pytorch - 💡(How to fix) Fix [functorch] popDynamicLayerStackToDepth double-pops the dynamic layer stack [1 comments, 2 participants]

This means when Dynamo's error recovery fires for compiled functorch transforms, the cleanup itself crashes, masking the original error with a cryptic internal assertion failure. _vmap_increment_nesting(3, "error")

Fix Plan

The fix involves modifying the popDynamicLayerStackToDepth function to correctly pop one layer per iteration.

Step-by-Step Solution

• Modify the popDynamicLayerStackToDepth function to use a while loop that checks the current depth and pops layers until it reaches the target depth.
• Ensure that the loop correctly handles the case where the target depth is greater than the current depth.

Example code:

def pop_dynamic_layer_stack_to_depth(target_depth):
    while get_dynamic_layer_stack_depth() > target_depth:
        # Pop one layer
        pop_dynamic_layer_stack_and_undo_to_depth(get_dynamic_layer_stack_depth() - 1)

• Replace the existing pop_dynamic_layer_stack_to_depth function with the modified version.

Verification

• Run the reproducer code to verify that the stack unwinds correctly to the target depth.
• Test the function with different target depths to ensure it works as expected.

Extra Tips

• Make sure to test the modified function thoroughly to catch any edge cases.
• Consider adding additional logging or error handling to help diagnose any issues that may arise.