[
https://issues.apache.org/jira/browse/ARROW-1701?page=com.atlassian.jira.plugin.system.issuetabpanels:comment-tabpanel&focusedCommentId=16345877#comment-16345877
]
ASF GitHub Bot commented on ARROW-1701:
---------------------------------------
ppwwyyxx commented on issue #1223: ARROW-1701: [Serialization] Support zero
copy PyTorch Tensor serialization
URL: https://github.com/apache/arrow/pull/1223#issuecomment-359892433
Could you make pyarrow NOT do `import torch` by default?
Pytorch is a giant library with multiple C extensions, so importing it has a
lot of side-effect especially about symbol version conflict, e.g. issue 3059,
926, 4111, 4809 in pytorch, and issue#91 in uber/horovod.
I've seen other conflicting issues created by `import torch` in addition to
those above. More recently I found this import came from pyarrow after
debugging my segfaults for a long time.
----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on GitHub and use the
URL above to go to the specific comment.
For queries about this service, please contact Infrastructure at:
[email protected]
> [Serialization] Support zero copy PyTorch Tensor serialization
> --------------------------------------------------------------
>
> Key: ARROW-1701
> URL: https://issues.apache.org/jira/browse/ARROW-1701
> Project: Apache Arrow
> Issue Type: Improvement
> Reporter: Philipp Moritz
> Assignee: Philipp Moritz
> Priority: Major
> Labels: pull-request-available
> Fix For: 0.8.0
>
>
> see http://pytorch.org/docs/master/tensors.html
> This should be optional and only included if the user has PyTorch installed.
--
This message was sent by Atlassian JIRA
(v7.6.3#76005)
