Fixing a Broken Deployment¶

Use the fix CLI command to repair a broken pipeline definition automatically.

Quick fix¶

inference-engine fix models/my_model/v1/

The CLI will: 1. Read the existing definition.py 2. Ask for a sample input 3. Validate the pipeline — if it passes, nothing to do 4. If it fails, send the error + code to the LLM for a fix 5. Re-validate; retry up to 3 times 6. Show a unified diff and ask for confirmation before writing

What gets rewritten¶

Only load() and predict() are ever rewritten. The rest of definition.py — imports, MODEL_NAME, build_pipeline structure — is preserved.

If all retries fail¶

The command exits with an error and the original file is left unchanged. You can then edit definition.py manually and re-run fix.

Manual repair¶

If you prefer to fix manually:

Open models/<name>/<version>/definition.py
Check the load() method — ensure the artifact path is correct and the model loads without error
Check the predict() method — ensure it accepts the preprocessor output format
Test with a direct Python call before restarting the server

Common issues¶

Symptom	Likely cause
`FileNotFoundError` in `load()`	Artifact path is wrong or file was not copied
`TypeError` in `predict()`	Input format mismatch — check preprocessor output
`ValidationError` on every request	Validator rejects the preprocessed input shape
`ModelNotFoundError`	`MODEL_NAME` / `MODEL_VERSION` don't match the request

See Troubleshooting for more.