Poc to use safetensors (#19175)

* Poc to use safetensors

* Typo

* Final version

* Add tests

* Save with the right name!

* Update tests/test_modeling_common.py

Co-authored-by: Julien Chaumond <julien@huggingface.co>

* Support for sharded checkpoints

* Test from Hub part 1

* Test from hub part 2

* Fix regular checkpoint sharding

* Bump for fixes

Co-authored-by: Julien Chaumond <julien@huggingface.co>

setup.py

@@ -148,6 +148,7 @@ _deps = [
     "rouge-score!=0.0.7,!=0.0.8,!=0.1,!=0.1.1",
     "sacrebleu>=1.4.12,<2.0.0",
     "sacremoses",
+    "safetensors>=0.2.1",
     "sagemaker>=2.31.0",
     "scikit-learn",
     "sentencepiece>=0.1.91,!=0.1.92",
@@ -300,6 +301,7 @@ extras["testing"] = (
         "protobuf",  # Can be removed once we can unpin protobuf
         "sacremoses",
         "rjieba",
+        "safetensors",
     )
     + extras["retrieval"]
     + extras["modelcreation"]

src/transformers/__init__.py

@@ -478,6 +478,7 @@ _import_structure = {
         "is_psutil_available",
         "is_py3nvml_available",
         "is_pyctcdecode_available",
+        "is_safetensors_available",
         "is_scipy_available",
         "is_sentencepiece_available",
         "is_sklearn_available",
@@ -3417,6 +3418,7 @@ if TYPE_CHECKING:
         is_psutil_available,
         is_py3nvml_available,
         is_pyctcdecode_available,
+        is_safetensors_available,
         is_scipy_available,
         is_sentencepiece_available,
         is_sklearn_available,

src/transformers/dependency_versions_table.py

@@ -54,6 +54,7 @@ deps = {
     "rouge-score": "rouge-score!=0.0.7,!=0.0.8,!=0.1,!=0.1.1",
     "sacrebleu": "sacrebleu>=1.4.12,<2.0.0",
     "sacremoses": "sacremoses",
+    "safetensors": "safetensors>=0.2.1",
     "sagemaker": "sagemaker>=2.31.0",
     "scikit-learn": "scikit-learn",
     "sentencepiece": "sentencepiece>=0.1.91,!=0.1.92",

src/transformers/modeling_utils.py

@@ -50,6 +50,8 @@ from .pytorch_utils import (  # noqa: F401
 from .utils import (
     DUMMY_INPUTS,
     FLAX_WEIGHTS_NAME,
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
     TF2_WEIGHTS_NAME,
     TF_WEIGHTS_NAME,
     WEIGHTS_INDEX_NAME,
@@ -65,6 +67,7 @@ from .utils import (
     is_bitsandbytes_available,
     is_offline_mode,
     is_remote_url,
+    is_safetensors_available,
     logging,
     replace_return_docstrings,
 )
@@ -86,6 +89,10 @@ if is_accelerate_available():
     else:
         get_balanced_memory = None
 
+if is_safetensors_available():
+    from safetensors import safe_open
+    from safetensors.torch import load_file as safe_load_file
+    from safetensors.torch import save_file as safe_save_file
 
 logger = logging.get_logger(__name__)
 
@@ -241,7 +248,9 @@ def dtype_byte_size(dtype):
     return bit_size // 8
 
 
-def shard_checkpoint(state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB"):
+def shard_checkpoint(
+    state_dict: Dict[str, torch.Tensor], max_shard_size: Union[int, str] = "10GB", weights_name: str = WEIGHTS_NAME
+):
     """
     Splits a model state dictionary in sub-checkpoints so that the final size of each sub-checkpoint does not exceed a
     given size.
@@ -263,6 +272,8 @@ def shard_checkpoint(state_dict: Dict[str, torch.Tensor], max_shard_size: Union[
         max_shard_size (`int` or `str`, *optional*, defaults to `"10GB"`):
             The maximum size of each sub-checkpoint. If expressed as a string, needs to be digits followed by a unit
             (like `"5MB"`).
+        weights_name (`str`, *optional*, defaults to `"pytorch_model.bin"`):
+            The name of the model save file.
     """
     max_shard_size = convert_file_size_to_int(max_shard_size)
 
@@ -289,13 +300,16 @@ def shard_checkpoint(state_dict: Dict[str, torch.Tensor], max_shard_size: Union[
 
     # If we only have one shard, we return it
     if len(sharded_state_dicts) == 1:
-        return {WEIGHTS_NAME: sharded_state_dicts[0]}, None
+        return {weights_name: sharded_state_dicts[0]}, None
 
     # Otherwise, let's build the index
     weight_map = {}
     shards = {}
     for idx, shard in enumerate(sharded_state_dicts):
-        shard_file = WEIGHTS_NAME.replace(".bin", f"-{idx+1:05d}-of-{len(sharded_state_dicts):05d}.bin")
+        shard_file = weights_name.replace(".bin", f"-{idx+1:05d}-of-{len(sharded_state_dicts):05d}.bin")
+        shard_file = shard_file.replace(
+            ".safetensors", f"-{idx + 1:05d}-of-{len(sharded_state_dicts):05d}.safetensors"
+        )
         shards[shard_file] = shard
         for key in shard.keys():
             weight_map[key] = shard_file
@@ -367,6 +381,20 @@ def load_state_dict(checkpoint_file: Union[str, os.PathLike]):
     """
     Reads a PyTorch checkpoint file, returning properly formatted errors if they arise.
     """
+    if checkpoint_file.endswith(".safetensors") and is_safetensors_available():
+        # Check format of the archive
+        with safe_open(checkpoint_file, framework="pt") as f:
+            metadata = f.metadata()
+        if metadata.get("format") not in ["pt", "tf", "flax"]:
+            raise OSError(
+                f"The safetensors archive passed at {checkpoint_file} does not contain the valid metadata. Make sure "
+                "you save your model with the `save_pretrained` method."
+            )
+        elif metadata["format"] != "pt":
+            raise NotImplementedError(
+                f"Conversion from a {metadata['format']} safetensors archive to PyTorch is not implemented yet."
+            )
+        return safe_load_file(checkpoint_file)
     try:
         return torch.load(checkpoint_file, map_location="cpu")
     except Exception as e:
@@ -1468,6 +1496,7 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
         save_function: Callable = torch.save,
         push_to_hub: bool = False,
         max_shard_size: Union[int, str] = "10GB",
+        safe_serialization: bool = False,
         **kwargs,
     ):
         """
@@ -1503,6 +1532,9 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
                 </Tip>
 
+            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.
         """
@@ -1511,6 +1543,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 "`save_config` is deprecated and will be removed in v5 of Transformers. Use `is_main_process` instead."
             )
             is_main_process = kwargs.pop("save_config")
+        if safe_serialization and not is_safetensors_available():
+            raise ImportError("`safe_serialization` requires the `safetensors library: `pip install safetensors`.")
 
         if os.path.isfile(save_directory):
             logger.error(f"Provided path ({save_directory}) should be a directory, not a file")
@@ -1560,15 +1594,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     del state_dict[ignore_key]
 
         # Shard the model if it is too big.
-        shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size)
+        weights_name = SAFE_WEIGHTS_NAME if safe_serialization else WEIGHTS_NAME
+        shards, index = shard_checkpoint(state_dict, max_shard_size=max_shard_size, weights_name=weights_name)
 
         # Clean the folder from a previous save
         for filename in os.listdir(save_directory):
             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(WEIGHTS_NAME[:-4])
+                filename.startswith(weights_no_suffix)
                 and os.path.isfile(full_filename)
                 and filename not in shards.keys()
                 and is_main_process
@@ -1577,12 +1613,18 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
 
         # Save the model
         for shard_file, shard in shards.items():
-            save_function(shard, os.path.join(save_directory, shard_file))
+            if safe_serialization:
+                # At some point we will need to deal better with save_function (used for TPU and other distributed
+                # joyfulness), but for now this enough.
+                safe_save_file(shard, os.path.join(save_directory, shard_file), metadata={"format": "pt"})
+            else:
+                save_function(shard, os.path.join(save_directory, shard_file))
 
         if index is None:
             logger.info(f"Model weights saved in {os.path.join(save_directory, WEIGHTS_NAME)}")
         else:
-            save_index_file = os.path.join(save_directory, WEIGHTS_INDEX_NAME)
+            save_index_file = SAFE_WEIGHTS_INDEX_NAME if safe_serialization else WEIGHTS_INDEX_NAME
+            save_index_file = os.path.join(save_directory, save_index_file)
             # Save the index as well
             with open(save_index_file, "w", encoding="utf-8") as f:
                 content = json.dumps(index, indent=2, sort_keys=True) + "\n"
@@ -1966,6 +2008,17 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                 ):
                     # Load from a Flax checkpoint in priority if from_flax
                     archive_file = os.path.join(pretrained_model_name_or_path, subfolder, FLAX_WEIGHTS_NAME)
+                elif is_safetensors_available() and os.path.isfile(
+                    os.path.join(pretrained_model_name_or_path, subfolder, SAFE_WEIGHTS_NAME)
+                ):
+                    # Load from a safetensors checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, SAFE_WEIGHTS_NAME)
+                elif is_safetensors_available() and os.path.isfile(
+                    os.path.join(pretrained_model_name_or_path, subfolder, SAFE_WEIGHTS_INDEX_NAME)
+                ):
+                    # Load from a sharded safetensors checkpoint
+                    archive_file = os.path.join(pretrained_model_name_or_path, subfolder, SAFE_WEIGHTS_INDEX_NAME)
+                    is_sharded = True
                 elif os.path.isfile(os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_NAME)):
                     # Load from a PyTorch checkpoint
                     archive_file = os.path.join(pretrained_model_name_or_path, subfolder, WEIGHTS_NAME)
@@ -2013,6 +2066,8 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     filename = TF2_WEIGHTS_NAME
                 elif from_flax:
                     filename = FLAX_WEIGHTS_NAME
+                elif is_safetensors_available():
+                    filename = SAFE_WEIGHTS_NAME
                 else:
                     filename = WEIGHTS_NAME
 
@@ -2033,8 +2088,21 @@ class PreTrainedModel(nn.Module, ModuleUtilsMixin, GenerationMixin, PushToHubMix
                     )
                     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
+                        else:
+                            # This repo has no safetensors file of any kind, we switch to PyTorch.
+                            filename = WEIGHTS_NAME
+                            resolved_archive_file = cached_file(
+                                pretrained_model_name_or_path, WEIGHTS_NAME, **cached_file_kwargs
+                            )
                     if resolved_archive_file is None and filename == WEIGHTS_NAME:
                         # Maybe the checkpoint is sharded, we try to grab the index name in this case.
                         resolved_archive_file = cached_file(

src/transformers/testing_utils.py

@@ -60,6 +60,7 @@ from .utils import (
     is_pytesseract_available,
     is_pytorch_quantization_available,
     is_rjieba_available,
+    is_safetensors_available,
     is_scatter_available,
     is_scipy_available,
     is_sentencepiece_available,
@@ -264,6 +265,13 @@ def require_accelerate(test_case):
     return unittest.skipUnless(is_accelerate_available(), "test requires accelerate")(test_case)
 
 
+def require_safetensors(test_case):
+    """
+    Decorator marking a test that requires safetensors. These tests are skipped when safetensors isn't installed.
+    """
+    return unittest.skipUnless(is_safetensors_available(), "test requires safetensors")(test_case)
+
+
 def require_rjieba(test_case):
     """
     Decorator marking a test that requires rjieba. These tests are skipped when rjieba isn't installed.

src/transformers/utils/__init__.py

@@ -111,6 +111,7 @@ from .import_utils import (
     is_pytorch_quantization_available,
     is_rjieba_available,
     is_sacremoses_available,
+    is_safetensors_available,
     is_sagemaker_dp_enabled,
     is_sagemaker_mp_enabled,
     is_scatter_available,
@@ -156,6 +157,8 @@ TF2_WEIGHTS_INDEX_NAME = "tf_model.h5.index.json"
 TF_WEIGHTS_NAME = "model.ckpt"
 FLAX_WEIGHTS_NAME = "flax_model.msgpack"
 FLAX_WEIGHTS_INDEX_NAME = "flax_model.msgpack.index.json"
+SAFE_WEIGHTS_NAME = "model.safetensors"
+SAFE_WEIGHTS_INDEX_NAME = "model.safetensors.index.json"
 CONFIG_NAME = "config.json"
 FEATURE_EXTRACTOR_NAME = "preprocessor_config.json"
 MODEL_CARD_NAME = "modelcard.json"

src/transformers/utils/import_utils.py

@@ -533,6 +533,10 @@ def is_accelerate_available():
     return importlib.util.find_spec("accelerate") is not None
 
 
+def is_safetensors_available():
+    return importlib.util.find_spec("safetensors") is not None
+
+
 def is_tokenizers_available():
     return importlib.util.find_spec("tokenizers") is not None
 

tests/test_modeling_common.py

@@ -53,6 +53,7 @@ from transformers.testing_utils import (
     is_pt_tf_cross_test,
     is_staging_test,
     require_accelerate,
+    require_safetensors,
     require_torch,
     require_torch_gpu,
     require_torch_multi_gpu,
@@ -61,6 +62,8 @@ from transformers.testing_utils import (
     torch_device,
 )
 from transformers.utils import (
+    SAFE_WEIGHTS_INDEX_NAME,
+    SAFE_WEIGHTS_NAME,
     WEIGHTS_INDEX_NAME,
     WEIGHTS_NAME,
     is_accelerate_available,
@@ -2980,6 +2983,57 @@ class ModelUtilsTest(TestCasePlus):
             "https://huggingface.co/hf-internal-testing/tiny-random-bert/resolve/main/pytorch_model.bin", config=config
         )
 
+    @require_safetensors
+    def test_safetensors_save_and_load(self):
+        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.save_pretrained(tmp_dir, safe_serialization=True)
+            # No pytorch_model.bin 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, WEIGHTS_NAME)))
+
+            new_model = BertModel.from_pretrained(tmp_dir)
+
+            # Check models are equal
+            for p1, p2 in zip(model.parameters(), new_model.parameters()):
+                self.assertTrue(torch.allclose(p1, p2))
+
+    @require_safetensors
+    def test_safetensors_load_from_hub(self):
+        safetensors_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-safetensors")
+        pytorch_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+
+        # Check models are equal
+        for p1, p2 in zip(safetensors_model.parameters(), pytorch_model.parameters()):
+            self.assertTrue(torch.allclose(p1, p2))
+
+    @require_safetensors
+    def test_safetensors_save_and_load_sharded(self):
+        model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert")
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.save_pretrained(tmp_dir, safe_serialization=True, max_shard_size="100kB")
+            # No pytorch_model.bin index file, only a model.safetensors index
+            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_INDEX_NAME)))
+            self.assertTrue(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_INDEX_NAME)))
+            # No regular weights file
+            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, WEIGHTS_NAME)))
+            self.assertFalse(os.path.isfile(os.path.join(tmp_dir, SAFE_WEIGHTS_NAME)))
+
+            new_model = BertModel.from_pretrained(tmp_dir)
+
+            # Check models are equal
+            for p1, p2 in zip(model.parameters(), new_model.parameters()):
+                self.assertTrue(torch.allclose(p1, p2))
+
+    @require_safetensors
+    def test_safetensors_load_from_hub_sharded(self):
+        safetensors_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded-safetensors")
+        pytorch_model = BertModel.from_pretrained("hf-internal-testing/tiny-random-bert-sharded")
+
+        # Check models are equal
+        for p1, p2 in zip(safetensors_model.parameters(), pytorch_model.parameters()):
+            self.assertTrue(torch.allclose(p1, p2))
+
 
 @require_torch
 @is_staging_test