DeepSpeed

[CCLBackend] Using parallel memcpy for inference_all_reduce

#4404

Merged