move baseten to a REST endpoint call

litellm/__init__.py

@@ -21,6 +21,7 @@ huggingface_key: Optional[str] = None
 vertex_project: Optional[str] = None
 vertex_location: Optional[str] = None
 togetherai_api_key: Optional[str] = None
+baseten_key: Optional[str] = None
 caching = False
 caching_with_models = False  # if you want the caching key to be model + prompt
 model_alias_map: Dict[str, str] = {}

litellm/integrations/litedebugger.py

@@ -29,7 +29,7 @@ class LiteDebugger:
             )
 
     def input_log_event(
-        self, model, messages, end_user, litellm_call_id, print_verbose
+        self, model, messages, end_user, litellm_call_id, print_verbose, litellm_params, optional_params
     ):
         try:
             print_verbose(
@@ -42,6 +42,8 @@ class LiteDebugger:
                 "status": "initiated",
                 "litellm_call_id": litellm_call_id,
                 "user_email": self.user_email,
+                "litellm_params": litellm_params,
+                "optional_params": optional_params
             }
             response = requests.post(
                 url=self.api_url,

litellm/llms/baseten.py

@@ -0,0 +1,129 @@
+import os, json
+from enum import Enum
+import requests
+import time
+from typing import Callable
+from litellm.utils import ModelResponse
+
+class BasetenError(Exception):
+    def __init__(self, status_code, message):
+        self.status_code = status_code
+        self.message = message
+        super().__init__(
+            self.message
+        )  # Call the base class constructor with the parameters it needs
+
+
+class BasetenLLM:
+    def __init__(
+        self, encoding, logging_obj, api_key=None
+    ):
+        self.encoding = encoding
+        self.completion_url_fragment_1 = "https://app.baseten.co/models/"
+        self.completion_url_fragment_2 = "/predict"
+        self.api_key = api_key
+        self.logging_obj = logging_obj
+        self.validate_environment(api_key=api_key)
+
+    def validate_environment(
+        self, api_key
+    ):  # set up the environment required to run the model
+        # set the api key
+        if self.api_key == None:
+            raise ValueError(
+                "Missing Baseten API Key - A call is being made to baseten but no key is set either in the environment variables or via params"
+            )
+        self.api_key = api_key
+        self.headers = {
+            "accept": "application/json",
+            "content-type": "application/json",
+            "Authorization": "Api-Key " + self.api_key,
+        }
+
+    def completion(
+        self,
+        model: str,
+        messages: list,
+        model_response: ModelResponse,
+        print_verbose: Callable,
+        optional_params=None,
+        litellm_params=None,
+        logger_fn=None,
+    ):  # logic for parsing in - calling - parsing out model completion calls
+        model = model
+        prompt = ""
+        for message in messages:
+            if "role" in message:
+                if message["role"] == "user":
+                    prompt += (
+                        f"{message['content']}"
+                    )
+                else:
+                    prompt += (
+                        f"{message['content']}"
+                    )
+            else:
+                prompt += f"{message['content']}"
+        data = {
+            "prompt": prompt,
+            # "instruction": prompt, # some baseten models require the prompt to be passed in via the 'instruction' kwarg
+            # **optional_params,
+        }
+
+        ## LOGGING
+        self.logging_obj.pre_call(
+            input=prompt,
+            api_key=self.api_key,
+            additional_args={"complete_input_dict": data},
+        )
+        ## COMPLETION CALL
+        response = requests.post(
+            self.completion_url_fragment_1 + model + self.completion_url_fragment_2, headers=self.headers, data=json.dumps(data)
+        )
+        if "stream" in optional_params and optional_params["stream"] == True:
+            return response.iter_lines()
+        else:
+            ## LOGGING
+            self.logging_obj.post_call(
+                input=prompt,
+                api_key=self.api_key,
+                original_response=response.text,
+                additional_args={"complete_input_dict": data},
+            )
+            print_verbose(f"raw model_response: {response.text}")
+            ## RESPONSE OBJECT
+            completion_response = response.json()
+            if "error" in completion_response:
+                raise BasetenError(
+                    message=completion_response["error"],
+                    status_code=response.status_code,
+                )
+            else:
+                if "model_output" in completion_response:
+                    if isinstance(completion_response["model_output"], str):
+                        model_response["choices"][0]["message"]["content"] = completion_response["model_output"]
+                    elif isinstance(completion_response["model_output"], dict) and "data" in completion_response["model_output"] and isinstance(completion_response["model_output"]["data"], list):
+                        model_response["choices"][0]["message"]["content"] = completion_response["model_output"]["data"][0]
+                    else:
+                        raise ValueError(f"Unable to parse response. Original response: {response.text}")
+            ## CALCULATING USAGE - baseten charges on time, not tokens - have some mapping of cost here. 
+            prompt_tokens = len(
+                self.encoding.encode(prompt)
+            ) 
+            completion_tokens = len(
+                self.encoding.encode(model_response["choices"][0]["message"]["content"])
+            )
+
+            model_response["created"] = time.time()
+            model_response["model"] = model
+            model_response["usage"] = {
+                "prompt_tokens": prompt_tokens,
+                "completion_tokens": completion_tokens,
+                "total_tokens": prompt_tokens + completion_tokens,
+            }
+            return model_response
+
+    def embedding(
+        self,
+    ):  # logic for parsing in - calling - parsing out model embedding calls
+        pass

litellm/main.py

@@ -21,6 +21,7 @@ from litellm.utils import (
 )
 from .llms.anthropic import AnthropicLLM
 from .llms.huggingface_restapi import HuggingfaceRestAPILLM
+from .llms.baseten import BasetenLLM
 import tiktoken
 from concurrent.futures import ThreadPoolExecutor
 
@@ -73,6 +74,7 @@ def completion(
     max_tokens=float("inf"),
     presence_penalty=0,
     frequency_penalty=0,
+    num_beams=1,
     logit_bias={},
     user="",
     deployment_id=None,
@@ -681,36 +683,31 @@ def completion(
             custom_llm_provider == "baseten"
             or litellm.api_base == "https://app.baseten.co"
         ):
-            import baseten
+            custom_llm_provider = "baseten"
-
+            baseten_key = (
-            base_ten_key = get_secret("BASETEN_API_KEY")
+                api_key
-            baseten.login(base_ten_key)
+                or litellm.baseten_key
-
+                or os.environ.get("BASETEN_API_KEY")
-            prompt = " ".join([message["content"] for message in messages])
-            ## LOGGING
-            logging.pre_call(input=prompt, api_key=base_ten_key, model=model)
-
-            base_ten__model = baseten.deployed_model_version_id(model)
-
-            completion_response = base_ten__model.predict({"prompt": prompt})
-            if type(completion_response) == dict:
-                completion_response = completion_response["data"]
-                if type(completion_response) == dict:
-                    completion_response = completion_response["generated_text"]
-
-            ## LOGGING
-            logging.post_call(
-                input=prompt,
-                api_key=base_ten_key,
-                original_response=completion_response,
             )
-
+            baseten_client = BasetenLLM(
-            ## RESPONSE OBJECT
+                encoding=encoding, api_key=baseten_key, logging_obj=logging
-            model_response["choices"][0]["message"]["content"] = completion_response
+            )
-            model_response["created"] = time.time()
+            model_response = baseten_client.completion(
-            model_response["model"] = model
+                model=model,
+                messages=messages,
+                model_response=model_response,
+                print_verbose=print_verbose,
+                optional_params=optional_params,
+                litellm_params=litellm_params,
+                logger_fn=logger_fn,
+            )
+            if "stream" in optional_params and optional_params["stream"] == True:
+                # don't try to access stream object,
+                response = CustomStreamWrapper(
+                    model_response, model, custom_llm_provider="huggingface"
+                )
+                return response
             response = model_response
-
         elif custom_llm_provider == "petals" or (
             litellm.api_base and "chat.petals.dev" in litellm.api_base
         ):

litellm/tests/test_model_alias_map.py

@@ -13,4 +13,5 @@ from litellm import embedding, completion
 litellm.set_verbose = True
 
 # Test: Check if the alias created via LiteDebugger is mapped correctly
-print(completion("wizard-lm", messages=[{"role": "user", "content": "Hey, how's it going?"}]))
+{"top_p": 0.75, "prompt": "What's the meaning of life?", "num_beams": 4, "temperature": 0.1}
+print(completion("llama-7b", messages=[{"role": "user", "content": "Hey, how's it going?"}], top_p=0.1, temperature=0, num_beams=4, max_tokens=60))

litellm/utils.py

@@ -154,6 +154,7 @@ class Logging:
         self.optional_params = optional_params
         self.litellm_params = litellm_params
         self.logger_fn = litellm_params["logger_fn"]
+        print_verbose(f"self.optional_params: {self.optional_params}")
         self.model_call_details = {
             "model": model,
             "messages": messages,
@@ -214,6 +215,8 @@ class Logging:
                             end_user=litellm._thread_context.user,
                             litellm_call_id=self.
                             litellm_params["litellm_call_id"],
+                            litellm_params=self.model_call_details["litellm_params"],
+                            optional_params=self.model_call_details["optional_params"],
                             print_verbose=print_verbose,
                         )
                 except Exception as e:
@@ -539,7 +542,7 @@ def get_litellm_params(
     return litellm_params
 
 
-def get_optional_params(
+def get_optional_params( # use the openai defaults
     # 12 optional params
     functions=[],
     function_call="",
@@ -552,6 +555,7 @@ def get_optional_params(
     presence_penalty=0,
     frequency_penalty=0,
     logit_bias={},
+    num_beams=1, 
     user="",
     deployment_id=None,
     model=None,
@@ -613,7 +617,13 @@ def get_optional_params(
         optional_params["temperature"] = temperature
         optional_params["top_p"] = top_p
         optional_params["top_k"] = top_k
-
+    elif custom_llm_provider == "baseten":
+        optional_params["temperature"] = temperature
+        optional_params["top_p"] = top_p
+        optional_params["top_k"] = top_k
+        optional_params["num_beams"] = num_beams
+        if max_tokens != float("inf"):
+            optional_params["max_new_tokens"] = max_tokens
     else:  # assume passing in params for openai/azure openai
         if functions != []:
             optional_params["functions"] = functions

pyproject.toml

@@ -1,6 +1,6 @@
 [tool.poetry]
 name = "litellm"
-version = "0.1.478"
+version = "0.1.479"
 description = "Library to easily interface with LLM API providers"
 authors = ["BerriAI"]
 license = "MIT License"