[Frontend] Fix Kimi K2 tool call IDs for required tool choice (#46344)

Signed-off-by: chaunceyjiang <chaunceyjiang@gmail.com>
This commit is contained in:
Chauncey
2026-06-24 19:59:40 +00:00
committed by GitHub
parent 49f2104c53
commit 84c2f9f0fb
11 changed files with 406 additions and 178 deletions
@@ -24,6 +24,7 @@ tools = [
"description": "Get the current weather in a given location",
"parameters": {
"type": "object",
"strict": True,
"properties": {
"city": {
"type": "string",
@@ -215,80 +216,6 @@ async def test_function_tool_use(
assert len(reasoning) > 0
@pytest.fixture(scope="module")
def k2_server():
args = [
# use half precision for speed and memory savings in CI environment
"--dtype",
"half",
"--enable-auto-tool-choice",
"--structured-outputs-config.backend",
"xgrammar",
"--tool-call-parser",
"hermes",
"--reasoning-parser",
"qwen3",
"--gpu-memory-utilization",
"0.4",
] + ROCM_EXTRA_ARGS
# Test kimi_k2 tool use tool_id format by overriding model_type.
# is_deepseek_mla safely returns False via getattr when kv_lora_rank
# is absent from the underlying config.
with RemoteOpenAIServer(
MODEL_NAME,
args,
env_dict=ROCM_ENV_OVERRIDES,
override_hf_configs={"model_type": "kimi_k2"},
) as remote_server:
yield remote_server
@pytest_asyncio.fixture
async def k2_client(k2_server):
async with k2_server.get_async_client() as async_client:
yield async_client
@pytest.mark.asyncio
@pytest.mark.skip(reason="Skipping Kimi K2 tool ID test")
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("stream", [True, False])
@pytest.mark.parametrize("tool_choice", ["required"])
async def test_tool_id_kimi_k2(
k2_client: openai.AsyncOpenAI, model_name: str, stream: bool, tool_choice: str
):
if not stream:
# Non-streaming test
chat_completion = await k2_client.chat.completions.create(
messages=messages, model=model_name, tools=tools, tool_choice=tool_choice
)
assert chat_completion.choices[0].message.tool_calls is not None
assert len(chat_completion.choices[0].message.tool_calls) > 0
assert chat_completion.choices[0].message.tool_calls[0].id in [
"functions.get_current_weather:0",
"functions.get_forecast:1",
]
else:
# Streaming test
output_stream = await k2_client.chat.completions.create(
messages=messages,
model=model_name,
tools=tools,
tool_choice=tool_choice,
stream=True,
)
output = []
async for chunk in output_stream:
if chunk.choices and chunk.choices[0].delta.tool_calls:
output.extend(chunk.choices[0].delta.tool_calls)
for o in output:
assert o.id is None or o.id in [
"functions.get_current_weather:0",
"functions.get_forecast:1",
]
@pytest.mark.asyncio
@pytest.mark.parametrize("model_name", [MODEL_NAME])
@pytest.mark.parametrize("arguments", ["{}", ""])
@@ -183,11 +183,19 @@ def _make_request_output(
def _make_context(parser_cls, **overrides):
# ParsableContext no longer lazily builds a parser from ``parser_cls``;
# the caller (here, the serving layer in production) must supply one.
request = overrides.get("request", _make_request())
response_parser = overrides.pop("response_parser", None)
if response_parser is None and parser_cls is not None:
response_parser = parser_cls(MagicMock(), request.tools)
defaults = dict(
tokenizer=MagicMock(),
parser_cls=parser_cls,
response_parser=response_parser,
response_messages=[],
request=_make_request(),
request=request,
available_tools=None,
chat_template=None,
chat_template_content_format="auto",
+149 -2
View File
@@ -3,6 +3,7 @@
import json
import os
from types import SimpleNamespace
import pytest
@@ -13,7 +14,9 @@ os.environ[_STRICT_TOOL_CALLING_ENV] = "0"
from vllm.entrypoints.openai.chat_completion.protocol import ( # noqa: E402
ChatCompletionRequest,
)
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest # noqa: E402
from vllm.parser.abstract_parser import DelegatingParser # noqa: E402
from vllm.parser.utils import count_history_tool_calls # noqa: E402
from vllm.reasoning.basic_parsers import ( # noqa: E402
BaseThinkingReasoningParser,
)
@@ -82,12 +85,55 @@ TOOLS = [
]
def make_parser(tokenizer, reasoning=False, tool=False):
KIMI_K2_MODEL_CONFIG = SimpleNamespace(
hf_text_config=SimpleNamespace(model_type="kimi_k2"),
hf_overrides=None,
)
HISTORY_MESSAGES = [
{"role": "user", "content": "first"},
{
"role": "assistant",
"content": None,
"tool_calls": [
{
"id": "functions.get_current_weather:0",
"type": "function",
"function": {
"name": "get_current_weather",
"arguments": "{}",
},
},
{
"id": "functions.get_forecast:1",
"type": "function",
"function": {
"name": "get_forecast",
"arguments": "{}",
},
},
],
},
{
"role": "tool",
"tool_call_id": "functions.get_current_weather:0",
"content": "{}",
},
{
"role": "tool",
"tool_call_id": "functions.get_forecast:1",
"content": "{}",
},
{"role": "user", "content": "again"},
]
def make_parser(tokenizer, reasoning=False, tool=False, **kwargs):
class TestParser(DelegatingParser):
reasoning_parser_cls = ThinkReasoningParser if reasoning else None
tool_parser_cls = Hermes2ProToolParser if tool else None
return TestParser(tokenizer)
return TestParser(tokenizer, **kwargs)
@pytest.mark.parametrize(
@@ -232,6 +278,107 @@ def test_parse_required_tool_choice(tokenizer):
assert json.loads(tool_calls[1].arguments) == {"timezone": "UTC"}
def test_parse_required_tool_choice_kimi_k2_ids(tokenizer):
parser = make_parser(
tokenizer, reasoning=False, tool=True, model_config=KIMI_K2_MODEL_CONFIG
)
functions_json = json.dumps(
[
{"name": "get_current_weather", "parameters": {"city": "Dallas"}},
{"name": "get_forecast", "parameters": {"city": "Dallas", "days": 2}},
]
)
request = make_request(tools=TOOLS, tool_choice="required")
_, content, tool_calls = parser.parse(
functions_json, request, enable_auto_tools=True
)
assert content is None
assert tool_calls is not None
assert [tc.id for tc in tool_calls] == [
"functions.get_current_weather:0",
"functions.get_forecast:1",
]
def test_parse_required_tool_choice_kimi_k2_ids_after_history(tokenizer):
parser = make_parser(
tokenizer, reasoning=False, tool=True, model_config=KIMI_K2_MODEL_CONFIG
)
functions_json = json.dumps(
[{"name": "get_current_weather", "parameters": {"city": "Dallas"}}]
)
request = make_request(
messages=HISTORY_MESSAGES,
tools=TOOLS,
tool_choice="required",
)
_, _, tool_calls = parser.parse(functions_json, request, enable_auto_tools=True)
assert tool_calls is not None
assert tool_calls[0].id == "functions.get_current_weather:2"
def test_count_history_tool_calls_responses_request():
request = ResponsesRequest.model_validate(
{
"model": "test-model",
"input": [
{
"type": "function_call",
"call_id": "call_0",
"name": "get_current_weather",
"arguments": "{}",
},
{
"type": "function_call",
"call_id": "call_1",
"name": "get_forecast",
"arguments": "{}",
},
],
}
)
assert count_history_tool_calls(request) == 2
def test_parse_required_tool_choice_random_ids_deferred(tokenizer):
parser = make_parser(tokenizer, reasoning=False, tool=True)
functions_json = json.dumps(
[{"name": "get_current_weather", "parameters": {"city": "Dallas"}}]
)
request = make_request(
messages=HISTORY_MESSAGES,
tools=TOOLS,
tool_choice="required",
)
_, _, tool_calls = parser.parse(functions_json, request, enable_auto_tools=True)
assert tool_calls is not None
assert tool_calls[0].id is None
def test_parse_named_tool_choice_kimi_k2_id(tokenizer):
parser = make_parser(
tokenizer, reasoning=False, tool=True, model_config=KIMI_K2_MODEL_CONFIG
)
request = make_request(
tools=TOOLS,
tool_choice={
"type": "function",
"function": {"name": "get_weather"},
},
)
_, content, tool_calls = parser.parse(
TOOL_ARGUMENTS, request, enable_auto_tools=True
)
assert content is None
assert tool_calls is not None
assert tool_calls[0].id == "functions.get_weather:0"
def test_parse_named_tool_choice_content_none(tokenizer):
parser = make_parser(tokenizer, reasoning=False, tool=True)
request = make_request(
+106 -2
View File
@@ -2,6 +2,7 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import json
from types import SimpleNamespace
import pytest
@@ -47,6 +48,36 @@ TOOLS = [
]
KIMI_K2_MODEL_CONFIG = SimpleNamespace(
hf_text_config=SimpleNamespace(model_type="kimi_k2"),
hf_overrides=None,
)
HISTORY_MESSAGES = [
{"role": "user", "content": "first"},
{
"role": "assistant",
"content": None,
"tool_calls": [
{
"id": "functions.get_current_weather:0",
"type": "function",
"function": {
"name": "get_current_weather",
"arguments": "{}",
},
}
],
},
{
"role": "tool",
"tool_call_id": "functions.get_current_weather:0",
"content": "{}",
},
{"role": "user", "content": "again"},
]
@pytest.fixture
def request_obj():
return ChatCompletionRequest(
@@ -57,12 +88,12 @@ def request_obj():
)
def make_parser(tokenizer, reasoning=False, tool=False):
def make_parser(tokenizer, reasoning=False, tool=False, **kwargs):
class TestParser(DelegatingParser):
reasoning_parser_cls = ThinkReasoningParser if reasoning else None
tool_parser_cls = Hermes2ProToolParser if tool else None
return TestParser(tokenizer)
return TestParser(tokenizer, **kwargs)
def stream_text(parser, tokenizer, text, request, prompt_token_ids=None):
@@ -365,3 +396,76 @@ def test_parse_delta_tool_choice_none_with_reasoning(tokenizer, request_obj):
assert len(tool_calls) == 0
assert "<tool_call>" in content
assert "get_weather" in content
def test_parse_delta_required_tool_choice_kimi_k2_ids(tokenizer, request_obj):
parser = make_parser(
tokenizer, reasoning=False, tool=True, model_config=KIMI_K2_MODEL_CONFIG
)
request = request_obj.model_copy(update={"tool_choice": "required"})
output = json.dumps(
[
{
"name": "get_current_weather",
"parameters": {"city": "Dallas"},
}
]
)
results: list[DeltaMessage | None] = []
prompt_token_ids: list[int] | None = []
for i in range(0, len(output), 3):
chunk = output[i : i + 3]
results.append(
parser.parse_delta(
chunk,
[],
request,
prompt_token_ids=prompt_token_ids,
finished=False,
)
)
prompt_token_ids = None
_, content, tool_calls = collect_fields(results)
assert content == ""
assert any(tc.id == "functions.get_current_weather:0" for tc in tool_calls)
assert all(tc.id in (None, "functions.get_current_weather:0") for tc in tool_calls)
def test_parse_delta_required_tool_choice_kimi_k2_ids_after_history(
tokenizer, request_obj
):
parser = make_parser(
tokenizer, reasoning=False, tool=True, model_config=KIMI_K2_MODEL_CONFIG
)
request = request_obj.model_copy(
update={"messages": HISTORY_MESSAGES, "tool_choice": "required"}
)
output = json.dumps(
[
{
"name": "get_current_weather",
"parameters": {"city": "Dallas"},
}
]
)
results: list[DeltaMessage | None] = []
prompt_token_ids: list[int] | None = []
for i in range(0, len(output), 3):
chunk = output[i : i + 3]
results.append(
parser.parse_delta(
chunk,
[],
request,
prompt_token_ids=prompt_token_ids,
finished=False,
)
)
prompt_token_ids = None
_, _, tool_calls = collect_fields(results)
assert any(tc.id == "functions.get_current_weather:1" for tc in tool_calls)
assert all(tc.id in (None, "functions.get_current_weather:1") for tc in tool_calls)