[Parser] Migrate ResponsesParser to unified Parser interface (#42977)

Signed-off-by: Alberto Perdomo <aperdomo@redhat.com>
This commit is contained in:
alberto
2026-06-02 08:50:05 +00:00
committed by GitHub
parent d247a9dc13
commit e30313220c
4 changed files with 442 additions and 94 deletions
@@ -10,10 +10,6 @@ from openai.types.responses.response_function_tool_call_output_item import (
from openai.types.responses.response_output_item import McpCall
from openai.types.responses.response_output_message import ResponseOutputMessage
from openai.types.responses.response_output_text import ResponseOutputText
from openai.types.responses.response_reasoning_item import (
Content,
ResponseReasoningItem,
)
from vllm.entrypoints.chat_utils import ChatTemplateContentFormatOption
from vllm.entrypoints.constants import MCP_PREFIX
@@ -22,9 +18,8 @@ from vllm.entrypoints.openai.responses.protocol import (
ResponsesRequest,
)
from vllm.outputs import CompletionOutput
from vllm.reasoning.abs_reasoning_parsers import ReasoningParser
from vllm.parser.abstract_parser import Parser
from vllm.tokenizers import TokenizerLike
from vllm.tool_parsers.abstract_tool_parser import ToolParser
from vllm.utils import random_uuid
logger = logging.getLogger(__name__)
@@ -37,12 +32,13 @@ class ResponsesParser:
self,
*,
tokenizer: TokenizerLike,
reasoning_parser_cls: type[ReasoningParser],
parser_cls: type[Parser] | None,
response_messages: list[ResponseInputOutputItem],
request: ResponsesRequest,
tool_parser_cls: type[ToolParser] | None,
chat_template: str | None,
chat_template_content_format: ChatTemplateContentFormatOption,
enable_auto_tools: bool = False,
tool_call_id_type: str = "random",
):
self.response_messages: list[ResponseInputOutputItem] = (
# TODO: initial messages may not be properly typed
@@ -52,17 +48,22 @@ class ResponsesParser:
self.tokenizer = tokenizer
self.request = request
self.reasoning_parser_instance = reasoning_parser_cls(
tokenizer,
chat_template_kwargs=_effective_chat_template_kwargs(
self.parser_instance: Parser | None = None
if parser_cls is not None:
chat_template_kwargs = _effective_chat_template_kwargs(
request,
chat_template=chat_template,
chat_template_content_format=chat_template_content_format,
),
)
self.tool_parser_instance = None
if tool_parser_cls is not None:
self.tool_parser_instance = tool_parser_cls(tokenizer, request.tools)
)
self.parser_instance = parser_cls(
tokenizer,
tools=request.tools,
chat_template_kwargs=chat_template_kwargs,
)
self.enable_auto_tools = enable_auto_tools
self.tool_call_id_type = tool_call_id_type
# Store the last finish_reason to determine response status
self.finish_reason: str | None = None
@@ -71,66 +72,34 @@ class ResponsesParser:
# Store the finish_reason from the output
self.finish_reason = output.finish_reason
reasoning, content = self.reasoning_parser_instance.extract_reasoning(
output.text, request=self.request
)
if reasoning:
self.response_messages.append(
ResponseReasoningItem(
type="reasoning",
id=f"rs_{random_uuid()}",
summary=[],
content=[
Content(
type="reasoning_text",
text=reasoning,
)
],
)
if self.parser_instance is not None:
output_items = self.parser_instance.extract_response_outputs(
model_output=output.text,
model_output_token_ids=output.token_ids,
request=self.request,
enable_auto_tools=self.enable_auto_tools,
tool_call_id_type=self.tool_call_id_type,
)
function_calls: list[ResponseFunctionToolCall] = []
if self.tool_parser_instance is not None:
tool_call_info = self.tool_parser_instance.extract_tool_calls(
content if content is not None else "",
request=self.request, # type: ignore
)
if tool_call_info is not None and tool_call_info.tools_called:
# extract_tool_calls() returns a list of tool calls.
function_calls.extend(
ResponseFunctionToolCall(
id=f"fc_{random_uuid()}",
call_id=f"call_{random_uuid()}",
type="function_call",
self.response_messages.extend(output_items)
else:
# No parser configured, treat entire output as text content
if output.text:
self.response_messages.append(
ResponseOutputMessage(
type="message",
id=f"msg_{random_uuid()}",
status="completed",
name=tool_call.function.name,
arguments=tool_call.function.arguments,
role="assistant",
content=[
ResponseOutputText(
annotations=[], # TODO
type="output_text",
text=output.text,
logprobs=None, # TODO
)
],
)
for tool_call in tool_call_info.tool_calls
)
content = tool_call_info.content
if content and content.strip() == "":
content = None
if content:
self.response_messages.append(
ResponseOutputMessage(
type="message",
id=f"msg_{random_uuid()}",
status="completed",
role="assistant",
content=[
ResponseOutputText(
annotations=[], # TODO
type="output_text",
text=content,
logprobs=None, # TODO
)
],
)
)
if len(function_calls) > 0:
self.response_messages.extend(function_calls)
return self
@@ -169,27 +138,29 @@ class ResponsesParser:
def get_responses_parser_for_simple_context(
*,
tokenizer: TokenizerLike,
reasoning_parser_cls: type[ReasoningParser],
parser_cls: type[Parser] | None,
response_messages: list[ResponseInputOutputItem],
request: ResponsesRequest,
tool_parser_cls,
chat_template: str | None,
chat_template_content_format: ChatTemplateContentFormatOption,
enable_auto_tools: bool = False,
tool_call_id_type: str = "random",
) -> ResponsesParser:
"""Factory function to create a ResponsesParser with
optional reasoning parser.
optional unified parser.
Returns:
ResponsesParser instance configured with the provided parser
"""
return ResponsesParser(
tokenizer=tokenizer,
reasoning_parser_cls=reasoning_parser_cls,
parser_cls=parser_cls,
response_messages=response_messages,
request=request,
tool_parser_cls=tool_parser_cls,
chat_template=chat_template,
chat_template_content_format=chat_template_content_format,
enable_auto_tools=enable_auto_tools,
tool_call_id_type=tool_call_id_type,
)