trunk/10a2f3aa6abef3a84ad0ba37007fbce4d4681a6b: [TokenSwitch] Add symmetric-memory zero-copy dispatch/combine (#187785)

发生了什么

When a dispatch/combine tensor is backed by NCCLSymmetricMemory registered under the process group's name, address it via the registered ncclWindow + offset (RDMA zero-copy path) instead of a plain device pointer, avoiding a staging copy into a regular device buffer. Applies to dispatch inputs/outputs and combine inputs. EpTensor gains a window-backed constructor ( win_hdl + win_offset ). make_ep_tensor() looks up t…

摘要按规则整理自下方来源原文

来源