README.md 12.3 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


## 环境配置
-v 路径、docker_name和imageID根据实际情况修改
Rayyyyy's avatar
Rayyyyy committed
15
**注意**:bitsandbytes库功能不全,暂不支持量化相关
Rayyyyy's avatar
Rayyyyy committed
16
17
18

### Docker(方法一)
```bash
Rayyyyy's avatar
Rayyyyy committed
19
20
docker pull image.sourcefind.cn:5000/dcu/admin/base/pytorch:2.1.0-centos7.6-dtk24.04-py310
docker run -it -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro --shm-size=80G --privileged=true --device=/dev/kfd --device=/dev/dri/ --group-add video --name docker_name imageID bash
Rayyyyy's avatar
Rayyyyy committed
21
22
23
24
25
26
27

cd /your_code_path/llama3_pytorch
pip install -e .
```

### Dockerfile(方法二)
```bash
Rayyyyy's avatar
Rayyyyy committed
28
cd docker
Rayyyyy's avatar
Rayyyyy committed
29
docker build --no-cache -t llama3:latest .
Rayyyyy's avatar
Rayyyyy committed
30
docker run -it -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro --shm-size=80G --privileged=true --device=/dev/kfd --device=/dev/dri/ --group-add video --name docker_name imageID bash
Rayyyyy's avatar
Rayyyyy committed
31
32
33
34
35
36
37
38

cd /your_code_path/llama3_pytorch
pip install -e .
```

### 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
39
DTK驱动: dtk24.04
Rayyyyy's avatar
Rayyyyy committed
40
python: python3.10
Rayyyyy's avatar
Rayyyyy committed
41
torch: 2.1.0
Rayyyyy's avatar
Rayyyyy committed
42
xtuner: 0.1.18
Rayyyyy's avatar
Rayyyyy committed
43
```
Rayyyyy's avatar
Rayyyyy committed
44
`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
45
46
47
48
49
50
51

其它非深度学习库安装方式如下:
```bash
pip install -e .
```

## 数据集
Rayyyyy's avatar
Rayyyyy committed
52
53
54
55
56
57
```
├── llama3_pytorch
│   ├── datasets
│       ├── alpaca_data.json
│       └── multi_turn_dataset_2.json
```
Rayyyyy's avatar
Rayyyyy committed
58
59

## 训练
Rayyyyy's avatar
Rayyyyy committed
60
### xtuner微调方法
Rayyyyy's avatar
Rayyyyy committed
61
1. 训练库安装(非llama3_pytorch目录下),请注意所需库版本
Rayyyyy's avatar
Rayyyyy committed
62
```bash
Rayyyyy's avatar
Rayyyyy committed
63
64
pip uninstall flash-attn # 2.0.4+82379d7.abi0.dtk2404.torch2.1
# docker环境含有deepspeed的可不进行安装, 需要对照版本是否一致即可
Rayyyyy's avatar
Rayyyyy committed
65
pip install deepspeed-0.12.3+das1.0+gita724046.abi0.dtk2404.torch2.1.0-cp310-cp310-manylinux2014_x86_64.whl
Rayyyyy's avatar
Rayyyyy committed
66
67
68
git clone -b v0.1.18 https://github.com/InternLM/xtuner.git
cd xtuner
pip install -e '.[all]'
Rayyyyy's avatar
Rayyyyy committed
69
pip install mmengine==0.10.3
Rayyyyy's avatar
Rayyyyy committed
70
71
# 注意bitsandbytes库版本,如果环境中一致可不安装,否则需要重新安装
pip install bitsandbytes-0.37.0+das1.0+gitd3d888f.abi0.dtk2404.torch2.1-py3-none-any.whl
Rayyyyy's avatar
Rayyyyy committed
72
73
```
2. 下载预训练模型,具体模型请修改`download_models.py`
Rayyyyy's avatar
Rayyyyy committed
74
75
76
77
78
```bash
cd /your_code_path/llama3_pytorch
pip install modelscope
python download_models.py
```
Rayyyyy's avatar
Rayyyyy committed
79
80
81
82
3. 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`为本地模型、数据地址;
4. 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;
5. ${DCU_NUM}参数修改为要使用的DCU卡数量,不同数据集需要修改llama3_8b_instruct_qlora_alpaca_e3_M.py中`SYSTEM``evaluation_inputs``dataset_map_fn``train_dataloader.sampler``train_cfg`参数设置,详情请参考代码注释项,当前默认alpaca数据集,**`--work-dir`设定保存模型路径**
6. 执行
Rayyyyy's avatar
Rayyyyy committed
83
```bash
Rayyyyy's avatar
Rayyyyy committed
84
85
bash finetune.sh
or
Rayyyyy's avatar
Rayyyyy committed
86
87
NPROC_PER_NODE=${DCU_NUM} xtuner train ./llama3_8b_instruct_qlora_alpaca_e3_M.py --deepspeed deepspeed_zero2
```
Rayyyyy's avatar
Rayyyyy committed
88
89

## 推理
Rayyyyy's avatar
Rayyyyy committed
90
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
91

Rayyyyy's avatar
Rayyyyy committed
92
93
94
|  Model | MP |
|--------|----|
| 8B     | 1  |
Rayyyyy's avatar
Rayyyyy committed
95
| 70B    | 8  |
Rayyyyy's avatar
Rayyyyy committed
96
97
98
99

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
100
- `–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
101
102
103
- `max_seq_len``max_batch_size`参数按需设置。

### Pretrained模型
Rayyyyy's avatar
Rayyyyy committed
104
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
105

Rayyyyy's avatar
Rayyyyy committed
106
- Meta-Llama-3-8B 模型示例,Meta-Llama-3-70B模型仅需替换–-nproc_per_node、--ckpt_dir、--tokenizer_path对应模型地址即可。
Rayyyyy's avatar
Rayyyyy committed
107
```bash
Rayyyyy's avatar
Rayyyyy committed
108
torchrun --nproc_per_node 1 example_text_completion.py \
Rayyyyy's avatar
Rayyyyy committed
109
    --ckpt_dir Meta-Llama-3-8B/original/ \
Rayyyyy's avatar
Rayyyyy committed
110
111
112
113
    --tokenizer_path Meta-Llama-3-8B/original/tokenizer.model \
    --max_seq_len 128 --max_batch_size 4
```

Rayyyyy's avatar
Rayyyyy committed
114
### Instruction-tuned模型
Rayyyyy's avatar
Rayyyyy committed
115
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:
Rayyyyy's avatar
Rayyyyy committed
116
117
- 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。
Rayyyyy's avatar
Rayyyyy committed
118
119
- 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
120
121
122

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
123
- Meta-Llama-3-8B-Instruct 模型示例,Meta-Llama-3-70B-Instruct模型仅需替换–-nproc_per_node、--ckpt_dir、--tokenizer_path对应模型地址即可。
Rayyyyy's avatar
Rayyyyy committed
124
125
126
127
128
129
```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
130

Rayyyyy's avatar
Rayyyyy committed
131
132
## 多轮对话
1. 确认环境安装及模型下载完毕;
Rayyyyy's avatar
Rayyyyy committed
133
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;
Rayyyyy's avatar
Rayyyyy committed
134
135
136
137
3. 执行:
```bash
bash chat.sh
```
Rayyyyy's avatar
Rayyyyy committed
138

Rayyyyy's avatar
Rayyyyy committed
139
## Evaluation
Rayyyyy's avatar
Rayyyyy committed
140
1. 安装`llama-recipes``lm-eval`
Rayyyyy's avatar
Rayyyyy committed
141
142
143
144
145
146
147
148
```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
149
150
# 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
151
152
153

# 返回根目录
cd ~
Rayyyyy's avatar
Rayyyyy committed
154

Rayyyyy's avatar
Rayyyyy committed
155
156
157
158
159
160
161
# lm-eval 下载
git clone http://developer.hpccube.com/codes/chenych/lm-evaluation-harness.git
cd ./lm-evaluation-harness/
pip install -e .
cd ../
```

Rayyyyy's avatar
Rayyyyy committed
162
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:
Rayyyyy's avatar
Rayyyyy committed
163
```bash
Rayyyyy's avatar
Rayyyyy committed
164
cd /path_of/llama-recipes/recipes/evaluation
Rayyyyy's avatar
Rayyyyy committed
165
166
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
167
168
169
<div align=center>
    <img src="./doc/evaluation.png"/>
</div>
Rayyyyy's avatar
Rayyyyy committed
170

Rayyyyy's avatar
Rayyyyy committed
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
## 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
200
201
202
203
204
205
206

- 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
207
208
209
- Meta-Llama-3-8B-Instruct 模型
```bash
mkdir Meta-Llama-3-8B-Instruct
Rayyyyy's avatar
Rayyyyy committed
210
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
211
```
Rayyyyy's avatar
Rayyyyy committed
212

Rayyyyy's avatar
Rayyyyy committed
213
214
215
216
217
218
219
220
221
222
223
224
- 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
225
226
模型目录结构如下:
```bash
Rayyyyy's avatar
Rayyyyy committed
227
├── model_save_path
Rayyyyy's avatar
Rayyyyy committed
228
│   ├── Meta-Llama-3-8B
Rayyyyy's avatar
Rayyyyy committed
229
│       ├── original
Rayyyyy's avatar
Rayyyyy committed
230
231
232
│           ├── consolidated.00.pth
│           ├── params.json
│           └── tokenizer.model
Rayyyyy's avatar
Rayyyyy committed
233
234
235
236
237
238
239
240
241
242
243
244
245
246
│       ├── config.json
│       ├── configuration.json
│       ├── generation_config.json
│       ├── LICENSE
│       ├── model-00001-of-00004.safetensors
│       ├── model-00002-of-00004.safetensors
│       ├── model-00003-of-00004.safetensors
│       ├── model-00004-of-00004.safetensors
│       ├── model.safetensors.index.json
│       ├── README.md
│       ├── special_tokens_map.json
│       ├── tokenizer_config.json
│       ├── tokenizer.json
│       └── USE_POLICY.md
Rayyyyy's avatar
Rayyyyy committed
247
│   ├── Meta-Llama-3-8B-Instruct
Rayyyyy's avatar
Rayyyyy committed
248
│       ├── original
Rayyyyy's avatar
Rayyyyy committed
249
250
251
│           ├── consolidated.00.pth
│           ├── params.json
│           └── tokenizer.model
Rayyyyy's avatar
Rayyyyy committed
252
253
254
255
256
257
258
259
260
261
262
263
264
265
│       ├── config.json
│       ├── configuration.json
│       ├── generation_config.json
│       ├── LICENSE
│       ├── model-00001-of-00004.safetensors
│       ├── model-00002-of-00004.safetensors
│       ├── model-00003-of-00004.safetensors
│       ├── model-00004-of-00004.safetensors
│       ├── model.safetensors.index.json
│       ├── README.md
│       ├── special_tokens_map.json
│       ├── tokenizer_config.json
│       ├── tokenizer.json
│       └── USE_POLICY.md
Rayyyyy's avatar
Rayyyyy committed
266
│   ├── Meta-Llama-3-70B
Rayyyyy's avatar
Rayyyyy committed
267
│       ├── original
Rayyyyy's avatar
Rayyyyy committed
268
│           ├── consolidated.00.pth
Rayyyyy's avatar
Rayyyyy committed
269
│           ...
Rayyyyy's avatar
Rayyyyy committed
270
271
272
│           ├── consolidated.07.pth
│           ├── params.json
│           └── tokenizer.model
Rayyyyy's avatar
Rayyyyy committed
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
│       ├── config.json
│       ├── generation_config.json
│       ├── LICENSE
│       ├── README.md
│       ├── model-00001-of-00030.safetensors
│       ├── model-00002-of-00030.safetensors
│       ...
│       ├── model-000029-of-00030.safetensors
│       ├── model-000030-of-00030.safetensors
│       ├── model.safetensors.index.json
│       ├── tokenizer_config.json
│       ├── tokenizer.json
│       ├── tokenizer_config.json
│       ├── special_tokens_map.json
│       └── USE_POLICY.md
Rayyyyy's avatar
Rayyyyy committed
288
│   └── Meta-Llama-3-70B-Instruct
Rayyyyy's avatar
Rayyyyy committed
289
│       ├── original
Rayyyyy's avatar
Rayyyyy committed
290
│           ├── consolidated.00.pth
Rayyyyy's avatar
Rayyyyy committed
291
│           ...
Rayyyyy's avatar
Rayyyyy committed
292
│           ├── consolidated.07.pth
Rayyyyy's avatar
Rayyyyy committed
293
294
│           ├── params.json
│           └── tokenizer.model
Rayyyyy's avatar
Rayyyyy committed
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
│       ├── config.json
│       ├── generation_config.json
│       ├── LICENSE
│       ├── README.md
│       ├── model-00001-of-00030.safetensors
│       ├── model-00002-of-00030.safetensors
│       ...
│       ├── model-000029-of-00030.safetensors
│       ├── model-000030-of-00030.safetensors
│       ├── model.safetensors.index.json
│       ├── tokenizer_config.json
│       ├── tokenizer.json
│       ├── tokenizer_config.json
│       ├── special_tokens_map.json
│       └── USE_POLICY.md
Rayyyyy's avatar
Rayyyyy committed
310
311
```

Rayyyyy's avatar
Rayyyyy committed
312
313
314
315
316
## 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
317
- https://github.com/InternLM/xtuner
Rayyyyy's avatar
Rayyyyy committed
318
- https://github.com/meta-llama/llama-recipes