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

4
vllm/model_executor/models/roberta.py
View File

@@ -23,7 +23,7 @@ from vllm.model_executor.models.utils import (AutoWeightsLoader, WeightsMapper,
from vllm.sequence import IntermediateTensors
from .bert_with_rope import BertWithRope, JinaRobertaModel
from .interfaces import SupportsCrossEncoding
from .interfaces import SupportsCrossEncoding, default_pooling_type
class RobertaEmbedding(nn.Module):
@@ -86,6 +86,7 @@ class RobertaClassificationHead(nn.Module):
return x
@default_pooling_type("CLS")
class RobertaEmbeddingModel(BertEmbeddingModel):
"""A model that uses Roberta to provide embedding functionalities.
@@ -149,6 +150,7 @@ class RobertaEmbeddingModel(BertEmbeddingModel):
return loader.load_weights(weights_list, mapper=mapper)
@default_pooling_type("CLS")
class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding):
"""A model that uses Roberta to provide embedding functionalities.