| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | from collections.abc import Iterable, Sequence |
| 4 | from typing import TYPE_CHECKING |
| 5 | |
| 6 | from transformers import PreTrainedTokenizerBase |
| 7 | |
| 8 | from vllm.entrypoints.generate.base.protocol import DeltaMessage |
| 9 | from vllm.reasoning import ReasoningParser |
| 10 | |
| 11 | if TYPE_CHECKING: |
| 12 | from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest |
| 13 | from vllm.entrypoints.openai.responses.protocol import ResponsesRequest |
| 14 | |
| 15 | |
| 16 | class GptOssReasoningParser(ReasoningParser): |
| 17 | """Reasoning parser for GptOss model. |
| 18 | |
| 19 | The GptOss model uses harmony to extract reasoning content and this parser |
| 20 | is only used for detecting the end of the reasoning content. |
| 21 | """ |
| 22 | |
| 23 | def __init__(self, tokenizer: PreTrainedTokenizerBase, *args, **kwargs): |
| 24 | super().__init__(tokenizer, *args, **kwargs) |
| 25 | |
| 26 | def is_reasoning_end(self, input_ids: Sequence[int]) -> bool: |
| 27 | return True |
| 28 | |
| 29 | def is_reasoning_end_streaming( |
| 30 | self, input_ids: Sequence[int], delta_ids: Iterable[int] |
| 31 | ) -> bool: |
| 32 | return True |
| 33 | |
| 34 | def extract_content_ids(self, input_ids: list[int]) -> list[int]: |
| 35 | raise NotImplementedError( |
| 36 | "GptOssReasoningParser only provides boundary detection. " |
| 37 | "Use HarmonyParser for output parsing." |
| 38 | ) |
| 39 | |
| 40 | def extract_reasoning_streaming( |
| 41 | self, |
| 42 | previous_text: str, |
| 43 | current_text: str, |
| 44 | delta_text: str, |
| 45 | previous_token_ids: Sequence[int], |
| 46 | current_token_ids: Sequence[int], |
| 47 | delta_token_ids: Sequence[int], |
| 48 | ) -> DeltaMessage | None: |
| 49 | raise NotImplementedError( |
| 50 | "GptOssReasoningParser only provides boundary detection. " |
| 51 | "Use HarmonyParser for output parsing." |
| 52 | ) |
| 53 | |
| 54 | def extract_reasoning( |
| 55 | self, |
| 56 | model_output: str, |
| 57 | request: "ChatCompletionRequest | ResponsesRequest", |
| 58 | ) -> tuple[str | None, str | None]: |
| 59 | raise NotImplementedError( |
| 60 | "GptOssReasoningParser only provides boundary detection. " |
| 61 | "Use HarmonyParser for output parsing." |
| 62 | ) |