Fix config deprecation (#3129)

* Better deprecation message

* Better deprecation message

* Better doc string

* Fixes

* fix more

* fix more

* Improve __getattr__

* correct more

* fix more

* fix

* Improve more

* more improvements

* fix more

* Apply suggestions from code review

Co-authored-by: Pedro Cuenca <[email protected]>

* make style

* Fix all rest & add tests & remove old deprecation fns

---------

Co-authored-by: Pedro Cuenca <[email protected]>

Author: patrickvonplaten

examples/community/unclip_image_interpolation.py

@@ -372,9 +372,9 @@ def __call__(
         self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
         decoder_timesteps_tensor = self.decoder_scheduler.timesteps
 
-        num_channels_latents = self.decoder.in_channels
-        height = self.decoder.sample_size
-        width = self.decoder.sample_size
+        num_channels_latents = self.decoder.config.in_channels
+        height = self.decoder.config.sample_size
+        width = self.decoder.config.sample_size
 
         decoder_latents = self.prepare_latents(
             (batch_size, num_channels_latents, height, width),
@@ -425,9 +425,9 @@ def __call__(
         self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
         super_res_timesteps_tensor = self.super_res_scheduler.timesteps
 
-        channels = self.super_res_first.in_channels // 2
-        height = self.super_res_first.sample_size
-        width = self.super_res_first.sample_size
+        channels = self.super_res_first.config.in_channels // 2
+        height = self.super_res_first.config.sample_size
+        width = self.super_res_first.config.sample_size
 
         super_res_latents = self.prepare_latents(
             (batch_size, channels, height, width),

examples/community/unclip_text_interpolation.py

@@ -452,9 +452,9 @@ def __call__(
         self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
         decoder_timesteps_tensor = self.decoder_scheduler.timesteps
 
-        num_channels_latents = self.decoder.in_channels
-        height = self.decoder.sample_size
-        width = self.decoder.sample_size
+        num_channels_latents = self.decoder.config.in_channels
+        height = self.decoder.config.sample_size
+        width = self.decoder.config.sample_size
 
         decoder_latents = self.prepare_latents(
             (batch_size, num_channels_latents, height, width),
@@ -505,9 +505,9 @@ def __call__(
         self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
         super_res_timesteps_tensor = self.super_res_scheduler.timesteps
 
-        channels = self.super_res_first.in_channels // 2
-        height = self.super_res_first.sample_size
-        width = self.super_res_first.sample_size
+        channels = self.super_res_first.config.in_channels // 2
+        height = self.super_res_first.config.sample_size
+        width = self.super_res_first.config.sample_size
 
         super_res_latents = self.prepare_latents(
             (batch_size, channels, height, width),

src/diffusers/configuration_utils.py

@@ -118,6 +118,24 @@ def register_to_config(self, **kwargs):
 
         self._internal_dict = FrozenDict(internal_dict)
 
+    def __getattr__(self, name: str) -> Any:
+        """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
+        config attributes directly. See https://github.com/huggingface/diffusers/pull/3129
+
+        Tihs funtion is mostly copied from PyTorch's __getattr__ overwrite:
+        https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+        """
+
+        is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
+        is_attribute = name in self.__dict__
+
+        if is_in_config and not is_attribute:
+            deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'scheduler.config.{name}'."
+            deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False)
+            return self._internal_dict[name]
+
+        raise AttributeError(f"'{type(self).__name__}' object has no attribute '{name}'")
+
     def save_config(self, save_directory: Union[str, os.PathLike], push_to_hub: bool = False, **kwargs):
         """
         Save a configuration object to the directory `save_directory`, so that it can be re-loaded using the

src/diffusers/models/autoencoder_kl.py

@@ -18,7 +18,7 @@
 import torch.nn as nn
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput, apply_forward_hook, deprecate
+from ..utils import BaseOutput, apply_forward_hook
 from .modeling_utils import ModelMixin
 from .vae import Decoder, DecoderOutput, DiagonalGaussianDistribution, Encoder
 
@@ -123,16 +123,6 @@ def __init__(
         self.tile_latent_min_size = int(sample_size / (2 ** (len(self.config.block_out_channels) - 1)))
         self.tile_overlap_factor = 0.25
 
-    @property
-    def block_out_channels(self):
-        deprecate(
-            "block_out_channels",
-            "1.0.0",
-            "Accessing `block_out_channels` directly via vae.block_out_channels is deprecated. Please use `vae.config.block_out_channels instead`",
-            standard_warn=False,
-        )
-        return self.config.block_out_channels
-
     def _set_gradient_checkpointing(self, module, value=False):
         if isinstance(module, (Encoder, Decoder)):
             module.gradient_checkpointing = value

src/diffusers/models/modeling_utils.py

@@ -17,7 +17,7 @@
 import inspect
 import os
 from functools import partial
-from typing import Callable, List, Optional, Tuple, Union
+from typing import Any, Callable, List, Optional, Tuple, Union
 
 import torch
 from torch import Tensor, device
@@ -32,6 +32,7 @@
     WEIGHTS_NAME,
     _add_variant,
     _get_model_file,
+    deprecate,
     is_accelerate_available,
     is_safetensors_available,
     is_torch_version,
@@ -156,6 +157,24 @@ class ModelMixin(torch.nn.Module):
     def __init__(self):
         super().__init__()
 
+    def __getattr__(self, name: str) -> Any:
+        """The only reason we overwrite `getattr` here is to gracefully deprecate accessing
+        config attributes directly. See https://github.com/huggingface/diffusers/pull/3129 We need to overwrite
+        __getattr__ here in addition so that we don't trigger `torch.nn.Module`'s __getattr__':
+        https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+        """
+
+        is_in_config = "_internal_dict" in self.__dict__ and hasattr(self.__dict__["_internal_dict"], name)
+        is_attribute = name in self.__dict__
+
+        if is_in_config and not is_attribute:
+            deprecation_message = f"Accessing config attribute `{name}` directly via '{type(self).__name__}' object attribute is deprecated. Please access '{name}' over '{type(self).__name__}'s config object instead, e.g. 'unet.config.{name}'."
+            deprecate("direct config name access", "1.0.0", deprecation_message, standard_warn=False, stacklevel=3)
+            return self._internal_dict[name]
+
+        # call PyTorch's https://pytorch.org/docs/stable/_modules/torch/nn/modules/module.html#Module
+        return super().__getattr__(name)
+
     @property
     def is_gradient_checkpointing(self) -> bool:
         """

src/diffusers/models/unet_1d.py

@@ -19,7 +19,7 @@
 import torch.nn as nn
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput, deprecate
+from ..utils import BaseOutput
 from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
 from .modeling_utils import ModelMixin
 from .unet_1d_blocks import get_down_block, get_mid_block, get_out_block, get_up_block
@@ -190,16 +190,6 @@ def __init__(
             fc_dim=block_out_channels[-1] // 4,
         )
 
-    @property
-    def in_channels(self):
-        deprecate(
-            "in_channels",
-            "1.0.0",
-            "Accessing `in_channels` directly via unet.in_channels is deprecated. Please use `unet.config.in_channels` instead",
-            standard_warn=False,
-        )
-        return self.config.in_channels
-
     def forward(
         self,
         sample: torch.FloatTensor,

src/diffusers/models/unet_2d.py

@@ -18,7 +18,7 @@
 import torch.nn as nn
 
 from ..configuration_utils import ConfigMixin, register_to_config
-from ..utils import BaseOutput, deprecate
+from ..utils import BaseOutput
 from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
 from .modeling_utils import ModelMixin
 from .unet_2d_blocks import UNetMidBlock2D, get_down_block, get_up_block
@@ -216,16 +216,6 @@ def __init__(
         self.conv_act = nn.SiLU()
         self.conv_out = nn.Conv2d(block_out_channels[0], out_channels, kernel_size=3, padding=1)
 
-    @property
-    def in_channels(self):
-        deprecate(
-            "in_channels",
-            "1.0.0",
-            "Accessing `in_channels` directly via unet.in_channels is deprecated. Please use `unet.config.in_channels` instead",
-            standard_warn=False,
-        )
-        return self.config.in_channels
-
     def forward(
         self,
         sample: torch.FloatTensor,

src/diffusers/models/unet_2d_condition.py

@@ -21,7 +21,7 @@
 
 from ..configuration_utils import ConfigMixin, register_to_config
 from ..loaders import UNet2DConditionLoadersMixin
-from ..utils import BaseOutput, deprecate, logging
+from ..utils import BaseOutput, logging
 from .attention_processor import AttentionProcessor, AttnProcessor
 from .embeddings import GaussianFourierProjection, TimestepEmbedding, Timesteps
 from .modeling_utils import ModelMixin
@@ -447,16 +447,6 @@ def __init__(
             block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
         )
 
-    @property
-    def in_channels(self):
-        deprecate(
-            "in_channels",
-            "1.0.0",
-            "Accessing `in_channels` directly via unet.in_channels is deprecated. Please use `unet.config.in_channels` instead",
-            standard_warn=False,
-        )
-        return self.config.in_channels
-
     @property
     def attn_processors(self) -> Dict[str, AttentionProcessor]:
         r"""

src/diffusers/pipelines/pipeline_utils.py

@@ -508,7 +508,7 @@ def register_modules(self, **kwargs):
             setattr(self, name, module)
 
     def __setattr__(self, name: str, value: Any):
-        if hasattr(self, name) and hasattr(self.config, name):
+        if name in self.__dict__ and hasattr(self.config, name):
             # We need to overwrite the config if name exists in config
             if isinstance(getattr(self.config, name), (tuple, list)):
                 if value is not None and self.config[name][0] is not None:
@@ -648,26 +648,25 @@ def module_is_offloaded(module):
             )
 
         module_names, _ = self._get_signature_keys(self)
-        module_names = [m for m in module_names if hasattr(self, m)]
+        modules = [getattr(self, n, None) for n in module_names]
+        modules = [m for m in modules if isinstance(m, torch.nn.Module)]
 
         is_offloaded = pipeline_is_offloaded or pipeline_is_sequentially_offloaded
-        for name in module_names:
-            module = getattr(self, name)
-            if isinstance(module, torch.nn.Module):
-                module.to(torch_device, torch_dtype)
-                if (
-                    module.dtype == torch.float16
-                    and str(torch_device) in ["cpu"]
-                    and not silence_dtype_warnings
-                    and not is_offloaded
-                ):
-                    logger.warning(
-                        "Pipelines loaded with `torch_dtype=torch.float16` cannot run with `cpu` device. It"
-                        " is not recommended to move them to `cpu` as running them will fail. Please make"
-                        " sure to use an accelerator to run the pipeline in inference, due to the lack of"
-                        " support for`float16` operations on this device in PyTorch. Please, remove the"
-                        " `torch_dtype=torch.float16` argument, or use another device for inference."
-                    )
+        for module in modules:
+            module.to(torch_device, torch_dtype)
+            if (
+                module.dtype == torch.float16
+                and str(torch_device) in ["cpu"]
+                and not silence_dtype_warnings
+                and not is_offloaded
+            ):
+                logger.warning(
+                    "Pipelines loaded with `torch_dtype=torch.float16` cannot run with `cpu` device. It"
+                    " is not recommended to move them to `cpu` as running them will fail. Please make"
+                    " sure to use an accelerator to run the pipeline in inference, due to the lack of"
+                    " support for`float16` operations on this device in PyTorch. Please, remove the"
+                    " `torch_dtype=torch.float16` argument, or use another device for inference."
+                )
         return self
 
     @property
@@ -677,12 +676,12 @@ def device(self) -> torch.device:
             `torch.device`: The torch device on which the pipeline is located.
         """
         module_names, _ = self._get_signature_keys(self)
-        module_names = [m for m in module_names if hasattr(self, m)]
+        modules = [getattr(self, n, None) for n in module_names]
+        modules = [m for m in modules if isinstance(m, torch.nn.Module)]
+
+        for module in modules:
+            return module.device
 
-        for name in module_names:
-            module = getattr(self, name)
-            if isinstance(module, torch.nn.Module):
-                return module.device
         return torch.device("cpu")
 
     @classmethod
@@ -1451,13 +1450,12 @@ def fn_recursive_set_mem_eff(module: torch.nn.Module):
             for child in module.children():
                 fn_recursive_set_mem_eff(child)
 
-        module_names, _, _ = self.extract_init_dict(dict(self.config))
-        module_names = [m for m in module_names if hasattr(self, m)]
+        module_names, _ = self._get_signature_keys(self)
+        modules = [getattr(self, n, None) for n in module_names]
+        modules = [m for m in modules if isinstance(m, torch.nn.Module)]
 
-        for module_name in module_names:
-            module = getattr(self, module_name)
-            if isinstance(module, torch.nn.Module):
-                fn_recursive_set_mem_eff(module)
+        for module in modules:
+            fn_recursive_set_mem_eff(module)
 
     def enable_attention_slicing(self, slice_size: Optional[Union[str, int]] = "auto"):
         r"""
@@ -1484,10 +1482,9 @@ def disable_attention_slicing(self):
         self.enable_attention_slicing(None)
 
     def set_attention_slice(self, slice_size: Optional[int]):
-        module_names, _, _ = self.extract_init_dict(dict(self.config))
-        module_names = [m for m in module_names if hasattr(self, m)]
+        module_names, _ = self._get_signature_keys(self)
+        modules = [getattr(self, n, None) for n in module_names]
+        modules = [m for m in modules if isinstance(m, torch.nn.Module) and hasattr(m, "set_attention_slice")]
 
-        for module_name in module_names:
-            module = getattr(self, module_name)
-            if isinstance(module, torch.nn.Module) and hasattr(module, "set_attention_slice"):
-                module.set_attention_slice(slice_size)
+        for module in modules:
+            module.set_attention_slice(slice_size)

src/diffusers/pipelines/text_to_video_synthesis/pipeline_text_to_video_zero.py

@@ -441,7 +441,7 @@ def __call__(
         timesteps = self.scheduler.timesteps
 
         # Prepare latent variables
-        num_channels_latents = self.unet.in_channels
+        num_channels_latents = self.unet.config.in_channels
         latents = self.prepare_latents(
             batch_size * num_videos_per_prompt,
             num_channels_latents,

src/diffusers/pipelines/unclip/pipeline_unclip.py

@@ -413,9 +413,9 @@ def __call__(
         self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
         decoder_timesteps_tensor = self.decoder_scheduler.timesteps
 
-        num_channels_latents = self.decoder.in_channels
-        height = self.decoder.sample_size
-        width = self.decoder.sample_size
+        num_channels_latents = self.decoder.config.in_channels
+        height = self.decoder.config.sample_size
+        width = self.decoder.config.sample_size
 
         decoder_latents = self.prepare_latents(
             (batch_size, num_channels_latents, height, width),
@@ -466,9 +466,9 @@ def __call__(
         self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
         super_res_timesteps_tensor = self.super_res_scheduler.timesteps
 
-        channels = self.super_res_first.in_channels // 2
-        height = self.super_res_first.sample_size
-        width = self.super_res_first.sample_size
+        channels = self.super_res_first.config.in_channels // 2
+        height = self.super_res_first.config.sample_size
+        width = self.super_res_first.config.sample_size
 
         super_res_latents = self.prepare_latents(
             (batch_size, channels, height, width),

src/diffusers/pipelines/unclip/pipeline_unclip_image_variation.py

@@ -339,9 +339,9 @@ def __call__(
         self.decoder_scheduler.set_timesteps(decoder_num_inference_steps, device=device)
         decoder_timesteps_tensor = self.decoder_scheduler.timesteps
 
-        num_channels_latents = self.decoder.in_channels
-        height = self.decoder.sample_size
-        width = self.decoder.sample_size
+        num_channels_latents = self.decoder.config.in_channels
+        height = self.decoder.config.sample_size
+        width = self.decoder.config.sample_size
 
         if decoder_latents is None:
             decoder_latents = self.prepare_latents(
@@ -393,9 +393,9 @@ def __call__(
         self.super_res_scheduler.set_timesteps(super_res_num_inference_steps, device=device)
         super_res_timesteps_tensor = self.super_res_scheduler.timesteps
 
-        channels = self.super_res_first.in_channels // 2
-        height = self.super_res_first.sample_size
-        width = self.super_res_first.sample_size
+        channels = self.super_res_first.config.in_channels // 2
+        height = self.super_res_first.config.sample_size
+        width = self.super_res_first.config.sample_size
 
         if super_res_latents is None:
             super_res_latents = self.prepare_latents(