README.md 9.41 KB
Newer Older
zhuwenwen's avatar
zhuwenwen committed
1
2
3
4
<!--
 * @Author: zhuww
 * @email: zhuww@sugon.com
 * @Date: 2024-04-25 10:38:07
zhuwenwen's avatar
zhuwenwen committed
5
 * @LastEditTime: 2024-12-11 17:18:01
zhuwenwen's avatar
zhuwenwen committed
6
-->
dcuai's avatar
dcuai committed
7
# LLaMA
zhuwenwen's avatar
zhuwenwen committed
8
9
10
11
12
13
14
15
16
17

## 论文
- [https://arxiv.org/pdf/2302.13971.pdf](https://arxiv.org/pdf/2302.13971.pdf)

## 模型结构
LLAMA网络基于 Transformer 架构。提出了各种改进,并用于不同的模型,例如 PaLM。以下是与原始架构的主要区别:
预归一化。为了提高训练稳定性,对每个transformer 子层的输入进行归一化,而不是对输出进行归一化。使用 RMSNorm 归一化函数。
SwiGLU 激活函数 [PaLM]。使用 SwiGLU 激活函数替换 ReLU 非线性以提高性能。使用 2 /3 4d 的维度而不是 PaLM 中的 4d。
旋转嵌入。移除了绝对位置嵌入,而是添加了旋转位置嵌入 (RoPE),在网络的每一层。

zhuwenwen's avatar
zhuwenwen committed
18
![img](./docs/llama_str.png)
zhuwenwen's avatar
zhuwenwen committed
19
20
21
22

## 算法原理
LLama是一个基础语言模型的集合,参数范围从7B到65B。在数万亿的tokens上训练出的模型,并表明可以专门使用公开可用的数据集来训练最先进的模型,而不依赖于专有的和不可访问的数据集。

zhuwenwen's avatar
zhuwenwen committed
23
![img](./docs/llama_pri.png)
zhuwenwen's avatar
zhuwenwen committed
24
25
26

## 环境配置

zhuwenwen's avatar
zhuwenwen committed
27
### Docker(方法一)
zhuwenwen's avatar
zhuwenwen committed
28
提供[光源](https://www.sourcefind.cn/#/image/dcu/custom)拉取推理的docker镜像:
zhuwenwen's avatar
zhuwenwen committed
29

zhuwenwen's avatar
zhuwenwen committed
30
```
31
docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.8.5-ubuntu22.04-dtk25.04.1-rc5-das1.6-py3.10-20250724
zhuwenwen's avatar
zhuwenwen committed
32
33
34
# <Image ID>用上面拉取docker镜像的ID替换
# <Host Path>主机端路径
# <Container Path>容器映射路径
zhuwenwen's avatar
zhuwenwen committed
35
# 若要在主机端和容器端映射端口需要删除--network host参数
zhuwenwen's avatar
zhuwenwen committed
36
docker run -it --name llama_vllm --privileged --shm-size=64G  --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal -v <Host Path>:<Container Path> <Image ID> /bin/bash
zhuwenwen's avatar
zhuwenwen committed
37
```
zhuwenwen's avatar
zhuwenwen committed
38
`Tips:若在K100/Z100L上使用,使用定制镜像docker pull image.sourcefind.cn:5000/dcu/admin/base/custom:vllm0.5.0-dtk24.04.1-ubuntu20.04-py310-zk-v1,K100/Z100L不支持awq量化`
zhuwenwen's avatar
zhuwenwen committed
39

zhuwenwen's avatar
zhuwenwen committed
40
41
42
43
44
45
46
47
48
49
50
51
### Dockerfile(方法二)
```
# <Host Path>主机端路径
# <Container Path>容器映射路径
docker build -t llama:latest .
docker run -it --name llama_vllm --privileged --shm-size=64G  --device=/dev/kfd --device=/dev/dri/ --cap-add=SYS_PTRACE --security-opt seccomp=unconfined --ulimit memlock=-1:-1 --ipc=host --network host --group-add video -v /opt/hyhal:/opt/hyhal -v <Host Path>:<Container Path> llama:latest /bin/bash
```

### Anaconda(方法三)
```
conda create -n llama_vllm python=3.10
```
52
53
54
55
56
57
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。

* DTK驱动:dtk25.04.01
* Pytorch: 2.4.0
* triton: 3.0.0
* lmslim: 0.2.1
zhuwenwen's avatar
zhuwenwen committed
58
* flash_attn: 2.6.1
59
60
* flash_mla: 1.0.0
* vllm: 0.8.5
zhuwenwen's avatar
zhuwenwen committed
61
* python: python3.10
zhuwenwen's avatar
zhuwenwen committed
62

laibao's avatar
laibao committed
63
64
`Tips:需先安装相关依赖,最后安装vllm包`
  
65
66
环境变量:  
export ALLREDUCE_STREAM_WITH_COMPUTE=1  
67
export VLLM_NUMA_BIND=1  
68
69
export VLLM_RANK0_NUMA=0  
export VLLM_RANK1_NUMA=1  
70
71
72
73
export VLLM_RANK2_NUMA=2  
export VLLM_RANK3_NUMA=3  
export VLLM_RANK4_NUMA=4  
export VLLM_RANK5_NUMA=5  
74
75
export VLLM_RANK6_NUMA=6  
export VLLM_RANK7_NUMA=7  
zhuwenwen's avatar
zhuwenwen committed
76
77
78
79
80
81
82
83

## 数据集


## 推理

### 模型下载

zhuwenwen's avatar
zhuwenwen committed
84
| 基座模型 | chat模型 | GPTQ模型 | AWQ模型 |
chenzk's avatar
chenzk committed
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
可从HF下载以下模型进行使用:
Llama-2-7b-hf
Llama-2-7b-chat-hf
Llama-2-7B-Chat-GPTQ
Llama-2-7B-AWQ
Llama-2-13b-hf
Llama-2-13b-chat-hf
Llama-2-13B-GPTQ
Llama-2-13B-AWQ
Llama-2-70b-hf
Llama-2-70B-Chat-GPTQ
Llama-2-70B-AWQ
Meta-Llama-3-8B
Meta-Llama-3-8B-Instruct
Meta-Llama-3-8B-Instruct-AWQ
Meta-Llama-3-70B
Meta-Llama-3-70B-Instruct
Meta-Llama-3-70B-Instruct-AWQ
zhuwenwen's avatar
zhuwenwen committed
103

zhuwenwen's avatar
zhuwenwen committed
104
### 离线批量推理
105

zhuwenwen's avatar
zhuwenwen committed
106
```bash
107
VLLM_USE_FLASH_ATTN_PA=1 python examples/offline_inference/basic/basic.py
zhuwenwen's avatar
zhuwenwen committed
108
```
zhuwenwen's avatar
zhuwenwen committed
109
其中,`prompts`为提示词;`temperature`为控制采样随机性的值,值越小模型生成越确定,值变高模型生成更随机,0表示贪婪采样,默认为1;`max_tokens=16`为生成长度,默认为1;
zhuwenwen's avatar
zhuwenwen committed
110
`model`为模型路径;`tensor_parallel_size=1`为使用卡数,默认为1;`dtype="float16"`为推理数据类型,如果模型权重是bfloat16,需要修改为float16推理,`quantization="gptq"`为使用gptq量化进行推理,需下载以上GPTQ模型。`quantization="awq"`为使用awq量化进行推理,需下载以上AWQ模型。
zhuwenwen's avatar
zhuwenwen committed
111

zhuwenwen's avatar
zhuwenwen committed
112

zhuwenwen's avatar
zhuwenwen committed
113
114
115
### 离线批量推理性能测试
1、指定输入输出
```bash
116
VLLM_USE_FLASH_ATTN_PA=1 python benchmarks/benchmark_throughput.py --num-prompts 1 --input-len 32 --output-len 128 --model meta-llama/Llama-2-7b-chat-hf -tp 1 --trust-remote-code --enforce-eager --dtype float16
zhuwenwen's avatar
zhuwenwen committed
117
```
zhuwenwen's avatar
zhuwenwen committed
118
其中`--num-prompts`是batch数,`--input-len`是输入seqlen,`--output-len`是输出token长度,`--model`为模型路径,`-tp`为使用卡数,`dtype="float16"`为推理数据类型,如果模型权重是bfloat16,需要修改为float16推理。若指定`--output-len  1`即为首字延迟。`-q gptq`为使用gptq量化模型进行推理。
zhuwenwen's avatar
zhuwenwen committed
119
120
121

2、使用数据集
下载数据集:
122
[sharegpt_v3_unfiltered_cleaned_split](https://huggingface.co/datasets/learnanything/sharegpt_v3_unfiltered_cleaned_split)
zhuwenwen's avatar
zhuwenwen committed
123
124

```bash
125
VLLM_USE_FLASH_ATTN_PA=1 python benchmarks/benchmark_throughput.py --num-prompts 1 --model meta-llama/Llama-2-7b-chat-hf --dataset-name sharegpt --dataset-path /path/to/ShareGPT_V3_unfiltered_cleaned_split.json -tp 1 --trust-remote-code --enforce-eager --dtype float16
zhuwenwen's avatar
zhuwenwen committed
126
127
```

128
其中`--num-prompts`是batch数,`--model`为模型路径,`--dataset`为使用的数据集,`-tp`为使用卡数,`dtype="float16"`为推理数据类型。若模型权重为 bfloat16,建议设置 `--dtype bfloat16` 或使用 `--dtype auto` 以匹配权重精度。
zhuwenwen's avatar
zhuwenwen committed
129

laibao's avatar
laibao committed
130
### openAI api服务推理性能测试
zhuwenwen's avatar
zhuwenwen committed
131
132
1、启动服务端:
```bash
133
VLLM_USE_FLASH_ATTN_PA=1 vllm serve --model meta-llama/Llama-2-7b-chat-hf --enforce-eager --dtype float16 --trust-remote-code --tensor-parallel-size 1
zhuwenwen's avatar
zhuwenwen committed
134
135
136
137
```

2、启动客户端:
```bash
zhuwenwen's avatar
zhuwenwen committed
138
python benchmarks/benchmark_serving.py --model meta-llama/Llama-2-7b-chat-hf --dataset ShareGPT_V3_unfiltered_cleaned_split.json  --num-prompts 1 --trust-remote-code
zhuwenwen's avatar
zhuwenwen committed
139
```
zhuwenwen's avatar
zhuwenwen committed
140
参数同使用数据集,离线批量推理性能测试,具体参考[benchmarks/benchmark_serving.py](benchmarks/benchmark_serving.py)
zhuwenwen's avatar
zhuwenwen committed
141
142


zhuwenwen's avatar
zhuwenwen committed
143
144
145
### OpenAI兼容服务
启动服务:
```bash
146
VLLM_USE_FLASH_ATTN_PA=1 vllm serve meta-llama/Llama-2-7b-chat-hf --enforce-eager --dtype float16 --trust-remote-code
zhuwenwen's avatar
zhuwenwen committed
147
```
laibao's avatar
laibao committed
148
这里serve之后为加载模型路径,`--dtype`为数据类型:float16,默认情况使用tokenizer中的预定义聊天模板,`--chat-template`可以添加新模板覆盖默认模板,`-q gptq`为使用gptq量化模型进行推理。
zhuwenwen's avatar
zhuwenwen committed
149
150
151
152
153
154
155
156
157
158
159

列出模型型号:
```bash
curl http://localhost:8000/v1/models
```

### OpenAI Completions API和vllm结合使用
```bash
curl http://localhost:8000/v1/completions \
    -H "Content-Type: application/json" \
    -d '{
zhuwenwen's avatar
zhuwenwen committed
160
        "model": "meta-llama/Llama-2-7b-hf",
zhuwenwen's avatar
zhuwenwen committed
161
162
163
164
165
        "prompt": "I believe the meaning of life is",
        "max_tokens": 7,
        "temperature": 0
    }'
```
zhuwenwen's avatar
zhuwenwen committed
166
或者使用[examples/openai_completion_client.py](examples/openai_completion_client.py)
zhuwenwen's avatar
zhuwenwen committed
167
168
169
170
171


### OpenAI Chat API和vllm结合使用
```bash
curl http://localhost:8000/v1/chat/completions \
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
  -X POST \
  -H "Content-Type: application/json" \
  -d '{
    "model": "meta-llama/Llama-2-7b-chat-hf",
    "max_tokens": 128,
    "messages": [
      {
        "role": "user",
        "content": "I believe the meaning of life is"
      }
    ]
  }'
```

或者使用[examples/online_serving/openai_chat_completion_client.py](examples/online_serving/openai_chat_completion_client.py)
laibao's avatar
laibao committed
187
### **gradio和vllm结合使用**
zhuwenwen's avatar
zhuwenwen committed
188

laibao's avatar
laibao committed
189
190
191
192
193
194
195
196
197
198
199
1.安装gradio

```
pip install gradio
```

2.安装必要文件

    2.1 启动gradio服务,根据提示操作

```
200
python examples/online_serving/gradio_openai_chatbot_webserver.py --model "meta-llama/Llama-2-7b-chat-hf" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids ""
laibao's avatar
laibao committed
201
202
203
204
205
206
207
208
209
```

    2.2 更改文件权限

打开提示下载文件目录,输入以下命令给予权限

```
chmod +x frpc_linux_amd64_v0.*
```
laibao's avatar
laibao committed
210
211
212
213
214
   2.3 端口映射

```
ssh -L 8000:计算节点IP:8000 -L 8001:计算节点IP:8001 用户名@登录节点 -p 登录节点端口
```  
laibao's avatar
laibao committed
215
216
217
218

3.启动OpenAI兼容服务

```
219
VLLM_USE_FLASH_ATTN_PA=1 vllm serve meta-llama/Llama-2-7b-chat-hf --enforce-eager --dtype float16 --trust-remote-code --host "0.0.0.0"
laibao's avatar
laibao committed
220
221
222
223
224
```

4.启动gradio服务

```
225
python examples/online_serving/gradio_openai_chatbot_webserver.py --model "meta-llama/Llama-2-7b-chat-hf" --model-url http://localhost:8000/v1 --temp 0.8 --stop-token-ids "" --host "0.0.0.0" --port 8001
laibao's avatar
laibao committed
226
227
228
229
230
```

5.使用对话服务

在浏览器中输入本地 URL,可以使用 Gradio 提供的对话服务。
zhuwenwen's avatar
zhuwenwen committed
231
## result
zhuwenwen's avatar
zhuwenwen committed
232
使用的加速卡:1张 DCU-K100_AI-64G
zhuwenwen's avatar
zhuwenwen committed
233
234
235
236
```
Prompt: 'I believe the meaning of life is', Generated text: ' to find purpose, happiness, and fulfillment. Here are some reasons why:\n\n1. Purpose: Having a sense of purpose gives life meaning and direction. It helps individuals set goals and work towards achieving them, which can lead to a sense of accomplishment and fulfillment.\n2. Happiness: Happiness is a fundamental aspect of life that brings joy and satisfaction.
```

zhuwenwen's avatar
zhuwenwen committed
237
### 精度
zhuwenwen's avatar
zhuwenwen committed
238
239
240
241
242
243
244
245
246
247
248


## 应用场景

### 算法类别
对话问答

### 热点应用行业
金融,科研,教育

## 源码仓库及问题反馈
chenzk's avatar
chenzk committed
249
* [https://developer.sourcefind.cn/codes/modelzoo/llama_vllm](https://developer.sourcefind.cn/codes/modelzoo/llama_vllm)
zhuwenwen's avatar
zhuwenwen committed
250
251
252

## 参考资料
* [https://github.com/vllm-project/vllm](https://github.com/vllm-project/vllm)