[Misc] Rename embedding classes to pooling (#10801)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2024-12-01 14:36:51 +08:00
committed by GitHub
parent f877a7d12a
commit d2f058e76c
25 changed files with 166 additions and 123 deletions
+5 -6
View File
@@ -1,15 +1,14 @@
from .interfaces import (HasInnerState, SupportsLoRA, SupportsMultiModal,
SupportsPP, has_inner_state, supports_lora,
supports_multimodal, supports_pp)
from .interfaces_base import (VllmModelForEmbedding,
VllmModelForTextGeneration, is_embedding_model,
is_text_generation_model)
from .interfaces_base import (VllmModelForPooling, VllmModelForTextGeneration,
is_pooling_model, is_text_generation_model)
from .registry import ModelRegistry
__all__ = [
"ModelRegistry",
"VllmModelForEmbedding",
"is_embedding_model",
"VllmModelForPooling",
"is_pooling_model",
"VllmModelForTextGeneration",
"is_text_generation_model",
"HasInnerState",
@@ -20,4 +19,4 @@ __all__ = [
"supports_multimodal",
"SupportsPP",
"supports_pp",
]
]
+3 -3
View File
@@ -4,7 +4,7 @@ from typing import Any, TypeVar
import torch
import torch.nn as nn
from .interfaces_base import VllmModelForEmbedding, is_embedding_model
from .interfaces_base import VllmModelForPooling, is_pooling_model
_T = TypeVar("_T", bound=type[nn.Module])
@@ -12,7 +12,7 @@ _T = TypeVar("_T", bound=type[nn.Module])
def as_embedding_model(cls: _T) -> _T:
"""Subclass an existing vLLM model to support embeddings."""
# Avoid modifying existing embedding models
if is_embedding_model(cls):
if is_pooling_model(cls):
return cls
# Lazy import
@@ -23,7 +23,7 @@ def as_embedding_model(cls: _T) -> _T:
from .utils import AutoWeightsLoader, WeightsMapper
class ModelForEmbedding(cls, VllmModelForEmbedding):
class ModelForEmbedding(cls, VllmModelForPooling):
def __init__(
self,
+2 -2
View File
@@ -7,7 +7,7 @@ from typing_extensions import TypeIs, TypeVar
from vllm.logger import init_logger
from vllm.utils import supports_kw
from .interfaces_base import is_embedding_model
from .interfaces_base import is_pooling_model
if TYPE_CHECKING:
from vllm.attention import AttentionMetadata
@@ -389,4 +389,4 @@ def _supports_cross_encoding(
def supports_cross_encoding(
model: Union[Type[object], object],
) -> Union[TypeIs[Type[SupportsCrossEncoding]], TypeIs[SupportsCrossEncoding]]:
return is_embedding_model(model) and _supports_cross_encoding(model)
return is_pooling_model(model) and _supports_cross_encoding(model)
@@ -141,7 +141,7 @@ def is_text_generation_model(
@runtime_checkable
class VllmModelForEmbedding(VllmModel[C_co, T], Protocol[C_co, T]):
class VllmModelForPooling(VllmModel[C_co, T], Protocol[C_co, T]):
def pooler(
self,
@@ -153,23 +153,22 @@ class VllmModelForEmbedding(VllmModel[C_co, T], Protocol[C_co, T]):
@overload
def is_embedding_model(
model: Type[object]) -> TypeIs[Type[VllmModelForEmbedding]]:
def is_pooling_model(model: Type[object]) -> TypeIs[Type[VllmModelForPooling]]:
...
@overload
def is_embedding_model(model: object) -> TypeIs[VllmModelForEmbedding]:
def is_pooling_model(model: object) -> TypeIs[VllmModelForPooling]:
...
def is_embedding_model(
def is_pooling_model(
model: Union[Type[object], object],
) -> Union[TypeIs[Type[VllmModelForEmbedding]], TypeIs[VllmModelForEmbedding]]:
) -> Union[TypeIs[Type[VllmModelForPooling]], TypeIs[VllmModelForPooling]]:
if not is_vllm_model(model):
return False
if isinstance(model, type):
return isinstance(model, VllmModelForEmbedding)
return isinstance(model, VllmModelForPooling)
return isinstance(model, VllmModelForEmbedding)
return isinstance(model, VllmModelForPooling)
+8 -8
View File
@@ -24,7 +24,7 @@ from .adapters import as_embedding_model
from .interfaces import (has_inner_state, is_attention_free,
supports_cross_encoding, supports_multimodal,
supports_pp)
from .interfaces_base import is_embedding_model, is_text_generation_model
from .interfaces_base import is_pooling_model, is_text_generation_model
logger = init_logger(__name__)
@@ -211,7 +211,7 @@ _ROCM_PARTIALLY_SUPPORTED_MODELS: Dict[str, str] = {
class _ModelInfo:
architecture: str
is_text_generation_model: bool
is_embedding_model: bool
is_pooling_model: bool
supports_cross_encoding: bool
supports_multimodal: bool
supports_pp: bool
@@ -220,19 +220,19 @@ class _ModelInfo:
@staticmethod
def from_model_cls(model: Type[nn.Module]) -> "_ModelInfo":
is_embedding_model_ = is_embedding_model(model)
if not is_embedding_model_:
is_pooling_model_ = is_pooling_model(model)
if not is_pooling_model_:
try:
as_embedding_model(model)
except Exception:
pass
else:
is_embedding_model_ = True
is_pooling_model_ = True
return _ModelInfo(
architecture=model.__name__,
is_text_generation_model=is_text_generation_model(model),
is_embedding_model=is_embedding_model_,
is_pooling_model=is_pooling_model_,
supports_cross_encoding=supports_cross_encoding(model),
supports_multimodal=supports_multimodal(model),
supports_pp=supports_pp(model),
@@ -441,12 +441,12 @@ class _ModelRegistry:
model_cls, _ = self.inspect_model_cls(architectures)
return model_cls.is_text_generation_model
def is_embedding_model(
def is_pooling_model(
self,
architectures: Union[str, List[str]],
) -> bool:
model_cls, _ = self.inspect_model_cls(architectures)
return model_cls.is_embedding_model
return model_cls.is_pooling_model
def is_cross_encoder_model(
self,