This is a patch release including the following bug fixes and minor features.
Documentation fixes
- #2081 Reorganize user guide and split chapters into multiple pages
- #2085 #2090 Fix links
- #2091 #2114 User guide on distributed training
- #2096 #2097 Other documentation fixes
- #2086
- #2123 Temporarily remove SSE MXNet tutorial
Bug fixes
- #2100 add_edges() crashes if the input tensor is empty
- #2084 Fix distributed GraphSage running with GPU
- #2107 Building with HDFS previously fails
- #2087 Cannot load the PTC dataset via
dgl.data.GINDataset
- #2076 Empty cuda graph raise error in create_formats_
- #2121 Disable hypersparse memory optimization due to incomplete COO graph support on GPU
- #2115 Fallback to CPU for graph traversal functions on GPU graphs
- #2108 Force
num_workers
andnum_samplers
to be the same for distributed training - #2118 (#2127 )
Bug fixes in examples
- #2119 Bug in using Layer Normalization in RGCN
New features
- #2102 5 utility functions that handle raw data features
- #1979 #2117 CUDA 11 support - We will not release binary builds of CUDA 11 for now
Release changes
- Linux now requires GCC 5 to build DGL from source.
- DGL now supports Mac 10.9+.