可靠性 - 重試、備援
LiteLLM 可透過 2 種方式防止請求失敗:
- 重試
- 備援:Context Window + General
輔助工具
LiteLLM 支援以下用於可靠性的函式:
litellm.longer_context_model_fallback_dict:具有對應關係的字典,對應於那些有更大等效模型的模型num_retries:使用 tenacity 重試completion()搭配備援:在發生錯誤時,在模型/金鑰/API base 之間切換。
重試失敗的請求
像這樣在 completion 中呼叫它 completion(..num_retries=2)。
以下簡單看看您可以如何使用它:
from litellm import completion
user_message = "Hello, whats the weather in San Francisco??"
messages = [{"content": user_message, "role": "user"}]
# normal call
response = completion(
model="gpt-3.5-turbo",
messages=messages,
num_retries=2
)
備援(SDK)
Context Window 備援(SDK)
from litellm import completion
fallback_dict = {"gpt-3.5-turbo": "gpt-3.5-turbo-16k"}
messages = [{"content": "how does a court case get to the Supreme Court?" * 500, "role": "user"}]
completion(model="gpt-3.5-turbo", messages=messages, context_window_fallback_dict=fallback_dict)
備援 - 切換模型/API 金鑰/API Bases(SDK)
LLM APIs 可能不穩定,帶有備援的 completion() 可確保您從請求中始終取得回應
用法
若要搭配 completion() 使用備援模型,請在 fallbacks 參數中指定模型清單。
fallbacks 清單應包含您要使用的主要模型,接著是其他可在主要模型無法提供回應時作為備份使用的模型。
切換模型
response = completion(model="bad-model", messages=messages,
fallbacks=["gpt-3.5-turbo" "command-nightly"])
切換 api keys/bases(例如 Azure deployment)
在同一個 Azure deployment 之間切換不同的金鑰,或也使用另一個 deployment。
api_key="bad-key"
response = completion(model="azure/gpt-4", messages=messages, api_key=api_key,
fallbacks=[{"api_key": "good-key-1"}, {"api_key": "good-key-2", "api_base": "good-api-base-2"}])
實作細節(SDK)
備援
呼叫的回應
Completion with 'bad-model': got exception Unable to map your input to a model. Check your input - {'model': 'bad-model'
completion call gpt-3.5-turbo
{
"id": "chatcmpl-7qTmVRuO3m3gIBg4aTmAumV1TmQhB",
"object": "chat.completion",
"created": 1692741891,
"model": "gpt-3.5-turbo-0613",
"choices": [
{
"index": 0,
"message": {
"role": "assistant",
"content": "I apologize, but as an AI, I do not have the capability to provide real-time weather updates. However, you can easily check the current weather in San Francisco by using a search engine or checking a weather website or app."
},
"finish_reason": "stop"
}
],
"usage": {
"prompt_tokens": 16,
"completion_tokens": 46,
"total_tokens": 62
}
}
備援如何運作
當您將 fallbacks 傳入 completion 時,它會先使用在 completion(model=model) 中指定為 model 的主要模型,進行第一次 completion 呼叫。如果主要模型失敗或發生錯誤,它會依指定順序自動嘗試 fallbacks 模型。這可確保即使主要模型無法使用,仍能取得回應。
Model Fallbacks 實作的關鍵元件:
- 依序遍歷
fallbacks - 速率限制模型的冷卻時間
依序遍歷 fallbacks
允許每個請求有 45seconds。在這 45 秒內,此函式會嘗試呼叫設定為 model 的主要模型。若模型失敗,它會依序遍歷備份的 fallbacks 模型,並嘗試在此處設定的分配 45s 時間內取得回應:
while response == None and time.time() - start_time < 45:
for model in fallbacks:
速率限制模型的冷卻時間
如果某次模型 API 呼叫導致錯誤,則允許其冷卻 60s
except Exception as e:
print(f"got exception {e} for model {model}")
rate_limited_models.add(model)
model_expiration_times[model] = (
time.time() + 60
) # cool down this selected model
pass
在進行 LLM API 呼叫之前,我們會檢查所選模型是否位於 rate_limited_models 中,若是則略過 API 呼叫
if (
model in rate_limited_models
): # check if model is currently cooling down
if (
model_expiration_times.get(model)
and time.time() >= model_expiration_times[model]
):
rate_limited_models.remove(
model
) # check if it's been 60s of cool down and remove model
else:
continue # skip model
具有備援的 completion() 完整程式碼
response = None
rate_limited_models = set()
model_expiration_times = {}
start_time = time.time()
fallbacks = [kwargs["model"]] + kwargs["fallbacks"]
del kwargs["fallbacks"] # remove fallbacks so it's not recursive
while response == None and time.time() - start_time < 45:
for model in fallbacks:
# loop thru all models
try:
if (
model in rate_limited_models
): # check if model is currently cooling down
if (
model_expiration_times.get(model)
and time.time() >= model_expiration_times[model]
):
rate_limited_models.remove(
model
) # check if it's been 60s of cool down and remove model
else:
continue # skip model
# delete model from kwargs if it exists
if kwargs.get("model"):
del kwargs["model"]
print("making completion call", model)
response = litellm.completion(**kwargs, model=model)
if response != None:
return response
except Exception as e:
print(f"got exception {e} for model {model}")
rate_limited_models.add(model)
model_expiration_times[model] = (
time.time() + 60
) # cool down this selected model
pass
return response