Skip to content

Conversation

@Chamberlain0w0
Copy link
Contributor

@Chamberlain0w0 Chamberlain0w0 commented Nov 5, 2025

  1. 按照 feat: support TP/SP #83 的 review 意见做了部分代码修改
  2. 加了多维并行度的 Layout 表示,在 nn::parallel::global 的 namespace 里面
  3. 修改了测试脚本,把测试项和路径配置单拎出来作为了一个 config.json 文件,这样以后修改路径或者加测试项会更方便。然后加了 TP 相关的两个测例(分别是 TP = 4 以及 TP = SP = 4),使用方法与之前相同


nn::parallel::global::InitAllEnv(FLAGS_nthread_per_process, FLAGS_tensor_parallel, FLAGS_sequence_parallel);

LOG(INFO) << nn::parallel::global::ProcessGroupOverview();
Copy link
Contributor Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

留了一个接口输出不同并行通信组下 rank 划分的情况,效果类似于:

Image

@Chamberlain0w0
Copy link
Contributor Author

先单开了两个新的表格(为了不影响原有的)填入了改后的测试脚本在本分支下的运行结果:https://gxtctab8no8.feishu.cn/drive/folder/C3KIfnRSGlz6HsdFW8Sc9EQ9nng

@kilinchange kilinchange self-requested a review November 6, 2025 02:14
@Chamberlain0w0
Copy link
Contributor Author

gpt2 bf16 测试结果:

test3(DP=8):
image

test4(DP=2,TP=4):
image

test5(DP=2,TP=4,SP=4):
image

@Chamberlain0w0
Copy link
Contributor Author

llama3 bf16 测试结果:

test3(DP=8):
image

test4(DP=2,TP=4):
image

test5(DP=2,TP=4,SP=4):
image

@kilinchange kilinchange merged commit 108928d into master Nov 10, 2025
2 checks passed
@kilinchange kilinchange deleted the fix/tp_patch branch November 10, 2025 09:23
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.

3 participants