add test phase build files under starter_kit/test_phase

#4
by htranx - opened
AI at Meta org

Adds the Test Phase build files under starter_kit/test_phase/: Containerfile, build_image.sh, validate_image.sh and a README covering build, validate and submit.

The build context is the starter kit directory, so build_image.sh works from here without a full clone of the dataset being tarred up before the build starts.

AI at Meta org

Some minor issues:

  1. starter_kit/test_phase/README.md, around line 90:

The preceding commands imply the user is running from starter_kit/test_phase, but there is no data/ directory here. The validation data is at the repository root and out/ may not exist. This copy-paste command will therefore fail or not find /data/egolongqa/val. Should we make it runnable from this directory, for example by first running mkdir -p out and mounting "$PWD/../..:/data:ro" instead?

  1. starter_kit/test_phase/build_image.sh, line 85:

This works with Podman, but Docker’s push command does not support --digestfile, even though the script selects Docker as a fallback above.

AI at Meta org

Both correct, thank you, and both fixed in the latest commit on this PR.

1. The end-to-end command. Right on all three counts: there is no data/ in this directory, out/ may not exist, and the val data is at the repository root. It now reads:

mkdir -p out
podman run --rm --device nvidia.com/gpu=all \
  -v "$PWD/../..:/data:ro" -v "$PWD/out:/output" \
  ...

with a line saying ../.. is the repository root. The in-container paths were already written against that root, so /data/egolongqa/val and the gold file now resolve.

2. --digestfile. Also right, and worse than cosmetic: the script picks docker whenever podman is absent, which is the common case on a participant's own machine, so the documented push path could not report a digest for most of them, and a digest is what a submission is registered with. Docker exits with an unknown-flag error rather than degrading. The push now branches on the engine: --digestfile under podman, and under docker the digest: sha256:... line is parsed out of the push output. If neither yields a digest the script fails and prints the inspect command to recover it, rather than reporting an empty one.

That path had no test at all. It has one now, driven by a fake docker that rejects --digestfile the way the real one does, so a fake that quietly accepted the flag cannot make it pass.

htranx changed pull request status to merged

Sign up or log in to comment