check as a CI gate

ormas-monitor check reads weights only, needs no GPU and no data, and returns a non-zero exit code when it finds a problem. That makes it a cheap gate before a checkpoint is published or deployed.

Exit codes

Code	Meaning
0	No finding at or above --fail-on.
1	At least one finding at or above --fail-on. The default is critical (for example NaN or Inf weights, or tensors missing from the shards).
2	The file could not be read, or was refused (for example a pickle file without --trust).

--fail-on takes critical (default), warning or never. --quiet prints nothing except errors and --json output.

For a CI gate, use --fail-on warning. The default, critical, fails only on a broken file (NaN or Inf weights, missing shards); dead channels, decayed channels and channels that died since the baseline are warnings, so with the default the check prints them and still exits 0. The last line of the output says why, for example exit 0: no critical findings (2 warnings; use --fail-on warning to fail on them).

ormas-monitor check model.safetensors --fail-on warning --quiet    # also fail on dead or decayed channels

GitHub Actions

name: checkpoint health
on: [pull_request]
jobs:
  check:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: "3.12"
      - run: pip install torch --index-url https://download.pytorch.org/whl/cpu
      - run: pip install "ormas-monitor[safetensors]"
      - run: ormas-monitor check artifacts/model.safetensors --against artifacts/base.safetensors --fail-on warning
      - uses: actions/upload-artifact@v4
        if: always()
        with:
          name: health-report
          path: artifacts/*.health.html

Machine-readable output

--json prints the full result, including every finding with its severity, kind, title and detail, so a script can apply its own policy:

ormas-monitor check model.pt --json > health.json

The finding kinds are listed in the findings catalogue.

Things to know

• Never pass --trust in CI on files from an untrusted source. It allows pickle loading, which can run arbitrary code.
• Pass --arch when you know the architecture. Without it, the check cannot see how layers are wired and qualifies dead-channel findings with "if a ReLU directly follows". A wrong --arch (under half of the file's BatchNorm layers found) fails with exit 2. The first --arch torchvision: run after installing torchvision is slow while Python compiles it; later runs add about a second.
• Decayed channels are common in published models. In our scan, 19 of 21 torchvision BN→ReLU models carried some. Failing on warning may be too strict for a model you did not train. Compare against a baseline with --against and gate on the newly_dead finding instead, for example with --json and a short script:

ormas-monitor check new.safetensors --against base.safetensors --json \
  | python -c "import json,sys; r=json.load(sys.stdin); sys.exit(any(f['kind']=='newly_dead' for f in r['findings']))"