forked from jiuyuan/CPM-9G-8B
完整版
This commit is contained in:
p18457032
parent
42e87760f5
commit
c8a6ae1a20
|
|
@ -40,7 +40,6 @@
|
|||
# 九格大模型使用文档
|
||||
## 目录
|
||||
|
||||
- [模型推理](https://www.osredm.com/jiuyuan/CPM-9G-8B/tree/master/quick_start_clean/readmes/README_ALL.md?tab=readme-ov-file#模型推理)
|
||||
<!-- - [仓库目录结构](#仓库目录结构) -->
|
||||
- [九格大模型使用文档](#九格大模型使用文档)
|
||||
- [目录](#目录)
|
||||
|
|
@ -53,6 +52,7 @@
|
|||
- [多机训练](#多机训练)
|
||||
- [参数详细介绍](#参数详细介绍)
|
||||
- [查看训练情况](#查看训练情况)
|
||||
- [模型推理](#模型推理)
|
||||
- [常见问题](#常见问题)
|
||||
|
||||
<!-- ## 仓库目录结构
|
||||
|
|
@ -106,10 +106,11 @@ pip install tensorboard
|
|||
pip install tensorboardX
|
||||
|
||||
|
||||
8.安装vllm
|
||||
我们提供了两种vllm的安装方式
|
||||
请直接安装/quick_start_clean/tools/vllm-0.5.0.dev0+cu122-cp310-cp310-linux_x86_64.whl
|
||||
如果不成功，请通过源码安装vllm，即通过/quick_start_clean/tools/vllm-0.5.0.dev0.tar中的setup.py安装
|
||||
8.安装vllm（模型推理）
|
||||
我们提供了python3.8、python3.10版本的vllm安装包，相关依赖均已封装，可直接安装后执行推理：
|
||||
[vllm-0.5.0.dev0+cu122-cp38-cp38-linux_x86_64.whl](https://qy-obs-6d58.obs.cn-north-4.myhuaweicloud.com/vllm-0.5.0.dev0%2Bcu122-cp38-cp38-linux_x86_64.whl)
|
||||
[vllm-0.5.0.dev0+cu122-cp310-cp310-linux_x86_64.whl](https://qy-obs-6d58.obs.cn-north-4.myhuaweicloud.com/vllm-0.5.0.dev0%2Bcu122-cp310-cp310-linux_x86_64.whl)
|
||||
同时，我们也提供了vllm源码,位于/quick_start_clean/tools/vllm-0.5.0.dev0.tar。
|
||||
```
|
||||
|
||||
## 开源模型
|
||||
|
|
@ -117,7 +118,7 @@ pip install tensorboardX
|
|||
[8b_sft_model_v1](https://qy-obs-6d58.obs.cn-north-4.myhuaweicloud.com/checkpoints-epoch-1.tar.gz), [8b_sft_model_v2](https://qy-obs-6d58.obs.cn-north-4.myhuaweicloud.com/sft_8b_v2.zip)
|
||||
|
||||
2. 端侧2B模型，下载链接：
|
||||
[2b—sft-model] # TODO
|
||||
[2b—sft-model](https://qy-obs-6d58.obs.cn-north-4.myhuaweicloud.com/fm9g_2b_hf_models.tar.gz)
|
||||
|
||||
## 数据处理流程
|
||||
### 单个数据集处理
|
||||
|
|
@ -410,39 +411,104 @@ tensorboard –-logdir /apps/fm9g_2b/data/tensorboard/2b_0701 #存放.events文
|
|||
TypeError: MessageToJson() got an unexpected keyword argument 'including_default_value_fields'
|
||||
```
|
||||
|
||||
<!-- ## 模型推理 TODO:需要补充
|
||||
```python
|
||||
import os
|
||||
from libcpm import CPM9G
|
||||
import argparse, json, os
|
||||
def main():
|
||||
parser = argparse.ArgumentParser()
|
||||
parser.add_argument("--pt", type=str, help="the path of ckpt")
|
||||
parser.add_argument("--config", type=str, help="the path of config file")
|
||||
parser.add_argument("--vocab", type=str, help="the path of vocab file")
|
||||
args = parser.parse_args()
|
||||
model_config = json.load(open(args.config, 'r'))
|
||||
model_config["new_vocab"] = True
|
||||
model = CPM9G(
|
||||
"",
|
||||
args.vocab,
|
||||
-1,
|
||||
memory_limit = 30 << 30,#memory limit 左边的参数根据gpu的显存设置，如果是A100，可以设置为 72 << 30，这样的话就可以用到更多的显存
|
||||
model_config=model_config,
|
||||
load_model=False,
|
||||
)
|
||||
model.load_model_pt(args.pt)
|
||||
datas = [
|
||||
'''<用户>马化腾是谁？<AI>''',
|
||||
'''<用户>你是谁？<AI>''',
|
||||
'''<用户>我要参加一个高性能会议，请帮我写一个致辞。<AI>''',
|
||||
]
|
||||
for data in datas:
|
||||
res = model.inference(data, max_length=4096)
|
||||
print(res['result'])
|
||||
if __name__ == "__main__":
|
||||
main()
|
||||
``` -->
|
||||
## 模型推理
|
||||
模型推理列举了两种推理方法：离线批量推理和部署OpenAI API服务推理
|
||||
|
||||
### 离线批量推理：
|
||||
离线批量推理可参考以下脚本：
|
||||
``` python
|
||||
# offline_inference.py
|
||||
from vllm import LLM, SamplingParams
|
||||
|
||||
# 提示用例
|
||||
prompts = [
|
||||
"Hello, my name is",
|
||||
"The president of the United States is",
|
||||
"The capital of France is",
|
||||
"The future of AI is",
|
||||
]
|
||||
|
||||
# 设置采样参数以控制生成文本，更多参数详细介绍可见/vllm/sampling_params.py
|
||||
# temperature越大，生成结果的随机性越强，top_p过滤掉生成词汇表中概率低于给定阈值的词汇，控制随机性
|
||||
sampling_params = SamplingParams(temperature=0.8, top_p=0.95)
|
||||
|
||||
# 初始化语言模型
|
||||
llm = LLM(model="../models/9G/", trust_remote_code=True)
|
||||
|
||||
# 根据提示生成文本
|
||||
outputs = llm.generate(prompts, sampling_params)
|
||||
|
||||
# 打印输出结果
|
||||
for output in outputs:
|
||||
prompt = output.prompt
|
||||
generated_text = output.outputs[0].text
|
||||
print(f"Prompt: {prompt!r}, Generated text: {generated_text!r}")
|
||||
```
|
||||
|
||||
在初始化语言模型部分，不同模型有所差异：
|
||||
端侧2B模型：
|
||||
``` python
|
||||
# 初始化语言模型，与Hugging Face Transformers库兼容，支持AWQ、GPTQ和GGUF量化格式转换
|
||||
llm = LLM(model="../models/FM9G/", tokenizer_mode="auto", trust_remote_code=True)
|
||||
```
|
||||
8B百亿SFT模型：
|
||||
``` python
|
||||
# 初始化语言模型，tokenizer_mode需指定为"cpm"，不支持AWQ、GPTQ和GGUF量化格式转换
|
||||
# 需要特别注意的是，由于8B模型训练分词方式差异，vllm库中代码有新增，需要按照“环境配置”安装指定版本vllm
|
||||
llm = LLM(model="../models/8b_sft_model/", tokenizer_mode="cpm", trust_remote_code=True)
|
||||
```
|
||||
|
||||
### 部署OpenAI API服务推理
|
||||
vLLM可以为 LLM 服务进行部署，这里提供了一个示例：
|
||||
1. 启动服务：
|
||||
端侧2B模型：
|
||||
```shell
|
||||
python -m vllm.entrypoints.openai.api_server \
|
||||
--model ../models/FM9G/ \
|
||||
--tokenizer-mode auto \
|
||||
--dtype auto \
|
||||
--trust-remote-code \
|
||||
--api-key CPMAPI
|
||||
|
||||
# 与离线批量推理类似，使用端侧2B模型，tokenizer-mode为"auto"
|
||||
# dtype为模型数据类型，设置为"auto"即可
|
||||
# api-key为可选项，可在此处指定你的api密钥
|
||||
```
|
||||
8B百亿SFT模型：
|
||||
```shell
|
||||
python -m vllm.entrypoints.openai.api_server \
|
||||
--model ../models/8b_sft_model/ \
|
||||
--tokenizer-mode cpm \
|
||||
--dtype auto \
|
||||
--api-key CPMAPI
|
||||
|
||||
# 与离线批量推理类似，使用8B百亿SFT模型，tokenizer-mode为"cpm"
|
||||
```
|
||||
|
||||
执行对应指令后，默认在http://localhost:8000地址上启动服务，启动成功后终端会出现如下提示：
|
||||
```shell
|
||||
INFO: Started server process [950965]
|
||||
INFO: Waiting for application startup.
|
||||
INFO: Application startup complete.
|
||||
INFO: Uvicorn running on http://0.0.0.0:8000 (Press CTRL+C to quit)
|
||||
```
|
||||
|
||||
2. 调用API：
|
||||
启动服务端成功后，重新打开一个终端，可参考执行以下python脚本：
|
||||
``` python
|
||||
from openai import OpenAI
|
||||
|
||||
# 如果启动服务时指定了api密钥，需要修改为对应的密钥，否则为"EMPTY"
|
||||
openai_api_key = "CPMAPI"
|
||||
openai_api_base = "http://localhost:8000/v1"
|
||||
client = OpenAI(
|
||||
api_key=openai_api_key,
|
||||
base_url=openai_api_base,
|
||||
)
|
||||
completion = client.completions.create(model="../models/9G/",
|
||||
prompt="San Francisco is a")
|
||||
print("Completion result:", completion)
|
||||
```
|
||||
|
||||
## 常见问题
|
||||
1. Conda安装pytorch时卡在solving environment：网络问题。
|
||||
|
|
|
|||
Loading…
Reference in New Issue