[MPS] Fix non-contig to contig tensor copy #86056
Conversation
🔗 Helpful Links🧪 See artifacts and rendered test results at hud.pytorch.org/pr/86056
Note: Links to docs will display an error until the docs builds have been completed. ❌ 3 FailuresAs of commit 557f10b: The following jobs have failed:
This comment was automatically generated by Dr. CI and updates every 15 minutes. |
pytorch-bot
bot
added
ciflow/mps
| @@ -186,7 +186,9 @@ void copy_cast_mps(at::Tensor& dst, const at::Tensor& src, | |||
| // For View tensors, the storage offset can be bigger than what's being reported by nbytes | |||
| src_total_size = at::detail::computeStorageNbytesContiguous(src.sizes(), src.element_size(), src.storage_offset()); | |||
| } else { | |||
| src = src_; | |||
| // Views are not the only source of non-contiguous tensors | |||
| // For example, tensor constructed from ndarray can be non-contagious | |||
There was a problem hiding this comment.
| // For example, tensor constructed from ndarray can be non-contagious | |
| // For example, tensor constructed from ndarray can be non-contiguous |
nit
`Tensor::contiguous()` is a no-op if tensor is contiguous Fixes #85967
malfet
force-pushed
the
malfet/mps-fix-non-contig-copy
branch
from
7dcd968
to
557f10b
Compare
| # See https://github.com/pytorch/pytorch/issues/85967 | ||
| def test_from_numpy_non_contiguous(self): | ||
| a = np.arange(9).reshape(3, 3)[:, :2] | ||
| t_cpu = torch.tensor(a, device="cpu") |
There was a problem hiding this comment.
does this tensor constructor copy the passed data and makes always contig tensors? should this not be torch.as_tensor instead?
There was a problem hiding this comment.
Indeed it does, but that's the point: MPS copy operator did not handled this case correctly
malfet
changed the title
|
@pytorchbot merge -f "MPS tests are green" |
|
@pytorchbot successfully started a merge job. Check the current status here. |
|
Hey @malfet. |
This handles a rare case when MPS tensor is constructed from non-contiguous CPU tensor. Fixes #85967 Pull Request resolved: #86056 Approved by: https://github.com/janeyx99
This handles a rare case when MPS tensor is constructed from non-contiguous CPU tensor.
Fixes #85967