[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
@@ -17,8 +17,6 @@ from vllm.engine.protocol import EngineClient
from vllm.entrypoints.chat_utils import (
ChatTemplateContentFormatOption,
ConversationMessage,
get_history_tool_calls_cnt,
get_tool_call_id_type,
make_tool_call_id,
)
from vllm.entrypoints.openai.chat_completion.protocol import (
@@ -170,8 +168,6 @@ class OpenAIServingChat(OpenAIServing):
if mc.generation_config not in ("auto", "vllm")
else getattr(mc, "override_generation_config", {}).get("max_new_tokens")
)
self.tool_call_id_type = get_tool_call_id_type(self.model_config)
# NOTE(woosuk): While OpenAI's chat completion API supports browsing
# for some models, currently vLLM doesn't support it. Please use the
# Responses API instead.
@@ -261,6 +257,7 @@ class OpenAIServingChat(OpenAIServing):
tokenizer,
request.tools,
chat_template_kwargs=chat_template_kwargs,
model_config=self.model_config,
)
result = await self.render_chat_request(request)
if isinstance(result, ErrorResponse):
@@ -433,11 +430,6 @@ class OpenAIServingChat(OpenAIServing):
else:
tool_choice_function_name = None
if self.tool_call_id_type == "kimi_k2":
history_tool_call_cnt = get_history_tool_calls_cnt(conversation)
else:
history_tool_call_cnt = 0
previous_texts = [""] * num_choices
try:
@@ -451,14 +443,10 @@ class OpenAIServingChat(OpenAIServing):
tokenizer,
request.tools,
chat_template_kwargs=chat_template_kwargs,
model_config=self.model_config,
)
for _ in range(num_choices)
]
for p in parsers:
if p is not None:
# NOTE: HarmonyParser ignores _stream_state (uses its own FSM).
p._stream_state.tool_call_id_type = self.tool_call_id_type
p._stream_state.history_tool_call_cnt = history_tool_call_cnt
else:
parsers = [None] * num_choices
except Exception as e:
@@ -842,10 +830,6 @@ class OpenAIServingChat(OpenAIServing):
)
choices: list[ChatCompletionResponseChoice] = []
if self.tool_call_id_type == "kimi_k2":
history_tool_call_cnt = get_history_tool_calls_cnt(conversation)
else:
history_tool_call_cnt = 0
role = self.get_chat_request_role(request)
tool_parser_cls = (
@@ -885,54 +869,26 @@ class OpenAIServingChat(OpenAIServing):
tool_calls = []
auto_tools_called = False
is_named_tool_choice = (
request.tool_choice is not None
and type(request.tool_choice) is ChatCompletionNamedToolChoiceParam
)
is_required_tool_choice = request.tool_choice == "required"
if (not self.enable_auto_tools or not tool_parser_cls) and (
not isinstance(request.tool_choice, ChatCompletionNamedToolChoiceParam)
and request.tool_choice != "required"
not is_named_tool_choice and not is_required_tool_choice
):
message = ChatMessage(role=role, reasoning=reasoning, content=content)
elif (
request.tool_choice
and type(request.tool_choice) is ChatCompletionNamedToolChoiceParam
):
tool_call_items = []
tool_calls = tool_calls or []
for tc in tool_calls:
if not tc.id:
tc.id = make_tool_call_id(
id_type=self.tool_call_id_type,
func_name=tc.name,
idx=history_tool_call_cnt,
)
tool_call_items.append(ToolCall(id=tc.id, function=tc))
history_tool_call_cnt += 1
elif is_named_tool_choice or is_required_tool_choice:
message = ChatMessage(
role=role,
reasoning=reasoning,
content=content or "",
tool_calls=tool_call_items,
)
elif request.tool_choice and request.tool_choice == "required":
tool_call_items = []
tool_calls = tool_calls or []
for tool_call in tool_calls:
if not tool_call.id:
tool_call.id = make_tool_call_id(
id_type=self.tool_call_id_type,
func_name=tool_call.name,
idx=history_tool_call_cnt,
)
tool_call_items.append(
ToolCall(id=tool_call.id, function=tool_call)
)
history_tool_call_cnt += 1
message = ChatMessage(
role=role,
content=content or "",
tool_calls=tool_call_items,
reasoning=reasoning,
tool_calls=[
ToolCall(id=tc.id or make_tool_call_id(), function=tc)
for tc in (tool_calls or [])
],
)
# if the request doesn't use tool choice
@@ -949,21 +905,14 @@ class OpenAIServingChat(OpenAIServing):
):
auto_tools_called = tool_calls is not None and len(tool_calls) > 0
if tool_calls:
tool_call_items = []
for tc in tool_calls:
if not tc.id:
tc.id = make_tool_call_id(
id_type=self.tool_call_id_type,
func_name=tc.name,
idx=history_tool_call_cnt,
)
tool_call_items.append(ToolCall(id=tc.id, function=tc))
history_tool_call_cnt += 1
message = ChatMessage(
role=role,
reasoning=reasoning,
content=content,
tool_calls=tool_call_items,
tool_calls=[
ToolCall(id=tc.id or make_tool_call_id(), function=tc)
for tc in tool_calls
],
)
else:
@@ -301,7 +301,6 @@ class ParsableContext(ConversationContext):
chat_template_content_format: ChatTemplateContentFormatOption,
response_parser: Parser | None = None,
enable_auto_tools: bool = False,
tool_call_id_type: str = "random",
):
self.num_prompt_tokens = 0
self.num_output_tokens = 0
@@ -314,20 +313,8 @@ class ParsableContext(ConversationContext):
self.num_init_messages = len(response_messages)
self.finish_reason: str | None = None
self.enable_auto_tools = enable_auto_tools
self.tool_call_id_type = tool_call_id_type
self.response_parser = response_parser
if self.response_parser is None and parser_cls is not None:
chat_template_kwargs = request.build_chat_params(
default_template=chat_template,
default_template_content_format=chat_template_content_format,
).chat_template_kwargs
self.response_parser = parser_cls(
tokenizer,
tools=request.tools,
chat_template_kwargs=chat_template_kwargs,
)
self.parser_cls = parser_cls
self.request = request
@@ -365,7 +352,6 @@ class ParsableContext(ConversationContext):
reasoning=reasoning,
content=content,
tool_calls=tool_calls,
tool_call_id_type=self.tool_call_id_type,
)
)
elif completion.text:
+1 -6
View File
@@ -30,7 +30,6 @@ from vllm.engine.protocol import EngineClient
from vllm.entrypoints.chat_utils import (
ChatCompletionMessageParam,
ChatTemplateContentFormatOption,
get_tool_call_id_type,
)
from vllm.entrypoints.mcp.tool_server import ToolServer
from vllm.entrypoints.openai.engine.protocol import (
@@ -222,9 +221,6 @@ class OpenAIServingResponses(OpenAIServing):
"For gpt-oss, we ignore --enable-auto-tool-choice "
"and always enable tool use."
)
self.tool_call_id_type = get_tool_call_id_type(self.model_config)
self.enable_auto_tools = enable_auto_tools
# HACK(woosuk): This is a hack. We should use a better store.
# FIXME: If enable_store=True, this may cause a memory leak since we
@@ -272,6 +268,7 @@ class OpenAIServingResponses(OpenAIServing):
tokenizer,
request.tools,
chat_template_kwargs=chat_template_kwargs,
model_config=self.model_config,
)
def _validate_generator_input(
@@ -493,7 +490,6 @@ class OpenAIServingResponses(OpenAIServing):
chat_template=self.chat_template,
chat_template_content_format=self.chat_template_content_format,
enable_auto_tools=self.enable_auto_tools,
tool_call_id_type=self.tool_call_id_type,
)
else:
context = SimpleContext(
@@ -1073,7 +1069,6 @@ class OpenAIServingResponses(OpenAIServing):
content=content,
tool_calls=tool_calls,
logprobs=logprobs,
tool_call_id_type=self.tool_call_id_type,
)
# Fallback when no parser is configured
+1 -7
View File
@@ -45,7 +45,6 @@ def build_response_output_items(
content: str | None,
tool_calls: list[FunctionCall] | None,
logprobs: list[Logprob] | None = None,
tool_call_id_type: str = "random",
) -> list[ResponseOutputItem]:
outputs: list[ResponseOutputItem] = []
@@ -86,12 +85,7 @@ def build_response_output_items(
ResponseFunctionToolCall(
id=f"fc_{random_uuid()}",
call_id=tool_call.id
if tool_call.id
else make_tool_call_id(
id_type=tool_call_id_type,
func_name=tool_call.name,
idx=idx,
),
or make_tool_call_id(func_name=tool_call.name, idx=idx),
type="function_call",
status="completed",
name=tool_call.name,
+46 -2
View File
@@ -11,6 +11,10 @@ from functools import cached_property
from openai.types.responses import ToolChoiceFunction
from pydantic import TypeAdapter, ValidationError
from vllm.entrypoints.chat_utils import (
get_tool_call_id_type,
make_tool_call_id,
)
from vllm.entrypoints.openai.chat_completion.protocol import (
ChatCompletionNamedToolChoiceParam,
ChatCompletionRequest,
@@ -24,6 +28,7 @@ from vllm.entrypoints.openai.engine.protocol import (
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.logger import init_logger
from vllm.parser.metrics import record_tool_parser_invocation
from vllm.parser.utils import count_history_tool_calls
from vllm.reasoning.abs_reasoning_parsers import ReasoningParser
from vllm.sampling_params import StructuredOutputsParams
from vllm.tokenizers import TokenizerLike
@@ -46,6 +51,7 @@ class StreamState:
previous_text: str = ""
previous_token_ids: list[int] = field(default_factory=list)
history_tool_call_cnt: int = 0
history_tool_call_cnt_initialized: bool = False
tool_call_id_type: str = "random"
# only used for "required" and "named tool" choices,
# tracks whether function name has been fully returned in the stream yet
@@ -108,6 +114,7 @@ class Parser:
tokenizer: TokenizerLike,
tools: list[Tool] | None = None,
*args,
model_config=None,
**kwargs,
):
self.model_tokenizer = tokenizer
@@ -124,7 +131,14 @@ class Parser:
self._reasoning_parser is None
or self._reasoning_parser.engine_based_streaming
) and (self._tool_parser is None or self._tool_parser.engine_based_streaming)
self._stream_state = StreamState(engine_based=self._engine_based)
self._stream_state = StreamState(
tool_call_id_type=(
get_tool_call_id_type(model_config)
if model_config is not None
else "random"
),
engine_based=self._engine_based,
)
@cached_property
def vocab(self) -> dict[str, int]:
@@ -149,6 +163,19 @@ class Parser:
def tool_parser(self, parser: ToolParser | None) -> None:
self._tool_parser = parser
def _initialize_history_tool_call_cnt(
self,
request: ChatCompletionRequest | ResponsesRequest,
) -> None:
state = self._stream_state
if state.history_tool_call_cnt_initialized:
return
if state.tool_call_id_type != "kimi_k2":
state.history_tool_call_cnt_initialized = True
return
state.history_tool_call_cnt = count_history_tool_calls(request)
state.history_tool_call_cnt_initialized = True
# ========== Reasoning Parser Methods ==========
@abstractmethod
@@ -375,6 +402,18 @@ class DelegatingParser(Parser):
return request.tool_choice.function.name
raise ValueError("Invalid tool_choice for function name extraction.")
def _make_tool_call_id(self, function_name: str) -> str | None:
state = self._stream_state
if state.tool_call_id_type != "kimi_k2":
return None
tool_call_id = make_tool_call_id(
id_type=state.tool_call_id_type,
func_name=function_name,
idx=state.history_tool_call_cnt,
)
state.history_tool_call_cnt += 1
return tool_call_id
def _extract_tool_calls(
self,
content: str | None,
@@ -404,9 +443,11 @@ class DelegatingParser(Parser):
if is_named_tool_choice and supports_required_and_named:
if content is None:
return [], None
function_name = self._get_function_name(request)
tool_calls.append(
FunctionCall(
name=self._get_function_name(request),
id=self._make_tool_call_id(function_name),
name=function_name,
arguments=content,
)
)
@@ -422,6 +463,7 @@ class DelegatingParser(Parser):
for tc in parsed_calls:
tool_calls.append(
FunctionCall(
id=self._make_tool_call_id(tc.name),
name=tc.name,
arguments=json.dumps(tc.parameters, ensure_ascii=False),
)
@@ -733,6 +775,7 @@ class DelegatingParser(Parser):
enable_auto_tools: bool = False,
model_output_token_ids: Sequence[int] = (),
) -> tuple[str | None, str | None, list[FunctionCall] | None]:
self._initialize_history_tool_call_cnt(request)
reasoning, content = self.extract_reasoning(model_output, request)
tool_calls, content = self._extract_tool_calls(
content=content,
@@ -750,6 +793,7 @@ class DelegatingParser(Parser):
*,
finished: bool,
) -> DeltaMessage | None:
self._initialize_history_tool_call_cnt(request)
state = self._stream_state
if not state.prompt_reasoning_checked and prompt_token_ids is not None:
+11 -2
View File
@@ -13,7 +13,7 @@ from typing import TYPE_CHECKING
import regex as re
from vllm.entrypoints.chat_utils import make_tool_call_id
from vllm.entrypoints.chat_utils import get_tool_call_id_type, make_tool_call_id
from vllm.entrypoints.openai.engine.protocol import (
DeltaFunctionCall,
DeltaMessage,
@@ -89,11 +89,18 @@ class ParserEngine(Parser):
tools: list[Tool] | None = None,
*,
parser_engine_config: ParserEngineConfig,
model_config=None,
**kwargs,
) -> None:
self.model_tokenizer = tokenizer
self._tools = tools
self._stream_state = StreamState()
self._stream_state = StreamState(
tool_call_id_type=(
get_tool_call_id_type(model_config)
if model_config is not None
else "random"
),
)
self._reasoning_parser = None
self._tool_parser = None
self.parser_engine_config = parser_engine_config
@@ -419,6 +426,7 @@ class ParserEngine(Parser):
*,
finished: bool,
) -> DeltaMessage | None:
self._initialize_history_tool_call_cnt(request)
if not self._prompt_streaming_prepared and prompt_token_ids is not None:
# NOTE: call the hook BEFORE setting the flag, because the hook
# may invoke ``_reset`` (e.g. via ``initialize_streaming``) which
@@ -658,6 +666,7 @@ class ParserEngine(Parser):
enable_auto_tools: bool = False,
model_output_token_ids: Sequence[int] = (),
) -> tuple[str | None, str | None, list[FunctionCall] | None]:
self._initialize_history_tool_call_cnt(request)
self._check_skip_tool_parsing(request)
reasoning, content, tool_call_info = self._single_pass_parse(
model_output,
+65
View File
@@ -0,0 +1,65 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from collections.abc import Iterable, Sequence
from openai.types.responses import ResponseFunctionToolCall
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.responses.protocol import (
ResponseInputOutputItem,
ResponsesRequest,
)
def count_tool_calls(tool_calls: object) -> int:
if tool_calls is None:
return 0
if isinstance(tool_calls, (str, bytes, dict)):
return 1
if isinstance(tool_calls, Iterable):
return sum(1 for _ in tool_calls)
return 1
def count_chat_history_tool_calls(
messages: Sequence[ChatCompletionMessageParam],
) -> int:
return sum(
count_tool_calls(msg.get("tool_calls"))
for msg in messages
if isinstance(msg, dict) and msg.get("role") == "assistant"
)
def count_response_history_tool_calls(
response_items: Sequence[ResponseInputOutputItem],
) -> int:
count = 0
for item in response_items:
if isinstance(item, ResponseFunctionToolCall):
count += 1
continue
if isinstance(item, dict):
item_type = item.get("type")
if item_type == "function_call":
count += 1
elif item.get("role") == "assistant":
count += count_tool_calls(item.get("tool_calls"))
return count
def count_history_tool_calls(
request: ChatCompletionRequest | ResponsesRequest,
) -> int:
if isinstance(request, ChatCompletionRequest):
return count_chat_history_tool_calls(request.messages)
request_input = request.input
if isinstance(request_input, str):
return 0
return count_response_history_tool_calls(request_input)