pytorch

1456515f - [JIT] Disallow plain List type annotation without arg (#38130)

[JIT] Disallow plain List type annotation without arg (#38130) Summary: **Summary** This commit detects and prohibits the case in which `typing.List` is used as an annotation without a type argument (i.e. `typing.List[T]`). At present, `typing.List` is always assumed to have one argument, and when it is used without one, `typing.List.__args__[0]` is nonempty and set to some `typing.TypeVar` instance, which has no JIT type equivalent. Consequently, trying to convert `typing.List` to a JIT type results in a `c10::ListType` with `nullptr` for its element type, which can cause a segmentation fault. This is fixed by returning a `ListType` from `jit.annotations.try_ann_to_type` only if the element type is converted successfully to a JIT type and returning `None` otherwise. **Test Plan** I ran the code from the issue (https://github.com/pytorch/pytorch/issues/37530) that reported this problem and also ran some unit tests. *Before* ``` $ python3 segfault.py Segmentation fault (core dumped) ``` *After* ``` $ python3 segfault.py Traceback (most recent call last): ... RuntimeError: Unknown type name 'List': File "segfault.py", line 9 classmethod def cat(cls, box_lists: List): ~~~~ <--- HERE return cls(torch.cat([x for x in box_lists])) 'Boxes.cat' is being compiled since it was called from 'Boxes' File "segfault.py", line 13 def f(t: torch.Tensor): b = Boxes(t) ~~~~~ <--- HERE c = Boxes(torch.tensor([3, 4])) return Boxes.cat([b, c]) 'Boxes' is being compiled since it was called from 'f' File "segfault.py", line 13 def f(t: torch.Tensor): b = Boxes(t) ~~~~~~~~~~~ <--- HERE c = Boxes(torch.tensor([3, 4])) return Boxes.cat([b, c]) ``` **Fixes** This pull request fixes https://github.com/pytorch/pytorch/issues/37530. Pull Request resolved: https://github.com/pytorch/pytorch/pull/38130 Differential Revision: D21485284 Pulled By: SplitInfinity fbshipit-source-id: 9b51ef6340485a24c8b7cfb85832d4668b8ac51a