[Frontend] Consolidate scale out entrypoints (#44512)

Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
wang.yuqi
2026-06-29 03:18:53 -07:00
committed by GitHub
parent eddfd4cf21
commit 3483240b7e
45 changed files with 422 additions and 289 deletions
@@ -17,10 +17,9 @@ from vllm.entrypoints.openai.chat_completion.serving import OpenAIServingChat
from vllm.entrypoints.openai.engine.protocol import GenerationError
from vllm.entrypoints.openai.models.protocol import BaseModelPath
from vllm.entrypoints.openai.models.serving import OpenAIServingModels
from vllm.entrypoints.serve.render.serving import ServingRender
from vllm.entrypoints.scale_out.render.serving import ServingRender
from vllm.outputs import CompletionOutput, RequestOutput
from vllm.renderers.hf import HfRenderer
from vllm.renderers.online_derenderer import OnlineDerenderer
from vllm.renderers.online_renderer import OnlineRenderer
from vllm.tokenizers.registry import cached_tokenizer_from_config
from vllm.v1.engine.async_llm import AsyncLLM
@@ -206,15 +205,8 @@ def _build_serving_render(engine: AsyncLLM) -> ServingRender:
chat_template=None,
chat_template_content_format="auto",
)
online_derenderer = OnlineDerenderer(
model_config=engine.model_config,
renderer=engine.renderer,
request_logger=None,
chat_template=None,
chat_template_content_format="auto",
)
serving_render = ServingRender(models, online_renderer, online_derenderer)
serving_render = ServingRender(models, online_renderer)
async def _fake_preprocess_chat(*args, **kwargs):
# return conversation, engine_inputs
@@ -14,10 +14,9 @@ from vllm.entrypoints.openai.completion.serving import OpenAIServingCompletion
from vllm.entrypoints.openai.engine.protocol import GenerationError
from vllm.entrypoints.openai.models.protocol import BaseModelPath
from vllm.entrypoints.openai.models.serving import OpenAIServingModels
from vllm.entrypoints.serve.render.serving import ServingRender
from vllm.entrypoints.scale_out.render.serving import ServingRender
from vllm.outputs import CompletionOutput, RequestOutput
from vllm.renderers.hf import HfRenderer
from vllm.renderers.online_derenderer import OnlineDerenderer
from vllm.renderers.online_renderer import OnlineRenderer
from vllm.tokenizers.registry import cached_tokenizer_from_config
from vllm.v1.engine.async_llm import AsyncLLM
@@ -191,15 +190,8 @@ def _build_serving_render(engine: AsyncLLM) -> ServingRender:
chat_template=None,
chat_template_content_format="auto",
)
online_derenderer = OnlineDerenderer(
model_config=engine.model_config,
renderer=engine.renderer,
request_logger=None,
chat_template=None,
chat_template_content_format="auto",
)
serving_render = ServingRender(models, online_renderer, online_derenderer)
serving_render = ServingRender(models, online_renderer)
async def _fake_preprocess_chat(*args, **kwargs):
# return conversation, engine_inputs
-111
View File
@@ -1,111 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Roundtrip tests for multimodal serde used by the disagg generate endpoint."""
import torch
from vllm.entrypoints.serve.disagg.mm_serde import (
decode_mm_kwargs_item,
encode_mm_kwargs_item,
)
from vllm.entrypoints.serve.disagg.protocol import (
MultiModalFeatures,
PlaceholderRangeInfo,
)
from vllm.multimodal.inputs import (
MultiModalBatchedField,
MultiModalFieldElem,
MultiModalFlatField,
MultiModalKwargsItem,
MultiModalSharedField,
)
def test_mm_kwargs_item_roundtrip():
"""Full roundtrip test with all three field types and multiple dtypes."""
e1 = MultiModalFieldElem(
data=torch.zeros(1000, dtype=torch.bfloat16),
field=MultiModalBatchedField(),
)
e2 = MultiModalFieldElem(
data=torch.ones(100, dtype=torch.int32),
field=MultiModalSharedField(batch_size=4),
)
e3 = MultiModalFieldElem(
data=torch.randn(20, dtype=torch.float32),
field=MultiModalFlatField(slices=[slice(0, 10), slice(10, 20)], dim=0),
)
item = MultiModalKwargsItem({"pixel_values": e1, "grid_thw": e2, "embeds": e3})
encoded = encode_mm_kwargs_item(item)
# Encoded result is a base64 string
assert isinstance(encoded, str)
decoded = decode_mm_kwargs_item(encoded)
assert set(decoded.keys()) == {"pixel_values", "grid_thw", "embeds"}
assert torch.equal(item["pixel_values"].data, decoded["pixel_values"].data)
assert torch.equal(item["grid_thw"].data, decoded["grid_thw"].data)
assert torch.equal(item["embeds"].data, decoded["embeds"].data)
assert isinstance(decoded["pixel_values"].field, MultiModalBatchedField)
assert isinstance(decoded["grid_thw"].field, MultiModalSharedField)
assert isinstance(decoded["embeds"].field, MultiModalFlatField)
def test_mm_kwargs_item_none_data():
"""Roundtrip with None data field."""
elem = MultiModalFieldElem(
data=None,
field=MultiModalSharedField(batch_size=2),
)
item = MultiModalKwargsItem({"empty": elem})
encoded = encode_mm_kwargs_item(item)
decoded = decode_mm_kwargs_item(encoded)
assert decoded["empty"].data is None
assert isinstance(decoded["empty"].field, MultiModalSharedField)
def test_mm_kwargs_item_nested_tensors():
"""Roundtrip with nested tensor data."""
nested = [torch.randn(3, 4), torch.randn(5, 4)]
elem = MultiModalFieldElem(
data=nested,
field=MultiModalBatchedField(),
)
item = MultiModalKwargsItem({"nested": elem})
encoded = encode_mm_kwargs_item(item)
decoded = decode_mm_kwargs_item(encoded)
decoded_data = decoded["nested"].data
assert len(decoded_data) == 2
assert torch.equal(nested[0], decoded_data[0])
assert torch.equal(nested[1], decoded_data[1])
def test_mm_features_with_kwargs_data():
"""Test that MultiModalFeatures can carry serialized tensor data."""
elem = MultiModalFieldElem(
data=torch.randn(5, 3, dtype=torch.float32),
field=MultiModalBatchedField(),
)
item = MultiModalKwargsItem({"pixel_values": elem})
encoded = encode_mm_kwargs_item(item)
features = MultiModalFeatures(
mm_hashes={"image": ["abc123"]},
mm_placeholders={"image": [PlaceholderRangeInfo(offset=0, length=10)]},
kwargs_data={"image": [encoded]},
)
# JSON roundtrip
json_str = features.model_dump_json()
features2 = MultiModalFeatures.model_validate_json(json_str)
assert features2.mm_hashes == {"image": ["abc123"]}
assert features2.kwargs_data is not None
assert len(features2.kwargs_data["image"]) == 1
decoded = decode_mm_kwargs_item(features2.kwargs_data["image"][0])
assert torch.equal(elem.data, decoded["pixel_values"].data)
@@ -3,7 +3,7 @@
"""Unit tests for the token-offsets request/response protocol wiring:
the request flag flowing into ``TokenizeParams`` and the ``GenerateRequest``
serialization boundary. End-to-end behavior is covered by
``tests/entrypoints/serve/render/test_render.py``; plain Pydantic field
``tests/entrypoints/scale_out/render/test_render.py``; plain Pydantic field
storage is not retested here.
"""
@@ -12,7 +12,7 @@ from unittest.mock import Mock
from vllm.config import ModelConfig
from vllm.entrypoints.openai.chat_completion.protocol import ChatCompletionRequest
from vllm.entrypoints.openai.completion.protocol import CompletionRequest
from vllm.entrypoints.serve.disagg.protocol import GenerateRequest
from vllm.entrypoints.scale_out.token_in_token_out.protocol import GenerateRequest
from vllm.sampling_params import SamplingParams