vllm.reasoning.cohere_command_reasoning_parser ¶

class BaseCohereCommandReasoningParser(ReasoningParser):
    """Expose is_reasoning_end for the structured-output engine.
    Everything else runs in
    :class:`vllm.parser.cohere_command.CohereCommandParser`.
    """

    melody_preset: ClassVar[str]

    def __init__(self, tokenizer: TokenizerLike, *args, **kwargs):
        super().__init__(tokenizer, *args, **kwargs)
        self.start_token_id = tokenizer.convert_tokens_to_ids("<|START_THINKING|>")
        self.end_token_id = tokenizer.convert_tokens_to_ids("<|END_THINKING|>")
        self.chatbot_token_id = tokenizer.convert_tokens_to_ids("<|CHATBOT_TOKEN|>")

    @property
    def reasoning_start_str(self) -> str | None:
        return "<|START_THINKING|>"

    @property
    def reasoning_end_str(self) -> str | None:
        return "<|END_THINKING|>"

    def is_reasoning_end(self, input_ids: Sequence[int]) -> bool:
        chatbot = self.chatbot_token_id
        start = self.start_token_id
        end = self.end_token_id
        has_end_token = False

        for i in reversed(range(len(input_ids))):
            tid = input_ids[i]
            if tid == start:
                return has_end_token
            if tid == chatbot:
                return False
            if tid == end:
                has_end_token = True

        return has_end_token