[Frontend] Support OpenAI Responses API namespace tools (#47024)

Signed-off-by: zhongjing123 <jimzhong5193@gmail.com>
Co-authored-by: zhongjing123 <jimzhong5193@gmail.com>
This commit is contained in:
Zhong
2026-07-06 06:21:27 +00:00
committed by GitHub
co-authored by zhongjing123
parent 990c2a0187
commit 98ba9b9583
8 changed files with 383 additions and 22 deletions
@@ -349,6 +349,7 @@ class ParsableContext(ConversationContext):
reasoning=reasoning,
content=content,
tool_calls=tool_calls,
tools=self.request.tools,
)
)
elif completion.text:
+13 -4
View File
@@ -592,10 +592,19 @@ class ResponsesRequest(OpenAIBaseModel):
)
elif is_named_tool_choice and tools is not None:
tool_name = tool_choice.get("name")
tool_names = {
t.get("name") if isinstance(t, dict) else getattr(t, "name", None)
for t in tools
}
tool_names = set()
for tool in tools:
if isinstance(tool, dict):
if tool.get("type") == "namespace":
namespace = tool.get("name")
for namespaced_tool in tool.get("tools", []):
namespaced_name = namespaced_tool.get("name")
tool_names.add(namespaced_name)
tool_names.add(f"{namespace}__{namespaced_name}")
else:
tool_names.add(tool.get("name"))
else:
tool_names.add(getattr(tool, "name", None))
if not tool_name or tool_name not in tool_names:
raise VLLMValidationError(
"Tool choice 'function' not found in 'tools' parameter.",
+2 -1
View File
@@ -1073,6 +1073,7 @@ class OpenAIServingResponses(GenerateBaseServing):
content=content,
tool_calls=tool_calls,
logprobs=logprobs,
tools=request.tools,
)
# Fallback when no parser is configured
@@ -1339,7 +1340,7 @@ class OpenAIServingResponses(GenerateBaseServing):
[StreamingResponsesResponse], StreamingResponsesResponse
],
) -> AsyncGenerator[StreamingResponsesResponse, None]:
processor = SimpleStreamingEventProcessor()
processor = SimpleStreamingEventProcessor(tools=request.tools)
def _get_logprobs(
output: CompletionOutput,
@@ -58,6 +58,7 @@ from openai.types.responses.response_output_item import McpCall
from openai.types.responses.response_reasoning_item import (
Content as ResponseReasoningTextContent,
)
from openai.types.responses.tool import Tool
from openai_harmony import Message as HarmonyMessage
from vllm.entrypoints.mcp.tool_server import ToolServer
@@ -71,6 +72,10 @@ from vllm.entrypoints.openai.responses.protocol import (
ResponseReasoningPartDoneEvent,
StreamingResponsesResponse,
)
from vllm.entrypoints.openai.responses.utils import (
build_responses_tool_call_name_map,
resolve_responses_tool_call_name,
)
from vllm.outputs import CompletionOutput
from vllm.parser.harmony import Segment
from vllm.utils import random_uuid
@@ -815,6 +820,7 @@ class SimpleStreamingState:
accumulated_text: str = ""
tool_call_id: str = ""
tool_call_name: str = ""
tool_call_namespace: str | None = None
tool_call_index: int | None = None
has_emitted_tool_call_delta: bool = False
current_state: _StateType = field(default_factory=lambda: _StateType.NONE)
@@ -1016,11 +1022,13 @@ def emit_simple_tool_call_open(
state: SimpleStreamingState,
name: str,
index: int | None,
namespace: str | None = None,
) -> list[StreamingResponsesResponse]:
state.current_state = _StateType.TOOL_CALL
state.current_item_id = random_uuid()
state.tool_call_id = f"call_{random_uuid()}"
state.tool_call_name = name
state.tool_call_namespace = namespace
state.tool_call_index = index
state.accumulated_text = ""
state.has_emitted_tool_call_delta = False
@@ -1034,6 +1042,7 @@ def emit_simple_tool_call_open(
id=state.current_item_id,
call_id=state.tool_call_id,
name=name,
namespace=namespace,
arguments="",
status="in_progress",
),
@@ -1081,6 +1090,7 @@ def emit_simple_tool_call_done(
item=ResponseFunctionToolCall(
type="function_call",
name=state.tool_call_name,
namespace=state.tool_call_namespace,
arguments=state.accumulated_text,
status="completed",
id=state.current_item_id,
@@ -1089,6 +1099,7 @@ def emit_simple_tool_call_done(
),
)
state.output_index += 1
state.tool_call_namespace = None
state.current_state = _StateType.NONE
return events
@@ -1166,8 +1177,13 @@ class SimpleStreamingEventProcessor:
),
}
def __init__(self, state: SimpleStreamingState | None = None) -> None:
def __init__(
self,
state: SimpleStreamingState | None = None,
tools: list[Tool] | None = None,
) -> None:
self.state = state or SimpleStreamingState()
self.tool_call_name_map = build_responses_tool_call_name_map(tools)
def resolve_target_state(
self, delta_message: DeltaMessage
@@ -1224,8 +1240,15 @@ class SimpleStreamingEventProcessor:
handlers = self._STATE_HANDLERS[target_state]
if target_state == _StateType.TOOL_CALL:
assert tool_call is not None
call_name = resolve_responses_tool_call_name(
tool_call.function.name,
tool_call_name_map=self.tool_call_name_map,
)
return handlers.open_fn(
self.state, tool_call.function.name, tool_call.index
self.state,
call_name.name,
tool_call.index,
call_name.namespace,
)
return handlers.open_fn(self.state)
+30 -5
View File
@@ -35,6 +35,12 @@ from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionMessa
from vllm.entrypoints.openai.engine.protocol import FunctionCall
from vllm.entrypoints.openai.responses.protocol import ResponseInputOutputItem
from vllm.logger import init_logger
from vllm.tool_parsers.utils import (
build_responses_tool_call_name_map,
flat_namespace_tool_name,
iter_response_function_tool_dicts,
resolve_responses_tool_call_name,
)
from vllm.utils import random_uuid
logger = init_logger(__name__)
@@ -45,8 +51,10 @@ def build_response_output_items(
content: str | None,
tool_calls: list[FunctionCall] | None,
logprobs: list[Logprob] | None = None,
tools: list[Tool] | None = None,
) -> list[ResponseOutputItem]:
outputs: list[ResponseOutputItem] = []
tool_call_name_map = build_responses_tool_call_name_map(tools)
if reasoning:
outputs.append(
@@ -81,6 +89,9 @@ def build_response_output_items(
if tool_calls:
for idx, tool_call in enumerate(tool_calls):
call_name = resolve_responses_tool_call_name(
tool_call.name, tool_call_name_map=tool_call_name_map
)
outputs.append(
ResponseFunctionToolCall(
id=f"fc_{random_uuid()}",
@@ -88,7 +99,8 @@ def build_response_output_items(
or make_tool_call_id(func_name=tool_call.name, idx=idx),
type="function_call",
status="completed",
name=tool_call.name,
name=call_name.name,
namespace=call_name.namespace,
arguments=tool_call.arguments,
)
)
@@ -219,10 +231,13 @@ def _construct_message_from_response_item(
)
if isinstance(item, ResponseFunctionToolCall):
tool_name = item.name
if item.namespace:
tool_name = flat_namespace_tool_name(item.namespace, item.name)
tool_call = ChatCompletionMessageToolCallParam(
id=item.call_id,
function=FunctionCallTool(
name=item.name,
name=tool_name,
arguments=item.arguments,
),
type="function",
@@ -318,7 +333,17 @@ def _construct_message_from_response_item(
def extract_function_tool_names(tools: list[Tool]) -> frozenset[str]:
return frozenset(tool.name for tool in tools if tool.type == "function")
names = []
for tool in tools:
if tool.type == "function":
names.append(tool.name)
elif tool.type == "namespace":
names.extend(
flat_namespace_tool_name(tool.name, namespaced_tool.name)
for namespaced_tool in tool.tools
if namespaced_tool.type == "function"
)
return frozenset(names)
def extract_tool_types(tools: list[Tool]) -> set[str]:
@@ -358,7 +383,7 @@ def construct_tool_dicts(
tool_dicts = None
else:
tool_dicts = [
convert_tool_responses_to_completions_format(tool.model_dump())
for tool in tools
convert_tool_responses_to_completions_format(tool)
for tool in iter_response_function_tool_dicts(tools)
]
return tool_dicts