[Frontend] Add /v1/messages/render endpoint for the Anthropic Messages API (#45803)
Signed-off-by: HyunKyun Moon <mhg5303@gmail.com>
This commit is contained in:
@@ -3,7 +3,7 @@
|
||||
"""Unit tests for Anthropic-to-OpenAI request conversion.
|
||||
|
||||
Tests the image source handling and tool_result content parsing in
|
||||
AnthropicServingMessages._convert_anthropic_to_openai_request().
|
||||
AnthropicServingMessages.to_chat_completion_request().
|
||||
|
||||
Also covers extended-thinking edge cases such as ``redacted_thinking``
|
||||
blocks echoed back by Anthropic clients, and streaming conversion in
|
||||
@@ -51,7 +51,7 @@ from vllm.entrypoints.serve.exception_handling.handlers.validation import (
|
||||
)
|
||||
from vllm.exceptions import VLLMValidationError
|
||||
|
||||
_convert = AnthropicServingMessages._convert_anthropic_to_openai_request
|
||||
_convert = AnthropicServingMessages.to_chat_completion_request
|
||||
_img_url = AnthropicServingMessages._convert_image_source_to_url
|
||||
|
||||
|
||||
|
||||
@@ -493,3 +493,115 @@ async def test_chat_completion_render_assistant_tokens_mask_with_generation_tags
|
||||
assert detok_rest.status_code == 200
|
||||
assert "Hi!" not in detok_rest.json()["prompt"]
|
||||
assert "Bye" in detok_rest.json()["prompt"]
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_messages_render_basic(client):
|
||||
"""Test basic Anthropic Messages render endpoint."""
|
||||
response = await client.post(
|
||||
"/v1/messages/render",
|
||||
json={
|
||||
"model": MODEL_NAME,
|
||||
"max_tokens": 16,
|
||||
"messages": [{"role": "user", "content": "Render this Anthropic message."}],
|
||||
},
|
||||
)
|
||||
|
||||
assert response.status_code == 200
|
||||
data = response.json()
|
||||
|
||||
# Single GenerateRequest, like chat render.
|
||||
assert isinstance(data, dict)
|
||||
assert "token_ids" in data
|
||||
assert "sampling_params" in data
|
||||
assert "model" in data
|
||||
assert data["model"] == MODEL_NAME
|
||||
|
||||
token_ids = data["token_ids"]
|
||||
assert isinstance(token_ids, list)
|
||||
assert len(token_ids) > 0
|
||||
assert all(isinstance(tid, int) for tid in token_ids)
|
||||
assert token_ids[0] == 1 # BOS
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_messages_render_system_and_multi_turn(client):
|
||||
"""System field + multi-turn messages render to a single prompt."""
|
||||
response = await client.post(
|
||||
"/v1/messages/render",
|
||||
json={
|
||||
"model": MODEL_NAME,
|
||||
"max_tokens": 16,
|
||||
"system": "You are a helpful assistant.",
|
||||
"messages": [
|
||||
{"role": "user", "content": "Hello"},
|
||||
{"role": "assistant", "content": "Hi! How can I help?"},
|
||||
{"role": "user", "content": "What is 2 + 2?"},
|
||||
],
|
||||
},
|
||||
)
|
||||
|
||||
assert response.status_code == 200
|
||||
data = response.json()
|
||||
assert isinstance(data, dict)
|
||||
assert len(data["token_ids"]) > 0
|
||||
assert data["token_ids"][0] == 1 # BOS
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_messages_render_merges_inline_system(client):
|
||||
"""Inline system messages merge into the leading system block.
|
||||
|
||||
Without a --chat-template arg the /v1/messages server path detects
|
||||
merge_inline_system=True, so render must produce the same tokens as
|
||||
the manually pre-merged request.
|
||||
"""
|
||||
inline = await client.post(
|
||||
"/v1/messages/render",
|
||||
json={
|
||||
"model": MODEL_NAME,
|
||||
"max_tokens": 16,
|
||||
"system": "You are a helpful assistant.",
|
||||
"messages": [
|
||||
{"role": "user", "content": "Hello"},
|
||||
{"role": "assistant", "content": "Hi! How can I help?"},
|
||||
{"role": "system", "content": "Be brief."},
|
||||
{"role": "user", "content": "What is 2 + 2?"},
|
||||
],
|
||||
},
|
||||
)
|
||||
assert inline.status_code == 200
|
||||
|
||||
merged = await client.post(
|
||||
"/v1/messages/render",
|
||||
json={
|
||||
"model": MODEL_NAME,
|
||||
"max_tokens": 16,
|
||||
"system": "You are a helpful assistant.Be brief.",
|
||||
"messages": [
|
||||
{"role": "user", "content": "Hello"},
|
||||
{"role": "assistant", "content": "Hi! How can I help?"},
|
||||
{"role": "user", "content": "What is 2 + 2?"},
|
||||
],
|
||||
},
|
||||
)
|
||||
assert merged.status_code == 200
|
||||
|
||||
assert inline.json()["token_ids"] == merged.json()["token_ids"]
|
||||
|
||||
|
||||
@pytest.mark.asyncio
|
||||
async def test_messages_render_error_invalid_model(client):
|
||||
"""Messages render with an invalid model returns an error."""
|
||||
response = await client.post(
|
||||
"/v1/messages/render",
|
||||
json={
|
||||
"model": "invalid-model-name",
|
||||
"max_tokens": 16,
|
||||
"messages": [{"role": "user", "content": "Hello"}],
|
||||
},
|
||||
)
|
||||
|
||||
assert response.status_code == 404
|
||||
data = response.json()
|
||||
assert "error" in data
|
||||
|
||||
@@ -191,7 +191,7 @@ class AnthropicServingMessages(OpenAIServingChat):
|
||||
return f"data:{media_type};base64,{data}"
|
||||
|
||||
@classmethod
|
||||
def _convert_anthropic_to_openai_request(
|
||||
def to_chat_completion_request(
|
||||
cls,
|
||||
anthropic_request: AnthropicMessagesRequest | AnthropicCountTokensRequest,
|
||||
*,
|
||||
@@ -602,7 +602,7 @@ class AnthropicServingMessages(OpenAIServingChat):
|
||||
"""
|
||||
if logger.isEnabledFor(logging.DEBUG):
|
||||
logger.debug("Received messages request %s", request.model_dump_json())
|
||||
chat_req = self._convert_anthropic_to_openai_request(
|
||||
chat_req = self.to_chat_completion_request(
|
||||
request,
|
||||
merge_inline_system=self._merge_inline_system,
|
||||
)
|
||||
@@ -1023,7 +1023,7 @@ class AnthropicServingMessages(OpenAIServingChat):
|
||||
raw_request: Request | None = None,
|
||||
) -> AnthropicCountTokensResponse | ErrorResponse:
|
||||
"""Implements Anthropic's messages.count_tokens endpoint."""
|
||||
chat_req = self._convert_anthropic_to_openai_request(
|
||||
chat_req = self.to_chat_completion_request(
|
||||
request,
|
||||
merge_inline_system=self._merge_inline_system,
|
||||
)
|
||||
|
||||
@@ -5,6 +5,7 @@ from http import HTTPStatus
|
||||
from fastapi import APIRouter, Depends, Request
|
||||
from fastapi.responses import JSONResponse
|
||||
|
||||
from vllm.entrypoints.anthropic.protocol import AnthropicMessagesRequest
|
||||
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
|
||||
from vllm.entrypoints.openai.completion.protocol import CompletionRequest
|
||||
from vllm.entrypoints.openai.engine.protocol import ErrorResponse
|
||||
@@ -49,6 +50,30 @@ async def render_chat_completion(request: ChatCompletionRequest, raw_request: Re
|
||||
return JSONResponse(content=result.model_dump())
|
||||
|
||||
|
||||
@router.post(
|
||||
"/v1/messages/render",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
response_model=GenerateRequest,
|
||||
responses={
|
||||
HTTPStatus.BAD_REQUEST.value: {"model": ErrorResponse},
|
||||
HTTPStatus.NOT_FOUND.value: {"model": ErrorResponse},
|
||||
HTTPStatus.NOT_IMPLEMENTED.value: {"model": ErrorResponse},
|
||||
HTTPStatus.INTERNAL_SERVER_ERROR.value: {"model": ErrorResponse},
|
||||
},
|
||||
)
|
||||
async def render_messages(request: AnthropicMessagesRequest, raw_request: Request):
|
||||
handler = render(raw_request)
|
||||
if handler is None:
|
||||
raise NotImplementedError("The model does not support Messages Render API")
|
||||
|
||||
result = await handler.render_messages_request(request)
|
||||
|
||||
if isinstance(result, ErrorResponse):
|
||||
return JSONResponse(content=result.model_dump(), status_code=result.error.code)
|
||||
|
||||
return JSONResponse(content=result.model_dump())
|
||||
|
||||
|
||||
@router.post(
|
||||
"/v1/completions/render",
|
||||
dependencies=[Depends(validate_json_request)],
|
||||
|
||||
@@ -2,6 +2,8 @@
|
||||
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
|
||||
from typing import cast
|
||||
|
||||
from vllm.entrypoints.anthropic.protocol import AnthropicMessagesRequest
|
||||
from vllm.entrypoints.anthropic.serving import AnthropicServingMessages
|
||||
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
|
||||
from vllm.entrypoints.openai.completion.protocol import CompletionRequest
|
||||
from vllm.entrypoints.openai.engine.protocol import ErrorResponse
|
||||
@@ -51,6 +53,12 @@ class ServingRender(BaseServing):
|
||||
|
||||
self.online_renderer = online_renderer
|
||||
|
||||
self._merge_inline_system = (
|
||||
AnthropicServingMessages._detect_merge_inline_system(
|
||||
online_renderer.chat_template
|
||||
)
|
||||
)
|
||||
|
||||
self.default_sampling_params = (
|
||||
online_renderer.model_config.get_diff_sampling_param()
|
||||
)
|
||||
@@ -149,6 +157,21 @@ class ServingRender(BaseServing):
|
||||
token_offsets=engine_input.get("prompt_token_offsets"),
|
||||
)
|
||||
|
||||
async def render_messages_request(
|
||||
self,
|
||||
request: AnthropicMessagesRequest,
|
||||
) -> GenerateRequest | ErrorResponse:
|
||||
"""Validate the model and preprocess an Anthropic Messages request.
|
||||
|
||||
Converts the request to the OpenAI chat format using the same
|
||||
conversion as the /v1/messages server path, then delegates to
|
||||
render_chat_request so the rendered tokens match the server exactly.
|
||||
"""
|
||||
chat_req = AnthropicServingMessages.to_chat_completion_request(
|
||||
request, merge_inline_system=self._merge_inline_system
|
||||
)
|
||||
return await self.render_chat_request(chat_req)
|
||||
|
||||
async def render_completion_request(
|
||||
self,
|
||||
request: CompletionRequest,
|
||||
|
||||
Reference in New Issue
Block a user