DeepSpeed

[CPU] add fp16 support to shm inference_all_reduce

#5669

Merged