fix(openai.p): adding support for exception mapping for openai-compatible apis via http calls

litellm/llms/openai.py

@@ -246,30 +246,53 @@ class OpenAIChatCompletion(BaseLLM):
                logger_fn=None):
         super().completion()
         headers = self.validate_environment(api_key=api_key)
-        data = {
-            "messages": messages, 
-            **optional_params
-        }
-        if "stream" in optional_params and optional_params["stream"] == True:
-            response = self._client_session.post(
-                url=f"{api_base}/chat/completions",
-                json=data,
-                headers=headers,
-                stream=optional_params["stream"]
-            )
-            if response.status_code != 200:
-                raise CustomOpenAIError(status_code=response.status_code, message=response.text)
-                
-            ## RESPONSE OBJECT
-            return response.iter_lines()
-        else:
-            response = self._client_session.post(
-                url=f"{api_base}/chat/completions",
-                json=data,
-                headers=headers,
-            )
-            if response.status_code != 200:
-                raise CustomOpenAIError(status_code=response.status_code, message=response.text)
-                
-            ## RESPONSE OBJECT
-            return self.convert_to_model_response_object(response_object=response.json(), model_response_object=model_response)
+
+        for _ in range(2): # if call fails due to alternating messages, retry with reformatted message
+            data = {
+                "model": model,
+                "messages": messages, 
+                **optional_params
+            }
+            try: 
+                if "stream" in optional_params and optional_params["stream"] == True:
+                    response = self._client_session.post(
+                        url=f"{api_base}/chat/completions",
+                        json=data,
+                        headers=headers,
+                        stream=optional_params["stream"]
+                    )
+                    if response.status_code != 200:
+                        raise CustomOpenAIError(status_code=response.status_code, message=response.text)
+                        
+                    ## RESPONSE OBJECT
+                    return response.iter_lines()
+                else:
+                    response = self._client_session.post(
+                        url=f"{api_base}/chat/completions",
+                        json=data,
+                        headers=headers,
+                    )
+                    if response.status_code != 200:
+                        raise CustomOpenAIError(status_code=response.status_code, message=response.text)
+                        
+                    ## RESPONSE OBJECT
+                    return self.convert_to_model_response_object(response_object=response.json(), model_response_object=model_response)
+            except Exception as e:
+                if "Conversation roles must alternate user/assistant" in str(e) or "user and assistant roles should be alternating" in str(e): 
+                    # reformat messages to ensure user/assistant are alternating, if there's either 2 consecutive 'user' messages or 2 consecutive 'assistant' message, add a blank 'user' or 'assistant' message to ensure compatibility
+                    new_messages = []
+                    for i in range(len(messages)-1): 
+                        new_messages.append(messages[i])
+                        if messages[i]["role"] == messages[i+1]["role"]:
+                            if messages[i]["role"] == "user":
+                                new_messages.append({"role": "assistant", "content": ""})
+                            else:
+                                new_messages.append({"role": "user", "content": ""})
+                    new_messages.append(messages[-1])
+                    messages = new_messages
+                elif "Last message must have role `user`" in str(e):
+                    new_messages = messages
+                    new_messages.append({"role": "user", "content": ""})
+                    messages = new_messages
+                else:
+                    raise e

litellm/main.py

@@ -445,7 +445,7 @@ def completion(
                 raise e
             
             if "stream" in optional_params and optional_params["stream"] == True:
-                response = CustomStreamWrapper(response, model, custom_llm_provider="openai", logging_obj=logging)
+                response = CustomStreamWrapper(response, model, custom_llm_provider=custom_llm_provider, logging_obj=logging)
                 return response
             ## LOGGING
             logging.post_call(

litellm/proxy/proxy_server.py

@@ -332,7 +332,6 @@ def logger(
     end_time=None    # start/end time
 ):
   log_event_type = kwargs['log_event_type']
-  print(f"REACHES LOGGER: {log_event_type}")
   try: 
     if log_event_type == 'pre_api_call':
         inference_params = copy.deepcopy(kwargs)
@@ -355,7 +354,6 @@ def logger(
         with open(log_file, 'w') as f:
             json.dump(existing_data, f, indent=2)
     elif log_event_type == 'post_api_call':
-        print(f"post api call kwargs: {kwargs}")
         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')
@@ -438,7 +436,6 @@ async def completion(request: Request):
 @router.post("/chat/completions")
 async def chat_completion(request: Request):
     data = await request.json()
-    print(f"data passed in: {data}")
     response = litellm_completion(data, type="chat_completion")
     return response
 

litellm/tests/test_completion.py

@@ -108,6 +108,28 @@ def test_completion_with_litellm_call_id():
     except Exception as e:
         pytest.fail(f"Error occurred: {e}")
 
+def test_completion_perplexity_api():
+    try:
+        litellm.set_verbose=True
+        messages=[{
+            "role": "system", 
+            "content": "You're a good bot"
+        },{
+            "role": "user", 
+            "content": "Hey", 
+        },{
+            "role": "user", 
+            "content": "Hey", 
+        }]
+        response = completion(
+            model="mistral-7b-instruct", 
+            messages=messages,
+            api_base="https://api.perplexity.ai")
+        print(response)
+    except Exception as e:
+        pytest.fail(f"Error occurred: {e}")
+
+test_completion_perplexity_api()
 # commenting out as this is a flaky test on circle ci
 # def test_completion_nlp_cloud():
 #     try:

litellm/utils.py

@@ -1419,7 +1419,9 @@ def get_llm_provider(model: str, custom_llm_provider: Optional[str] = None, api_
         if api_base: 
             for endpoint in litellm.openai_compatible_endpoints:
                 if endpoint in api_base:
-                    custom_llm_provider = "openai"
+                    custom_llm_provider = "custom_openai"
+                    if endpoint == "api.perplexity.ai": 
+                        litellm.api_key = os.getenv("PERPLEXITYAI_API_KEY")
                     return model, custom_llm_provider
 
         # check if model in known model provider list  -> for huggingface models, raise exception as they don't have a fixed provider (can be togetherai, anyscale, baseten, runpod, et.)
@@ -2936,6 +2938,45 @@ def exception_type(
             elif custom_llm_provider == "ollama":
                 if "no attribute 'async_get_ollama_response_stream" in error_str:
                     raise ImportError("Import error - trying to use async for ollama. import async_generator failed. Try 'pip install async_generator'")
+            elif custom_llm_provider == "custom_openai":
+                if hasattr(original_exception, "status_code"):
+                    exception_mapping_worked = True
+                    if original_exception.status_code == 401:
+                        exception_mapping_worked = True
+                        raise AuthenticationError(
+                            message=f"CustomOpenAIException - {original_exception.message}",
+                            llm_provider="custom_openai",
+                            model=model
+                        )
+                    elif original_exception.status_code == 408:
+                        exception_mapping_worked = True
+                        raise Timeout(
+                            message=f"CustomOpenAIException - {original_exception.message}",
+                            model=model,
+                            llm_provider="custom_openai"
+                        )
+                    if original_exception.status_code == 422:
+                        exception_mapping_worked = True
+                        raise InvalidRequestError(
+                            message=f"CustomOpenAIException - {original_exception.message}",
+                            model=model,
+                            llm_provider="custom_openai",
+                        )
+                    elif original_exception.status_code == 429:
+                        exception_mapping_worked = True
+                        raise RateLimitError(
+                            message=f"CustomOpenAIException - {original_exception.message}",
+                            model=model,
+                            llm_provider="custom_openai",
+                        )
+                    else:
+                        exception_mapping_worked = True
+                        raise APIError(
+                            status_code=original_exception.status_code, 
+                            message=f"CustomOpenAIException - {original_exception.message}",
+                            llm_provider="custom_openai",
+                            model=model
+                        )
         exception_mapping_worked = True
         raise APIError(status_code=500, message=str(original_exception), llm_provider=custom_llm_provider, model=model)
     except Exception as e:
@@ -3205,6 +3246,30 @@ class CustomStreamWrapper:
         except:
             raise ValueError(f"Unable to parse response. Original response: {chunk}")
     
+    def handle_custom_openai_chat_completion_chunk(self, chunk): 
+        try: 
+            str_line = chunk.decode("utf-8")  # Convert bytes to string
+            text = "" 
+            is_finished = False
+            finish_reason = None
+            if str_line.startswith("data:"):
+                data_json = json.loads(str_line[5:])
+                print(f"delta content: {data_json['choices'][0]['delta']}")
+                text = data_json["choices"][0]["delta"].get("content", "") 
+                if data_json["choices"][0].get("finish_reason", None): 
+                    is_finished = True
+                    finish_reason = data_json["choices"][0]["finish_reason"]
+                return {"text": text, "is_finished": is_finished, "finish_reason": finish_reason}
+            elif "error" in str_line:
+                raise ValueError(f"Unable to parse response. Original response: {str_line}")
+            else:
+                return {"text": text, "is_finished": is_finished, "finish_reason": finish_reason}
+
+        except:
+            traceback.print_exc()
+            pass
+
+
     def handle_openai_text_completion_chunk(self, chunk):
         try:
             return chunk["choices"][0]["text"]
@@ -3401,6 +3466,13 @@ class CustomStreamWrapper:
                     if "error" in chunk:
                         exception_type(model=self.model, custom_llm_provider=self.custom_llm_provider, original_exception=chunk["error"])
                     completion_obj = chunk
+                elif self.custom_llm_provider == "custom_openai":
+                    chunk = next(self.completion_stream)
+                    response_obj = self.handle_custom_openai_chat_completion_chunk(chunk)
+                    completion_obj["content"] = response_obj["text"]
+                    print(f"completion obj content: {completion_obj['content']}")
+                    if response_obj["is_finished"]: 
+                        model_response.choices[0].finish_reason = response_obj["finish_reason"]
                 else: # openai chat/azure models
                     chunk = next(self.completion_stream)
                     model_response = chunk