Retry 503 OpenAI errors

autogpt/llm/utils/__init__.py

@@ -36,8 +36,8 @@
                 response.usage.completion_tokens if "completion_tokens" in usage else 0,
                 response.model,
             )
         except Exception as err:
             logger.warn(f"Failed to update API costs: {err.__class__.__name__}: {err}")
 
     def metering_wrapper(*args, **kwargs):
         openai_obj = openai_obj_processor(*args, **kwargs)
@@ -89,20 +89,22 @@
 
                 except RateLimitError:
                     if attempt == num_attempts:
                         raise
 
                     logger.debug(retry_limit_msg)
                     if not user_warned:
                         logger.double_check(api_key_error_msg)
                         user_warned = True
 
                 except APIError as e:
-                    if (e.http_status not in [502, 429]) or (attempt == num_attempts):
+                    if (e.http_status not in [429, 502, 503]) or (
+                        attempt == num_attempts
+                    ):
                         raise
 
                 backoff = backoff_base ** (attempt + 2)
                 logger.debug(backoff_msg.format(backoff=backoff))
                 time.sleep(backoff)
 
         return _wrapped
 
@@ -131,13 +133,13 @@
         str: The response from the function
     """
     if model is None:
         model = config.smart_llm_model
     # For each arg, if any are None, convert to "None":
     args = [str(arg) if arg is not None else "None" for arg in args]
     # parse args to comma separated string
     arg_str: str = ", ".join(args)
 
     prompt = ChatSequence.for_model(
         model,
         [
             Message(
@@ -148,7 +150,7 @@
             Message("user", arg_str),
         ],
     )
     return create_chat_completion(prompt=prompt, temperature=0)
 
 
 @metered
@@ -159,25 +161,25 @@
     temperature: Optional[float],
     max_output_tokens: Optional[int],
 ) -> str:
     cfg = Config()
     if model is None:
         model = cfg.fast_llm_model
     if temperature is None:
         temperature = cfg.temperature
 
     if cfg.use_azure:
         kwargs = {"deployment_id": cfg.get_azure_deployment_id_for_model(model)}
     else:
         kwargs = {"model": model}
 
     response = openai.Completion.create(
         **kwargs,
         prompt=prompt,
         temperature=temperature,
         max_tokens=max_output_tokens,
         api_key=cfg.openai_api_key,
     )
     return response.choices[0].text
 
 
 # Overly simple abstraction until we create something better
@@ -219,19 +221,19 @@
             temperature=temperature,
             max_tokens=max_tokens,
         ):
             message = plugin.handle_chat_completion(
                 messages=prompt.raw(),
                 model=model,
                 temperature=temperature,
                 max_tokens=max_tokens,
             )
             if message is not None:
                 return message
     api_manager = ApiManager()
     response = None
 
     if cfg.use_azure:
         kwargs = {"deployment_id": cfg.get_azure_deployment_id_for_model(model)}
     else:
         kwargs = {"model": model}
 
@@ -245,8 +247,8 @@
     resp = response.choices[0].message.content
     for plugin in cfg.plugins:
         if not plugin.can_handle_on_response():
             continue
         resp = plugin.on_response(resp)
     return resp