| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | from collections.abc import Iterable, Sequence |
| 5 | from itertools import islice |
| 6 | from typing import TYPE_CHECKING |
| 7 | |
| 8 | import regex as re |
| 9 | from transformers import PreTrainedTokenizerBase |
| 10 | |
| 11 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 12 | from vllm.reasoning import ReasoningParser |
| 13 | |
| 14 | if TYPE_CHECKING: |
| 15 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 16 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 17 | |
| 18 | |
| 19 | class Step3ReasoningParser(ReasoningParser): |
| 20 | """Reasoning parser for Step3 model. |
| 21 | |
| 22 | The Step3 model uses </think> token to denote the end of reasoning |
| 23 | text. This parser extracts all content before </think> as reasoning content. |
| 24 | """ |
| 25 | |
| 26 | def __init__(self, tokenizer: PreTrainedTokenizerBase, *args, **kwargs): |
| 27 | super().__init__(tokenizer, *args, **kwargs) |
| 28 | self.think_start_token = "<think>" |
| 29 | self.think_end_token = "</think>" |
| 30 | |
| 31 | self.reasoning_regex = re.compile(rf"(.*?){self.think_end_token}", re.DOTALL) |
| 32 | |
| 33 | if not self.model_tokenizer: |
| 34 | raise ValueError( |
| 35 | "The model tokenizer must be passed to the ReasoningParser " |
| 36 | "constructor during construction." |
| 37 | ) |
| 38 | |
| 39 | think_end_token_id = self.vocab.get(self.think_end_token) |
| 40 | if think_end_token_id is None: |
| 41 | raise RuntimeError( |
| 42 | "Step3 reasoning parser could not locate think end " |
| 43 | "token in the tokenizer!" |
| 44 | ) |
| 45 | self.think_end_token_id: int = think_end_token_id |
| 46 | |
| 47 | @property |
| 48 | def reasoning_start_str(self) -> str: |
| 49 | return self.think_start_token |
| 50 | |
| 51 | @property |
| 52 | def reasoning_end_str(self) -> str: |
| 53 | return self.think_end_token |
| 54 | |
| 55 | def extract_reasoning_streaming( |
| 56 | self, |
| 57 | previous_text: str, |
| 58 | current_text: str, |
| 59 | delta_text: str, |
| 60 | previous_token_ids: Sequence[int], |
| 61 | current_token_ids: Sequence[int], |
| 62 | delta_token_ids: Sequence[int], |
| 63 | ) -> DeltaMessage | None: |
| 64 | """Extract reasoning content from a delta message. |
| 65 | Handles streaming output where previous + delta = current. |
| 66 | Uses token IDs for faster processing. |
| 67 | For text "abc</think>xyz": |
| 68 | - 'abc' goes to reasoning |
| 69 | - 'xyz' goes to content |
| 70 | """ |
| 71 | # Skip single special token |
| 72 | if len(delta_token_ids) == 1 and delta_token_ids[0] == self.think_end_token_id: |
| 73 | return None |
| 74 | |
| 75 | if self.think_end_token_id in delta_token_ids: |
| 76 | # </think> in delta, extract reasoning content and remaining content |
| 77 | end_index = delta_text.find(self.think_end_token) |
| 78 | reasoning = delta_text[:end_index] |
| 79 | content = delta_text[end_index + len(self.think_end_token) :] |
| 80 | return DeltaMessage( |
| 81 | reasoning=reasoning, |
| 82 | content=content if content else None, |
| 83 | ) |
| 84 | elif self.think_end_token_id in previous_token_ids: |
| 85 | # </think> already seen in previous text, everything is content |
| 86 | return DeltaMessage(content=delta_text) |
| 87 | else: |
| 88 | # No </think> seen yet, everything is reasoning |
| 89 | return DeltaMessage(reasoning=delta_text) |
| 90 | |
| 91 | def extract_reasoning( |
| 92 | self, model_output: str, request: "ChatCompletionRequest | ResponsesRequest" |
| 93 | ) -> tuple[str | None, str | None]: |
| 94 | # Check if the model output contains the </think> token |
| 95 | if self.think_end_token not in model_output: |
| 96 | # If no </think> token, everything is reasoning content |
| 97 | return model_output, None |
| 98 | else: |
| 99 | # Find the first occurrence of </think> |
| 100 | end_index = model_output.find(self.think_end_token) |
| 101 | reasoning = model_output[:end_index] |
| 102 | |
| 103 | # Content after </think> token |
| 104 | content = model_output[end_index + len(self.think_end_token) :] or None |
| 105 | |
| 106 | return reasoning, content |
| 107 | |
| 108 | def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: |
| 109 | return self.think_end_token_id in input_ids |
| 110 | |
| 111 | def is_reasoning_end_streaming( |
| 112 | self, input_ids: Sequence[int], delta_ids: Iterable[int] |
| 113 | ) -> bool: |
| 114 | end_token_id = self.think_end_token_id |
| 115 | return end_token_id in delta_ids |
| 116 | |
| 117 | def extract_content_ids(self, input_ids: list[int]) -> list[int]: |
| 118 | if self.think_end_token_id not in islice( |
| 119 | input_ids, 0, max(0, len(input_ids) - 1) |
| 120 | ): |
| 121 | return [] |
| 122 | else: |
| 123 | return input_ids[input_ids.index(self.think_end_token_id) + 1 :] |
| 124 | |
| 125 | def count_reasoning_tokens(self, token_ids: Sequence[int]) -> int: |
| 126 | if self.think_end_token_id not in token_ids: |
| 127 | return len(token_ids) |
| 128 | return list(token_ids).index(self.think_end_token_id) |