[Frontend] Add stateless /v1/responses/render endpoint (#50195)

Signed-off-by: Francisco Javier Arceo <farceo@redhat.com>
Co-authored-by: OpenAI Codex <codex@openai.com>
This commit is contained in:
Francisco Javier Arceo
2026-09-07 12:50:42 +00:00
committed by GitHub
co-authored by OpenAI Codex
parent 5e6f6a8ed4
commit 6a2a2bb02b
19 changed files with 1228 additions and 298 deletions
+267 -2
View File
@@ -1,13 +1,19 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import time
from collections.abc import Sequence
from dataclasses import dataclass
from http import HTTPStatus
from typing import Any
from typing import TYPE_CHECKING, Any, TypeAlias
from openai.types.responses import ResponseFunctionToolCall, ResponseOutputItem
from openai.types.responses.tool import Mcp, Tool
from openai_harmony import Message as OpenAIMessage
from openai_harmony import ToolNamespaceConfig
from vllm.config import ModelConfig
from vllm.entrypoints.chat_utils import (
ChatCompletionMessageParam,
ChatTemplateContentFormatOption,
ConversationMessage,
)
@@ -19,12 +25,24 @@ from vllm.entrypoints.openai.completion.protocol import (
CompletionRequest,
)
from vllm.entrypoints.openai.parser.harmony_utils import (
BUILTIN_TOOL_TO_MCP_SERVER_LABEL,
build_harmony_preamble,
extract_instructions_from_messages,
get_user_message,
has_custom_tools,
parse_chat_inputs_to_harmony_messages,
render_for_completion,
)
from vllm.entrypoints.openai.responses.harmony import (
construct_harmony_previous_input_messages,
response_input_to_harmony,
)
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.entrypoints.openai.responses.utils import (
construct_input_messages,
construct_tool_dicts,
extract_tool_types,
)
from vllm.entrypoints.serve import create_error_response
from vllm.entrypoints.serve.engine.protocol import ErrorResponse
from vllm.entrypoints.serve.utils.request_logger import RequestLogger
@@ -32,11 +50,12 @@ from vllm.inputs import (
EngineInput,
PromptType,
SingletonPrompt,
TokensPrompt,
tokens_input,
)
from vllm.logger import init_logger
from vllm.parser import Parser, ParserManager
from vllm.renderers import BaseRenderer, ChatParams, merge_kwargs
from vllm.renderers import BaseRenderer, ChatParams, TokenizeParams, merge_kwargs
from vllm.renderers.inputs.preprocess import (
parse_model_prompt,
prompt_to_seq,
@@ -46,6 +65,39 @@ from vllm.utils.mistral import mt as _mt
logger = init_logger(__name__)
if TYPE_CHECKING:
from vllm.entrypoints.mcp.tool_server import ToolServer
ResponsesPreviousMessages: TypeAlias = (
list[ChatCompletionMessageParam] | list[OpenAIMessage]
)
@dataclass(frozen=True)
class ResponsesRenderResult:
messages: ResponsesPreviousMessages
engine_input: EngineInput
def _extract_allowed_tools_from_mcp_requests(
tools: list[Tool],
) -> dict[str, list[str] | None]:
allowed_tools_map: dict[str, list[str] | None] = {}
for tool in tools:
if not isinstance(tool, Mcp):
continue
allowed_tools = None
if isinstance(tool.allowed_tools, list):
allowed_tools = tool.allowed_tools
elif tool.allowed_tools is not None:
allowed_tools = tool.allowed_tools.tool_names
if allowed_tools is not None and "*" in allowed_tools:
allowed_tools = None
allowed_tools_map[tool.server_label] = allowed_tools
return allowed_tools_map
def _reused_prompt_token_ids(request: Any) -> list[int] | None:
"""Pop prompt token ids forwarded for decode-side reuse, if any.
@@ -217,6 +269,219 @@ class OnlineRenderer:
return conversation, engine_inputs
async def render_responses(
self,
request: ResponsesRequest,
*,
previous_messages: ResponsesPreviousMessages | None = None,
previous_response_outputs: list[ResponseOutputItem] | None = None,
tool_server: "ToolServer | None" = None,
skip_mm_cache: bool = False,
) -> ResponsesRenderResult | ErrorResponse:
"""Render a Responses request using only explicitly supplied history."""
template_error = self.validate_chat_template(
request_chat_template=None,
chat_template_kwargs=request.chat_template_kwargs,
trust_request_chat_template=self.trust_request_chat_template,
)
if template_error is not None:
return template_error
if self.use_harmony:
return self._render_responses_with_harmony(
request,
previous_messages=previous_messages,
previous_response_outputs=previous_response_outputs,
tool_server=tool_server,
)
if previous_messages is not None and any(
isinstance(message, OpenAIMessage) for message in previous_messages
):
return self.create_error_response(
"Non-Harmony Responses history must use chat messages.",
err_type="invalid_request_error",
param="previous_response_id",
)
tool_dicts = construct_tool_dicts(
request.tools,
request.tool_choice,
exclude_tools_when_tool_choice_none=(
self.exclude_tools_when_tool_choice_none
),
)
messages = construct_input_messages(
request_instructions=request.instructions,
request_input=request.input,
prev_msg=list(previous_messages) if previous_messages is not None else None,
prev_response_output=(
list(previous_response_outputs)
if previous_response_outputs is not None
else None
),
)
chat_template_kwargs = (
request.build_chat_params(
self.chat_template,
self.chat_template_content_format,
)
.with_defaults(self.default_chat_template_kwargs)
.chat_template_kwargs
)
_, engine_inputs = await self.preprocess_chat(
request,
messages,
default_template=self.chat_template,
default_template_content_format=self.chat_template_content_format,
default_template_kwargs=chat_template_kwargs,
tool_dicts=tool_dicts,
parser=self.parser,
skip_mm_cache=skip_mm_cache,
)
return self._responses_render_result(messages, engine_inputs)
def _render_responses_with_harmony(
self,
request: ResponsesRequest,
*,
previous_messages: ResponsesPreviousMessages | None,
previous_response_outputs: list[ResponseOutputItem] | None,
tool_server: "ToolServer | None",
) -> ResponsesRenderResult | ErrorResponse:
if self.parser is not None:
# HarmonyParser doesn't need chat_template_kwargs
# TODO: Unify adjust_request() call with non-harmony branch
self.parser(
self.renderer.get_tokenizer(),
request.tools,
model_config=self.model_config,
).adjust_request(request=request)
if previous_messages is not None and any(
not isinstance(message, OpenAIMessage) for message in previous_messages
):
return self.create_error_response(
"Harmony Responses history must use Harmony messages.",
err_type="invalid_request_error",
param="previous_response_id",
)
messages: list[OpenAIMessage] = []
request_input = request.input
if previous_messages is None:
tool_types = extract_tool_types(request.tools)
with_custom_tools = has_custom_tools(tool_types)
instructions = request.instructions
if instructions is None and isinstance(request_input, list):
instructions, request_input = extract_instructions_from_messages(
request_input
)
descriptions = self._get_harmony_builtin_tool_descriptions(
request.tools,
tool_types,
tool_server,
)
messages.extend(
build_harmony_preamble(
instructions=instructions,
tools=request.tools if with_custom_tools else None,
reasoning_effort=(
request.reasoning.effort if request.reasoning else None
),
with_custom_tools=with_custom_tools,
**descriptions,
)
)
messages.extend(construct_harmony_previous_input_messages(request))
else:
messages.extend(previous_messages)
previous_outputs = list(previous_response_outputs or ())
try:
if isinstance(request_input, str):
if request_input or not request.previous_input_messages:
messages.append(get_user_message(request_input))
else:
for response_message in request_input:
new_message = response_input_to_harmony(
response_message,
previous_outputs,
)
if new_message is not None:
messages.append(new_message)
if isinstance(response_message, ResponseFunctionToolCall):
previous_outputs.append(response_message)
except ValueError as exc:
return self.create_error_response(
str(exc),
err_type="invalid_request_error",
param="input",
)
return ResponsesRenderResult(
messages=messages,
engine_input=self.render_responses_harmony_messages(
messages,
cache_salt=request.cache_salt,
tok_params=request.build_tok_params(self.model_config),
),
)
def _get_harmony_builtin_tool_descriptions(
self,
tools: list[Tool],
tool_types: set[str],
tool_server: "ToolServer | None",
) -> dict[str, ToolNamespaceConfig | None]:
allowed_tools = _extract_allowed_tools_from_mcp_requests(tools)
descriptions: dict[str, ToolNamespaceConfig | None] = {}
for server_name, request_name in BUILTIN_TOOL_TO_MCP_SERVER_LABEL.items():
description = (
tool_server.get_tool_description(
server_name,
allowed_tools.get(request_name),
)
if request_name in tool_types
and tool_server is not None
and tool_server.has_tool(server_name)
else None
)
descriptions[f"{server_name}_description"] = description
return descriptions
def render_responses_harmony_messages(
self,
messages: list[OpenAIMessage],
*,
cache_salt: str | None,
tok_params: TokenizeParams | None = None,
) -> EngineInput:
arrival_time = time.time()
prompt = TokensPrompt(prompt_token_ids=render_for_completion(messages))
if tok_params is not None:
tok_params.apply_post_tokenization(
self.renderer.tokenizer,
prompt,
)
engine_input = tokens_input(prompt["prompt_token_ids"], cache_salt=cache_salt)
engine_input["arrival_time"] = arrival_time
return engine_input
def _responses_render_result(
self,
messages: list[ChatCompletionMessageParam],
engine_inputs: list[EngineInput],
) -> ResponsesRenderResult | ErrorResponse:
if len(engine_inputs) != 1:
return self.create_error_response(
f"Expected exactly 1 engine prompt, got {len(engine_inputs)}"
)
return ResponsesRenderResult(
messages=messages,
engine_input=engine_inputs[0],
)
def _make_request_with_harmony(
self,
request: ChatCompletionRequest,