[Frontend] Add Streaming Parser Engine and new Qwen3 Parser (#45413)

Signed-off-by: Ben Browning <bbrownin@redhat.com>
Co-authored-by: Flora Feng <4florafeng@gmail.com>
This commit is contained in:
Ben Browning
2026-06-15 11:59:05 +08:00
committed by GitHub
co-authored by Flora Feng
parent e3e3cd5458
commit c4a3f9d137
33 changed files with 8494 additions and 904 deletions
+114
View File
@@ -0,0 +1,114 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Declarative configuration for model tool-call and reasoning formats.
Each model format is described by a :class:`ParserEngineConfig` that specifies:
* **terminals** – literal strings or regex patterns that delimit the format
(e.g. ``<tool_call>``, ``</think>``).
* **token_id_terminals** – terminals that should be matched by token ID
rather than (or in addition to) text.
* **transitions** – a state machine mapping
``(state, terminal) → (new_state, events_to_emit)`` that drives semantic
event generation during streaming.
* **content_events** – what :class:`EventType` to emit for plain content
(non-terminal text) in each state.
"""
from __future__ import annotations
from collections.abc import Callable
from dataclasses import dataclass, field
from enum import Enum, auto
from functools import cached_property
from vllm.parser.engine.events import EventType
STRUCTURAL_DROP_TOKENS: frozenset[str] = frozenset(
{
"<eos>",
"<bos>",
"<pad>",
"<unk>",
"<mask>",
}
)
class ParserState(Enum):
CONTENT = auto()
REASONING = auto()
TOOL_PREAMBLE = auto()
TOOL_NAME = auto()
TOOL_ARGS = auto()
TOOL_BETWEEN = auto()
@dataclass(frozen=True, slots=True)
class Transition:
next_state: ParserState
events: tuple[EventType, ...] = field(default_factory=tuple)
skip_in_token_id_mode: bool = False
@dataclass(frozen=True)
class ParserEngineConfig:
"""Declarative description of a model's tool-call / reasoning format.
The engine feeds terminals from the incremental lexer into the
transition table and emits the corresponding semantic events.
Content tokens (text between terminals) are classified by the
current state via ``content_events``.
"""
name: str
terminals: dict[str, str] = field(default_factory=dict)
token_id_terminals: dict[str, str] = field(default_factory=dict)
transitions: dict[tuple[ParserState, str], Transition] = field(
default_factory=dict,
)
content_events: dict[ParserState, EventType] = field(
default_factory=lambda: {
ParserState.CONTENT: EventType.TEXT_CHUNK,
ParserState.REASONING: EventType.REASONING_CHUNK,
ParserState.TOOL_NAME: EventType.TOOL_NAME,
ParserState.TOOL_ARGS: EventType.ARG_VALUE_CHUNK,
},
)
initial_state: ParserState = ParserState.CONTENT
arg_converter: Callable[[str, bool], str] | None = None
stream_arg_deltas: bool = True
tool_args_json: bool = True
arg_structural_chars: frozenset[str] | None = None
# Prevents trailing-whitespace accumulation across multi-turn conversations.
strip_trailing_reasoning_whitespace: bool = True
# Drop content that is entirely whitespace when tool calls follow.
drop_whitespace_only_content_before_tools: bool = True
# .strip() content text when tool calls are present.
strip_content_whitespace_with_tools: bool = True
drop_tokens: frozenset[str] = field(default_factory=frozenset)
@cached_property
def terminal_defs(self):
from vllm.parser.engine.incremental_lexer import terminals_from_literals
return terminals_from_literals(self.terminals)
@cached_property
def lexer_shape(self):
from vllm.parser.engine.incremental_lexer import LexerShape
return LexerShape(self.terminal_defs)