Skip to content

fix(validator): isolate each NCCL benchmark run in its own namespace #2443

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Open
mikecook wants to merge 13 commits into
base: main
Choose a base branch
from
Open

fix(validator): isolate each NCCL benchmark run in its own namespace #2443

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
13 commits
Select commit Hold shift + click to select a range
1cd1255
fix(validator): isolate each NCCL benchmark run in its own namespace
mikecook Aug 28, 2026
9e54920
fix(validator): preserve structured errors from namespace lifecycle
mikecook Aug 28, 2026
2e76cc2
fix: verify NCCL namespace ownership before reuse or delete
mikecook Aug 28, 2026
39fbf35
fix: run namespace cleanup before Trainer teardown on self-install path
mikecook Aug 28, 2026
8290cb7
fix: don't fail the benchmark on a namespace-termination wait timeout
mikecook Aug 28, 2026
0a364b7
fix: fold NCCL variant into the per-run namespace, not runID alone
mikecook Aug 28, 2026
23a23a2
fix: prune stale orphaned NCCL benchmark namespaces on next run
mikecook Aug 29, 2026
5964fe1
test: harden the finalizer-wait regression guard's timing check
mikecook Aug 28, 2026
da236f9
fix: ignore stale Deleted and terminal pod events in the launcher watch
mikecook Aug 29, 2026
a1b8e57
fix: reclaim a stale TrainJob by delete-and-recreate, not update
mikecook Aug 29, 2026
5503000
fix: don't prune a stale-looking NCCL namespace that still has a live…
mikecook Aug 29, 2026
6092710
fix: reject an empty owning UID in cleanupNCCLResources
mikecook Aug 29, 2026
a105267
docs: correct a stale doc comment on the namespace-gone timeout test
mikecook Aug 29, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
6 changes: 6 additions & 0 deletions pkg/defaults/timeouts.go
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -686,6 +686,12 @@ const (
// NCCLTrainJobTimeout is the maximum time to wait for the NCCL all-reduce TrainJob to complete.
NCCLTrainJobTimeout = 30 * time.Minute

// NCCLStaleNamespacePruneAge is how old a leftover aicr-nccl-perf-* namespace
// must be before the best-effort prune in runNCCLTrainJob deletes it. Sized
// well past NCCLTrainJobTimeout so a namespace still owned by an in-progress
// sibling variant, or a run that is simply slow, is never touched.
NCCLStaleNamespacePruneAge = 2 * NCCLTrainJobTimeout

// NCCLLauncherPodTimeout is the maximum time to wait for the NCCL launcher pod to be created.
NCCLLauncherPodTimeout = 5 * time.Minute

Expand Down
Loading
Loading