[Model] Pooling models default to using chunked prefill & prefix caching if supported. (#20930)
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
@@ -25,8 +25,8 @@ from vllm.logger import init_logger
|
||||
from vllm.transformers_utils.dynamic_module import (
|
||||
try_get_class_from_dynamic_module)
|
||||
|
||||
from .interfaces import (has_inner_state, has_noops, is_attention_free,
|
||||
is_hybrid, supports_cross_encoding,
|
||||
from .interfaces import (get_default_pooling_type, has_inner_state, has_noops,
|
||||
is_attention_free, is_hybrid, supports_cross_encoding,
|
||||
supports_multimodal, supports_multimodal_raw_input,
|
||||
supports_pp, supports_transcription, supports_v0_only)
|
||||
from .interfaces_base import is_pooling_model, is_text_generation_model
|
||||
@@ -305,6 +305,7 @@ class _ModelInfo:
|
||||
architecture: str
|
||||
is_text_generation_model: bool
|
||||
is_pooling_model: bool
|
||||
default_pooling_type: str
|
||||
supports_cross_encoding: bool
|
||||
supports_multimodal: bool
|
||||
supports_multimodal_raw_input: bool
|
||||
@@ -323,6 +324,7 @@ class _ModelInfo:
|
||||
architecture=model.__name__,
|
||||
is_text_generation_model=is_text_generation_model(model),
|
||||
is_pooling_model=is_pooling_model(model),
|
||||
default_pooling_type=get_default_pooling_type(model),
|
||||
supports_cross_encoding=supports_cross_encoding(model),
|
||||
supports_multimodal=supports_multimodal(model),
|
||||
supports_multimodal_raw_input=supports_multimodal_raw_input(model),
|
||||
|
||||
Reference in New Issue
Block a user