add accelerate to load models with smaller memory footprint (huggingface#361)

* add accelerate to load models with smaller memory footprint

* remove low_cpu_mem_usage as it is reduntant

* move accelerate init weights context to modelling utils

* add test to ensure results are the same when loading with accelerate

* add tests to ensure ram usage gets lower when using accelerate

* move accelerate logic to single snippet under modelling utils and remove it from configuration utils

* format code using to pass quality check

* fix imports with isor

* add accelerate to test extra deps

* only import accelerate if device_map is set to auto

* move accelerate availability check to diffusers import utils

* format code

Co-authored-by: Patrick von Platen <[email protected]>

Author: 2 people

setup.py

@@ -104,6 +104,7 @@
     "torch>=1.4",
     "torchvision",
     "transformers>=4.21.0",
+    "accelerate>=0.12.0"
 ]
 
 # this is a lookup table with items like:

src/diffusers/modeling_utils.py

@@ -21,6 +21,7 @@
 import torch
 from torch import Tensor, device
 
+from diffusers.utils import is_accelerate_available
 from huggingface_hub import hf_hub_download
 from huggingface_hub.utils import EntryNotFoundError, RepositoryNotFoundError, RevisionNotFoundError
 from requests import HTTPError
@@ -293,33 +294,13 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
         from_auto_class = kwargs.pop("_from_auto", False)
         torch_dtype = kwargs.pop("torch_dtype", None)
         subfolder = kwargs.pop("subfolder", None)
+        device_map = kwargs.pop("device_map", None)
 
         user_agent = {"file_type": "model", "framework": "pytorch", "from_auto_class": from_auto_class}
 
         # Load config if we don't provide a configuration
         config_path = pretrained_model_name_or_path
-        model, unused_kwargs = cls.from_config(
-            config_path,
-            cache_dir=cache_dir,
-            return_unused_kwargs=True,
-            force_download=force_download,
-            resume_download=resume_download,
-            proxies=proxies,
-            local_files_only=local_files_only,
-            use_auth_token=use_auth_token,
-            revision=revision,
-            subfolder=subfolder,
-            **kwargs,
-        )
 
-        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
-            raise ValueError(
-                f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
-            )
-        elif torch_dtype is not None:
-            model = model.to(torch_dtype)
-
-        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
         # This variable will flag if we're loading a sharded checkpoint. In this case the archive file is just the
         # Load model
         pretrained_model_name_or_path = str(pretrained_model_name_or_path)
@@ -391,25 +372,81 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 )
 
             # restore default dtype
-        state_dict = load_state_dict(model_file)
-        model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
-            model,
-            state_dict,
-            model_file,
-            pretrained_model_name_or_path,
-            ignore_mismatched_sizes=ignore_mismatched_sizes,
-        )
 
-        # Set model in evaluation mode to deactivate DropOut modules by default
-        model.eval()
+        if device_map == "auto":
+            if is_accelerate_available():
+                import accelerate
+            else:
+                raise ImportError("Please install accelerate via `pip install accelerate`")
+
+            with accelerate.init_empty_weights():
+                model, unused_kwargs = cls.from_config(
+                    config_path,
+                    cache_dir=cache_dir,
+                    return_unused_kwargs=True,
+                    force_download=force_download,
+                    resume_download=resume_download,
+                    proxies=proxies,
+                    local_files_only=local_files_only,
+                    use_auth_token=use_auth_token,
+                    revision=revision,
+                    subfolder=subfolder,
+                    device_map=device_map,
+                    **kwargs,
+                )
+
+            accelerate.load_checkpoint_and_dispatch(model, model_file, device_map)
+
+            loading_info = {
+                "missing_keys": [],
+                "unexpected_keys": [],
+                "mismatched_keys": [],
+                "error_msgs": [],
+            }
+        else:
+            model, unused_kwargs = cls.from_config(
+                config_path,
+                cache_dir=cache_dir,
+                return_unused_kwargs=True,
+                force_download=force_download,
+                resume_download=resume_download,
+                proxies=proxies,
+                local_files_only=local_files_only,
+                use_auth_token=use_auth_token,
+                revision=revision,
+                subfolder=subfolder,
+                device_map=device_map,
+                **kwargs,
+            )
+
+            state_dict = load_state_dict(model_file)
+            model, missing_keys, unexpected_keys, mismatched_keys, error_msgs = cls._load_pretrained_model(
+                model,
+                state_dict,
+                model_file,
+                pretrained_model_name_or_path,
+                ignore_mismatched_sizes=ignore_mismatched_sizes,
+            )
 
-        if output_loading_info:
             loading_info = {
                 "missing_keys": missing_keys,
                 "unexpected_keys": unexpected_keys,
                 "mismatched_keys": mismatched_keys,
                 "error_msgs": error_msgs,
             }
+
+        if torch_dtype is not None and not isinstance(torch_dtype, torch.dtype):
+            raise ValueError(
+                f"{torch_dtype} needs to be of type `torch.dtype`, e.g. `torch.float16`, but is {type(torch_dtype)}."
+            )
+        elif torch_dtype is not None:
+            model = model.to(torch_dtype)
+
+        model.register_to_config(_name_or_path=pretrained_model_name_or_path)
+
+        # Set model in evaluation mode to deactivate DropOut modules by default
+        model.eval()
+        if output_loading_info:
             return model, loading_info
 
         return model

src/diffusers/utils/__init__.py

@@ -23,6 +23,7 @@
     USE_TF,
     USE_TORCH,
     DummyObject,
+    is_accelerate_available,
     is_flax_available,
     is_inflect_available,
     is_modelcards_available,

src/diffusers/utils/import_utils.py

@@ -159,6 +159,13 @@
 except importlib_metadata.PackageNotFoundError:
     _scipy_available = False
 
+_accelerate_available = importlib.util.find_spec("accelerate") is not None
+try:
+    _accelerate_version = importlib_metadata.version("accelerate")
+    logger.debug(f"Successfully imported accelerate version {_accelerate_version}")
+except importlib_metadata.PackageNotFoundError:
+    _accelerate_available = False
+
 
 def is_torch_available():
     return _torch_available
@@ -196,6 +203,10 @@ def is_scipy_available():
     return _scipy_available
 
 
+def is_accelerate_available():
+    return _accelerate_available
+
+
 # docstyle-ignore
 FLAX_IMPORT_ERROR = """
 {0} requires the FLAX library but it was not found in your environment. Checkout the instructions on the

tests/test_models_unet.py

@@ -13,7 +13,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 
+import gc
 import math
+import tracemalloc
 import unittest
 
 import torch
@@ -133,6 +135,74 @@ def test_from_pretrained_hub(self):
 
         assert image is not None, "Make sure output is not None"
 
+    @unittest.skipIf(torch_device == "cpu", "This test is supposed to run on GPU")
+    def test_from_pretrained_accelerate(self):
+        model, _ = UNet2DModel.from_pretrained(
+            "fusing/unet-ldm-dummy-update", output_loading_info=True, device_map="auto"
+        )
+        model.to(torch_device)
+        image = model(**self.dummy_input).sample
+
+        assert image is not None, "Make sure output is not None"
+
+    @unittest.skipIf(torch_device == "cpu", "This test is supposed to run on GPU")
+    def test_from_pretrained_accelerate_wont_change_results(self):
+        model_accelerate, _ = UNet2DModel.from_pretrained(
+            "fusing/unet-ldm-dummy-update", output_loading_info=True, device_map="auto"
+        )
+        model_accelerate.to(torch_device)
+        model_accelerate.eval()
+
+        noise = torch.randn(
+            1,
+            model_accelerate.config.in_channels,
+            model_accelerate.config.sample_size,
+            model_accelerate.config.sample_size,
+            generator=torch.manual_seed(0),
+        )
+        noise = noise.to(torch_device)
+        time_step = torch.tensor([10] * noise.shape[0]).to(torch_device)
+
+        arr_accelerate = model_accelerate(noise, time_step)["sample"]
+
+        # two models don't need to stay in the device at the same time
+        del model_accelerate
+        torch.cuda.empty_cache()
+        gc.collect()
+
+        model_normal_load, _ = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)
+        model_normal_load.to(torch_device)
+        model_normal_load.eval()
+        arr_normal_load = model_normal_load(noise, time_step)["sample"]
+
+        assert torch.allclose(arr_accelerate, arr_normal_load, rtol=1e-3)
+
+    @unittest.skipIf(torch_device == "cpu", "This test is supposed to run on GPU")
+    def test_memory_footprint_gets_reduced(self):
+        torch.cuda.empty_cache()
+        gc.collect()
+
+        tracemalloc.start()
+        model_accelerate, _ = UNet2DModel.from_pretrained(
+            "fusing/unet-ldm-dummy-update", output_loading_info=True, device_map="auto"
+        )
+        model_accelerate.to(torch_device)
+        model_accelerate.eval()
+        _, peak_accelerate = tracemalloc.get_traced_memory()
+
+        del model_accelerate
+        torch.cuda.empty_cache()
+        gc.collect()
+
+        model_normal_load, _ = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update", output_loading_info=True)
+        model_normal_load.to(torch_device)
+        model_normal_load.eval()
+        _, peak_normal = tracemalloc.get_traced_memory()
+
+        tracemalloc.stop()
+
+        assert peak_accelerate < peak_normal
+
     def test_output_pretrained(self):
         model = UNet2DModel.from_pretrained("fusing/unet-ldm-dummy-update")
         model.eval()