[Frontend] Support strict MiMo-V2.6 tool calling (#58019)

Co-authored-by: Codex <noreply@openai.com>
Co-authored-by: yuchuan <yuchuan.7streams@gmail.com>
Co-authored-by: Bugen Zhao <i@bugenzhao.com>
Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>
Signed-off-by: Ubospica <ubospica@gmail.com>
Signed-off-by: yuchuan <yuchuan.7streams@gmail.com>
Signed-off-by: Bugen Zhao <i@bugenzhao.com>
This commit is contained in:
Yixin Dong
2026-09-29 12:47:47 +00:00
committed by GitHub
co-authored by Codex yuchuan Bugen Zhao Claude Opus 5.5
parent 79a0307d74
commit dfc8e0f3e2
17 changed files with 243 additions and 15 deletions
+11
View File
@@ -458,6 +458,17 @@ Supported models:
Flags: `--tool-call-parser qwen3_xml` Flags: `--tool-call-parser qwen3_xml`
### MiMo-V2.6 Models (`mimo`)
Supported models:
* `XiaomiMiMo/MiMo-V2.6-Pro-RL`
* `XiaomiMiMo/MiMo-V2.6-Flash-RL`
Use the checkpoint's chat template with `--tool-call-parser mimo --reasoning-parser mimo --enable-auto-tool-choice`.
For schema-constrained tool arguments, set `strict: true` inside each tool's `function`, or use `--tool-strict-level parameter`.
The parser supports `auto`, `required`, named tools, and `parallel_tool_calls=False`, preserving leading/trailing whitespace in string arguments.
### Olmo 3 Models (`olmo3`) ### Olmo 3 Models (`olmo3`)
Olmo 3 models output tool calls in a format that is very similar to the one expected by the `pythonic` parser (see below), with a few differences. Each tool call is a pythonic string, but the parallel tool calls are newline-delimited, and the calls are wrapped within XML tags as `<function_calls>..</function_calls>`. In addition, the parser also allows JSON boolean and null literals (`true`, `false`, and `null`) in addition to the pythonic ones (`True`, `False`, and `None`). Olmo 3 models output tool calls in a format that is very similar to the one expected by the `pythonic` parser (see below), with a few differences. Each tool call is a pythonic string, but the parallel tool calls are newline-delimited, and the calls are wrapped within XML tags as `<function_calls>..</function_calls>`. In addition, the parser also allows JSON boolean and null literals (`true`, `false`, and `null`) in addition to the pythonic ones (`True`, `False`, and `None`).
+1 -1
View File
@@ -26,7 +26,7 @@ lm-format-enforcer == 0.11.3
llguidance >= 1.7.0, < 1.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "ppc64le" or platform_machine == "s390x" llguidance >= 1.7.0, < 1.8.0; platform_machine == "x86_64" or platform_machine == "arm64" or platform_machine == "aarch64" or platform_machine == "ppc64le" or platform_machine == "s390x"
outlines_core == 0.2.14 outlines_core == 0.2.14
lark == 1.2.2 lark == 1.2.2
xgrammar == 0.2.7; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x" or platform_machine == "ppc64le" xgrammar == 0.2.8; platform_machine == "x86_64" or platform_machine == "aarch64" or platform_machine == "arm64" or platform_machine == "s390x" or platform_machine == "ppc64le"
typing_extensions >= 4.10 typing_extensions >= 4.10
filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317 filelock >= 3.16.1 # need to contain https://github.com/tox-dev/filelock/pull/317
partial-json-parser # used for parsing partial JSON outputs partial-json-parser # used for parsing partial JSON outputs
+1 -1
View File
@@ -24,7 +24,7 @@ timm>=1.0.17
amd-quark==0.12.post1 amd-quark==0.12.post1
tilelang==0.1.10 tilelang==0.1.10
# Required apache-tvm-ffi matching tilelang version # Required apache-tvm-ffi matching tilelang version
apache-tvm-ffi==0.1.10 apache-tvm-ffi==0.1.11
# Required for faster safetensors model loading # Required for faster safetensors model loading
fastsafetensors >= 0.3.3 fastsafetensors >= 0.3.3
# Mooncake transfer engine for KV cache offloading/transfer (ROCm build, published on PyPI) # Mooncake transfer engine for KV cache offloading/transfer (ROCm build, published on PyPI)
+1 -1
View File
@@ -1319,7 +1319,7 @@ word2number==1.1
# via lm-eval # via lm-eval
wrapt==1.17.2 wrapt==1.17.2
# via smart-open # via smart-open
xgrammar==0.2.7 xgrammar==0.2.8
# via -r requirements/test/../common.txt # via -r requirements/test/../common.txt
xxhash==3.5.0 xxhash==3.5.0
# via # via
+1 -1
View File
@@ -1429,7 +1429,7 @@ word2number==1.1
# via lm-eval # via lm-eval
wrapt==1.17.2 wrapt==1.17.2
# via smart-open # via smart-open
xgrammar==0.2.7 xgrammar==0.2.8
# via # via
# -c requirements/common.txt # -c requirements/common.txt
# -r requirements/test/../common.txt # -r requirements/test/../common.txt
+2 -2
View File
@@ -43,7 +43,7 @@ anyio==4.13.0
# sse-starlette # sse-starlette
# starlette # starlette
# watchfiles # watchfiles
apache-tvm-ffi==0.1.10 apache-tvm-ffi==0.1.11
# via # via
# -c requirements/rocm.txt # -c requirements/rocm.txt
# tilelang # tilelang
@@ -1380,7 +1380,7 @@ word2number==1.1
# via lm-eval # via lm-eval
wrapt==2.1.2 wrapt==2.1.2
# via smart-open # via smart-open
xgrammar==0.2.7 xgrammar==0.2.8
# via # via
# -c requirements/common.txt # -c requirements/common.txt
# -r requirements/test/../common.txt # -r requirements/test/../common.txt
+1 -1
View File
@@ -1218,7 +1218,7 @@ werkzeug==3.1.8
# via schemathesis # via schemathesis
word2number==1.1 word2number==1.1
# via lm-eval # via lm-eval
xgrammar==0.2.7 xgrammar==0.2.8
# via # via
# -c requirements/common.txt # -c requirements/common.txt
# -r requirements/test/../common.txt # -r requirements/test/../common.txt
+24
View File
@@ -1175,3 +1175,27 @@ class TestNestedSchemaCoercion:
assert questions[0]["question"] == "Pick a color" assert questions[0]["question"] == "Pick a color"
assert questions[0]["multiSelect"] is False assert questions[0]["multiSelect"] is False
assert questions[0]["answer"] is None assert questions[0]["answer"] is None
@pytest.mark.parametrize("value", ["\n北京 &amp;\n", "null", "42", "", "a\nb"])
@pytest.mark.parametrize("chunk_size", [1, 7, 1000])
def test_mimo_preserves_verbatim_parameter_values(
mock_tokenizer, mock_request, value, chunk_size
):
from vllm.parser.mimo import MiMoParser
text = (
f"<tool_call><function=run><parameter=text>{value}</parameter>"
"</function></tool_call>"
)
parser = MiMoParser(mock_tokenizer, chat_template_kwargs={"enable_thinking": False})
result = parser.extract_tool_calls(text, mock_request)
assert json.loads(result.tool_calls[0].function.arguments) == {"text": value}
parser = MiMoParser(mock_tokenizer, chat_template_kwargs={"enable_thinking": False})
results = simulate_tool_streaming(
parser,
mock_request,
[text[i : i + chunk_size] for i in range(0, len(text), chunk_size)],
)
assert collect_function_name(results) == "run"
assert json.loads(collect_tool_arguments(results)) == {"text": value}
+39 -3
View File
@@ -15,6 +15,7 @@ from __future__ import annotations
import functools import functools
import json import json
from collections.abc import Callable
from dataclasses import dataclass, replace from dataclasses import dataclass, replace
from typing import Any from typing import Any
@@ -39,6 +40,7 @@ from vllm.parser.engine.registered_adapters import (
GraniteThinkingParser, GraniteThinkingParser,
InklingParser, InklingParser,
KimiK2Parser, KimiK2Parser,
MiMoParser,
MinimaxM2Parser, MinimaxM2Parser,
NemotronV3Parser, NemotronV3Parser,
Plamo3Parser, Plamo3Parser,
@@ -356,7 +358,12 @@ def _qwen3_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]:
] ]
def _qwen3_segments(scenario: Scenario) -> list[tuple[str, bool]]: def _qwen3_segments(
scenario: Scenario,
tool_segments: Callable[[ToolCallSpec], list[tuple[str, bool]]] = (
_qwen3_tool_segments
),
) -> list[tuple[str, bool]]:
segs: list[tuple[str, bool]] = [] segs: list[tuple[str, bool]] = []
if scenario.reasoning is not None: if scenario.reasoning is not None:
segs.append((scenario.reasoning, False)) segs.append((scenario.reasoning, False))
@@ -369,7 +376,7 @@ def _qwen3_segments(scenario: Scenario) -> list[tuple[str, bool]]:
segs.append((scenario.content, False)) segs.append((scenario.content, False))
if scenario.tool_calls: if scenario.tool_calls:
for tc in scenario.tool_calls: for tc in scenario.tool_calls:
segs.extend(_qwen3_tool_segments(tc)) segs.extend(tool_segments(tc))
return segs return segs
@@ -389,6 +396,9 @@ def _build_qwen3(
parser_cls: type = Qwen3Parser, parser_cls: type = Qwen3Parser,
strip_trailing_ws: bool = False, strip_trailing_ws: bool = False,
validate: bool = True, validate: bool = True,
tool_segments: Callable[[ToolCallSpec], list[tuple[str, bool]]] = (
_qwen3_tool_segments
),
) -> Sample: ) -> Sample:
expected_reasoning: str | None expected_reasoning: str | None
if scenario.reasoning is not None: if scenario.reasoning is not None:
@@ -403,7 +413,7 @@ def _build_qwen3(
sample_id=f"{name}-{scenario.id}", sample_id=f"{name}-{scenario.id}",
description=scenario.description, description=scenario.description,
vocab=_QWEN3_VOCAB, vocab=_QWEN3_VOCAB,
segments=_qwen3_segments(scenario), segments=_qwen3_segments(scenario, tool_segments),
expected_reasoning=expected_reasoning, expected_reasoning=expected_reasoning,
expected_content=_qwen3_expected_content(scenario), expected_content=_qwen3_expected_content(scenario),
expected_tool_calls=_expected_tc(scenario), expected_tool_calls=_expected_tc(scenario),
@@ -414,6 +424,31 @@ def _build_qwen3(
return sample return sample
# ── MiMo (compact Qwen3 XML, no newlines between tags) ───────────────
def _mimo_tool_segments(tc: ToolCallSpec) -> list[tuple[str, bool]]:
parts = [f"<function={tc.name}>"]
for key, value in tc.arguments.items():
parts.append(f"<parameter={key}>{_qwen3_arg_value(value)}</parameter>")
parts.append("</function>")
return [
("<tool_call>", True),
("".join(parts), False),
("</tool_call>", True),
]
def _build_mimo(scenario: Scenario, validate: bool = True) -> Sample:
return _build_qwen3(
scenario,
name="mimo",
parser_cls=MiMoParser,
validate=validate,
tool_segments=_mimo_tool_segments,
)
# ── MiniMax M2 (XML invoke format, starts in REASONING) ────────────── # ── MiniMax M2 (XML invoke format, starts in REASONING) ──────────────
_MINIMAX_M2_VOCAB: dict[str, int] = { _MINIMAX_M2_VOCAB: dict[str, int] = {
@@ -1232,6 +1267,7 @@ _BUILDERS: dict[str, Any] = {
"glm47_moe": _build_glm47_moe, "glm47_moe": _build_glm47_moe,
"kimi_k2": _build_kimi_k2, "kimi_k2": _build_kimi_k2,
"qwen3": _build_qwen3, "qwen3": _build_qwen3,
"mimo": _build_mimo,
"inkling": _build_inkling, "inkling": _build_inkling,
"plamo3": _build_plamo3, "plamo3": _build_plamo3,
} }
@@ -40,6 +40,7 @@ from vllm.tool_parsers.hermes_tool_parser import Hermes2ProToolParser
from vllm.tool_parsers.kimi_k2_tool_parser import KimiK2ToolParser from vllm.tool_parsers.kimi_k2_tool_parser import KimiK2ToolParser
from vllm.tool_parsers.kimi_k3_tool_parser import KimiK3ToolParser from vllm.tool_parsers.kimi_k3_tool_parser import KimiK3ToolParser
from vllm.tool_parsers.llama_tool_parser import Llama3JsonToolParser from vllm.tool_parsers.llama_tool_parser import Llama3JsonToolParser
from vllm.tool_parsers.mimo_tool_parser import MiMoToolParser
from vllm.tool_parsers.minimax_m2_tool_parser import MinimaxM2ToolParser from vllm.tool_parsers.minimax_m2_tool_parser import MinimaxM2ToolParser
from vllm.tool_parsers.plamo3_engine_tool_parser import Plamo3EngineToolParser from vllm.tool_parsers.plamo3_engine_tool_parser import Plamo3EngineToolParser
from vllm.tool_parsers.qwen3_engine_tool_parser import Qwen3EngineToolParser from vllm.tool_parsers.qwen3_engine_tool_parser import Qwen3EngineToolParser
@@ -665,6 +666,7 @@ def test_get_model_structural_tag_supports_named_tool_choice(
(Llama3JsonToolParser, "llama"), (Llama3JsonToolParser, "llama"),
(MinimaxM2ToolParser, "minimax"), (MinimaxM2ToolParser, "minimax"),
(Qwen3EngineToolParser, "qwen_3_coder"), (Qwen3EngineToolParser, "qwen_3_coder"),
(MiMoToolParser, "mimo"),
], ],
) )
def test_tool_parsers_declare_matching_xgrammar_builtin_model(parser_cls, model): def test_tool_parsers_declare_matching_xgrammar_builtin_model(parser_cls, model):
@@ -727,7 +729,7 @@ def test_get_structural_tag_disables_reasoning(
@pytest.mark.parametrize( @pytest.mark.parametrize(
"parser_cls", [Qwen3EngineToolParser, DeepSeekV41EngineToolParser] "parser_cls", [Qwen3EngineToolParser, DeepSeekV41EngineToolParser, MiMoToolParser]
) )
def test_unified_parser_get_structural_tag_disables_reasoning( def test_unified_parser_get_structural_tag_disables_reasoning(
parser_cls, parser_cls,
@@ -1167,3 +1169,78 @@ def test_tool_strict_level_from_name():
ToolStrictLevel.from_name("strict") ToolStrictLevel.from_name("strict")
with pytest.raises(ValueError, match="expected one of auto, function, parameter"): with pytest.raises(ValueError, match="expected one of auto, function, parameter"):
ToolStrictLevel.from_name("off") ToolStrictLevel.from_name("off")
@pytest.mark.parametrize("policy", ["auto", "required", "named"])
def test_mimo_strict_compact_xml(policy):
tools = [
ChatCompletionToolsParam(
type="function",
function={
"name": "run",
"strict": True,
"parameters": {
"type": "object",
"properties": {"n": {"type": "integer", "minimum": 1}},
"required": ["n"],
"additionalProperties": False,
},
},
)
]
choice = (
ChatCompletionNamedToolChoiceParam(
function=ChatCompletionNamedFunction(name="run")
)
if policy == "named"
else policy
)
tag = get_model_structural_tag("mimo", tools, choice, reasoning=False)
grammar = Grammar.from_structural_tag(tag)
call = "<tool_call><function=run><parameter=n>2</parameter></function></tool_call>"
assert _is_grammar_accept_string(grammar, call)
assert _is_grammar_accept_string(grammar, call * 2) == (policy != "named")
assert _is_grammar_accept_string(grammar, "answer") == (policy == "auto")
for invalid in [
call.replace("function=run", "function=unknown"),
call.replace("<parameter=n>2</parameter>", ""),
call.replace(">2</parameter>", ">bad</parameter>"),
call.replace(">2</parameter>", ">0</parameter>"),
call.replace("<parameter=n>", "<parameter=other>"),
call.replace("</function>", "<parameter=n>3</parameter></function>"),
"<think>plan</think>" + call,
]:
assert not _is_grammar_accept_string(grammar, invalid)
tools[0].function.strict = False
relaxed = Grammar.from_structural_tag(
get_model_structural_tag("mimo", tools, "required", reasoning=False)
)
assert _is_grammar_accept_string(
relaxed, call.replace("<parameter=n>", "<parameter=other>")
)
@pytest.mark.parametrize("policy", ["auto", "required"])
def test_mimo_request_limits_parallel_calls(sample_tools_strict, policy):
from vllm.parser.parser_manager import ParserManager
request = ChatCompletionRequest(
model="mimo",
messages=[],
tools=[t.model_dump() for t in sample_tools_strict],
tool_choice=policy,
parallel_tool_calls=False,
)
parser_cls = ParserManager.get_parser(
tool_parser_name="mimo", enable_auto_tools=True
)
parser = parser_cls(MagicMock(), tools=sample_tools_strict)
adjusted = parser.adjust_request(request)
grammar = Grammar.from_structural_tag(adjusted.structured_outputs.structural_tag)
call = (
"<tool_call><function=get_weather><parameter=city>北京</parameter>"
"</function></tool_call>"
)
assert _is_grammar_accept_string(grammar, call)
assert not _is_grammar_accept_string(grammar, call * 2)
assert not _is_grammar_accept_string(grammar, call + "extra text")
@@ -17,6 +17,7 @@ from vllm.parser.granite import GraniteParser
from vllm.parser.granite_thinking import GraniteThinkingParser from vllm.parser.granite_thinking import GraniteThinkingParser
from vllm.parser.inkling import InklingParser from vllm.parser.inkling import InklingParser
from vllm.parser.kimi_k2 import KimiK2Parser from vllm.parser.kimi_k2 import KimiK2Parser
from vllm.parser.mimo import MiMoParser
from vllm.parser.minimax_m2 import MinimaxM2Parser from vllm.parser.minimax_m2 import MinimaxM2Parser
from vllm.parser.mistral import MistralParser from vllm.parser.mistral import MistralParser
from vllm.parser.nemotron_v3 import NemotronV3Parser from vllm.parser.nemotron_v3 import NemotronV3Parser
@@ -69,6 +70,11 @@ from vllm.parser.seed_oss import SeedOssParser
Qwen3ParserToolAdapter, Qwen3ParserToolAdapter,
) = make_adapters(Qwen3Parser) ) = make_adapters(Qwen3Parser)
(
MiMoParserReasoningAdapter,
MiMoParserToolAdapter,
) = make_adapters(MiMoParser)
( (
SeedOssParserReasoningAdapter, SeedOssParserReasoningAdapter,
SeedOssParserToolAdapter, SeedOssParserToolAdapter,
+42
View File
@@ -0,0 +1,42 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""MiMo's compact XML parameters contain verbatim string values."""
import json
from dataclasses import replace
from vllm.parser.qwen3 import (
_PARAM_RE,
_PARTIAL_PARAM_RE,
CHATML_TURN_BOUNDARIES,
Qwen3Parser,
qwen3_config,
)
def _mimo_arg_converter(raw_args: str, partial: bool) -> str:
params = {m.group(1): m.group(2) for m in _PARAM_RE.finditer(raw_args)}
if partial:
match = _PARTIAL_PARAM_RE.search(_PARAM_RE.sub("", raw_args))
if match and match.group(1):
params[match.group(1)] = match.group(2)
return json.dumps(params, ensure_ascii=False)
class MiMoParser(Qwen3Parser):
def __init__(self, tokenizer, tools=None, **kwargs):
thinking = (kwargs.get("chat_template_kwargs") or {}).get(
"enable_thinking", True
)
kwargs.setdefault(
"parser_engine_config",
replace(
qwen3_config(
thinking=thinking,
name="mimo",
turn_boundary_tokens=CHATML_TURN_BOUNDARIES,
),
arg_converter=_mimo_arg_converter,
),
)
super().__init__(tokenizer, tools, **kwargs)
+2 -2
View File
@@ -109,8 +109,8 @@ _REASONING_PARSERS_TO_REGISTER = {
"K2HorizonReasoningParser", "K2HorizonReasoningParser",
), ),
"mimo": ( "mimo": (
"qwen3_engine_reasoning_parser", "mimo_engine_reasoning_parser",
"Qwen3ParserReasoningAdapter", "MiMoParserReasoningAdapter",
), ),
"minimax_m2": ( "minimax_m2": (
"minimax_m2_reasoning_parser", "minimax_m2_reasoning_parser",
@@ -0,0 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.parser.engine.registered_adapters import MiMoParserReasoningAdapter
__all__ = ["MiMoParserReasoningAdapter"]
+2 -2
View File
@@ -143,8 +143,8 @@ _TOOL_PARSERS_TO_REGISTER = {
"LongcatFlashToolParser", "LongcatFlashToolParser",
), ),
"mimo": ( "mimo": (
"qwen3_engine_tool_parser", "mimo_tool_parser",
"Qwen3EngineToolParser", "MiMoToolParser",
), ),
"minimax_m2": ( "minimax_m2": (
"minimax_m2_tool_parser", "minimax_m2_tool_parser",
+25
View File
@@ -0,0 +1,25 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from xgrammar.structural_tag import TriggeredTagsFormat
from vllm.parser.engine.registered_adapters import MiMoParserToolAdapter
from vllm.tool_parsers.tool_strict_level import ToolStrictLevel
class MiMoToolParser(MiMoParserToolAdapter): # type: ignore[valid-type, misc]
structural_tag_model = "mimo"
def get_structural_tag(
self, request, *, reasoning=False, strict_level=ToolStrictLevel.AUTO
):
tag = super().get_structural_tag(
request, reasoning=reasoning, strict_level=strict_level
)
if (
tag is not None
and request.parallel_tool_calls is False
and isinstance(tag.format, TriggeredTagsFormat)
):
tag.format.stop_after_first = True
return tag
@@ -66,6 +66,7 @@ XGRAMMAR_BUILTIN_STRUCTURAL_TAG_MODELS = frozenset(
"deepseek_v3_1", "deepseek_v3_1",
"qwen_3_5", "qwen_3_5",
"qwen_3_coder", "qwen_3_coder",
"mimo",
"qwen_3", "qwen_3",
"deepseek_v3_2", "deepseek_v3_2",
"glm_4_7", "glm_4_7",