[mortenoh]

Summary

chapkit mlproject export writes the module path into the MLproject header. When the module was not in the MLproject's folder, that was the absolute path, for example /Users/<name>/.../src/my_model/service.py. A committed MLproject then carried a machine-specific path, and export --check failed on every other machine, CI included, because it compares the whole file. A src/-layout service had to keep a root main.py only to get a portable header.

• The header names the module relative to the MLproject's folder, for example src/my_model/service.py. The absolute path remains only where no relative path exists, such as a module on another drive on Windows.
• When the MLproject is not next to the module, the header's regenerate command adds -o <name>, because export otherwise writes next to the module.
• A module next to its MLproject, the usual main.py case, is unchanged: the header still reads main.py.

The MLproject guide describes exporting a src/-layout service to the project root.

Test plan

• New test: a module in src/pkg/ exported to the project root gets a relative header with no temporary path in it, and the header's command run from the project folder passes --check.
• Existing export tests, including -o into another folder and into the module's folder.
• make lint; full test suite: 1095 passed, 12 skipped.
• Exported chapkit_deep_auto_regressive_model from src/chapkit_deep_auto_regressive/service.py: the header is relative and the MLproject body is unchanged.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!