This is a binary rebuild of 0.6.0 release that adds support on PyTorch 1.8 + CUDA 11.1. Please install with either of the following:
conda install dgl-cuda11.1 -c dglteam
pip install dgl-cu111
No feature changes are incorporated.
Currently there is an issue in CUB when building with CUDA 11.1 from source where DGL will crash with various CUDA errors or freeze when using with PyTorch 1.8. You will need to define CUB_CPP_DIALECT=2003 in the C++ and NVCC flags as a work around. Consequently, CUDA 11.1 binaries are built with the macro CUB_CPP_DIALECT=2003 while CUDA 11.0- are built without the macro.