litellm/docs/my-website/docs/simple_proxy.md

import Image from '@theme/IdealImage'; import Tabs from '@theme/Tabs'; import TabItem from '@theme/TabItem';

💥 OpenAI Proxy Server

LiteLLM Server manages:

• Calling 100+ LLMs Huggingface/Bedrock/TogetherAI/etc. in the OpenAI ChatCompletions & Completions format
• Set custom prompt templates + model-specific configs (temperature, max_tokens, etc.)

Quick Start

$ litellm --model huggingface/bigcode/starcoder

#INFO: Proxy running on http://0.0.0.0:8000

Test

In a new shell, run, this will make an openai.ChatCompletion request

litellm --test

This will now automatically route any requests for gpt-3.5-turbo to bigcode starcoder, hosted on huggingface inference endpoints.

Output

{
  "object": "chat.completion",
  "choices": [
    {
      "finish_reason": "length",
      "index": 0,
      "message": {
        "content": ", and create a new test page.\n\n### Test data\n\n- A user named",
        "role": "assistant"
      }
    }
  ],
  "id": "chatcmpl-56634359-d4ce-4dbc-972c-86a640e3a5d8",
  "created": 1699308314.054251,
  "model": "huggingface/bigcode/starcoder",
  "usage": {
    "completion_tokens": 16,
    "prompt_tokens": 10,
    "total_tokens": 26
  }
}

Supported LLMs

$ export AWS_ACCESS_KEY_ID=""
$ export AWS_REGION_NAME="" # e.g. us-west-2
$ export AWS_SECRET_ACCESS_KEY=""

$ litellm --model bedrock/anthropic.claude-v2

$ export HUGGINGFACE_API_KEY=my-api-key #[OPTIONAL]

$ litellm --model huggingface/<huggingface-model-name> --api_base https://<your-hf-endpoint># e.g. huggingface/mistralai/Mistral-7B-v0.1

$ export ANTHROPIC_API_KEY=my-api-key

$ litellm --model claude-instant-1

Assuming you're running vllm locally

$ litellm --model vllm/facebook/opt-125m

$ litellm --model openai/<model_name> --api_base <your-api-base>

$ export TOGETHERAI_API_KEY=my-api-key

$ litellm --model together_ai/lmsys/vicuna-13b-v1.5-16k

$ export REPLICATE_API_KEY=my-api-key

$ litellm \
  --model replicate/meta/llama-2-70b-chat:02e509c789964a7ea8736978a43525956ef40397be9033abf9fd2badfe68c9e3

$ litellm --model petals/meta-llama/Llama-2-70b-chat-hf

$ export PALM_API_KEY=my-palm-key

$ litellm --model palm/chat-bison

$ export AZURE_API_KEY=my-api-key
$ export AZURE_API_BASE=my-api-base

$ litellm --model azure/my-deployment-name

$ export AI21_API_KEY=my-api-key

$ litellm --model j2-light

$ export COHERE_API_KEY=my-api-key

$ litellm --model command-nightly

Server Endpoints

• POST /chat/completions - chat completions endpoint to call 100+ LLMs
• POST /completions - completions endpoint
• POST /embeddings - embedding endpoint for Azure, OpenAI, Huggingface endpoints
• GET /models - available models on server

Advanced

Tutorial: Use with Multiple LLMs + Aider/AutoGen/Langroid/etc.

$ litellm

#INFO: litellm proxy running on http://0.0.0.0:8000

Send a request to your proxy

import openai 

openai.api_key = "any-string-here"
openai.api_base = "http://0.0.0.0:8080" # your proxy url

# call gpt-3.5-turbo
response = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=[{"role": "user", "content": "Hey"}])

print(response)

# call ollama/llama2
response = openai.ChatCompletion.create(model="ollama/llama2", messages=[{"role": "user", "content": "Hey"}])

print(response)

Continue-Dev brings ChatGPT to VSCode. See how to install it here.

In the config.py set this as your default model.

  default=OpenAI(
      api_key="IGNORED",
      model="fake-model-name",
      context_length=2048, # customize if needed for your model
      api_base="http://localhost:8000" # your proxy server url
  ),

Credits @vividfog for this tutorial.

$ pip install aider 

$ aider --openai-api-base http://0.0.0.0:8000 --openai-api-key fake-key

pip install pyautogen

from autogen import AssistantAgent, UserProxyAgent, oai
config_list=[
    {
        "model": "my-fake-model",
        "api_base": "http://localhost:8000",  #litellm compatible endpoint
        "api_type": "open_ai",
        "api_key": "NULL", # just a placeholder
    }
]

response = oai.Completion.create(config_list=config_list, prompt="Hi")
print(response) # works fine

llm_config={
    "config_list": config_list,
}

assistant = AssistantAgent("assistant", llm_config=llm_config)
user_proxy = UserProxyAgent("user_proxy")
user_proxy.initiate_chat(assistant, message="Plot a chart of META and TESLA stock price change YTD.", config_list=config_list)

Credits @victordibia for this tutorial.

from autogen import AssistantAgent, GroupChatManager, UserProxyAgent
from autogen.agentchat import GroupChat
config_list = [
    {
        "model": "ollama/mistralorca",
        "api_base": "http://localhost:8000",  # litellm compatible endpoint
        "api_type": "open_ai",
        "api_key": "NULL",  # just a placeholder
    }
]
llm_config = {"config_list": config_list, "seed": 42}

code_config_list = [
    {
        "model": "ollama/phind-code",
        "api_base": "http://localhost:8000",  # litellm compatible endpoint
        "api_type": "open_ai",
        "api_key": "NULL",  # just a placeholder
    }
]

code_config = {"config_list": code_config_list, "seed": 42}

admin = UserProxyAgent(
    name="Admin",
    system_message="A human admin. Interact with the planner to discuss the plan. Plan execution needs to be approved by this admin.",
    llm_config=llm_config,
    code_execution_config=False,
)


engineer = AssistantAgent(
    name="Engineer",
    llm_config=code_config,
    system_message="""Engineer. You follow an approved plan. You write python/shell code to solve tasks. Wrap the code in a code block that specifies the script type. The user can't modify your code. So do not suggest incomplete code which requires others to modify. Don't use a code block if it's not intended to be executed by the executor.
Don't include multiple code blocks in one response. Do not ask others to copy and paste the result. Check the execution result returned by the executor.
If the result indicates there is an error, fix the error and output the code again. Suggest the full code instead of partial code or code changes. If the error can't be fixed or if the task is not solved even after the code is executed successfully, analyze the problem, revisit your assumption, collect additional info you need, and think of a different approach to try.
""",
)
planner = AssistantAgent(
    name="Planner",
    system_message="""Planner. Suggest a plan. Revise the plan based on feedback from admin and critic, until admin approval.
The plan may involve an engineer who can write code and a scientist who doesn't write code.
Explain the plan first. Be clear which step is performed by an engineer, and which step is performed by a scientist.
""",
    llm_config=llm_config,
)
executor = UserProxyAgent(
    name="Executor",
    system_message="Executor. Execute the code written by the engineer and report the result.",
    human_input_mode="NEVER",
    llm_config=llm_config,
    code_execution_config={"last_n_messages": 3, "work_dir": "paper"},
)
critic = AssistantAgent(
    name="Critic",
    system_message="Critic. Double check plan, claims, code from other agents and provide feedback. Check whether the plan includes adding verifiable info such as source URL.",
    llm_config=llm_config,
)
groupchat = GroupChat(
    agents=[admin, engineer, planner, executor, critic],
    messages=[],
    max_round=50,
)
manager = GroupChatManager(groupchat=groupchat, llm_config=llm_config)


admin.initiate_chat(
    manager,
    message="""
""",
)

Credits @Nathan for this tutorial.

GPT-Pilot helps you build apps with AI Agents. [For more](https://github.com/Pythagora-io/gpt-pilot)

In your .env set the openai endpoint to your local server.

OPENAI_ENDPOINT=http://0.0.0.0:8000
OPENAI_API_KEY=my-fake-key

A guidance language for controlling large language models. https://github.com/guidance-ai/guidance

NOTE: Guidance sends additional params like stop_sequences which can cause some models to fail if they don't support it.

Fix: Start your proxy using the --drop_params flag

litellm --model ollama/codellama --temperature 0.3 --max_tokens 2048 --drop_params

import guidance

# set api_base to your proxy
# set api_key to anything
gpt4 = guidance.llms.OpenAI("gpt-4", api_base="http://0.0.0.0:8000", api_key="anything")

experts = guidance('''
{{#system~}}
You are a helpful and terse assistant.
{{~/system}}

{{#user~}}
I want a response to the following question:
{{query}}
Name 3 world-class experts (past or present) who would be great at answering this?
Don't answer the question yet.
{{~/user}}

{{#assistant~}}
{{gen 'expert_names' temperature=0 max_tokens=300}}
{{~/assistant}}
''', llm=gpt4)

result = experts(query='How can I be more productive?')
print(result)

[TUTORIAL] LM-Evaluation Harness with TGI

Evaluate LLMs 20x faster with TGI via litellm proxy's /completions endpoint.

This tutorial assumes you're using lm-evaluation-harness

Step 1: Start the local proxy

$ litellm --model huggingface/bigcode/starcoder

OpenAI Compatible Endpoint at http://0.0.0.0:8000

Step 2: Set OpenAI API Base

$ export OPENAI_API_BASE="http://0.0.0.0:8000"

Step 3: Run LM-Eval-Harness

$ python3 main.py \
  --model gpt3 \
  --model_args engine=huggingface/bigcode/starcoder \
  --tasks hellaswag

Caching

Control caching per completion request

Caching can be switched on/off per /chat/completions request

• Caching on for completion - pass caching=True:

  curl http://0.0.0.0:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
     "model": "gpt-3.5-turbo",
     "messages": [{"role": "user", "content": "write a poem about litellm!"}],
     "temperature": 0.7,
     "caching": true
   }'
  

• Caching off for completion - pass caching=False:

  curl http://0.0.0.0:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
     "model": "gpt-3.5-turbo",
     "messages": [{"role": "user", "content": "write a poem about litellm!"}],
     "temperature": 0.7,
     "caching": false
   }'
  

Set Custom Prompt Templates

LiteLLM by default checks if a model has a prompt template and applies it (e.g. if a huggingface model has a saved chat template in it's tokenizer_config.json). However, you can also set a custom prompt template on your proxy in the config.yaml:

Step 1: Save your prompt template in a config.yaml

# Model-specific parameters
model_list:
  - model_name: mistral-7b # model alias
    litellm_params: # actual params for litellm.completion()
      model: "huggingface/mistralai/Mistral-7B-Instruct-v0.1" 
      api_base: "<your-api-base>"
      api_key: "<your-api-key>" # [OPTIONAL] for hf inference endpoints
      initial_prompt_value: "\n"
      roles: {"system":{"pre_message":"<|im_start|>system\n", "post_message":"<|im_end|>"}, "assistant":{"pre_message":"<|im_start|>assistant\n","post_message":"<|im_end|>"}, "user":{"pre_message":"<|im_start|>user\n","post_message":"<|im_end|>"}}
      final_prompt_value: "\n"
      bos_token: "<s>"
      eos_token: "</s>"
      max_tokens: 4096

Step 2: Start server with config

$ litellm --config /path/to/config.yaml

Multiple Models

If you have 1 model running on a local GPU and another that's hosted (e.g. on Runpod), you can call both via the same litellm server by listing them in your config.yaml.

model_list:
  - model_name: zephyr-alpha
    litellm_params: # params for litellm.completion() - https://docs.litellm.ai/docs/completion/input#input---request-body
      model: huggingface/HuggingFaceH4/zephyr-7b-alpha
      api_base: http://0.0.0.0:8001
  - model_name: zephyr-beta
    litellm_params:
      model: huggingface/HuggingFaceH4/zephyr-7b-beta
      api_base: https://<my-hosted-endpoint>

$ litellm --config /path/to/config.yaml

Evaluate model

If you're repo let's you set model name, you can call the specific model by just passing in that model's name -

import openai 
openai.api_base = "http://0.0.0.0:8000" 

completion = openai.ChatCompletion.create(model="zephyr-alpha", messages=[{"role": "user", "content": "Hello world"}])
print(completion.choices[0].message.content)

If you're repo only let's you specify api base, then you can add the model name to the api base passed in -

import openai 
openai.api_base = "http://0.0.0.0:8000/openai/deployments/zephyr-alpha/chat/completions" # zephyr-alpha will be used 

completion = openai.ChatCompletion.create(model="gpt-3.5-turbo", messages=[{"role": "user", "content": "Hello world"}])
print(completion.choices[0].message.content)

Save Model-specific params (API Base, API Keys, Temperature, etc.)

Use the router_config_template.yaml to save model-specific information like api_base, api_key, temperature, max_tokens, etc.

Step 1: Create a config.yaml file

model_list:
  - model_name: gpt-3.5-turbo
    litellm_params: # params for litellm.completion() - https://docs.litellm.ai/docs/completion/input#input---request-body
      model: azure/chatgpt-v-2 # azure/<your-deployment-name>
      api_key: your_azure_api_key
      api_version: your_azure_api_version
      api_base: your_azure_api_base
  - model_name: mistral-7b
    litellm_params:
      model: ollama/mistral
      api_base: your_ollama_api_base

Step 2: Start server with config

$ litellm --config /path/to/config.yaml

Model Alias

Set a model alias for your deployments.

In the config.yaml the model_name parameter is the user-facing name to use for your deployment.

E.g.: If we want to save a Huggingface TGI Mistral-7b deployment, as 'mistral-7b' for our users, we might save it as:

model_list:
  - model_name: mistral-7b # ALIAS
    litellm_params:
      model: huggingface/mistralai/Mistral-7B-Instruct-v0.1 # ACTUAL NAME
      api_key: your_huggingface_api_key # [OPTIONAL] if deployed on huggingface inference endpoints
      api_base: your_api_base # url where model is deployed