Setup rccl for PyTorch

[jithunnair-amd]

No description provided.

[jithunnair-amd]

@pytorchbot retest this please

[jithunnair-amd]

@iotamudelta unrelated issue, but thought I'd bring it to your attention: the caffe2 test build failed with segfault for test_Conv2d_groups_thnn. Is it another candidate for skipping?

[iotamudelta]

@jithunnair-amd yeah, it's been giving us trouble. I'd like to understand why all of these tests suddenly have issues.

@lcskrishna could you have a look into this? thanks!

[iotamudelta]

@pytorchbot retest this please

[jithunnair-amd]

@iotamudelta

03:19:23 + python setup.py install --user
03:19:24 Traceback (most recent call last):
03:19:24   File "setup.py", line 191, in <module>
03:19:24     from tools.setup_helpers.mkldnn import (USE_MKLDNN, MKLDNN_LIBRARY,
03:19:24 ImportError: No module named mkldnn

This seems like a conflict merge gone wrong? I think they removed mkldnn.py upstream.

[iotamudelta]

@jithunnair-amd yes, i needed to do some merging after the last IFU in this PR - please do check.

[jithunnair-amd]

@iotamudelta At this point, I think all the changes needed to integrate rccl into "c10d->rccl" path are in this PR. Once the CI docker gets updated to have the rccl package in it, please let me know and I can make it a REQUIRED dependency in the cmake.

[iotamudelta]

@jithunnair-amd if you can resolve the conflicts, this should now be mergable with the CI update :-)

[jithunnair-amd]

I decided to reapply the changes at the top of tree instead of resolving conflicts, since they changed a few setup files. I'll file a new PR once they respond to my query about a recent change that causes a compilation issue when building distributed library on ROCm. Get Outlook for Android<https://aka.ms/ghei36>

…

________________________________ From: iotamudelta <notifications@github.com> Sent: Saturday, February 9, 2019 10:38:23 PM To: ROCmSoftwarePlatform/pytorch Cc: Nair, Jithun; Mention Subject: Re: [ROCmSoftwarePlatform/pytorch] [WIP] Setup rccl for PyTorch (#309) @jithunnair-amd<https://github.com/jithunnair-amd> if you can resolve the conflicts, this should now be mergable with the CI update :-) — You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub<#309 (comment)>, or mute the thread<https://github.com/notifications/unsubscribe-auth/AkIT-KRzBspLu7HjDQERbXibngFFWHzWks5vL6I_gaJpZM4YPVOR>.

[iotamudelta]

@jithunnair-amd IFU completed.

[jithunnair-amd]

Reworked changes in #347, so closing this PR