Skip to content

[ENHANCEMENT] MegaDPP:动态流水线并行优化 #27

Description

@jennifer88huang
  1. 高效 NCCL P2P 通信替代 Shared Memory

    • 通过 C++ 封装的 NCCL 点对点通信,显著提升跨 GPU 带宽、降低延迟,消除 CPU 拷贝与共享内存限制。(预计需要1-2month)
  2. 引入多线程通信–计算重叠

    • 利用 C++ 多线程与条件变量,将前向/反向通信与计算解耦并流水线化,提高整体吞吐。(预计需要1-2month)

Activity

Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Metadata

Metadata

Assignees

No one assigned

    Labels

    Type

    No type

    Projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions