Deeplearning4j with GPU/CUDA on SLURM

job_with_gpu.slurm

#!/bin/bash
#SBATCH -J bom_with_gpu
#SBATCH --mail-user=<??>
#SBATCH --mail-type=ALL
# Bitte achten Sie auf vollständige Pfad-Angaben:
#SBATCH -e /work/scratch/?user name?/job_with_gpu.err.%j
#SBATCH -o /work/scratch/?user name?/job_with_gpu.out.%j
#
#SBATCH -t 00:30:00     # in Stunden, Minuten und Sekunden, oder '#SBATCH -t 10' - nur Minuten
#SBATCH -C nvd
#SBATCH --ntasks=16
#SBATCH --mem-per-cpu=1000
#
# -------------------------------
# Anschließend schreiben Sie Ihre eigenen Befehle, wie z.B.
module load cuda/8.0
cd /home/?user name?/Project
java -cp target/Project-0.5-nd4j-cuda-8.0-platform.jar project.TestClass

pom.xml

...
  <!-- <nd4j.backend>nd4j-native-platform</nd4j.backend> -->
  <nd4j.backend>nd4j-cuda-8.0-platform</nd4j.backend> <!-- activate cuda platform, deactivate native platform -->
...
    <dependencyManagement>
        <dependencies>
            <dependency>
                <groupId>org.nd4j</groupId>
                <artifactId>nd4j-native-platform</artifactId>
                <version>0.6.0</version>
            </dependency>
          <!-- add following: -->
            <dependency>
                <groupId>org.nd4j</groupId>
                <artifactId>nd4j-cuda-8.0-platform</artifactId>
                <version>0.6.0</version>
            </dependency>
        </dependencies>
    </dependencyManagement>
...

TestClass.java

import org.nd4j.jita.conf.CudaEnvironment;
import org.nd4j.jita.conf.Configuration;

public class TestClass {
    public static void main(String[] args) throws Exception {

        CudaEnvironment.getInstance().getConfiguration().allowMultiGPU(true); //Use all available GPUs
        DataTypeUtil.setDTypeForContext(DataBuffer.Type.HALF); //Use half precision, boost learning up to 200%
        CudaEnvironment.getInstance().getConfiguration()
                .setMaximumDeviceCacheableLength(1024 * 1024 * 1024L)
                .setMaximumDeviceCache(6L * 1024 * 1024 * 1024L)
                .setMaximumHostCacheableLength(1024 * 1024 * 1024L)
                .setMaximumHostCache(6L * 1024 * 1024 * 1024L); //Allow to cache up to 6GB of GPU RAM
      
      
      ....
        
    }
}