Skip to content

Document sub-minute H200 startup from a network volume - #69

Merged
kbespalov merged 1 commit into
mainfrom
docs/runpod-volume-restore
Nov 1, 2026
Merged

kbespalov merged 1 commit into
mainfrom
docs/runpod-volume-restore

Conversation

@kbespalov

Copy link
Copy Markdown
Contributor

Fresh H200 replica startup previously spent time re-extracting the Python environment and contending over network prefetch. Document the released deployment that reuses a verified runtime and snapshot on the network volume, uses 16 prefetch workers and sufficient CPU capacity, and returns the first external token in 52.14/52.94 seconds including test-only cache advice. An ordinary fresh pod with warm host cache completed in 32.93 seconds.

Include the local-build/verified-copy lifecycle, per-stage timing, failed attempts, cache/resource boundaries and the unresolved numerical canary and publication-recovery limitations. This changes documentation only; the measured binaries are warmctl v2.0.0-beta.17 and DMTCP v5.0.0-warmctl.1.

Validation: raw per-pod logs and monotonic-clock results checked; three final replicas passed weight fingerprints, streaming, tool calls and 32 concurrent requests each; cleanup verified through Runpod API; git diff --check.

Closes #68

@kbespalov
kbespalov merged commit c370dc5 into main Nov 1, 2026
8 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

Qualify fresh H200 replicas from a network volume below 60 seconds

1 participant