README.md 6.38 KB
Newer Older
chenych's avatar
chenych committed
1
2
3
4
5
# DeepSeek-V3.2-Exp
## 论文
[DeepSeek_V3.2](./DeepSeek_V3_2.pdf)

## 模型结构
chenych's avatar
chenych committed
6
7
DeepSeek-V3.2-Exp模型是一个实验版本,作为迈向下一代架构的中间步骤,V3.2-Exp 在 V3.1-Terminus 的基础上引入了 DeepSeek 稀疏注意力机制--一种旨在探索和验证在长上下文场景中训练和推理效率优化的稀疏注意力机制。
这个实验版本代表了deepseek团队对更高效变压器架构的持续研究,特别关注在处理扩展文本序列时提高计算效率。
chenych's avatar
chenych committed
8

chenych's avatar
chenych committed
9
10
11
<div align=center>
    <img src="./doc/arch.png"/>
</div>
chenych's avatar
chenych committed
12
13

## 算法原理
chenych's avatar
chenych committed
14
DeepSeek 稀疏注意力机制(DSA)首次实现了细粒度的稀疏注意力,在保持几乎相同的模型输出质量的同时,显著提高了长上下文训练和推理效率。
chenych's avatar
chenych committed
15
16
17
18
19
20
21
22
23

## 环境配置
### 硬件需求
DCU型号:K100AI,节点数量:4台,卡数:32 张。

`-v 路径``docker_name``imageID`根据实际情况修改

### Docker(方法一)
```bash
chenych's avatar
chenych committed
24
docker pull image.sourcefind.cn:5000/dcu/admin/base/vllm:0.9.2-ubuntu22.04-dtk25.04.1-rc5-rocblas104381-0915-das1.6-py3.10-20250916-rc2-ds3.2
chenych's avatar
chenych committed
25
26
docker run -it --shm-size 200g --network=host --name {docker_name} --privileged --device=/dev/kfd --device=/dev/dri --device=/dev/mkfd --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -u root -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro {imageID} bash

chenych's avatar
chenych committed
27
cd /your_code_path/deepseek-v3.2-exp_vllm
chenych's avatar
chenych committed
28
29
30
31
32
33
34
35
36
```

### Dockerfile(方法二)
```bash
cd docker
docker build --no-cache -t deepseek-v3.2-exp:latest .

docker run -it --shm-size 200g --network=host --name {docker_name} --privileged --device=/dev/kfd --device=/dev/dri --device=/dev/mkfd --group-add video --cap-add=SYS_PTRACE --security-opt seccomp=unconfined -u root -v /path/your_code_data/:/path/your_code_data/ -v /opt/hyhal/:/opt/hyhal/:ro {imageID} bash

chenych's avatar
chenych committed
37
cd /your_code_path/deepseek-v3.2-exp_vllm
chenych's avatar
chenych committed
38
39
40
41
42
43
44
45
```

### 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.sourcefind.cn/tool/)开发者社区下载安装。
```bash
DTK: 25.04.1
python: 3.10.12
torch: 2.5.1+das.opt1.dtk25041
chenych's avatar
chenych committed
46
transformers: 4.56.1
chenych's avatar
chenych committed
47
```
chenych's avatar
chenych committed
48
`Tips:以上dtk驱动、pytorch等DCU相关工具版本需要严格一一对应`,其他包安装如下:
chenych's avatar
chenych committed
49
```bash
chenych's avatar
chenych committed
50
51
wget http://112.11.119.99:18000/temp/vllm-0.9.2%2Bdas.opt1.rc2.51af08a.dtk25041-cp310-cp310-linux_x86_64.whl
pip install vllm-0.9.2+das.opt1.rc2.51af08a.dtk25041-cp310-cp310-linux_x86_64.whl
chenych's avatar
chenych committed
52
```
chenych's avatar
chenych committed
53
54
55
56
57
58
59
60

## 数据集


## 训练
暂无

## 推理
chenych's avatar
chenych committed
61
62
样例模型:[DeepSeek-V3.2-Exp](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp)

chenych's avatar
chenych committed
63
首先将模型转换成bf16格式,转换命令如下:
chenych's avatar
chenych committed
64
65
```bash
# fp8转bf16
chenych's avatar
chenych committed
66
python inference/fp8_cast_bf16.py --input-fp8-hf-path /path/to/DeepSeek-V3.2-Exp --output-bf16-hf-path /path/to/DeepSeek-V3.2-Exp-bf16
chenych's avatar
chenych committed
67
68
# 拷贝config文件
cp inference/config.json /path/to/DeepSeek-V3.2-Exp-bf16
chenych's avatar
chenych committed
69
```
chenych's avatar
chenych committed
70
转换完成后,将原模型中的 `generation_config.json`, `tokenizer_config.json`, `tokenizer.json`拷贝到`/path/to/DeepSeek-V3.2-Exp-bf16`中。
chenych's avatar
chenych committed
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92

### vllm推理方法
#### server 多机
1. 加入环境变量
> 请注意:
> 每个节点上的环境变量都写到.sh文件中,保存后各个计算节点分别source `.sh` 文件
>
> VLLM_HOST_IP:节点本地通信口ip,尽量选择IB网卡的IP,**避免出现rccl超时问题**
>
> NCCL_SOCKET_IFNAME和GLOO_SOCKET_IFNAME:节点本地通信网口ip对应的名称
>
> 通信口和ip查询方法:ifconfig
>
> IB口状态查询:ibstat  !!!一定要active激活状态才可用,各个节点要保持统一

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

```bash
export ALLREDUCE_STREAM_WITH_COMPUTE=1
export VLLM_HOST_IP=x.x.x.x # 对应计算节点的IP,建议选择IB口SOCKET_IFNAME对应IP地址
chenych's avatar
chenych committed
93
94
export HIP_VISIBLE_DEVICES=0,1,2,3,4,5,6,7
export HSA_FORCE_FINE_GRAIN_PCIE=1
chenych's avatar
chenych committed
95
96
export NCCL_SOCKET_IFNAME=ibxxxx
export GLOO_SOCKET_IFNAME=ibxxxx
chenych's avatar
chenych committed
97
export NCCL_IB_HCA=mlx5_0:1
chenych's avatar
chenych committed
98
unset NCCL_ALGO
chenych's avatar
chenych committed
99
export NCCL_IB_DISABLE=0
chenych's avatar
chenych committed
100
export NCCL_MAX_NCHANNELS=16
chenych's avatar
chenych committed
101
export NCCL_MIN_NCHANNELS=16
chenych's avatar
chenych committed
102
export NCCL_NET_GDR_READ=1
chenych's avatar
chenych committed
103
104
105
106
107
108
109
110
export NCCL_DEBUG=INFO
export NCCL_MIN_P2P_NCHANNELS=16
export NCCL_NCHANNELS_PER_PEER=16
export HIP_USE_GRAPH_QUEUE_POOL=1
export VLLM_ENABLE_MOE_FUSED_GATE=0
export VLLM_ENFORCE_EAGER_BS_THRESHOLD=44
export VLLM_RPC_TIMEOUT=1800000
export VLLM_USE_FLASH_MLA=1
chenych's avatar
chenych committed
111

chenych's avatar
chenych committed
112
# 海光CPU绑定核,intel cpu可不加
chenych's avatar
chenych committed
113
114
115
116
117
118
119
120
121
export VLLM_NUMA_BIND=1
export VLLM_RANK0_NUMA=0
export VLLM_RANK1_NUMA=1
export VLLM_RANK2_NUMA=2
export VLLM_RANK3_NUMA=3
export VLLM_RANK4_NUMA=4
export VLLM_RANK5_NUMA=5
export VLLM_RANK6_NUMA=6
export VLLM_RANK7_NUMA=7
chenych's avatar
chenych committed
122
123

#BW集群需要额外设置的环境变量
chenych's avatar
chenych committed
124
125
export NCCL_NET_GDR_LEVEL=7
export NCCL_SDMA_COPY_ENABLE=0
chenych's avatar
chenych committed
126
```
chenych's avatar
chenych committed
127

chenych's avatar
chenych committed
128
129
130
2. 启动RAY集群
> x.x.x.x 对应第一步 Master节点的 VLLM_HOST_IP

chenych's avatar
chenych committed
131
```bash
chenych's avatar
chenych committed
132
133
134
135
# head节点执行
ray start --head --node-ip-address=x.x.x.x --port=6379 --num-gpus=8 --num-cpus=32
# worker节点执行
ray start --address='x.x.x.x:6379' --num-gpus=8 --num-cpus=32
chenych's avatar
chenych committed
136
137
```

chenych's avatar
chenych committed
138
139
140
3. 启动vllm server
> intel cpu 需要加参数:`--enforce-eager`

chenych's avatar
chenych committed
141
```bash
chenych's avatar
chenych committed
142
vllm serve /path/to/DeepSeek-V3.2-Exp-bf16 \
chenych's avatar
chenych committed
143
144
145
146
    --trust-remote-code \
    --distributed-executor-backend ray \
    --dtype bfloat16 \
    --tensor-parallel-size 32 \
chenych's avatar
chenych committed
147
    --max-model-len 32768 \
chenych's avatar
chenych committed
148
149
    --no-enable-chunked-prefill \
    --no-enable-prefix-caching \
chenych's avatar
chenych committed
150
    --port 8001
chenych's avatar
chenych committed
151
152
```

chenych's avatar
chenych committed
153
154
启动完成后可通过以下方式访问:
```bash
chenych's avatar
chenych committed
155
156
curl http://127.0.0.1:8001/v1/chat/completions   \
    -H "Content-Type: application/json"  \
chenych's avatar
chenych committed
157
    -d '{
chenych's avatar
chenych committed
158
        "model": "/path/to/DeepSeek-V3.2-Exp-bf16",
chenych's avatar
chenych committed
159
160
161
        "messages": [
            {
                "role": "user",
chenych's avatar
chenych committed
162
                "content": "请介绍下你自己。"
chenych's avatar
chenych committed
163
164
            }
        ],
chenych's avatar
chenych committed
165
166
167
168
169
170
        "max_tokens": 1024,
        "temperature": 0.7,
        "chat_template_kwargs": {
            "thinking": false
        }
    }'
chenych's avatar
chenych committed
171
```
chenych's avatar
add  
chenych committed
172

chenych's avatar
chenych committed
173
## result
chenych's avatar
chenych committed
174
<div align=center>
chenych's avatar
chenych committed
175
    <img src="./doc/results_dcu.png"/>
chenych's avatar
chenych committed
176
</div>
chenych's avatar
chenych committed
177
178

### 精度
chenych's avatar
chenych committed
179
DCU与GPU精度一致,推理框架:vllm。
chenych's avatar
chenych committed
180
181
182
183
184
185
186
187
188
189

## 应用场景
### 算法类别
`对话问答`

### 热点应用行业
`制造,金融,教育,广媒`

## 预训练权重
- [DeepSeek-V3.2-Exp](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp)
chenych's avatar
chenych committed
190
- [DeepSeek-V3.2-Exp-Base](https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp-Base)
chenych's avatar
chenych committed
191
192

## 源码仓库及问题反馈
chenych's avatar
chenych committed
193
- https://developer.sourcefind.cn/codes/modelzoo/deepseek-v3.2-exp_vllm
chenych's avatar
chenych committed
194
195
196
197

## 参考资料
- https://huggingface.co/deepseek-ai/DeepSeek-V3.2-Exp
- https://github.com/deepseek-ai/DeepSeek-V3.2-Exp