aliyun / aliyun/SimAI

aicb生成的txt文件问题

Open
#271 4 comments 0 reactions 0 assignees View on GitHub

Nobody has claimed this yet.

Dominant language
Python
Stars
1.2k
Forks
184
Avg merge
13h 4m
Merged PRs (30d)
1

Description

你好!我用以下指令
sh ./scripts/megatron_workload_with_aiob.sh -m 7
--world_size 128 --tensor_model_parallel_size 4 --pipeline_model_parallel 4
--frame Megatron --global_batch 512
--micro_batch 1 --seq_length 4096 --swiglu
--use_flash_attn --aiob_enable
--comp_filepath workload/aiob_inputs/Example.txt
生成了一个txt文件,这里 -m 7的模型用官方给的模型,在aicb/scripts/megatron_workload_with_aiob.sh中写到的
7)model_name=gpt_7B
num_layers=36
hidden_size=4096
num_attention_heads=32
tensor_model_parallel_size=4;
意思就是这个模型是36层的,我设定-tensor_model_parallel_size 4 --pipeline_model_parallel 4 \ 也就是说DP=128\4\4=8的,pp为4意思是模型被分为了4部分,一部分里面有36/4=9层,而我的--global_batch 512 ,micro_batch 1,也就是说8个DP组分别要处理512/8=64个微批次,所以生成的文件中ga=64.
请问我生成的文件刚好是有64组重复的 以embedding_layer开头且attention_layer和mlp_layer共交替重复18行的数据,所以我想问这个txt文件只是显现出了模型的第一个pp组中的一张单卡gpu的计算和通信情况吗,也就是模型的前9层的计算和通信的情况吗?如果这样意思是模型后面几层的pp组里的所有的单卡gpu计算和通信的情况和这个txt文件中第一个pp组(模型前9层)单卡gpu的情况是一样的吗?

None-gpt_7B-world_size128-tp4-pp4-ep1-gbs512-mbs1-seq4096-MOE-False-GEMM-False-flash_attn-True.txt

Contributor guide

No contributing guide indexed for this repository

First steps

  1. Read the whole issue, then the project's contributing guide.
  2. Comment on the issue to say you are picking it up — it saves two people doing the same work.
  3. Fork the repository and make your change on a branch.
  4. Open a pull request that references the issue number.

Research direction

Start with aicb/scripts/megatron_workload_with_aiob.sh and the supplied workload/aiob_inputs/Example.txt output, tracing how the -m 7, tensor-model-parallel, pipeline-model-parallel, and batch settings determine the generated entries. Compare the output’s repeated layers with the stated PP and DP configuration. Done means documenting whether it represents one PP group or all groups, and whether later PP groups have identical per-GPU computation and communication.

Written by the indexing model from the issue text.

Assessment

Tech stack
bash, python
Domain
distributed-systems, performance
Issue type
Bug
Difficulty
4/5
Estimated time
3-5 days
Activity status
Quiet
Clarity
Needs clarification
Newbie friendliness
30/100

Get new issues in your inbox

A short digest of beginner-friendly GitHub issues.