[gpt-oss][2/N] Support input_messages in responsesRequest (#26962)
Signed-off-by: Andrew Xia <axia@fb.com> Co-authored-by: Andrew Xia <axia@fb.com>
This commit is contained in:
@@ -38,11 +38,14 @@ from openai_harmony import (
|
||||
ToolDescription,
|
||||
load_harmony_encoding,
|
||||
)
|
||||
from openai_harmony import Message as OpenAIHarmonyMessage
|
||||
from openai_harmony import Role as OpenAIHarmonyRole
|
||||
|
||||
from vllm import envs
|
||||
from vllm.entrypoints.openai.protocol import (
|
||||
ChatCompletionToolsParam,
|
||||
ResponseInputOutputItem,
|
||||
ResponsesRequest,
|
||||
)
|
||||
from vllm.utils import random_uuid
|
||||
|
||||
@@ -228,7 +231,7 @@ def parse_response_input(
|
||||
return msg
|
||||
|
||||
|
||||
def parse_chat_input(chat_msg) -> list[Message]:
|
||||
def parse_input_to_harmony_message(chat_msg) -> list[Message]:
|
||||
if not isinstance(chat_msg, dict):
|
||||
# Handle Pydantic models
|
||||
chat_msg = chat_msg.model_dump(exclude_none=True)
|
||||
@@ -279,6 +282,40 @@ def parse_chat_input(chat_msg) -> list[Message]:
|
||||
return [msg]
|
||||
|
||||
|
||||
def construct_harmony_previous_input_messages(
|
||||
request: ResponsesRequest,
|
||||
) -> list[OpenAIHarmonyMessage]:
|
||||
messages: list[OpenAIHarmonyMessage] = []
|
||||
if request.previous_input_messages:
|
||||
for message in request.previous_input_messages:
|
||||
# Handle both OpenAIHarmonyMessage objects and dictionary inputs
|
||||
if isinstance(message, OpenAIHarmonyMessage):
|
||||
message_role = message.author.role
|
||||
# To match OpenAI, instructions, reasoning and tools are
|
||||
# always taken from the most recent Responses API request
|
||||
# not carried over from previous requests
|
||||
if (
|
||||
message_role == OpenAIHarmonyRole.SYSTEM
|
||||
or message_role == OpenAIHarmonyRole.DEVELOPER
|
||||
):
|
||||
continue
|
||||
messages.append(message)
|
||||
else:
|
||||
harmony_messages = parse_input_to_harmony_message(message)
|
||||
for harmony_msg in harmony_messages:
|
||||
message_role = harmony_msg.author.role
|
||||
# To match OpenAI, instructions, reasoning and tools are
|
||||
# always taken from the most recent Responses API request
|
||||
# not carried over from previous requests
|
||||
if (
|
||||
message_role == OpenAIHarmonyRole.SYSTEM
|
||||
or message_role == OpenAIHarmonyRole.DEVELOPER
|
||||
):
|
||||
continue
|
||||
messages.append(harmony_msg)
|
||||
return messages
|
||||
|
||||
|
||||
def render_for_completion(messages: list[Message]) -> list[int]:
|
||||
conversation = Conversation.from_messages(messages)
|
||||
token_ids = get_encoding().render_conversation_for_completion(
|
||||
|
||||
@@ -47,6 +47,7 @@ from openai.types.responses import (
|
||||
from openai.types.responses.response_reasoning_item import (
|
||||
Content as ResponseReasoningTextContent,
|
||||
)
|
||||
from openai_harmony import Message as OpenAIHarmonyMessage
|
||||
|
||||
from vllm.utils.serial_utils import (
|
||||
EmbedDType,
|
||||
@@ -383,10 +384,15 @@ class ResponsesRequest(OpenAIBaseModel):
|
||||
default=False,
|
||||
description=(
|
||||
"Dictates whether or not to return messages as part of the "
|
||||
"response object. Currently only supported for non-streaming "
|
||||
"response object. Currently only supported for"
|
||||
"non-background and gpt-oss only. "
|
||||
),
|
||||
)
|
||||
# similar to input_messages / output_messages in ResponsesResponse
|
||||
# we take in previous_input_messages (ie in harmony format)
|
||||
# this cannot be used in conjunction with previous_response_id
|
||||
# TODO: consider supporting non harmony messages as well
|
||||
previous_input_messages: list[OpenAIHarmonyMessage | dict] | None = None
|
||||
# --8<-- [end:responses-extra-params]
|
||||
|
||||
_DEFAULT_SAMPLING_PARAMS = {
|
||||
|
||||
@@ -27,8 +27,8 @@ from vllm.entrypoints.harmony_utils import (
|
||||
get_stop_tokens_for_assistant_actions,
|
||||
get_streamable_parser_for_assistant,
|
||||
get_system_message,
|
||||
parse_chat_input,
|
||||
parse_chat_output,
|
||||
parse_input_to_harmony_message,
|
||||
render_for_completion,
|
||||
)
|
||||
from vllm.entrypoints.logger import RequestLogger
|
||||
@@ -1351,11 +1351,13 @@ class OpenAIServingChat(OpenAIServing):
|
||||
index=output.index,
|
||||
message=message,
|
||||
logprobs=logprobs,
|
||||
finish_reason="tool_calls"
|
||||
if (tool_call_info is not None and tool_call_info.tools_called)
|
||||
else output.finish_reason
|
||||
if output.finish_reason
|
||||
else "stop",
|
||||
finish_reason=(
|
||||
"tool_calls"
|
||||
if (tool_call_info is not None and tool_call_info.tools_called)
|
||||
else output.finish_reason
|
||||
if output.finish_reason
|
||||
else "stop"
|
||||
),
|
||||
stop_reason=output.stop_reason,
|
||||
)
|
||||
choices.append(choice_data)
|
||||
@@ -1522,11 +1524,13 @@ class OpenAIServingChat(OpenAIServing):
|
||||
index=output.index,
|
||||
message=message,
|
||||
logprobs=logprobs,
|
||||
finish_reason="tool_calls"
|
||||
if auto_tools_called
|
||||
else output.finish_reason
|
||||
if output.finish_reason
|
||||
else "stop",
|
||||
finish_reason=(
|
||||
"tool_calls"
|
||||
if auto_tools_called
|
||||
else output.finish_reason
|
||||
if output.finish_reason
|
||||
else "stop"
|
||||
),
|
||||
stop_reason=output.stop_reason,
|
||||
token_ids=(
|
||||
as_list(output.token_ids) if request.return_token_ids else None
|
||||
@@ -1685,9 +1689,11 @@ class OpenAIServingChat(OpenAIServing):
|
||||
should_return_as_token_id,
|
||||
),
|
||||
logprob=max(step_token.logprob, -9999.0),
|
||||
bytes=None
|
||||
if step_decoded is None
|
||||
else list(step_decoded.encode("utf-8", errors="replace")),
|
||||
bytes=(
|
||||
None
|
||||
if step_decoded is None
|
||||
else list(step_decoded.encode("utf-8", errors="replace"))
|
||||
),
|
||||
top_logprobs=self._get_top_logprobs(
|
||||
step_top_logprobs,
|
||||
num_output_top_logprobs,
|
||||
@@ -1764,7 +1770,7 @@ class OpenAIServingChat(OpenAIServing):
|
||||
|
||||
# Add user message.
|
||||
for chat_msg in request.messages:
|
||||
messages.extend(parse_chat_input(chat_msg))
|
||||
messages.extend(parse_input_to_harmony_message(chat_msg))
|
||||
|
||||
# Render prompt token ids.
|
||||
prompt_token_ids = render_for_completion(messages)
|
||||
|
||||
@@ -63,6 +63,7 @@ from vllm.entrypoints.context import (
|
||||
StreamingHarmonyContext,
|
||||
)
|
||||
from vllm.entrypoints.harmony_utils import (
|
||||
construct_harmony_previous_input_messages,
|
||||
get_developer_message,
|
||||
get_stop_tokens_for_assistant_actions,
|
||||
get_system_message,
|
||||
@@ -248,6 +249,13 @@ class OpenAIServingResponses(OpenAIServing):
|
||||
),
|
||||
status_code=HTTPStatus.BAD_REQUEST,
|
||||
)
|
||||
if request.previous_input_messages and request.previous_response_id:
|
||||
return self.create_error_response(
|
||||
err_type="invalid_request_error",
|
||||
message="Only one of `previous_input_messages` and "
|
||||
"`previous_response_id` can be set.",
|
||||
status_code=HTTPStatus.BAD_REQUEST,
|
||||
)
|
||||
return None
|
||||
|
||||
async def create_responses(
|
||||
@@ -941,6 +949,8 @@ class OpenAIServingResponses(OpenAIServing):
|
||||
instructions=request.instructions, tools=request.tools
|
||||
)
|
||||
messages.append(dev_msg)
|
||||
messages += construct_harmony_previous_input_messages(request)
|
||||
|
||||
else:
|
||||
# Continue the previous conversation.
|
||||
# FIXME(woosuk): Currently, request params like reasoning and
|
||||
|
||||
Reference in New Issue
Block a user