How to get CUDA to work?

[thhart]

Hi, I have problems to get cuda support running, CUDA libraries are available as 11.7 and 12.0, however neither mxnet nor pytorch can find the correct cuda version. Always I get a error message like this:

PyTorch:
"cuda flavor for linux-x86_64 found: cu120."
MXnet:
"No matching cuda flavor for linux found: cu120mkl/sm_75."

Other libraries like onnx native (outside ai.djl) have no problems finding and using CUDA.

pom.xml contains this:

    <dependency>
        <groupId>ai.djl</groupId>
        <artifactId>api</artifactId>
        <version>0.21.0</version>
    </dependency>
    <!-- https://mvnrepository.com/artifact/ai.djl.mxnet/mxnet-engine -->
    <dependency>
        <groupId>ai.djl.mxnet</groupId>
        <artifactId>mxnet-engine</artifactId>
        <version>0.21.0</version>
        <scope>runtime</scope>
    </dependency>
    <!-- https://mvnrepository.com/artifact/ai.djl.pytorch/pytorch-engine -->
    <dependency>
        <groupId>ai.djl.pytorch</groupId>
        <artifactId>pytorch-engine</artifactId>
        <version>0.21.0</version>
    </dependency>
    <!-- https://mvnrepository.com/artifact/ai.djl.mxnet/mxnet-native-cu110mkl -->
    <dependency>
        <groupId>ai.djl.mxnet</groupId>
        <artifactId>mxnet-native-cu110mkl</artifactId>
        <version>1.9.0</version>
    </dependency>
    <dependency>
        <groupId>ai.djl.pytorch</groupId>
        <artifactId>pytorch-jni</artifactId>
        <version>1.13.1-0.21.0</version>
        <scope>runtime</scope>
    </dependency>
    <dependency>
        <groupId>ai.djl.pytorch</groupId>
        <classifier>linux-x86_64</classifier>
        <artifactId>pytorch-native-cu117</artifactId>
        <version>1.13.1</version>
        <scope>runtime</scope>
    </dependency>

[frankfliu]

MXNext only works for cu102 and cu112, see: https://docs.djl.ai/master/engines/mxnet/mxnet-engine/index.html#linux-gpu

PyTorch should work on cu117, what you need to do is set current cuda version to 11.7. DJL try to load default libcudart.so file and detect the current cuda version. Make sure cu117 version of libcudart.so is before cu120.

If you use DJL 0.22.0-SNAPSHOT version, you should be able to use cuda 12.0. See: #2485

[thhart]

Thanks for the pointer, 11.2 is working and finding the GPU with mxnet.