litellm/cookbook/llm-ab-test-server/readme.md

🚅 LiteLLM - A/B Testing LLMs in Production

Call all LLM APIs using the OpenAI format [Anthropic, Huggingface, Cohere, Azure OpenAI etc.]

100+ Supported Models | Docs | Demo Website

LiteLLM allows you to call 100+ LLMs using completion

This template server allows you to define LLMs with their A/B test ratios

llm_dict = {
    "gpt-4": 0.2,
    "together_ai/togethercomputer/llama-2-70b-chat": 0.4,
    "claude-2": 0.2,
    "claude-1.2": 0.2
}

Easily call selected model during completion

selected_llm = random.choices(list(llm_dict.keys()), weights=list(llm_dict.values()))[0]

response = completion(model=selected_model, messages=[{ "content": "Hello, how are you?","role": "user"}])

This server allows you to view responses, costs and latency on your LiteLLM dashboard

Using LiteLLM A/B Testing Server

Setup

Install LiteLLM

pip install litellm

Stable version

pip install litellm==0.1.424

Clone LiteLLM Git Repo

git clone https://github.com/BerriAI/litellm/

Navigate to LiteLLM-A/B Test Server

cd litellm/cookbook/llm-ab-test-server

Run the Server

python3 main.py

Set your LLM Configs

Set your LLMs and LLM weights you want to run A/B testing with In main.py set your selected LLMs you want to AB test in llm_dict You can A/B test more than 100+ LLMs using LiteLLM https://docs.litellm.ai/docs/completion/supported

llm_dict = {
    "gpt-4": 0.2,
    "together_ai/togethercomputer/llama-2-70b-chat": 0.4,
    "claude-2": 0.2,
    "claude-1.2": 0.2
}

Setting your API Keys

Set your LLM API keys in a .env file in the directory or set them as os.environ variables.

See https://docs.litellm.ai/docs/completion/supported for the format of API keys

LiteLLM generalizes api keys to follow the following format PROVIDER_API_KEY

Making Requests to the LiteLLM Server Locally

The server follows the Input/Output format set by the OpenAI Chat Completions API Here is an example request made the LiteLLM Server

Python

import requests
import json

url = "http://localhost:5000/chat/completions"

payload = json.dumps({
  "messages": [
    {
      "content": "who is CTO of litellm",
      "role": "user"
    }
  ]
})
headers = {
  'Content-Type': 'application/json'
}

response = requests.request("POST", url, headers=headers, data=payload)

print(response.text)

Curl Command

curl --location 'http://localhost:5000/chat/completions' \
--header 'Content-Type: application/json' \
--data '{
    "messages": [
                    { 
                        "content": "who is CTO of litellm",
                        "role": "user"
                    }
                ]
    
}
'

Viewing Logs

After running your first completion() call litellm autogenerates a new logs dashboard for you. Link to your Logs dashboard is generated in the terminal / console.

Example Terminal Output with Log Dashboard

support / talk with founders

• Schedule Demo 👋
• Community Discord 💭
• Our numbers 📞 +1 (770) 8783-106 / ‭+1 (412) 618-6238‬
• Our emails ✉️ ishaan@berri.ai / krrish@berri.ai

why did we build this

• Need for simplicity: Our code started to get extremely complicated managing & translating calls between Azure, OpenAI, Cohere