fix(utils): fall back to INFO for a blank or unknown LOG_LEVEL

[Lesereingrape]

PR type

• Bug Fix
• New Feature
• Document Updates
• More Models or Datasets Support

LOG_LEVEL is read twice in swift/utils/logger.py, and the two read sites disagree about what to do with a value that is not a level name.

• get_logger() normalises it with getattr(logging, log_level, logging.INFO), so an unusable value falls back to INFO.
• The module-level statement passed the raw string straight to ms_logger.setLevel(log_level), and Logger.setLevel() raises for anything that is not a level name.

So the tolerant path is the one that never gets to run:

$ LOG_LEVEL= python -c "from swift.utils import get_logger"
Traceback (most recent call last):
  ...
  File ".../swift/utils/logger.py", line 121, in <module>
    ms_logger.setLevel(log_level)
  ...
ValueError: Unknown level: ''

An empty LOG_LEVEL is easy to get by accident — ENV LOG_LEVEL= in a Dockerfile, LOG_LEVEL= in a launcher script or .env, or a variable that a scheduler still exports while empty. In that state every module that imports the logger fails, so the job dies at import time over a cosmetic setting, and the same happens for a value padded with blanks or misspelled. The documented contract is only "LOG_LEVEL: The log level, default is 'INFO'. You can set it to 'WARNING', 'ERROR', etc." — nothing promises an exception, and get_logger() already treats a bad value as INFO.

What changed

Both read sites now share one _get_log_level() helper that strips the value, upper-cases it and always returns a numeric level. Valid settings keep working exactly as before, including the spellings that already worked:

LOG_LEVEL	before	after
unset	INFO	INFO
WARNING	WARNING	WARNING
warning	WARNING	WARNING
''	ValueError at import	INFO
' '	ValueError at import	INFO
VERBOSE	ValueError at import	INFO

How it was checked

tests/utils/test_log_level.py resolves both loggers in a fresh interpreter, because the failure happens at import time and cannot be reproduced in-process. Run with the same runner CI uses:

python -m unittest tests.utils.test_log_level -v

• without the patch: Ran 6 tests ... FAILED (failures=3) — the blank, the whitespace-padded and the unknown value, with the traceback above;
• with the patch: Ran 6 tests ... OK, including the two controls that pin WARNING/warning to level 30 and unset to 20.

pre-commit run --files swift/utils/logger.py tests/utils/test_log_level.py passes (flake8, isort, yapf and the formatting hooks). The rest of tests/utils is unaffected: in my environment it reports errors=38, skipped=22 both with and without this patch — those errors are optional dependencies missing locally (e.g. pydantic), not this change.

No linked issue.

---

Disclosure: this PR was prepared, tested and submitted by an AI agent working on behalf of the account owner.