引导式测验 · 硬 等级
MoE 服务测验的专家并行性
测试 MoE 专家放置、代币分配和负载不平衡,包括所选专家和通信如何影响分布式推理。
问题 1 的 8
专家并行在 GPU 上的分布是什么?
预览每个问题8 问题
当您参加测验时,就会出现答案和解释。
- 专家并行在 GPU 上的分布是什么?
- 完整模型的副本以实现冗余
- 仅限个人关注头
- MoE层的不同前馈专家
- 嵌入表的层数
- 哪种通信模式对于 MoE 专家并行性至关重要?
- 两次全对全交换(调度和合并)
- 单次广播
- 点对点梯度同步
- 无需沟通
- 尽管总参数巨大,为什么 MoE 仍保持每个代币的计算量较低?
- 它将权重压缩到 1 位
- 它跳过了注意力层
- 专家分享他们所有的权重
- 每个代币仅运行前 k 个选定的专家
- 由于路由依赖于数据会产生什么问题?
- 梯度爆炸
- 负载不平衡,一些专家获得更多代币
- 路由器无法训练
- 代币失去了位置嵌入
- Why are capacity limits and padding used in some MoE implementations?
- To control expert buffer sizes and make computation shapes more predictable
- To require every expert to process every token
- To eliminate all token routing
- To guarantee that every MoE system drops tokens
- 系统如何隐藏所有步骤的延迟?
- 通过在 CPU 上运行
- 通过删除组合步骤
- 通过提高学习率
- 通过与专家计算重叠通信
- A routed token chooses two MoE experts on different devices. What must expert-parallel execution do?
- Run every expert in the model regardless of routing
- Dispatch its representation to the selected experts and combine their outputs
- Replace both expert networks with an attention head
- Drop the token solely because it selected two devices
- One MoE expert receives far more tokens than others. Why can the layer’s latency increase?
- The total expert parameter count automatically doubles
- The router’s outputs become attention positions
- The overloaded expert becomes a straggler while other devices wait
- All other experts are removed from memory