diff --git a/vllm/model_executor/models/bert.py b/vllm/model_executor/models/bert.py index 7348ec9383..8f061b7284 100644 --- a/vllm/model_executor/models/bert.py +++ b/vllm/model_executor/models/bert.py @@ -461,6 +461,7 @@ class BertEmbeddingModel(nn.Module, SupportsQuant): """ is_pooling_model = True + embedding_class: type[nn.Module] = BertEmbedding def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -504,7 +505,7 @@ class BertEmbeddingModel(nn.Module, SupportsQuant): def _build_model(self, vllm_config: VllmConfig, prefix: str = "") -> BertModel: return BertModel( - vllm_config=vllm_config, prefix=prefix, embedding_class=BertEmbedding + vllm_config=vllm_config, prefix=prefix, embedding_class=self.embedding_class ) def _build_pooler(self, pooler_config: PoolerConfig) -> Pooler: @@ -781,6 +782,7 @@ class BertForSequenceClassification(nn.Module, SupportsCrossEncoding, SupportsQu """ is_pooling_model = True + embedding_class: type[nn.Module] = BertEmbedding def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -790,7 +792,7 @@ class BertForSequenceClassification(nn.Module, SupportsCrossEncoding, SupportsQu self.bert = BertPoolingModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "bert"), - embedding_class=BertEmbedding, + embedding_class=self.embedding_class, ) self.classifier = nn.Linear( config.hidden_size, @@ -840,6 +842,7 @@ class BertForSequenceClassification(nn.Module, SupportsCrossEncoding, SupportsQu @default_pooling_type(tok_pooling_type="ALL") class BertForTokenClassification(nn.Module): is_pooling_model = True + embedding_class: type[nn.Module] = BertEmbedding def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -849,7 +852,7 @@ class BertForTokenClassification(nn.Module): self.bert = BertModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "bert"), - embedding_class=BertEmbedding, + embedding_class=self.embedding_class, ) self.classifier = nn.Linear( config.hidden_size, config.num_labels, dtype=self.head_dtype @@ -927,13 +930,15 @@ class BertForMaskedLM(nn.Module): } ) + embedding_class: type[nn.Module] = BertEmbedding + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config self.bert = BertModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "bert"), - embedding_class=BertEmbedding, + embedding_class=self.embedding_class, ) self.mlm_head = BertMLMHead( hidden_size=config.hidden_size, diff --git a/vllm/model_executor/models/roberta.py b/vllm/model_executor/models/roberta.py index 8e141229f0..a694e7c3ce 100644 --- a/vllm/model_executor/models/roberta.py +++ b/vllm/model_executor/models/roberta.py @@ -117,6 +117,8 @@ class RobertaClassificationHead(nn.Module): class RobertaEmbeddingModel(BertEmbeddingModel): """A model that uses Roberta to provide embedding functionalities.""" + embedding_class: type[nn.Module] = RobertaEmbedding + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__(vllm_config=vllm_config, prefix=prefix) self.padding_idx: int = vllm_config.model_config.hf_config.pad_token_id @@ -143,7 +145,7 @@ class RobertaEmbeddingModel(BertEmbeddingModel): return BertModel( vllm_config=vllm_config, prefix=prefix, - embedding_class=RobertaEmbedding, + embedding_class=self.embedding_class, ) else: return JinaRobertaModel(vllm_config=vllm_config, prefix=prefix) @@ -293,6 +295,8 @@ class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding, Support } ) + embedding_class: type[nn.Module] = RobertaEmbedding + def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() config = vllm_config.model_config.hf_config @@ -303,7 +307,7 @@ class RobertaForSequenceClassification(nn.Module, SupportsCrossEncoding, Support self.roberta = BertModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "bert"), - embedding_class=RobertaEmbedding, + embedding_class=self.embedding_class, ) self.classifier = RobertaClassificationHead(vllm_config.model_config) @@ -355,6 +359,7 @@ class RobertaForTokenClassification(nn.Module): """ is_pooling_model = True + embedding_class: type[nn.Module] = RobertaEmbedding def __init__(self, *, vllm_config: VllmConfig, prefix: str = ""): super().__init__() @@ -364,7 +369,7 @@ class RobertaForTokenClassification(nn.Module): self.roberta = BertModel( vllm_config=vllm_config, prefix=maybe_prefix(prefix, "roberta"), - embedding_class=RobertaEmbedding, + embedding_class=self.embedding_class, ) self.classifier = nn.Linear( config.hidden_size, config.num_labels, dtype=self.head_dtype