This example:

mylist = []

@torchdynamo.optimize(...)
def foo():
   mylist.append(torch.randn(10) + 1)


foo()
foo()
foo()

We will compile 3 versions of foo:

• One guarding that mylist is len 0
• One guarding that mylist is len 1 (and has one matching tensor)
• One guarding that mylist is len 2 (and has two matching tensor)

Really, only 1 is version is needed.

Similarly. A different function calling mylist.clear() will guard on the contents of the list, despite never looking at the contents.

This may cause an issue for python autograd, as currently TorchDynamo will guard on the entire contents of the tape.

cc @ezyang @bdhirsh @anijain2305 @chauhang @voznesenskym @penguinwu @EikanWang @jgong5 @Guobing-Chen @XiaobingSuper @zhuhaozhe @blzheng @wenzhe-nrv @jiayisunx @chenyang78 @kadeng @soumith @msaroufim @wconstab @ngimel @mlazos @yanboliang @Xia-Weiwen @desertfire @zdevito