Problem

When specifying devices_list in TrainerConfig, users encountered an OmegaConf validation error:

from pytorch_tabular.config import TrainerConfig
from omegaconf import OmegaConf
# This fails with ValidationError
trainer_config = TrainerConfig(devices_list=[1, 2, 3, 4])
config = OmegaConf.structured(trainer_config)

Error:

omegaconf.errors.ValidationError: Value '[1, 2, 3, 4]' of type 'list' could not be converted to Integer
 full_key: devices
 object_type=TrainerConfig

Root Cause

The devices field in TrainerConfig was typed as Optional[int], but the __post_init__ method assigns the devices_list (a list) to devices when provided. When OmegaConf wraps the config with OmegaConf.structured(), it validates types and rejects the list assignment.

Solution

Changed the type annotation of the devices field from Optional[int] to Any. This approach:

• Allows both integer and list values, matching PyTorch Lightning's Trainer API which accepts both
• Works correctly with OmegaConf (note: Union[int, List[int]] is not supported by OmegaConf for container types)
• Maintains full backward compatibility with existing code

Changes

src/pytorch_tabular/config/config.py:

• Changed devices field type from Optional[int] to Any
• Updated docstring to indicate devices accepts Union[int, List[int]]

tests/test_config.py (new file):

• Added comprehensive test coverage for devices/devices_list functionality
• Tests cover the documented use case (devices_list=[0, 1])
• Tests verify backward compatibility with integer devices values
• Tests validate config merging scenarios as used in TabularModel

Testing

All scenarios now work correctly:

# Multiple GPUs (from the issue)
TrainerConfig(devices_list=[1, 2, 3, 4]) # ✅ Works
# Documented example
TrainerConfig(devices_list=[0, 1]) # ✅ Works
# Backward compatibility
TrainerConfig(devices=2) # ✅ Still works
TrainerConfig() # ✅ Default devices=-1 still works

Backward Compatibility

✅ Fully backward compatible - all existing code using devices=<int> continues to work unchanged.

Fixes #issue_number

✨ Let Copilot coding agent set things up for you — coding agent works faster and does higher quality work when set up for your repo.