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

12
tests/models/language/pooling/mteb_utils.py
View File

@@ -177,9 +177,12 @@ def mteb_test_embed_models(hf_runner,
max_model_len=None,
**vllm_extra_kwargs) as vllm_model:
model_config = vllm_model.llm.llm_engine.model_config
if model_info.architecture:
assert (model_info.architecture
in vllm_model.llm.llm_engine.model_config.architectures)
assert model_info.architecture in model_config.architectures
assert (model_config._model_info.default_pooling_type ==
model_info.default_pooling_type)
vllm_main_score = run_mteb_embed_task(VllmMtebEncoder(vllm_model),
MTEB_EMBED_TASKS)
@@ -286,7 +289,12 @@ def mteb_test_rerank_models(hf_runner,
**vllm_extra_kwargs) as vllm_model:
model_config = vllm_model.llm.llm_engine.model_config
if model_info.architecture:
assert (model_info.architecture in model_config.architectures)
assert model_config.hf_config.num_labels == 1
assert (model_config._model_info.default_pooling_type ==
model_info.default_pooling_type)
vllm_main_score = run_mteb_rerank(vllm_mteb_encoder(vllm_model),
tasks=MTEB_RERANK_TASKS,