[Model] Enable optional prefix when loading embedding models (vllm-pr…

…oject#10639) Signed-off-by: DarkLight1337 <[email protected]>
neuralmagic · Nov 25, 2024 · cf73f0c · cf73f0c
1 parent b1d9205
commit cf73f0c
Show file tree

Hide file tree

Showing 5 changed files with 20 additions and 13 deletions.
diff --git a/vllm/model_executor/models/bert.py b/vllm/model_executor/models/bert.py
@@ -14,18 +14,17 @@
                                                RowParallelLinear)
 from vllm.model_executor.layers.pooler import (CrossEncodingPooler, Pooler,
                                                PoolingType)
-from vllm.model_executor.layers.quantization.base_config import (
-    QuantizationConfig)
+from vllm.model_executor.layers.quantization import 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 @@ def pooler(
         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,

diff --git a/vllm/model_executor/models/gemma2.py b/vllm/model_executor/models/gemma2.py
@@ -42,7 +42,7 @@
 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 @@ def pooler(
         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)
diff --git a/vllm/model_executor/models/llama.py b/vllm/model_executor/models/llama.py
@@ -53,7 +53,8 @@
 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 @@ def pooler(
         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:

diff --git a/vllm/model_executor/models/qwen2.py b/vllm/model_executor/models/qwen2.py
@@ -50,7 +50,8 @@
 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 @@ def pooler(
     ) -> Optional[PoolerOutput]:
         return self._pooler(hidden_states, pooling_metadata)
 
-    def load_weights(self, weights: Iterable[Tuple[str,
-                                                   torch.Tensor]]) -> Set[str]:
-        loader = AutoWeightsLoader(self,
-                                   ignore_unexpected_prefixes=["lm_head."])
-        return loader.load_weights(weights)
+    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)
diff --git a/vllm/model_executor/models/roberta.py b/vllm/model_executor/models/roberta.py
@@ -11,13 +11,14 @@
     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):