NumNodes) into the JobSet template spec's Parallelism/Completions
fields. Previously, external consumers like Kueue had to post-process
the results because ObjApply stayed stale after ML policy plugins
updated PodSets.Count.

Adds a BuildParallelCountPlugin interface to the plugin framework.
The JobSet plugin implements it with name-based matching between
PodSets and ReplicatedJobs, called from RuntimeInfo() after
RunEnforceMLPolicyPlugins.

Fixes kubeflow/trainer#3042
Ref: kubernetes-sigs/kueue#8287

Signed-off-by: Dmytro Hryshchenko <dhryshch@redhat.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

* refactor(runtimes): fold SyncParallelCount into ComponentBuilderPlugin

Move parallel count sync from standalone BuildParallelCountPlugin
interface into ComponentBuilderPlugin per #3057 consensus. Remove
duplicate index-based sync from Build() - SyncParallelCount is now
the single source of truth, called before Build() in
RunComponentBuilderPlugins.

Signed-off-by: Dmytro Hryshchenko <dhryshch@redhat.com>
Co-Authored-By: Claude Opus 4.6 <noreply@anthropic.com>

---------

Signed-off-by: Dmytro Hryshchenko <dhryshch@redhat.com>
Co-authored-by: Claude Opus 4.6 <noreply@anthropic.com>

Coverage Stats

19 of 26 new or added lines in 7 files covered. (73.08%)

1 existing line in 1 file now uncovered.

2563 of 3755 relevant lines covered (68.26%)

0.8 hits per line