djl
djl copied to clipboard
deepjavalibrary
Adding ppc64le support for PyTorch engine
Description
Adding support to ppc64le architecture to pytorch-engine. By referring these instructions have built and test jni locally on my ppc64le VM with opence conda pytorch. Build and tests completed successfully.
We don't have plan to ship ppc64le native libraries for PyTorch engine.
Since you are able to build JNI, you can package your JNI and the .so files in a jar or save them in DJL cache folder. DJL should be able to pick it up.
@frankfliu DJL is dependency for https://github.com/opensearch-project/ml-commons We are working on enabling ppc64le support for opensearch. Here is issue https://github.com/opensearch-project/opensearch-build/issues/3122. If you can consider this request, we are happy to work with you.
The challenge part is how to create a release pipeline for ppc. And how to maintain this pipeline for new version of PyTorch
The challenge part is how to create a release pipeline for ppc. And how to maintain this pipeline for new version of PyTorch
@frankfliu you can use multiarch/qemu-user-static container image for implementing workflow for ppc64le. We can support you to implement the same.