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

[Model] Reorganize pooling layers (#31973)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-09 19:02:14 +08:00
committed by GitHub
parent 020732800c
commit c8ed39b9dd
34 changed files with 1290 additions and 1143 deletions
Download Patch File Download Diff File Expand all files Collapse all files

24
vllm/model_executor/models/roberta.py
View File

@@ -8,12 +8,8 @@ from torch import nn
from transformers import RobertaConfig
from vllm.config import ModelConfig, VllmConfig
from vllm.model_executor.layers.pooler import (
ClassifierPooler,
CLSPool,
DispatchPooler,
Pooler,
)
from vllm.model_executor.layers.pooler import DispatchPooler
from vllm.model_executor.layers.pooler.seqwise import CLSPool
from vllm.model_executor.layers.vocab_parallel_embedding import VocabParallelEmbedding
from vllm.model_executor.models.bert import (
TOKEN_TYPE_SHIFT,
@@ -196,18 +192,10 @@ class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding):
pooler_config = vllm_config.model_config.pooler_config
assert pooler_config is not None
self.pooler = DispatchPooler(
{
"token_classify": Pooler.for_token_classify(
pooler_config=pooler_config, classifier=self.classifier
),
"classify": ClassifierPooler(
pooling=CLSPool(), classifier=self.classifier, act_fn="classify"
),
"score": ClassifierPooler(
pooling=CLSPool(), classifier=self.classifier, act_fn="score"
),
}
self.pooler = DispatchPooler.for_seq_cls(
pooler_config,
pooling=CLSPool(),
classifier=self.classifier,
)
def load_weights(self, weights: Iterable[tuple[str, torch.Tensor]]):