diff --git a/vllm/reasoning/deepseek_v3_reasoning_parser.py b/vllm/reasoning/deepseek_v3_reasoning_parser.py index d2f7f50a32..bb79afd8de 100644 --- a/vllm/reasoning/deepseek_v3_reasoning_parser.py +++ b/vllm/reasoning/deepseek_v3_reasoning_parser.py @@ -40,6 +40,14 @@ class DeepSeekV3ReasoningParser(ReasoningParser): else: self._parser = IdentityReasoningParser(tokenizer, *args, **kwargs) + @property + def reasoning_start_str(self) -> str | None: + return self._parser.reasoning_start_str + + @property + def reasoning_end_str(self) -> str | None: + return self._parser.reasoning_end_str + def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: return self._parser.is_reasoning_end(input_ids) diff --git a/vllm/reasoning/identity_reasoning_parser.py b/vllm/reasoning/identity_reasoning_parser.py index b02a9d3184..c6f117e2f9 100644 --- a/vllm/reasoning/identity_reasoning_parser.py +++ b/vllm/reasoning/identity_reasoning_parser.py @@ -33,6 +33,14 @@ class IdentityReasoningParser(ReasoningParser): "constructor during construction." ) + @property + def reasoning_start_str(self) -> str | None: + return None + + @property + def reasoning_end_str(self) -> str | None: + return None + def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: # Always return True, since we never treat reasoning specially return True diff --git a/vllm/reasoning/kimi_k2_reasoning_parser.py b/vllm/reasoning/kimi_k2_reasoning_parser.py index 8ee05ffd23..7a92703426 100644 --- a/vllm/reasoning/kimi_k2_reasoning_parser.py +++ b/vllm/reasoning/kimi_k2_reasoning_parser.py @@ -65,6 +65,14 @@ class KimiK2ReasoningParser(ReasoningParser): "tokens in the tokenizer!" ) + @property + def reasoning_start_str(self) -> str | None: + return self._start_token + + @property + def reasoning_end_str(self) -> str | None: + return self._end_token + def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: """ Check if the reasoning content ends in the input_ids. diff --git a/vllm/reasoning/olmo3_reasoning_parser.py b/vllm/reasoning/olmo3_reasoning_parser.py index 9697b50044..b685aa2318 100644 --- a/vllm/reasoning/olmo3_reasoning_parser.py +++ b/vllm/reasoning/olmo3_reasoning_parser.py @@ -237,6 +237,14 @@ class Olmo3ReasoningParser(ReasoningParser): think_start=self.think_start, think_end=self.think_end ) + @property + def reasoning_start_str(self) -> str: + return self.think_start + + @property + def reasoning_end_str(self) -> str: + return self.think_end + def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: text = self.model_tokenizer.decode(input_ids) return self.think_end in text diff --git a/vllm/reasoning/step3_reasoning_parser.py b/vllm/reasoning/step3_reasoning_parser.py index 5837f0673b..a50fcf02db 100644 --- a/vllm/reasoning/step3_reasoning_parser.py +++ b/vllm/reasoning/step3_reasoning_parser.py @@ -29,6 +29,7 @@ class Step3ReasoningParser(ReasoningParser): def __init__(self, tokenizer: PreTrainedTokenizerBase, *args, **kwargs): super().__init__(tokenizer, *args, **kwargs) + self.think_start_token = "" self.think_end_token = "" self.reasoning_regex = re.compile(rf"(.*?){self.think_end_token}", re.DOTALL) @@ -47,6 +48,14 @@ class Step3ReasoningParser(ReasoningParser): ) self.think_end_token_id: int = think_end_token_id + @property + def reasoning_start_str(self) -> str: + return self.think_start_token + + @property + def reasoning_end_str(self) -> str: + return self.think_end_token + def extract_reasoning_streaming( self, previous_text: str,