Fix aten.copy device mismatch bug in FakeTensor #102664
Conversation
[ghstack-poisoned]
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/102664
Note: Links to docs will display an error until the docs builds have been completed. ✅ No FailuresAs of commit 6187fbc: This comment was automatically generated by Dr. CI and updates every 15 minutes. |
Fixes `pytest ./generated/test_yizhou_wang_RODNet.py -k test_000` failure in #92670. FakeTensor would raise an error upon trying to run `aten.copy` with inputs with different devices, although this is allowed behavior. [ghstack-poisoned]
Fixes `pytest ./generated/test_yizhou_wang_RODNet.py -k test_000` failure in #92670. FakeTensor would raise an error upon trying to run `aten.copy` with inputs with different devices, although this is allowed behavior. [ghstack-poisoned]
|
Can you coordinate with @davidberard98 to add other ops that support different devices? Thanks! |
Fixes `pytest ./generated/test_yizhou_wang_RODNet.py -k test_000` failure in #92670. FakeTensor would raise an error upon trying to run `aten.copy` with inputs with different devices, although this is allowed behavior. [ghstack-poisoned]
|
@pytorchbot merge |
Merge failedReason: This PR needs a If not, please add the To add a label, you can comment to pytorchbot, for example For more information, see Details for Dev Infra teamRaised by workflow job |
|
@pytorchbot merge |
Merge startedYour change will be merged once all checks pass (ETA 0-4 Hours). Learn more about merging in the wiki. Questions? Feedback? Please reach out to the PyTorch DevX Team |
| j1 = torch.tensor([2], device="cuda") | ||
| j2 = torch.tensor([2], device="cpu") | ||
| r3 = torch.ops.aten.index_put.default(x1, j1, y1) | ||
| r4 = torch.ops.aten.index_put.default(x2, j2, y2) |
There was a problem hiding this comment.
This code doesn't actually pass on eager though?
(/home/ezyang/local/b/pytorch-env) [ezyang@devgpu005.nha1 ~/local/b/pytorch (fc2aa23c)]$ python test.py
Traceback (most recent call last):
File "/data/users/ezyang/b/pytorch/test.py", line 7, in <module>
r1 = torch.ops.aten.index(x1, i1)
File "/data/users/ezyang/b/pytorch/torch/_ops.py", line 1018, in __call__
return self_._op(*args, **(kwargs or {}))
RuntimeError: indices should be either on cpu or on the same device as the indexed tensor (cpu)
(/home/ezyang/local/b/pytorch-env) [ezyang@devgpu005.nha1 ~/local/b/pytorch (fc2aa23c)]$ cat test.py
import torch
x1 = torch.rand(4, 4, device="cpu")
x2 = torch.rand(4, 4, device="cuda")
i1 = torch.tensor([0, 1], device="cuda")
i2 = torch.tensor([0, 1], device="cpu")
r1 = torch.ops.aten.index(x1, i1)
r2 = torch.ops.aten.index(x2, i2)
y1 = torch.rand(4, device="cpu")
y2 = torch.rand(4, device="cuda")
j1 = torch.tensor([2], device="cuda")
j2 = torch.tensor([2], device="cpu")
r3 = torch.ops.aten.index_put.default(x1, j1, y1)
r4 = torch.ops.aten.index_put.default(x2, j2, y2)
5 participants
Stack from ghstack (oldest at bottom):
Fixes
pytest ./generated/test_yizhou_wang_RODNet.py -k test_000failure in #92670.FakeTensor would raise an error upon trying to run
aten.copywith inputs with different devices, although this is allowed behavior.Also fix
aten.slice_scatter, since it also takes args with different devices.