| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Sequence |
| 5 | |
| 6 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 7 | from vllm.reasoning.basic_parsers import BaseThinkingReasoningParser |
| 8 | |
| 9 | |
| 10 | class DeepSeekR1ReasoningParser(BaseThinkingReasoningParser): |
| 11 | """Reasoning parser for DeepSeek R1 model. |
| 12 | |
| 13 | The DeepSeek R1 model uses <think>...</think> tokens to denote reasoning |
| 14 | text. This parser extracts the reasoning content from the model output. |
| 15 | """ |
| 16 | |
| 17 | @property |
| 18 | def start_token(self) -> str: |
| 19 | """The token that starts reasoning content.""" |
| 20 | return "<think>" |
| 21 | |
| 22 | @property |
| 23 | def end_token(self) -> str: |
| 24 | """The token that ends reasoning content.""" |
| 25 | return "</think>" |
| 26 | |
| 27 | def extract_reasoning_streaming( |
| 28 | self, |
| 29 | previous_text: str, |
| 30 | current_text: str, |
| 31 | delta_text: str, |
| 32 | previous_token_ids: Sequence[int], |
| 33 | current_token_ids: Sequence[int], |
| 34 | delta_token_ids: Sequence[int], |
| 35 | ) -> DeltaMessage | None: |
| 36 | ret = super().extract_reasoning_streaming( |
| 37 | previous_text, |
| 38 | current_text, |
| 39 | delta_text, |
| 40 | previous_token_ids, |
| 41 | current_token_ids, |
| 42 | delta_token_ids, |
| 43 | ) |
| 44 | if ( |
| 45 | ret is not None |
| 46 | and self.start_token_id not in previous_token_ids |
| 47 | and self.start_token_id not in delta_token_ids |
| 48 | ): |
| 49 | if self.end_token_id in delta_token_ids: |
| 50 | # end token in delta with more tokens, |
| 51 | # extract reasoning content and content |
| 52 | end_index = delta_text.find(self.end_token) |
| 53 | reasoning = delta_text[:end_index] |
| 54 | content = delta_text[end_index + len(self.end_token) :] |
| 55 | return DeltaMessage( |
| 56 | reasoning=reasoning, |
| 57 | content=content if content else None, |
| 58 | ) |
| 59 | elif self.end_token_id in previous_token_ids: |
| 60 | # end token in previous, thinking content ends |
| 61 | return DeltaMessage(content=delta_text) |
| 62 | else: |
| 63 | # no end token in previous or delta, reasoning content continues |
| 64 | return DeltaMessage(reasoning=delta_text) |
| 65 | |
| 66 | return ret |