| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from abc import abstractmethod |
| 5 | from collections.abc import Iterable, Sequence |
| 6 | from itertools import islice |
| 7 | from typing import TYPE_CHECKING |
| 8 | |
| 9 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 10 | from vllm.reasoning.abs_reasoning_parsers import ReasoningParser |
| 11 | from vllm.tokenizers import TokenizerLike |
| 12 | |
| 13 | if TYPE_CHECKING: |
| 14 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 15 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 16 | |
| 17 | |
| 18 | class BaseThinkingReasoningParser(ReasoningParser): |
| 19 | """Base class for reasoning parsers that use thinking tokens. |
| 20 | |
| 21 | This class provides common functionality for parsers that use start and end |
| 22 | tokens to delimit reasoning content ( |
| 23 | e.g., <think>...</think>, <seed:think>...</seed:think>). |
| 24 | |
| 25 | Subclasses must implement the start and end tokens via abstract |
| 26 | properties. |
| 27 | """ |
| 28 | |
| 29 | @property |
| 30 | @abstractmethod |
| 31 | def start_token(self) -> str: |
| 32 | """The token that starts reasoning content.""" |
| 33 | raise NotImplementedError |
| 34 | |
| 35 | @property |
| 36 | @abstractmethod |
| 37 | def end_token(self) -> str: |
| 38 | """The token that ends reasoning content.""" |
| 39 | raise NotImplementedError |
| 40 | |
| 41 | @property |
| 42 | def reasoning_start_str(self) -> str: |
| 43 | return self.start_token |
| 44 | |
| 45 | @property |
| 46 | def reasoning_end_str(self) -> str: |
| 47 | return self.end_token |
| 48 | |
| 49 | def __init__(self, tokenizer: TokenizerLike, *args, **kwargs): |
| 50 | super().__init__(tokenizer, *args, **kwargs) |
| 51 | |
| 52 | if not self.model_tokenizer: |
| 53 | raise ValueError( |
| 54 | "The model tokenizer must be passed to the ReasoningParser " |
| 55 | "constructor during construction." |
| 56 | ) |
| 57 | |
| 58 | if not self.start_token or not self.end_token: |
| 59 | raise ValueError("start_token and end_token must be defined in subclasses") |
| 60 | |
| 61 | start_token_id = self.vocab.get(self.start_token) |
| 62 | end_token_id = self.vocab.get(self.end_token) |
| 63 | if start_token_id is None or end_token_id is None: |
| 64 | raise RuntimeError( |
| 65 | f"{self.__class__.__name__} reasoning parser could not locate " |
| 66 | "think start/end tokens in the tokenizer!" |
| 67 | ) |
| 68 | self.start_token_id: int = start_token_id |
| 69 | self.end_token_id: int = end_token_id |
| 70 | |
| 71 | def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: |
| 72 | start_token_id = self.start_token_id |
| 73 | end_token_id = self.end_token_id |
| 74 | |
| 75 | for i in range(len(input_ids) - 1, -1, -1): |
| 76 | if input_ids[i] == start_token_id: |
| 77 | return False |
| 78 | if input_ids[i] == end_token_id: |
| 79 | return True |
| 80 | return False |
| 81 | |
| 82 | def is_reasoning_end_streaming( |
| 83 | self, input_ids: Sequence[int], delta_ids: Iterable[int] |
| 84 | ) -> bool: |
| 85 | end_token_id = self.end_token_id |
| 86 | return end_token_id in delta_ids |
| 87 | |
| 88 | def extract_content_ids(self, input_ids: list[int]) -> list[int]: |
| 89 | """Extract the content after the end tokens.""" |
| 90 | if self.end_token_id not in islice(input_ids, 0, max(0, len(input_ids) - 1)): |
| 91 | return [] |
| 92 | else: |
| 93 | return input_ids[input_ids.index(self.end_token_id) + 1 :] |
| 94 | |
| 95 | def extract_reasoning_streaming( |
| 96 | self, |
| 97 | previous_text: str, |
| 98 | current_text: str, |
| 99 | delta_text: str, |
| 100 | previous_token_ids: Sequence[int], |
| 101 | current_token_ids: Sequence[int], |
| 102 | delta_token_ids: Sequence[int], |
| 103 | ) -> DeltaMessage | None: |
| 104 | """Extract reasoning content from a delta message. |
| 105 | Handles streaming output where previous + delta = current. |
| 106 | Uses token IDs for faster processing. |
| 107 | """ |
| 108 | # Skip single special tokens |
| 109 | if len(delta_token_ids) == 1 and ( |
| 110 | delta_token_ids[0] in [self.start_token_id, self.end_token_id] |
| 111 | ): |
| 112 | return None |
| 113 | |
| 114 | # Check if start token is present in previous or delta. |
| 115 | # Keep compatibility with models that don't generate start tokens. |
| 116 | if self.start_token_id in previous_token_ids: |
| 117 | if self.end_token_id in delta_token_ids: |
| 118 | # start token in previous, end token in delta, |
| 119 | # extract reasoning content |
| 120 | end_index = delta_text.find(self.end_token) |
| 121 | reasoning = delta_text[:end_index] |
| 122 | content = delta_text[end_index + len(self.end_token) :] |
| 123 | return DeltaMessage( |
| 124 | reasoning=reasoning, content=content if content else None |
| 125 | ) |
| 126 | elif self.end_token_id in previous_token_ids: |
| 127 | # start token in previous, end token in previous, |
| 128 | # reasoning content continues |
| 129 | return DeltaMessage(content=delta_text) |
| 130 | else: |
| 131 | # start token in previous, no end token in previous or delta, |
| 132 | # reasoning content continues |
| 133 | return DeltaMessage(reasoning=delta_text) |
| 134 | elif self.start_token_id in delta_token_ids: |
| 135 | if self.end_token_id in delta_token_ids: |
| 136 | # start token in delta, end token in delta, |
| 137 | # extract reasoning content |
| 138 | start_index = delta_text.find(self.start_token) |
| 139 | end_index = delta_text.find(self.end_token) |
| 140 | reasoning = delta_text[start_index + len(self.start_token) : end_index] |
| 141 | content = delta_text[end_index + len(self.end_token) :] |
| 142 | return DeltaMessage( |
| 143 | reasoning=reasoning, content=content if content else None |
| 144 | ) |
| 145 | else: |
| 146 | # start token in delta, no end token in delta, |
| 147 | # reasoning content continues |
| 148 | return DeltaMessage(reasoning=delta_text) |
| 149 | else: |
| 150 | # not find thinking start token |
| 151 | return DeltaMessage(content=delta_text) |
| 152 | |
| 153 | def extract_reasoning( |
| 154 | self, model_output: str, request: "ChatCompletionRequest | ResponsesRequest" |
| 155 | ) -> tuple[str | None, str | None]: |
| 156 | """Extract reasoning content from the model output. |
| 157 | |
| 158 | This is the base implementation that works for most models. |
| 159 | Subclasses can override this method for specific behavior. |
| 160 | """ |
| 161 | # Check if the start token is present in the model output, remove it |
| 162 | # if it is present. |
| 163 | model_output_parts = model_output.partition(self.start_token) |
| 164 | model_output = ( |
| 165 | model_output_parts[2] if model_output_parts[1] else model_output_parts[0] |
| 166 | ) |
| 167 | |
| 168 | # For models that may not generate start token, |
| 169 | # assume the reasoning content is always at the start. |
| 170 | if self.end_token not in model_output: |
| 171 | return model_output, None |
| 172 | else: |
| 173 | reasoning, _, content = model_output.partition(self.end_token) |
| 174 | # If generation stops right after end-of-think, return null content |
| 175 | final_content = content or None |
| 176 | return reasoning, final_content |
| 177 | |
| 178 | def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int: |
| 179 | """Count tokens that fall within start/end thinking markers. |
| 180 | |
| 181 | Uses a depth counter so nested spans are handled safely and stray end |
| 182 | tokens do not drive the counter negative. |
| 183 | """ |
| 184 | count = 0 |
| 185 | depth = 0 |
| 186 | for token_id in token_ids: |
| 187 | if token_id == self.start_token_id: |
| 188 | depth += 1 |
| 189 | continue |
| 190 | if token_id == self.end_token_id: |
| 191 | if depth > 0: |
| 192 | depth -= 1 |
| 193 | continue |
| 194 | if depth > 0: |
| 195 | count += 1 |
| 196 | return count |