DeepSpeed

db8289ad - rollout: fix hybrid engine device index bug, vllm import path, and smoke configs

rollout: fix hybrid engine device index bug, vllm import path, and smoke configs - hybrid_engine_rollout: fix get_accelerator().device() receiving torch.device instead of int in graph capture path; accept cfg= parameter; default continuous_batching_size to 0 (no limit) - vllm_rollout: fix StatelessProcessGroup import for vllm 0.22.0 (moved to vllm.distributed.utils); use 127.0.0.1 for GDR channel rendezvous - smoke configs: use ZeRO-0 + torch_adam for portability, add vllm_python and weight_transfer_backend options - test: update default continuous_batching_size assertion (0 instead of 1) Signed-off-by: Guokai Ma <guokai.ma@intel.com>