I am looking for extending PyTorch with the feature of using persistent memory for allocating tensors. As of now, we can use CPU-Memory or Cuda-Memory for the allocations of tensors. What should be the starting path for using persistent memory? When some new allocation of tensor will be made that allocation will be in Persistent memory, not in CPU-DRAM. What are files need to be modified? I went through the tutorials of extending PyTorch. However, they talked about things like custom c++ extensions, custom c++ classes, new backend dispatcher, etc. I am not sure how should I proceed with it.
Could you explain what exactly you mean by “persistent memory”?