[2/N] move responses/serving _make_response_output_items logic to parser (#33281)
Signed-off-by: Andrew Xia <axia@fb.com> Signed-off-by: Andrew Xia <axia@meta.com> Co-authored-by: Andrew Xia <axia@fb.com>
This commit is contained in:
Andrew Xia
@@ -63,7 +63,6 @@ from vllm.engine.protocol import EngineClient
|
||||
from vllm.entrypoints.chat_utils import (
|
||||
ChatCompletionMessageParam,
|
||||
ChatTemplateContentFormatOption,
|
||||
make_tool_call_id,
|
||||
)
|
||||
from vllm.entrypoints.logger import RequestLogger
|
||||
from vllm.entrypoints.mcp.tool_server import ToolServer
|
||||
@@ -915,114 +914,57 @@ class OpenAIServingResponses(OpenAIServing):
|
||||
final_output: CompletionOutput,
|
||||
tokenizer: TokenizerLike,
|
||||
) -> list[ResponseOutputItem]:
|
||||
if self.parser and self.parser.reasoning_parser_cls:
|
||||
try:
|
||||
reasoning_parser = self.parser.reasoning_parser_cls(tokenizer)
|
||||
except RuntimeError as e:
|
||||
logger.exception("Error in reasoning parser creation.")
|
||||
raise e
|
||||
|
||||
reasoning, content = reasoning_parser.extract_reasoning(
|
||||
final_output.text, request=request
|
||||
)
|
||||
else:
|
||||
reasoning = None
|
||||
content = final_output.text
|
||||
|
||||
# Log complete response if output logging is enabled
|
||||
if self.enable_log_outputs and self.request_logger:
|
||||
output_text = ""
|
||||
if content:
|
||||
output_text = content
|
||||
elif reasoning:
|
||||
output_text = f"[reasoning: {reasoning}]"
|
||||
|
||||
if output_text:
|
||||
self.request_logger.log_outputs(
|
||||
request_id=request.request_id,
|
||||
outputs=output_text,
|
||||
output_token_ids=final_output.token_ids,
|
||||
finish_reason=final_output.finish_reason,
|
||||
is_streaming=False,
|
||||
delta=False,
|
||||
)
|
||||
|
||||
reasoning_item = None
|
||||
message_item = None
|
||||
if reasoning:
|
||||
reasoning_item = ResponseReasoningItem(
|
||||
id=f"rs_{random_uuid()}",
|
||||
summary=[],
|
||||
type="reasoning",
|
||||
content=[
|
||||
ResponseReasoningTextContent(text=reasoning, type="reasoning_text")
|
||||
],
|
||||
status=None, # NOTE: Only the last output item has status.
|
||||
self.request_logger.log_outputs(
|
||||
request_id=request.request_id,
|
||||
outputs=final_output.text,
|
||||
output_token_ids=final_output.token_ids,
|
||||
finish_reason=final_output.finish_reason,
|
||||
is_streaming=False,
|
||||
delta=False,
|
||||
)
|
||||
tool_calls, content = self._parse_tool_calls_from_content(
|
||||
request=request,
|
||||
tokenizer=tokenizer,
|
||||
content=content,
|
||||
enable_auto_tools=self.enable_auto_tools,
|
||||
tool_parser_cls=self.parser.tool_parser_cls if self.parser else None,
|
||||
)
|
||||
|
||||
if content or (self.use_harmony and tool_calls):
|
||||
res_text_part = None
|
||||
if content:
|
||||
res_text_part = ResponseOutputText(
|
||||
text=content,
|
||||
annotations=[], # TODO
|
||||
type="output_text",
|
||||
logprobs=(
|
||||
self._create_response_logprobs(
|
||||
token_ids=final_output.token_ids,
|
||||
logprobs=final_output.logprobs,
|
||||
tokenizer=tokenizer,
|
||||
top_logprobs=request.top_logprobs,
|
||||
)
|
||||
if request.is_include_output_logprobs()
|
||||
else None
|
||||
),
|
||||
)
|
||||
message_item = ResponseOutputMessage(
|
||||
# Compute logprobs if requested
|
||||
logprobs = None
|
||||
if request.is_include_output_logprobs() and final_output.logprobs:
|
||||
logprobs = self._create_response_logprobs(
|
||||
token_ids=final_output.token_ids,
|
||||
logprobs=final_output.logprobs,
|
||||
tokenizer=tokenizer,
|
||||
top_logprobs=request.top_logprobs,
|
||||
)
|
||||
|
||||
# Use parser to extract and create response output items
|
||||
if self.parser:
|
||||
parser = self.parser(tokenizer)
|
||||
return parser.extract_response_outputs(
|
||||
model_output=final_output.text,
|
||||
request=request,
|
||||
enable_auto_tools=self.enable_auto_tools,
|
||||
tool_call_id_type=self.tool_call_id_type,
|
||||
logprobs=logprobs,
|
||||
)
|
||||
|
||||
# Fallback when no parser is configured
|
||||
return [
|
||||
ResponseOutputMessage(
|
||||
id=f"msg_{random_uuid()}",
|
||||
content=[res_text_part] if res_text_part else [],
|
||||
content=[
|
||||
ResponseOutputText(
|
||||
text=final_output.text,
|
||||
annotations=[],
|
||||
type="output_text",
|
||||
logprobs=logprobs,
|
||||
)
|
||||
]
|
||||
if final_output.text
|
||||
else [],
|
||||
role="assistant",
|
||||
status="completed",
|
||||
type="message",
|
||||
)
|
||||
outputs = []
|
||||
|
||||
if reasoning_item:
|
||||
outputs.append(reasoning_item)
|
||||
if message_item:
|
||||
outputs.append(message_item)
|
||||
if tool_calls:
|
||||
# We use a simple counter for history_tool_call_count because
|
||||
# we don't track the history of tool calls in the Responses API yet.
|
||||
# This means that the tool call index will start from 0 for each
|
||||
# request.
|
||||
tool_call_items = []
|
||||
for history_tool_call_cnt, tool_call in enumerate(tool_calls):
|
||||
tool_call_items.append(
|
||||
ResponseFunctionToolCall(
|
||||
id=f"fc_{random_uuid()}",
|
||||
call_id=tool_call.id
|
||||
if tool_call.id
|
||||
else make_tool_call_id(
|
||||
id_type=self.tool_call_id_type,
|
||||
func_name=tool_call.name,
|
||||
idx=history_tool_call_cnt,
|
||||
),
|
||||
type="function_call",
|
||||
status="completed",
|
||||
name=tool_call.name,
|
||||
arguments=tool_call.arguments,
|
||||
)
|
||||
)
|
||||
outputs.extend(tool_call_items)
|
||||
return outputs
|
||||
]
|
||||
|
||||
def _make_response_output_items_with_harmony(
|
||||
self,
|
||||
|
||||
@@ -1,23 +1,46 @@
|
||||
# SPDX-License-Identifier: Apache-2.0
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
|
||||
import json
|
||||
from abc import abstractmethod
|
||||
from collections.abc import Sequence
|
||||
from functools import cached_property
|
||||
|
||||
from openai.types.responses import (
|
||||
ResponseFunctionToolCall,
|
||||
ResponseOutputItem,
|
||||
ResponseOutputMessage,
|
||||
ResponseOutputText,
|
||||
ResponseReasoningItem,
|
||||
ToolChoiceFunction,
|
||||
)
|
||||
from openai.types.responses.response_output_text import Logprob
|
||||
from openai.types.responses.response_reasoning_item import (
|
||||
Content as ResponseReasoningTextContent,
|
||||
)
|
||||
from pydantic import TypeAdapter
|
||||
|
||||
from vllm.entrypoints.chat_utils import make_tool_call_id
|
||||
from vllm.entrypoints.openai.chat_completion.protocol import (
|
||||
ChatCompletionNamedToolChoiceParam,
|
||||
ChatCompletionRequest,
|
||||
)
|
||||
from vllm.entrypoints.openai.engine.protocol import (
|
||||
DeltaMessage,
|
||||
ExtractedToolCallInformation,
|
||||
FunctionCall,
|
||||
FunctionDefinition,
|
||||
)
|
||||
from vllm.entrypoints.openai.responses.protocol import (
|
||||
ResponsesRequest,
|
||||
)
|
||||
from vllm.logger import init_logger
|
||||
from vllm.reasoning.abs_reasoning_parsers import ReasoningParser
|
||||
from vllm.tokenizers import TokenizerLike
|
||||
from vllm.tool_parsers.abstract_tool_parser import ToolParser
|
||||
from vllm.utils import random_uuid
|
||||
|
||||
logger = init_logger(__name__)
|
||||
|
||||
|
||||
class Parser:
|
||||
@@ -128,6 +151,33 @@ class Parser:
|
||||
The extracted content token IDs.
|
||||
"""
|
||||
|
||||
@abstractmethod
|
||||
def extract_response_outputs(
|
||||
self,
|
||||
model_output: str,
|
||||
request: ResponsesRequest,
|
||||
enable_auto_tools: bool = False,
|
||||
tool_call_id_type: str = "random",
|
||||
logprobs: list[Logprob] | None = None,
|
||||
) -> list[ResponseOutputItem]:
|
||||
"""
|
||||
Extract reasoning, content, and tool calls from a complete
|
||||
model-generated string and return as ResponseOutputItem objects.
|
||||
|
||||
Used for non-streaming responses where we have the entire model
|
||||
response available before sending to the client.
|
||||
|
||||
Args:
|
||||
model_output: The complete model-generated string.
|
||||
request: The request object used to generate the output.
|
||||
enable_auto_tools: Whether to enable automatic tool call parsing.
|
||||
tool_call_id_type: Type of tool call ID generation ("random", etc).
|
||||
logprobs: Pre-computed logprobs for the output text, if any.
|
||||
|
||||
Returns:
|
||||
A list of ResponseOutputItem objects.
|
||||
"""
|
||||
|
||||
@abstractmethod
|
||||
def extract_reasoning(
|
||||
self,
|
||||
@@ -260,6 +310,156 @@ class DelegatingParser(Parser):
|
||||
return None, model_output
|
||||
return self._reasoning_parser.extract_reasoning(model_output, request)
|
||||
|
||||
def extract_response_outputs(
|
||||
self,
|
||||
model_output: str,
|
||||
request: ResponsesRequest,
|
||||
enable_auto_tools: bool = False,
|
||||
tool_call_id_type: str = "random",
|
||||
logprobs: list[Logprob] | None = None,
|
||||
) -> list[ResponseOutputItem]:
|
||||
# First extract reasoning
|
||||
reasoning, content = self.extract_reasoning(model_output, request)
|
||||
|
||||
# Then parse tool calls from the content
|
||||
tool_calls, content = self._parse_tool_calls(
|
||||
request=request,
|
||||
content=content,
|
||||
enable_auto_tools=enable_auto_tools,
|
||||
)
|
||||
|
||||
# Build output items
|
||||
outputs: list[ResponseOutputItem] = []
|
||||
|
||||
# Add reasoning item if present
|
||||
if reasoning:
|
||||
reasoning_item = ResponseReasoningItem(
|
||||
id=f"rs_{random_uuid()}",
|
||||
summary=[],
|
||||
type="reasoning",
|
||||
content=[
|
||||
ResponseReasoningTextContent(text=reasoning, type="reasoning_text")
|
||||
],
|
||||
status=None, # NOTE: Only the last output item has status.
|
||||
)
|
||||
outputs.append(reasoning_item)
|
||||
|
||||
# Add message item if there's content
|
||||
if content:
|
||||
res_text_part = ResponseOutputText(
|
||||
text=content,
|
||||
annotations=[],
|
||||
type="output_text",
|
||||
logprobs=logprobs,
|
||||
)
|
||||
message_item = ResponseOutputMessage(
|
||||
id=f"msg_{random_uuid()}",
|
||||
content=[res_text_part],
|
||||
role="assistant",
|
||||
status="completed",
|
||||
type="message",
|
||||
)
|
||||
outputs.append(message_item)
|
||||
|
||||
if tool_calls:
|
||||
# We use a simple counter for history_tool_call_count because
|
||||
# we don't track the history of tool calls in the Responses API yet.
|
||||
# This means that the tool call index will start from 0 for each
|
||||
# request.
|
||||
for history_tool_call_cnt, tool_call in enumerate(tool_calls):
|
||||
tool_call_item = 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=history_tool_call_cnt,
|
||||
),
|
||||
type="function_call",
|
||||
status="completed",
|
||||
name=tool_call.name,
|
||||
arguments=tool_call.arguments,
|
||||
)
|
||||
outputs.append(tool_call_item)
|
||||
|
||||
return outputs
|
||||
|
||||
def _parse_tool_calls(
|
||||
self,
|
||||
request: ResponsesRequest,
|
||||
content: str | None,
|
||||
enable_auto_tools: bool,
|
||||
) -> tuple[list[FunctionCall], str | None]:
|
||||
"""
|
||||
TODO(qandrew): merge _parse_tool_calls_from_content
|
||||
for ChatCompletions into this function
|
||||
Parse tool calls from content based on request tool_choice settings.
|
||||
|
||||
Returns:
|
||||
A tuple of (function_calls, remaining_content) if tool calls
|
||||
were parsed
|
||||
"""
|
||||
function_calls: list[FunctionCall] = []
|
||||
|
||||
if request.tool_choice and isinstance(request.tool_choice, ToolChoiceFunction):
|
||||
# Forced Function Call (Responses API style)
|
||||
assert content is not None
|
||||
function_calls.append(
|
||||
FunctionCall(name=request.tool_choice.name, arguments=content)
|
||||
)
|
||||
return function_calls, None # Clear content since tool is called.
|
||||
|
||||
if request.tool_choice and isinstance(
|
||||
request.tool_choice, ChatCompletionNamedToolChoiceParam
|
||||
):
|
||||
# Forced Function Call (Chat Completion API style)
|
||||
assert content is not None
|
||||
function_calls.append(
|
||||
FunctionCall(name=request.tool_choice.function.name, arguments=content)
|
||||
)
|
||||
return function_calls, None # Clear content since tool is called.
|
||||
|
||||
if request.tool_choice == "required":
|
||||
# Required tool calls - parse JSON
|
||||
assert content is not None
|
||||
tool_calls = TypeAdapter(list[FunctionDefinition]).validate_json(content)
|
||||
function_calls.extend(
|
||||
FunctionCall(
|
||||
name=tool_call.name,
|
||||
arguments=json.dumps(tool_call.parameters, ensure_ascii=False),
|
||||
)
|
||||
for tool_call in tool_calls
|
||||
)
|
||||
return function_calls, None # Clear content since tool is called.
|
||||
|
||||
if (
|
||||
self._tool_parser is not None
|
||||
and enable_auto_tools
|
||||
and (request.tool_choice == "auto" or request.tool_choice is None)
|
||||
):
|
||||
# Automatic Tool Call Parsing
|
||||
tool_call_info = self._tool_parser.extract_tool_calls(
|
||||
content if content is not None else "",
|
||||
request=request, # type: ignore
|
||||
)
|
||||
if tool_call_info is not None and tool_call_info.tools_called:
|
||||
function_calls.extend(
|
||||
FunctionCall(
|
||||
id=tool_call.id,
|
||||
name=tool_call.function.name,
|
||||
arguments=tool_call.function.arguments,
|
||||
)
|
||||
for tool_call in tool_call_info.tool_calls
|
||||
)
|
||||
remaining_content = tool_call_info.content
|
||||
if remaining_content and remaining_content.strip() == "":
|
||||
remaining_content = None
|
||||
return function_calls, remaining_content
|
||||
|
||||
# No tool calls
|
||||
return [], content
|
||||
|
||||
def extract_reasoning_streaming(
|
||||
self,
|
||||
previous_text: str,
|
||||
|
||||
Reference in New Issue
Block a user