Skip to content

fix(relearn): pin proven CUDA scoring eval image digest - #205

Merged
echobt merged 2 commits into
mainfrom
cursor/pin-relearn-eval-8ffbe8a
Sep 3, 2026
Merged

fix(relearn): pin proven CUDA scoring eval image digest#205
echobt merged 2 commits into
mainfrom
cursor/pin-relearn-eval-8ffbe8a

Conversation

@echobt

@echobt echobt commented Sep 3, 2026

Copy link
Copy Markdown
Contributor

Summary

  • Pin config/relearn-pin.toml eval_image_digest to the runtime CUDA scoring image proven by CortexLM/relearn publish-eval-image on merge SHA 8ffbe8a0481cc62c1e2d6ed80a2211befa98dd3f.
  • Digest sha256:4806db4bd6650415c3705290117fd0190cd2a6d387cef871dd3d9ff1a8bb6d7e. Runtime job pulled it, ran the harvest-PATH check, and proved import vllm, torchvision plus relearn-eval selftest.
  • Publish: https://github.com/CortexLM/relearn/actions/runs/33730126042
  • Not the contract slim python image (sha256:3ab0f45e…). Not sha256:cbc4bbb8…. Dead harvest failures stay named in the pin.

A live host still 503s until harvest + champion baseline are recorded. No terraform/helm.

Test plan

  • cargo test -p relearn-eval --test committed_pin
  • Confirm eval_image_digest is sha256:4806db4b… and relearn_git_sha is 8ffbe8a0…
  • Confirm no cbc4bbb8 in the digest field
  • CI green on this PR before merge

Pin eval_image_digest to sha256:4806db4b from CortexLM/relearn@8ffbe8a
publish-eval-image run 33730126042, which pulled that digest and proved
harvest-PATH plus import vllm, torchvision. Not the contract slim image
and not cbc4bbb8.
@greptile-apps

greptile-apps Bot commented Sep 3, 2026

Copy link
Copy Markdown

Greptile Summary

Pins Relearn live evaluation to a specific CUDA scoring image and records the matching upstream source revision, with tests and operator documentation updated to match. Execution checks disproved two potential regressions: the new pin does not allow live scoring without a wired harvest, and the configured Git revision and GHCR image manifest are reachable and match the configured values.

Confidence Score: 5/5

Safe to merge based on the exercised Relearn readiness and public image-provenance checks.

No defects remain: the changed pin parses and validates, preserves the harvest prerequisite for live scoring, and points to a reachable upstream revision and image manifest.

Files Needing Attention: No files need further attention.

T-Rex T-Rex Logs

What T-Rex did

  • I ran a focused Relearn readiness check before and after the commit, observing that before the change the empty digest produced can_rent=false and Lium scoring returned EvalImageUnpinned, while after the change the pinned digest was rentable but Lium scoring without harvest returned LiveHarvestUnavailable and did not fall back to simulation.
  • I validated the configured image and source revision by parsing TOML, validating the pin, resolving CortexLM/relearn@8ffbe8a, and confirming GHCR HTTP 200 for sha256:4806db...6d7e.
  • I disproved the hypothesis that a newly non-empty pin would immediately enable an unsafe scorer, since scoring_readiness(pin, Lium, None) stayed at LiveHarvestUnavailable and did not fall back to Sim.
  • I disproved the hypothesis that the digest/SHA was malformed or unreachable, because the TOML parser and pin validator passed and Git resolved the exact SHA while GHCR authenticated manifest lookups returned the expected digest.
  • I authored the test source and reproducible runner and uploaded them with their captured output.

View all artifacts

T-Rex Ran code and verified through T-Rex

Reviews (1): Last reviewed commit: "fix(relearn): pin proven CUDA scoring ev..." | Re-trigger Greptile

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.

2 participants