[Frontend] Add Streaming Parser Engine and new MinimaxM2 Parser (#45701)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
Chauncey
2026-06-16 13:38:17 -04:00
committed by GitHub
parent 44b2512767
commit f00e163f35
12 changed files with 588 additions and 481 deletions
+18 -5
View File
@@ -127,12 +127,14 @@ class IncrementalLexer:
def flush(self) -> list[LexToken]:
tokens: list[LexToken] = []
if self.buffer:
tokens.extend(self._drain(final=True))
if self.buffer:
tokens.append(LexToken(self.content_terminal, self.buffer))
self.buffer = ""
return tokens
def _drain(self) -> list[LexToken]:
def _drain(self, *, final: bool = False) -> list[LexToken]:
tokens: list[LexToken] = []
first_chars = self._literal_first_chars
content_terminal = self.content_terminal
@@ -161,11 +163,22 @@ class IncrementalLexer:
):
best_match = (name, lit, len(lit))
if self.buffer in prefix_set:
# If the current buffer is both a complete literal and the prefix
# of a longer literal, wait for the next chunk. For example,
# "<invoke name=" should not be emitted before the next chunk
# proves whether this is the quoted form '<invoke name="'.
if self.buffer in prefix_set and not final:
if best_match is not None:
tokens.append(LexToken(best_match[0], best_match[1]))
self.buffer = self.buffer[best_match[2] :]
continue
longer_match = False
for lit, _ in literals_by_first.get(first, ()):
if len(lit) > best_match[2] and lit.startswith(self.buffer):
longer_match = True
break
if not longer_match:
tokens.append(LexToken(best_match[0], best_match[1]))
self.buffer = self.buffer[best_match[2] :]
continue
break
else:
break
+16 -4
View File
@@ -30,6 +30,7 @@ from vllm.parser.engine.streaming_parser_engine import StreamingParserEngine
from vllm.tool_parsers.utils import (
coerce_to_schema_type,
extract_types_from_schema,
find_tool_name,
find_tool_properties,
)
@@ -319,15 +320,24 @@ class ParserEngine(Parser):
return json.dumps(args, ensure_ascii=False)
return args_json
def _is_valid_tool_name(self, name: str) -> bool:
if not self.parser_engine_config.validate_tool_names:
return True
if not self._tools:
return True
return find_tool_name(self._tools, name)
# ── Private helpers ─────────────────────────────────────────────
def _check_skip_tool_parsing(
self,
request: ChatCompletionRequest | ResponsesRequest,
) -> None:
tools = getattr(request, "tools", None)
if tools:
self._tools = tools
if not self.skip_tool_parsing:
tool_choice = getattr(request, "tool_choice", None)
tools = getattr(request, "tools", None)
if tool_choice == "none" and tools:
self.skip_tool_parsing = True
@@ -467,6 +477,7 @@ class ParserEngine(Parser):
output, this method starts the parser engine in ``CONTENT`` state
so it can parse content that has already had reasoning stripped.
"""
self._check_skip_tool_parsing(request)
_, parsed_content, tool_call_info = self._single_pass_parse(
content,
[],
@@ -586,6 +597,7 @@ class ParserEngine(Parser):
enable_auto_tools: bool = False,
model_output_token_ids: Sequence[int] = (),
) -> tuple[str | None, str | None, list[FunctionCall] | None]:
self._check_skip_tool_parsing(request)
reasoning, content, tool_call_info = self._single_pass_parse(
model_output,
model_output_token_ids,
@@ -705,7 +717,7 @@ class ParserEngine(Parser):
deltas: list[DeltaToolCall],
name: str | None,
) -> None:
if not name:
if not name or not self._is_valid_tool_name(name):
return
slot = self._tool_slots[idx]
slot.name = name
@@ -762,7 +774,7 @@ class ParserEngine(Parser):
if not slot.name_sent:
name = slot.name or self._try_extract_name(idx)
if name:
if name and self._is_valid_tool_name(name):
slot.name = name
slot.name_sent = True
self._ensure_tool_id(slot, name)
@@ -934,7 +946,7 @@ class ParserEngine(Parser):
else:
args_json = "{}"
if name:
if name and self._is_valid_tool_name(name):
self._ensure_tool_id(slot, name)
args_json = self._fix_arg_types(args_json, name)
tool_calls.append(
@@ -99,6 +99,9 @@ class ParserEngineConfig:
# .strip() content text when tool calls are present.
strip_content_whitespace_with_tools: bool = True
# Reject tool calls whose names are absent from the request tools.
validate_tool_names: bool = False
drop_tokens: frozenset[str] = field(default_factory=frozenset)
@cached_property
@@ -9,9 +9,15 @@ names so that :class:`ReasoningParserManager` and
from vllm.parser.engine.adapters import make_adapters
from vllm.parser.gemma4 import Gemma4Parser
from vllm.parser.minimax_m2 import MinimaxM2Parser
from vllm.parser.nemotron_v3 import NemotronV3Parser
from vllm.parser.qwen3 import Qwen3Parser
(
MinimaxM2ParserReasoningAdapter,
MinimaxM2ParserToolAdapter,
) = make_adapters(MinimaxM2Parser)
(
Gemma4ParserReasoningAdapter,
Gemma4ParserToolAdapter,
+180
View File
@@ -0,0 +1,180 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""MiniMax M2 parser for XML-style tool calls.
MiniMax M2 tool call format::
<minimax:tool_call><invoke name="get_weather">
<parameter name="city">Seattle</parameter>
</invoke></minimax:tool_call>
Each ``<invoke>`` block becomes one tool call. The argument body consists
of ``<parameter name="...">...</parameter>`` tags.
"""
from __future__ import annotations
import functools
import json
import regex as re
from vllm.parser.engine.events import EventType
from vllm.parser.engine.parser_engine import ParserEngine
from vllm.parser.engine.parser_engine_config import (
ParserEngineConfig,
ParserState,
Transition,
)
TOOL_CALL_START = "<minimax:tool_call>"
TOOL_CALL_END = "</minimax:tool_call>"
THINK_START = "<think>"
THINK_END = "</think>"
INVOKE_PREFIX_DQ = '<invoke name="'
INVOKE_PREFIX_SQ = "<invoke name='"
INVOKE_PREFIX_UNQUOTED = "<invoke name="
INVOKE_END = "</invoke>"
NAME_END_DQ = '">'
NAME_END_SQ = "'>"
NAME_END_UNQUOTED = ">"
_PARAM_RE = re.compile(
r"<\s*parameter\s+name\s*=\s*"
r"(?:\"(?P<dq_name>[^\"]*)\"|'(?P<sq_name>[^']*)'|(?P<bare_name>[^>\s]+))"
r"\s*>"
r"(?P<value>.*?)"
r"<\s*/\s*parameter\s*>",
re.DOTALL,
)
def _minimax_m2_arg_converter(raw_args: str, partial: bool) -> str:
params: dict[str, object] = {}
for match in _PARAM_RE.finditer(raw_args):
name = (
match.group("dq_name")
or match.group("sq_name")
or match.group("bare_name")
or ""
).strip()
if not name:
continue
params[name] = match.group("value").strip()
return json.dumps(params, ensure_ascii=False)
@functools.cache
def minimax_m2_config() -> ParserEngineConfig:
return ParserEngineConfig(
name="minimax_m2",
initial_state=ParserState.REASONING,
terminals={
"THINK_START": THINK_START,
"THINK_END": THINK_END,
"TOOL_START": TOOL_CALL_START,
"TOOL_END": TOOL_CALL_END,
"INVOKE_PREFIX_DQ": INVOKE_PREFIX_DQ,
"INVOKE_PREFIX_SQ": INVOKE_PREFIX_SQ,
"INVOKE_PREFIX_UNQUOTED": INVOKE_PREFIX_UNQUOTED,
"INVOKE_END": INVOKE_END,
"NAME_END_DQ": NAME_END_DQ,
"NAME_END_SQ": NAME_END_SQ,
"NAME_END_UNQUOTED": NAME_END_UNQUOTED,
},
token_id_terminals={
"THINK_START": THINK_START,
"THINK_END": THINK_END,
"TOOL_START": TOOL_CALL_START,
"TOOL_END": TOOL_CALL_END,
},
transitions={
(ParserState.REASONING, "THINK_START"): Transition(
ParserState.REASONING,
(),
),
(ParserState.REASONING, "THINK_END"): Transition(
ParserState.CONTENT,
(EventType.REASONING_END,),
),
(ParserState.CONTENT, "THINK_END"): Transition(
ParserState.CONTENT,
(),
),
(ParserState.REASONING, "TOOL_START"): Transition(
ParserState.TOOL_PREAMBLE,
(EventType.REASONING_END,),
),
(ParserState.CONTENT, "TOOL_START"): Transition(
ParserState.TOOL_PREAMBLE,
(),
),
(ParserState.TOOL_PREAMBLE, "TOOL_END"): Transition(
ParserState.CONTENT,
(),
),
(ParserState.TOOL_BETWEEN, "TOOL_END"): Transition(
ParserState.CONTENT,
(),
),
(ParserState.CONTENT, "TOOL_END"): Transition(
ParserState.CONTENT,
(),
),
(ParserState.TOOL_ARGS, "INVOKE_END"): Transition(
ParserState.TOOL_BETWEEN,
(EventType.TOOL_CALL_END,),
),
**{
(state, terminal): Transition(
ParserState.TOOL_NAME,
(EventType.TOOL_CALL_START,),
)
for state in (
ParserState.CONTENT,
ParserState.TOOL_PREAMBLE,
ParserState.TOOL_BETWEEN,
)
for terminal in (
"INVOKE_PREFIX_DQ",
"INVOKE_PREFIX_SQ",
"INVOKE_PREFIX_UNQUOTED",
)
},
**{
(ParserState.TOOL_NAME, terminal): Transition(
ParserState.TOOL_ARGS,
(),
)
for terminal in (
"NAME_END_DQ",
"NAME_END_SQ",
"NAME_END_UNQUOTED",
)
},
},
arg_converter=_minimax_m2_arg_converter,
stream_arg_deltas=True,
tool_args_json=False,
validate_tool_names=True,
)
class MinimaxM2Parser(ParserEngine):
"""MiniMax M2 parser backed by the declarative parser engine."""
def __init__(self, tokenizer, tools=None, **kwargs) -> None:
kwargs.setdefault("parser_engine_config", minimax_m2_config())
super().__init__(tokenizer, tools, **kwargs)
self._think_end_token_id = self.vocab.get(THINK_END)
def extract_content_ids(self, input_ids: list[int]) -> list[int]:
end_id = self._think_end_token_id
if end_id is None:
return []
for i in range(len(input_ids) - 1, -1, -1):
if input_ids[i] == end_id:
return input_ids[i + 1 :]
return []