feat - try using hf tokenizer

litellm/proxy/_types.py

@@ -79,6 +79,8 @@ class LiteLLMRoutes(enum.Enum):
         "/v1/models",
     ]
 
+    llm_utils_routes: List = ["utils/token_counter"]
+
     info_routes: List = [
         "/key/info",
         "/team/info",
@@ -1012,3 +1014,5 @@ class TokenCountRequest(LiteLLMBase):
 class TokenCountResponse(LiteLLMBase):
     total_tokens: int
     model: str
+    base_model: str
+    tokenizer_type: str

litellm/proxy/proxy_server.py

@@ -4775,21 +4775,38 @@ async def token_counter(request: TokenCountRequest):
     """ """
     from litellm import token_counter
 
+    global llm_router
+
     prompt = request.prompt
     messages = request.messages
 
+    if llm_router is not None:
+        # get 1 deployment corresponding to the model
+        for _model in llm_router.model_list:
+            if _model["model_name"] == request.model:
+                deployment = _model
+                break
+
+    litellm_model_name = deployment.get("litellm_params", {}).get("model")
+    # remove the custom_llm_provider_prefix in the litellm_model_name
+    if "/" in litellm_model_name:
+        litellm_model_name = litellm_model_name.split("/", 1)[1]
+
     if prompt is None and messages is None:
         raise HTTPException(
             status_code=400, detail="prompt or messages must be provided"
         )
-    total_tokens = token_counter(
-        model=request.model,
+    total_tokens, tokenizer_used = token_counter(
+        model=litellm_model_name,
         text=prompt,
         messages=messages,
+        return_tokenizer_used=True,
     )
     return TokenCountResponse(
         total_tokens=total_tokens,
         model=request.model,
+        base_model=litellm_model_name,
+        tokenizer_type=tokenizer_used,
     )
 
 

litellm/utils.py

@@ -3860,6 +3860,11 @@ def _select_tokenizer(model: str):
         return {"type": "huggingface_tokenizer", "tokenizer": tokenizer}
     # default - tiktoken
     else:
+        tokenizer = None
+        try:
+            tokenizer = Tokenizer.from_pretrained(model)
+            return {"type": "huggingface_tokenizer", "tokenizer": tokenizer}
+        except:
             return {"type": "openai_tokenizer", "tokenizer": encoding}
 
 
@@ -4097,6 +4102,7 @@ def token_counter(
     text: Optional[Union[str, List[str]]] = None,
     messages: Optional[List] = None,
     count_response_tokens: Optional[bool] = False,
+    return_tokenizer_used: Optional[bool] = False,
 ):
     """
     Count the number of tokens in a given text using a specified model.
@@ -4189,7 +4195,10 @@ def token_counter(
                 )
     else:
         num_tokens = len(encoding.encode(text, disallowed_special=()))  # type: ignore
-
+    _tokenizer_type = tokenizer_json["type"]
+    if return_tokenizer_used:
+        # used by litellm proxy server ->  POST /utils/token_counter
+        return num_tokens, _tokenizer_type
     return num_tokens