FastDeploy/fastdeploy/reasoning/ernie_45_vl_thinking_reasoning_parser.py

"""
# Copyright (c) 2025  PaddlePaddle Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License"
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
"""

from collections.abc import Sequence
from typing import Optional, Union

from fastdeploy.entrypoints.openai.protocol import ChatCompletionRequest, DeltaMessage
from fastdeploy.reasoning import ReasoningParser, ReasoningParserManager


@ReasoningParserManager.register_module("ernie-45-vl-thinking")
class Ernie45VLThinkingReasoningParser(ReasoningParser):
    """
    Reasoning parser for ernie_vl model.

    The ernie_vl model uses ...</think>... tokens to denote reasoning text
    within its output. The model provides a strict switch to disable reasoning
    output via the 'enable_thinking=False' parameter. This parser extracts the
    reasoning content enclosed by <think> and </think> tokens from the model's
    output.
    """

    def __init__(self, tokenizer):
        super().__init__(tokenizer)
        token_definitions = {
            "think_start_token": "<think>",
            "think_end_token": "</think>",
            "tool_call_start_token": "<tool_call>",
            "tool_call_end_token": "</tool_call>",
        }

        if not self.model_tokenizer:
            raise ValueError(
                "The model tokenizer must be passed to the ReasoningParser " "constructor during construction."
            )
        missing_tokens = []
        for name, token_value in token_definitions.items():
            setattr(self, name, token_value)
            token_id = self.vocab.get(token_value)
            setattr(self, f"{name}_id", token_id)
            if token_id is None:
                missing_tokens.append(f"{name.replace('_', ' ')} token")

        if missing_tokens:
            raise RuntimeError(
                f"ernie vl reasoning parser could not find the following token ids in tokenizer vocabulary: {', '.join(missing_tokens)}"
            )
        self.token_status_mapping = {
            self.think_start_token_id: "think_start",
            self.think_end_token_id: "think_end",
            self.tool_call_start_token_id: "tool_call_start",
            self.tool_call_end_token_id: "tool_call_end",
        }

    def is_reasoning_end(self, input_ids: list[int]) -> bool:
        return self.think_end_token_id in input_ids

    def find_last_special_token(self, prompt_token_ids: list[int]) -> int:
        for i in range(len(prompt_token_ids) - 1, -1, -1):
            if prompt_token_ids[i] in self.token_status_mapping:
                return prompt_token_ids[i]
        return -1

    def get_model_status(self, prompt_token_ids: list[int]):
        special_token_id = self.find_last_special_token(prompt_token_ids)

        if special_token_id == -1:
            return "think_start"

        return self.token_status_mapping[special_token_id]

    def extract_reasoning_content_streaming(
        self,
        previous_text: str,
        current_text: str,
        delta_text: str,
        previous_token_ids: Sequence[int],
        current_token_ids: Sequence[int],
        delta_token_ids: Sequence[int],
        model_status: str,
    ) -> Union[DeltaMessage, None]:
        """
        Extract reasoning content from a delta message.
        Handles streaming output where previous + delta = current.
        Uses token IDs for faster processing.
        For text abc</think>xyz:
        - 'abc' goes to reasoning_content
        - 'xyz' goes to content
        """
        if model_status == "think_start":
            if self.think_end_token not in current_text:
                return DeltaMessage(reasoning_content=delta_text)
            # Skip single special tokens
            if len(delta_token_ids) == 1 and delta_token_ids[0] == self.think_end_token_id:
                return None
            if self._is_with_tool(current_text=current_text, current_token_ids=current_token_ids):
                if self.think_end_token in delta_text:
                    think_begin = delta_text.find(self.think_end_token)
                    reasoning_content = delta_text[:think_begin]
                    return DeltaMessage(reasoning_content=reasoning_content)
                return None
            if self.think_end_token in delta_text:
                reasoning_content, _, content = delta_text.partition(self.think_end_token)
                striped_content = content.strip("\n")
                if len(striped_content) == 0:
                    return DeltaMessage(reasoning_content=reasoning_content) if reasoning_content else None
                return (
                    DeltaMessage(reasoning_content=reasoning_content, content=content)
                    if reasoning_content
                    else DeltaMessage(content=content)
                )
            think_end = current_text.find(self.think_end_token) + len(self.think_end_token)
            suffix = current_text[think_end:]
            striped_suffix = suffix.strip("\n")
            if len(striped_suffix) == 0:
                return None
            return DeltaMessage(content=delta_text)
        elif model_status == "think_end":
            if current_text.lstrip("\n").startswith(self.tool_call_start_token):
                return None
            return DeltaMessage(content=delta_text)
        else:
            return None

    def extract_reasoning_content(
        self,
        model_output: str,
        request: ChatCompletionRequest,
        model_status: str,
    ) -> tuple[Optional[str], Optional[str]]:
        """
        Extract reasoning content from the model output.

        For text abc</think>xyz:
        - 'abc' goes to reasoning_content
        - 'xyz' goes to content

        Returns:
            tuple[Optional[str], Optional[str]]: reasoning content and content
        """

        # Check if the model output contains the </think> tokens.
        if model_status == "think_start":
            if self.think_end_token not in model_output:
                return model_output, ""
            reasoning_content, _, content = model_output.partition(self.think_end_token)
            if self.tool_call_start_token in content:
                prefix, _, _ = content.partition(self.tool_call_start_token)
                prefix_strip = prefix.lstrip("\n")
                if len(prefix_strip) > 0:
                    return reasoning_content, content
                return reasoning_content, ""
            return reasoning_content, content
        elif model_status == "think_end":
            if model_output.lstrip("\n").startswith(self.tool_call_start_token):
                return "", ""
            return "", model_output
        else:
            return "", ""

    def _is_with_tool(self, current_text: str, current_token_ids: Sequence[int]) -> bool:
        think_end_index = current_text.find(self.think_end_token)
        think_end = think_end_index + len(self.think_end_token)
        middle_str = current_text[think_end:]
        if self.tool_call_start_token_id in current_token_ids:
            prefix, _, _ = middle_str.partition(self.tool_call_start_token)
            striped_prefix = prefix.strip("\n")
            if len(striped_prefix) > 0:
                return False
            return True
        return False