[Mini-VLLM] Remove LoRA, plugins, speech-to-text and scale-out surfaces

Production runs one static model behind the OpenAI-compatible API. Every
adapter path and non-OpenAI HTTP surface is unreachable there.

- Delete the LoRA stack: vllm/lora, LoRARequest on the engine wire, the
  scheduler, worker, cudagraph and MoE plumbing, LoRA metrics, the lora
  CLI group, and the LoRA tests.
- Drop the plugin loader in vllm/plugins. Platform resolution is now
  CUDA-only and the general/stat-logger plugin hooks are gone.
- Delete the speech-to-text and realtime entrypoints, the scale-out
  render routes, and vllm/config/speech_to_text.py. GenerationTask is
  generate-only.
- Cut DecodeBenchConnector and the disagg-prefill doc and image. The
  Multi, Offloading and Example connectors stay; production uses the
  CPU KV offload tier.
- Delete the ROCm-only fused KDA chunk kernel (2346 lines) with its CMake
  arch gate, declarations and op registrations. The CUDA build never
  defined the guard and no Python caller existed.
- Delete the orphan Mistral and Pixtral chat fixtures and examples/.
- Fix _check_model: removing the LoRA early return made every request
  fail with "The model ... does not exist. Valid aliases: ...".

Verified on an RTX 3090: import sweep reports 1242 modules with 7 known
optional-dependency failures; a live server answers chat, completions,
responses, tokenize and detokenize with zero log errors and 404 for the
removed surfaces; collection yields 8558 tests with the 3 pre-existing
errors; ruff adds no new finding; the incremental _C build succeeds.

Co-authored-by: Hinata <hinata@idihore.id>
Signed-off-by: MOVZX <movzx@yahoo.com>
This commit is contained in:
2026-10-19 21:06:19 +07:00
co-authored by Hinata
parent 5f3c7bd8ab
commit 5aa2ac64f8
269 changed files with 417 additions and 29932 deletions
@@ -27,15 +27,7 @@ MODEL_NAME = "HuggingFaceH4/zephyr-7b-beta"
@pytest.fixture(scope="module")
def zephyr_lora_files():
"""Download zephyr LoRA files once per test session."""
from vllm.transformers_utils.repo_utils import hf_api
return hf_api().snapshot_download(repo_id="typeof/zephyr-7b-beta-lora")
@pytest.fixture(scope="module")
def server(zephyr_lora_files):
def server():
args = [
# use half precision for speed and memory savings in CI environment
"--dtype",
@@ -43,14 +35,6 @@ def server(zephyr_lora_files):
"--max-model-len",
"8192",
"--enforce-eager",
# lora config below
"--enable-lora",
"--lora-modules",
f"zephyr-lora={zephyr_lora_files}",
"--max-lora-rank",
"64",
"--max-cpu-loras",
"2",
"--max-num-seqs",
"128",
]
@@ -67,9 +51,8 @@ async def client(server):
@pytest.mark.asyncio
@pytest.mark.parametrize(
# first test base model, then test loras
"model_name",
[MODEL_NAME, "zephyr-lora"],
[MODEL_NAME],
)
async def test_no_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
messages = [
@@ -91,9 +74,8 @@ async def test_no_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
@pytest.mark.asyncio
@pytest.mark.parametrize(
# just test 1 lora hereafter
"model_name",
[MODEL_NAME, "zephyr-lora"],
[MODEL_NAME],
)
async def test_zero_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
messages = [
@@ -119,7 +101,7 @@ async def test_zero_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
@pytest.mark.asyncio
@pytest.mark.parametrize(
"model_name",
[MODEL_NAME, "zephyr-lora"],
[MODEL_NAME],
)
async def test_some_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
messages = [
@@ -145,7 +127,7 @@ async def test_some_logprobs_chat(client: openai.AsyncOpenAI, model_name: str):
@pytest.mark.asyncio
@pytest.mark.parametrize(
"model_name",
[MODEL_NAME, "zephyr-lora"],
[MODEL_NAME],
)
async def test_too_many_chat_logprobs(client: openai.AsyncOpenAI, model_name: str):
messages = [
@@ -254,7 +236,7 @@ async def test_more_than_one_prompt_logprobs_chat(
@pytest.mark.asyncio
@pytest.mark.parametrize(
"model_name",
[MODEL_NAME, "zephyr-lora"],
[MODEL_NAME],
)
async def test_single_chat_session(client: openai.AsyncOpenAI, model_name: str):
messages = [
@@ -297,9 +279,8 @@ async def test_single_chat_session(client: openai.AsyncOpenAI, model_name: str):
@pytest.mark.asyncio
@pytest.mark.parametrize(
# just test 1 lora hereafter
"model_name",
[MODEL_NAME, "zephyr-lora"],
[MODEL_NAME],
)
async def test_chat_streaming(client: openai.AsyncOpenAI, model_name: str):
messages = [
@@ -345,7 +326,7 @@ async def test_chat_streaming(client: openai.AsyncOpenAI, model_name: str):
@pytest.mark.asyncio
@pytest.mark.parametrize(
"model_name",
["HuggingFaceH4/zephyr-7b-beta", "zephyr-lora"],
["HuggingFaceH4/zephyr-7b-beta"],
)
async def test_chat_completion_stream_options(
client: openai.AsyncOpenAI, model_name: str
@@ -9,7 +9,6 @@ import openai # use the official client for correctness check
import pytest
import pytest_asyncio
# downloading lora to test lora requests
from tests.utils import ROCM_EXTRA_ARGS, RemoteOpenAIServer
# any model with a chat template should work here
+2 -15
View File
@@ -9,12 +9,10 @@ from tests.utils import RemoteOpenAIServer
# any model with a chat template should work here
MODEL_NAME = "Qwen/Qwen3-0.6B"
# technically this needs Mistral-7B-v0.1 as base, but we're not testing
# generation quality here
@pytest.fixture(scope="module")
def server(qwen3_lora_files):
def server():
args = [
# use half precision for speed and memory savings in CI environment
"--dtype",
@@ -22,14 +20,6 @@ def server(qwen3_lora_files):
"--max-model-len",
"8192",
"--enforce-eager",
# lora config below
"--enable-lora",
"--lora-modules",
f"qwen3-lora={qwen3_lora_files}",
"--max-lora-rank",
"64",
"--max-cpu-loras",
"2",
"--max-num-seqs",
"128",
]
@@ -45,12 +35,9 @@ async def client(server):
@pytest.mark.asyncio
async def test_check_models(client: openai.AsyncOpenAI, qwen3_lora_files):
async def test_check_models(client: openai.AsyncOpenAI):
models = await client.models.list()
models = models.data
served_model = models[0]
lora_models = models[1:]
assert served_model.id == MODEL_NAME
assert served_model.root == MODEL_NAME
assert all(lora_model.root == qwen3_lora_files for lora_model in lora_models)
assert lora_models[0].id == "qwen3-lora"
@@ -1,78 +0,0 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
"""Unit tests for the token-offsets request/response protocol wiring:
the request flag flowing into ``TokenizeParams`` and the ``GenerateRequest``
serialization boundary. Plain Pydantic field storage is not retested here.
"""
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.scale_out.token_in_token_out.protocol import GenerateRequest
from vllm.sampling_params import SamplingParams
def _model_config() -> Mock:
model_config = Mock(spec=ModelConfig)
model_config.max_model_len = 128
return model_config
def test_completion_flag_forwarded_to_tok_params():
"""build_tok_params must forward return_token_offsets, defaulting to
False (zero behavioral change for existing callers) and coercing JSON
null to False via the bool() guard."""
cfg = _model_config()
default = CompletionRequest(model="m", prompt="hi")
assert default.build_tok_params(cfg).return_token_offsets is False
on = CompletionRequest(model="m", prompt="hi", return_token_offsets=True)
assert on.build_tok_params(cfg).return_token_offsets is True
null = CompletionRequest(model="m", prompt="hi", return_token_offsets=None)
assert null.build_tok_params(cfg).return_token_offsets is False
def test_chat_flag_forwarded_to_tok_params():
"""Chat build_tok_params has its own (max_completion_tokens) branch, so
its return_token_offsets forwarding is verified independently."""
cfg = _model_config()
messages = [{"role": "user", "content": "hi"}]
default = ChatCompletionRequest(model="m", messages=messages)
assert default.build_tok_params(cfg).return_token_offsets is False
on = ChatCompletionRequest(model="m", messages=messages, return_token_offsets=True)
assert on.build_tok_params(cfg).return_token_offsets is True
null = ChatCompletionRequest(
model="m", messages=messages, return_token_offsets=None
)
assert null.build_tok_params(cfg).return_token_offsets is False
def test_generate_request_token_offsets_default_none():
"""Defaults to None so existing /v1/.../render responses are unchanged."""
req = GenerateRequest(token_ids=[1, 2, 3], sampling_params=SamplingParams())
assert req.token_offsets is None
def test_generate_request_token_offsets_survive_json_round_trip():
"""GenerateRequest crosses the disagg serialization boundary; the
tuple[int, int] offsets must survive model_dump and re-validate."""
req = GenerateRequest(
token_ids=[10, 20],
sampling_params=SamplingParams(),
token_offsets=[(0, 1), (1, 3)],
)
dumped = req.model_dump()
assert dumped["token_offsets"] == [(0, 1), (1, 3)]
# Re-validate from the dumped dict (sampling_params doesn't round-trip
# cleanly via dump, so re-inject a fresh instance).
again = GenerateRequest.model_validate(
{**dumped, "sampling_params": SamplingParams()}
)
assert again.token_offsets == [(0, 1), (1, 3)]