search

csarofeen / pytorch

Tensors and Dynamic neural networks in Python with strong GPU acceleration
http://pytorch.org
Other
26 stars 7 forks source link

Drop contiguity info for size 1 dimensions for Python's define_tensor(sizes=..., strides=...) #2551

Closed IvanYashchuk closed 1 year ago

IvanYashchuk commented 1 year ago

Fixes https://github.com/csarofeen/pytorch/issues/2549.

jjsjann123 commented 1 year ago

Not trying to block this PR.

think about a tensor torch.randn(5, 5).expand(-1).expand((5, 5, 5))

We still have issue with explicitly broadcast dimensions that I don't think is handled by this PR yet. With the API on TensorOpRecord, we can't properly pass correct contiguity information for an expanded dimension. I'm trying to keep the contiguity information more comprehensible on the python side (not skipping contiguity for broadcast dims).

Wondering if you want to patch that in this PR as well? otherwise I'll start one once this merges. cc'ing @IvanYashchuk

IvanYashchuk commented 1 year ago

Closing in favor of https://github.com/csarofeen/pytorch/pull/2561