README.md 10.6 KB
Newer Older
Rayyyyy's avatar
Rayyyyy committed
1
2
3
4
5
# llama3
## 论文
[llama3](https://llama.meta.com/llama3/)

## 模型结构
Rayyyyy's avatar
Rayyyyy committed
6
7
8
9
10
Llama-3中选择了一个相对标准的decoder-only的transformer架构。与Llama-2相比,做了几个关键的改进:
- 基于超过15T token训练数据,大小相当于Llama 2数据集的7倍还多,增强了推理、代码生成和指令跟随等方面的能力;
- 支持8K长文本(之前是4k),改进的tokenizer具有128K tokens的词汇量,可以更有效地对语言进行编码,从而大大提高了模型的性能;
- 采用分组查询注意力(grouped query attention,GQA)、掩码等技术,帮助开发者以最低的能耗获取绝佳的性能。
- 在8,192个tokens的序列上训练模型,使用掩码来确保self-attention不会跨越文档边界。
Rayyyyy's avatar
Rayyyyy committed
11
12
13
14
15
16
17
18
19

## 算法原理

<div align=center>
    <img src="./doc/method.png"/>
</div>

## 环境配置
-v 路径、docker_name和imageID根据实际情况修改
Rayyyyy's avatar
Rayyyyy committed
20
**注意**:bitsandbytes库功能不全,暂不支持4bits
Rayyyyy's avatar
Rayyyyy committed
21
22
23
24

### Docker(方法一)

```bash
Rayyyyy's avatar
Rayyyyy committed
25
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-ubuntu22.04-dtk23.10.1-py310
Rayyyyy's avatar
Rayyyyy committed
26
27
28
docker run -it -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro --shm-size=32G --privileged=true --device=/dev/kfd --device=/dev/dri/ --group-add video --name docker_name imageID bash

cd /your_code_path/llama3_pytorch
Rayyyyy's avatar
Rayyyyy committed
29

Rayyyyy's avatar
Rayyyyy committed
30
pip install -e .
Rayyyyy's avatar
Rayyyyy committed
31

Rayyyyy's avatar
Rayyyyy committed
32
pip install deepspeed-0.12.3+gitfe61783.abi0.dtk2310.torch2.1.0a0-cp310-cp310-manylinux2014_x86_64.whl
Rayyyyy's avatar
Rayyyyy committed
33
pip install bitsandbytes-0.43.0-py3-none-any.whl
Rayyyyy's avatar
Rayyyyy committed
34
35
pip install -U xtuner # 0.1.18
pip install mmengine==0.10.3
Rayyyyy's avatar
Rayyyyy committed
36
37
38
39
40
```

### Dockerfile(方法二)

```bash
Rayyyyy's avatar
Rayyyyy committed
41
cd docker
Rayyyyy's avatar
Rayyyyy committed
42
43
44
45
docker build --no-cache -t llama3:latest .
docker run -it -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro --shm-size=32G --privileged=true --device=/dev/kfd --device=/dev/dri/ --group-add video --name docker_name imageID bash

cd /your_code_path/llama3_pytorch
Rayyyyy's avatar
Rayyyyy committed
46

Rayyyyy's avatar
Rayyyyy committed
47
pip install -e .
Rayyyyy's avatar
Rayyyyy committed
48

Rayyyyy's avatar
Rayyyyy committed
49
pip install deepspeed-0.12.3+gitfe61783.abi0.dtk2310.torch2.1.0a0-cp310-cp310-manylinux2014_x86_64.whl
Rayyyyy's avatar
Rayyyyy committed
50
pip install bitsandbytes-0.43.0-py3-none-any.whl
Rayyyyy's avatar
Rayyyyy committed
51
52
pip install -U xtuner # 0.1.18
pip install mmengine==0.10.3
Rayyyyy's avatar
Rayyyyy committed
53
54
55
56
57
```

### Anaconda(方法三)
关于本项目DCU显卡所需的特殊深度学习库可从[光合](https://developer.hpccube.com/tool/)开发者社区下载安装。
```bash
Rayyyyy's avatar
Rayyyyy committed
58
DTK驱动: dtk23.10.1
Rayyyyy's avatar
Rayyyyy committed
59
python: python3.10
Rayyyyy's avatar
Rayyyyy committed
60
torch: 2.1.0
Rayyyyy's avatar
Rayyyyy committed
61
xtuner: 0.1.18
Rayyyyy's avatar
Rayyyyy committed
62
```
Rayyyyy's avatar
Rayyyyy committed
63
`Tips:以上dtk驱动、python、torch等DCU相关工具版本需要严格一一对应`
Rayyyyy's avatar
Rayyyyy committed
64
65
66
67

其它非深度学习库安装方式如下:
```bash
pip install -e .
Rayyyyy's avatar
Rayyyyy committed
68

Rayyyyy's avatar
Rayyyyy committed
69
pip install deepspeed-0.12.3+gitfe61783.abi0.dtk2310.torch2.1.0a0-cp310-cp310-manylinux2014_x86_64.whl
Rayyyyy's avatar
Rayyyyy committed
70
pip install bitsandbytes-0.43.0-py3-none-any.whl
Rayyyyy's avatar
Rayyyyy committed
71
72
pip install -U xtuner # 0.1.18
pip install mmengine==0.10.3
Rayyyyy's avatar
Rayyyyy committed
73
74
75
```

## 数据集
Rayyyyy's avatar
Rayyyyy committed
76
77
78
79
80
81
```
├── llama3_pytorch
│   ├── datasets
│       ├── alpaca_data.json
│       └── multi_turn_dataset_2.json
```
Rayyyyy's avatar
Rayyyyy committed
82
83

## 训练
Rayyyyy's avatar
Rayyyyy committed
84
### xtuner微调方法
Rayyyyy's avatar
Rayyyyy committed
85
86
87
88
89
90
91
92
93
94
95
1. 下载预训练模型,具体模型请修改 `download_models.py`
```bash
cd /your_code_path/llama3_pytorch
pip install modelscope
python download_models.py
mv ~/.cache/modelscope/hub/LLM-Research ./
```
2. 修改[llama3_8b_instruct_qlora_alpaca_e3_M.py](./llama3_8b_instruct_qlora_alpaca_e3_M.py)代码中的`pretrained_model_name_or_path``data_path`为本地对应数据地址;
3. 根据硬件环境和自身训练需求来调整 `max_length``batch_size``accumulative_counts``max_epochs``lr``save_steps``evaluation_freq`、model.lora中的`r``lora_alpha`参数,默认参数支持4*32G;
4. ${DCU_NUM}参数修改为要使用的DCU卡数量,不同数据集需要修改llama3_8b_instruct_qlora_alpaca_e3_M.py中`SYSTEM``evaluation_inputs``dataset_map_fn``train_dataloader.sampler``train_cfg`参数设置,详情请参考代码注释项,当前默认alpaca数据集。
5. 执行
Rayyyyy's avatar
Rayyyyy committed
96
```bash
Rayyyyy's avatar
Rayyyyy committed
97
98
bash finetune.sh
or
Rayyyyy's avatar
Rayyyyy committed
99
100
NPROC_PER_NODE=${DCU_NUM} xtuner train ./llama3_8b_instruct_qlora_alpaca_e3_M.py --deepspeed deepspeed_zero2
```
Rayyyyy's avatar
Rayyyyy committed
101
102

## 推理
Rayyyyy's avatar
Rayyyyy committed
103
预训练模型下载方法请参考下面的[预训练权重](#预训练权重)章节,不同的模型需要不同的模型并行(MP)值,如下表所示:
Rayyyyy's avatar
Rayyyyy committed
104

Rayyyyy's avatar
Rayyyyy committed
105
106
107
|  Model | MP |
|--------|----|
| 8B     | 1  |
Rayyyyy's avatar
Rayyyyy committed
108
| 70B    | 8  |
Rayyyyy's avatar
Rayyyyy committed
109
110
111
112

所有模型都支持序列长度高达8192个tokens,但我们根据max_seq_len和max_batch_size值预先分配缓存。根据你的硬件设置。

**Tips:**
Rayyyyy's avatar
Rayyyyy committed
113
- `–nproc_per_node`需要根据模型的MP值进行设置(参考上表)。
Rayyyyy's avatar
Rayyyyy committed
114
115
116
- `max_seq_len``max_batch_size`参数按需设置。

### Pretrained模型
Rayyyyy's avatar
Rayyyyy committed
117
这些模型都没有针对聊天或者Q&A进行微调。可以参考`example_text_completion.py`里的用例。
Rayyyyy's avatar
Rayyyyy committed
118

Rayyyyy's avatar
Rayyyyy committed
119
- Meta-Llama-3-8B 模型示例,Meta-Llama-3-70B模型仅需替换--ckpt_dir、--tokenizer_path对应模型地址即可。
Rayyyyy's avatar
Rayyyyy committed
120
```bash
Rayyyyy's avatar
Rayyyyy committed
121
torchrun --nproc_per_node 8 example_text_completion.py \
Rayyyyy's avatar
Rayyyyy committed
122
    --ckpt_dir Meta-Llama-3-8B/original/ \
Rayyyyy's avatar
Rayyyyy committed
123
124
125
126
    --tokenizer_path Meta-Llama-3-8B/original/tokenizer.model \
    --max_seq_len 128 --max_batch_size 4
```

Rayyyyy's avatar
Rayyyyy committed
127
### Instruction-tuned模型
Rayyyyy's avatar
Rayyyyy committed
128
129
130
131
132
经过微调的模型被训练用于对话应用程序。为了获得模型的预期特性和性能,需要遵循 [`ChatFormat`](llama/tokenizer.py#L202)中定义的特定格式:
- 提示以特殊令牌 <|begin_of_text|> 开始,之后跟随一个或多个消息。
- 每条消息以标签`<|start_header_id|>`开始,角色为`system``user`或者`assistant`、并以标签 `<|end_header_id|>`  结束。
- 在双换行符`\n\n`之后,消息的内容随之而来。
- 每条消息的结尾由`<|eot_id|>`令牌标记。
Rayyyyy's avatar
Rayyyyy committed
133
134
135

您还可以部署额外的分类器来过滤被认为不安全的输入和输出。有关如何向推理代码的输入和输出添加安全检查器,请参阅[llama-recipes repo](https://github.com/meta-llama/llama-recipes/blob/main/recipes/inference/local_inference/inference.py)

Rayyyyy's avatar
Rayyyyy committed
136
- Meta-Llama-3-8B-Instruct 模型示例,Meta-Llama-3-70B-Instruct模型仅需替换--ckpt_dir、--tokenizer_path对应模型地址即可。
Rayyyyy's avatar
Rayyyyy committed
137
138
139
140
141
142
```bash
torchrun --nproc_per_node 1 example_chat_completion.py \
    --ckpt_dir Meta-Llama-3-8B-Instruct/original/ \
    --tokenizer_path Meta-Llama-3-8B-Instruct/original/tokenizer.model \
    --max_seq_len 512 --max_batch_size 6
```
Rayyyyy's avatar
Rayyyyy committed
143
144
145
146
147
148
149
## 多轮对话
1. 确认环境安装及模型下载完毕;
2. 修改[chat.sh](./chat.sh)文件中的 `--ckpt_dir``--tokenizer_path` 参数为本地模型地址,`--max_seq_len` 根据自身需求进行修改,调整该值可以增加多轮对话模型的记忆长度,不过需要注意的是这可能会增加模型运算的时间和内存需求;
3. 执行:
```bash
bash chat.sh
```
Rayyyyy's avatar
Rayyyyy committed
150

Rayyyyy's avatar
Rayyyyy committed
151
## Evaluation
Rayyyyy's avatar
Rayyyyy committed
152
153
154
155
156
157
158
159
160
1. 安装 `llama-recipes``lm-eval`
```bash
# llama-recipes 下载
git clone https://github.com/meta-llama/llama-recipes.git
cd ./llama-recipes/recipes/evaluation/
# 修改eval.py第15行代码,将from lm_eval.utils import make_table 改为
from lm_eval.evaluator import make_table
# 修改eval.py第121行代码,num_fewshot参数的默认值改为0
default=0
Rayyyyy's avatar
Rayyyyy committed
161
162
# 修改eval.py第215行代码,use_cache=args.use_cache 修改为
no_cache=args.use_cache
Rayyyyy's avatar
Rayyyyy committed
163
164
165

# 返回根目录
cd ~
Rayyyyy's avatar
Rayyyyy committed
166

Rayyyyy's avatar
Rayyyyy committed
167
168
169
170
171
172
173
174
175
# lm-eval 下载
git clone http://developer.hpccube.com/codes/chenych/lm-evaluation-harness.git
cd ./lm-evaluation-harness/
pip install -e .
cd ../
```

2. 修改待测模型**pretrained**参数地址,例如 `/home/Meta-Llama-3-8B-Instruct`,特别地,当前仅支持`hellaswag`数据集进行测试验证。执行以下命令:
```bash
Rayyyyy's avatar
Rayyyyy committed
176
cd /path_of/llama-recipes/recipes/evaluation
Rayyyyy's avatar
Rayyyyy committed
177
178
python eval.py --model hf --model_args pretrained=/home/llama3/Meta-Llama-3-8B-Instruct,dtype="float" --tasks hellaswag --device cuda --batch_size 8
```
Rayyyyy's avatar
Rayyyyy committed
179
180
181
<div align=center>
    <img src="./doc/evaluation.png"/>
</div>
Rayyyyy's avatar
Rayyyyy committed
182

Rayyyyy's avatar
Rayyyyy committed
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
## result
- Meta-Llama-3-8B-Instruct
<div align=center>
    <img src="./doc/Meta-Llama-3-8B-Instruct.png"/>
</div>

- Meta-Llama-3-8B
<div align=center>
    <img src="./doc/Meta-Llama-3-8B.png"/>
</div>

### 精度
暂无


## 应用场景
### 算法类别
对话问答

### 热点应用行业
制造,广媒,家居,教育

## 预训练权重
1. 环境安装
```bash
pip install -U huggingface_hub hf_transfer
export HF_ENDPOINT=https://hf-mirror.com
```

2. 预训练模型下载,**token**参数通过huggingface账号获取
Rayyyyy's avatar
Rayyyyy committed
213
214
215
216
217
218
219

- Meta-Llama-3-8B 模型
```bash
mkdir Meta-Llama-3-8B
huggingface-cli download meta-llama/Meta-Llama-3-8B --include "original/*" --local-dir Meta-Llama-3-8B --token hf_*
```

Rayyyyy's avatar
Rayyyyy committed
220
221
222
- Meta-Llama-3-8B-Instruct 模型
```bash
mkdir Meta-Llama-3-8B-Instruct
Rayyyyy's avatar
Rayyyyy committed
223
huggingface-cli download meta-llama/Meta-Llama-3-8B-Instruct --include "original/*" --local-dir Meta-Llama-3-8B-Instruct --token hf_*
Rayyyyy's avatar
Rayyyyy committed
224
```
Rayyyyy's avatar
Rayyyyy committed
225

Rayyyyy's avatar
Rayyyyy committed
226
227
228
229
230
231
232
233
234
235
236
237
- Meta-Llama-3-70B 模型
```bash
mkdir Meta-Llama-3-70B
huggingface-cli download meta-llama/Meta-Llama-3-70B --include "original/*" --local-dir Meta-Llama-3-70B --token hf_*
```

- Meta-Llama-3-70B-Instruct 模型
```bash
mkdir Meta-Llama-3-70B-Instruct
huggingface-cli download meta-llama/Meta-Llama-3-70B-Instruct --include "original/*" --local-dir Meta-Llama-3-70B-Instruct --token hf_*
```

Rayyyyy's avatar
Rayyyyy committed
238
239
240
模型目录结构如下:
```bash
├── llama3_pytorch
Rayyyyy's avatar
Rayyyyy committed
241
│   ├── Meta-Llama-3-8B
Rayyyyy's avatar
Rayyyyy committed
242
│       └── original
Rayyyyy's avatar
Rayyyyy committed
243
244
245
│           ├── consolidated.00.pth
│           ├── params.json
│           └── tokenizer.model
Rayyyyy's avatar
Rayyyyy committed
246
│   ├── Meta-Llama-3-8B-Instruct
Rayyyyy's avatar
Rayyyyy committed
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
│       └── original
│           ├── consolidated.00.pth
│           ├── params.json
│           └── tokenizer.model
│   ├── Meta-Llama-3-70B
│       └── original
│           ├── consolidated.00.pth
│           ├── consolidated.01.pth
│           ├── consolidated.02.pth
│           ├── consolidated.03.pth
│           ├── consolidated.04.pth
│           ├── consolidated.05.pth
│           ├── consolidated.06.pth
│           ├── consolidated.07.pth
│           ├── params.json
│           └── tokenizer.model
│   └── Meta-Llama-3-70B-Instruct
│       └── original
Rayyyyy's avatar
Rayyyyy committed
265
│           ├── consolidated.00.pth
Rayyyyy's avatar
Rayyyyy committed
266
267
268
269
270
271
272
│           ├── consolidated.01.pth
│           ├── consolidated.02.pth
│           ├── consolidated.03.pth
│           ├── consolidated.04.pth
│           ├── consolidated.05.pth
│           ├── consolidated.06.pth
│           ├── consolidated.07.pth
Rayyyyy's avatar
Rayyyyy committed
273
274
275
276
│           ├── params.json
│           └── tokenizer.model
```

Rayyyyy's avatar
Rayyyyy committed
277
278
279
280
281
## 源码仓库及问题反馈
- https://developer.hpccube.com/codes/modelzoo/llama3_pytorch

## 参考资料
- https://github.com/meta-llama/llama3
Rayyyyy's avatar
Rayyyyy committed
282
283
- https://github.com/InternLM/xtuner
- https://github.com/SmartFlowAI/EmoLLM
Rayyyyy's avatar
Rayyyyy committed
284
- https://github.com/meta-llama/llama-recipes