v0.7.5

Built with PyTorch 1.11.0

v0.7.4

• Added experimental functions for tensor partitioning
• Improved efficiency of allreduce
• Fixed for PyTorch 1.11

v0.7.3

• Added function to switch dropout mode
• Use NCCL to allreduce batch sizes and bcast tensors

v0.7.2

• Updated for PyTorch 1.10.1
• Added watchdog to monitor comm errors
• Removed dependency on mpicxx

v0.7.1

Fixed a bug of parameter synchronization in state_dict()

v0.7.0

This release contains the following new features:

• Integration with zero redundancy optimizer proposed by DeepZero
• Data offloading to host memory