OSSForks/pytorch
1
0
Fork 0
mirror of https://github.com/zebrajr/pytorch.git synced 2025-12-06 12:20:52 +01:00
Code Issues Actions 38 Packages Projects Releases Wiki Activity

Drop dynamo-specific type hints on Tensor in favor of type-ignores (#113720)

Browse Source 
Per [this][1] discussion, plus some offline discussion. The summary:
@albanD considers the core PyTorch types like Tensor to be extremely
brittle, and does not think the risk of adding these typed attributes to
be worth it.

@eellison mentioned that we could use `WeakTensorKeyDictionary` instead.
However, based on the sparse usage of these bonus attributes, I think
that would be overkill. So I've opted to go with a few more type-ignore
comments instead.

[1]: https://github.com/pytorch/pytorch/pull/113610#discussion_r1392907367

Pull Request resolved: https://github.com/pytorch/pytorch/pull/113720
Approved by: https://github.com/ezyang, https://github.com/albanD, https://github.com/eellison
ghstack dependencies: #113534, #113610
This commit is contained in:
Jez Ng 2023-11-14 17:43:59 -08:00 committed by PyTorch MergeBot
parent 605d274300
commit a3b859fc67
3 changed files with 5 additions and 8 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

4
torch/_dynamo/decorators.py
View File

@ -249,9 +249,9 @@ def mark_static_address(t, guard=True):
raise TypeError(f"mark_static_address expects a tensor but recieved {type(t)}") raise TypeError(f"mark_static_address expects a tensor but recieved {type(t)}")
if guard: if guard:
t._dynamo_static_input_type = "guarded" t._dynamo_static_input_type = "guarded" # type: ignore[attr-defined]
else: else:
t._dynamo_static_input_type = "unguarded" t._dynamo_static_input_type = "unguarded" # type: ignore[attr-defined]
# Note: this carefully avoids eagerly import einops. # Note: this carefully avoids eagerly import einops.

4
torch/_dynamo/utils.py
View File

@ -636,7 +636,7 @@ def clone_input(x, *, dtype=None):
if x.is_leaf and x.grad is not None: if x.is_leaf and x.grad is not None:
y.grad = clone_input(x.grad, dtype=dtype) y.grad = clone_input(x.grad, dtype=dtype)
if hasattr(x, "_dynamo_dynamic_indices"): if hasattr(x, "_dynamo_dynamic_indices"):
y._dynamo_dynamic_indices = x._dynamo_dynamic_indices.copy() y._dynamo_dynamic_indices = x._dynamo_dynamic_indices.copy() # type: ignore[attr-defined]
return y return y
with torch.no_grad(): with torch.no_grad():
@ -669,7 +669,7 @@ def clone_input(x, *, dtype=None):
# performing the operation. # performing the operation.
return torch_clone(x) return torch_clone(x)
if hasattr(x, "_dynamo_dynamic_indices"): if hasattr(x, "_dynamo_dynamic_indices"):
result._dynamo_dynamic_indices = x._dynamo_dynamic_indices.copy() result._dynamo_dynamic_indices = x._dynamo_dynamic_indices.copy() # type: ignore[attr-defined]
return result return result

5
torch/_tensor.py
View File

@ -5,7 +5,7 @@ import warnings
from collections import OrderedDict from collections import OrderedDict
from copy import deepcopy from copy import deepcopy
from numbers import Number from numbers import Number
from typing import Any, Dict, Optional, Set, Tuple, Union from typing import Any, Dict, Optional, Tuple, Union
import torch import torch
import torch._C as _C import torch._C as _C
@ -79,9 +79,6 @@ def _rebuild_from_type_v2(func, new_type, args, state):
# torch/_C/__init__.pyi.in to add a type annotation for your method; # torch/_C/__init__.pyi.in to add a type annotation for your method;
# otherwise, it will not show up in autocomplete. # otherwise, it will not show up in autocomplete.
class Tensor(torch._C.TensorBase): class Tensor(torch._C.TensorBase):
_dynamo_static_input_type: str # Set in `mark_static_address`
_dynamo_dynamic_indices: Set[int] # Set in `mark_dynamic`
def __deepcopy__(self, memo): def __deepcopy__(self, memo):
if has_torch_function_unary(self): if has_torch_function_unary(self):
return handle_torch_function(Tensor.__deepcopy__, (self,), self, memo) return handle_torch_function(Tensor.__deepcopy__, (self,), self, memo)