[Bugfix] Add use_cross_encoder flag to use correct activation in ClassifierPooler (#20527)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2025-07-06 14:01:48 -07:00
committed by GitHub
parent 9528e3a05e
commit c18b3b8e8b
8 changed files with 56 additions and 41 deletions
+31 -11
View File
@@ -15,6 +15,7 @@ from vllm.model_executor.pooling_metadata import ( # noqa: E501
from vllm.model_executor.pooling_metadata import PoolingTensors
from vllm.sequence import PoolerOutput, PoolingSequenceGroupOutput
from vllm.transformers_utils.config import (
get_classification_activation_function,
get_cross_encoder_activation_function)
from vllm.v1.pool.metadata import PoolingMetadata as V1PoolingMetadata
@@ -388,15 +389,14 @@ class ClassifierPooler(nn.Module):
self.classifier = classifier
self.pooler = pooler
if config.task == "score":
self.default_activation_function = \
get_cross_encoder_activation_function(config.hf_config)
elif config.task == "classify":
self.default_activation_function = nn.Sigmoid() \
if config.hf_config.num_labels == 1 else nn.Softmax()
else:
raise NotImplementedError(f"task={config.task!r} is not supported"
" with the classification pooler")
self.classification_act_fn = get_classification_activation_function(
config.hf_config)
self.cross_encoder_act_fn = get_cross_encoder_activation_function(
config.hf_config)
def _get_act_fn(self, use_cross_encoder: bool):
return (self.cross_encoder_act_fn
if use_cross_encoder else self.classification_act_fn)
def get_prompt_lens(
self,
@@ -446,8 +446,28 @@ class ClassifierPooler(nn.Module):
# apply classifier once on the full batch if possible
pooled_output = self.classifier(pooled_output)
# shape: (batch_size, num_labels)
scores = self.default_activation_function(pooled_output)
if isinstance(pooling_metadata, V0PoolingMetadata):
use_cross_encoder_list = [
pooling_param.use_cross_encoder
for _, pooling_param in pooling_metadata.seq_groups
]
else:
use_cross_encoder_list = [
pooling_param.use_cross_encoder
for pooling_param in pooling_metadata.pooling_params
]
# shape of scores: (batch_size, num_labels)
if all(use_cross_encoder == use_cross_encoder_list[0]
for use_cross_encoder in use_cross_encoder_list):
act_fn = self._get_act_fn(use_cross_encoder_list[0])
scores = act_fn(pooled_output)
else:
scores = torch.stack([
self._get_act_fn(use_cross_encoder)(vecs)
for use_cross_encoder, vecs in zip(use_cross_encoder_list,
pooled_output)
])
pooled_outputs = [PoolingSequenceGroupOutput(data) for data in scores]
return PoolerOutput(outputs=pooled_outputs)
-5
View File
@@ -25,8 +25,6 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
from vllm.model_executor.model_loader.weight_utils import default_weight_loader
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.sequence import IntermediateTensors, PoolerOutput
from vllm.transformers_utils.config import (
get_cross_encoder_activation_function)
from .interfaces import SupportsCrossEncoding, SupportsQuant, SupportsV0Only
from .utils import WeightsMapper, maybe_prefix
@@ -462,9 +460,6 @@ class BertForSequenceClassification(nn.Module, SupportsV0Only,
super().__init__()
config = vllm_config.model_config.hf_config
self.default_activation_function = \
get_cross_encoder_activation_function(config)
self.num_labels = config.num_labels
self.bert = BertModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "bert"),
-5
View File
@@ -18,8 +18,6 @@ from vllm.model_executor.models.bert import BertEmbeddingModel, BertModel
from vllm.model_executor.models.utils import WeightsMapper, maybe_prefix
from vllm.model_executor.pooling_metadata import PoolingMetadata
from vllm.sequence import IntermediateTensors, PoolerOutput
from vllm.transformers_utils.config import (
get_cross_encoder_activation_function)
from .bert_with_rope import BertWithRope, JinaRobertaModel
from .interfaces import SupportsCrossEncoding, SupportsV0Only
@@ -178,9 +176,6 @@ class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding,
super().__init__()
config = vllm_config.model_config.hf_config
self.default_activation_function = \
get_cross_encoder_activation_function(config)
self.num_labels = config.num_labels
self.roberta = BertModel(vllm_config=vllm_config,
prefix=maybe_prefix(prefix, "bert"),