pytorch - 💡(How to fix) Fix DTensor.local_map inconsistently handles pytree on input versus placements [1 participants]

TL;DR

Modify the comparison logic to handle placements as a nested structure instead of a flat list.

Guidance

• Review the current implementation of pytree flattening and placement comparison to identify the root cause of the issue.
• Consider modifying the comparison logic to recursively handle placements as a nested structure, rather than flattening them into a list.
• Verify that the modified comparison logic correctly handles different scenarios, including nested placements and empty lists.
• Test the changes with various input cases to ensure the new behavior is consistent and accurate.

Notes

The exact implementation details are not provided, so the suggested modifications are high-level and may require additional changes to the underlying code.

Recommendation

Apply workaround: Modify the comparison logic to handle placements as a nested structure, as this approach addresses the root cause of the issue and provides a more accurate comparison.