Bases: ReasoningParser
Expose is_reasoning_end for the structured-output engine.
Everything else runs in
:class:vllm.parser.cohere_command.CohereCommandParser.
Source code in vllm/reasoning/cohere_command_reasoning_parser.py
| 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:
# Reasoning ended if an end token follows the newest start token,
# unless a chatbot token comes after that start.
i = rfind_tokens(input_ids, (self.start_token_id, self.chatbot_token_id))
if i >= 0 and input_ids[i] == self.chatbot_token_id:
return False
return rfind_tokens(input_ids, (self.end_token_id,), start=i + 1) >= 0
|