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

[Model] Pooling models default to using chunked prefill & prefix caching if supported. (#20930)

Browse Source 
Signed-off-by: wang.yuqi <noooop@126.com>
This commit is contained in:
wang.yuqi
2025-08-12 00:41:37 +08:00
committed by GitHub
parent 16fb668b61
commit 84cf78acee
31 changed files with 452 additions and 261 deletions
Download Patch File Download Diff File Expand all files Collapse all files

18
tests/models/utils.py
View File

@@ -345,16 +345,34 @@ class EmbedModelInfo(NamedTuple):
matryoshka_dimensions: Optional[list[int]] = None
architecture: str = ""
dtype: str = "auto"
default_pooling_type: str = ""
enable_test: bool = True
class CLSPoolingEmbedModelInfo(EmbedModelInfo):
default_pooling_type: str = "CLS"
class LASTPoolingEmbedModelInfo(EmbedModelInfo):
default_pooling_type: str = "LAST"
class RerankModelInfo(NamedTuple):
name: str
architecture: str = ""
dtype: str = "auto"
default_pooling_type: str = ""
enable_test: bool = True
class CLSPoolingRerankModelInfo(RerankModelInfo):
default_pooling_type: str = "CLS"
class LASTPoolingRerankModelInfo(RerankModelInfo):
default_pooling_type: str = "LAST"
def dummy_hf_overrides(
hf_config: PretrainedConfig,
*,