[core] Layerwise Upcasting

docs/source/en/optimization/memory.md

@@ -158,6 +158,39 @@ In order to properly offload models after they're called, it is required to run
 
 </Tip>
 
+## FP8 layerwise weight-casting
+
+PyTorch supports `torch.float8_e4m3fn` and `torch.float8_e5m2` as weight storage dtypes. This precision cannot be used for performing computation directly for many different tensor operations due to unimplemented kernel support. However, one can still use these dtypes for storing model weights in lower FP8 precision. For computation, the weights can be upcasted on-the-fly as and when layers are invoked in the forward pass.
+
+Typically, inference on most models is done with `torch.float16` or `torch.bfloat16` weight/computation precision. Applying layerwise weight-casting, by storing the weights in FP8 precision, cuts down the memory footprint of the model weights by half approximately.
+
+```python
+import torch
+from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
+
+pipe = CogVideoXPipeline.from_pretrained("THUDM/CogVideoX-5b", torch_dtype=torch.bfloat16)
+pipe.to("cuda")
+pipe.transformer.enable_layerwise_upcasting(storage_dtype=torch.float8_e4m3fn, compute_dtype=torch.bfloat16)
+
+prompt = (
+    "A panda, dressed in a small, red jacket and a tiny hat, sits on a wooden stool in a serene bamboo forest. "
+    "The panda's fluffy paws strum a miniature acoustic guitar, producing soft, melodic tunes. Nearby, a few other "
+    "pandas gather, watching curiously and some clapping in rhythm. Sunlight filters through the tall bamboo, "
+    "casting a gentle glow on the scene. The panda's face is expressive, showing concentration and joy as it plays. "
+    "The background includes a small, flowing stream and vibrant green foliage, enhancing the peaceful and magical "
+    "atmosphere of this unique musical performance."
+)
+video = pipe(prompt=prompt, guidance_scale=6, num_inference_steps=50).frames[0]
+export_to_video(video, "output.mp4", fps=8)
+```
+
+In the above example, we enable layerwise upcasting on the transformer component of the pipeline. By default, certain layers are skipped from the FP8 weight casting because it can lead to significant degradation of generation quality. For most cases, skipping the normalization and modulation related weight parameters is a good choice (which is also the default choice). However, more control and flexibility can be obtained by directly utilizing the [`~hooks.layerwise_upcasting.apply_layerwise_upcasting`] function instead of using [`~ModelMixin.enable_layerwise_upcasting`].
+
+[[autodoc]] ModelMixin.enable_layerwise_upcasting
+
+[[autodoc]] hooks.layerwise_upcasting.apply_layerwise_upcasting
+
 ## Channels-last memory format
 
 The channels-last memory format is an alternative way of ordering NCHW tensors in memory to preserve dimension ordering. Channels-last tensors are ordered in such a way that the channels become the densest dimension (storing images pixel-per-pixel). Since not all operators currently support the channels-last format, it may result in worst performance but you should still try and see if it works for your model.

src/diffusers/hooks/__init__.py

@@ -0,0 +1,5 @@
+from ..utils import is_torch_available
+
+
+if is_torch_available():
+    from .layerwise_upcasting import apply_layerwise_upcasting, apply_layerwise_upcasting_hook

src/diffusers/hooks/hooks.py

@@ -0,0 +1,188 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import functools
+from typing import Any, Dict, Optional, Tuple
+
+import torch
+
+from ..utils.logging import get_logger
+
+
+logger = get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class ModelHook:
+    r"""
+    A hook that contains callbacks to be executed just before and after the forward method of a model.
+    """
+
+    _is_stateful = False
+
+    def initialize_hook(self, module: torch.nn.Module) -> torch.nn.Module:
+        r"""
+        Hook that is executed when a model is initialized.
+
+        Args:
+            module (`torch.nn.Module`):
+                The module attached to this hook.
+        """
+        return module
+
+    def deinitalize_hook(self, module: torch.nn.Module) -> torch.nn.Module:
+        r"""
+        Hook that is executed when a model is deinitalized.
+
+        Args:
+            module (`torch.nn.Module`):
+                The module attached to this hook.
+        """
+        module.forward = module._old_forward
+        del module._old_forward
+        return module
+
+    def pre_forward(self, module: torch.nn.Module, *args, **kwargs) -> Tuple[Tuple[Any], Dict[str, Any]]:
+        r"""
+        Hook that is executed just before the forward method of the model.
+
+        Args:
+            module (`torch.nn.Module`):
+                The module whose forward pass will be executed just after this event.
+            args (`Tuple[Any]`):
+                The positional arguments passed to the module.
+            kwargs (`Dict[Str, Any]`):
+                The keyword arguments passed to the module.
+        Returns:
+            `Tuple[Tuple[Any], Dict[Str, Any]]`:
+                A tuple with the treated `args` and `kwargs`.
+        """
+        return args, kwargs
+
+    def post_forward(self, module: torch.nn.Module, output: Any) -> Any:
+        r"""
+        Hook that is executed just after the forward method of the model.
+
+        Args:
+            module (`torch.nn.Module`):
+                The module whose forward pass been executed just before this event.
+            output (`Any`):
+                The output of the module.
+        Returns:
+            `Any`: The processed `output`.
+        """
+        return output
+
+    def detach_hook(self, module: torch.nn.Module) -> torch.nn.Module:
+        r"""
+        Hook that is executed when the hook is detached from a module.
+
+        Args:
+            module (`torch.nn.Module`):
+                The module detached from this hook.
+        """
+        return module
+
+    def reset_state(self, module: torch.nn.Module):
+        if self._is_stateful:
+            raise NotImplementedError("This hook is stateful and needs to implement the `reset_state` method.")
+        return module
+
+
+class HookRegistry:
+    def __init__(self, module_ref: torch.nn.Module) -> None:
+        super().__init__()
+
+        self.hooks: Dict[str, ModelHook] = {}
+
+        self._module_ref = module_ref
+        self._hook_order = []
+
+    def register_hook(self, hook: ModelHook, name: str) -> None:
+        if name in self.hooks.keys():
+            logger.warning(f"Hook with name {name} already exists, replacing it.")
+
+        if hasattr(self._module_ref, "_old_forward"):
+            old_forward = self._module_ref._old_forward
+        else:
+            old_forward = self._module_ref.forward
+            self._module_ref._old_forward = self._module_ref.forward
+
+        self._module_ref = hook.initialize_hook(self._module_ref)
+
+        if hasattr(hook, "new_forward"):
+            rewritten_forward = hook.new_forward
+
+            def new_forward(module, *args, **kwargs):
+                args, kwargs = hook.pre_forward(module, *args, **kwargs)
+                output = rewritten_forward(module, *args, **kwargs)
+                return hook.post_forward(module, output)
+        else:
+
+            def new_forward(module, *args, **kwargs):
+                args, kwargs = hook.pre_forward(module, *args, **kwargs)
+                output = old_forward(*args, **kwargs)
+                return hook.post_forward(module, output)
+
+        self._module_ref.forward = functools.update_wrapper(
+            functools.partial(new_forward, self._module_ref), old_forward
+        )
+
+        self.hooks[name] = hook
+        self._hook_order.append(name)
+
+    def get_hook(self, name: str) -> Optional[ModelHook]:
+        if name not in self.hooks.keys():
+            return None
+        return self.hooks[name]
+
+    def remove_hook(self, name: str, recurse: bool = True) -> None:
+        if name in self.hooks.keys():
+            hook = self.hooks[name]
+            self._module_ref = hook.deinitalize_hook(self._module_ref)
+            del self.hooks[name]
+            self._hook_order.remove(name)
+
+        if recurse:
+            for module_name, module in self._module_ref.named_modules():
+                if module_name == "":
+                    continue
+                if hasattr(module, "_diffusers_hook"):
+                    module._diffusers_hook.remove_hook(name, recurse=False)
+
+    def reset_stateful_hooks(self, recurse: bool = True) -> None:
+        for hook_name in self._hook_order:
+            hook = self.hooks[hook_name]
+            if hook._is_stateful:
+                hook.reset_state(self._module_ref)
+
+        if recurse:
+            for module_name, module in self._module_ref.named_modules():
+                if module_name == "":
+                    continue
+                if hasattr(module, "_diffusers_hook"):
+                    module._diffusers_hook.reset_stateful_hooks(recurse=False)
+
+    @classmethod
+    def check_if_exists_or_initialize(cls, module: torch.nn.Module) -> "HookRegistry":
+        if not hasattr(module, "_diffusers_hook"):
+            module._diffusers_hook = cls(module)
+        return module._diffusers_hook
+
+    def __repr__(self) -> str:
+        hook_repr = ""
+        for i, hook_name in enumerate(self._hook_order):
+            hook_repr += f"  ({i}) {hook_name} - ({self.hooks[hook_name].__class__.__name__})"
+            if i < len(self._hook_order) - 1:
+                hook_repr += "\n"
+        return f"HookRegistry(\n{hook_repr}\n)"