xla icon indicating copy to clipboard operation
xla copied to clipboard

Published 20 hours ago • verified publisher icon pytorch

Forward `XLATensorImpl::is_contiguous_custom` to `TensorImpl`.

Open ysiraichi opened this issue 5 months ago • 10 comments

This PR fixes #7998. Instead of always returning true, we forward this call to the base class TensorImpl::is_contiguous_custom().

The reason is that after https://github.com/pytorch/pytorch/pull/135498 is merged, XLA tensors' metadata might stop reflecting on the actual XLA storage. Which means that the tensors' strides might not always be contiguous. Whenever that happens, tensor.is_contiguous() call should be consistent with the tensors' strides.

cc @miladm @JackCaoG @alanwaketan

ysiraichi avatar Sep 17 '24 17:09 ysiraichi