Xin Yao
Xin Yao
> @yaox12 Can you add cuda versions? Added.
I don't think this is an issue related to data alignment. Because in this GAT model, the node features are projected from 602 to 256-dimension tensors before invoking GSpMM. We...
> Thanks for the suggestion. Just be curious. Does PyTorch support bfloat16 natively? Yes. PyTorch supports bfloat16 for both CPU and GPU.
bfloat16 requires compute capability >= 8.0 and CUDA >= 11.
Do we need fallbacks for `__CUDA_ARCH__ < 800`? cc @nv-dlasalle For PyTorch, 1. bf16 arithmetic functions are supported on all CUDA architectures. For example, the following code is valid. ```python...
@BarclayII Cannot repro with the GraphSAGE example and dgl 0.9.0. Multi-worker CPU sampling and CUDA dataloader device should have been covered in the unit test now. https://github.com/dmlc/dgl/blob/5ba5106acab6a642e9b790e5331ee519112a5623/tests/pytorch/test_dataloader.py#L185-L187 @samvanstroud Are you...
> @yaox12 Do you know why pytorch 1.12.1 would cause this? This looks like an issue of a forked cuda context, not related to the TensorAdapter (which has issues when...
@mufeili I can reproduce this issue with PyTorch 1.12.1, but haven't found the root cause. Regarding the error message, it seems not related to the tensoradaptor so I'm not sure...
@nv-dlasalle Good catch! This is my fault. Should be fixed in #4450.
Can you provide env information such as: - DGL Version (e.g., 1.0): - Backend Library & Version (e.g., PyTorch 0.4.1, MXNet/Gluon 1.3): - OS (e.g., Linux): - How you installed...