Skip to content

wangyuxinwhy/generate

BranchesTags

Repository files navigation

Generate Logo

Generate

A Python Package to Access World-Class Generative Models.

中文文档 ｜ 交互式教程

Python Version CI Status CD Status License Documentation Made with Love



Important

generate v0.5.0 版本的设计思路发生较大变化。由于个人精力有限，generate 不再追求支持更多的平台，而是更多的围绕模型的辅助功能进行开发。另外，国内大部分平台已经提供了适配 OpenAI SDK 的 API，如果你只需要基础的文本生成功能，建议直接使用 OpenAI SDK。

简介

Generate 允许用户通过统一的 api 访问多平台的生成式模型，当前支持：

平台 🤖 同步 🔄 异步 ⏳ 流式 🌊 Vision 👀 Tools 🛠️
OpenAI ✅ ✅ ✅ ✅ ✅
Azure ✅ ✅ ❌ ✅ ✅
Anthropic ✅ ✅ ✅ ✅ ❌
文心 Wenxin ✅ ✅ ✅ ❌ ✅
灵积/百炼 DashScope ✅ ✅ ✅ ✅ ✅
百川智能 Baichuan ✅ ✅ ✅ ❌ ✅
Minimax ✅ ✅ ✅ ❌ ✅
混元 Hunyuan ✅ ✅ ✅ ❌ ❌
智谱 Zhipu ✅ ✅ ✅ ✅ ✅
月之暗面 Moonshot ✅ ✅ ✅ ❌ ✅
DeepSeek ✅ ✅ ✅ ❌ ❌
零一万物 Yi ✅ ✅ ✅ ✅ ❌
阶跃星辰 StepFun ✅ ✅ ✅ ✅ ❌

v0.5.0-beta 版本中，混元，文心尚未适配

Features

  • 多模态，支持文本生成，多模态文本生成，结构体生成，图像生成，语音生成...
  • 跨平台，支持 OpenAI，Azure，Minimax，智谱，月之暗面，文心一言 在内的国内外 10+ 平台
  • One API，统一了不同平台的消息格式，推理参数，接口封装，返回解析，让用户无需关心不同平台的差异
  • 异步，流式和并发，提供流式调用，非流式调用，同步调用，异步调用，异步批量并发调用，适配不同的应用场景
  • 自带电池，提供 chainlit UI，输入检查，参数检查，计费，速率控制，Disk Cache，Agent, Tool call 等
  • 轻量，最小化依赖，不同平台的请求和鉴权逻辑均为原生内置功能
  • 高质量代码，100% typehints，pylance strict, ruff lint & format, test coverage > 85% ...

基础使用

Open In Colab

安装

pip install generate-core

查看模型列表

from generate.chat_completion import ChatModelRegistry

print('\n'.join([model_cls.__name__ for model_cls, _ in ChatModelRegistry.values()]))

# ----- Output -----
AzureChat
AnthropicChat
OpenAIChat
MinimaxProChat
MinimaxChat
ZhipuChat
ZhipuCharacterChat
WenxinChat
HunyuanChat
BaichuanChat
BailianChat
DashScopeChat
DashScopeMultiModalChat
MoonshotChat
DeepSeekChat
YiChat

配置模型 API

from generate import WenxinChat

# 获取如何配置文心一言，其他模型同理
print(WenxinChat.how_to_settings())

# ----- Output -----
WenxinChat Settings

# Platform
Qianfan

# Required Environment Variables
['QIANFAN_API_KEY', 'QIANFAN_SECRET_KEY']

# Optional Environment Variables
['QIANFAN_PLATFORM_URL', 'QIANFAN_COMLPETION_API_BASE', 'QIANFAN_IMAGE_GENERATION_API_BASE', 'QIANFAN_ACCESS_TOKEN_API']

You can get more information from this link: https://cloud.baidu.com/doc/WENXINWORKSHOP/s/Dlkm79mnx

tips: You can also set these variables in the .env file, and generate will automatically load them.

对话补全模型

文本生成

from generate import OpenAIChat

model = OpenAIChat()
model.generate('你好，GPT！', temperature=0, seed=2023)

# ----- Output -----
ChatCompletionOutput(
    model_info=ModelInfo(task='chat_completion', type='openai', name='gpt-3.5-turbo-0613'),
    cost=0.000343,
    extra={'usage': {'prompt_tokens': 13, 'completion_tokens': 18, 'total_tokens': 31}},
    message=AssistantMessage(
        role='assistant',
        name=None,
        content='你好！有什么我可以帮助你的吗？',
        function_call=None,
        tool_calls=None
    ),
    finish_reason='stop'
)

多模态文本生成

from generate import OpenAIChat

model = OpenAIChat(model='gpt-4-vision-preview')
user_message = {
    'role': 'user',
    'content': [
        {'text': '这个图片是哪里？'},
        {'image_url': {'url': 'https://dashscope.oss-cn-beijing.aliyuncs.com/images/dog_and_girl.jpeg'}},
    ],
}
model.generate(user_message, max_tokens=1000)

# ----- Output -----
ChatCompletionOutput(
    model_info=ModelInfo(task='chat_completion', type='openai', name='gpt-4-1106-vision-preview'),
    cost=0.10339000000000001,
    extra={'usage': {'prompt_tokens': 1120, 'completion_tokens': 119, 'total_tokens': 1239}},
    message=AssistantMessage(
        role='assistant',
        name=None,
        content='这张图片显示的是一名女士和一只狗在沙滩上。他们似乎在享受日落时分的宁静时刻',
        function_call=None,
        tool_calls=None
    ),
    finish_reason='stop'
)

派生功能

结构体生成

from generate import OpenAIChat
from pydantic import BaseModel

class Country(BaseModel):
    name: str
    capital: str

model = OpenAIChat().structure(output_structure_type=Country)
model.generate('Paris is the capital of France and also the largest city in the country.')
# ----- Output -----
StructureModelOutput(
    model_info=ModelInfo(task='chat_completion', type='openai', name='gpt-3.5-turbo-0613'),
    cost=0.000693,
    extra={'usage': {'prompt_tokens': 75, 'completion_tokens': 12, 'total_tokens': 87}},
    structure=Country(name='France', capital='Paris')
)

速率限制

import time
from generate import OpenAIChat

# 限制速率，每 10 秒最多 4 次请求
limit_model = OpenAIChat().limit(max_generates_per_time_window=2, num_seconds_in_time_window=10)
start_time = time.time()
for i in limit_model.batch_generate([f'1 + {i} = ?' for i in range(4)]):
    print(i.reply)
    print(f'elapsed time: {time.time() - start_time:.2f} seconds')

# ----- Output -----
1
elapsed time: 0.70 seconds
2
elapsed time: 1.34 seconds
3
elapsed time: 11.47 seconds
4
elapsed time: 12.15 seconds

对话历史保持

from generate import OpenAIChat

session_model = OpenAIChat().session()
session_model.generate('i am bob')
print(session_model.generate('What is my name?').reply)

# ----- Output -----
Your name is Bob.

工具调用

from generate import OpenAIChat, tool

@tool
def get_weather(location: str) -> str:
    return f'{location}, 27°C, Sunny'

agent = OpenAIChat().agent(tools=get_weather)
print(agent.generate('what is the weather in Beijing?').reply)

# ----- Output -----
The weather in Beijing is currently 27°C and sunny.

图像生成模型

from generate import OpenAIImageGeneration

model = OpenAIImageGeneration()
model.generate('black hole')

# ----- Output -----
ImageGenerationOutput(
    model_info=ModelInfo(task='image_generation', type='openai', name='dall-e-3'),
    cost=0.56,
    extra={},
    images=[
        GeneratedImage(
            url='https://oaidalleapiprodscus.blob.core.windows.net/...',
            prompt='Visualize an astronomical illustration featuring a black hole at its core. The black hole
should be portrayed with strong gravitational lensing effect that distorts the light around it. Include a
surrounding accretion disk, glowing brightly with blue and white hues, streaked with shades of red and orange,
indicating heat and intense energy. The cosmos in the background should be filled with distant stars, galaxies, and
nebulas, illuminating the vast, infinite space with specks of light.',
            image_format='png',
            content=b'<image bytes>'
        )
    ]
)

语音生成模型

from generate import MinimaxSpeech

model = MinimaxSpeech()
model.generate('你好，世界！')

# ----- Output -----
TextToSpeechOutput(
    model_info=ModelInfo(task='text_to_speech', type='minimax', name='speech-01'),
    cost=0.01,
    extra={},
    audio=b'<audio bytes>',
    audio_format='mp3'
)

多种调用方式

from generate import OpenAIChat

model = OpenAIChat()
for stream_output in model.stream_generate('介绍一下唐朝'):
    print(stream_output.stream.delta, end='', flush=True)

# 同步调用，model.generate
# 异步调用，model.async_generate
# 流式调用，model.stream_generate
# 异步流式调用，model.async_stream_generate
# 批量调研，model.batch_generate
# 异步批量调用，model.async_batch_generate

启动 chainlit UI

python -m generate.ui
# help
# python -m generate.ui --help

About

A Python Package to Access World-Class Generative Models

wangyuxinwhy.github.io/generate/

Topics

ai openai minimax litellm

Resources

Readme

License

MIT license
Activity

Stars

126 stars

Watchers

1 watching

Forks

15 forks
Report repository

Releases 18

v0.5.0-beta Latest
May 23, 2024
+ 17 releases

Packages

No packages published

Contributors 5

Languages