DeepSpeed

skip torch.zeros and tensor.copy_ when model parallel is not used

#2479

Merged