rebase and merge

Signed-off-by: Keyun Tong <[email protected]>

vllm/entrypoints/llm.py

@@ -1174,27 +1174,28 @@ def ensure_str(prompt: SingletonPrompt):
         if isinstance(text_1, (str, dict)):
             # Convert a single prompt to a list.
             text_1 = [text_1]
-        text_1 = [ensure_str(t) for t in text_1]
+        input_text_1: List[str] = [ensure_str(t) for t in text_1]
 
         if isinstance(text_2, (str, dict)):
             # Convert a single prompt to a list.
             text_2 = [text_2]
-        text_2 = [ensure_str(t) for t in text_2]
+        input_text_2: List[str] = [ensure_str(t) for t in text_2]
 
-        if len(text_1) > 1 and len(text_1) != len(text_2):
+        if len(input_text_1) > 1 and len(input_text_1) != len(input_text_2):
             raise ValueError("Input lengths must be either 1:1, 1:N or N:N")
-        if len(text_1) == 0:
+        if len(input_text_1) == 0:
             raise ValueError("At least one text element must be given")
-        if len(text_2) == 0:
+        if len(input_text_2) == 0:
             raise ValueError("At least one text_pair element must be given")
 
         if self.llm_engine.model_config.is_cross_encoder:
-            return self._cross_encoding_score(tokenizer, text_1, text_2,
+            return self._cross_encoding_score(tokenizer, input_text_1,
+                                              input_text_2,
                                               truncate_prompt_tokens, use_tqdm,
                                               lora_request,
                                               prompt_adapter_request)
         else:
-            return self._embedding_score(tokenizer, text_1, text_2,
+            return self._embedding_score(tokenizer, input_text_1, input_text_2,
                                          truncate_prompt_tokens, use_tqdm,
                                          lora_request, prompt_adapter_request)
 

vllm/entrypoints/openai/serving_score.py

@@ -110,7 +110,7 @@ async def create_score(
 
             input_pairs = make_pairs(request.text_1, request.text_2)
             for q, t in input_pairs:
-                request_prompt = f"{q}{tokenizer.sep_token}{t}"
+                request_prompt = f"{q}{tokenizer.sep_token_id}{t}"
 
                 tokenization_kwargs: Dict[str, Any] = {}
                 if truncate_prompt_tokens is not None:

vllm/transformers_utils/tokenizer.py

@@ -47,11 +47,7 @@ def encode_tokens(
     Backend-agnostic equivalent of HF's
     :code:`tokenizer.encode(text, add_special_tokens=...)`.
     """
-    if isinstance(tokenizer, MistralTokenizer):
-        return tokenizer.tokenizer.encode(text,
-                                          bos=add_special_tokens,
-                                          eos=add_special_tokens)
-    elif add_special_tokens is not None:
+    if add_special_tokens is not None:
         return tokenizer.encode(text, add_special_tokens=add_special_tokens)
     return tokenizer.encode(text)
 

vllm/transformers_utils/tokenizers/mistral.py

@@ -207,7 +207,7 @@ def bos_token_id(self) -> int:
     @property
     def eos_token_id(self) -> int:
         return self.tokenizer.eos_id
-    
+
     @property
     def sep_token_id(self) -> int:
         raise NotImplementedError()
@@ -283,7 +283,12 @@ def encode(self,
         # `encode` should only be used for prompt completion
         # it should never be used for chat_completion.
         # For chat completion use `apply_chat_template`
-        return self.tokenizer.encode(text, bos=True, eos=False)
+        if add_special_tokens is not None:
+            return self.tokenizer.encode(text,
+                                         bos=add_special_tokens,
+                                         eos=add_special_tokens)
+        else:
+            return self.tokenizer.encode(text, bos=True, eos=False)
 
     def apply_chat_template(self,
                             messages: List["ChatCompletionMessageParam"],