[Dynamo][Bug] _debug_get_cache_entry_list causes core dump for resume functions

[youkaichao]

🐛 Describe the bug

A minimal working example:

import torch

@torch.compile
def f(x):
    s = "Hello" + x
    print(s)

f(", world!")

from torch._dynamo.eval_frame import _debug_get_cache_entry_list

name = ""
for x in list(globals().keys()):
    if x.startswith("__resume"):
        name = x # should be __resume_at_14_0

code = globals()[name]

print(_debug_get_cache_entry_list(code))

Versions

PyTorch 2.3.0.dev20240225.

cc @ezyang @msaroufim @bdhirsh @anijain2305 @zou3519

[youkaichao]

cc @jansel @ezyang @anijain2305 @williamwen42

[youkaichao]

Hi, friends, I have been away for about 1 month. This bug seems to be related with the refactoring of Dynamo. It would be invaluable if anyone can provide some background or pointers on what's going on now :)

[williamwen42]

We moved a bunch of code from eval_frame.c to C++ here: #118438.

[youkaichao]

Thanks for the very fast fix!