| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Sequence |
| 5 | from typing import TYPE_CHECKING |
| 6 | |
| 7 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 8 | from vllm.parser.engine.registered_adapters import MinimaxM2ParserReasoningAdapter |
| 9 | from vllm.reasoning.abs_reasoning_parsers import ReasoningParser |
| 10 | from vllm.tokenizers import TokenizerLike |
| 11 | |
| 12 | if TYPE_CHECKING: |
| 13 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 14 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 15 | |
| 16 | |
| 17 | class MiniMaxM2ReasoningParser(MinimaxM2ParserReasoningAdapter): # type: ignore[valid-type, misc] |
| 18 | """Reasoning parser for MiniMax M2 model. |
| 19 | |
| 20 | MiniMax M2 models don't generate <think> start token, only </think> end |
| 21 | token. All content before </think> is reasoning, content after is the |
| 22 | actual response. |
| 23 | """ |
| 24 | |
| 25 | |
| 26 | class MiniMaxM2AppendThinkReasoningParser(ReasoningParser): |
| 27 | """Reasoning parser for MiniMax M2 model.""" |
| 28 | |
| 29 | def __init__(self, tokenizer: TokenizerLike, *args, **kwargs): |
| 30 | super().__init__(tokenizer, *args, **kwargs) |
| 31 | self.end_token_id = self.vocab.get("</think>") |
| 32 | self.start_token_id = self.vocab.get("<think>") |
| 33 | |
| 34 | def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: |
| 35 | end_token_id = self.end_token_id |
| 36 | start_token_id = self.start_token_id |
| 37 | for input_id in reversed(input_ids): |
| 38 | if input_id in (end_token_id, start_token_id): |
| 39 | return input_id == end_token_id |
| 40 | return False |
| 41 | |
| 42 | def extract_content_ids(self, input_ids: list[int]) -> list[int]: |
| 43 | return input_ids |
| 44 | |
| 45 | def extract_reasoning_streaming( |
| 46 | self, |
| 47 | previous_text: str, |
| 48 | current_text: str, |
| 49 | delta_text: str, |
| 50 | previous_token_ids: Sequence[int], |
| 51 | current_token_ids: Sequence[int], |
| 52 | delta_token_ids: Sequence[int], |
| 53 | ) -> DeltaMessage | None: |
| 54 | if len(previous_token_ids) == 0: |
| 55 | delta_text = "<think>" + delta_text |
| 56 | return DeltaMessage(content=delta_text) |
| 57 | |
| 58 | def extract_reasoning( |
| 59 | self, model_output: str, request: "ChatCompletionRequest | ResponsesRequest" |
| 60 | ) -> tuple[str | None, str | None]: |
| 61 | return None, "<think>" + model_output |