jenkins build for cuda 10.0

[rongou]

Also removed the hack to work around the cuda 10.1 compiler bug. We can try 10.1 again once update 1 is out and the bug is fixed.

@RAMitchell @trivialfis

[RAMitchell]

The dockerfile searches for nccl binaries here (

xgboost/tests/ci_build/Dockerfile.gpu

Line 24 in 09bd9e6

wget https://developer.download.nvidia.com/compute/redist/nccl/v2.2/nccl_2.2.13-1%2Bcuda${CUDA_SHORT}_x86_64.txz && \

).

I'm not sure if 10.0 binaries are available at this link. We may have to move back to a nccl submodule to have a more sustainable way of building multi-GPU xgboost.

[rongou]

OK, I guess Jenkins doesn't actually use the Jenkinsfile in the PR?

Is there a reason to stick with CUDA 8 and Centos 6? If we drop CUDA 8 and move to Centos 7 there would be more options.

[RAMitchell]

I think updates to the Jenkinsfile do not immediately take effect due to security reasons. Possibly it works for members, I'm not sure.

I set up the builds on Centos 6 and cuda 8.0 for maximum compatibility binary wheels (https://opensource.com/article/19/2/manylinux-python-wheels). Potentially we could upgrade cuda soon. What particular features do you need?

[hcho3]

@RAMitchell I changed the security configuration for Jenkinsfile, so that changes are visible. Let me try again.

[rongou]

Looks like Jenkins has passed with CUDA 10.0? Please take another look. @RAMitchell @hcho3

[hcho3]

LGTM