osdc/hf-cache: run the mount on GPU nodes only - #884
Draft
huydhn wants to merge 2 commits into
Draft
Conversation
CPU runners don't pull models, so the per-node hf-cache rclone mount is wasted
on CPU nodes. Restrict the mount DaemonSet to GPU nodes via a `nvidia.com/gpu:
"true"` nodeSelector.
Shipped with the startup-taint gate (required): the `node-init.osdc.io/hf-cache`
taint is emitted on runner nodepools and was cleared by the mount DS running
everywhere. With the DS now GPU-only, gate the taint to GPU nodepools
(`applies_when: d.get("gpu")`) — otherwise CPU nodes would be tainted with
nothing to clear it and never schedule runner pods (same pattern cache-enforcer
uses for release-runner pools).
Packing: the templated hf-cache overhead flips to `gpu_only=True`, so
`analyze_node_utilization` no longer reserves it on CPU nodes.
Job-pod `/mnt/hf_cache` mount on CPU runners is left as-is (harmless empty
read-only dir via hostPath DirectoryOrCreate; `hf_cache_enabled` is cluster-wide).
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What
CPU runners don't pull models, so the per-node hf-cache rclone mount is wasted on CPU nodes. Restrict the mount DaemonSet to GPU nodes via a
nvidia.com/gpu: "true"nodeSelector.Rebased onto current
main(on top of #883's per-GPU-count tiers + #889); GPU-only is expressed purely via the nodeSelector, tier sizes unchanged.Changes
mount-daemonset.yaml.tpl: addnvidia.com/gpu: "true"nodeSelector.node-init.osdc.io/hf-cacheto GPU nodepools (applies_when: d.get("gpu")) — with the DS GPU-only, a CPU node would otherwise be tainted with nothing to clear it and never schedule runner pods.gpu_only=Trueso it's no longer reserved on CPU nodes.The CPU job-pod
/mnt/hf_cachebind-mount stays (cluster-widehf_cache_enabled); harmless because the hostPath istype: DirectoryOrCreate— with no rclone DS on the node, kubelet just creates an empty read-only dir.daemonset_overhead.py(hf-cache now GPU-only)CPU nodes no longer reserve the hf-cache 256Mi; no runner regressed to 0 pods.
Deploy
Redeploy
hf-cache+arc-runners(+arc-runners-opt). Drain so nodes recycle onto the GPU-gated nodepool taint.Independent of #890; overlapping lines in
deploy.sh/daemonset_overhead.pymerge trivially.