Merge pull request #52 from lightonai/loading_logic

Loading logic rework

pylate/models/Dense.py

@@ -2,10 +2,12 @@
 import os
 
 import torch
+from safetensors import safe_open
 from safetensors.torch import load_model as load_safetensors_model
 from sentence_transformers.models import Dense as DenseSentenceTransformer
 from sentence_transformers.util import import_from_string
 from torch import nn
+from transformers.utils import cached_file
 
 __all__ = ["Dense"]
 
@@ -77,6 +79,61 @@ def from_sentence_transformers(dense: DenseSentenceTransformer) -> "Dense":
         model.load_state_dict(dense.state_dict())
         return model
 
+    @staticmethod
+    def from_stanford_weights(
+        model_name_or_path: str | os.PathLike,
+        cache_folder: str | os.PathLike | None = None,
+        revision: str | None = None,
+        local_files_only: bool | None = None,
+        token: str | bool | None = None,
+        use_auth_token: str | bool | None = None,
+    ) -> "Dense":
+        """Load the weight of the Dense layer using weights from a stanford-nlp checkpoint.
+
+        Parameters
+        ----------
+        model_name_or_path (`str` or `os.PathLike`):
+            This can be either:
+            - a string, the *model id* of a model repo on huggingface.co.
+            - a path to a *directory* potentially containing the file.
+        cache_folder (`str` or `os.PathLike`, *optional*):
+            Path to a directory in which a downloaded pretrained model configuration should be cached if the standard
+            cache should not be used.
+        token (`str` or *bool*, *optional*):
+            The token to use as HTTP bearer authorization for remote files. If `True`, will use the token generated
+            when running `huggingface-cli login` (stored in `~/.huggingface`).
+        revision (`str`, *optional*, defaults to `"main"`):
+            The specific model version to use. It can be a branch name, a tag name, or a commit id, since we use a
+            git-based system for storing models and other artifacts on huggingface.co, so `revision` can be any
+            identifier allowed by git.
+        local_files_only (`bool`, *optional*, defaults to `False`):
+            If `True`, will only try to load the tokenizer configuration from local files.
+        """
+        # Check if the model is locally available
+        if not (os.path.exists(os.path.join(model_name_or_path))):
+            # Else download the model/use the cached version
+            model_name_or_path = cached_file(
+                model_name_or_path,
+                filename="model.safetensors",
+                cache_dir=cache_folder,
+                revision=revision,
+                local_files_only=local_files_only,
+                token=token,
+                use_auth_token=use_auth_token,
+            )
+        with safe_open(model_name_or_path, framework="pt", device="cpu") as f:
+            state_dict = {"linear.weight": f.get_tensor("linear.weight")}
+
+        # Determine input and output dimensions
+        in_features = state_dict["linear.weight"].shape[1]
+        out_features = state_dict["linear.weight"].shape[0]
+
+        # Create Dense layer instance
+        model = Dense(in_features=in_features, out_features=out_features, bias=False)
+
+        model.load_state_dict(state_dict)
+        return model
+
     @staticmethod
     def load(input_path) -> "Dense":
         """Load a Dense layer."""

pylate/models/colbert.py

@@ -219,9 +219,6 @@ def __init__(
         config_kwargs: dict | None = None,
         model_card_data: Optional[SentenceTransformerModelCardData] = None,
     ) -> None:
-        model_kwargs = {} if model_kwargs is None else model_kwargs
-        model_kwargs["add_pooling_layer"] = False
-
         self.query_prefix = query_prefix
         self.document_prefix = document_prefix
         self.query_length = query_length
@@ -248,20 +245,35 @@ def __init__(
             config_kwargs=config_kwargs,
             model_card_data=model_card_data,
         )
-
         hidden_size = self[0].get_word_embedding_dimension()
 
         # Add a linear projection layer to the model in order to project the embeddings to the desired size.
         if len(self) < 2:
-            # Add a linear projection layer to the model in order to project the embeddings to the desired size
-            embedding_size = embedding_size or 128
+            # If the model is a stanford-nlp ColBERT, load the weights of the dense layer
+            if self[0].auto_model.config.architectures[0] == "HF_ColBERT":
+                self.append(
+                    Dense.from_stanford_weights(
+                        model_name_or_path,
+                        cache_folder,
+                        revision,
+                        local_files_only,
+                        token,
+                        use_auth_token,
+                    )
+                )
+                logger.warning("Loaded the ColBERT model from Stanford NLP.")
+            else:
+                # Add a linear projection layer to the model in order to project the embeddings to the desired size
+                embedding_size = embedding_size or 128
 
-            logger.warning(
-                f"The checkpoint does not contain a linear projection layer. Adding one with output dimensions ({hidden_size}, {embedding_size})."
-            )
-            self.append(
-                Dense(in_features=hidden_size, out_features=embedding_size, bias=bias)
-            )
+                logger.warning(
+                    f"The checkpoint does not contain a linear projection layer. Adding one with output dimensions ({hidden_size}, {embedding_size})."
+                )
+                self.append(
+                    Dense(
+                        in_features=hidden_size, out_features=embedding_size, bias=bias
+                    )
+                )
 
         elif (
             embedding_size is not None
@@ -282,6 +294,9 @@ def __init__(
         else:
             logger.warning("Pylate model loaded successfully.")
 
+        if model_kwargs is not None and "torch_dtype" in model_kwargs:
+            self[1].to(model_kwargs["torch_dtype"])
+
         self.to(device)
         self.is_hpu_graph_enabled = False
 

setup.py

@@ -6,7 +6,7 @@
     long_description = fh.read()
 
 base_packages = [
-    "sentence-transformers >= 3.0.1",
+    "sentence-transformers == 3.0.1",
     "datasets >= 2.20.0",
     "accelerate >= 0.31.0",
     "voyager >= 2.0.9",