biondizzle/vllm
1
0
Fork 0
Code Issues Pull Requests Actions 2 Packages Projects Releases Wiki Activity

[Model][7/N] Improve all pooling task | Deprecation as_reward_model. Extract hidden states prefer using new multi-vector retrieval API (#26686)

Browse Source 
Signed-off-by: wang.yuqi <yuqi.wang@daocloud.io>
This commit is contained in:
wang.yuqi
2025-12-08 16:10:09 +08:00
committed by GitHub
parent bcb6f5947f
commit 9e77ffca3f
8 changed files with 88 additions and 58 deletions
Download Patch File Download Diff File Expand all files Collapse all files

4
vllm/model_executor/model_loader/utils.py
View File

@@ -167,7 +167,6 @@ _MODEL_ARCH_BY_HASH = dict[int, tuple[type[nn.Module], str]]()
def _get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module], str]:
from vllm.model_executor.models.adapters import (
as_embedding_model,
as_reward_model,
as_seq_cls_model,
try_create_mm_pooling_model_cls,
)
@@ -207,9 +206,6 @@ def _get_model_architecture(model_config: ModelConfig) -> tuple[type[nn.Module],
elif convert_type == "classify":
logger.debug_once("Converting to sequence classification model.")
model_cls = as_seq_cls_model(model_cls)
elif convert_type == "reward":
logger.debug_once("Converting to reward model.")
model_cls = as_reward_model(model_cls)
else:
assert_never(convert_type)

38
vllm/model_executor/models/adapters.py
View File

@@ -346,44 +346,6 @@ def as_seq_cls_model(cls: _T) -> _T:
return ModelForSequenceClassification # type: ignore
def as_reward_model(cls: _T) -> _T:
"""
Subclass an existing vLLM model to support reward modeling.
By default, we return the hidden states of each token directly.
Note:
We assume that no extra layers are added to the original model;
please implement your own model if this is not the case.
"""
# Avoid modifying existing reward models
if is_pooling_model(cls):
return cls
# Lazy import
from vllm.model_executor.layers.pooler import DispatchPooler, Pooler
from .interfaces_base import default_pooling_type
@default_pooling_type("ALL")
class ModelForReward(_create_pooling_model_cls(cls)):
def _init_pooler(self, vllm_config: "VllmConfig", prefix: str = ""):
pooler_config = vllm_config.model_config.pooler_config
assert pooler_config is not None
self.pooler = DispatchPooler(
{
"token_classify": Pooler.for_token_classify(
pooler_config=pooler_config
)
}
)
ModelForReward.__name__ = _get_pooling_model_name(cls.__name__, "ForReward")
return ModelForReward # type: ignore
class SequenceClassificationConfig(VerifyAndUpdateConfig):
@staticmethod
def verify_and_update_config(vllm_config: "VllmConfig") -> None: