diff --git a/vllm/entrypoints/openai/chat_completion/serving.py b/vllm/entrypoints/openai/chat_completion/serving.py index 21bc0f4424..48fb666484 100644 --- a/vllm/entrypoints/openai/chat_completion/serving.py +++ b/vllm/entrypoints/openai/chat_completion/serving.py @@ -679,6 +679,7 @@ class OpenAIServingChat(OpenAIServing): # For reasoning parser and tool call all enabled added_content_delta_arr = [False] * num_choices reasoning_end_arr = [False] * num_choices + prompt_is_reasoning_end_arr: list[bool | None] = [None] * num_choices else: all_previous_token_ids = None @@ -824,6 +825,16 @@ class OpenAIServingChat(OpenAIServing): i = output.index tool_parser = tool_parsers[i] + if ( + self.reasoning_parser + and res.prompt_token_ids + and prompt_is_reasoning_end_arr[i] is None + ): + # only check once per choice, because prompt_token_ids + # are the same for all deltas in that choice + prompt_is_reasoning_end_arr[i] = ( + reasoning_parser.is_reasoning_end(res.prompt_token_ids) + ) if finish_reason_sent[i]: continue @@ -926,13 +937,11 @@ class OpenAIServingChat(OpenAIServing): # i.e {"enable_thinking": False}, # set reasoning status to end. # Only keep 'content', remove 'reasoning'. - if reasoning_parser.is_reasoning_end( - as_list(output.token_ids) - ) or ( - res.prompt_token_ids - and reasoning_parser.is_reasoning_end( - res.prompt_token_ids + if ( + reasoning_parser.is_reasoning_end( + as_list(output.token_ids) ) + or prompt_is_reasoning_end_arr[i] ): reasoning_end_arr[i] = True if delta_message and delta_message.content: @@ -991,8 +1000,7 @@ class OpenAIServingChat(OpenAIServing): if ( self.reasoning_parser is not None and not reasoning_end_arr[i] - and res.prompt_token_ids - and reasoning_parser.is_reasoning_end(res.prompt_token_ids) + and prompt_is_reasoning_end_arr[i] ): reasoning_end_arr[i] = True @@ -1049,12 +1057,7 @@ class OpenAIServingChat(OpenAIServing): # When encountering think end id in prompt_token_ids # i.e {"enable_thinking": False}, # set reasoning status to end. - if ( - res.prompt_token_ids - and reasoning_parser.is_reasoning_end( - res.prompt_token_ids - ) - ): + if prompt_is_reasoning_end_arr[i]: reasoning_end_arr[i] = True current_token_ids = output_token_ids # Don't update current_text, keep it as is from delta