Hacktoberfest 2026：维护者为十月标记出来的 issue，仍然开放、适合新手。 浏览 Hacktoberfest issue

Expert Parallelism for MoEs

未关闭
  • 哪里不对？
#165 1 条评论 0 个 reaction 已指派 0 人 在 GitHub 查看

还没有人认领这个 Issue。

评估

难度
5/5
预计耗时
一周以上
新手友好度
8/100
Issue 类型
功能
描述清晰度
基本清楚
活跃度
停滞
技术栈
python
领域
distributed-systems, machine-learning

调研方向

Start with the existing MoE layer and router code in Fast-LLM, then read the EPLB repo to see what eplb.rebalance_experts expects as input. The issue calls for a 2-layer, 12-expert test with 4 redundant experts across 8 GPUs, with expert load checked for balance. The design is open and the issue is nearly two years old with no assignee, so confirm scope with maintainers first.

由索引模型根据 Issue 内容生成。

描述

enhancement need update

🎯 Goal (What & Why)

Integrate Expert Parallelism (EP) into Fast-LLM's MoE architecture.

Motivations:

  • Support Larger MoE Models: Enable training and inference for bigger models than Mixtral-8x7B, exceeding single-GPU memory constraints.
  • Improve GPU Load Balancing: Use DeepSeek's EPLB (eplb.rebalance_experts) to dynamically balance GPU loads via redundant expert replication and heuristic packing, see https://github.com/deepseek-ai/EPLB/tree/main.
  • Reduce Communication Overhead: Initially apply global load balancing, and eventually implement hierarchical load balancing to colocate expert groups.

🚀 Execution Plan

Step 1: Minimal Viable Product (MVP)

  1. Integrate eplb.rebalance_experts with Fast-LLM:

    • Add support for expert parallelism only in MoE layers.
    • Implement global load balancing for distributing expert replicas across GPUs.
    • Provide a minimal configuration interface, reusing existing MoE router logic in Fast-LLM.

  2. Validation & Testing:

    • Test with a small MoE, say 2-layers containing 12 experts per layer, using 4 redundant experts and distributing them across 8 GPUs.
    • Verify that the expert load is balanced and that performance is maintained during inference.

  3. Documentation:

    • Update Fast-LLM's documentation to include usage examples and configuration guidelines for enabling EP.
Step 2: Additional Optimizations (Post-MVP)
  • Support for Hierarchical Load Balancing: When the number of server nodes divides the number of expert groups.
  • Advanced Routing Strategies: Introduce DeepSeek's group-limited expert routing.
  • Performance Enhancements: Explore fused operations and other optimizations for speed.

📌 Acceptance Criteria (Must-Haves for Completion)

  • EP must be functional and tested in Fast-LLM's MoE layers.
  • The implementation must be documented clearly, with examples and edge cases covered.
  • The PR must include a performance/impact summary, demonstrating memory and compute efficiency gains.
  • No refactors unless directly necessary for feature completion.

🛠️ Project Management

  • Assign the project to the Fast-LLM project.
  • Set the Estimate field (in days) in the GitHub project.
  • Use the Size field to categorize the PR size (Small/Medium/Large).
  • Assign an owner when opening the issue.
主要语言
Python
星标
337
派生
52
30 天内没有已合并 PR

环境准备

从这里开始

  1. 先读完整个 Issue，再读项目的贡献指南。
  2. 在 Issue 下留言说明你要接手 —— 这能避免两个人做同样的事。
  3. Fork 仓库，在一个分支上完成修改。
  4. 提交 Pull Request，并在描述里引用这个 Issue 编号。

ServiceNow/Fast-LLM 的其他 Issue

查看 ServiceNow/Fast-LLM 的全部 Issue

相似的 Issue

更多 Python Issue

把新 issue 发到你的邮箱

精选适合新手参与的 GitHub issue 摘要。