README.md 5.85 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
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
# 更新日志

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
```bash
PROFILE_ARGS=(
    --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
)
APP="... \
    ${PROFILE_ARGS[@]} \
"
${APP}
```


liangjing's avatar
v1  
liangjing committed
38
39
# 环境配置
1. 安装基础依赖包
Neel Kant's avatar
Neel Kant committed
40
<pre>
liangjing's avatar
v1  
liangjing committed
41
pip install -r requirements.txt
Neel Kant's avatar
Neel Kant committed
42
</pre>
wxj's avatar
wxj committed
43
2. 安装HCU相关whl包
Neel Kant's avatar
Neel Kant committed
44

wxj's avatar
wxj committed
45
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
46

wxj's avatar
wxj committed
47
pytorch whl包:pytorch ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
48
根据python版本,下载对应pytorch的whl包
Neel Kant's avatar
Neel Kant committed
49
50

<pre>
liangjing's avatar
v1  
liangjing committed
51
pip install torch* (下载的torch的whl包)
Neel Kant's avatar
Neel Kant committed
52
</pre>
wxj's avatar
wxj committed
53
torchvision whl包:vision ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
54
根据python版本,下载对应torchvision的whl包
Mohammad's avatar
Mohammad committed
55
56

<pre>
liangjing's avatar
v1  
liangjing committed
57
pip install torchvision* (下载的torchvision的whl包)
Mohammad's avatar
Mohammad committed
58
</pre>
wxj's avatar
wxj committed
59
apex whl包:apex ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
60
根据python版本,下载对应apex的whl包
Mohammad's avatar
Mohammad committed
61
62

<pre>
liangjing's avatar
v1  
liangjing committed
63
pip install apex* (下载的apex的whl包)
64
</pre>
wxj's avatar
wxj committed
65

liangjing's avatar
v1  
liangjing committed
66
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
67

wxj's avatar
wxj committed
68
69
70
# 预训练
## GPT
### 下载词汇文件
71

Mohammad's avatar
Mohammad committed
72
<pre>
liangjing's avatar
v1  
liangjing committed
73
74
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
75
</pre>
76

wxj's avatar
wxj committed
77
### 下载训练数据
liangjing's avatar
v1  
liangjing committed
78
使用1GB 79K jsonl数据集
Mohammad's avatar
Mohammad committed
79
<pre>
liangjing's avatar
v1  
liangjing committed
80
81
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
82
</pre>
wxj's avatar
wxj committed
83
解压后为单个`oscar-1GB.jsonl`文件
Mohammad's avatar
Mohammad committed
84

wxj's avatar
wxj committed
85
### 数据预处理
Mohammad's avatar
Mohammad committed
86

wxj's avatar
wxj committed
87
```shell
liangjing's avatar
v1  
liangjing committed
88
89
python tools/preprocess_data.py \
    --input oscar-1GB.jsonl \ 
wxj's avatar
wxj committed
90
91
    --output-prefix ./dataset/oscar-1GB-gpt \
    --vocab-file gpt2-vocab.json \
liangjing's avatar
v1  
liangjing committed
92
93
94
95
    --tokenizer-type GPT2BPETokenizer \
    --merge-file gpt2-merges.txt \
    --append-eod \
    --workers 8
Mohammad's avatar
Mohammad committed
96

wxj's avatar
wxj committed
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
# 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
脚本: `GPT_pretraining.sh`

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
124
  ```
wxj's avatar
wxj committed
125
126
127
128
129
130
131
132
133
134
135
  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
136

wxj's avatar
wxj committed
137
138

  **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
139

wxj's avatar
wxj committed
140
141
142
143
  hosts文件:
  ```txt
  192.168.1.1 slots=8 
  192.168.1.2 slots=8
liangjing's avatar
v1  
liangjing committed
144
  ```
wxj's avatar
wxj committed
145
146
147
148

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
149
  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
150
  ```
wxj's avatar
wxj committed
151
152
153
154
155
156
157
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
187
188
189
190
191
192
193
194
195
`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
脚本: `Llama_pretraining.sh`

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
196
  **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
197
198
199
200
201
202

  hosts配置如上文GTP所示

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
203
  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
204
205
206
  ```

`Llama_pretraining.log`中查看训练日志
207

liangjing's avatar
v1  
liangjing committed
208
# 参考
209

silencealiang's avatar
silencealiang committed
210
- [README_ORIGIN](README_ORIGIN.md)