When fullgraph=True and the eager_on_recompile stance is active, recompile attempts cause the "torch.compile with fullgraph=True found no compiled frames." error. This is not good behavior since in this stance, the user likely wants to allow fallbacks to eager. fn(torch.ones(3), 2) # fullgraph=True causes error

1. Warn instead of error for the 2.12 release.
2. Do not do the fallback-to-eager error if a non-default stance is active.

TL;DR

Modify the torch.compile behavior to warn instead of error when fullgraph=True and eager_on_recompile stance is active.

Guidance

• Identify the torch.compile function calls with fullgraph=True and adjust the error handling to warn instead of error when eager_on_recompile stance is active.
• Verify that the fallback to eager mode works as expected by checking the logs for skip reasons.
• Consider implementing a comprehensive logging system to track why tracing is skipped.
• Review the code for any instances where the fallback-to-eager error is triggered when a non-default stance is active and modify the behavior accordingly.

Example

import torch

@torch.compile(fullgraph=True, backend="eager", dynamic=False)
def fn(x, n):
  try:
    return x + n
  except torch.errors.CompileError as e:
    # Handle the error or warn instead
    print(f"Warning: {e}")

Notes

The provided solution assumes that the torch.compile function and eager_on_recompile stance are correctly implemented and configured. The example code snippet is a minimal illustration and may require adjustments based on the actual codebase.

Recommendation

Apply workaround: Modify the torch.compile behavior to warn instead of error when fullgraph=True and eager_on_recompile stance is active, as this allows for fallbacks to eager mode and provides a more robust error handling mechanism.