[Bugfix] Reject beam search in completion render (#61212)

Signed-off-by: bzsuni <bingzhe.sun@daocloud.io>
This commit is contained in:
bzsuni
2026-10-12 13:49:46 +00:00
committed by GitHub
parent 4c36d21280
commit f022975964
2 changed files with 37 additions and 5 deletions
@@ -12,6 +12,7 @@ import pytest
import pytest_asyncio
from tests.utils import RemoteLaunchRenderServer
from vllm.entrypoints.openai.completion.protocol import CompletionRequest
from vllm.entrypoints.openai.responses.protocol import ResponsesRequest
from vllm.entrypoints.scale_out.render.api_router import router
from vllm.entrypoints.scale_out.render.serving import ServingRender
@@ -21,7 +22,7 @@ from vllm.renderers.online_renderer import OnlineRenderer
MODEL_NAME = "hmellor/tiny-random-LlamaForCausalLM"
def _build_responses_serving_render() -> ServingRender:
def _build_serving_render() -> ServingRender:
serving = ServingRender.__new__(ServingRender)
serving.model_config = SimpleNamespace(
max_model_len=100,
@@ -43,6 +44,31 @@ def _build_responses_serving_render() -> ServingRender:
return serving
@pytest.mark.asyncio
@pytest.mark.skip_global_cleanup
async def test_render_completion_rejects_beam_search():
serving = _build_serving_render()
serving.online_renderer.render_completion = AsyncMock(
return_value=[{"prompt_token_ids": [7, 8, 9]}]
)
request = CompletionRequest(
model=MODEL_NAME,
prompt="Test prompt",
use_beam_search=True,
n=2,
length_penalty=0.7,
)
response = await serving.render_completion_request(request)
assert isinstance(response, ErrorResponse)
assert response.error.code == 400
assert (
response.error.message == "Beam search is not supported by the render endpoint"
)
serving.online_renderer.render_completion.assert_not_awaited()
@pytest.mark.skip_global_cleanup
def test_responses_render_route_is_registered():
assert any(route.path == "/v1/responses/render" for route in router.routes)
@@ -60,7 +86,7 @@ def test_responses_render_route_documents_not_implemented():
@pytest.mark.asyncio
@pytest.mark.skip_global_cleanup
async def test_render_responses_returns_generate_request_without_stored_state():
serving = _build_responses_serving_render()
serving = _build_serving_render()
serving.online_renderer.render_responses = AsyncMock(
return_value=MagicMock(
messages=[{"role": "user", "content": "Test prompt"}],
@@ -101,7 +127,7 @@ async def test_render_responses_returns_generate_request_without_stored_state():
@pytest.mark.asyncio
@pytest.mark.skip_global_cleanup
async def test_render_responses_rejects_previous_response_id():
serving = _build_responses_serving_render()
serving = _build_serving_render()
serving.online_renderer.render_responses = AsyncMock()
request = ResponsesRequest(
model=MODEL_NAME,
@@ -120,7 +146,7 @@ async def test_render_responses_rejects_previous_response_id():
@pytest.mark.asyncio
@pytest.mark.skip_global_cleanup
async def test_render_responses_rejects_empty_token_ids():
serving = _build_responses_serving_render()
serving = _build_serving_render()
serving.online_renderer.render_responses = AsyncMock(
return_value=MagicMock(
messages=[],
@@ -141,7 +167,7 @@ async def test_render_responses_rejects_empty_token_ids():
async def test_render_responses_carries_reasoning_parser_kwargs():
"""/v1/responses gives the engine's reasoning parser the effective template
kwargs, so the rendered request must carry them to generate."""
serving = _build_responses_serving_render()
serving = _build_serving_render()
online = serving.online_renderer
online.parser = MagicMock(reasoning_parser_cls=object)
online.chat_template = None
@@ -186,6 +186,12 @@ class ServingRender(BaseServing):
error_check_ret = await self._check_model(request)
if error_check_ret is not None:
return error_check_ret
if request.use_beam_search:
return self.create_error_response(
"Beam search is not supported by the render endpoint"
)
result = await self.online_renderer.render_completion(
request, skip_mm_cache=True
)