frozenleaves/pytorch
1
0
Fork 0
mirror of https://github.com/pytorch/pytorch.git synced 2025-10-20 21:14:14 +08:00
Code Issues Packages Projects Releases Wiki Activity

removed zero dim cpu logic from fake_tensor.py (#147501)

Browse Source 
Fixes #144748
In #144748, the inconsistency between the eager mode and the inductor mode is reported as a bug.
The root cause is fake_tenosr.py's find-common-device method, 0b0da81021/torch/_subclasses/fake_tensor.py (L833), takes zero dim cpu tensor into account but  the device check in adaption.h doesn't.

This fix is to add a list for some ops to bypass zero-dim-cpu-tensor check to align with the eager mode.

Pull Request resolved: https://github.com/pytorch/pytorch/pull/147501
Approved by: https://github.com/ezyang
This commit is contained in:
zero000064
2025-07-21 21:11:06 +00:00
committed by PyTorch MergeBot
parent 5e17932c22
commit 9e0473b566
2 changed files with 27 additions and 2 deletions
Download Patch File Download Diff File Expand all files Collapse all files

13
torch/_subclasses/fake_tensor.py
View File

@ -889,6 +889,11 @@ class FakeTensor(Tensor):
aten._foreach_copy.default,
)
# list of ops not using zero dim cpu tensor logic to align with the eager mode.
bypass_zero_dim_cpu_tensor_check_ops = ordered_set(
aten.nextafter.default,
)
def check_cpu_device(device: torch.device) -> bool:
return device.type == "cpu"
@ -912,13 +917,17 @@ class FakeTensor(Tensor):
is_cpu_zero_dim = t_is_cpu_zero_dim
return
is_bypass_zero_dim_cpu_tensor_check_op = (
func in bypass_zero_dim_cpu_tensor_check_ops
)
# mismatching devices !
# if current tensor is cpu 0 dim, defer to existing device
if t_is_cpu_zero_dim:
if t_is_cpu_zero_dim and not is_bypass_zero_dim_cpu_tensor_check_op:
return
# current device is from cpu 0 dim tensor, overwrite
if is_cpu_zero_dim:
if is_cpu_zero_dim and not is_bypass_zero_dim_cpu_tensor_check_op:
common_device = t.device
is_cpu_zero_dim = t_is_cpu_zero_dim
return