Does the input tensor have to be from CPU?

K-Wu / pytorch-direct_dgl

PyTorch-Direct code on top of PyTorch-1.8.0nightly (e152ca5) for Large Graph Convolutional Network Training with GPU-Oriented Data Communication Architecture (accepted by PVLDB)

https://arxiv.org/abs/2103.03330

45 stars 5 forks source link

Does the input tensor have to be from CPU? #3

Open srikanthmalla opened 1 year ago

srikanthmalla commented 1 year ago

Hi, Thank you for the interesting work.

In your readme examples you mentioned this: input_tensor: A given input 2D tensor in CPU

But in the paper, you have also mentioned very large GNN models can give OOM. In this case what if we store the node features on NVMe? Do you have a simple example then? Does the library avoid CPU and go to GPU in that case, like the GPUDirect or DALI? Please let me know.

K-Wu commented 1 year ago

Hi there, thank you for your interest. Sorry that we didn't look into and evaluate the direct access to NVMe in this project.

LukeLIN-web commented 1 year ago

Usually, we assume CPU memory is big enough. You can refer to https://marius-project.org/publications/ to find some papers which store the node in the disk.