Skip to main content

Split traffic betwen GPT-4 and Llama2 in Production!

In this tutorial, we'll walk through A/B testing between GPT-4 and Llama2 in production. We'll assume you've deployed Llama2 on Huggingface Inference Endpoints (but any of TogetherAI, Baseten, Ollama, Petals, Openrouter should work as well).

Relevant Resources:

Code Walkthrough

In production, we don't know if Llama2 is going to provide:

  • good results
  • quickly

💡 Route 20% traffic to Llama2​

If Llama2 returns poor answers / is extremely slow, we want to roll-back this change, and use GPT-4 instead.

Instead of routing 100% of our traffic to Llama2, let's start by routing 20% traffic to it and see how it does. 

## route 20% of responses to Llama2
split_per_model = {
    "gpt-4": 0.8, 
    "huggingface/https://my-unique-endpoint.us-east-1.aws.endpoints.huggingface.cloud": 0.2
}

👨‍💻 Complete Code​

a) For Local​

If we're testing this in a script - this is what our complete code looks like.

from litellm import completion_with_split_tests
import os 

## set ENV variables
os.environ["OPENAI_API_KEY"] = "openai key"
os.environ["HUGGINGFACE_API_KEY"] = "huggingface key"

## route 20% of responses to Llama2
split_per_model = {
    "gpt-4": 0.8, 
    "huggingface/https://my-unique-endpoint.us-east-1.aws.endpoints.huggingface.cloud": 0.2
}

messages = [{ "content": "Hello, how are you?","role": "user"}]

completion_with_split_tests(
  models=split_per_model, 
  messages=messages, 
)

b) For Production​

If we're in production, we don't want to keep going to code to change model/test details (prompt, split%, etc.) for our completion function and redeploying changes.

LiteLLM exposes a client dashboard to do this in a UI - and instantly updates our completion function in prod.

Relevant Code​

completion_with_split_tests(..., use_client=True, id="my-unique-id")

Complete Code​

from litellm import completion_with_split_tests
import os 

## set ENV variables
os.environ["OPENAI_API_KEY"] = "openai key"
os.environ["HUGGINGFACE_API_KEY"] = "huggingface key"

## route 20% of responses to Llama2
split_per_model = {
    "gpt-4": 0.8, 
    "huggingface/https://my-unique-endpoint.us-east-1.aws.endpoints.huggingface.cloud": 0.2
}

messages = [{ "content": "Hello, how are you?","role": "user"}]

completion_with_split_tests(
  models=split_per_model, 
  messages=messages, 
  use_client=True, 
  id="my-unique-id" # Auto-create this @ https://admin.litellm.ai/
)