Functionalities that require pyg_lib/torch_sparse/torch_cluster

General Questions
1

Since version 2.3, PyG can be installed and used with only raw PyTorch.
Gaudi HPUs support PyTorch, and thus, PyG’s basic functionalities can be used.

However, many other functionalities require pyg_lib/torch_sparse/torch_cluster, while those libraries only support CPU and CUDA GPUs.

I am wondering whether there are general workarounds for using such functionalities on Gaudi HPUs.
If not, what should I do to make it possible?

Thank you for your help in advance.

2

@vezenbu I looked at these libraries.
pyg_lib is provider of custom ops. They have special implementations for CPU and CUDA. To enable it on Gaudi, you will need to provide to either: implement custom HPU ops or (better) implement these ops using standard PyTorch ops.

The same comment applies to torch_sparse and torch_cluster.