Skip to content

Latest commit

 

History

History
283 lines (222 loc) · 11 KB

README.md

File metadata and controls

283 lines (222 loc) · 11 KB
Raw

OneAPI

The LLM calling tool is designed for researchers to interact with the language model. It can be accessed through a user interface (UI), code, or command.

You can engage in multi-turn conversations with ChatGPT or other LLMs APIs and automatically save them in a training-specific data format.

Step 1: Installation (requires Python environment and python >= 3.11): pip install one-api-tool

Step 2: Start the command: one-api

Step 3: Select the API type and set the key or other information following the guide.

Alt text

Step 4: Initiate the chat dialogue and start a conversation:

Alt text

Step 5: Edit the conversation without restart:

  • : + clear to clear the conversation history
  • : + d_clear to clear the conversation history and system prompt
  • : + undo to remove the latest message
  • : + save to save the current session history (the program also saves automatically upon exit)
  • : + load to load the last conversation from cache file. Specify a numerical index to load a specific conversation, e.g., :load -2.
  • : + system to set the system prompt

Alt text

The currently supported APIs include:

  • OpenAI Official API.

    • ChatGPT: GPT-3.5-turbo/GPT-4.
    • Token number counting.
    • Embedding generation.

  • Microsoft Azure OpenAI Resource endpoint API.

    • ChatGPT: GPT-3.5-turbo/GPT-4.
    • Token number counting.
    • Embedding generation.

  • Anthropic Claude series model API.

    • Claude-v1.3-100k, etc.
    • Token number counting.

  • Huggingface LLMs.

  • vLLM deployed Inference Endpoint.

    Note: If you deploy vLLM with OpenAI-Compatible API as follows, you should set your api config as api_type="openai" and api_base="http://ip:port/v1" and api_key="EMPTY.

    Deploy vLLM with OpenAI-Compatible API:

    python -m vllm.entrypoints.openai.api_server \
--model /path_to_the_model/facebook/opt-125m \
--chat-template ./examples/template_chatml.jinja

Installation

Requirements Python >=3.9

pip install -U one-api-tool

Usage

1. With python.

OpenAI config:

{
    "api_key": "YOUR_API_KEY",
    "api_base": "https://api.openai.com/v1",
    "api_type": "openai"
}

Azure OpenAI config:

{
    "api_key": "YOUR_API_KEY",
    "api_base": "Replace with your Azure OpenAI resource's endpoint value.",
    "api_type": "azure",
    "api_version": "2023-03-15-preview" 
}

Anthropic config:

{
    "api_key": "YOUR_API_KEY",
    "api_base": "https://api.anthropic.com",
    "api_type": "anthropic"
}

Huggingface config:

{
    "api_key": "",
    "api_base": "http://ip:port",
    "api_type": "huggingface",
    "chat_template": "your_jinja2_template"
}

VLLM config:

{
    "api_key": "",
    "api_base": "http://ip:port/generate",
    "api_type": "vllm",
    "chat_template": "your_jinja2_template"
}

api_key: Obtain OpenAI API key from the OpenAI website and Claude API key from the Anthropic website.

api_base: This is the base API that is used to send requests. You can also specify a proxy URL, such as "https://your_proxy_domain/v1". For example, you can use Cloudflare workers to proxy the OpenAI site.

If you are using Azure APIs, you can find relevant information on the Azure resource dashboard. The API format typically follows this pattern: https://{your_organization}.openai.azure.com/.

api_type: Currently supported values are "open_ai", "azure", or "anthropic".

api_version: This field is optional. Azure provides several versions of APIs, such as "2023-03-15-preview". However, the OpenAI SDK always has a default value set for this field. Therefore, you should only specify a specific value if you want to use that particular version of APIs.

chat_template: This field is optional. When using a local endpoint server, you can pass a Jinja2 template specifically designed for that model, mathing the template during training. The template render function takes prompt and system as parameters: template.render(prompt=prompt, system=system). The default template differs based on the type of prompt. For string input, it is as follows:

DEFAULT_STR_TEMP_SYSTEM_USER_ASSISTANT = """{% if system != '' %}{{'<s>System:\n'+system+'\n\nHuman\n'+prompt+'\n\nAssistant:\n'}}{% else %}{{'<s>Human:\n'+prompt+'\n\nAssistant:\n'}}{% endif %}"""

For a list of dictionary messages in OpenAI format, the default template is:

DEFAULT_LIST_MSG_TEMP_SYSTEM_USER_ASSISTANT = """{% for message in prompt %}{% if loop.first %}{% if message['role'] == 'user' %}{% if loop.length != 1 %}{{ '<s>Human:\n' + message['content'] }}{% else %}{{ '<s>Human:\n' + message['content'] + '\n\nAssistant:\n' }}{% endif %}{% elif message['role'] == 'system' %}{{ '<s>System:\n' + message['content'] }}{% endif %}{% elif message['role'] == 'user' %}{% if loop.last %}{{ '\n\nHuman:\n' + message['content'] + '\n\nAssistant:\n'}}{% else %}{{ '\n\nHuman:\n' + message['content']}}{% endif %}{% elif message['role'] == 'assistant' %}{{ '\n\nAssistant:\n' + message['content'] }}{% endif %}{% endfor %}"""

Chat example:

There are three acceptable types of inputs for function chat():

  • list of dicts
  • string
  • list of string
from oneapi import OneLLM
import asyncio
# Two ways to initialize the OneAPITool object  
# llm = OneAPITool.from_config(api_key=api_key, api_base=api_base, api_type=api_type)
llm = OneLLM.from_config("your_config_file.json")

# There are three acceptable types of inputs.
conversations_openai_style = [{"role": "user", "content": "hello"}, {"role": "assistant", "content": "Hello, how can i assistant you today?"}, {"role": "user", "content": "I want to know the weather of tomorrow"}]
conversation_with_system_msg = [{"role": "system", "content": "Now you are a weather forecast assistant."},{"role": "user", "content": "hello"}, {"role": "assistant", "content": "Hello, how can i assistant you today?"}, {"role": "user", "content": "I want to know the weather of tomorrow"}]
string_message = "Hello AI!"
list_of_string_messages = ["Hello AI!", "Hello, how can i assistant you today?", "I want to know the weather of tomorrow"]

for msg in [conversations_sharegpt_style, conversations_openai_style, conversation_with_system_msg, string_message, list_of_string_messages]:
    res = llm(msg)
    print(res)

# Pass system message independently
res = llm("Hello AI!", system="Now you are a helpful assistant.")
print(res)

#Set `vebose=True` to print the detail of args passing to LLMs
res = llm("Hello AI!", verbose=True) 

# Async chat 
res = asyncio.run(llm.achat("How\'s the weather today?", model="gpt-4", stream=False))
print(res)

# Get embeddings of some sentences for further usage, e.g., clustering
embeddings = llm.get_embeddings(["Hello AI!", "Hello world!"])
print(len(embeddings))

# Count the number of tokens
print(llm.count_tokens(["Hello AI!", "Hello world!"]))

Note: Currently, the get_embeddings function only supports OpenAI, Microsoft Azure API, and locally deployed Inference Endpoint using Text Embeddings Inference (Which is really fast).

Batch request with asyncio, python >= 3.11 required.

from oneapi import batch_chat
import asyncio
import time

configs = [
    {"api_key": "EMPTY", "api_base": "http://0.0.0.0:8000/v1", "api_type": "openai"},
    {"api_key": "EMPTY", "api_base": "http://0.0.0.0:8000", "api_type": "vllm"},
    {"api_key": "", "api_base": "http://0.0.0.0:8999", "api_type": "huggingface"},
    {"api_key": "", "api_base": "http://0.0.0.0:8998", "api_type": "huggingface"},
    "azure_openai_config.json",
    "anthropic_config.json",
    "openapi_config.json",
]
prompts = [
    "What is the capital of France?",
    "What is the capital of Germany?",
    "What is the capital of Italy?",
    "What is the capital of Spain?",
    "What is the capital of Portugal?",
    "What is the capital of Belgium?",
    "What is the capital of Netherlands?",
    "What is the capital of Luxembourg?",
    "What is the capital of Denmark?",
    "What is the capital of Sweden?",
    "What is the capital of Norway?",
    "What is the capital of Finland?",
    "Tell me a joke about Trump.",
    "Tell me a joke about Biden.",
    "Tell me a joke about Obama.",
    "Tell me a joke about Bush.",
    "Tell me a joke about Clinton.",
    "Tell me a joke about Reagan.",
    "Tell me a joke about Carter.",
    "Tell me a joke about Soviet Union.",
    [{"role": "user", "content": "Where is Mars?"}]
]
tic = time.time()
models = ["path_to_vllm_deployed_model", "path_to_vllm_deployed_model", "", "", "gpt-4", "claude-v2.1", "gpt-3.5-turbo"]
res = asyncio.run(
    batch_chat(configs, prompts, models, request_interval=0.01, min_process_num=200)
)
print(len(res))
print("Time: ", time.time() - tic)

2. Using command line

Interactive

one-api

Non-interactive

open-api --config_file CHANGE_TO_YOUR_CONFIG_PATH \
--model gpt-3.5-turbo \
--prompt "1+1=?"
Output detail 
-------------------- prompt detail 🚀  --------------------

1+1=?

-------------------- gpt-3.5-turbo response ⭐️ --------------------

2

Arguments detail:

--config_file string ${\color{orange}\text{Required}}$
A local configuration file containing API key information.

--prompt string ${\color{orange}\text{Required}}$
The question that would be predicted by LLMs, e.g., A math question would be like: "1+1=?".

--system string ${\color{grey}\text{Optional}}$ Defaults to null
System message to instruct chatGPT, e.g., You are a helpful assistant.

--model string ${\color{grey}\text{Optional}}$ Defaults to GPT-3.5-turbo or Claude-v1.3 depends on api_type
Which model to use, e.g., gpt-4.

--temperature int ${\color{grey}\text{Optional}}$ Defaults to 1
What sampling temperature to use. Higher values like 0.9 will make the output more random, while lower values like 0.1 will make it more focused and deterministic.

--max_tokens int ${\color{grey}\text{Optional}}$ Defaults to 2048
The maximum number of tokens to generate in the chat completion. The total length of input tokens and generated tokens is limited by the model's context length.

--save_to_file bool ${\color{grey}\text{Optional}}$ Defaults to True
Save the prompt and response to local file at directory "~/.cache/history_cache_{date_of_month}" with the format style of shareGPT.

ToDo

  • Batch requests.
  • Token number counting.
  • Async requests.
  • Custom LLMs.
  • Custom token budget.
  • Using tools.

Architecture

img