Fix Plan

To fix the torch._dynamo.exc.Unsupported: Bad import result error, we need to modify the torch.compile function to recognize types.ModuleType subclasses as valid Python modules.

Here are the concrete steps:

• Modify the torch._dynamo module to check if the imported module is an instance of types.ModuleType using isinstance.
• If it is, proceed with the import without raising an error.

Example code:

import sys
import types
import torch

class _ConfigModule(types.ModuleType):
    x = 1

_ConfigModule.__module__ = __name__
sys.modules["my_config"] = _ConfigModule("my_config")

# Patch torch._dynamo to recognize types.ModuleType subclasses
import torch._dynamo as dynamo
original_import = dynamo.import_module
def patched_import(name, *args, **kwargs):
    module = original_import(name, *args, **kwargs)
    if isinstance(module, types.ModuleType):
        return module
    raise dynamo.exc.Unsupported(f"Bad import result: {name}")
dynamo.import_module = patched_import

@torch.compile(fullgraph=True)
def f():
    import my_config  # noqa: F401
    return torch.tensor(1)

f()

Verification

To verify that the fix worked, run the modified code and check that the torch._dynamo.exc.Unsupported: Bad import result error is no longer raised.

Extra Tips

Note that this fix is a temporary workaround and may need to be revisited in future versions of PyTorch. Additionally, be cautious when modifying internal PyTorch modules, as this can have unintended consequences.