From e0705731be04aac90838f318fcdae15c56a53080 Mon Sep 17 00:00:00 2001
From: Flora Feng <4florafeng@gmail.com>
Date: Wed, 16 Sep 2026 17:00:48 -0400
Subject: [PATCH] [Refactor] Remove unused interface methods (#57194)
Signed-off-by: sfeng33 <4florafeng@gmail.com>
---
.../responses/test_parsable_context_unit.py | 9 -
.../openai/test_tool_choice_content_none.py | 3 -
tests/entrypoints/unit_tests/test_context.py | 4 +-
.../test_structural_tag_registry.py | 2 +-
vllm/parser/abstract_parser.py | 199 ++----------------
5 files changed, 18 insertions(+), 199 deletions(-)
diff --git a/tests/entrypoints/openai/responses/test_parsable_context_unit.py b/tests/entrypoints/openai/responses/test_parsable_context_unit.py
index 01293f62b2..85bf6f2020 100644
--- a/tests/entrypoints/openai/responses/test_parsable_context_unit.py
+++ b/tests/entrypoints/openai/responses/test_parsable_context_unit.py
@@ -36,9 +36,6 @@ class _NoOpParser(DelegatingParser):
def is_reasoning_end(self, input_ids: list[int]) -> bool:
return False
- def extract_content_ids(self, input_ids: list[int]) -> list[int]:
- return input_ids
-
def extract_reasoning(self, model_output, request):
return None, model_output
@@ -63,9 +60,6 @@ class _ReasoningOnlyParser(DelegatingParser):
def is_reasoning_end(self, input_ids: list[int]) -> bool:
return False
- def extract_content_ids(self, input_ids: list[int]) -> list[int]:
- return input_ids
-
def extract_reasoning(self, model_output, request):
if "" in model_output and "" in model_output:
start = model_output.index("") + len("")
@@ -131,9 +125,6 @@ class _ToolCallingParser(DelegatingParser):
def is_reasoning_end(self, input_ids: list[int]) -> bool:
return False
- def extract_content_ids(self, input_ids: list[int]) -> list[int]:
- return input_ids
-
def extract_reasoning(self, model_output, request):
return None, model_output
diff --git a/tests/entrypoints/openai/test_tool_choice_content_none.py b/tests/entrypoints/openai/test_tool_choice_content_none.py
index 7db815c45f..a0a8f38e9f 100644
--- a/tests/entrypoints/openai/test_tool_choice_content_none.py
+++ b/tests/entrypoints/openai/test_tool_choice_content_none.py
@@ -31,9 +31,6 @@ class _DummyDelegatingParser(DelegatingParser):
def is_reasoning_end(self, input_ids: list[int]) -> bool:
return False
- def extract_content_ids(self, input_ids: list[int]) -> list[int]:
- return input_ids
-
def extract_reasoning(self, model_output: str, request):
return None, model_output
diff --git a/tests/entrypoints/unit_tests/test_context.py b/tests/entrypoints/unit_tests/test_context.py
index c92813ea66..8c3d1467ed 100644
--- a/tests/entrypoints/unit_tests/test_context.py
+++ b/tests/entrypoints/unit_tests/test_context.py
@@ -71,8 +71,8 @@ async def generate_mock_outputs(
class FakeHarmonyParser(HarmonyParser):
def __init__(self):
# Skip HarmonyParser initialization and script outputs directly.
- self.reasoning_parser = None
- self.tool_parser = None
+ self._reasoning_parser = None
+ self._tool_parser = None
self._chunk_results: list[ChunkResult] = []
self._flush_results: list[list[Segment]] = []
self.processed_chunks: list[list[int]] = []
diff --git a/tests/tool_parsers/test_structural_tag_registry.py b/tests/tool_parsers/test_structural_tag_registry.py
index 88e1f3f383..06c8fb6129 100644
--- a/tests/tool_parsers/test_structural_tag_registry.py
+++ b/tests/tool_parsers/test_structural_tag_registry.py
@@ -623,7 +623,7 @@ def test_unified_parser_get_structural_tag_disables_reasoning(
tool_choice="auto",
)
parser = TestParser(MagicMock(), tools=sample_tools_strict)
- parser.reasoning_parser = MagicMock(adjust_request=lambda request: request)
+ parser._reasoning_parser = MagicMock(adjust_request=lambda request: request)
parser.adjust_request(request)
diff --git a/vllm/parser/abstract_parser.py b/vllm/parser/abstract_parser.py
index d51c4455be..af448e0951 100644
--- a/vllm/parser/abstract_parser.py
+++ b/vllm/parser/abstract_parser.py
@@ -6,7 +6,6 @@ import json
from abc import abstractmethod
from collections.abc import Sequence
from dataclasses import dataclass, field
-from functools import cached_property
from openai.types.responses import ToolChoiceFunction
from pydantic import TypeAdapter, ValidationError
@@ -84,23 +83,12 @@ class StreamState:
class Parser:
- """Abstract Parser class that unifies ReasoningParser and ToolParser into
- a single interface for parsing model output.
+ """Parse model output into reasoning, content and tool calls.
- This class provides a unified way to handle both reasoning extraction
- (e.g., chain-of-thought content in tags) and tool call extraction
- (e.g., function calls in XML/JSON format) from model outputs.
-
- Subclasses can either:
- 1. Override the abstract methods directly for custom parsing logic
- 2. Set `reasoning_parser` and `tool_parser` properties to delegate to
- existing parser implementations
-
- Class Attributes:
- reasoning_parser_cls: The ReasoningParser class to use (for compatibility
- with code that needs the class, not instance).
- tool_parser_cls: The ToolParser class to use (for compatibility with
- code that needs the class, not instance).
+ The serving layer holds one ``Parser`` per request and calls only the
+ members defined here. ``ParserEngine`` implements them over a single
+ declarative engine; ``DelegatingParser`` composes a legacy
+ ``ReasoningParser`` / ``ToolParser`` pair.
"""
# Class-level parser classes for compatibility with existing patterns
@@ -150,29 +138,16 @@ class Parser:
engine_based=self._engine_based,
)
- @cached_property
- def vocab(self) -> dict[str, int]:
- """Get the vocabulary mapping from tokens to IDs."""
- return self.model_tokenizer.get_vocab()
-
@property
def reasoning_parser(self) -> ReasoningParser | None:
"""The underlying reasoning parser, if any."""
return self._reasoning_parser
- @reasoning_parser.setter
- def reasoning_parser(self, parser: ReasoningParser | None) -> None:
- self._reasoning_parser = parser
-
@property
def tool_parser(self) -> ToolParser | None:
"""The underlying tool parser, if any."""
return self._tool_parser
- @tool_parser.setter
- def tool_parser(self, parser: ToolParser | None) -> None:
- self._tool_parser = parser
-
def _initialize_history_tool_call_cnt(
self,
request: ChatCompletionRequest | ResponsesRequest,
@@ -186,119 +161,6 @@ class Parser:
state.history_tool_call_cnt = count_history_tool_calls(request)
state.history_tool_call_cnt_initialized = True
- # ========== Reasoning Parser Methods ==========
-
- @abstractmethod
- def is_reasoning_end(self, input_ids: list[int]) -> bool:
- """Check if the reasoning content ends in the input_ids.
-
- Used by structured engines like `xgrammar` to check if the
- reasoning content ends in the model output.
-
- Must be a pure function of the input_ids.
-
- Args:
- input_ids: The token IDs of the model output.
-
- Returns:
- True if the reasoning content ends in the input_ids.
-
- """
-
- def is_reasoning_end_streaming(
- self, input_ids: list[int], delta_ids: list[int]
- ) -> bool:
- """Check if the reasoning content ends during a decode step.
-
- Must be a pure function of the input_ids and delta_ids.
-
- Args:
- input_ids: The entire model output token IDs.
- delta_ids: The last few computed tokens at the current decode step.
-
- Returns:
- True if the reasoning content ends in the delta_ids.
-
- """
- return self.is_reasoning_end(input_ids)
-
- def find_reasoning_end_offset(self, token_ids: Sequence[int]) -> int | None:
- """Locate the token that ends reasoning within one window of tokens.
-
- Args:
- token_ids: The tokens to examine.
-
- Returns:
- The offset within ``token_ids`` of the last reasoning token, or
- ``len(token_ids)`` when reasoning does not end in this window. Parsers
- that cannot answer from a window alone always return ``None``; callers
- fall back to :meth:`is_reasoning_end_streaming`.
-
- """
- return None
-
- @abstractmethod
- def extract_content_ids(self, input_ids: list[int]) -> list[int]:
- """Extract content token IDs from the input_ids.
-
- This extracts the non-reasoning content (e.g., everything after
- the tag).
-
- Args:
- input_ids: The token IDs of the model output.
-
- Returns:
- The extracted content token IDs.
-
- """
-
- @abstractmethod
- def extract_reasoning(
- self,
- model_output: str,
- request: ChatCompletionRequest | ResponsesRequest,
- ) -> tuple[str | None, str | None]:
- """Extract reasoning content from a complete model-generated string.
-
- Used for non-streaming responses where we have the entire model
- response available before sending to the client.
-
- Args:
- model_output: The complete model-generated string.
- request: The request object used to generate the output.
-
- Returns:
- A tuple of (reasoning, response_content).
-
- """
-
- @abstractmethod
- def extract_reasoning_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],
- ) -> DeltaMessage | None:
- """Extract reasoning content from a streaming delta message.
-
- Args:
- previous_text: Text from all previous tokens.
- current_text: Text including the current delta.
- delta_text: The new text in this delta.
- previous_token_ids: Token IDs from previous generation.
- current_token_ids: All token IDs including current.
- delta_token_ids: The new token IDs in this delta.
-
- Returns:
- A DeltaMessage with reasoning and/or content fields, or None.
-
- """
-
- # ========== Tool Parser Methods ==========
-
def adjust_request(
self, request: ChatCompletionRequest | ResponsesRequest
) -> ChatCompletionRequest | ResponsesRequest:
@@ -317,48 +179,17 @@ class Parser:
return request
@abstractmethod
- def extract_tool_calls(
- self,
- model_output: str,
- request: ChatCompletionRequest | ResponsesRequest,
- ) -> ExtractedToolCallInformation:
- """Extract tool calls from a complete model-generated string.
+ def is_reasoning_end(self, input_ids: list[int]) -> bool:
+ """Check if the reasoning content ends in the input_ids.
- Used for non-streaming responses.
+ Called with the rendered prompt to decide whether generation starts
+ after reasoning. Must be a pure function of the input_ids.
Args:
- model_output: The complete model-generated string.
- request: The request object used to generate the output.
+ input_ids: The token IDs of the model output.
Returns:
- ExtractedToolCallInformation containing the tool calls.
-
- """
-
- @abstractmethod
- def extract_tool_calls_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],
- request: ChatCompletionRequest | ResponsesRequest,
- ) -> DeltaMessage | None:
- """Extract tool calls from a streaming delta message.
-
- Args:
- previous_text: Text from all previous tokens.
- current_text: Text including the current delta.
- delta_text: The new text in this delta.
- previous_token_ids: Token IDs from previous generation.
- current_token_ids: All token IDs including current.
- delta_token_ids: The new token IDs in this delta.
- request: The request object.
-
- Returns:
- A DeltaMessage with tool_calls field, or None.
+ True if the reasoning content ends in the input_ids.
"""
@@ -764,14 +595,14 @@ class DelegatingParser(Parser):
return False
return self._reasoning_parser.is_reasoning_end(input_ids)
- def is_reasoning_end_streaming(
+ def _is_reasoning_end_streaming(
self, input_ids: list[int], delta_ids: list[int]
) -> bool:
if self._reasoning_parser is None:
return False
return self._reasoning_parser.is_reasoning_end_streaming(input_ids, delta_ids)
- def extract_content_ids(self, input_ids: list[int]) -> list[int]:
+ def _extract_content_ids(self, input_ids: list[int]) -> list[int]:
if self._reasoning_parser is None:
return input_ids
return self._reasoning_parser.extract_content_ids(input_ids)
@@ -885,13 +716,13 @@ class DelegatingParser(Parser):
reasoning_parser.has_engine_confirmed_reasoning_end()
)
else:
- should_transition = self.is_reasoning_end_streaming(
+ should_transition = self._is_reasoning_end_streaming(
current_token_ids, delta_token_ids
)
if should_transition:
state.reasoning_ended = True
reasoning_transitioned = True
- current_token_ids = self.extract_content_ids(delta_token_ids)
+ current_token_ids = self._extract_content_ids(delta_token_ids)
# Flush whenever the reasoning parser is engine-based (not only
# when _engine_based is True): it buffers the post-marker text
# (e.g. the "<" of ""), surfaced via finish_streaming().