Safetensors tf (#19900)

* Wip

* Add safetensors support for TensorFlow

* First tests

* Add final test for now

* Retrigger CI like this

* Update src/transformers/modeling_tf_utils.py

Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

Co-authored-by: Lysandre Debut <lysandre.debut@reseau.eseo.fr>

src/transformers/modeling_tf_pytorch_utils.py

@@ -21,7 +21,7 @@ import re
 
 import numpy
 
-from .utils import ExplicitEnum, expand_dims, is_numpy_array, is_torch_tensor, logging, reshape, squeeze
+from .utils import ExplicitEnum, expand_dims, is_numpy_array, is_torch_tensor, logging, reshape, squeeze, tensor_size
 from .utils import transpose as transpose_func
 
 
@@ -273,7 +273,7 @@ def load_pytorch_state_dict_in_tf2_model(
 
         array = apply_transpose(transpose, pt_state_dict[name], symbolic_weight.shape)
 
-        tf_loaded_numel += array.size
+        tf_loaded_numel += tensor_size(array)
 
         weight_value_tuples.append((symbolic_weight, array))
         all_pytorch_weights.discard(name)

src/transformers/modeling_tf_utils.py

@@ -47,6 +47,8 @@ from .generation_tf_utils import TFGenerationMixin
 from .tf_utils import shape_list
 from .utils import (
     DUMMY_INPUTS,
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
     TF2_WEIGHTS_INDEX_NAME,
     TF2_WEIGHTS_NAME,
     WEIGHTS_INDEX_NAME,
@@ -59,12 +61,18 @@ from .utils import (
     has_file,
     is_offline_mode,
     is_remote_url,
+    is_safetensors_available,
     logging,
     requires_backends,
     working_or_temp_dir,
 )
 
 
+if is_safetensors_available():
+    from safetensors import safe_open
+    from safetensors.tensorflow import load_file as safe_load_file
+    from safetensors.tensorflow import save_file as safe_save_file
+
 if TYPE_CHECKING:
     from . import PreTrainedTokenizerBase
 
@@ -612,6 +620,14 @@ def dtype_byte_size(dtype):
     return bit_size // 8
 
 
+def format_weight_name(name, _prefix=None):
+    if "model." not in name and len(name.split("/")) > 1:
+        name = "/".join(name.split("/")[1:])
+    if _prefix is not None:
+        name = _prefix + "/" + name
+    return name
+
+
 def tf_shard_checkpoint(weights, max_shard_size="10GB"):
     """
     Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
@@ -849,6 +865,17 @@ def load_tf_weights(model, resolved_archive_file, ignore_mismatched_sizes=False,
         Three lists, one for the missing layers, another one for the unexpected layers, and a last one for the
         mismatched layers.
     """
+    if resolved_archive_file.endswith(".safetensors"):
+        load_function = load_tf_weights_from_safetensors
+    else:
+        load_function = load_tf_weights_from_h5
+
+    return load_function(
+        model, resolved_archive_file, ignore_mismatched_sizes=ignore_mismatched_sizes, _prefix=_prefix
+    )
+
+
+def load_tf_weights_from_h5(model, resolved_archive_file, ignore_mismatched_sizes=False, _prefix=None):
     missing_layers = []
     unexpected_layers = []
     mismatched_layers = []
@@ -952,6 +979,47 @@ def load_tf_weights(model, resolved_archive_file, ignore_mismatched_sizes=False,
     return missing_layers, unexpected_layers, mismatched_layers
 
 
+def load_tf_weights_from_safetensors(model, resolved_archive_file, ignore_mismatched_sizes=False, _prefix=None):
+    # Read the safetensors file
+    state_dict = safe_load_file(resolved_archive_file)
+
+    weight_value_tuples = []
+    mismatched_layers = []
+
+    weight_names = [format_weight_name(w.name, _prefix=_prefix) for w in model.weights]
+    loaded_weight_names = list(state_dict.keys())
+
+    # Find the missing layers from the high level list of layers
+    missing_layers = list(set(weight_names) - set(loaded_weight_names))
+    # Find the unexpected layers from the high level list of layers
+    unexpected_layers = list(set(loaded_weight_names) - set(weight_names))
+
+    weight_value_tuples = []
+    for weight in model.weights:
+        weight_name = format_weight_name(weight.name, _prefix=_prefix)
+        if weight_name in state_dict:
+            weight_value = state_dict[weight_name]
+            # Check if the shape of the current weight and the one from the H5 file are different
+            if K.int_shape(weight) != weight_value.shape:
+                # If yes we reshape the weight from the H5 file accordingly to the current weight
+                # If the two shapes are not compatible we raise an issue
+                try:
+                    weight_value = tf.reshape(weight_value, K.int_shape(weight))
+                except ValueError as e:
+                    if ignore_mismatched_sizes:
+                        mismatched_layers.append((weight_name, weight_value.shape, K.int_shape(weight)))
+                        continue
+                    else:
+                        raise e
+
+            weight_value_tuples.append((weight, weight_value))
+
+    # Load all the weights
+    K.batch_set_value(weight_value_tuples)
+
+    return missing_layers, unexpected_layers, mismatched_layers
+
+
 def init_copy_embeddings(old_embeddings, new_num_tokens):
     r"""
     This function aims to reduce the embeddings in case new_num_tokens < old_num_tokens or to pad with -1 in case
@@ -2118,6 +2186,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
         signatures=None,
         max_shard_size: Union[int, str] = "10GB",
         create_pr: bool = False,
+        safe_serialization: bool = False,
         **kwargs
     ):
         """
@@ -2152,6 +2221,8 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
 
             create_pr (`bool`, *optional*, defaults to `False`):
                 Whether or not to create a PR with the uploaded files or directly commit.
+            safe_serialization (`bool`, *optional*, defaults to `False`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way (that uses `pickle`).
 
             kwargs:
                 Additional key word arguments passed along to the [`~utils.PushToHubMixin.push_to_hub`] method.
@@ -2186,7 +2257,8 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
         self.config.save_pretrained(save_directory)
 
         # If we save using the predefined names, we can load using `from_pretrained`
-        output_model_file = os.path.join(save_directory, TF2_WEIGHTS_NAME)
+        weights_name = SAFE_WEIGHTS_NAME if safe_serialization else TF2_WEIGHTS_NAME
+        output_model_file = os.path.join(save_directory, weights_name)
 
         shards, index = tf_shard_checkpoint(self.weights, max_shard_size)
 
@@ -2195,15 +2267,20 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
             full_filename = os.path.join(save_directory, filename)
             # If we have a shard file that is not going to be replaced, we delete it, but only from the main process
             # in distributed settings to avoid race conditions.
+            weights_no_suffix = weights_name.replace(".bin", "").replace(".safetensors", "")
             if (
-                filename.startswith(TF2_WEIGHTS_NAME[:-4])
+                filename.startswith(weights_no_suffix)
                 and os.path.isfile(full_filename)
                 and filename not in shards.keys()
             ):
                 os.remove(full_filename)
 
         if index is None:
-            self.save_weights(output_model_file)
+            if safe_serialization:
+                state_dict = {format_weight_name(w.name): w.value() for w in self.weights}
+                safe_save_file(state_dict, output_model_file, metadata={"format": "tf"})
+            else:
+                self.save_weights(output_model_file)
             logger.info(f"Model weights saved in {output_model_file}")
         else:
             save_index_file = os.path.join(save_directory, TF2_WEIGHTS_INDEX_NAME)
@@ -2427,6 +2504,18 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
                     # Load from a sharded PyTorch checkpoint
                     archive_file = os.path.join(pretrained_model_name_or_path, WEIGHTS_INDEX_NAME)
                     is_sharded = True
+                elif is_safetensors_available() and os.path.isfile(
+                    os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_NAME)
+                ):
+                    # Load from a safetensors checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_NAME)
+                elif is_safetensors_available() and os.path.isfile(
+                    os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_INDEX_NAME)
+                ):
+                    # Load from a sharded safetensors checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, SAFE_WEIGHTS_INDEX_NAME)
+                    is_sharded = True
+                    raise NotImplementedError("Support for sharded checkpoints using safetensors is coming soon!")
                 elif os.path.isfile(os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)):
                     # Load from a TF 2.0 checkpoint
                     archive_file = os.path.join(pretrained_model_name_or_path, TF2_WEIGHTS_NAME)
@@ -2457,7 +2546,12 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
                 resolved_archive_file = download_url(pretrained_model_name_or_path)
             else:
                 # set correct filename
-                filename = WEIGHTS_NAME if from_pt else TF2_WEIGHTS_NAME
+                if from_pt:
+                    filename = WEIGHTS_NAME
+                elif is_safetensors_available():
+                    filename = SAFE_WEIGHTS_NAME
+                else:
+                    filename = TF2_WEIGHTS_NAME
 
                 try:
                     # Load from URL or cache if already cached
@@ -2476,8 +2570,24 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
                     )
                     resolved_archive_file = cached_file(pretrained_model_name_or_path, filename, **cached_file_kwargs)
 
-                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an expection but a None
+                    # Since we set _raise_exceptions_for_missing_entries=False, we don't get an exception but a None
                     # result when internet is up, the repo and revision exist, but the file does not.
+                    if resolved_archive_file is None and filename == SAFE_WEIGHTS_NAME:
+                        # Maybe the checkpoint is sharded, we try to grab the index name in this case.
+                        resolved_archive_file = cached_file(
+                            pretrained_model_name_or_path, SAFE_WEIGHTS_INDEX_NAME, **cached_file_kwargs
+                        )
+                        if resolved_archive_file is not None:
+                            is_sharded = True
+                            raise NotImplementedError(
+                                "Support for sharded checkpoints using safetensors is coming soon!"
+                            )
+                        else:
+                            # This repo has no safetensors file of any kind, we switch to TensorFlow.
+                            filename = TF2_WEIGHTS_NAME
+                            resolved_archive_file = cached_file(
+                                pretrained_model_name_or_path, TF2_WEIGHTS_NAME, **cached_file_kwargs
+                            )
                     if resolved_archive_file is None and filename == TF2_WEIGHTS_NAME:
                         # Maybe the checkpoint is sharded, we try to grab the index name in this case.
                         resolved_archive_file = cached_file(
@@ -2521,6 +2631,7 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
             if is_local:
                 logger.info(f"loading weights file {archive_file}")
                 resolved_archive_file = archive_file
+                filename = resolved_archive_file.split(os.path.sep)[-1]
             else:
                 logger.info(f"loading weights file {filename} from cache at {resolved_archive_file}")
         else:
@@ -2543,6 +2654,17 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
                 _commit_hash=commit_hash,
             )
 
+        safetensors_from_pt = False
+        if filename == SAFE_WEIGHTS_NAME:
+            with safe_open(resolved_archive_file, framework="tf") as f:
+                safetensors_metadata = f.metadata()
+            if safetensors_metadata is None or safetensors_metadata.get("format") not in ["pt", "tf", "flax"]:
+                raise OSError(
+                    f"The safetensors archive passed at {resolved_archive_file} does not contain the valid metadata."
+                    " Make sure you save your model with the `save_pretrained` method."
+                )
+            safetensors_from_pt = safetensors_metadata.get("format") == "pt"
+
         config.name_or_path = pretrained_model_name_or_path
 
         # composed models, *e.g.* TFRag, require special treatment when it comes to loading
@@ -2560,6 +2682,14 @@ class TFPreTrainedModel(tf.keras.Model, TFModelUtilsMixin, TFGenerationMixin, Pu
             return load_pytorch_checkpoint_in_tf2_model(
                 model, resolved_archive_file, allow_missing_keys=True, output_loading_info=output_loading_info
             )
+        elif safetensors_from_pt:
+            from .modeling_tf_pytorch_utils import load_pytorch_state_dict_in_tf2_model
+
+            state_dict = safe_load_file(resolved_archive_file)
+            # Load from a PyTorch checkpoint
+            return load_pytorch_state_dict_in_tf2_model(
+                model, state_dict, allow_missing_keys=True, output_loading_info=output_loading_info
+            )
 
         # we might need to extend the variable scope for composite models
         if load_weight_prefix is not None:

src/transformers/utils/__init__.py

@@ -49,6 +49,7 @@ from .generic import (
     is_torch_tensor,
     reshape,
     squeeze,
+    tensor_size,
     to_numpy,
     to_py_obj,
     transpose,

src/transformers/utils/generic.py

@@ -445,3 +445,19 @@ def expand_dims(array, axis):
         return jnp.expand_dims(array, axis=axis)
     else:
         raise ValueError(f"Type not supported for expand_dims: {type(array)}.")
+
+
+def tensor_size(array):
+    """
+    Framework-agnostic version of `numpy.size` that will work on torch/TensorFlow/Jax tensors as well as NumPy arrays.
+    """
+    if is_numpy_array(array):
+        return np.size(array)
+    elif is_torch_tensor(array):
+        return array.numel()
+    elif is_tf_tensor(array):
+        return tf.size(array)
+    elif is_jax_tensor(array):
+        return array.size
+    else:
+        raise ValueError(f"Type not supported for expand_dims: {type(array)}.")

tests/test_modeling_tf_common.py

@@ -43,13 +43,14 @@ from transformers.testing_utils import (  # noqa: F401
     _tf_gpu_memory_limit,
     is_pt_tf_cross_test,
     is_staging_test,
+    require_safetensors,
     require_tf,
     require_tf2onnx,
     slow,
     tooslow,
     torch_device,
 )
-from transformers.utils import logging
+from transformers.utils import SAFE_WEIGHTS_NAME, TF2_WEIGHTS_INDEX_NAME, TF2_WEIGHTS_NAME, logging
 from transformers.utils.generic import ModelOutput
 
 
@@ -94,12 +95,7 @@ if is_tf_available():
         TFSampleDecoderOnlyOutput,
         TFSampleEncoderDecoderOutput,
     )
-    from transformers.modeling_tf_utils import (
-        TF2_WEIGHTS_INDEX_NAME,
-        TF2_WEIGHTS_NAME,
-        tf_shard_checkpoint,
-        unpack_inputs,
-    )
+    from transformers.modeling_tf_utils import tf_shard_checkpoint, unpack_inputs
     from transformers.tf_utils import stable_softmax
 
     if _tf_gpu_memory_limit is not None:
@@ -119,6 +115,8 @@ if is_tf_available():
 if is_torch_available():
     import torch
 
+    from transformers import BertModel
+
 
 def _config_zero_init(config):
     configs_no_init = copy.deepcopy(config)
@@ -2168,7 +2166,7 @@ class UtilsFunctionsTest(unittest.TestCase):
             )
 
     @slow
-    def test_special_layer_name_shardind(self):
+    def test_special_layer_name_sharding(self):
         retriever = RagRetriever.from_pretrained("facebook/rag-token-nq", index_name="exact", use_dummy_dataset=True)
         model = TFRagModel.from_pretrained("facebook/rag-token-nq", retriever=retriever)
 
@@ -2268,6 +2266,54 @@ class UtilsFunctionsTest(unittest.TestCase):
             self.assertTrue("custom_signature_1" in list(model_loaded.signatures.keys()))
             self.assertTrue("custom_signature_2" in list(model_loaded.signatures.keys()))
 
+    @require_safetensors
+    def test_safetensors_save_and_load(self):
+        model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.save_pretrained(tmp_dir, safe_serialization=True)
+            # No tf_model.h5 file, only a model.safetensors
+            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
+            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, TF2_WEIGHTS_NAME)))
+
+            new_model = TFBertModel.from_pretrained(tmp_dir)
+
+            # Check models are equal
+            for p1, p2 in zip(model.weights, new_model.weights):
+                self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
+
+    @is_pt_tf_cross_test
+    def test_safetensors_save_and_load_pt_to_tf(self):
+        model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+        pt_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            pt_model.save_pretrained(tmp_dir, safe_serialization=True)
+            # Check we have a model.safetensors file
+            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
+
+            new_model = TFBertModel.from_pretrained(tmp_dir)
+
+            # Check models are equal
+            for p1, p2 in zip(model.weights, new_model.weights):
+                self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
+
+    @require_safetensors
+    def test_safetensors_load_from_hub(self):
+        tf_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+
+        # Can load from the TF-formatted checkpoint
+        safetensors_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors-tf")
+
+        # Check models are equal
+        for p1, p2 in zip(safetensors_model.weights, tf_model.weights):
+            self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
+
+        # Can load from the PyTorch-formatted checkpoint
+        safetensors_model = TFBertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors")
+
+        # Check models are equal
+        for p1, p2 in zip(safetensors_model.weights, tf_model.weights):
+            self.assertTrue(np.allclose(p1.numpy(), p2.numpy()))
+
 
 @require_tf
 @is_staging_test