Compiling deeplearning4j with -Pcuda does not produce cuda artifacts

1

Hello,

I am working on compiling deeplearning4j for my specific architecture. In the top level directory I am running the command:

mvn clean install -Pcuda -DskipTests -PskipTestResourceEnforcement=true -Djavacpp.platform=linux-x86_64

But the following dependencies that I thought should exist are not produced:

  • org.nd4j:nd4j-cuda-11.2-platform:jar:chriscustom1.0
  • org.deeplearning4j:deeplearning4j-cuda-11.2:jar:chriscustom1.0

Other artifacts are produced, but not this one. I am branched off the 1.0.0-M1.1 tag.

What am I doing wrong?

I can successfully produce a -Pcpu build, FYI, and run on the CPU with it.

2

@chris2 you can just download cuda 11.2 support from maven central now. This was fixed last week. Note that going forward we’ll mainly support cuda 11.6 and 11.7 though.

3

Amazing! Can I get cuda 11.6 support from maven central or your baseline or something? I don’t see it yet: Maven Central Repository Search I’m actually only compiling because I must use 11.6. I was trying wit h 11.2 just to learn how to build it from a simpler test case. @agibsonccc

4

Also, part of why I have had to recompile is that the linux_x86 *.so files in the nd4j native packages didn’t work on my older operating system/architecture. They worked after I recompiled them. I have run into the same problem with your new cuda artifacts. Do you have some instructions floating around for compiling the -Pcuda project?

5

Hello! I have discovered a couple of critical facts that I think have taken me 90% of the way to a solution here. For the record:

  1. I discovered the shell script change-cuda-versions.sh and have updated it for the proper versions of cuda/cudann/etc
  2. You gave me a clue when you said you were all changing the build to build the nd4j backends from the root. When I moved to the nd4j-backends directory and started building there, it started creating the artifacts.

It will be some time before the build is complete, but I think this is going to work.

6

@chris2 yes that’s all you need to do is change the base versions. Basically cuda version support just consists of changing the versions in the pom and ensuring that libnd4j c++ compilation matches that. Cuda 11.6 is on the way with the next release.