README.md 6.31 KB
Newer Older
wxj's avatar
wxj committed
1
2
3
该仓库计划即将停止维护, 后续维护仓库地址: https://developer.sourcefind.cn/codes/OpenDAS/dcu_megatron


liangjing's avatar
v1  
liangjing committed
4
5
6
# 内容
- [内容](#内容)
- [环境配置](#环境配置)
wxj's avatar
wxj committed
7
8
9
10
11
12
13
14
15
16
17
- [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
18
19
- [参考](#参考)

wxj's avatar
wxj committed
20
21
22
23
24
# 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
silencealiang's avatar
silencealiang committed
25
26
27
28
29
30
31
32
33

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

# 采集hipprof
mpirun -np 8 --allow-run-as-root train_mixtral_8x7B_1nodes.sh localhost --profiling=hip
```

wxj's avatar
wxj committed
34
```bash
silencealiang's avatar
silencealiang committed
35
36
# prof相关参数
TORCH_PROFIE_ARGS=(
wxj's avatar
wxj committed
37
38
39
40
41
42
43
    --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
)
silencealiang's avatar
silencealiang committed
44
45
46
47
48
49
50
51

HIP_PROFIE_ARGS=(
    --profile
    --profile-ranks 0 1 2 3 4 5 6 7
    --profile-step-start 4
    --profile-step-end 5
    --use-hip-profiler
)
wxj's avatar
wxj committed
52
53
54
```


liangjing's avatar
v1  
liangjing committed
55
56
# 环境配置
1. 安装基础依赖包
Neel Kant's avatar
Neel Kant committed
57
<pre>
liangjing's avatar
v1  
liangjing committed
58
pip install -r requirements.txt
Neel Kant's avatar
Neel Kant committed
59
</pre>
wxj's avatar
wxj committed
60
2. 安装HCU相关whl包
Neel Kant's avatar
Neel Kant committed
61

wxj's avatar
wxj committed
62
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
63

wxj's avatar
wxj committed
64
pytorch whl包:pytorch ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
65
根据python版本,下载对应pytorch的whl包
Neel Kant's avatar
Neel Kant committed
66
67

<pre>
liangjing's avatar
v1  
liangjing committed
68
pip install torch* (下载的torch的whl包)
Neel Kant's avatar
Neel Kant committed
69
</pre>
wxj's avatar
wxj committed
70
torchvision whl包:vision ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
71
根据python版本,下载对应torchvision的whl包
Mohammad's avatar
Mohammad committed
72
73

<pre>
liangjing's avatar
v1  
liangjing committed
74
pip install torchvision* (下载的torchvision的whl包)
Mohammad's avatar
Mohammad committed
75
</pre>
wxj's avatar
wxj committed
76
apex whl包:apex ---> dtk-24.04.1
liangjing's avatar
v1  
liangjing committed
77
根据python版本,下载对应apex的whl包
Mohammad's avatar
Mohammad committed
78
79

<pre>
liangjing's avatar
v1  
liangjing committed
80
pip install apex* (下载的apex的whl包)
81
</pre>
wxj's avatar
wxj committed
82

liangjing's avatar
v1  
liangjing committed
83
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
84

wxj's avatar
wxj committed
85
86
87
# 预训练
## GPT
### 下载词汇文件
88

Mohammad's avatar
Mohammad committed
89
<pre>
liangjing's avatar
v1  
liangjing committed
90
91
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
92
</pre>
93

wxj's avatar
wxj committed
94
### 下载训练数据
liangjing's avatar
v1  
liangjing committed
95
使用1GB 79K jsonl数据集
Mohammad's avatar
Mohammad committed
96
<pre>
liangjing's avatar
v1  
liangjing committed
97
98
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
99
</pre>
wxj's avatar
wxj committed
100
解压后为单个`oscar-1GB.jsonl`文件
Mohammad's avatar
Mohammad committed
101

wxj's avatar
wxj committed
102
### 数据预处理
Mohammad's avatar
Mohammad committed
103

wxj's avatar
wxj committed
104
```shell
liangjing's avatar
v1  
liangjing committed
105
106
python tools/preprocess_data.py \
    --input oscar-1GB.jsonl \ 
wxj's avatar
wxj committed
107
108
    --output-prefix ./dataset/oscar-1GB-gpt \
    --vocab-file gpt2-vocab.json \
liangjing's avatar
v1  
liangjing committed
109
110
111
112
    --tokenizer-type GPT2BPETokenizer \
    --merge-file gpt2-merges.txt \
    --append-eod \
    --workers 8
Mohammad's avatar
Mohammad committed
113

wxj's avatar
wxj committed
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
# 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
141
  ```
wxj's avatar
wxj committed
142
143
144
145
146
147
148
149
150
151
152
  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
153

wxj's avatar
wxj committed
154
155

  **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
156

wxj's avatar
wxj committed
157
158
159
160
  hosts文件:
  ```txt
  192.168.1.1 slots=8 
  192.168.1.2 slots=8
liangjing's avatar
v1  
liangjing committed
161
  ```
wxj's avatar
wxj committed
162
163
164
165

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
166
  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
167
  ```
wxj's avatar
wxj committed
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
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
`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
213
  **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
214
215
216
217
218
219

  hosts配置如上文GTP所示

  在master节点执行命令

  ```shell
wxj's avatar
wxj committed
220
  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
221
222
223
  ```

`Llama_pretraining.log`中查看训练日志
224

liangjing's avatar
v1  
liangjing committed
225
# 参考
226

silencealiang's avatar
silencealiang committed
227
- [README_ORIGIN](README_ORIGIN.md)