[Core] introduce _no_split_modules to ModelMixin (huggingface#6396)

* introduce _no_split_modules.

* unnecessary spaces.

* remove unnecessary kwargs and style

* fix: accelerate imports.

* change to _determine_device_map

* add the blocks that have residual connections.

* add: CrossAttnUpBlock2D

* add: testin

* style

* line-spaces

* quality

* add disk offload test without safetensors.

* checking disk offloading percentages.

* change model split

* add: utility for checking multi-gpu requirement.

* model parallelism test

* splits.

* splits.

* splits

* splits.

* splits.

* splits.

* offload folder to test_disk_offload_with_safetensors

* add _no_split_modules

* fix-copies

Author: sayakpaul

src/diffusers/models/autoencoders/autoencoder_kl.py

@@ -65,6 +65,7 @@ class AutoencoderKL(ModelMixin, ConfigMixin, FromOriginalVAEMixin):
     """
 
     _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D"]
 
     @register_to_config
     def __init__(

src/diffusers/models/modeling_utils.py

@@ -57,7 +57,8 @@
 
 if is_accelerate_available():
     import accelerate
-    from accelerate.utils import set_module_tensor_to_device
+    from accelerate import infer_auto_device_map
+    from accelerate.utils import get_balanced_memory, get_max_memory, set_module_tensor_to_device
     from accelerate.utils.versions import is_torch_version
 
 
@@ -99,6 +100,29 @@ def find_tensor_attributes(module: torch.nn.Module) -> List[Tuple[str, Tensor]]:
         return first_tuple[1].dtype
 
 
+# Adapted from `transformers` (see modeling_utils.py)
+def _determine_device_map(model: "ModelMixin", device_map, max_memory, torch_dtype):
+    if isinstance(device_map, str):
+        no_split_modules = model._get_no_split_modules(device_map)
+        device_map_kwargs = {"no_split_module_classes": no_split_modules}
+
+        if device_map != "sequential":
+            max_memory = get_balanced_memory(
+                model,
+                dtype=torch_dtype,
+                low_zero=(device_map == "balanced_low_0"),
+                max_memory=max_memory,
+                **device_map_kwargs,
+            )
+        else:
+            max_memory = get_max_memory(max_memory)
+
+        device_map_kwargs["max_memory"] = max_memory
+        device_map = infer_auto_device_map(model, dtype=torch_dtype, **device_map_kwargs)
+
+    return device_map
+
+
 def load_state_dict(checkpoint_file: Union[str, os.PathLike], variant: Optional[str] = None):
     """
     Reads a checkpoint file, returning properly formatted errors if they arise.
@@ -201,6 +225,7 @@ class ModelMixin(torch.nn.Module, PushToHubMixin):
     _automatically_saved_args = ["_diffusers_version", "_class_name", "_name_or_path"]
     _supports_gradient_checkpointing = False
     _keys_to_ignore_on_load_unexpected = None
+    _no_split_modules = None
 
     def __init__(self):
         super().__init__()
@@ -560,6 +585,36 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 " dispatching. Please make sure to set `low_cpu_mem_usage=True`."
             )
 
+        # change device_map into a map if we passed an int, a str or a torch.device
+        if isinstance(device_map, torch.device):
+            device_map = {"": device_map}
+        elif isinstance(device_map, str) and device_map not in ["auto", "balanced", "balanced_low_0", "sequential"]:
+            try:
+                device_map = {"": torch.device(device_map)}
+            except RuntimeError:
+                raise ValueError(
+                    "When passing device_map as a string, the value needs to be a device name (e.g. cpu, cuda:0) or "
+                    f"'auto', 'balanced', 'balanced_low_0', 'sequential' but found {device_map}."
+                )
+        elif isinstance(device_map, int):
+            if device_map < 0:
+                raise ValueError(
+                    "You can't pass device_map as a negative int. If you want to put the model on the cpu, pass device_map = 'cpu' "
+                )
+            else:
+                device_map = {"": device_map}
+
+        if device_map is not None:
+            if low_cpu_mem_usage is None:
+                low_cpu_mem_usage = True
+            elif not low_cpu_mem_usage:
+                raise ValueError("Passing along a `device_map` requires `low_cpu_mem_usage=True`")
+
+        if low_cpu_mem_usage:
+            if device_map is not None and not is_torch_version(">=", "1.10"):
+                # The max memory utils require PyTorch >= 1.10 to have torch.cuda.mem_get_info.
+                raise ValueError("`low_cpu_mem_usage` and `device_map` require PyTorch >= 1.10.")
+
         # Load config if we don't provide a configuration
         config_path = pretrained_model_name_or_path
 
@@ -582,10 +637,6 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
             token=token,
             revision=revision,
             subfolder=subfolder,
-            device_map=device_map,
-            max_memory=max_memory,
-            offload_folder=offload_folder,
-            offload_state_dict=offload_state_dict,
             user_agent=user_agent,
             **kwargs,
         )
@@ -690,6 +741,7 @@ def from_pretrained(cls, pretrained_model_name_or_path: Optional[Union[str, os.P
                 else:  # else let accelerate handle loading and dispatching.
                     # Load weights and dispatch according to the device_map
                     # by default the device_map is None and the weights are loaded on the CPU
+                    device_map = _determine_device_map(model, device_map, max_memory, torch_dtype)
                     try:
                         accelerate.load_checkpoint_and_dispatch(
                             model,
@@ -881,6 +933,36 @@ def _find_mismatched_keys(
 
         return model, missing_keys, unexpected_keys, mismatched_keys, error_msgs
 
+    # Adapted from `transformers` modeling_utils.py
+    def _get_no_split_modules(self, device_map: str):
+        """
+        Get the modules of the model that should not be spit when using device_map. We iterate through the modules to
+        get the underlying `_no_split_modules`.
+
+        Args:
+            device_map (`str`):
+                The device map value. Options are ["auto", "balanced", "balanced_low_0", "sequential"]
+
+        Returns:
+            `List[str]`: List of modules that should not be split
+        """
+        _no_split_modules = set()
+        modules_to_check = [self]
+        while len(modules_to_check) > 0:
+            module = modules_to_check.pop(-1)
+            # if the module does not appear in _no_split_modules, we also check the children
+            if module.__class__.__name__ not in _no_split_modules:
+                if isinstance(module, ModelMixin):
+                    if module._no_split_modules is None:
+                        raise ValueError(
+                            f"{module.__class__.__name__} does not support `device_map='{device_map}'`. To implement support, the model "
+                            "class needs to implement the `_no_split_modules` attribute."
+                        )
+                    else:
+                        _no_split_modules = _no_split_modules | set(module._no_split_modules)
+                modules_to_check += list(module.children())
+        return list(_no_split_modules)
+
     @property
     def device(self) -> torch.device:
         """

src/diffusers/models/transformers/transformer_2d.py

@@ -72,6 +72,7 @@ class Transformer2DModel(ModelMixin, ConfigMixin):
     """
 
     _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock"]
 
     @register_to_config
     def __init__(

src/diffusers/models/unets/unet_2d_condition.py

@@ -161,6 +161,7 @@ class conditioning with `class_embed_type` equal to `None`.
     """
 
     _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
 
     @register_to_config
     def __init__(

src/diffusers/pipelines/deprecated/versatile_diffusion/modeling_text_unet.py

@@ -363,6 +363,7 @@ class conditioning with `class_embed_type` equal to `None`.
     """
 
     _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlockFlat", "CrossAttnUpBlockFlat"]
 
     @register_to_config
     def __init__(

tests/models/test_modeling_common.py

@@ -24,6 +24,7 @@
 import numpy as np
 import requests_mock
 import torch
+from accelerate.utils import compute_module_sizes
 from huggingface_hub import ModelCard, delete_repo
 from huggingface_hub.utils import is_jinja_available
 from requests.exceptions import HTTPError
@@ -39,6 +40,7 @@
     require_torch_2,
     require_torch_accelerator_with_training,
     require_torch_gpu,
+    require_torch_multi_gpu,
     run_test_in_subprocess,
     torch_device,
 )
@@ -200,6 +202,21 @@ class ModelTesterMixin:
     main_input_name = None  # overwrite in model specific tester class
     base_precision = 1e-3
     forward_requires_fresh_args = False
+    model_split_percents = [0.5, 0.7, 0.9]
+
+    def check_device_map_is_respected(self, model, device_map):
+        for param_name, param in model.named_parameters():
+            # Find device in device_map
+            while len(param_name) > 0 and param_name not in device_map:
+                param_name = ".".join(param_name.split(".")[:-1])
+            if param_name not in device_map:
+                raise ValueError("device map is incomplete, it does not contain any device for `param_name`.")
+
+            param_device = device_map[param_name]
+            if param_device in ["cpu", "disk"]:
+                self.assertEqual(param.device, torch.device("meta"))
+            else:
+                self.assertEqual(param.device, torch.device(param_device))
 
     def test_from_save_pretrained(self, expected_max_diff=5e-5):
         if self.forward_requires_fresh_args:
@@ -670,6 +687,117 @@ def test_deprecated_kwargs(self):
                 " from `_deprecated_kwargs = [<deprecated_argument>]`"
             )
 
+    @require_torch_gpu
+    def test_cpu_offload(self):
+        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        # We test several splits of sizes to make sure it works.
+        max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.cpu().save_pretrained(tmp_dir)
+
+            for max_size in max_gpu_sizes:
+                max_memory = {0: max_size, "cpu": model_size * 2}
+                new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
+                # Making sure part of the model will actually end up offloaded
+                self.assertSetEqual(set(new_model.hf_device_map.values()), {0, "cpu"})
+
+                self.check_device_map_is_respected(new_model, new_model.hf_device_map)
+                torch.manual_seed(0)
+                new_output = new_model(**inputs_dict)
+
+                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
+
+    @require_torch_gpu
+    def test_disk_offload_without_safetensors(self):
+        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.cpu().save_pretrained(tmp_dir, safe_serialization=False)
+
+            with self.assertRaises(ValueError):
+                max_size = int(self.model_split_percents[1] * model_size)
+                max_memory = {0: max_size, "cpu": max_size}
+                # This errors out because it's missing an offload folder
+                new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
+
+            max_size = int(self.model_split_percents[1] * model_size)
+            max_memory = {0: max_size, "cpu": max_size}
+            new_model = self.model_class.from_pretrained(
+                tmp_dir, device_map="auto", max_memory=max_memory, offload_folder=tmp_dir
+            )
+
+            self.check_device_map_is_respected(new_model, new_model.hf_device_map)
+            torch.manual_seed(0)
+            new_output = new_model(**inputs_dict)
+
+            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
+
+    @require_torch_gpu
+    def test_disk_offload_with_safetensors(self):
+        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.cpu().save_pretrained(tmp_dir)
+
+            max_size = int(self.model_split_percents[1] * model_size)
+            max_memory = {0: max_size, "cpu": max_size}
+            new_model = self.model_class.from_pretrained(
+                tmp_dir, device_map="auto", offload_folder=tmp_dir, max_memory=max_memory
+            )
+
+            self.check_device_map_is_respected(new_model, new_model.hf_device_map)
+            torch.manual_seed(0)
+            new_output = new_model(**inputs_dict)
+
+            self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
+
+    @require_torch_multi_gpu
+    def test_model_parallelism(self):
+        config, inputs_dict = self.prepare_init_args_and_inputs_for_common()
+        model = self.model_class(**config).eval()
+        model = model.to(torch_device)
+
+        torch.manual_seed(0)
+        base_output = model(**inputs_dict)
+
+        model_size = compute_module_sizes(model)[""]
+        # We test several splits of sizes to make sure it works.
+        max_gpu_sizes = [int(p * model_size) for p in self.model_split_percents[1:]]
+        with tempfile.TemporaryDirectory() as tmp_dir:
+            model.cpu().save_pretrained(tmp_dir)
+
+            for max_size in max_gpu_sizes:
+                max_memory = {0: max_size, 1: model_size * 2, "cpu": model_size * 2}
+                new_model = self.model_class.from_pretrained(tmp_dir, device_map="auto", max_memory=max_memory)
+                # Making sure part of the model will actually end up offloaded
+                self.assertSetEqual(set(new_model.hf_device_map.values()), {0, 1})
+
+                self.check_device_map_is_respected(new_model, new_model.hf_device_map)
+
+                torch.manual_seed(0)
+                new_output = new_model(**inputs_dict)
+
+                self.assertTrue(torch.allclose(base_output[0], new_output[0], atol=1e-5))
+
 
 @is_staging_test
 class ModelPushToHubTester(unittest.TestCase):

tests/models/unets/test_models_unet_2d_condition.py

@@ -300,6 +300,8 @@ def create_custom_diffusion_layers(model, mock_weights: bool = True):
 class UNet2DConditionModelTests(ModelTesterMixin, UNetTesterMixin, unittest.TestCase):
     model_class = UNet2DConditionModel
     main_input_name = "sample"
+    # We override the items here because the unet under consideration is small.
+    model_split_percents = [0.5, 0.3, 0.4]
 
     @property
     def dummy_input(self):