Lazy Import for Diffusers (#4829)

* initial commit

* move modules to import struct

* add dummy objects and _LazyModule

* add lazy import to schedulers

* clean up unused imports

* lazy import on models module

* lazy import for schedulers module

* add lazy import to pipelines module

* lazy import altdiffusion

* lazy import audio diffusion

* lazy import audioldm

* lazy import consistency model

* lazy import controlnet

* lazy import dance diffusion ddim ddpm

* lazy import deepfloyd

* lazy import kandinksy

* lazy imports

* lazy import semantic diffusion

* lazy imports

* lazy import stable diffusion

* move sd output to its own module

* clean up

* lazy import t2iadapter

* lazy import unclip

* lazy import versatile and vq diffsuion

* lazy import vq diffusion

* helper to fetch objects from modules

* lazy import sdxl

* lazy import txt2vid

* lazy import stochastic karras

* fix model imports

* fix bug

* lazy import

* clean up

* clean up

* fixes for tests

* fixes for tests

* clean up

* remove import of torch_utils from utils module

* clean up

* clean up

* fix mistake import statement

* dedicated modules for exporting and loading

* remove testing utils from utils module

* fixes from  merge conflicts

* Update src/diffusers/pipelines/kandinsky2_2/__init__.py

* fix docs

* fix alt diffusion copied from

* fix check dummies

* fix more docs

* remove accelerate import from utils module

* add type checking

* make style

* fix check dummies

* remove torch import from xformers check

* clean up error message

* fixes after upstream merges

* dummy objects fix

* fix tests

* remove unused module import

---------
Co-authored-by: Patrick von Platen <patrick.v.platen@gmail.com>

docs/source/en/api/utilities.md

@@ -2,30 +2,26 @@
 
 Utility and helper functions for working with 🤗 Diffusers.
 
-## randn_tensor
-
-[[autodoc]] diffusers.utils.randn_tensor
-
 ## numpy_to_pil
 
-[[autodoc]] utils.pil_utils.numpy_to_pil
+[[autodoc]] utils.numpy_to_pil
 
 ## pt_to_pil
 
-[[autodoc]] utils.pil_utils.pt_to_pil
+[[autodoc]] utils.pt_to_pil
 
 ## load_image
 
-[[autodoc]] utils.testing_utils.load_image
+[[autodoc]] utils.load_image
 
 ## export_to_gif
 
-[[autodoc]] utils.testing_utils.export_to_gif
+[[autodoc]] utils.export_to_gif
 
 ## export_to_video
 
-[[autodoc]] utils.testing_utils.export_to_video
+[[autodoc]] utils.export_to_video
 
 ## make_image_grid
 
-[[autodoc]] utils.pil_utils.make_image_grid
\ No newline at end of file
+[[autodoc]] utils.pil_utils.make_image_grid

docs/source/en/using-diffusers/reproducibility.md

@@ -28,7 +28,7 @@ This is why it's important to understand how to control sources of randomness in
 
 ## Control randomness
 
-During inference, pipelines rely heavily on random sampling operations which include creating the 
+During inference, pipelines rely heavily on random sampling operations which include creating the
 Gaussian noise tensors to denoise and adding noise to the scheduling step.
 
 Take a look at the tensor values in the [`DDIMPipeline`] after two inference steps:
@@ -47,7 +47,7 @@ image = ddim(num_inference_steps=2, output_type="np").images
 print(np.abs(image).sum())
 ```
 
-Running the code above prints one value, but if you run it again you get a different value. What is going on here? 
+Running the code above prints one value, but if you run it again you get a different value. What is going on here?
 
 Every time the pipeline is run, [`torch.randn`](https://pytorch.org/docs/stable/generated/torch.randn.html) uses a different random seed to create Gaussian noise which is denoised stepwise. This leads to a different result each time it is run, which is great for diffusion pipelines since it generates a different random image each time.
 
@@ -81,16 +81,16 @@ If you run this code example on your specific hardware and PyTorch version, you
 
 <Tip>
 
-💡 It might be a bit unintuitive at first to pass `Generator` objects to the pipeline instead of 
-just integer values representing the seed, but this is the recommended design when dealing with 
-probabilistic models in PyTorch as `Generator`'s are *random states* that can be 
+💡 It might be a bit unintuitive at first to pass `Generator` objects to the pipeline instead of
+just integer values representing the seed, but this is the recommended design when dealing with
+probabilistic models in PyTorch as `Generator`'s are *random states* that can be
 passed to multiple pipelines in a sequence.
 
 </Tip>
 
 ### GPU
 
-Writing a reproducible pipeline on a GPU is a bit trickier, and full reproducibility across different hardware is not guaranteed because matrix multiplication - which diffusion pipelines require a lot of - is less deterministic on a GPU than a CPU. For example, if you run the same code example above on a GPU: 
+Writing a reproducible pipeline on a GPU is a bit trickier, and full reproducibility across different hardware is not guaranteed because matrix multiplication - which diffusion pipelines require a lot of - is less deterministic on a GPU than a CPU. For example, if you run the same code example above on a GPU:
 
 ```python
 import torch
@@ -113,7 +113,7 @@ print(np.abs(image).sum())
 
 The result is not the same even though you're using an identical seed because the GPU uses a different random number generator than the CPU.
 
-To circumvent this problem, 🧨 Diffusers has a [`~diffusers.utils.randn_tensor`] function for creating random noise on the CPU, and then moving the tensor to a GPU if necessary. The `randn_tensor` function is used everywhere inside the pipeline, allowing the user to **always** pass a CPU `Generator` even if the pipeline is run on a GPU. 
+To circumvent this problem, 🧨 Diffusers has a [`~diffusers.utils.torch_utils.randn_tensor`] function for creating random noise on the CPU, and then moving the tensor to a GPU if necessary. The `randn_tensor` function is used everywhere inside the pipeline, allowing the user to **always** pass a CPU `Generator` even if the pipeline is run on a GPU.
 
 You'll see the results are much closer now!
 
@@ -139,14 +139,14 @@ print(np.abs(image).sum())
 <Tip>
 
 💡 If reproducibility is important, we recommend always passing a CPU generator.
-The performance loss is often neglectable, and you'll generate much more similar 
+The performance loss is often neglectable, and you'll generate much more similar
 values than if the pipeline had been run on a GPU.
 
 </Tip>
 
-Finally, for more complex pipelines such as [`UnCLIPPipeline`], these are often extremely 
-susceptible to precision error propagation. Don't expect similar results across 
-different GPU hardware or PyTorch versions. In this case, you'll need to run 
+Finally, for more complex pipelines such as [`UnCLIPPipeline`], these are often extremely
+susceptible to precision error propagation. Don't expect similar results across
+different GPU hardware or PyTorch versions. In this case, you'll need to run
 exactly the same hardware and PyTorch version for full reproducibility.
 
 ## Deterministic algorithms

examples/community/clip_guided_images_mixing_stable_diffusion.py

@@ -19,10 +19,8 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.utils import (
-    PIL_INTERPOLATION,
-    randn_tensor,
-)
+from diffusers.utils import PIL_INTERPOLATION
+from diffusers.utils.torch_utils import randn_tensor
 
 
 def preprocess(image, w, h):

examples/community/clip_guided_stable_diffusion_img2img.py

@@ -19,11 +19,8 @@ from diffusers import (
     UNet2DConditionModel,
 )
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.utils import (
-    PIL_INTERPOLATION,
-    deprecate,
-    randn_tensor,
-)
+from diffusers.utils import PIL_INTERPOLATION, deprecate
+from diffusers.utils.torch_utils import randn_tensor
 
 
 EXAMPLE_DOC_STRING = """

examples/community/ddim_noise_comparative_analysis.py

@@ -20,7 +20,7 @@ from torchvision import transforms
 
 from diffusers.pipeline_utils import DiffusionPipeline, ImagePipelineOutput
 from diffusers.schedulers import DDIMScheduler
-from diffusers.utils import randn_tensor
+from diffusers.utils.torch_utils import randn_tensor
 
 
 trans = transforms.Compose(

examples/community/lpw_stable_diffusion.py

@@ -21,8 +21,8 @@ from diffusers.utils import (
     is_accelerate_available,
     is_accelerate_version,
     logging,
-    randn_tensor,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 # ------------------------------------------------------------------------------

examples/community/lpw_stable_diffusion_xl.py

@@ -30,9 +30,9 @@ from diffusers.utils import (
     is_accelerate_version,
     is_invisible_watermark_available,
     logging,
-    randn_tensor,
     replace_example_docstring,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 if is_invisible_watermark_available():

examples/community/pipeline_fabric.py

@@ -14,6 +14,7 @@
 from typing import List, Optional, Union
 
 import torch
+from diffuser.utils.torch_utils import randn_tensor
 from packaging import version
 from PIL import Image
 from transformers import CLIPTextModel, CLIPTokenizer
@@ -30,7 +31,6 @@ from diffusers.schedulers import EulerAncestralDiscreteScheduler, KarrasDiffusio
 from diffusers.utils import (
     deprecate,
     logging,
-    randn_tensor,
     replace_example_docstring,
 )
 

examples/community/pipeline_zero1to3.py

@@ -35,9 +35,9 @@ from diffusers.utils import (
     is_accelerate_available,
     is_accelerate_version,
     logging,
-    randn_tensor,
     replace_example_docstring,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/run_onnx_controlnet.py

@@ -8,6 +8,7 @@ from typing import Any, Callable, Dict, List, Optional, Union
 import numpy as np
 import PIL.Image
 import torch
+from diffuser.utils.torch_utils import randn_tensor
 from PIL import Image
 from transformers import CLIPTokenizer
 
@@ -19,7 +20,6 @@ from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     deprecate,
     logging,
-    randn_tensor,
     replace_example_docstring,
 )
 

examples/community/run_tensorrt_controlnet.py

@@ -11,6 +11,7 @@ import PIL.Image
 import pycuda.driver as cuda
 import tensorrt as trt
 import torch
+from diffuser.utils.torch_utils import randn_tensor
 from PIL import Image
 from pycuda.tools import make_default_context
 from transformers import CLIPTokenizer
@@ -23,7 +24,6 @@ from diffusers.schedulers import KarrasDiffusionSchedulers
 from diffusers.utils import (
     deprecate,
     logging,
-    randn_tensor,
     replace_example_docstring,
 )
 

examples/community/stable_diffusion_controlnet_img2img.py

@@ -16,9 +16,9 @@ from diffusers.utils import (
     PIL_INTERPOLATION,
     is_accelerate_available,
     is_accelerate_version,
-    randn_tensor,
     replace_example_docstring,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_controlnet_inpaint.py

@@ -17,9 +17,9 @@ from diffusers.utils import (
     PIL_INTERPOLATION,
     is_accelerate_available,
     is_accelerate_version,
-    randn_tensor,
     replace_example_docstring,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_controlnet_inpaint_img2img.py

@@ -16,9 +16,9 @@ from diffusers.utils import (
     PIL_INTERPOLATION,
     is_accelerate_available,
     is_accelerate_version,
-    randn_tensor,
     replace_example_docstring,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_controlnet_reference.py

@@ -11,7 +11,8 @@ from diffusers.models.attention import BasicTransformerBlock
 from diffusers.models.unet_2d_blocks import CrossAttnDownBlock2D, CrossAttnUpBlock2D, DownBlock2D, UpBlock2D
 from diffusers.pipelines.controlnet.multicontrolnet import MultiControlNetModel
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
-from diffusers.utils import is_compiled_module, logging, randn_tensor
+from diffusers.utils import logging
+from diffusers.utils.torch_utils import is_compiled_module, randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_ipex.py

@@ -31,9 +31,9 @@ from diffusers.utils import (
     is_accelerate_available,
     is_accelerate_version,
     logging,
-    randn_tensor,
     replace_example_docstring,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_reference.py

@@ -10,7 +10,8 @@ from diffusers.models.attention import BasicTransformerBlock
 from diffusers.models.unet_2d_blocks import CrossAttnDownBlock2D, CrossAttnUpBlock2D, DownBlock2D, UpBlock2D
 from diffusers.pipelines.stable_diffusion import StableDiffusionPipelineOutput
 from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion import rescale_noise_cfg
-from diffusers.utils import PIL_INTERPOLATION, logging, randn_tensor
+from diffusers.utils import PIL_INTERPOLATION, logging
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_repaint.py

@@ -33,8 +33,8 @@ from diffusers.utils import (
     is_accelerate_available,
     is_accelerate_version,
     logging,
-    randn_tensor,
 )
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_diffusion_xl_reference.py

@@ -15,7 +15,8 @@ from diffusers.models.unet_2d_blocks import (
     UpBlock2D,
 )
 from diffusers.pipelines.stable_diffusion_xl import StableDiffusionXLPipelineOutput
-from diffusers.utils import PIL_INTERPOLATION, logging, randn_tensor
+from diffusers.utils import PIL_INTERPOLATION, logging
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name

examples/community/stable_unclip.py

@@ -8,7 +8,8 @@ from transformers.models.clip.modeling_clip import CLIPTextModelOutput
 from diffusers.models import PriorTransformer
 from diffusers.pipelines import DiffusionPipeline, StableDiffusionImageVariationPipeline
 from diffusers.schedulers import UnCLIPScheduler
-from diffusers.utils import logging, randn_tensor
+from diffusers.utils import logging
+from diffusers.utils.torch_utils import randn_tensor
 
 
 logger = logging.get_logger(__name__)  # pylint: disable=invalid-name