DeepSpeed

[Zero2] Reduce the unnecessary all-reduce when tensor size is 0.

#5868

Merged