[Model] Enable optional prefix when loading embedding models (#10639)

Signed-off-by: DarkLight1337 <tlleungac@connect.ust.hk>

vllm/model_executor/models/bert.py

@@ -14,18 +14,17 @@ from vllm.model_executor.layers.linear import (ColumnParallelLinear,
                                                RowParallelLinear)
 from vllm.model_executor.layers.pooler import (CrossEncodingPooler, Pooler,
                                                PoolingType)
-from vllm.model_executor.layers.quantization.base_config import (
+from vllm.model_executor.layers.quantization import QuantizationConfig
-    QuantizationConfig)
 from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
-from vllm.model_executor.models.interfaces import SupportsCrossEncoding
 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 .utils import maybe_prefix
+from .interfaces import SupportsCrossEncoding
+from .utils import WeightsMapper, maybe_prefix
 
 
 class BertEmbedding(nn.Module):
@@ -442,6 +441,8 @@ class BertEmbeddingModel(nn.Module):
         return self._pooler(hidden_states, pooling_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
+        weights = hf_to_vllm_mapper.apply(weights)
         self.model.load_weights(weights)
 
     def _build_model(self,

vllm/model_executor/models/gemma2.py

@@ -42,7 +42,7 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors, PoolerOutput
 
 from .interfaces import SupportsLoRA, SupportsPP
-from .utils import (AutoWeightsLoader, extract_layer_index,
+from .utils import (AutoWeightsLoader, WeightsMapper, extract_layer_index,
                     is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
@@ -511,4 +511,6 @@ class Gemma2EmbeddingModel(nn.Module, SupportsPP):
         return self._pooler(hidden_states, pooling_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
+        weights = hf_to_vllm_mapper.apply(weights)
         self.model.load_weights(weights)

vllm/model_executor/models/llama.py

@@ -53,7 +53,8 @@ from vllm.platforms import current_platform
 from vllm.sequence import IntermediateTensors, PoolerOutput
 
 from .interfaces import SupportsLoRA, SupportsPP
-from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
+                    is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -689,6 +690,8 @@ class LlamaEmbeddingModel(nn.Module, SupportsLoRA, SupportsPP):
         return self._pooler(hidden_states, pooling_metadata)
 
     def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
+        hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
+        weights = hf_to_vllm_mapper.apply(weights)
         self.model.load_weights(weights)
 
     def load_kv_cache_scales(self, quantization_param_path: str) -> None:

vllm/model_executor/models/qwen2.py

@@ -50,7 +50,8 @@ from vllm.model_executor.sampling_metadata import SamplingMetadata
 from vllm.sequence import IntermediateTensors, PoolerOutput
 
 from .interfaces import SupportsLoRA, SupportsPP
-from .utils import (AutoWeightsLoader, PPMissingLayer, is_pp_missing_parameter,
+from .utils import (AutoWeightsLoader, PPMissingLayer, WeightsMapper,
+                    is_pp_missing_parameter,
                     make_empty_intermediate_tensors_factory, make_layers,
                     maybe_prefix)
 
@@ -585,8 +586,7 @@ class Qwen2EmbeddingModel(nn.Module, SupportsLoRA, SupportsPP):
     ) -> Optional[PoolerOutput]:
         return self._pooler(hidden_states, pooling_metadata)
 
-    def load_weights(self, weights: Iterable[Tuple[str,
+    def load_weights(self, weights: Iterable[Tuple[str, torch.Tensor]]):
-                                                   torch.Tensor]]) -> Set[str]:
+        hf_to_vllm_mapper = WeightsMapper(orig_to_new_prefix={"model.": ""})
-        loader = AutoWeightsLoader(self,
+        weights = hf_to_vllm_mapper.apply(weights)
-                                   ignore_unexpected_prefixes=["lm_head."])
+        self.model.load_weights(weights)
-        return loader.load_weights(weights)

vllm/model_executor/models/roberta.py

@@ -11,13 +11,14 @@ from vllm.model_executor.layers.vocab_parallel_embedding import (
     VocabParallelEmbedding)
 from vllm.model_executor.model_loader.weight_utils import default_weight_loader
 from vllm.model_executor.models.bert import BertEmbeddingModel, BertModel
-from vllm.model_executor.models.interfaces import SupportsCrossEncoding
 from vllm.model_executor.models.utils import 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 .interfaces import SupportsCrossEncoding
+
 
 class RobertaEmbedding(nn.Module):