mirror of
https://github.com/PaddlePaddle/FastDeploy.git
synced 2025-12-24 13:28:13 +08:00
109 lines
4.4 KiB
Python
109 lines
4.4 KiB
Python
from collections.abc import Sequence
|
|
from typing import Tuple, Union
|
|
|
|
from fastdeploy.entrypoints.openai.protocol import ChatCompletionRequest, DeltaMessage
|
|
from fastdeploy.reasoning import ReasoningParser, ReasoningParserManager
|
|
|
|
|
|
@ReasoningParserManager.register_module("ernie_x1")
|
|
class ErnieX1ReasoningParser(ReasoningParser):
|
|
"""
|
|
Reasoning parser for ernie_x1 model with stricter boundary checking.
|
|
|
|
Unified rules:
|
|
- Do not strip newline before </think>
|
|
- Do not strip newline after <response>
|
|
- Do not strip newline before </response>
|
|
"""
|
|
|
|
def __init__(self, tokenizer):
|
|
super().__init__(tokenizer)
|
|
self.think_end_token = "</think>"
|
|
self.response_start_token = "<response>"
|
|
self.response_end_token = "</response>"
|
|
self.tool_call_start_token = "<tool_call>"
|
|
self.tool_call_end_token = "</tool_call>"
|
|
|
|
if not self.model_tokenizer:
|
|
raise ValueError("The model tokenizer must be passed to the ReasoningParser constructor.")
|
|
|
|
self.think_end_token_id = self.vocab.get("</think>")
|
|
if self.think_end_token_id is None:
|
|
raise RuntimeError("Could not find think end token id in tokenizer vocabulary")
|
|
self.tool_call_start_token_id = self.vocab.get("<tool_call>")
|
|
|
|
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],
|
|
) -> Union[DeltaMessage, None]:
|
|
# Ignore the single </think> token
|
|
if len(delta_token_ids) == 1 and delta_token_ids[0] == self.think_end_token_id:
|
|
return None
|
|
|
|
# --- Thinking stage handling ---
|
|
if not previous_text.endswith(self.think_end_token) and self.think_end_token not in previous_text:
|
|
# If delta is </think>, stop thinking, do not return
|
|
if delta_text.startswith(self.think_end_token):
|
|
return None
|
|
# Otherwise, return thinking content (keep \n as-is)
|
|
return DeltaMessage(reasoning_content=delta_text)
|
|
|
|
# --- After thinking ends, check tool_call or response ---
|
|
remaining_text = previous_text + delta_text
|
|
after_think = remaining_text[remaining_text.find(self.think_end_token) + len(self.think_end_token) :]
|
|
after_think = after_think.lstrip("\n")
|
|
|
|
# Handle tool_call case: skip it
|
|
if after_think.startswith(self.tool_call_start_token):
|
|
return None
|
|
|
|
# Handle response case
|
|
if after_think.startswith(self.response_start_token):
|
|
# Do not return when <response> tag itself appears
|
|
if delta_text == self.response_start_token:
|
|
return None
|
|
# Do not return </response> itself
|
|
elif delta_text == self.response_end_token:
|
|
return None
|
|
# Otherwise, return response content (keep \n as-is)
|
|
else:
|
|
return DeltaMessage(content=delta_text)
|
|
|
|
# Default case: return nothing
|
|
return None
|
|
|
|
def extract_reasoning_content(self, model_output: str, request: ChatCompletionRequest) -> Tuple[str, str]:
|
|
reasoning_content = ""
|
|
response_content = ""
|
|
|
|
think_end_pos = model_output.find(self.think_end_token)
|
|
if think_end_pos != -1:
|
|
reasoning_content = model_output[:think_end_pos]
|
|
|
|
remaining = model_output[think_end_pos + len(self.think_end_token) :]
|
|
|
|
# find <response> or <tool>
|
|
response_pos = remaining.find(self.response_start_token)
|
|
tool_pos = remaining.find(self.tool_call_start_token)
|
|
|
|
# <response> first
|
|
if response_pos != -1 and (tool_pos == -1 or response_pos < tool_pos):
|
|
# The content after the response_start position
|
|
remaining_response = remaining[response_pos + len(self.response_start_token) :]
|
|
response_end_pos = remaining_response.find(self.response_end_token)
|
|
if response_end_pos != -1:
|
|
response_content = remaining_response[:response_end_pos]
|
|
else:
|
|
response_content = remaining_response
|
|
# The content after the response_start position is tool_call
|
|
else:
|
|
reasoning_content = model_output
|
|
response_content = ""
|
|
|
|
return reasoning_content, response_content
|