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

[Bugfix] Replace PoolingParams.normalize with use_activation (#32243)

Browse Source 
Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>
This commit is contained in:
Cyrus Leung
2026-01-13 18:45:42 +08:00
committed by GitHub
parent eb28e8068d
commit 232214b2ae
21 changed files with 68 additions and 70 deletions
Download Patch File Download Diff File Expand all files Collapse all files

2
vllm/model_executor/layers/pooler/seqwise/heads.py
View File

@@ -83,7 +83,7 @@ class EmbeddingPoolerHead(SequencePoolerHead):
# for normalize
if self.activation is not None:
flags = [p.normalize for p in pooling_params]
flags = [p.use_activation for p in pooling_params]
if len(set(flags)) == 1:
if flags[0]:
pooled_data = self.activation(pooled_data)

4
vllm/model_executor/layers/pooler/seqwise/poolers.py
View File

@@ -95,8 +95,8 @@ def pooler_for_embed(pooler_config: PoolerConfig):
vllm_config = get_current_vllm_config()
model_config = vllm_config.model_config
head = EmbeddingPoolerHead(
projector=_load_st_projector(model_config),
head_dtype=model_config.head_dtype,
projector=_load_st_projector(model_config),
activation=PoolerNormalize(),
)
@@ -116,9 +116,9 @@ def pooler_for_classify(
vllm_config = get_current_vllm_config()
model_config = vllm_config.model_config
head = ClassifierPoolerHead(
head_dtype=model_config.head_dtype,
classifier=classifier,
logit_bias=model_config.pooler_config.logit_bias,
head_dtype=model_config.head_dtype,
activation=resolve_classifier_act_fn(
model_config, static_num_labels=True, act_fn=act_fn
),

6
vllm/model_executor/layers/pooler/tokwise/heads.py
View File

@@ -44,14 +44,14 @@ class TokenPoolerHead(nn.Module, ABC):
class TokenEmbeddingPoolerHead(TokenPoolerHead):
def __init__(
self,
projector: ProjectorFn | None = None,
head_dtype: torch.dtype | str | None = None,
projector: ProjectorFn | None = None,
activation: ActivationFn | None = None,
) -> None:
super().__init__()
self.projector = projector
self.head_dtype = head_dtype
self.projector = projector
self.activation = activation
def get_supported_tasks(self) -> Set[PoolingTask]:
@@ -79,7 +79,7 @@ class TokenEmbeddingPoolerHead(TokenPoolerHead):
pooled_data = pooled_data[..., : pooling_param.dimensions]
# for normalize
if self.activation is not None and pooling_param.normalize:
if self.activation is not None and pooling_param.use_activation:
pooled_data = self.activation(pooled_data)
# pooled_data shape: [n_tokens, embedding_dimension]

4
vllm/model_executor/layers/pooler/tokwise/poolers.py
View File

@@ -95,8 +95,8 @@ def pooler_for_token_embed(pooler_config: PoolerConfig):
vllm_config = get_current_vllm_config()
model_config = vllm_config.model_config
head = TokenEmbeddingPoolerHead(
projector=_load_st_projector(model_config),
head_dtype=model_config.head_dtype,
projector=_load_st_projector(model_config),
activation=PoolerNormalize(),
)
@@ -116,9 +116,9 @@ def pooler_for_token_classify(
vllm_config = get_current_vllm_config()
model_config = vllm_config.model_config
head = TokenClassifierPoolerHead(
head_dtype=model_config.head_dtype,
classifier=classifier,
logit_bias=model_config.pooler_config.logit_bias,
head_dtype=model_config.head_dtype,
activation=resolve_classifier_act_fn(
model_config, static_num_labels=False, act_fn=act_fn
),