• What's the plan for getting PPC machines connected to Jenkins, with enough capacity so that they can keep up with the volume of builds we request?
• Can we cross-compile for PPC? That will help reduce the number of PPC machines we need for testing.
• Can we avoid copy pasting the scripts?

@ezyang answer to your questions:

1. There is currently a CI build environment from https://powerci.osuosl.org/ . It is a shared environment for other jobs. I'm already doing build/test there for pytorch master. When I was working with @yf225 , he said pytorch PR builds should at least accommodate 4 simultaneous builds (the environment should be able to accommodate that). I believe @yf225 also already has the hooks from your ci to be able to trigger builds on our CI. The plan is to start with a single build/test image using Ubuntu CUDA image from nvidia, CUDA 9.1 , cudnn 7.x , python3 from anaconda. When the environment is ready, we can switch to cuda 9.2. I am still trying to eliminate random environment failures before we can integrate with your CI. We can talk about it more when I am ready. For now I hope to get the scripts ready for ppc64le.
2. Cross-compile may work (I've never tried it or set it up with nvcc). We do have a toolchain cross compiler that we can setup for x86 (I've used this on other non-CUDA projects). If you're willing to try it, I can help test the produced binaries. Just FYI , when building pytorch on ppc64le it takes between 7-14 minutes to finish when using ccache (so it does not take a lot of resources).
3. I'll take your suggestion on using your build.sh and test.sh and ifdef for ppc64le to avoid duplication of code in different files.

So, I looked at your modified patch: you're still doing what's effectively a copy paste, since you have a giant if-block conditioned on ppc, versus not. If you're going to do that, you might as well just have another file.

What I would like to see is a localized block that sets the extra environment variables that PPC needs, and no reindenting of the build script. It would be even better if we had a PPC docker image (similar to the docker images we have for our Linux builds), which can have these environment variables pre-set. The definitions for these currently live in https://github.com/pietern/pytorch-dockerfiles

@ezyang Ok , I got it. For build.sh I can use it as is except for the line WERROR=1 python setup.py install . I am not able to use the line with WERROR=1 . For test.sh , how are you able to run the test while in the pytorch directory? I have to do a "cd .." out of the directory. Otherwise I get an error like below.

test_python_nn
python test/run_test.py --include nn --verbose
Traceback (most recent call last):
File "test/run_test.py", line 14, in
import torch
File "/home/freddie/builder/jenkins/pytorch/pytorch/torch/init.py", line 84, in
from torch._C import *
ModuleNotFoundError: No module named 'torch._C'

this is because there is a torch directory in the pytorch repo. Does this not happen on x86 ?

@ezyang, when you get the chance please review the latest revision.

@pytorchbot retest this please

@ezyang when you get the chance , please review. I tried to do the export WERROR=1 in a separate conditional , but it created a problem when building the ATen install. So it looks like it needs to be in line with the python setup.py install.

@ezyang please review when you get the chance, hope the changes are OK