[Refactor][Mistral] Extract parsing logic into MistralParser (#44596)

Signed-off-by: sfeng33 <4florafeng@gmail.com>
This commit is contained in:
Flora Feng
2026-06-10 00:12:23 +00:00
committed by GitHub
parent dac9e9a640
commit d82ac00923
7 changed files with 136 additions and 729 deletions
+3 -20
View File
@@ -43,7 +43,6 @@ from vllm.tool_parsers.streaming import (
extract_required_tool_call_streaming,
)
from vllm.utils import random_uuid
from vllm.utils.mistral import is_mistral_tool_parser
logger = init_logger(__name__)
@@ -546,14 +545,6 @@ class DelegatingParser(Parser):
if tool_parser is None:
return [], content
# When the Mistral grammar factory injected structured outputs,
# let the parser handle the output.
use_mistral_tool_parser = (
is_mistral_tool_parser(type(tool_parser))
and isinstance(request, ChatCompletionRequest)
and request._grammar_from_tool_parser
)
supports_required_and_named = tool_parser.supports_required_and_named
is_named_tool_choice = request.tool_choice and isinstance(
request.tool_choice,
@@ -570,11 +561,7 @@ class DelegatingParser(Parser):
)
tool_calls = list[FunctionCall]()
if (
is_named_tool_choice
and supports_required_and_named
and not use_mistral_tool_parser
):
if is_named_tool_choice and supports_required_and_named:
if content is None:
return [], None
tool_calls.append(
@@ -584,11 +571,7 @@ class DelegatingParser(Parser):
)
)
content = None
elif (
is_required_tool_choice
and supports_required_and_named
and not use_mistral_tool_parser
):
elif is_required_tool_choice and supports_required_and_named:
# "required" with standard JSON-based parsing
parsed_calls = []
with contextlib.suppress(ValidationError):
@@ -604,7 +587,7 @@ class DelegatingParser(Parser):
)
)
content = None
elif is_auto_tool_choice or use_mistral_tool_parser:
elif is_auto_tool_choice:
# Automatic Tool Call Parsing (also used as fallback for
# required/named when supports_required_and_named=False)
tool_call_info = tool_parser.extract_tool_calls(
+77
View File
@@ -0,0 +1,77 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from __future__ import annotations
from typing import TYPE_CHECKING
from vllm.entrypoints.openai.engine.protocol import DeltaMessage, FunctionCall
from vllm.parser.abstract_parser import DelegatingParser
if TYPE_CHECKING:
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionRequest,
)
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
class MistralParser(DelegatingParser):
def __init__(self, tokenizer, tools=None, *args, **kwargs):
super().__init__(tokenizer, tools, *args, **kwargs)
from vllm.tool_parsers.mistral_tool_parser import MistralToolParser
if not isinstance(self._tool_parser, MistralToolParser):
raise ValueError(
"MistralParser requires --tool-call-parser mistral, "
f"got {self._tool_parser.__class__.__name__}."
)
def _maybe_force_auto_tool_parsing(
self, request: ChatCompletionRequest | ResponsesRequest
) -> None:
# When the Mistral grammar factory injected structured outputs,
# the model emits v11+ format ([TOOL_CALLS]name{args}) that the
# named/required parsers can't handle. Disable them so all
# tool_choice modes fall back to auto tool parsing via
# extract_tool_calls.
if getattr(request, "_grammar_from_tool_parser", False):
assert self._tool_parser is not None
self._tool_parser.supports_required_and_named = False
def parse(
self,
model_output: str,
request: ChatCompletionRequest | ResponsesRequest,
enable_auto_tools: bool = False,
) -> tuple[str | None, str | None, list[FunctionCall] | None]:
self._maybe_force_auto_tool_parsing(request)
reasoning, content, tool_calls = super().parse(
model_output, request, enable_auto_tools
)
if tool_calls:
from vllm.tool_parsers.mistral_tool_parser import MistralToolCall
# Named/required tool_choice builds FunctionCalls without
# ID, backfill with Mistral-format IDs.
for tc in tool_calls:
if not tc.id:
tc.id = MistralToolCall.generate_random_id()
return reasoning, content, tool_calls
def parse_delta(
self,
delta_text: str,
delta_token_ids: list[int],
request: ChatCompletionRequest | ResponsesRequest,
prompt_token_ids: list[int] | None = None,
*,
finished: bool,
) -> DeltaMessage | None:
self._maybe_force_auto_tool_parsing(request)
return super().parse_delta(
delta_text,
delta_token_ids,
request,
prompt_token_ids,
finished=finished,
)
+9
View File
@@ -106,6 +106,15 @@ class ParserManager:
if reasoning_parser_cls is None and tool_parser_cls is None:
return None
from vllm.utils.mistral import is_mistral_tool_parser
if is_mistral_tool_parser(tool_parser_cls):
from vllm.parser.mistral import MistralParser
MistralParser.reasoning_parser_cls = reasoning_parser_cls
MistralParser.tool_parser_cls = tool_parser_cls
return MistralParser
from vllm.parser.abstract_parser import DelegatingParser
r_cls = reasoning_parser_cls