DeepSpeed

9a262aee - feat: abstract weight transfer backend (gdr/http) for VLLMRollout

feat: abstract weight transfer backend (gdr/http) for VLLMRollout - Rename nccl -> gdr (GPU-direct) to decouple from NCCL specifically. - Add weight_transfer_backend to RolloutConfig: auto | gdr | http. - auto mode detects NCCL availability and falls back to HTTP. - HTTP path sends base64-encoded tensor data per parameter. - GDR path uses StatelessProcessGroup + NCCL broadcast (unchanged). Signed-off-by: Guokai Ma <guokai.ma@intel.com>