litellm-mirror/litellm/proxy/proxy_server.py

import sys, os, platform, time, copy
import threading
import shutil, random, traceback
sys.path.insert(
    0, os.path.abspath("../..")
)  # Adds the parent directory to the system path - for litellm local dev


try:
    import uvicorn
    import fastapi
    import tomli as tomllib
    import appdirs
except ImportError:
    import subprocess
    import sys

    subprocess.check_call([sys.executable, "-m", "pip", "install", "uvicorn", "fastapi", "tomli", "appdirs"])
    import uvicorn
    import fastapi
    import tomli as tomllib
    import appdirs
    

import random
list_of_messages = [
    "'The thing I wish you improved is...'",
    "'A feature I really want is...'",
    "'The worst thing about this product is...'",
    "'This product would be better if...'",
    "'I don't like how this works...'",
    "'It would help me if you could add...'",
    "'This feature doesn't meet my needs because...'",
    "'I get frustrated when the product...'",  
]

def generate_feedback_box():
  box_width = 60

  # Select a random message
  message = random.choice(list_of_messages)

  print()
  print('\033[1;37m' + '#' + '-'*box_width + '#\033[0m')
  print('\033[1;37m' + '#' + ' '*box_width + '#\033[0m')
  print('\033[1;37m' + '# {:^59} #\033[0m'.format(message))
  print('\033[1;37m' + '# {:^59} #\033[0m'.format('https://github.com/BerriAI/litellm/issues/new'))
  print('\033[1;37m' + '#' + ' '*box_width + '#\033[0m') 
  print('\033[1;37m' + '#' + '-'*box_width + '#\033[0m')
  print()
  print(' Thank you for using LiteLLM! - Krrish & Ishaan')
  print()
  print()

generate_feedback_box()


print()
print("\033[1;31mGive Feedback / Get Help: https://github.com/BerriAI/litellm/issues/new\033[0m")
print()
print("\033[1;34mDocs: https://docs.litellm.ai/docs/proxy_server\033[0m")
print() 

import litellm
from fastapi import FastAPI, Request
from fastapi.routing import APIRouter
from fastapi.responses import StreamingResponse, FileResponse
from fastapi.middleware.cors import CORSMiddleware
import json
import logging

app = FastAPI(docs_url="/", title="LiteLLM API")
router = APIRouter()
origins = ["*"]

app.add_middleware(
    CORSMiddleware,
    allow_origins=origins,
    allow_credentials=True,
    allow_methods=["*"],
    allow_headers=["*"],
)

user_api_base = None
user_model = None
user_debug = False
user_max_tokens = None
user_temperature = None
user_telemetry = True
user_config = None
config_filename = "litellm.secrets.toml"
config_dir = os.getcwd()
config_dir = appdirs.user_config_dir("litellm")
user_config_path = os.path.join(config_dir, config_filename)
log_file = 'api_log.json'
#### HELPER FUNCTIONS ####
def print_verbose(print_statement):
    global user_debug 
    if user_debug: 
         print(print_statement)

def usage_telemetry(feature: str): # helps us know if people are using this feature. Set `litellm --telemetry False` to your cli call to turn this off
    if user_telemetry: 
        data = {
            "feature": feature # "local_proxy_server"
        }
        threading.Thread(target=litellm.utils.litellm_telemetry, args=(data,), daemon=True).start()

def load_config():
    try: 
        global user_config, user_api_base, user_max_tokens, user_temperature, user_model
        # As the .env file is typically much simpler in structure, we use load_dotenv here directly
        with open(user_config_path, "rb") as f:
            user_config = tomllib.load(f)

        ## load keys
        if "keys" in user_config:
            for key in user_config["keys"]:
                if key == "HUGGINGFACE_API_KEY":
                    litellm.huggingface_key = user_config["keys"][key]
                elif key == "OPENAI_API_KEY":
                    litellm.openai_key = user_config["keys"][key]
                elif key == "TOGETHERAI_API_KEY": 
                    litellm.togetherai_api_key = user_config["keys"][key]
                elif key == "NLP_CLOUD_API_KEY": 
                    litellm.nlp_cloud_key = user_config["keys"][key]
                elif key == "ANTHROPIC_API_KEY":
                    litellm.anthropic_key = user_config["keys"][key]
                elif key == "REPLICATE_API_KEY":
                    litellm.replicate_key = user_config["keys"][key]
                elif key == "AWS_ACCESS_KEY_ID":
                    os.environ["AWS_ACCESS_KEY_ID"] = user_config["keys"][key]
                elif key == "AWS_SECRET_ACCESS_KEY":
                    os.environ["AWS_SECRET_ACCESS_KEY"] = user_config["keys"][key]

        ## settings 
        if "general" in user_config:
            litellm.add_function_to_prompt = user_config["general"].get("add_function_to_prompt", True) # by default add function to prompt if unsupported by provider
            litellm.drop_params = user_config["general"].get("drop_params", True) # by default drop params if unsupported by provider
            litellm.model_fallbacks = user_config["general"].get("fallbacks", None) # fallback models in case initial completion call fails 
            default_model = user_config["general"].get("default_model", None) # route all requests to this model. 

            if user_model is None: # `litellm --model <model-name>`` > default_model.
                user_model = default_model

        ## load model config - to set this run `litellm --config`
        model_config = None
        if "model" in user_config:
            if user_model in user_config["model"]: 
                model_config = user_config["model"][user_model]
        
        print_verbose(f"user_config: {user_config}")
        print_verbose(f"model_config: {model_config}")
        print_verbose(f"user_model: {user_model}")
        if model_config is None:
            return

        user_max_tokens = model_config.get("max_tokens", None)
        user_temperature = model_config.get("temperature", None)
        user_api_base = model_config.get("api_base", None)
        
        ## custom prompt template
        if "prompt_template" in model_config:
            model_prompt_template = model_config["prompt_template"]
            if len(model_prompt_template.keys()) > 0: # if user has initialized this at all
                litellm.register_prompt_template(
                    model=user_model,
                    initial_prompt_value=model_prompt_template.get("MODEL_PRE_PROMPT", ""),
                    roles={
                        "system": {
                            "pre_message": model_prompt_template.get("MODEL_SYSTEM_MESSAGE_START_TOKEN", ""),
                            "post_message": model_prompt_template.get("MODEL_SYSTEM_MESSAGE_END_TOKEN", ""), 
                        }, 
                        "user": {
                            "pre_message": model_prompt_template.get("MODEL_USER_MESSAGE_START_TOKEN", ""),
                            "post_message": model_prompt_template.get("MODEL_USER_MESSAGE_END_TOKEN", ""), 
                        }, 
                        "assistant": {
                            "pre_message": model_prompt_template.get("MODEL_ASSISTANT_MESSAGE_START_TOKEN", ""),
                            "post_message": model_prompt_template.get("MODEL_ASSISTANT_MESSAGE_END_TOKEN", ""), 
                        }
                    }, 
                    final_prompt_value=model_prompt_template.get("MODEL_POST_PROMPT", ""),
                )
    except Exception as e:
        pass

def initialize(model, api_base, debug, temperature, max_tokens, max_budget, telemetry, drop_params, add_function_to_prompt):
    global user_model, user_api_base, user_debug, user_max_tokens, user_temperature, user_telemetry
    user_model = model
    user_debug = debug
    
    load_config()
    user_api_base = api_base
    user_max_tokens = max_tokens
    user_temperature = temperature
    user_telemetry = telemetry
    usage_telemetry(feature="local_proxy_server")
    if drop_params == True: 
        litellm.drop_params = True
    if add_function_to_prompt == True: 
        litellm.add_function_to_prompt = True
    if max_budget: 
        litellm.max_budget = max_budget


def deploy_proxy(model, api_base, debug, temperature, max_tokens, telemetry, deploy):
    import requests
    # Load .env file

    # Prepare data for posting
    data = {
        "model": model,
        "api_base": api_base,
        "temperature": temperature,
        "max_tokens": max_tokens,
    }

    # print(data)

    # Make post request to the url
    url = "https://litellm-api.onrender.com/deploy"
    # url = "http://0.0.0.0:4000/deploy"

    with open(".env", "w") as env_file:
        for row in data:
            env_file.write(f"{row.upper()}='{data[row]}'\n")
        env_file.write("\n\n")
        for key in os.environ:
            value = os.environ[key]
            env_file.write(f"{key}='{value}'\n")
        # env_file.write(str(os.environ))

    files = {"file": open(".env", "rb")}
    # print(files)



    response = requests.post(url, data=data, files=files)
    # print(response)
    # Check the status of the request
    if response.status_code != 200:
        return f"Request to url: {url} failed with status: {response.status_code}"

    # Reading the response
    response_data = response.json()
    # print(response_data)
    url = response_data["url"]
    # # Do something with response_data

    return url


# for streaming
def data_generator(response):
    print_verbose("inside generator")
    for chunk in response:
        print_verbose(f"returned chunk: {chunk}")
        yield f"data: {json.dumps(chunk)}\n\n"

def track_cost_callback(
    kwargs,                 # kwargs to completion
    completion_response,    # response from completion
    start_time, end_time    # start/end time
):
    # track cost like this 
    # {
    #     "Oct12": {
    #         "gpt-4": 10,
    #         "claude-2": 12.01, 
    #     },
    #     "Oct 15": {
    #         "ollama/llama2": 0.0,
    #         "gpt2": 1.2
    #     }
    # }
    try:

        # for streaming responses
        if "complete_streaming_response" in kwargs:
            # for tracking streaming cost we pass the "messages" and the output_text to litellm.completion_cost 
            completion_response=kwargs["complete_streaming_response"]
            input_text = kwargs["messages"]
            output_text = completion_response["choices"][0]["message"]["content"]
            response_cost = litellm.completion_cost(
                model = kwargs["model"],
                messages = input_text,
                completion=output_text
            )
            model = kwargs['model']
            print("streaming response_cost", response_cost)

        # for non streaming responses
        else:
            # we pass the completion_response obj
            if kwargs["stream"] != True:
                response_cost = litellm.completion_cost(completion_response=completion_response)
                print("regular response_cost", response_cost)
                model = completion_response["model"]

        # read/write from json for storing daily model costs 
        cost_data = {}
        try:
            with open("costs.json") as f:
                cost_data = json.load(f)
        except FileNotFoundError:
            cost_data = {} 
        import datetime
        date = datetime.datetime.now().strftime("%b-%d-%Y")
        if date not in cost_data:
            cost_data[date] = {}

        if kwargs["model"] in cost_data[date]:
            cost_data[date][kwargs["model"]]["cost"] += response_cost
            cost_data[date][kwargs["model"]]["num_requests"] += 1
        else:
            cost_data[date][kwargs["model"]] = {
                "cost": response_cost,
                "num_requests": 1
            }

        with open("costs.json", "w") as f:
            json.dump(cost_data, f, indent=2)

    except:
        pass

def logger(
    kwargs,                 # kwargs to completion
    completion_response=None,    # response from completion
    start_time=None, 
    end_time=None    # start/end time
):
  log_event_type = kwargs['log_event_type']
  try: 
    if log_event_type == 'pre_api_call':
        inference_params = copy.deepcopy(kwargs)
        timestamp = inference_params.pop('start_time')
        dt_key = timestamp.strftime("%Y%m%d%H%M%S%f")[:23]
        log_data = {
            dt_key: {
                'pre_api_call': inference_params
            }
        }
        
        try:
            with open(log_file, 'r') as f:
                existing_data = json.load(f)
        except FileNotFoundError:
            existing_data = {}
            
        existing_data.update(log_data)
        
        with open(log_file, 'w') as f:
            json.dump(existing_data, f, indent=2)
    elif log_event_type == 'post_api_call':
        if "stream" not in kwargs["optional_params"] or kwargs["optional_params"]["stream"] is False or kwargs.get("complete_streaming_response", False):
            inference_params = copy.deepcopy(kwargs)
            timestamp = inference_params.pop('start_time')
            dt_key = timestamp.strftime("%Y%m%d%H%M%S%f")[:23]
            
            with open(log_file, 'r') as f:
                existing_data = json.load(f)
            
            existing_data[dt_key]['post_api_call'] = inference_params
            
            with open(log_file, 'w') as f:
                json.dump(existing_data, f, indent=2)
  except: 
      traceback.print_exc()

litellm.input_callback = [logger]
litellm.success_callback = [logger]
litellm.failure_callback = [logger]

def litellm_completion(data, type): 
    try: 
        if user_model:
            data["model"] = user_model
        # override with user settings
        if user_temperature: 
            data["temperature"] = user_temperature
        if user_max_tokens: 
            data["max_tokens"] = user_max_tokens
        if user_api_base: 
            data["api_base"] = user_api_base
        if type == "completion": 
            response = litellm.text_completion(**data)
        elif type == "chat_completion": 
            response = litellm.completion(**data)
        if 'stream' in data and data['stream'] == True: # use generate_responses to stream responses
            return StreamingResponse(data_generator(response), media_type='text/event-stream')
        print_verbose(f"response: {response}")
        return response
    except Exception as e: 
        if "Invalid response object from API" in str(e): 
            completion_call_details = {}
            if user_model: 
                completion_call_details["model"] = user_model
            else: 
                completion_call_details["model"] = data['model']
            
            if user_api_base: 
                completion_call_details["api_base"] = user_api_base
            else: 
                completion_call_details["api_base"] = None
            print(f"\033[1;31mLiteLLM.Exception: Invalid API Call. Call details: Model: \033[1;37m{completion_call_details['model']}\033[1;31m; LLM Provider: \033[1;37m{e.llm_provider}\033[1;31m; Custom API Base - \033[1;37m{completion_call_details['api_base']}\033[1;31m\033[0m")
            if completion_call_details["api_base"] == "http://localhost:11434": 
                print()
                print("Trying to call ollama? Try `litellm --model ollama/llama2 --api_base http://localhost:11434`")
                print()
        else: 
            print(f"\033[1;31mLiteLLM.Exception: {str(e)}\033[0m")
        return {"message": "An error occurred"}, 500

#### API ENDPOINTS ####
@router.get("/models") # if project requires model list 
def model_list():
    if user_model != None:
        return dict(
            data=[{"id": user_model, "object": "model", "created": 1677610602, "owned_by": "openai"}],
            object="list",
        )
    else:
        all_models = litellm.model_list
        return dict(
            data = [{"id": model, "object": "model", "created": 1677610602, "owned_by": "openai"} for model in all_models],
            object="list",
        )

@router.post("/completions")
async def completion(request: Request):
    data = await request.json()
    return litellm_completion(data=data, type="completion")

@router.post("/chat/completions")
async def chat_completion(request: Request):
    data = await request.json()
    response = litellm_completion(data, type="chat_completion")
    return response


# V1 Endpoints - some apps expect a v1 endpoint - these call the regular function
@router.post("/v1/completions")
async def v1_completion(request: Request):
    data = await request.json()
    return litellm_completion(data=data, type="completion")

@router.post("/v1/chat/completions")
async def v1_chat_completion(request: Request):
    data = await request.json()
    print_verbose(f"data passed in: {data}")
    response = litellm_completion(data, type="chat_completion")
    return response

def print_cost_logs():
    with open('costs.json', 'r') as f:
        # print this in green
        print("\033[1;32m")
        print(f.read())
        print("\033[0m")
    return

@router.get("/ollama_logs")
async def retrieve_server_log(request: Request):
    filepath = os.path.expanduser('~/.ollama/logs/server.log')
    return FileResponse(filepath)

@router.get("/")
async def home(request: Request):
    return "LiteLLM: RUNNING"

app.include_router(router)