DeepSpeed

[CPU] Support SHM based inference_all_reduce in TorchBackend

#5391

Merged