Skip to content

feat: issue1 SM/QP预算模型 + issue3 AlltoAllv分阶段overlap优化 - #23

Open
Mezca1 wants to merge 1 commit into
Tencent:mainfrom
Mezca1:main
Open

feat: issue1 SM/QP预算模型 + issue3 AlltoAllv分阶段overlap优化#23
Mezca1 wants to merge 1 commit into
Tencent:mainfrom
Mezca1:main

Conversation

@Mezca1

@Mezca1 Mezca1 commented Jul 23, 2026

Copy link
Copy Markdown

Issue 1: 分析式SM/QP预算模型

  • C++独立复现DeepEP get_theoretical_num_sms()算法
  • 组合数学计算期望topk,按阶段累加HBM/RDMA/NVLink流量
  • 瓶颈链路识别 + SM反推 + QP预算
  • 三种模式CLI: single/sweep/compare
  • 12个单元测试全部通过

Issue 3: AlltoAllv send/recv分阶段overlap

  • 对齐DeepEP return_recv_hook机制
  • CUDA kernel phases位掩码 (SEND/RECV/FULL)
  • 快慢卡场景模拟 + 人为延迟注入
  • 分阶段 vs 基线端到端对比 + recv SM扫描
  • 6个单元测试全部通过

Issue 1: 分析式SM/QP预算模型
- C++独立复现DeepEP get_theoretical_num_sms()算法
- 组合数学计算期望topk,按阶段累加HBM/RDMA/NVLink流量
- 瓶颈链路识别 + SM反推 + QP预算
- 三种模式CLI: single/sweep/compare
- 12个单元测试全部通过

Issue 3: AlltoAllv send/recv分阶段overlap
- 对齐DeepEP return_recv_hook机制
- CUDA kernel phases位掩码 (SEND/RECV/FULL)
- 快慢卡场景模拟 + 人为延迟注入
- 分阶段 vs 基线端到端对比 + recv SM扫描
- 6个单元测试全部通过
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant