[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
@@ -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
@@ -12,11 +12,11 @@ from vllm.config.multimodal import MultiModalConfig
from vllm.entrypoints.openai.engine.protocol import StreamOptions
from vllm.entrypoints.openai.models.protocol import BaseModelPath
from vllm.entrypoints.openai.models.serving import OpenAIServingModels
from vllm.entrypoints.serve.disagg.protocol import (
from vllm.entrypoints.scale_out.token_in_token_out.protocol import (
GenerateRequest,
GenerateResponse,
)
from vllm.entrypoints.serve.disagg.serving import ServingTokens
from vllm.entrypoints.scale_out.token_in_token_out.serving import ServingTokens
from vllm.logprobs import Logprob
from vllm.outputs import CompletionOutput, RequestOutput
from vllm.renderers import renderer_from_config
@@ -1,14 +1,17 @@
# 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."""
"""
Roundtrip tests for multimodal serde used by the
token_in_token_out generate endpoint.
"""
import torch
from vllm.entrypoints.serve.disagg.mm_serde import (
from vllm.entrypoints.scale_out.token_in_token_out.mm_serde import (
decode_mm_kwargs_item,
encode_mm_kwargs_item,
)
from vllm.entrypoints.serve.disagg.protocol import (
from vllm.entrypoints.scale_out.token_in_token_out.protocol import (
MultiModalFeatures,
PlaceholderRangeInfo,
)
@@ -1,6 +1,6 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for the disagg request/response protocol.
"""Unit tests for the token_in_token_out request/response protocol.
These tests intentionally avoid spinning up a server — they exercise the
pydantic validators on ``GenerateRequest`` directly so they run fast and
@@ -9,7 +9,7 @@ fail loudly if the validator semantics ever drift.
import json
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
@@ -1,7 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from vllm.entrypoints.serve.disagg.serving import ServingTokens
from vllm.entrypoints.scale_out.token_in_token_out.serving import ServingTokens
from vllm.logprobs import Logprob
+1 -1
View File
@@ -5,7 +5,7 @@
These exercise ``_tokenize_prompt`` (offset extraction + capability/MM
gating) and the ``_tokenize_prompt -> _process_tokens -> TokensInput``
forwarding chain. Endpoint-level coverage lives in
``tests/entrypoints/serve/render/test_render.py``.
``tests/entrypoints/scale_out/render/test_render.py``.
"""
import pytest