Is the RTX A4000 supported?

[mmz520]

when I use mxnet engine,it is loaded by class which name is LibUtils.But the fuction which name is "supported" prevented me from using the GPU and show error: "no matching cuda flavor for linux found:cu112mkl/sm_86".source code like this:

private static boolean supported(Platform platform) {  
        // mxnet-native-cu102mkl:1.8.0: 3.0, 5.0, 6.0, 7.0, 7.5  
        // mxnet-native-cu110mkl:1.8.0: 5.0, 6.0, 7.0, 8.0  
        String version = platform.getVersion();  
        if (version.startsWith("1.8.") || version.startsWith("1.9.")) {  
            String flavor = platform.getFlavor();  
            String cudaArch = platform.getCudaArch();  
            if (flavor.startsWith("cu11")) {  
                if (version.startsWith("1.8.")) {  
                    return Arrays.asList("50", "60", "70", "80").contains(cudaArch);  
                }  
                return Arrays.asList("50", "60", "70", "75", "80").contains(cudaArch);  
            } else if (flavor.startsWith("cu10")) {  
                return Arrays.asList("30", "50", "60", "70", "75").contains(cudaArch);  
            }  
        }  
        return true;  
}  

Why does this function tell me that djl does not support GPUs with 86 hashrate？What should I do?
env :

Ubuntu20.4
cuda11.2
cudnn8.6.0
RTX A4000

pom:

<dependency>
    <groupId>ai.djl</groupId>
    <artifactId>bom</artifactId>
    <version>0.20.0</version>
    <type>pom</type>
    <scope>import</scope>
</dependency>
<dependency>
    <groupId>ai.djl.mxnet</groupId>
    <artifactId>mxnet-engine</artifactId>
    <scope>runtime</scope>
</dependency>
<dependency>
    <groupId>ai.djl.mxnet</groupId>
    <artifactId>mxnet-native-cu112mkl</artifactId>
    <scope>runtime</scope>
    <version>1.9.1</version>
</dependency>
<dependency>
    <groupId>ai.djl.mxnet</groupId>
    <artifactId>mxnet-model-zoo</artifactId>
</dependency>
<dependency>
    <groupId>ai.djl</groupId>
    <artifactId>basicdataset</artifactId>
</dependency>
<dependency>
    <groupId>ai.djl</groupId>
    <artifactId>model-zoo</artifactId>
</dependency>

[frankfliu]

You can try to use PyTorch, it should work on Ampere GPU.

I just created a PR to allow MXNet run on sm_86 GPU: #2313.

To reduce the MXNet binary files size, it only compiles with selected CUDA architecture. If you load the MXNet with unsupported sm arch (eg. sm_52), the performance is really bad. We added that check to prevent this happen.

[mmz520]

Thanks!Now I know that my GPU is really not supported.

[frankfliu]

Thanks!Now I know that my GPU is really not supported.

You can try it again once my PR is merged, or use PyTorch instead of MXNet