[Bugfix] Reject removed pooling parameters (#48984)

Signed-off-by: Taneem Ibrahim <taneem.ibrahim@gmail.com>
This commit is contained in:
Taneem Ibrahim
2026-07-19 05:18:03 -07:00
committed by GitHub
parent ac5f38a0f7
commit e6d1310b2a
9 changed files with 108 additions and 17 deletions
+38
View File
@@ -1,12 +1,18 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from dataclasses import dataclass
from typing import Any
import pytest
from pydantic import TypeAdapter, ValidationError
from tests.models.utils import EmbedModelInfo
from vllm import PoolingParams
from vllm.config import ModelConfig, PoolerConfig
from vllm.entrypoints.pooling.classify.protocol import ClassificationRequest
from vllm.entrypoints.pooling.embed.protocol import EmbeddingRequest
from vllm.entrypoints.pooling.pooling.protocol import PoolingRequest
from vllm.exceptions import VLLMValidationError
EMBEDDING_MODELS = [
EmbedModelInfo("intfloat/multilingual-e5-small", is_matryoshka=False),
@@ -27,6 +33,38 @@ class MockModelConfig:
pooler_config: PoolerConfig
@pytest.mark.parametrize(
("parameter", "value", "message"),
[
(
"normalize",
False,
"Parameter `normalize` was removed; use `use_activation` instead.",
),
("task", "score", "`score` task was removed; use `classify` instead."),
(
"task",
"encode",
"`encode` task was removed; use `token_embed` or `token_classify` instead.",
),
],
)
def test_removed_pooling_parameters(parameter: str, value: Any, message: str):
data = {"input": "hello", parameter: value}
for request_type in (EmbeddingRequest, ClassificationRequest, PoolingRequest):
with pytest.raises(ValidationError, match=message) as exc_info:
TypeAdapter(request_type).validate_python(data)
assert len(exc_info.value.errors()) == 1
with pytest.raises(ValidationError, match=message) as exc_info:
TypeAdapter(PoolerConfig).validate_python({parameter: value})
assert len(exc_info.value.errors()) == 1
if parameter == "task":
with pytest.raises(VLLMValidationError, match=message):
PoolingParams(task=value)
def test_embed():
task = "embed"
model_config = MockModelConfig(pooler_config=PoolerConfig(seq_pooling_type="CLS"))
+17 -1
View File
@@ -3,9 +3,12 @@
from typing import Any, Literal, get_args
from pydantic import model_validator
from pydantic_core import ArgsKwargs
from vllm.config.utils import config
from vllm.logger import init_logger
from vllm.tasks import PoolingTask
from vllm.tasks import PoolingTask, check_removed_pooling_task
from vllm.utils.hashing import safe_hash
logger = init_logger(__name__)
@@ -112,6 +115,19 @@ class PoolerConfig:
`math-shepherd-mistral-7b-prm` model.
"""
@model_validator(mode="before")
@classmethod
def reject_removed_parameters(cls, data):
values = data.kwargs if isinstance(data, ArgsKwargs) else data
if not isinstance(values, dict):
return data
if "normalize" in values:
raise ValueError(
"Parameter `normalize` was removed; use `use_activation` instead."
)
check_removed_pooling_task(values.get("task"))
return data
def __post_init__(self) -> None:
if self.logit_sigma is not None and self.logit_sigma == 0:
raise ValueError("logit_sigma cannot be 0 (division by zero)")
+13
View File
@@ -14,10 +14,23 @@ from vllm.entrypoints.chat_utils import (
from vllm.entrypoints.openai.engine.protocol import OpenAIBaseModel
from vllm.exceptions import VLLMValidationError
from vllm.renderers import ChatParams, TokenizeParams, merge_kwargs
from vllm.tasks import check_removed_pooling_task
from vllm.utils import random_uuid
from vllm.utils.serial_utils import EmbedDType, EncodingFormat, Endianness
def reject_removed_pooling_parameters(data):
if not isinstance(data, dict):
return data
if "normalize" in data:
raise VLLMValidationError(
"Parameter `normalize` was removed; use `use_activation` instead.",
parameter="normalize",
)
check_removed_pooling_task(data.get("task"))
return data
class PoolingBasicRequestMixin(OpenAIBaseModel):
# --8<-- [start:pooling-common-params]
model: str | None = None
@@ -2,9 +2,9 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import time
from typing import TypeAlias
from typing import Annotated, TypeAlias
from pydantic import Field
from pydantic import BeforeValidator, Field
from vllm import PoolingParams
from vllm.entrypoints.openai.engine.protocol import OpenAIBaseModel, UsageInfo
@@ -17,6 +17,7 @@ from ..base.protocol import (
CompletionRequestMixin,
FixedMaxLenTokenizeParamsMixin,
PoolingBasicRequestMixin,
reject_removed_pooling_parameters,
)
logger = init_logger(__name__)
@@ -48,9 +49,10 @@ class ClassificationChatRequest(
)
ClassificationRequest: TypeAlias = (
ClassificationCompletionRequest | ClassificationChatRequest
)
ClassificationRequest: TypeAlias = Annotated[
ClassificationCompletionRequest | ClassificationChatRequest,
BeforeValidator(reject_removed_pooling_parameters),
]
class ClassificationData(OpenAIBaseModel):
+6 -4
View File
@@ -13,7 +13,7 @@ from collections.abc import Sequence
from typing import Annotated, Any, Literal, TypeAlias
import pybase64 as base64
from pydantic import BaseModel, Field, model_validator
from pydantic import BaseModel, BeforeValidator, Field, model_validator
from vllm import PoolingParams
from vllm.entrypoints.chat_utils import ChatCompletionMessageParam
@@ -27,6 +27,7 @@ from ..base.protocol import (
EmbeddingTokenizeParamsMixin,
EmbedRequestMixin,
PoolingBasicRequestMixin,
reject_removed_pooling_parameters,
)
@@ -154,13 +155,14 @@ class EmbeddingBatchChatInputRequest(EmbeddingBatchChatRequest):
return normalized
EmbeddingRequest: TypeAlias = (
EmbeddingRequest: TypeAlias = Annotated[
EmbeddingCompletionRequest
| EmbeddingChatRequest
| EmbeddingBatchChatRequest
| EmbeddingChatInputRequest
| EmbeddingBatchChatInputRequest
)
| EmbeddingBatchChatInputRequest,
BeforeValidator(reject_removed_pooling_parameters),
]
# ---------------------------------------------------------------------------
+7 -5
View File
@@ -1,9 +1,9 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
import time
from typing import Generic, TypeAlias, TypeVar
from typing import Annotated, Generic, TypeAlias, TypeVar
from pydantic import Field
from pydantic import BeforeValidator, Field
from vllm import PoolingParams
from vllm.config import ModelConfig
@@ -20,6 +20,7 @@ from ..base.protocol import (
EncodingRequestMixin,
FixedMaxLenTokenizeParamsMixin,
PoolingBasicRequestMixin,
reject_removed_pooling_parameters,
)
@@ -92,9 +93,10 @@ class IOProcessorResponse(OpenAIBaseModel, Generic[T]):
"""
PoolingRequest: TypeAlias = (
PoolingCompletionRequest | PoolingChatRequest | IOProcessorRequest
)
PoolingRequest: TypeAlias = Annotated[
PoolingCompletionRequest | PoolingChatRequest | IOProcessorRequest,
BeforeValidator(reject_removed_pooling_parameters),
]
class PoolingResponseData(OpenAIBaseModel):
+3 -1
View File
@@ -1,5 +1,7 @@
# SPDX-License-Identifier: Apache-2.0
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import cast
from fastapi.responses import JSONResponse, Response, StreamingResponse
from typing_extensions import assert_never
@@ -52,7 +54,7 @@ class ServingPooling(PoolingBaseServing):
self.json_response_cls = get_json_response_cls()
def get_io_processor(self, request: AnyPoolingRequest) -> PoolingIOProcessor:
assert isinstance(request, PoolingRequest)
request = cast(PoolingRequest, request)
pooling_task = self._verify_pooling_task(request)
return self.io_processors[pooling_task]
+2 -1
View File
@@ -9,7 +9,7 @@ import msgspec
from vllm.config import ModelConfig, PoolerConfig
from vllm.logger import init_logger
from vllm.sampling_params import RequestOutputKind
from vllm.tasks import PoolingTask
from vllm.tasks import PoolingTask, check_removed_pooling_task
logger = init_logger(__name__)
@@ -229,6 +229,7 @@ class PoolingParams(
)
def __post_init__(self) -> None:
check_removed_pooling_task(self.task)
if self.output_kind != RequestOutputKind.FINAL_ONLY:
raise ValueError(
"For pooling output_kind has to be FINAL_ONLY, "
+15
View File
@@ -2,6 +2,8 @@
# SPDX-FileCopyrightText: Copyright contributors to the vLLM project
from typing import Literal, get_args
from vllm.exceptions import VLLMValidationError
GenerationTask = Literal["generate", "transcription", "realtime"]
GENERATION_TASKS: tuple[GenerationTask, ...] = get_args(GenerationTask)
@@ -15,6 +17,19 @@ PoolingTask = Literal[
]
POOLING_TASKS: tuple[PoolingTask, ...] = get_args(PoolingTask)
_REMOVED_POOLING_TASK_MESSAGES = {
"score": "`score` task was removed; use `classify` instead.",
"encode": (
"`encode` task was removed; use `token_embed` or `token_classify` instead."
),
}
def check_removed_pooling_task(task: object) -> None:
if isinstance(task, str) and (message := _REMOVED_POOLING_TASK_MESSAGES.get(task)):
raise VLLMValidationError(message, parameter="task")
ScoreType = Literal["bi-encoder", "cross-encoder", "late-interaction"]
SCORE_TYPE_MAP: dict[PoolingTask, ScoreType] = {
"embed": "bi-encoder",