* Use shared memory for grad sync when NCCL is not avaliable as PyTorch distributed backend.
Fix small bugs and update unitests
* Fix bug
* update test
* update test
* Fix unitest
* Fix unitest
* Fix test
* Fix
* simple update
Co-authored-by: Ubuntu <ubuntu@ip-172-31-24-212.ec2.internal>
* Split from NCCL PR
* Fix type in comment
* Expand documentation for sparse_all_to_all_push
* Restore previous behavior in example
* Re-work optimizer to use NCCL based on gradient location
* Allow for running with embedding on CPU but using NCCL for gradient exchange
* Optimize single partition case
* Fix pylint errors
* Add missing include
* fix gradient indexing
* Fix line continuation
* Migrate 'first_step'
* Skip tests without enough GPUs to run NCCL
* Improve empty tensor handling for pytorch 1.5
* Fix indentation
* Allow multiple NCCL communicator to coexist
* Improve handling of empty message
* Update python/dgl/nn/pytorch/sparse_emb.py
Co-authored-by: xiang song(charlie.song) <classicxsong@gmail.com>
* Update python/dgl/nn/pytorch/sparse_emb.py
Co-authored-by: xiang song(charlie.song) <classicxsong@gmail.com>
* Keepy empty tensor dimensionaless
* th.empty -> th.tensor
* Preserve shape for empty non-zero dimension tensors
* Use shared state, when embedding is shared
* Add support for gathering an embedding
* Fix typo
* Fix more typos
* Fix backend call
* Use NodeDataLoader to take advantage of ddp
* Update training script to share memory
* Only squeeze last dimension
* Better handle empty message
* Keep embedding on the target device GPU if dgl_sparse if false in RGCN example
* Fix typo in comment
* Add asserts
* Improve documentation in example
Co-authored-by: xiang song(charlie.song) <classicxsong@gmail.com>
xiang song(charlie.song)