optimize muon 2D sharding - #79763
Conversation
There was a problem hiding this comment.
Paddle-Bot Review Board (review完成)
| 序号 | 位置 | 优先级 | 规则来源 | 状态 |
|---|---|---|---|---|
| 1 | PR 验证信息缺失 | 仓库规则:PR 信息与验证 | 🟡 | |
| 2 | 多进程通信路径缺少验证 | 仓库规则:分布式测试与验证 | 🚧 |
Codecov Report❌ Patch coverage is
❌ Your patch status has failed because the patch coverage (72.89%) is below the target coverage (90.00%). You can increase the patch coverage or adjust the target coverage. Additional details and impacted files@@ Coverage Diff @@
## develop #79763 +/- ##
==========================================
Coverage ? 72.89%
==========================================
Files ? 2
Lines ? 107
Branches ? 0
==========================================
Hits ? 78
Misses ? 29
Partials ? 0 ☔ View full report in Codecov by Harness. 🚀 New features to boost your workflow:
|
Failed CI看板
日志分析报告失败的测试 case: 根本原因分析:
修复建议:
Powered by Nyanpasu with gpt-5.6-luna 默认推理级别, please check the suggestions carefully. |
|
/re-run all-failed |
| if world_size <= 1: | ||
| # No partition needed, all params stay on rank 0 | ||
| self._rank2params_2d_by_color[color_key] = {0: list(params_2d)} | ||
| if self.machine_balanced_2d_partition: |
Paddle-Bot
left a comment
There was a problem hiding this comment.
此前指出的多进程通信测试缺口在当前提交中仍未解决。
PR Category
Performance Optimization
PR Types
Improvements
Description
优化muon 2d参数的切分
步骤如下:
a. 找到压力最小的机器
b. 找到拥挤程度最低的机器
c. 找到该机器上压力最小的rank
【测试结果】:在8机情况下,开启ep=4,pp=0的测试条件下,生成checkpoint
a. 其中checkpoint-10、checkpoint-40为老版本代码生成;checkpoint-20、checkpoint-30为新版本代码生成。
b. 四次训练的loss值如下,正常接续
文件大小:
老版本的所有权重均聚集于排序靠前的rank
新版本权重按照机器均分,高负载机器尽可能远离(1号机器和8号机器)

是否引起精度变化
否