Commit Graph

2 Commits

Author SHA1 Message Date
Reece O'Mahoney f65f3f7a4a Fix policy.path in YAML configs (PR #3145 followup) (#3597)
PR #3145 added YAML support for policy.path but left two bugs:

1. extract_path_fields_from_config only deleted config_data[field] when
   no sibling overrides existed. With siblings, the dict stayed in place
   and draccus crashed decoding it as PreTrainedConfig (no 'type' key).
   Sibling overrides go into _config_yaml_overrides and are applied later
   by from_pretrained(), so the field can always be removed.

2. wrap() updated config_path_cli to the cleaned temp file path but
   never propagated it to the draccus.parse fallback branch. cli_args
   still contained --config_path=<original>, so draccus read the
   original YAML with path: still present.

Tests passed because they (a) called extract_path_fields_from_config
directly and (b) included type: alongside path: in the YAML, sidestepping
both bugs.

Co-authored-by: Steven Palma <imstevenpmwork@ieee.org>
2026-05-26 14:01:19 +02:00
Jash Shah fe96b28c74 Fix policy.path not working in YAML config files (#3145)
* fix(config): support policy.path in YAML config files

policy.path was only handled via CLI args (filtered from sys.argv before
draccus, then retrieved in validate()). When specified in YAML, draccus
would crash because 'path' is not a valid field on PreTrainedConfig.

Extract path fields from the YAML/JSON config before draccus processes
it, store them in a module-level dict, and fall back to it in
get_path_arg() when the CLI doesn't have the path.

Fixes #2957

* fix(parser): preserve YAML policy overrides when loading from pretrained

When policy.path is set in YAML, validate() was calling from_pretrained
with only CLI overrides, discarding any YAML policy fields (e.g. lr,
batch_size) that draccus had already parsed. Fix by capturing the
remaining YAML fields as CLI-style args in _config_yaml_overrides and
merging them into the overrides passed to from_pretrained in train.py,
eval.py, and lerobot_record.py (CLI args still take precedence).

Also fix the NamedTemporaryFile SIM115 ruff warning and add types-PyYAML
to the mypy pre-commit hook.

* fix(parser): serialize bool/None values correctly in YAML policy overrides

Bool values from YAML configs (e.g. push_to_hub: true) were passed as
Python "True"/"False" strings instead of lowercase "true"/"false" that
draccus expects. Also skip None values to avoid passing "None" strings.

* revert: remove types-PyYAML from .pre-commit-config.yaml

* chore: fix quality check caused by untyped YAML import

Co-authored-by: masato-ka <jp6uzv@gmail.com>
Signed-off-by: Khalil Meftah <khalil.meftah@huggingface.co>

---------

Signed-off-by: Khalil Meftah <khalil.meftah@huggingface.co>
Co-authored-by: Khalil Meftah <khalil.meftah@huggingface.co>
Co-authored-by: masato-ka <jp6uzv@gmail.com>
2026-05-13 09:45:27 +02:00