# +-----------------------------------------------+
# | |
# | Give Feedback / Get Help |
# | https://github.com/BerriAI/litellm/issues/new |
# | |
# +-----------------------------------------------+
#
# Thank you ! We ❤️ you! - Krrish & Ishaan
from datetime import datetime
from typing import Dict, List, Optional, Union, Literal
import random, threading, time, traceback
import litellm, openai
from litellm.caching import RedisCache, InMemoryCache, DualCache
import logging, asyncio
import inspect, concurrent
from openai import AsyncOpenAI
from collections import defaultdict
class Router:
"""
Example usage:
```python
from litellm import Router
model_list = [
{
"model_name": "azure-gpt-3.5-turbo", # model alias
"litellm_params": { # params for litellm completion/embedding call
"model": "azure/<your-deployment-name-1>",
"api_key": <your-api-key>,
"api_version": <your-api-version>,
"api_base": <your-api-base>
},
},
{
"model_name": "azure-gpt-3.5-turbo", # model alias
"litellm_params": { # params for litellm completion/embedding call
"model": "azure/<your-deployment-name-2>",
"api_key": <your-api-key>,
"api_version": <your-api-version>,
"api_base": <your-api-base>
},
},
{
"model_name": "openai-gpt-3.5-turbo", # model alias
"litellm_params": { # params for litellm completion/embedding call
"model": "gpt-3.5-turbo",
"api_key": <your-api-key>,
},
]
router = Router(model_list=model_list, fallbacks=[{"azure-gpt-3.5-turbo": "openai-gpt-3.5-turbo"}])
```
"""
model_names: List = []
cache_responses: bool = False
default_cache_time_seconds: int = 1 * 60 * 60 # 1 hour
num_retries: int = 0
tenacity = None
def __init__(self,
model_list: Optional[list] = None,
redis_host: Optional[str] = None,
redis_port: Optional[int] = None,
redis_password: Optional[str] = None,
cache_responses: bool = False,
num_retries: int = 0,
timeout: Optional[float] = None,
default_litellm_params = {}, # default params for Router.chat.completion.create
set_verbose: bool = False,
fallbacks: List = [],
allowed_fails: Optional[int] = None,
context_window_fallbacks: List = [],
routing_strategy: Literal["simple-shuffle", "least-busy", "usage-based-routing", "latency-based-routing"] = "simple-shuffle") -> None:
self.set_verbose = set_verbose
if model_list:
self.set_model_list(model_list)
self.healthy_deployments: List = self.model_list
self.deployment_latency_map = {}
for m in model_list:
self.deployment_latency_map[m["litellm_params"]["model"]] = 0
self.allowed_fails = allowed_fails or litellm.allowed_fails
self.failed_calls = InMemoryCache() # cache to track failed call per deployment, if num failed calls within 1 minute > allowed fails, then add it to cooldown
self.num_retries = num_retries or litellm.num_retries or 0
self.timeout = timeout or litellm.request_timeout
self.routing_strategy = routing_strategy
self.fallbacks = fallbacks or litellm.fallbacks
self.context_window_fallbacks = context_window_fallbacks or litellm.context_window_fallbacks
self.model_exception_map: dict = {} # dict to store model: list exceptions. self.exceptions = {"gpt-3.5": ["API KEY Error", "Rate Limit Error", "good morning error"]}
self.total_calls: defaultdict = defaultdict(int) # dict to store total calls made to each model
self.fail_calls: defaultdict = defaultdict(int) # dict to store fail_calls made to each model
self.success_calls: defaultdict = defaultdict(int) # dict to store success_calls made to each model
# make Router.chat.completions.create compatible for openai.chat.completions.create
self.chat = litellm.Chat(params=default_litellm_params)
# default litellm args
self.default_litellm_params = default_litellm_params
self.default_litellm_params.setdefault("timeout", timeout)
self.default_litellm_params.setdefault("max_retries", 0)
### HEALTH CHECK THREAD ###
if self.routing_strategy == "least-busy":
self._start_health_check_thread()
### CACHING ###
redis_cache = None
if redis_host is not None and redis_port is not None and redis_password is not None:
cache_config = {
'type': 'redis',
'host': redis_host,
'port': redis_port,
'password': redis_password
}
redis_cache = RedisCache(host=redis_host, port=redis_port, password=redis_password)
else: # use an in-memory cache
cache_config = {
"type": "local"
}
if cache_responses:
litellm.cache = litellm.Cache(**cache_config) # use Redis for caching completion requests
self.cache_responses = cache_responses
self.cache = DualCache(redis_cache=redis_cache, in_memory_cache=InMemoryCache()) # use a dual cache (Redis+In-Memory) for tracking cooldowns, usage, etc.
## USAGE TRACKING ##
if isinstance(litellm.success_callback, list):
litellm.success_callback.append(self.deployment_callback)
else:
litellm.success_callback = [self.deployment_callback]
if isinstance(litellm.failure_callback, list):
litellm.failure_callback.append(self.deployment_callback_on_failure)
else:
litellm.failure_callback = [self.deployment_callback_on_failure]
self.print_verbose(f"Intialized router with Routing strategy: {self.routing_strategy}\n")
### COMPLETION + EMBEDDING FUNCTIONS
def completion(self,
model: str,
messages: List[Dict[str, str]],
**kwargs):
"""
Example usage:
response = router.completion(model="gpt-3.5-turbo", messages=[{"role": "user", "content": "Hey, how's it going?"}]
"""
try:
kwargs["model"] = model
kwargs["messages"] = messages
kwargs["original_function"] = self._completion
timeout = kwargs.get("request_timeout", self.timeout)
kwargs["num_retries"] = kwargs.get("num_retries", self.num_retries)
kwargs.setdefault("metadata", {}).update({"model_group": model})
with concurrent.futures.ThreadPoolExecutor(max_workers=1) as executor:
# Submit the function to the executor with a timeout
future = executor.submit(self.function_with_fallbacks, **kwargs)
response = future.result(timeout=timeout) # type: ignore
return response
except Exception as e:
raise e
def _completion(
self,
model: str,
messages: List[Dict[str, str]],
**kwargs):
try:
# pick the one that is available (lowest TPM/RPM)
deployment = self.get_available_deployment(model=model, messages=messages)
kwargs.setdefault("metadata", {}).update({"deployment": deployment["litellm_params"]["model"]})
data = deployment["litellm_params"].copy()
for k, v in self.default_litellm_params.items():
if k not in data: # prioritize model-specific params > default router params
data[k] = v
########## remove -ModelID-XXXX from model ##############
original_model_string = data["model"]
# Find the index of "ModelID" in the string
self.print_verbose(f"completion model: {original_model_string}")
index_of_model_id = original_model_string.find("-ModelID")
# Remove everything after "-ModelID" if it exists
if index_of_model_id != -1:
data["model"] = original_model_string[:index_of_model_id]
else:
data["model"] = original_model_string
model_client = self._get_client(deployment=deployment, kwargs=kwargs)
return litellm.completion(**{**data, "messages": messages, "caching": self.cache_responses, "client": model_client, **kwargs})
except Exception as e:
raise e
async def acompletion(self,
model: str,
messages: List[Dict[str, str]],
**kwargs):
try:
kwargs["model"] = model
kwargs["messages"] = messages
kwargs["original_function"] = self._acompletion
kwargs["num_retries"] = kwargs.get("num_retries", self.num_retries)
timeout = kwargs.get("request_timeout", self.timeout)
kwargs.setdefault("metadata", {}).update({"model_group": model})
# response = await asyncio.wait_for(self.async_function_with_fallbacks(**kwargs), timeout=timeout)
response = await self.async_function_with_fallbacks(**kwargs)
return response
except Exception as e:
raise e
async def _acompletion(
self,
model: str,
messages: List[Dict[str, str]],
**kwargs):
try:
self.print_verbose(f"Inside _acompletion()- model: {model}; kwargs: {kwargs}")
original_model_string = None # set a default for this variable
deployment = self.get_available_deployment(model=model, messages=messages)
kwargs.setdefault("metadata", {}).update({"deployment": deployment["litellm_params"]["model"]})
data = deployment["litellm_params"].copy()
for k, v in self.default_litellm_params.items():
if k not in data: # prioritize model-specific params > default router params
data[k] = v
########## remove -ModelID-XXXX from model ##############
original_model_string = data["model"]
# Find the index of "ModelID" in the string
index_of_model_id = original_model_string.find("-ModelID")
# Remove everything after "-ModelID" if it exists
if index_of_model_id != -1:
data["model"] = original_model_string[:index_of_model_id]
else:
data["model"] = original_model_string
model_client = self._get_client(deployment=deployment, kwargs=kwargs, client_type="async")
self.total_calls[original_model_string] +=1
response = await litellm.acompletion(**{**data, "messages": messages, "caching": self.cache_responses, "client": model_client, **kwargs})
self.success_calls[original_model_string] +=1
return response
except Exception as e:
if original_model_string is not None:
self.fail_calls[original_model_string] +=1
raise e
def text_completion(self,
model: str,
prompt: str,
is_retry: Optional[bool] = False,
is_fallback: Optional[bool] = False,
is_async: Optional[bool] = False,
**kwargs):
try:
kwargs.setdefault("metadata", {}).update({"model_group": model})
messages=[{"role": "user", "content": prompt}]
# pick the one that is available (lowest TPM/RPM)
deployment = self.get_available_deployment(model=model, messages=messages)
data = deployment["litellm_params"].copy()
for k, v in self.default_litellm_params.items():
if k not in data: # prioritize model-specific params > default router params
data[k] = v
########## remove -ModelID-XXXX from model ##############
original_model_string = data["model"]
# Find the index of "ModelID" in the string
index_of_model_id = original_model_string.find("-ModelID")
# Remove everything after "-ModelID" if it exists
if index_of_model_id != -1:
data["model"] = original_model_string[:index_of_model_id]
else:
data["model"] = original_model_string
# call via litellm.completion()
return litellm.text_completion(**{**data, "prompt": prompt, "caching": self.cache_responses, **kwargs}) # type: ignore
except Exception as e:
if self.num_retries > 0:
kwargs["model"] = model
kwargs["messages"] = messages
kwargs["original_exception"] = e
kwargs["original_function"] = self.completion
return self.function_with_retries(**kwargs)
else:
raise e
def embedding(self,
model: str,
input: Union[str, List],
is_async: Optional[bool] = False,
**kwargs) -> Union[List[float], None]:
# pick the one that is available (lowest TPM/RPM)
deployment = self.get_available_deployment(model=model, input=input)
kwargs.setdefault("metadata", {}).update({"deployment": deployment["litellm_params"]["model"]})
data = deployment["litellm_params"].copy()
for k, v in self.default_litellm_params.items():
if k not in data: # prioritize model-specific params > default router params
data[k] = v
########## remove -ModelID-XXXX from model ##############
original_model_string = data["model"]
# Find the index of "ModelID" in the string
index_of_model_id = original_model_string.find("-ModelID")
# Remove everything after "-ModelID" if it exists
if index_of_model_id != -1:
data["model"] = original_model_string[:index_of_model_id]
else:
data["model"] = original_model_string
model_client = self._get_client(deployment=deployment, kwargs=kwargs)
# call via litellm.embedding()
return litellm.embedding(**{**data, "input": input, "caching": self.cache_responses, "client": model_client, **kwargs})
async def aembedding(self,
model: str,
input: Union[str, List],
is_async: Optional[bool] = True,
**kwargs) -> Union[List[float], None]:
# pick the one that is available (lowest TPM/RPM)
deployment = self.get_available_deployment(model=model, input=input)
kwargs.setdefault("metadata", {}).update({"deployment": deployment["litellm_params"]["model"]})
data = deployment["litellm_params"].copy()
for k, v in self.default_litellm_params.items():
if k not in data: # prioritize model-specific params > default router params
data[k] = v
########## remove -ModelID-XXXX from model ##############
original_model_string = data["model"]
# Find the index of "ModelID" in the string
index_of_model_id = original_model_string.find("-ModelID")
# Remove everything after "-ModelID" if it exists
if index_of_model_id != -1:
data["model"] = original_model_string[:index_of_model_id]
else:
data["model"] = original_model_string
model_client = self._get_client(deployment=deployment, kwargs=kwargs, client_type="async")
return await litellm.aembedding(**{**data, "input": input, "caching": self.cache_responses, "client": model_client, **kwargs})
async def async_function_with_fallbacks(self, *args, **kwargs):
"""
Try calling the function_with_retries
If it fails after num_retries, fall back to another model group
"""
model_group = kwargs.get("model")
fallbacks = kwargs.get("fallbacks", self.fallbacks)
context_window_fallbacks = kwargs.get("context_window_fallbacks", self.context_window_fallbacks)
try:
response = await self.async_function_with_retries(*args, **kwargs)
self.print_verbose(f'Async Response: {response}')
return response
except Exception as e:
self.print_verbose(f"An exception occurs")
original_exception = e
try:
self.print_verbose(f"Trying to fallback b/w models")
if isinstance(e, litellm.ContextWindowExceededError) and context_window_fallbacks is not None:
fallback_model_group = None
for item in context_window_fallbacks: # [{"gpt-3.5-turbo": ["gpt-4"]}]
if list(item.keys())[0] == model_group:
fallback_model_group = item[model_group]
break
if fallback_model_group is None:
raise original_exception
for mg in fallback_model_group:
"""
Iterate through the model groups and try calling that deployment
"""
try:
kwargs["model"] = mg
response = await self.async_function_with_retries(*args, **kwargs)
return response
except Exception as e:
pass
elif fallbacks is not None:
self.print_verbose(f"inside model fallbacks: {fallbacks}")
for item in fallbacks:
if list(item.keys())[0] == model_group:
fallback_model_group = item[model_group]
break
for mg in fallback_model_group:
"""
Iterate through the model groups and try calling that deployment
"""
try:
kwargs["model"] = mg
kwargs["metadata"]["model_group"] = mg
response = await self.async_function_with_retries(*args, **kwargs)
return response
except Exception as e:
raise e
except Exception as e:
self.print_verbose(f"An exception occurred - {str(e)}")
traceback.print_exc()
raise original_exception
async def async_function_with_retries(self, *args, **kwargs):
self.print_verbose(f"Inside async function with retries: args - {args}; kwargs - {kwargs}")
backoff_factor = 1
original_function = kwargs.pop("original_function")
fallbacks = kwargs.pop("fallbacks", self.fallbacks)
context_window_fallbacks = kwargs.pop("context_window_fallbacks", self.context_window_fallbacks)
self.print_verbose(f"async function w/ retries: original_function - {original_function}")
num_retries = kwargs.pop("num_retries")
try:
# if the function call is successful, no exception will be raised and we'll break out of the loop
response = await original_function(*args, **kwargs)
return response
except Exception as e:
original_exception = e
### CHECK IF RATE LIMIT / CONTEXT WINDOW ERROR w/ fallbacks available
if ((isinstance(original_exception, litellm.ContextWindowExceededError) and context_window_fallbacks is None)
or (isinstance(original_exception, openai.RateLimitError) and fallbacks is not None)):
raise original_exception
### RETRY
#### check if it should retry + back-off if required
if "No models available" in str(e):
timeout = litellm._calculate_retry_after(remaining_retries=num_retries, max_retries=num_retries)
await asyncio.sleep(timeout)
elif hasattr(original_exception, "status_code") and hasattr(original_exception, "response") and litellm._should_retry(status_code=original_exception.status_code):
if hasattr(original_exception.response, "headers"):
timeout = litellm._calculate_retry_after(remaining_retries=num_retries, max_retries=num_retries, response_headers=original_exception.response.headers)
else:
timeout = litellm._calculate_retry_after(remaining_retries=num_retries, max_retries=num_retries)
await asyncio.sleep(timeout)
else:
raise original_exception
for current_attempt in range(num_retries):
self.print_verbose(f"retrying request. Current attempt - {current_attempt}; num retries: {num_retries}")
try:
# if the function call is successful, no exception will be raised and we'll break out of the loop
response = await original_function(*args, **kwargs)
if inspect.iscoroutinefunction(response): # async errors are often returned as coroutines
response = await response
return response
except Exception as e:
remaining_retries = num_retries - current_attempt
if "No models available" in str(e):
timeout = litellm._calculate_retry_after(remaining_retries=remaining_retries, max_retries=num_retries, min_timeout=1)
await asyncio.sleep(timeout)
elif hasattr(e, "status_code") and hasattr(e, "response") and litellm._should_retry(status_code=e.status_code):
if hasattr(e.response, "headers"):
timeout = litellm._calculate_retry_after(remaining_retries=remaining_retries, max_retries=num_retries, response_headers=e.response.headers)
else:
timeout = litellm._calculate_retry_after(remaining_retries=remaining_retries, max_retries=num_retries)
await asyncio.sleep(timeout)
else:
raise e
raise original_exception
def function_with_fallbacks(self, *args, **kwargs):
"""
Try calling the function_with_retries
If it fails after num_retries, fall back to another model group
"""
model_group = kwargs.get("model")
fallbacks = kwargs.get("fallbacks", self.fallbacks)
context_window_fallbacks = kwargs.get("context_window_fallbacks", self.context_window_fallbacks)
try:
response = self.function_with_retries(*args, **kwargs)
return response
except Exception as e:
original_exception = e
self.print_verbose(f"An exception occurs {original_exception}")
try:
self.print_verbose(f"Trying to fallback b/w models. Initial model group: {model_group}")
if isinstance(e, litellm.ContextWindowExceededError) and context_window_fallbacks is not None:
self.print_verbose(f"inside context window fallbacks: {context_window_fallbacks}")
fallback_model_group = None
for item in context_window_fallbacks: # [{"gpt-3.5-turbo": ["gpt-4"]}]
if list(item.keys())[0] == model_group:
fallback_model_group = item[model_group]
break
if fallback_model_group is None:
raise original_exception
for mg in fallback_model_group:
"""
Iterate through the model groups and try calling that deployment
"""
try:
kwargs["model"] = mg
response = self.function_with_fallbacks(*args, **kwargs)
return response
except Exception as e:
pass
elif fallbacks is not None:
self.print_verbose(f"inside model fallbacks: {fallbacks}")
fallback_model_group = None
for item in fallbacks:
if list(item.keys())[0] == model_group:
fallback_model_group = item[model_group]
break
if fallback_model_group is None:
raise original_exception
for mg in fallback_model_group:
"""
Iterate through the model groups and try calling that deployment
"""
try:
kwargs["model"] = mg
response = self.function_with_fallbacks(*args, **kwargs)
return response
except Exception as e:
pass
except Exception as e:
raise e
raise original_exception
def function_with_retries(self, *args, **kwargs):
"""
Try calling the model 3 times. Shuffle between available deployments.
"""
self.print_verbose(f"Inside function with retries: args - {args}; kwargs - {kwargs}")
backoff_factor = 1
original_function = kwargs.pop("original_function")
num_retries = kwargs.pop("num_retries")
fallbacks = kwargs.pop("fallbacks", self.fallbacks)
context_window_fallbacks = kwargs.pop("context_window_fallbacks", self.context_window_fallbacks)
try:
# if the function call is successful, no exception will be raised and we'll break out of the loop
response = original_function(*args, **kwargs)
return response
except Exception as e:
original_exception = e
self.print_verbose(f"num retries in function with retries: {num_retries}")
### CHECK IF RATE LIMIT / CONTEXT WINDOW ERROR
if ((isinstance(original_exception, litellm.ContextWindowExceededError) and context_window_fallbacks is None)
or (isinstance(original_exception, openai.RateLimitError) and fallbacks is not None)):
raise original_exception
### RETRY
for current_attempt in range(num_retries):
self.print_verbose(f"retrying request. Current attempt - {current_attempt}; retries left: {num_retries}")
try:
# if the function call is successful, no exception will be raised and we'll break out of the loop
response = original_function(*args, **kwargs)
return response
except openai.RateLimitError as e:
if num_retries > 0:
remaining_retries = num_retries - current_attempt
timeout = litellm._calculate_retry_after(remaining_retries=remaining_retries, max_retries=num_retries)
# on RateLimitError we'll wait for an exponential time before trying again
time.sleep(timeout)
else:
raise e
except Exception as e:
# for any other exception types, immediately retry
if num_retries > 0:
pass
else:
raise e
raise original_exception
### HELPER FUNCTIONS
def deployment_callback(
self,
kwargs, # kwargs to completion
completion_response, # response from completion
start_time, end_time # start/end time
):
"""
Function LiteLLM submits a callback to after a successful
completion. Purpose of this is to update TPM/RPM usage per model
"""
model_name = kwargs.get('model', None) # i.e. gpt35turbo
custom_llm_provider = kwargs.get("litellm_params", {}).get('custom_llm_provider', None) # i.e. azure
if custom_llm_provider:
model_name = f"{custom_llm_provider}/{model_name}"
if kwargs["stream"] is True:
if kwargs.get("complete_streaming_response"):
total_tokens = kwargs.get("complete_streaming_response")['usage']['total_tokens']
self._set_deployment_usage(model_name, total_tokens)
else:
total_tokens = completion_response['usage']['total_tokens']
self._set_deployment_usage(model_name, total_tokens)
self.deployment_latency_map[model_name] = (end_time - start_time).total_seconds()
def deployment_callback_on_failure(
self,
kwargs, # kwargs to completion
completion_response, # response from completion
start_time, end_time # start/end time
):
try:
exception = kwargs.get("exception", None)
exception_type = type(exception)
exception_status = getattr(exception, 'status_code', "")
exception_cause = getattr(exception, '__cause__', "")
exception_message = getattr(exception, 'message', "")
exception_str = str(exception_type) + "Status: " + str(exception_status) + "Message: " + str(exception_cause) + str(exception_message) + "Full exception" + str(exception)
model_name = kwargs.get('model', None) # i.e. gpt35turbo
custom_llm_provider = kwargs.get("litellm_params", {}).get('custom_llm_provider', None) # i.e. azure
metadata = kwargs.get("litellm_params", {}).get('metadata', None)
if metadata:
deployment = metadata.get("deployment", None)
self._set_cooldown_deployments(deployment)
deployment_exceptions = self.model_exception_map.get(deployment, [])
deployment_exceptions.append(exception_str)
self.model_exception_map[deployment] = deployment_exceptions
self.print_verbose("\nEXCEPTION FOR DEPLOYMENTS\n")
self.print_verbose(self.model_exception_map)
for model in self.model_exception_map:
self.print_verbose(f"Model {model} had {len(self.model_exception_map[model])} exception")
if custom_llm_provider:
model_name = f"{custom_llm_provider}/{model_name}"
except Exception as e:
raise e
def _set_cooldown_deployments(self,
deployment: str):
"""
Add a model to the list of models being cooled down for that minute, if it exceeds the allowed fails / minute
"""
current_minute = datetime.now().strftime("%H-%M")
# get current fails for deployment
# update the number of failed calls
# if it's > allowed fails
# cooldown deployment
current_fails = self.failed_calls.get_cache(key=deployment) or 0
updated_fails = current_fails + 1
self.print_verbose(f"Attempting to add {deployment} to cooldown list. updated_fails: {updated_fails}; self.allowed_fails: {self.allowed_fails}")
if updated_fails > self.allowed_fails:
# get the current cooldown list for that minute
cooldown_key = f"{current_minute}:cooldown_models" # group cooldown models by minute to reduce number of redis calls
cached_value = self.cache.get_cache(key=cooldown_key)
self.print_verbose(f"adding {deployment} to cooldown models")
# update value
try:
if deployment in cached_value:
pass
else:
cached_value = cached_value + [deployment]
# save updated value
self.cache.set_cache(value=cached_value, key=cooldown_key, ttl=1)
except:
cached_value = [deployment]
# save updated value
self.cache.set_cache(value=cached_value, key=cooldown_key, ttl=1)
else:
self.failed_calls.set_cache(key=deployment, value=updated_fails, ttl=1)
def _get_cooldown_deployments(self):
"""
Get the list of models being cooled down for this minute
"""
current_minute = datetime.now().strftime("%H-%M")
# get the current cooldown list for that minute
cooldown_key = f"{current_minute}:cooldown_models"
# ----------------------
# Return cooldown models
# ----------------------
cooldown_models = self.cache.get_cache(key=cooldown_key) or []
self.print_verbose(f"retrieve cooldown models: {cooldown_models}")
return cooldown_models
def get_usage_based_available_deployment(self,
model: str,
messages: Optional[List[Dict[str, str]]] = None,
input: Optional[Union[str, List]] = None):
"""
Returns a deployment with the lowest TPM/RPM usage.
"""
# get list of potential deployments
potential_deployments = []
for item in self.model_list:
if item["model_name"] == model:
potential_deployments.append(item)
# get current call usage
token_count = 0
if messages is not None:
token_count = litellm.token_counter(model=model, messages=messages)
elif input is not None:
if isinstance(input, List):
input_text = "".join(text for text in input)
else:
input_text = input
token_count = litellm.token_counter(model=model, text=input_text)
# -----------------------
# Find lowest used model
# ----------------------
lowest_tpm = float("inf")
deployment = None
# return deployment with lowest tpm usage
for item in potential_deployments:
item_tpm, item_rpm = self._get_deployment_usage(deployment_name=item["litellm_params"]["model"])
if item_tpm == 0:
return item
elif ("tpm" in item and item_tpm + token_count > item["tpm"]
or "rpm" in item and item_rpm + 1 >= item["rpm"]): # if user passed in tpm / rpm in the model_list
continue
elif item_tpm < lowest_tpm:
lowest_tpm = item_tpm
deployment = item
# if none, raise exception
if deployment is None:
raise ValueError("No models available.")
# return model
return deployment
def _get_deployment_usage(
self,
deployment_name: str
):
# ------------
# Setup values
# ------------
current_minute = datetime.now().strftime("%H-%M")
tpm_key = f'{deployment_name}:tpm:{current_minute}'
rpm_key = f'{deployment_name}:rpm:{current_minute}'
# ------------
# Return usage
# ------------
tpm = self.cache.get_cache(key=tpm_key) or 0
rpm = self.cache.get_cache(key=rpm_key) or 0
return int(tpm), int(rpm)
def increment(self, key: str, increment_value: int):
# get value
cached_value = self.cache.get_cache(key=key)
# update value
try:
cached_value = cached_value + increment_value
except:
cached_value = increment_value
# save updated value
self.cache.set_cache(value=cached_value, key=key, ttl=self.default_cache_time_seconds)
def _set_deployment_usage(
self,
model_name: str,
total_tokens: int
):
# ------------
# Setup values
# ------------
current_minute = datetime.now().strftime("%H-%M")
tpm_key = f'{model_name}:tpm:{current_minute}'
rpm_key = f'{model_name}:rpm:{current_minute}'
# ------------
# Update usage
# ------------
self.increment(tpm_key, total_tokens)
self.increment(rpm_key, 1)
def _start_health_check_thread(self):
"""
Starts a separate thread to perform health checks periodically.
"""
health_check_thread = threading.Thread(target=self._perform_health_checks, daemon=True)
health_check_thread.start()
def _perform_health_checks(self):
"""
Periodically performs health checks on the servers.
Updates the list of healthy servers accordingly.
"""
while True:
self.healthy_deployments = self._health_check()
# Adjust the time interval based on your needs
time.sleep(15)
def _health_check(self):
"""
Performs a health check on the deployments
Returns the list of healthy deployments
"""
healthy_deployments = []
for deployment in self.model_list:
litellm_args = deployment["litellm_params"]
try:
start_time = time.time()
litellm.completion(messages=[{"role": "user", "content": ""}], max_tokens=1, **litellm_args) # hit the server with a blank message to see how long it takes to respond
end_time = time.time()
response_time = end_time - start_time
logging.debug(f"response_time: {response_time}")
healthy_deployments.append((deployment, response_time))
healthy_deployments.sort(key=lambda x: x[1])
except Exception as e:
pass
return healthy_deployments
def weighted_shuffle_by_latency(self, items):
# Sort the items by latency
sorted_items = sorted(items, key=lambda x: x[1])
# Get only the latencies
latencies = [i[1] for i in sorted_items]
# Calculate the sum of all latencies
total_latency = sum(latencies)
# Calculate the weight for each latency (lower latency = higher weight)
weights = [total_latency-latency for latency in latencies]
# Get a weighted random item
if sum(weights) == 0:
chosen_item = random.choice(sorted_items)[0]
else:
chosen_item = random.choices(sorted_items, weights=weights, k=1)[0][0]
return chosen_item
def set_model_list(self, model_list: list):
self.model_list = model_list
# we add api_base/api_key each model so load balancing between azure/gpt on api_base1 and api_base2 works
import os
for model in self.model_list:
litellm_params = model.get("litellm_params", {})
model_name = litellm_params.get("model")
#### for OpenAI / Azure we need to initalize the Client for High Traffic ########
custom_llm_provider = litellm_params.get("custom_llm_provider")
if custom_llm_provider is None:
custom_llm_provider = model_name.split("/",1)[0]
if (
model_name in litellm.open_ai_chat_completion_models
or custom_llm_provider == "custom_openai"
or custom_llm_provider == "deepinfra"
or custom_llm_provider == "perplexity"
or custom_llm_provider == "anyscale"
or custom_llm_provider == "openai"
or custom_llm_provider == "azure"
or "ft:gpt-3.5-turbo" in model_name
or model_name in litellm.open_ai_embedding_models
):
# glorified / complicated reading of configs
# user can pass vars directly or they can pas os.environ/AZURE_API_KEY, in which case we will read the env
# we do this here because we init clients for Azure, OpenAI and we need to set the right key
api_key = litellm_params.get("api_key")
if api_key and api_key.startswith("os.environ/"):
api_key_env_name = api_key.replace("os.environ/", "")
api_key = litellm.get_secret(api_key_env_name)
api_base = litellm_params.get("api_base")
base_url = litellm_params.get("base_url")
api_base = api_base or base_url # allow users to pass in `api_base` or `base_url` for azure
if api_base and api_base.startswith("os.environ/"):
api_base_env_name = api_base.replace("os.environ/", "")
api_base = litellm.get_secret(api_base_env_name)
api_version = litellm_params.get("api_version")
if api_version and api_version.startswith("os.environ/"):
api_version_env_name = api_version.replace("os.environ/", "")
api_version = litellm.get_secret(api_version_env_name)
timeout = litellm_params.pop("timeout", None)
if isinstance(timeout, str) and timeout.startswith("os.environ/"):
timeout_env_name = api_version.replace("os.environ/", "")
timeout = litellm.get_secret(timeout_env_name)
stream_timeout = litellm_params.pop("stream_timeout", timeout) # if no stream_timeout is set, default to timeout
if isinstance(stream_timeout, str) and stream_timeout.startswith("os.environ/"):
stream_timeout_env_name = api_version.replace("os.environ/", "")
stream_timeout = litellm.get_secret(stream_timeout_env_name)
max_retries = litellm_params.pop("max_retries", 2)
if isinstance(max_retries, str) and max_retries.startswith("os.environ/"):
max_retries_env_name = api_version.replace("os.environ/", "")
max_retries = litellm.get_secret(max_retries_env_name)
self.print_verbose(f"Initializing OpenAI Client for {model_name}, {str(api_base)}")
if "azure" in model_name:
self.print_verbose(f"Initializing Azure OpenAI Client for {model_name}, {str(api_base)}, {api_key}")
if api_version is None:
api_version = "2023-07-01-preview"
if "gateway.ai.cloudflare.com" in api_base:
if not api_base.endswith("/"):
api_base += "/"
azure_model = model_name.replace("azure/", "")
api_base += f"{azure_model}"
model["async_client"] = openai.AsyncAzureOpenAI(
api_key=api_key,
base_url=api_base,
api_version=api_version,
timeout=timeout,
max_retries=max_retries
)
model["client"] = openai.AzureOpenAI(
api_key=api_key,
base_url=api_base,
api_version=api_version,
timeout=timeout,
max_retries=max_retries
)
# streaming clients can have diff timeouts
model["stream_async_client"] = openai.AsyncAzureOpenAI(
api_key=api_key,
base_url=api_base,
api_version=api_version,
timeout=stream_timeout,
max_retries=max_retries
)
model["stream_client"] = openai.AzureOpenAI(
api_key=api_key,
base_url=api_base,
api_version=api_version,
timeout=stream_timeout,
max_retries=max_retries
)
else:
model["async_client"] = openai.AsyncAzureOpenAI(
api_key=api_key,
azure_endpoint=api_base,
api_version=api_version,
timeout=timeout,
max_retries=max_retries
)
model["client"] = openai.AzureOpenAI(
api_key=api_key,
azure_endpoint=api_base,
api_version=api_version,
timeout=timeout,
max_retries=max_retries
)
# streaming clients should have diff timeouts
model["stream_async_client"] = openai.AsyncAzureOpenAI(
api_key=api_key,
base_url=api_base,
api_version=api_version,
timeout=stream_timeout,
max_retries=max_retries
)
model["stream_client"] = openai.AzureOpenAI(
api_key=api_key,
base_url=api_base,
api_version=api_version,
timeout=stream_timeout,
max_retries=max_retries
)
else:
self.print_verbose(f"Initializing OpenAI Client for {model_name}, {str(api_base)}")
model["async_client"] = openai.AsyncOpenAI(
api_key=api_key,
base_url=api_base,
timeout=timeout,
max_retries=max_retries
)
model["client"] = openai.OpenAI(
api_key=api_key,
base_url=api_base,
timeout=timeout,
max_retries=max_retries
)
# streaming clients should have diff timeouts
model["stream_async_client"] = openai.AsyncOpenAI(
api_key=api_key,
base_url=api_base,
timeout=stream_timeout,
max_retries=max_retries
)
# streaming clients should have diff timeouts
model["stream_client"] = openai.OpenAI(
api_key=api_key,
base_url=api_base,
timeout=stream_timeout,
max_retries=max_retries
)
############ End of initializing Clients for OpenAI/Azure ###################
model_id = ""
for key in model["litellm_params"]:
if key != "api_key":
model_id+= str(model["litellm_params"][key])
model["litellm_params"]["model"] += "-ModelID-" + model_id
############ Users can either pass tpm/rpm as a litellm_param or a router param ###########
# for get_available_deployment, we use the litellm_param["rpm"]
# in this snippet we also set rpm to be a litellm_param
if model["litellm_params"].get("rpm") is None and model.get("rpm") is not None:
model["litellm_params"]["rpm"] = model.get("rpm")
if model["litellm_params"].get("tpm") is None and model.get("tpm") is not None:
model["litellm_params"]["tpm"] = model.get("tpm")
self.model_names = [m["model_name"] for m in model_list]
def get_model_names(self):
return self.model_names
def _get_client(self, deployment, kwargs, client_type=None):
"""
Returns the appropriate client based on the given deployment, kwargs, and client_type.
Parameters:
deployment (dict): The deployment dictionary containing the clients.
kwargs (dict): The keyword arguments passed to the function.
client_type (str): The type of client to return.
Returns:
The appropriate client based on the given client_type and kwargs.
"""
if client_type == "async":
if kwargs.get("stream") == True:
return deployment["stream_async_client"]
else:
return deployment["async_client"]
else:
if kwargs.get("stream") == True:
return deployment["stream_client"]
else:
return deployment["client"]
def print_verbose(self, print_statement):
if self.set_verbose or litellm.set_verbose:
print(f"LiteLLM.Router: {print_statement}") # noqa
def get_available_deployment(self,
model: str,
messages: Optional[List[Dict[str, str]]] = None,
input: Optional[Union[str, List]] = None):
"""
Returns the deployment based on routing strategy
"""
## get healthy deployments
### get all deployments
### filter out the deployments currently cooling down
healthy_deployments = [m for m in self.model_list if m["model_name"] == model]
if len(healthy_deployments) == 0:
# check if the user sent in a deployment name instead
healthy_deployments = [m for m in self.model_list if m["litellm_params"]["model"] == model]
self.print_verbose(f"initial list of deployments: {healthy_deployments}")
deployments_to_remove = []
cooldown_deployments = self._get_cooldown_deployments()
self.print_verbose(f"cooldown deployments: {cooldown_deployments}")
### FIND UNHEALTHY DEPLOYMENTS
for deployment in healthy_deployments:
deployment_name = deployment["litellm_params"]["model"]
if deployment_name in cooldown_deployments:
deployments_to_remove.append(deployment)
### FILTER OUT UNHEALTHY DEPLOYMENTS
for deployment in deployments_to_remove:
healthy_deployments.remove(deployment)
self.print_verbose(f"healthy deployments: length {len(healthy_deployments)} {healthy_deployments}")
if len(healthy_deployments) == 0:
raise ValueError("No models available")
if litellm.model_alias_map and model in litellm.model_alias_map:
model = litellm.model_alias_map[
model
] # update the model to the actual value if an alias has been passed in
if self.routing_strategy == "least-busy":
if len(self.healthy_deployments) > 0:
for item in self.healthy_deployments:
if item[0]["model_name"] == model: # first one in queue will be the one with the most availability
return item[0]
else:
raise ValueError("No models available.")
elif self.routing_strategy == "simple-shuffle":
# if users pass rpm or tpm, we do a random weighted pick - based on rpm/tpm
############## Check if we can do a RPM/TPM based weighted pick #################
rpm = healthy_deployments[0].get("litellm_params").get("rpm", None)
if rpm is not None:
# use weight-random pick if rpms provided
rpms = [m["litellm_params"].get("rpm", 0) for m in healthy_deployments]
self.print_verbose(f"\nrpms {rpms}")
total_rpm = sum(rpms)
weights = [rpm / total_rpm for rpm in rpms]
self.print_verbose(f"\n weights {weights}")
# Perform weighted random pick
selected_index = random.choices(range(len(rpms)), weights=weights)[0]
self.print_verbose(f"\n selected index, {selected_index}")
deployment = healthy_deployments[selected_index]
return deployment or deployment[0]
############## Check if we can do a RPM/TPM based weighted pick #################
tpm = healthy_deployments[0].get("litellm_params").get("tpm", None)
if tpm is not None:
# use weight-random pick if rpms provided
tpms = [m["litellm_params"].get("tpm", 0) for m in healthy_deployments]
self.print_verbose(f"\ntpms {tpms}")
total_tpm = sum(tpms)
weights = [tpm / total_tpm for tpm in tpms]
self.print_verbose(f"\n weights {weights}")
# Perform weighted random pick
selected_index = random.choices(range(len(tpms)), weights=weights)[0]
self.print_verbose(f"\n selected index, {selected_index}")
deployment = healthy_deployments[selected_index]
return deployment or deployment[0]
############## No RPM/TPM passed, we do a random pick #################
item = random.choice(healthy_deployments)
return item or item[0]
elif self.routing_strategy == "latency-based-routing":
returned_item = None
lowest_latency = float('inf')
### shuffles with priority for lowest latency
# items_with_latencies = [('A', 10), ('B', 20), ('C', 30), ('D', 40)]
items_with_latencies = []
for item in healthy_deployments:
items_with_latencies.append((item, self.deployment_latency_map[item["litellm_params"]["model"]]))
returned_item = self.weighted_shuffle_by_latency(items_with_latencies)
return returned_item
elif self.routing_strategy == "usage-based-routing":
return self.get_usage_based_available_deployment(model=model, messages=messages, input=input)
raise ValueError("No models available.")
def flush_cache(self):
self.cache.flush_cache()
def reset(self):
## clean up on close
litellm.success_callback = []
litellm.failure_callback = []
self.flush_cache()