This is a bugfix release including the following bugfixes:
Quality-of-life updates
- Python 3.10 support.
- PyTorch 1.11 support.
- CUDA 11.5 support on Linux. Please install with
pip install dgl-cu115 -f https://data.dgl.ai/wheels/repo.html # if using pip conda install dgl-cuda11.5 -c dglteam # if using conda - Compatibility to DLPack 0.6 in tensoradapter (#3803) for PyTorch 1.11
- Set stacklevel=2 for dgl_warning (#3816)
- Support custom datasets in DataLoader that are not necessarily tensors (#3810 @yinpeiqi )
Bug fixes
- Pass ntype/etype into partition book when node/edge_split (#3828)
- Fix multi-GPU RGCN example (#3871 @yaox12)
- Send rpc messages blockingly in case of congestion (#3867). Note that this fix would probably cause speed regression in distributed DGL training. We were still finding the root cause of the underlying issue in #3881.
- Fix CopyToSharedMem assuming that all relation graphs are homogeneous (#3841)
- Fix HAN example crashing with CUDA (#3841)
- Fix UVA sampling crash without specifying prefetching features (#3862)
- Fix documentation display issue of node/edge_split (#3858)
- Fix device mismatch error in GraphSAGE distributed training example under multi-node multi-GPU (#3870)
- Use
torch.distributed.algorithms.join.Jointo deal with uneven training sets in distributed training (#3870) - Dataloader documentation fixes (#3886)
- Remove redundant reference of networkx package in pagerank.py (#3888 @AzureLeon1 )
- Make source build work for systems where the default is Python 2 (#3718)
- Fix UVA sampling with partially specified node types (#3897)