README.md 6.13 KB
Newer Older
liangjing's avatar
v1  
liangjing committed
1
2
3
# 内容
- [内容](#内容)
- [环境配置](#环境配置)
wxj's avatar
wxj committed
4
5
6
7
8
9
10
11
12
13
14
- [预训练](#预训练)
  - [GPT](##GPT)
    - [下载词汇文件](###下载词汇文件)
    - [下载训练数据](###下载训练数据)
    - [数据预处理](###数据预处理)
    - [GPT预训练](###GPT预训练)
  - [Llama](##Llama)
    - [下载tokenizer文件](###下载tokenizer文件)
    - [下载训练数据](###下载训练数据)
    - [数据预处理](###数据预处理)
    - [Llama预训练](###Llama预训练)
liangjing's avatar
v1  
liangjing committed
15
16
- [参考](#参考)

wxj's avatar
wxj committed
17
18
# 更新日志

silencealiang's avatar
fix bug  
silencealiang committed
19
20
21
2025.3.14适配最新代码,shell启动脚本在examples对应模型目录下,模型相关数据集

[下载]: https://r0ddbu55vzx.feishu.cn/drive/folder/ZxHHfCoX4lg75td2hTqcmiAin3g
silencealiang's avatar
silencealiang committed
22

wxj's avatar
wxj committed
23
24
25
2024.12.16适配了torch prof

使用方法: 启动脚本中添加下列参数, 即可采集对应的prof信息
silencealiang's avatar
silencealiang committed
26
27
28
29
30
31

```python
# 采集torchprof
mpirun -np 8 --allow-run-as-root train_mixtral_8x7B_1nodes.sh localhost --profiling=torch
```

wxj's avatar
wxj committed
32
```bash
silencealiang's avatar
silencealiang committed
33
34
# prof相关参数
TORCH_PROFIE_ARGS=(
wxj's avatar
wxj committed
35
36
37
38
39
40
41
42
43
44
    --profile # 开启profile
    --profile-step-start 4 # skip前3个iter, warm第4个iter
    --profile-step-end 5 # 采集第5个iter
    --use-pytorch-profiler # 使用torch prof
    --profile-ranks 0 3 # 采集全局rank 第0和3
    --profile-dir ./prof_data # prof文件的保存目录
)
```


liangjing's avatar
v1  
liangjing committed
45
46
# 环境配置
1. 安装基础依赖包
Neel Kant's avatar
Neel Kant committed
47
<pre>
liangjing's avatar
v1  
liangjing committed
48
pip install -r requirements.txt
Neel Kant's avatar
Neel Kant committed
49
</pre>
wxj's avatar
wxj committed
50
2. 安装HCU相关whl包
Neel Kant's avatar
Neel Kant committed
51

wxj's avatar
wxj committed
52
HCU相关包下载目录:[https://cancon.hpccube.com:65024/4/main](https://cancon.hpccube.com:65024/4/main)
Neel Kant's avatar
Neel Kant committed
53

wxj's avatar
wxj committed
54
pytorch whl包:pytorch ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
55
根据python版本,下载对应pytorch的whl包
Neel Kant's avatar
Neel Kant committed
56
57

<pre>
liangjing's avatar
v1  
liangjing committed
58
pip install torch* (下载的torch的whl包)
Neel Kant's avatar
Neel Kant committed
59
</pre>
wxj's avatar
wxj committed
60
torchvision whl包:vision ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
61
根据python版本,下载对应torchvision的whl包
Mohammad's avatar
Mohammad committed
62
63

<pre>
liangjing's avatar
v1  
liangjing committed
64
pip install torchvision* (下载的torchvision的whl包)
Mohammad's avatar
Mohammad committed
65
</pre>
wxj's avatar
wxj committed
66
apex whl包:apex ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
67
根据python版本,下载对应apex的whl包
Mohammad's avatar
Mohammad committed
68
69

<pre>
liangjing's avatar
v1  
liangjing committed
70
pip install apex* (下载的apex的whl包)
71
</pre>
wxj's avatar
wxj committed
72

liangjing's avatar
v1  
liangjing committed
73
若使用 pip install 下载安装过慢,可添加源:-i https://pypi.tuna.tsinghua.edu.cn/simple/
Mohammad's avatar
Mohammad committed
74

wxj's avatar
wxj committed
75
76
77
# 预训练
## GPT
### 下载词汇文件
78

Mohammad's avatar
Mohammad committed
79
<pre>
liangjing's avatar
v1  
liangjing committed
80
81
wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-vocab.json
wget https://s3.amazonaws.com/models.huggingface.co/bert/gpt2-merges.txt
Mohammad's avatar
Mohammad committed
82
</pre>
83

wxj's avatar
wxj committed
84
### 下载训练数据
liangjing's avatar
v1  
liangjing committed
85
使用1GB 79K jsonl数据集
Mohammad's avatar
Mohammad committed
86
<pre>
liangjing's avatar
v1  
liangjing committed
87
88
wget https://huggingface.co/bigscience/misc-test-data/resolve/main/stas/oscar-1GB.jsonl.xz
xz -d oscar-1GB.jsonl.xz
Mohammad's avatar
Mohammad committed
89
</pre>
wxj's avatar
wxj committed
90
解压后为单个`oscar-1GB.jsonl`文件
Mohammad's avatar
Mohammad committed
91

wxj's avatar
wxj committed
92
### 数据预处理
Mohammad's avatar
Mohammad committed
93

wxj's avatar
wxj committed
94
```shell
liangjing's avatar
v1  
liangjing committed
95
96
python tools/preprocess_data.py \
    --input oscar-1GB.jsonl \ 
wxj's avatar
wxj committed
97
98
    --output-prefix ./dataset/oscar-1GB-gpt \
    --vocab-file gpt2-vocab.json \
liangjing's avatar
v1  
liangjing committed
99
100
101
102
    --tokenizer-type GPT2BPETokenizer \
    --merge-file gpt2-merges.txt \
    --append-eod \
    --workers 8
Mohammad's avatar
Mohammad committed
103

wxj's avatar
wxj committed
104
105
106
107
108
109
110
111
112
113
114
115
# 参数说明
# --input				输入数据集路径,即oscar-1GB.jsonl.xz解压后的文件路径
# --output-prefix		输出数据路径(需要输出目录已创建),处理后会自动加上_text_document后缀
# --vocab-file				下载的gpt2-vocab.json词表文件路径
# --tokenizer-type 	tokenizer类型
# --merge-file		下载的gpt2-merges.txt文件路径		
# --append-eod		添加结束标志符		
# --workers			进程数
```


### GPT预训练
wxj's avatar
wxj committed
116
脚本目录: `examples/gpt3/`
wxj's avatar
wxj committed
117
118
119
120
121
122
123
124
125
126
127
128
129
130

修改数据集与词汇文件路径
```shell
VOCAB_FILE=gpt2-vocab.json
MERGE_FILE=gpt2-merges.txt
DATA_PATH="./dataset/oscar-1GB-gpt_text_document"
```
- 单机多卡训练
  ```shell
  # 修改脚本中的分布式启动参数
  # 单机可以使用localhost指定通信地址为本地
  # -np 8指定8进程\(8卡\)并行
  # --allow-run-as-root以root权限启动
  mpirun --allow-run-as-root -np 8 GPT_pretraining.sh localhost >& GPT_pretraining.log
liangjing's avatar
v1  
liangjing committed
131
  ```
wxj's avatar
wxj committed
132
133
134
135
136
137
138
139
140
141
142
  注: 这里的`localhost`参数会传到脚本中的`--dist-url`

`GPT_pretraining.log`中查看训练日志

- 多机多卡训练
  
  多节点docker设置:
  1. 容器内执行/usr/sbin/sshd -p 12345,启动一个端口
  2. 容器间可通过该端口ssh登录,ssh ip -p 12345
  3. 如果需要免密,docker run容器时,docker -v /root/.ssh 挂载.ssh目录
  4. 容器间mpirun执行: `mpirun -np .. --hostfile hosts -mca plm_rsh_args "-p 12345" ./xx.sh master_ip`
Raul Puri's avatar
Raul Puri committed
143

wxj's avatar
wxj committed
144
145

  **案例**: 设有节点192.168.1.1和192.168.1.2两个节点, 每个节点上8张卡, 192.168.1.1作为master节点
146

wxj's avatar
wxj committed
147
148
149
150
  hosts文件:
  ```txt
  192.168.1.1 slots=8 
  192.168.1.2 slots=8
liangjing's avatar
v1  
liangjing committed
151
  ```
wxj's avatar
wxj committed
152
153
154
155

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
156
  mpirun --allow-run-as-root -np 16 --hostfile hosts -mca plm_rsh_no_tree_spawn 1 -mca plm_rsh_args "-p 12345" --bind-to none ./GPT_pretraining.sh 192.168.1.1 >& GPT_pretraining.log
liangjing's avatar
v1  
liangjing committed
157
  ```
wxj's avatar
wxj committed
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
`GPT_pretraining.log`中查看训练日志

## Llama
### 下载tokenizer文件

链接: https://www.modelscope.cn/models/shakechen/Llama-2-7b-hf/files
下载其中的tokenizer*文件

### 下载训练数据
使用1GB 79K jsonl数据集
<pre>
wget https://huggingface.co/bigscience/misc-test-data/resolve/main/stas/oscar-1GB.jsonl.xz
xz -d oscar-1GB.jsonl.xz
</pre>
解压后为单个`oscar-1GB.jsonl`文件

### 数据预处理

```shell
python tools/preprocess_data.py \
  --input oscar-1GB.jsonl \
  --output-prefix /datasets/oscar-1GB-llama\
  --tokenizer-type Llama2Tokenizer \
  --tokenizer-model /path/to/llama2_7b_hf/tokenizer.model \
  --workers 16 \
  --append-eod
```

### Llama预训练
wxj's avatar
wxj committed
187
脚本: `examples/llama`
wxj's avatar
wxj committed
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202

修改数据集与tokenizer路径
```shell
DATA_PATH="/datasets/oscar-1GB-llama_text_document"
--tokenizer-model /path/to/llama2_7b_hf/tokenizer.model
```
- 单机多卡训练
  ```shell
  # 具体参数说明参考上文GPT
  mpirun --allow-run-as-root -np 8 Llama_pretraining.sh localhost >& Llama_pretraining.log
  ```
`Llama_pretraining.log`中查看训练日志

- 多机多卡训练
  
wxj's avatar
wxj committed
203
  **案例**: 设有节点192.168.1.1和192.168.1.2两个节点, 每个节点上8张卡, 192.168.1.1作为master节点
wxj's avatar
wxj committed
204
205
206
207
208
209

  hosts配置如上文GTP所示

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
210
  mpirun --allow-run-as-root -np 16 --hostfile hosts -mca plm_rsh_no_tree_spawn 1 -mca plm_rsh_args "-p 12345" --bind-to none ./Llama_pretraining.sh 192.168.1.1 >& Llama_pretraining.log
wxj's avatar
wxj committed
211
212
213
  ```

`Llama_pretraining.log`中查看训练日志
214

liangjing's avatar
v1  
liangjing committed
215
# 参考
216

silencealiang's avatar
silencealiang committed
217
- [README_ORIGIN](README_ORIGIN.md)