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

Prevent the cross-encoder logic from being applied to classification tasks (#18838)

Browse Source 
Signed-off-by: Max de Bayser <mbayser@br.ibm.com>
Co-authored-by: Cyrus Leung <cyrus.tl.leung@gmail.com>
This commit is contained in:
Maximilien de Bayser
2025-05-28 23:16:17 -03:00
committed by GitHub
parent 269d901734
commit 515b413ebf
5 changed files with 40 additions and 24 deletions
Download Patch File Download Diff File Expand all files Collapse all files

6
vllm/model_executor/models/bert.py
View File

@@ -16,7 +16,7 @@ from vllm.model_executor.layers.activation import get_act_fn
from vllm.model_executor.layers.linear import (ColumnParallelLinear,
QKVParallelLinear,
RowParallelLinear)
from vllm.model_executor.layers.pooler import (CrossEncodingPooler, Pooler,
from vllm.model_executor.layers.pooler import (ClassifierPooler, Pooler,
PoolingType)
from vllm.model_executor.layers.quantization import QuantizationConfig
from vllm.model_executor.layers.vocab_parallel_embedding import (
@@ -470,8 +470,8 @@ class BertForSequenceClassification(nn.Module, SupportsCrossEncoding,
embedding_class=BertEmbedding,
add_pooling_layer=True)
self.classifier = nn.Linear(config.hidden_size, config.num_labels)
self._pooler = CrossEncodingPooler(config, self.classifier,
self.bert.pooler)
self._pooler = ClassifierPooler(vllm_config.model_config,
self.classifier, self.bert.pooler)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):