README.md 6.02 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
silencealiang committed
19
20
2025.3.14适配最新代码,shell启动脚本在examples对应模型目录下

wxj's avatar
wxj committed
21
22
23
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
24
25
26
27
28
29

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

wxj's avatar
wxj committed
30
```bash
silencealiang's avatar
silencealiang committed
31
32
# prof相关参数
TORCH_PROFIE_ARGS=(
wxj's avatar
wxj committed
33
34
35
36
37
38
39
40
41
42
    --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
43
44
# 环境配置
1. 安装基础依赖包
Neel Kant's avatar
Neel Kant committed
45
<pre>
liangjing's avatar
v1  
liangjing committed
46
pip install -r requirements.txt
Neel Kant's avatar
Neel Kant committed
47
</pre>
wxj's avatar
wxj committed
48
2. 安装HCU相关whl包
Neel Kant's avatar
Neel Kant committed
49

wxj's avatar
wxj committed
50
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
51

wxj's avatar
wxj committed
52
pytorch whl包:pytorch ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
53
根据python版本,下载对应pytorch的whl包
Neel Kant's avatar
Neel Kant committed
54
55

<pre>
liangjing's avatar
v1  
liangjing committed
56
pip install torch* (下载的torch的whl包)
Neel Kant's avatar
Neel Kant committed
57
</pre>
wxj's avatar
wxj committed
58
torchvision whl包:vision ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
59
根据python版本,下载对应torchvision的whl包
Mohammad's avatar
Mohammad committed
60
61

<pre>
liangjing's avatar
v1  
liangjing committed
62
pip install torchvision* (下载的torchvision的whl包)
Mohammad's avatar
Mohammad committed
63
</pre>
wxj's avatar
wxj committed
64
apex whl包:apex ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
65
根据python版本,下载对应apex的whl包
Mohammad's avatar
Mohammad committed
66
67

<pre>
liangjing's avatar
v1  
liangjing committed
68
pip install apex* (下载的apex的whl包)
69
</pre>
wxj's avatar
wxj committed
70

liangjing's avatar
v1  
liangjing committed
71
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
72

wxj's avatar
wxj committed
73
74
75
# 预训练
## GPT
### 下载词汇文件
76

Mohammad's avatar
Mohammad committed
77
<pre>
liangjing's avatar
v1  
liangjing committed
78
79
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
80
</pre>
81

wxj's avatar
wxj committed
82
### 下载训练数据
liangjing's avatar
v1  
liangjing committed
83
使用1GB 79K jsonl数据集
Mohammad's avatar
Mohammad committed
84
<pre>
liangjing's avatar
v1  
liangjing committed
85
86
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
87
</pre>
wxj's avatar
wxj committed
88
解压后为单个`oscar-1GB.jsonl`文件
Mohammad's avatar
Mohammad committed
89

wxj's avatar
wxj committed
90
### 数据预处理
Mohammad's avatar
Mohammad committed
91

wxj's avatar
wxj committed
92
```shell
liangjing's avatar
v1  
liangjing committed
93
94
python tools/preprocess_data.py \
    --input oscar-1GB.jsonl \ 
wxj's avatar
wxj committed
95
96
    --output-prefix ./dataset/oscar-1GB-gpt \
    --vocab-file gpt2-vocab.json \
liangjing's avatar
v1  
liangjing committed
97
98
99
100
    --tokenizer-type GPT2BPETokenizer \
    --merge-file gpt2-merges.txt \
    --append-eod \
    --workers 8
Mohammad's avatar
Mohammad committed
101

wxj's avatar
wxj committed
102
103
104
105
106
107
108
109
110
111
112
113
# 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
114
脚本目录: `examples/gpt3/`
wxj's avatar
wxj committed
115
116
117
118
119
120
121
122
123
124
125
126
127
128

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
129
  ```
wxj's avatar
wxj committed
130
131
132
133
134
135
136
137
138
139
140
  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
141

wxj's avatar
wxj committed
142
143

  **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
144

wxj's avatar
wxj committed
145
146
147
148
  hosts文件:
  ```txt
  192.168.1.1 slots=8 
  192.168.1.2 slots=8
liangjing's avatar
v1  
liangjing committed
149
  ```
wxj's avatar
wxj committed
150
151
152
153

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
154
  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
155
  ```
wxj's avatar
wxj committed
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
`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
185
脚本: `examples/llama`
wxj's avatar
wxj committed
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200

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
201
  **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
202
203
204
205
206
207

  hosts配置如上文GTP所示

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
208
  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
209
210
211
  ```

`Llama_pretraining.log`中查看训练日志
212

liangjing's avatar
v1  
liangjing committed
213
# 参考
214

silencealiang's avatar
silencealiang committed
215
- [README_ORIGIN](README_ORIGIN.md)