[Frontend] OpenAI API server: Do not add bos token by default when encoding (vllm-project#4688)

Author: bofenghuang

vllm/entrypoints/openai/serving_chat.py

@@ -158,7 +158,7 @@ async def create_chat_completion(
         try:
             # Tokenize/detokenize depending on prompt format (string/token list)
             prompt_ids, prompt_text = self._validate_prompt_and_tokenize(
-                request, prompt=prompt)
+                request, prompt=prompt, add_special_tokens=False)
             sampling_params = request.to_sampling_params()
             lora_request = self._maybe_get_lora(request)
             decoding_config = await self.engine.get_decoding_config()

vllm/entrypoints/openai/serving_engine.py

@@ -1,7 +1,7 @@
 import json
 from dataclasses import dataclass
 from http import HTTPStatus
-from typing import Dict, List, Optional, Tuple, Union
+from typing import Any, Dict, List, Optional, Tuple, Union
 
 from pydantic import Field
 from typing_extensions import Annotated
@@ -165,24 +165,34 @@ def _maybe_get_lora(
         raise ValueError(f"The model `{request.model}` does not exist.")
 
     def _validate_prompt_and_tokenize(
-        self,
-        request: Union[ChatCompletionRequest, CompletionRequest,
-                       EmbeddingRequest],
-        prompt: Optional[str] = None,
-        prompt_ids: Optional[List[int]] = None,
-        truncate_prompt_tokens: Optional[Annotated[int, Field(ge=1)]] = None
-    ) -> Tuple[List[int], str]:
+            self,
+            request: Union[ChatCompletionRequest, CompletionRequest,
+                           EmbeddingRequest],
+            prompt: Optional[str] = None,
+            prompt_ids: Optional[List[int]] = None,
+            truncate_prompt_tokens: Optional[Annotated[int,
+                                                       Field(ge=1)]] = None,
+            add_special_tokens: bool = True) -> Tuple[List[int], str]:
         if not (prompt or prompt_ids):
             raise ValueError("Either prompt or prompt_ids should be provided.")
         if (prompt and prompt_ids):
             raise ValueError(
                 "Only one of prompt or prompt_ids should be provided.")
 
         if prompt_ids is None:
-            tokenizer_kwargs = {} if truncate_prompt_tokens is None else {
-                "truncation": True,
-                "max_length": truncate_prompt_tokens,
+            # When using OpenAIServingChat for chat completions, the
+            # special tokens (e.g., BOS) have already been added by the
+            # chat template. Therefore, we do not need to add them again.
+            # Set add_special_tokens to False to avoid adding the BOS tokens
+            # again.
+            tokenizer_kwargs: Dict[str, Any] = {
+                "add_special_tokens": add_special_tokens
             }
+            if truncate_prompt_tokens is not None:
+                tokenizer_kwargs.update({
+                    "truncation": True,
+                    "max_length": truncate_prompt_tokens,
+                })
             input_ids = self.tokenizer(prompt, **tokenizer_kwargs).input_ids
         elif truncate_prompt_tokens is not None:
             input_ids = prompt_ids[-truncate_prompt_tokens:]