ppc64le/build-scripts

Hope there will be mxnet build scripts for ppc64le + redhat 7.5

Closed this issue · 7 comments

Hope there will be mxnet build scripts for ppc64le + redhat 7.5

@sphinx-jiang - are you building mxnet on CPU or GPU based setup? Have you tried using the RHEL 7.4 script .. in general I would expect it to run without issues on RHEL 7.5 as well ... could you please try?

I need GPU based setup on RHEL, so I tried CentOS gpu shell.

It failed here:
If needed I could paste whole log later.
a - build/src/kvstore/gradient_compression_gpu.o
a - build/src/kvstore/kvstore_utils_gpu.o
/bin/ld: warning: libfribidi.so.0, needed by //usr/lib64/libpango-1.0.so.0, not found (try using -rpath or -rpath-link)
//usr/lib64/libpango-1.0.so.0: undefined reference to fribidi_get_bidi_type' //usr/lib64/libpango-1.0.so.0: undefined reference to fribidi_get_bracket'
//usr/lib64/libpangoft2-1.0.so.0: undefined reference to g_log_structured_standard' //usr/lib64/libpango-1.0.so.0: undefined reference to [fribidi_get_par_embedding_levels_ex'](url)
collect2: error: ld returned 1 exit status
make: *** [bin/im2rec] Error 1

  • rm -rf build

Another question is the install shell just use system python env. However we need work with different python version in anaconda.

@seth-priya can you look at the errors when you get a chance? Looks like some dependencies are missing. If you can identify, maybe @sphinx-jiang can update and re-test?

@sphinx-jiang
https://github.com/ppc64le/build-scripts/blob/master/mxnet/Dockerfiles/latest_centos_7_gpu/Dockerfile These steps are working fine for me on centos+GPU with latest mxnet source code. Can you please try ?

FYI - I have used the nvidia/cuda-ppc64le:9.0-cudnn7-devel-centos7 as a base image.

Centos version used in nvidia/cuda-ppc64le:9.0-cudnn7-devel-centos7 image is :

$ cat /etc/redhat-release
CentOS Linux release 7.5.1804 (AltArch)

It has been solved already, user need select commands they need from the script, not run the script directly. Thanks all of you.

Hope all packages of ppc64le supported in anaconda officially. Currently for data scientists, using ppc64le on redhat, is totally nightmare:)