update

docker/diffusers-pytorch-compile-cuda/Dockerfile

@@ -0,0 +1,50 @@
+FROM nvidia/cuda:12.1.0-runtime-ubuntu20.04
+LABEL maintainer="Hugging Face"
+LABEL repository="diffusers"
+
+ENV DEBIAN_FRONTEND=noninteractive
+
+RUN apt-get -y update \
+    && apt-get install -y software-properties-common \
+    && add-apt-repository ppa:deadsnakes/ppa
+
+RUN apt install -y bash \
+    build-essential \
+    git \
+    git-lfs \
+    curl \
+    ca-certificates \
+    libsndfile1-dev \
+    libgl1 \
+    python3.10 \
+    python3.10-dev \
+    python3-pip \
+    python3.10-venv && \
+    rm -rf /var/lib/apt/lists
+
+# make sure to use venv
+RUN python3.10 -m venv /opt/venv
+ENV PATH="/opt/venv/bin:$PATH"
+
+# pre-install the heavy dependencies (these can later be overridden by the deps from setup.py)
+RUN python3.10 -m pip install --no-cache-dir --upgrade pip uv==0.1.11 && \
+    python3.10 -m uv pip install --no-cache-dir \
+    torch \
+    torchvision \
+    torchaudio \
+    invisible_watermark && \
+    python3.10 -m pip install --no-cache-dir \
+    accelerate \
+    datasets \
+    hf-doc-builder \
+    huggingface-hub \
+    hf_transfer \
+    Jinja2 \
+    librosa \
+    numpy==1.26.4 \
+    scipy \
+    tensorboard \
+    transformers \
+    hf_transfer
+
+CMD ["/bin/bash"]

docs/source/en/optimization/torch2.0.md

@@ -0,0 +1,421 @@
+<!--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.
+-->
+
+# PyTorch 2.0
+
+🤗 Diffusers supports the latest optimizations from [PyTorch 2.0](https://pytorch.org/get-started/pytorch-2.0/) which include:
+
+1. A memory-efficient attention implementation, scaled dot product attention, without requiring any extra dependencies such as xFormers.
+2. [`torch.compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html), a just-in-time (JIT) compiler to provide an extra performance boost when individual models are compiled.
+
+Both of these optimizations require PyTorch 2.0 or later and 🤗 Diffusers > 0.13.0.
+
+```bash
+pip install --upgrade torch diffusers
+```
+
+## Scaled dot product attention
+
+[`torch.nn.functional.scaled_dot_product_attention`](https://pytorch.org/docs/master/generated/torch.nn.functional.scaled_dot_product_attention) (SDPA) is an optimized and memory-efficient attention (similar to xFormers) that automatically enables several other optimizations depending on the model inputs and GPU type. SDPA is enabled by default if you're using PyTorch 2.0 and the latest version of 🤗 Diffusers, so you don't need to add anything to your code.
+
+However, if you want to explicitly enable it, you can set a [`DiffusionPipeline`] to use [`~models.attention_processor.AttnProcessor2_0`]:
+
+```diff
+  import torch
+  from diffusers import DiffusionPipeline
++ from diffusers.models.attention_processor import AttnProcessor2_0
+
+  pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
++ pipe.unet.set_attn_processor(AttnProcessor2_0())
+
+  prompt = "a photo of an astronaut riding a horse on mars"
+  image = pipe(prompt).images[0]
+```
+
+SDPA should be as fast and memory efficient as `xFormers`; check the [benchmark](#benchmark) for more details.
+
+In some cases - such as making the pipeline more deterministic or converting it to other formats - it may be helpful to use the vanilla attention processor, [`~models.attention_processor.AttnProcessor`]. To revert to [`~models.attention_processor.AttnProcessor`], call the [`~UNet2DConditionModel.set_default_attn_processor`] function on the pipeline:
+
+```diff
+  import torch
+  from diffusers import DiffusionPipeline
+
+  pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
++ pipe.unet.set_default_attn_processor()
+
+  prompt = "a photo of an astronaut riding a horse on mars"
+  image = pipe(prompt).images[0]
+```
+
+## torch.compile
+
+The `torch.compile` function can often provide an additional speed-up to your PyTorch code. In 🤗 Diffusers, it is usually best to wrap the UNet with `torch.compile` because it does most of the heavy lifting in the pipeline.
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+pipe = DiffusionPipeline.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True).to("cuda")
+pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+images = pipe(prompt, num_inference_steps=steps, num_images_per_prompt=batch_size).images[0]
+```
+
+Depending on GPU type, `torch.compile` can provide an *additional speed-up* of **5-300x** on top of SDPA! If you're using more recent GPU architectures such as Ampere (A100, 3090), Ada (4090), and Hopper (H100), `torch.compile` is able to squeeze even more performance out of these GPUs.
+
+Compilation requires some time to complete, so it is best suited for situations where you prepare your pipeline once and then perform the same type of inference operations multiple times. For example, calling the compiled pipeline on a different image size triggers compilation again which can be expensive.
+
+For more information and different options about `torch.compile`, refer to the [`torch_compile`](https://pytorch.org/tutorials/intermediate/torch_compile_tutorial.html) tutorial.
+
+> [!TIP]
+> Learn more about other ways PyTorch 2.0 can help optimize your model in the [Accelerate inference of text-to-image diffusion models](../tutorials/fast_diffusion) tutorial.
+
+## Benchmark
+
+We conducted a comprehensive benchmark with PyTorch 2.0's efficient attention implementation and `torch.compile` across different GPUs and batch sizes for five of our most used pipelines. The code is benchmarked on 🤗 Diffusers v0.17.0.dev0 to optimize `torch.compile` usage (see [here](https://github.com/huggingface/diffusers/pull/3313) for more details).
+
+Expand the dropdown below to find the code used to benchmark each pipeline:
+
+<details>
+
+### Stable Diffusion text-to-image
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+path = "stable-diffusion-v1-5/stable-diffusion-v1-5"
+
+run_compile = True  # Set True / False
+
+pipe = DiffusionPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True)
+pipe = pipe.to("cuda")
+pipe.unet.to(memory_format=torch.channels_last)
+
+if run_compile:
+    print("Run torch compile")
+    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+prompt = "ghibli style, a fantasy landscape with castles"
+
+for _ in range(3):
+    images = pipe(prompt=prompt).images
+```
+
+### Stable Diffusion image-to-image
+
+```python
+from diffusers import StableDiffusionImg2ImgPipeline
+from diffusers.utils import load_image
+import torch
+
+url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+init_image = load_image(url)
+init_image = init_image.resize((512, 512))
+
+path = "stable-diffusion-v1-5/stable-diffusion-v1-5"
+
+run_compile = True  # Set True / False
+
+pipe = StableDiffusionImg2ImgPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True)
+pipe = pipe.to("cuda")
+pipe.unet.to(memory_format=torch.channels_last)
+
+if run_compile:
+    print("Run torch compile")
+    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+prompt = "ghibli style, a fantasy landscape with castles"
+
+for _ in range(3):
+    image = pipe(prompt=prompt, image=init_image).images[0]
+```
+
+### Stable Diffusion inpainting
+
+```python
+from diffusers import StableDiffusionInpaintPipeline
+from diffusers.utils import load_image
+import torch
+
+img_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo.png"
+mask_url = "https://raw.githubusercontent.com/CompVis/latent-diffusion/main/data/inpainting_examples/overture-creations-5sI6fQgYIuo_mask.png"
+
+init_image = load_image(img_url).resize((512, 512))
+mask_image = load_image(mask_url).resize((512, 512))
+
+path = "runwayml/stable-diffusion-inpainting"
+
+run_compile = True  # Set True / False
+
+pipe = StableDiffusionInpaintPipeline.from_pretrained(path, torch_dtype=torch.float16, use_safetensors=True)
+pipe = pipe.to("cuda")
+pipe.unet.to(memory_format=torch.channels_last)
+
+if run_compile:
+    print("Run torch compile")
+    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+
+prompt = "ghibli style, a fantasy landscape with castles"
+
+for _ in range(3):
+    image = pipe(prompt=prompt, image=init_image, mask_image=mask_image).images[0]
+```
+
+### ControlNet
+
+```python
+from diffusers import StableDiffusionControlNetPipeline, ControlNetModel
+from diffusers.utils import load_image
+import torch
+
+url = "https://raw.githubusercontent.com/CompVis/stable-diffusion/main/assets/stable-samples/img2img/sketch-mountains-input.jpg"
+
+init_image = load_image(url)
+init_image = init_image.resize((512, 512))
+
+path = "stable-diffusion-v1-5/stable-diffusion-v1-5"
+
+run_compile = True  # Set True / False
+controlnet = ControlNetModel.from_pretrained("lllyasviel/sd-controlnet-canny", torch_dtype=torch.float16, use_safetensors=True)
+pipe = StableDiffusionControlNetPipeline.from_pretrained(
+    path, controlnet=controlnet, torch_dtype=torch.float16, use_safetensors=True
+)
+
+pipe = pipe.to("cuda")
+pipe.unet.to(memory_format=torch.channels_last)
+pipe.controlnet.to(memory_format=torch.channels_last)
+
+if run_compile:
+    print("Run torch compile")
+    pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
+    pipe.controlnet = torch.compile(pipe.controlnet, mode="reduce-overhead", fullgraph=True)
+
+prompt = "ghibli style, a fantasy landscape with castles"
+
+for _ in range(3):
+    image = pipe(prompt=prompt, image=init_image).images[0]
+```
+
+### DeepFloyd IF text-to-image + upscaling
+
+```python
+from diffusers import DiffusionPipeline
+import torch
+
+run_compile = True  # Set True / False
+
+pipe_1 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-I-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16, use_safetensors=True)
+pipe_1.to("cuda")
+pipe_2 = DiffusionPipeline.from_pretrained("DeepFloyd/IF-II-M-v1.0", variant="fp16", text_encoder=None, torch_dtype=torch.float16, use_safetensors=True)
+pipe_2.to("cuda")
+pipe_3 = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-x4-upscaler", torch_dtype=torch.float16, use_safetensors=True)
+pipe_3.to("cuda")
+
+
+pipe_1.unet.to(memory_format=torch.channels_last)
+pipe_2.unet.to(memory_format=torch.channels_last)
+pipe_3.unet.to(memory_format=torch.channels_last)
+
+if run_compile:
+    pipe_1.unet = torch.compile(pipe_1.unet, mode="reduce-overhead", fullgraph=True)
+    pipe_2.unet = torch.compile(pipe_2.unet, mode="reduce-overhead", fullgraph=True)
+    pipe_3.unet = torch.compile(pipe_3.unet, mode="reduce-overhead", fullgraph=True)
+
+prompt = "the blue hulk"
+
+prompt_embeds = torch.randn((1, 2, 4096), dtype=torch.float16)
+neg_prompt_embeds = torch.randn((1, 2, 4096), dtype=torch.float16)
+
+for _ in range(3):
+    image_1 = pipe_1(prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images
+    image_2 = pipe_2(image=image_1, prompt_embeds=prompt_embeds, negative_prompt_embeds=neg_prompt_embeds, output_type="pt").images
+    image_3 = pipe_3(prompt=prompt, image=image_1, noise_level=100).images
+```
+</details>
+
+The graph below highlights the relative speed-ups for the [`StableDiffusionPipeline`] across five GPU families with PyTorch 2.0 and `torch.compile` enabled. The benchmarks for the following graphs are measured in *number of iterations/second*.
+
+![t2i_speedup](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/pt2_benchmarks/t2i_speedup.png)
+
+To give you an even better idea of how this speed-up holds for the other pipelines, consider the following
+graph for an A100 with PyTorch 2.0 and `torch.compile`:
+
+![a100_numbers](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/pt2_benchmarks/a100_numbers.png)
+
+In the following tables, we report our findings in terms of the *number of iterations/second*.
+
+### A100 (batch size: 1)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 21.66 | 23.13 | 44.03 | 49.74 |
+| SD - img2img | 21.81 | 22.40 | 43.92 | 46.32 |
+| SD - inpaint | 22.24 | 23.23 | 43.76 | 49.25 |
+| SD - controlnet | 15.02 | 15.82 | 32.13 | 36.08 |
+| IF | 20.21 / <br>13.84 / <br>24.00 | 20.12 / <br>13.70 / <br>24.03 | ❌ | 97.34 / <br>27.23 / <br>111.66 |
+| SDXL - txt2img | 8.64 | 9.9 | - | - |
+
+### A100 (batch size: 4)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 11.6 | 13.12 | 14.62 | 17.27 |
+| SD - img2img | 11.47 | 13.06 | 14.66 | 17.25 |
+| SD - inpaint | 11.67 | 13.31 | 14.88 | 17.48 |
+| SD - controlnet | 8.28 | 9.38 | 10.51 | 12.41 |
+| IF | 25.02 | 18.04 | ❌ | 48.47 |
+| SDXL - txt2img | 2.44 | 2.74 | - | - |
+
+### A100 (batch size: 16)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 3.04 | 3.6 | 3.83 | 4.68 |
+| SD - img2img | 2.98 | 3.58 | 3.83 | 4.67 |
+| SD - inpaint | 3.04 | 3.66 | 3.9 | 4.76 |
+| SD - controlnet | 2.15 | 2.58 | 2.74 | 3.35 |
+| IF | 8.78 | 9.82 | ❌ | 16.77 |
+| SDXL - txt2img | 0.64 | 0.72 | - | - |
+
+### V100 (batch size: 1)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 18.99 | 19.14 | 20.95 | 22.17 |
+| SD - img2img | 18.56 | 19.18 | 20.95 | 22.11 |
+| SD - inpaint | 19.14 | 19.06 | 21.08 | 22.20 |
+| SD - controlnet | 13.48 | 13.93 | 15.18 | 15.88 |
+| IF |  20.01 / <br>9.08 / <br>23.34 | 19.79 / <br>8.98 / <br>24.10 | ❌ | 55.75 / <br>11.57 / <br>57.67 |
+
+### V100 (batch size: 4)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 5.96 | 5.89 | 6.83 | 6.86 |
+| SD - img2img | 5.90 | 5.91 | 6.81 | 6.82 |
+| SD - inpaint | 5.99 | 6.03 | 6.93 | 6.95 |
+| SD - controlnet | 4.26 | 4.29 | 4.92 | 4.93 |
+| IF | 15.41 | 14.76 | ❌ | 22.95 |
+
+### V100 (batch size: 16)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 1.66 | 1.66 | 1.92 | 1.90 |
+| SD - img2img | 1.65 | 1.65 | 1.91 | 1.89 |
+| SD - inpaint | 1.69 | 1.69 | 1.95 | 1.93 |
+| SD - controlnet | 1.19 | 1.19 | OOM after warmup | 1.36 |
+| IF | 5.43 | 5.29 | ❌ | 7.06 |
+
+### T4 (batch size: 1)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 6.9 | 6.95 | 7.3 | 7.56 |
+| SD - img2img | 6.84 | 6.99 | 7.04 | 7.55 |
+| SD - inpaint | 6.91 | 6.7 | 7.01 | 7.37 |
+| SD - controlnet | 4.89 | 4.86 | 5.35 | 5.48 |
+| IF | 17.42 / <br>2.47 / <br>18.52 | 16.96 / <br>2.45 / <br>18.69 | ❌ | 24.63 / <br>2.47 / <br>23.39 |
+| SDXL - txt2img | 1.15 | 1.16 | - | - |
+
+### T4 (batch size: 4)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 1.79 | 1.79 | 2.03 | 1.99 |
+| SD - img2img | 1.77 | 1.77 | 2.05 | 2.04 |
+| SD - inpaint | 1.81 | 1.82 | 2.09 | 2.09 |
+| SD - controlnet | 1.34 | 1.27 | 1.47 | 1.46 |
+| IF | 5.79 |  5.61 | ❌ | 7.39 |
+| SDXL - txt2img | 0.288 | 0.289 | - | - |
+
+### T4 (batch size: 16)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 2.34s | 2.30s | OOM after 2nd iteration | 1.99s |
+| SD - img2img | 2.35s | 2.31s | OOM after warmup | 2.00s |
+| SD - inpaint | 2.30s | 2.26s | OOM after 2nd iteration | 1.95s |
+| SD - controlnet | OOM after 2nd iteration | OOM after 2nd iteration | OOM after warmup | OOM after warmup |
+| IF * | 1.44 | 1.44 | ❌ | 1.94 |
+| SDXL - txt2img | OOM | OOM | - | - |
+
+### RTX 3090 (batch size: 1)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 22.56 | 22.84 | 23.84 | 25.69 |
+| SD - img2img | 22.25 | 22.61 | 24.1 | 25.83 |
+| SD - inpaint | 22.22 | 22.54 | 24.26 | 26.02 |
+| SD - controlnet | 16.03 | 16.33 | 17.38 | 18.56 |
+| IF | 27.08 / <br>9.07 / <br>31.23 | 26.75 / <br>8.92 / <br>31.47 | ❌ | 68.08 / <br>11.16 / <br>65.29 |
+
+### RTX 3090 (batch size: 4)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 6.46 | 6.35 | 7.29 | 7.3 |
+| SD - img2img | 6.33 | 6.27 | 7.31 | 7.26 |
+| SD - inpaint | 6.47 | 6.4 | 7.44 | 7.39 |
+| SD - controlnet | 4.59 | 4.54 | 5.27 | 5.26 |
+| IF | 16.81 | 16.62 | ❌ | 21.57 |
+
+### RTX 3090 (batch size: 16)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 1.7 | 1.69 | 1.93 | 1.91 |
+| SD - img2img | 1.68 | 1.67 | 1.93 | 1.9 |
+| SD - inpaint | 1.72 | 1.71 | 1.97 | 1.94 |
+| SD - controlnet | 1.23 | 1.22 | 1.4 | 1.38 |
+| IF | 5.01 | 5.00 | ❌ | 6.33 |
+
+### RTX 4090 (batch size: 1)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 40.5 | 41.89 | 44.65 | 49.81 |
+| SD - img2img | 40.39 | 41.95 | 44.46 | 49.8 |
+| SD - inpaint | 40.51 | 41.88 | 44.58 | 49.72 |
+| SD - controlnet | 29.27 | 30.29 | 32.26 | 36.03 |
+| IF | 69.71 / <br>18.78 / <br>85.49 | 69.13 / <br>18.80 / <br>85.56 | ❌ | 124.60 / <br>26.37 / <br>138.79 |
+| SDXL - txt2img | 6.8 | 8.18 | - | - |
+
+### RTX 4090 (batch size: 4)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 12.62 | 12.84 | 15.32 | 15.59 |
+| SD - img2img | 12.61 | 12,.79 | 15.35 | 15.66 |
+| SD - inpaint | 12.65 | 12.81 | 15.3 | 15.58 |
+| SD - controlnet | 9.1 | 9.25 | 11.03 | 11.22 |
+| IF | 31.88 | 31.14 | ❌ | 43.92 |
+| SDXL - txt2img | 2.19 | 2.35 | - | - |
+
+### RTX 4090 (batch size: 16)
+
+| **Pipeline** | **torch 2.0 - <br>no compile** | **torch nightly - <br>no compile** | **torch 2.0 - <br>compile** | **torch nightly - <br>compile** |
+|:---:|:---:|:---:|:---:|:---:|
+| SD - txt2img | 3.17 | 3.2 | 3.84 | 3.85 |
+| SD - img2img | 3.16 | 3.2 | 3.84 | 3.85 |
+| SD - inpaint | 3.17 | 3.2 | 3.85 | 3.85 |
+| SD - controlnet | 2.23 | 2.3 | 2.7 | 2.75 |
+| IF | 9.26 | 9.2 | ❌ | 13.31 |
+| SDXL - txt2img | 0.52 | 0.53 | - | - |
+
+## Notes
+
+* Follow this [PR](https://github.com/huggingface/diffusers/pull/3313) for more details on the environment used for conducting the benchmarks.
+* For the DeepFloyd IF pipeline where batch sizes > 1, we only used a batch size of > 1 in the first IF pipeline for text-to-image generation and NOT for upscaling. That means the two upscaling pipelines received a batch size of 1.
+
+*Thanks to [Horace He](https://github.com/Chillee) from the PyTorch team for their support in improving our support of `torch.compile()` in Diffusers.*

docs/source/en/tutorials/fast_diffusion.md

@@ -0,0 +1,322 @@
+<!--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.
+-->
+
+# Accelerate inference of text-to-image diffusion models
+
+Diffusion models are slower than their GAN counterparts because of the iterative and sequential reverse diffusion process. There are several techniques that can address this limitation such as progressive timestep distillation ([LCM LoRA](../using-diffusers/inference_with_lcm_lora)), model compression ([SSD-1B](https://huggingface.co/segmind/SSD-1B)), and reusing adjacent features of the denoiser ([DeepCache](../optimization/deepcache)).
+
+However, you don't necessarily need to use these techniques to speed up inference. With PyTorch 2 alone, you can accelerate the inference latency of text-to-image diffusion pipelines by up to 3x. This tutorial will show you how to progressively apply the optimizations found in PyTorch 2 to reduce inference latency. You'll use the [Stable Diffusion XL (SDXL)](../using-diffusers/sdxl) pipeline in this tutorial, but these techniques are applicable to other text-to-image diffusion pipelines too.
+
+Make sure you're using the latest version of Diffusers:
+
+```bash
+pip install -U diffusers
+```
+
+Then upgrade the other required libraries too:
+
+```bash
+pip install -U transformers accelerate peft
+```
+
+Install [PyTorch nightly](https://pytorch.org/) to benefit from the latest and fastest kernels:
+
+```bash
+pip3 install --pre torch --index-url https://download.pytorch.org/whl/nightly/cu121
+```
+
+> [!TIP]
+> The results reported below are from a 80GB 400W A100 with its clock rate set to the maximum.
+> If you're interested in the full benchmarking code, take a look at [huggingface/diffusion-fast](https://github.com/huggingface/diffusion-fast).
+
+
+## Baseline
+
+Let's start with a baseline. Disable reduced precision and the [`scaled_dot_product_attention` (SDPA)](../optimization/torch2.0#scaled-dot-product-attention) function which is automatically used by Diffusers:
+
+```python
+from diffusers import StableDiffusionXLPipeline
+
+# Load the pipeline in full-precision and place its model components on CUDA.
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0"
+).to("cuda")
+
+# Run the attention ops without SDPA.
+pipe.unet.set_default_attn_processor()
+pipe.vae.set_default_attn_processor()
+
+prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+image = pipe(prompt, num_inference_steps=30).images[0]
+```
+
+This default setup takes 7.36 seconds.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/progressive-acceleration-sdxl/SDXL%2C_Batch_Size%3A_1%2C_Steps%3A_30_0.png" width=500>
+</div>
+
+## bfloat16
+
+Enable the first optimization, reduced precision or more specifically bfloat16. There are several benefits of using reduced precision:
+
+* Using a reduced numerical precision (such as float16 or bfloat16) for inference doesn’t affect the generation quality but significantly improves latency.
+* The benefits of using bfloat16 compared to float16 are hardware dependent, but modern GPUs tend to favor bfloat16.
+* bfloat16 is much more resilient when used with quantization compared to float16, but more recent versions of the quantization library ([torchao](https://github.com/pytorch-labs/ao)) we used don't have numerical issues with float16.
+
+```python
+from diffusers import StableDiffusionXLPipeline
+import torch
+
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.bfloat16
+).to("cuda")
+
+# Run the attention ops without SDPA.
+pipe.unet.set_default_attn_processor()
+pipe.vae.set_default_attn_processor()
+
+prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+image = pipe(prompt, num_inference_steps=30).images[0]
+```
+
+bfloat16 reduces the latency from 7.36 seconds to 4.63 seconds.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/progressive-acceleration-sdxl/SDXL%2C_Batch_Size%3A_1%2C_Steps%3A_30_1.png" width=500>
+</div>
+
+<Tip>
+
+In our later experiments with float16, recent versions of torchao do not incur numerical problems from float16.
+
+</Tip>
+
+Take a look at the [Speed up inference](../optimization/fp16) guide to learn more about running inference with reduced precision.
+
+## SDPA
+
+Attention blocks are intensive to run. But with PyTorch's [`scaled_dot_product_attention`](../optimization/torch2.0#scaled-dot-product-attention) function, it is a lot more efficient. This function is used by default in Diffusers so you don't need to make any changes to the code.
+
+```python
+from diffusers import StableDiffusionXLPipeline
+import torch
+
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.bfloat16
+).to("cuda")
+
+prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+image = pipe(prompt, num_inference_steps=30).images[0]
+```
+
+Scaled dot product attention improves the latency from 4.63 seconds to 3.31 seconds.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/progressive-acceleration-sdxl/SDXL%2C_Batch_Size%3A_1%2C_Steps%3A_30_2.png" width=500>
+</div>
+
+## torch.compile
+
+PyTorch 2 includes `torch.compile` which uses fast and optimized kernels. In Diffusers, the UNet and VAE are usually compiled because these are the most compute-intensive modules. First, configure a few compiler flags (refer to the [full list](https://github.com/pytorch/pytorch/blob/main/torch/_inductor/config.py) for more options):
+
+```python
+from diffusers import StableDiffusionXLPipeline
+import torch
+
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+```
+
+It is also important to change the UNet and VAE's memory layout to "channels_last" when compiling them to ensure maximum speed.
+
+```python
+pipe.unet.to(memory_format=torch.channels_last)
+pipe.vae.to(memory_format=torch.channels_last)
+```
+
+Now compile and perform inference:
+
+```python
+# Compile the UNet and VAE.
+pipe.unet = torch.compile(pipe.unet, mode="max-autotune", fullgraph=True)
+pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
+
+prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+
+# First call to `pipe` is slow, subsequent ones are faster.
+image = pipe(prompt, num_inference_steps=30).images[0]
+```
+
+`torch.compile` offers different backends and modes. For maximum inference speed, use "max-autotune" for the inductor backend. “max-autotune” uses CUDA graphs and optimizes the compilation graph specifically for latency. CUDA graphs greatly reduces the overhead of launching GPU operations by using a mechanism to launch multiple GPU operations through a single CPU operation.
+
+Using SDPA attention and compiling both the UNet and VAE cuts the latency from 3.31 seconds to 2.54 seconds.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/progressive-acceleration-sdxl/SDXL%2C_Batch_Size%3A_1%2C_Steps%3A_30_3.png" width=500>
+</div>
+
+> [!TIP]
+> From PyTorch 2.3.1, you can control the caching behavior of `torch.compile()`. This is particularly beneficial for compilation modes like `"max-autotune"` which performs a grid-search over several compilation flags to find the optimal configuration. Learn more in the [Compile Time Caching in torch.compile](https://pytorch.org/tutorials/recipes/torch_compile_caching_tutorial.html) tutorial.
+
+### Prevent graph breaks
+
+Specifying `fullgraph=True` ensures there are no graph breaks in the underlying model to take full advantage of `torch.compile` without any performance degradation. For the UNet and VAE, this means changing how you access the return variables.
+
+```diff
+- latents = unet(
+-   latents, timestep=timestep, encoder_hidden_states=prompt_embeds
+-).sample
+
++ latents = unet(
++   latents, timestep=timestep, encoder_hidden_states=prompt_embeds, return_dict=False
++)[0]
+```
+
+### Remove GPU sync after compilation
+
+During the iterative reverse diffusion process, the `step()` function is [called](https://github.com/huggingface/diffusers/blob/1d686bac8146037e97f3fd8c56e4063230f71751/src/diffusers/pipelines/stable_diffusion_xl/pipeline_stable_diffusion_xl.py#L1228) on the scheduler each time after the denoiser predicts the less noisy latent embeddings. Inside `step()`, the `sigmas` variable is [indexed](https://github.com/huggingface/diffusers/blob/1d686bac8146037e97f3fd8c56e4063230f71751/src/diffusers/schedulers/scheduling_euler_discrete.py#L476) which when placed on the GPU, causes a communication sync between the CPU and GPU. This introduces latency and it becomes more evident when the denoiser has already been compiled.
+
+But if the `sigmas` array always [stays on the CPU](https://github.com/huggingface/diffusers/blob/35a969d297cba69110d175ee79c59312b9f49e1e/src/diffusers/schedulers/scheduling_euler_discrete.py#L240), the CPU and GPU sync doesn’t occur and you don't get any latency. In general, any CPU and GPU communication sync should be none or be kept to a bare minimum because it can impact inference latency.
+
+## Combine the attention block's projection matrices
+
+The UNet and VAE in SDXL use Transformer-like blocks which consists of attention blocks and feed-forward blocks.
+
+In an attention block, the input is projected into three sub-spaces using three different projection matrices – Q, K, and V. These projections are performed separately on the input. But we can horizontally combine the projection matrices into a single matrix and perform the projection in one step. This increases the size of the matrix multiplications of the input projections and improves the impact of quantization.
+
+You can combine the projection matrices with just a single line of code:
+
+```python
+pipe.fuse_qkv_projections()
+```
+
+This provides a minor improvement from 2.54 seconds to 2.52 seconds.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/progressive-acceleration-sdxl/SDXL%2C_Batch_Size%3A_1%2C_Steps%3A_30_4.png" width=500>
+</div>
+
+<Tip warning={true}>
+
+Support for [`~StableDiffusionXLPipeline.fuse_qkv_projections`] is limited and experimental. It's not available for many non-Stable Diffusion pipelines such as [Kandinsky](../using-diffusers/kandinsky). You can refer to this [PR](https://github.com/huggingface/diffusers/pull/6179) to get an idea about how to enable this for the other pipelines.
+
+</Tip>
+
+## Dynamic quantization
+
+You can also use the ultra-lightweight PyTorch quantization library, [torchao](https://github.com/pytorch-labs/ao) (commit SHA `54bcd5a10d0abbe7b0c045052029257099f83fd9`), to apply [dynamic int8 quantization](https://pytorch.org/tutorials/recipes/recipes/dynamic_quantization.html) to the UNet and VAE. Quantization adds additional conversion overhead to the model that is hopefully made up for by faster matmuls (dynamic quantization). If the matmuls are too small, these techniques may degrade performance.
+
+First, configure all the compiler tags:
+
+```python
+from diffusers import StableDiffusionXLPipeline
+import torch
+
+# Notice the two new flags at the end.
+torch._inductor.config.conv_1x1_as_mm = True
+torch._inductor.config.coordinate_descent_tuning = True
+torch._inductor.config.epilogue_fusion = False
+torch._inductor.config.coordinate_descent_check_all_directions = True
+torch._inductor.config.force_fuse_int_mm_with_mul = True
+torch._inductor.config.use_mixed_mm = True
+```
+
+Certain linear layers in the UNet and VAE don’t benefit from dynamic int8 quantization. You can filter out those layers with the [`dynamic_quant_filter_fn`](https://github.com/huggingface/diffusion-fast/blob/0f169640b1db106fe6a479f78c1ed3bfaeba3386/utils/pipeline_utils.py#L16) shown below.
+
+```python
+def dynamic_quant_filter_fn(mod, *args):
+    return (
+        isinstance(mod, torch.nn.Linear)
+        and mod.in_features > 16
+        and (mod.in_features, mod.out_features)
+        not in [
+            (1280, 640),
+            (1920, 1280),
+            (1920, 640),
+            (2048, 1280),
+            (2048, 2560),
+            (2560, 1280),
+            (256, 128),
+            (2816, 1280),
+            (320, 640),
+            (512, 1536),
+            (512, 256),
+            (512, 512),
+            (640, 1280),
+            (640, 1920),
+            (640, 320),
+            (640, 5120),
+            (640, 640),
+            (960, 320),
+            (960, 640),
+        ]
+    )
+
+
+def conv_filter_fn(mod, *args):
+    return (
+        isinstance(mod, torch.nn.Conv2d) and mod.kernel_size == (1, 1) and 128 in [mod.in_channels, mod.out_channels]
+    )
+```
+
+Finally, apply all the optimizations discussed so far:
+
+```python
+# SDPA + bfloat16.
+pipe = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.bfloat16
+).to("cuda")
+
+# Combine attention projection matrices.
+pipe.fuse_qkv_projections()
+
+# Change the memory layout.
+pipe.unet.to(memory_format=torch.channels_last)
+pipe.vae.to(memory_format=torch.channels_last)
+```
+
+Since dynamic quantization is only limited to the linear layers, convert the appropriate pointwise convolution layers into linear layers to maximize its benefit.
+
+```python
+from torchao import swap_conv2d_1x1_to_linear
+
+swap_conv2d_1x1_to_linear(pipe.unet, conv_filter_fn)
+swap_conv2d_1x1_to_linear(pipe.vae, conv_filter_fn)
+```
+
+Apply dynamic quantization:
+
+```python
+from torchao import apply_dynamic_quant
+
+apply_dynamic_quant(pipe.unet, dynamic_quant_filter_fn)
+apply_dynamic_quant(pipe.vae, dynamic_quant_filter_fn)
+```
+
+Finally, compile and perform inference:
+
+```python
+pipe.unet = torch.compile(pipe.unet, mode="max-autotune", fullgraph=True)
+pipe.vae.decode = torch.compile(pipe.vae.decode, mode="max-autotune", fullgraph=True)
+
+prompt = "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+image = pipe(prompt, num_inference_steps=30).images[0]
+```
+
+Applying dynamic quantization improves the latency from 2.52 seconds to 2.43 seconds.
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/sayakpaul/sample-datasets/resolve/main/progressive-acceleration-sdxl/SDXL%2C_Batch_Size%3A_1%2C_Steps%3A_30_5.png" width=500>
+</div>

docs/source/en/tutorials/inference_with_big_models.md

@@ -0,0 +1,139 @@
+<!--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.
+-->
+
+# Working with big models
+
+A modern diffusion model, like [Stable Diffusion XL (SDXL)](../using-diffusers/sdxl), is not just a single model, but a collection of multiple models. SDXL has four different model-level components:
+
+* A variational autoencoder (VAE)
+* Two text encoders
+* A UNet for denoising
+
+Usually, the text encoders and the denoiser are much larger compared to the VAE.
+
+As models get bigger and better, it’s possible your model is so big that even a single copy won’t fit in memory. But that doesn’t mean it can’t be loaded. If you have more than one GPU, there is more memory available to store your model. In this case, it’s better to split your model checkpoint into several smaller *checkpoint shards*.
+
+When a text encoder checkpoint has multiple shards, like [T5-xxl for SD3](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers/tree/main/text_encoder_3), it is automatically handled by the [Transformers](https://huggingface.co/docs/transformers/index) library as it is a required dependency of Diffusers when using the [`StableDiffusion3Pipeline`]. More specifically, Transformers will automatically handle the loading of multiple shards within the requested model class and get it ready so that inference can be performed.
+
+The denoiser checkpoint can also have multiple shards and supports inference thanks to the [Accelerate](https://huggingface.co/docs/accelerate/index) library.
+
+> [!TIP]
+> Refer to the [Handling big models for inference](https://huggingface.co/docs/accelerate/main/en/concept_guides/big_model_inference) guide for general guidance when working with big models that are hard to fit into memory.
+
+For example, let's save a sharded checkpoint for the [SDXL UNet](https://huggingface.co/stabilityai/stable-diffusion-xl-base-1.0/tree/main/unet):
+
+```python
+from diffusers import UNet2DConditionModel
+
+unet = UNet2DConditionModel.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", subfolder="unet"
+)
+unet.save_pretrained("sdxl-unet-sharded", max_shard_size="5GB")
+```
+
+The size of the fp32 variant of the SDXL UNet checkpoint is ~10.4GB. Set the `max_shard_size` parameter to 5GB to create 3 shards. After saving, you can load them in [`StableDiffusionXLPipeline`]:
+
+```python
+from diffusers import UNet2DConditionModel, StableDiffusionXLPipeline
+import torch
+
+unet = UNet2DConditionModel.from_pretrained(
+    "sayakpaul/sdxl-unet-sharded", torch_dtype=torch.float16
+)
+pipeline = StableDiffusionXLPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", unet=unet, torch_dtype=torch.float16
+).to("cuda")
+
+image = pipeline("a cute dog running on the grass", num_inference_steps=30).images[0]
+image.save("dog.png")
+```
+
+If placing all the model-level components on the GPU at once is not feasible, use [`~DiffusionPipeline.enable_model_cpu_offload`] to help you:
+
+```diff
+- pipeline.to("cuda")
++ pipeline.enable_model_cpu_offload()
+```
+
+In general, we recommend sharding when a checkpoint is more than 5GB (in fp32).
+
+## Device placement
+
+On distributed setups, you can run inference across multiple GPUs with Accelerate.
+
+> [!WARNING]
+> This feature is experimental and its APIs might change in the future.
+
+With Accelerate, you can use the `device_map` to determine how to distribute the models of a pipeline across multiple devices. This is useful in situations where you have more than one GPU.
+
+For example, if you have two 8GB GPUs, then using [`~DiffusionPipeline.enable_model_cpu_offload`] may not work so well because:
+
+* it only works on a single GPU
+* a single model might not fit on a single GPU ([`~DiffusionPipeline.enable_sequential_cpu_offload`] might work but it will be extremely slow and it is also limited to a single GPU)
+
+To make use of both GPUs, you can use the "balanced" device placement strategy which splits the models across all available GPUs.
+
+> [!WARNING]
+> Only the "balanced" strategy is supported at the moment, and we plan to support additional mapping strategies in the future.
+
+```diff
+from diffusers import DiffusionPipeline
+import torch
+
+pipeline = DiffusionPipeline.from_pretrained(
+-    "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True,
++    "stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16, use_safetensors=True, device_map="balanced"
+)
+image = pipeline("a dog").images[0]
+image
+```
+
+You can also pass a dictionary to enforce the maximum GPU memory that can be used on each device:
+
+```diff
+from diffusers import DiffusionPipeline
+import torch
+
+max_memory = {0:"1GB", 1:"1GB"}
+pipeline = DiffusionPipeline.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+    device_map="balanced",
++   max_memory=max_memory
+)
+image = pipeline("a dog").images[0]
+image
+```
+
+If a device is not present in `max_memory`, then it will be completely ignored and will not participate in the device placement.
+
+By default, Diffusers uses the maximum memory of all devices. If the models don't fit on the GPUs, they are offloaded to the CPU. If the CPU doesn't have enough memory, then you might see an error. In that case, you could defer to using [`~DiffusionPipeline.enable_sequential_cpu_offload`] and [`~DiffusionPipeline.enable_model_cpu_offload`].
+
+Call [`~DiffusionPipeline.reset_device_map`] to reset the `device_map` of a pipeline. This is also necessary if you want to use methods like `to()`, [`~DiffusionPipeline.enable_sequential_cpu_offload`], and [`~DiffusionPipeline.enable_model_cpu_offload`] on a pipeline that was device-mapped.
+
+```py
+pipeline.reset_device_map()
+```
+
+Once a pipeline has been device-mapped, you can also access its device map via `hf_device_map`:
+
+```py
+print(pipeline.hf_device_map)
+```
+
+An example device map would look like so:
+
+
+```bash
+{'unet': 1, 'vae': 1, 'safety_checker': 0, 'text_encoder': 0}
+```

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

@@ -0,0 +1,120 @@
+<!--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.
+-->
+# CogVideoX
+
+CogVideoX is a text-to-video generation model focused on creating more coherent videos aligned with a prompt. It achieves this using several methods.
+
+- a 3D variational autoencoder that compresses videos spatially and temporally, improving compression rate and video accuracy.
+
+- an expert transformer block to help align text and video, and a 3D full attention module for capturing and creating spatially and temporally accurate videos.
+ 
+
+
+## Load model checkpoints
+Model weights may be stored in separate subfolders on the Hub or locally, in which case, you should use the [`~DiffusionPipeline.from_pretrained`] method.
+
+
+```py
+from diffusers import CogVideoXPipeline, CogVideoXImageToVideoPipeline
+pipe = CogVideoXPipeline.from_pretrained(
+    "THUDM/CogVideoX-2b",
+    torch_dtype=torch.float16
+)
+
+pipe = CogVideoXImageToVideoPipeline.from_pretrained(
+    "THUDM/CogVideoX-5b-I2V",
+    torch_dtype=torch.bfloat16
+)
+
+```
+
+## Text-to-Video
+For text-to-video, pass a text prompt. By default, CogVideoX generates a 720x480 video for the best results.
+
+```py
+import torch
+from diffusers import CogVideoXPipeline
+from diffusers.utils import export_to_video
+
+prompt = "An elderly gentleman, with a serene expression, sits at the water's edge, a steaming cup of tea by his side. He is engrossed in his artwork, brush in hand, as he renders an oil painting on a canvas that's propped up against a small, weathered table. The sea breeze whispers through his silver hair, gently billowing his loose-fitting white shirt, while the salty air adds an intangible element to his masterpiece in progress. The scene is one of tranquility and inspiration, with the artist's canvas capturing the vibrant hues of the setting sun reflecting off the tranquil sea."
+
+pipe = CogVideoXPipeline.from_pretrained(
+    "THUDM/CogVideoX-5b",
+    torch_dtype=torch.bfloat16
+)
+
+pipe.enable_model_cpu_offload()
+pipe.vae.enable_tiling()
+
+video = pipe(
+    prompt=prompt,
+    num_videos_per_prompt=1,
+    num_inference_steps=50,
+    num_frames=49,
+    guidance_scale=6,
+    generator=torch.Generator(device="cuda").manual_seed(42),
+).frames[0]
+
+export_to_video(video, "output.mp4", fps=8)
+
+```
+
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_out.gif" alt="generated image of an astronaut in a jungle"/>
+</div>
+
+
+## Image-to-Video
+
+
+You'll use the [THUDM/CogVideoX-5b-I2V](https://huggingface.co/THUDM/CogVideoX-5b-I2V)  checkpoint for this guide.
+
+```py
+import torch
+from diffusers import CogVideoXImageToVideoPipeline
+from diffusers.utils import export_to_video, load_image
+
+prompt = "A vast, shimmering ocean flows gracefully under a twilight sky, its waves undulating in a mesmerizing dance of blues and greens. The surface glints with the last rays of the setting sun, casting golden highlights that ripple across the water. Seagulls soar above, their cries blending with the gentle roar of the waves. The horizon stretches infinitely, where the ocean meets the sky in a seamless blend of hues. Close-ups reveal the intricate patterns of the waves, capturing the fluidity and dynamic beauty of the sea in motion."
+image = load_image(image="cogvideox_rocket.png")
+pipe = CogVideoXImageToVideoPipeline.from_pretrained(
+    "THUDM/CogVideoX-5b-I2V",
+    torch_dtype=torch.bfloat16
+)
+ 
+pipe.vae.enable_tiling()
+pipe.vae.enable_slicing()
+
+video = pipe(
+    prompt=prompt,
+    image=image,
+    num_videos_per_prompt=1,
+    num_inference_steps=50,
+    num_frames=49,
+    guidance_scale=6,
+    generator=torch.Generator(device="cuda").manual_seed(42),
+).frames[0]
+
+export_to_video(video, "output.mp4", fps=8)
+```
+
+<div class="flex gap-4">
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_rocket.png"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">initial image</figcaption>
+  </div>
+  <div>
+    <img class="rounded-xl" src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/cogvideox/cogvideox_outrocket.gif"/>
+    <figcaption class="mt-2 text-center text-sm text-gray-500">generated video</figcaption>
+  </div>
+</div>
+

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

@@ -0,0 +1,363 @@
+<!--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.
+-->
+
+# Load adapters
+
+[[open-in-colab]]
+
+There are several [training](../training/overview) techniques for personalizing diffusion models to generate images of a specific subject or images in certain styles. Each of these training methods produces a different type of adapter. Some of the adapters generate an entirely new model, while other adapters only modify a smaller set of embeddings or weights. This means the loading process for each adapter is also different.
+
+This guide will show you how to load DreamBooth, textual inversion, and LoRA weights.
+
+<Tip>
+
+Feel free to browse the [Stable Diffusion Conceptualizer](https://huggingface.co/spaces/sd-concepts-library/stable-diffusion-conceptualizer), [LoRA the Explorer](https://huggingface.co/spaces/multimodalart/LoraTheExplorer), and the [Diffusers Models Gallery](https://huggingface.co/spaces/huggingface-projects/diffusers-gallery) for checkpoints and embeddings to use.
+
+</Tip>
+
+## DreamBooth
+
+[DreamBooth](https://dreambooth.github.io/) finetunes an *entire diffusion model* on just several images of a subject to generate images of that subject in new styles and settings. This method works by using a special word in the prompt that the model learns to associate with the subject image. Of all the training methods, DreamBooth produces the largest file size (usually a few GBs) because it is a full checkpoint model.
+
+Let's load the [herge_style](https://huggingface.co/sd-dreambooth-library/herge-style) checkpoint, which is trained on just 10 images drawn by Hergé, to generate images in that style. For it to work, you need to include the special word `herge_style` in your prompt to trigger the checkpoint:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("sd-dreambooth-library/herge-style", torch_dtype=torch.float16).to("cuda")
+prompt = "A cute herge_style brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_dreambooth.png" />
+</div>
+
+## Textual inversion
+
+[Textual inversion](https://textual-inversion.github.io/) is very similar to DreamBooth and it can also personalize a diffusion model to generate certain concepts (styles, objects) from just a few images. This method works by training and finding new embeddings that represent the images you provide with a special word in the prompt. As a result, the diffusion model weights stay the same and the training process produces a relatively tiny (a few KBs) file.
+
+Because textual inversion creates embeddings, it cannot be used on its own like DreamBooth and requires another model.
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
+```
+
+Now you can load the textual inversion embeddings with the [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`] method and generate some images. Let's load the [sd-concepts-library/gta5-artwork](https://huggingface.co/sd-concepts-library/gta5-artwork) embeddings and you'll need to include the special word `<gta5-artwork>` in your prompt to trigger it:
+
+```py
+pipeline.load_textual_inversion("sd-concepts-library/gta5-artwork")
+prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration, <gta5-artwork> style"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_txt_embed.png" />
+</div>
+
+Textual inversion can also be trained on undesirable things to create *negative embeddings* to discourage a model from generating images with those undesirable things like blurry images or extra fingers on a hand. This can be an easy way to quickly improve your prompt. You'll also load the embeddings with [`~loaders.TextualInversionLoaderMixin.load_textual_inversion`], but this time, you'll need two more parameters:
+
+- `weight_name`: specifies the weight file to load if the file was saved in the 🤗 Diffusers format with a specific name or if the file is stored in the A1111 format
+- `token`: specifies the special word to use in the prompt to trigger the embeddings
+
+Let's load the [sayakpaul/EasyNegative-test](https://huggingface.co/sayakpaul/EasyNegative-test) embeddings:
+
+```py
+pipeline.load_textual_inversion(
+    "sayakpaul/EasyNegative-test", weight_name="EasyNegative.safetensors", token="EasyNegative"
+)
+```
+
+Now you can use the `token` to generate an image with the negative embeddings:
+
+```py
+prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration, EasyNegative"
+negative_prompt = "EasyNegative"
+
+image = pipeline(prompt, negative_prompt=negative_prompt, num_inference_steps=50).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png" />
+</div>
+
+## LoRA
+
+[Low-Rank Adaptation (LoRA)](https://huggingface.co/papers/2106.09685) is a popular training technique because it is fast and generates smaller file sizes (a couple hundred MBs). Like the other methods in this guide, LoRA can train a model to learn new styles from just a few images. It works by inserting new weights into the diffusion model and then only the new weights are trained instead of the entire model. This makes LoRAs faster to train and easier to store.
+
+<Tip>
+
+LoRA is a very general training technique that can be used with other training methods. For example, it is common to train a model with DreamBooth and LoRA. It is also increasingly common to load and merge multiple LoRAs to create new and unique images. You can learn more about it in the in-depth [Merge LoRAs](merge_loras) guide since merging is outside the scope of this loading guide.
+
+</Tip>
+
+LoRAs also need to be used with another model:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+```
+
+Then use the [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method to load the [ostris/super-cereal-sdxl-lora](https://huggingface.co/ostris/super-cereal-sdxl-lora) weights and specify the weights filename from the repository:
+
+```py
+pipeline.load_lora_weights("ostris/super-cereal-sdxl-lora", weight_name="cereal_box_sdxl_v1.safetensors")
+prompt = "bears, pizza bites"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_lora.png" />
+</div>
+
+The [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method loads LoRA weights into both the UNet and text encoder. It is the preferred way for loading LoRAs because it can handle cases where:
+
+- the LoRA weights don't have separate identifiers for the UNet and text encoder
+- the LoRA weights have separate identifiers for the UNet and text encoder
+
+To directly load (and save) a LoRA adapter at the *model-level*, use [`~PeftAdapterMixin.load_lora_adapter`], which builds and prepares the necessary model configuration for the adapter. Like [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`], [`PeftAdapterMixin.load_lora_adapter`] can load LoRAs for both the UNet and text encoder. For example, if you're loading a LoRA for the UNet, [`PeftAdapterMixin.load_lora_adapter`] ignores the keys for the text encoder.
+
+Use the `weight_name` parameter to specify the specific weight file and the `prefix` parameter to filter for the appropriate state dicts (`"unet"` in this case) to load.
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.unet.load_lora_adapter("jbilcke-hf/sdxl-cinematic-1", weight_name="pytorch_lora_weights.safetensors", prefix="unet")
+
+# use cnmt in the prompt to trigger the LoRA
+prompt = "A cute cnmt eating a slice of pizza, stunning color scheme, masterpiece, illustration"
+image = pipeline(prompt).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_attn_proc.png" />
+</div>
+
+Save an adapter with [`~PeftAdapterMixin.save_lora_adapter`].
+
+To unload the LoRA weights, use the [`~loaders.StableDiffusionLoraLoaderMixin.unload_lora_weights`] method to discard the LoRA weights and restore the model to its original weights:
+
+```py
+pipeline.unload_lora_weights()
+```
+
+### Adjust LoRA weight scale
+
+For both [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] and [`~loaders.UNet2DConditionLoadersMixin.load_attn_procs`], you can pass the `cross_attention_kwargs={"scale": 0.5}` parameter to adjust how much of the LoRA weights to use. A value of `0` is the same as only using the base model weights, and a value of `1` is equivalent to using the fully finetuned LoRA.
+
+For more granular control on the amount of LoRA weights used per layer, you can use [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`] and pass a dictionary specifying by how much to scale the weights in each layer by.
+```python
+pipe = ... # create pipeline
+pipe.load_lora_weights(..., adapter_name="my_adapter")
+scales = {
+    "text_encoder": 0.5,
+    "text_encoder_2": 0.5,  # only usable if pipe has a 2nd text encoder
+    "unet": {
+        "down": 0.9,  # all transformers in the down-part will use scale 0.9
+        # "mid"  # in this example "mid" is not given, therefore all transformers in the mid part will use the default scale 1.0
+        "up": {
+            "block_0": 0.6,  # all 3 transformers in the 0th block in the up-part will use scale 0.6
+            "block_1": [0.4, 0.8, 1.0],  # the 3 transformers in the 1st block in the up-part will use scales 0.4, 0.8 and 1.0 respectively
+        }
+    }
+}
+pipe.set_adapters("my_adapter", scales)
+```
+
+This also works with multiple adapters - see [this guide](https://huggingface.co/docs/diffusers/tutorials/using_peft_for_inference#customize-adapters-strength) for how to do it.
+
+<Tip warning={true}>
+
+Currently, [`~loaders.StableDiffusionLoraLoaderMixin.set_adapters`] only supports scaling attention weights. If a LoRA has other parts (e.g., resnets or down-/upsamplers), they will keep a scale of 1.0.
+
+</Tip>
+
+### Kohya and TheLastBen
+
+Other popular LoRA trainers from the community include those by [Kohya](https://github.com/kohya-ss/sd-scripts/) and [TheLastBen](https://github.com/TheLastBen/fast-stable-diffusion). These trainers create different LoRA checkpoints than those trained by 🤗 Diffusers, but they can still be loaded in the same way.
+
+<hfoptions id="other-trainers">
+<hfoption id="Kohya">
+
+To load a Kohya LoRA, let's download the [Blueprintify SD XL 1.0](https://civitai.com/models/150986/blueprintify-sd-xl-10) checkpoint from [Civitai](https://civitai.com/) as an example:
+
+```sh
+!wget https://civitai.com/api/download/models/168776 -O blueprintify-sd-xl-10.safetensors
+```
+
+Load the LoRA checkpoint with the [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method, and specify the filename in the `weight_name` parameter:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("path/to/weights", weight_name="blueprintify-sd-xl-10.safetensors")
+```
+
+Generate an image:
+
+```py
+# use bl3uprint in the prompt to trigger the LoRA
+prompt = "bl3uprint, a highly detailed blueprint of the eiffel tower, explaining how to build all parts, many txt, blueprint grid backdrop"
+image = pipeline(prompt).images[0]
+image
+```
+
+<Tip warning={true}>
+
+Some limitations of using Kohya LoRAs with 🤗 Diffusers include:
+
+- Images may not look like those generated by UIs - like ComfyUI - for multiple reasons, which are explained [here](https://github.com/huggingface/diffusers/pull/4287/#issuecomment-1655110736).
+- [LyCORIS checkpoints](https://github.com/KohakuBlueleaf/LyCORIS) aren't fully supported. The [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method loads LyCORIS checkpoints with LoRA and LoCon modules, but Hada and LoKR are not supported.
+
+</Tip>
+
+</hfoption>
+<hfoption id="TheLastBen">
+
+Loading a checkpoint from TheLastBen is very similar. For example, to load the [TheLastBen/William_Eggleston_Style_SDXL](https://huggingface.co/TheLastBen/William_Eggleston_Style_SDXL) checkpoint:
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("TheLastBen/William_Eggleston_Style_SDXL", weight_name="wegg.safetensors")
+
+# use by william eggleston in the prompt to trigger the LoRA
+prompt = "a house by william eggleston, sunrays, beautiful, sunlight, sunrays, beautiful"
+image = pipeline(prompt=prompt).images[0]
+image
+```
+
+</hfoption>
+</hfoptions>
+
+## IP-Adapter
+
+[IP-Adapter](https://ip-adapter.github.io/) is a lightweight adapter that enables image prompting for any diffusion model. This adapter works by decoupling the cross-attention layers of the image and text features. All the other model components are frozen and only the embedded image features in the UNet are trained. As a result, IP-Adapter files are typically only ~100MBs.
+
+You can learn more about how to use IP-Adapter for different tasks and specific use cases in the [IP-Adapter](../using-diffusers/ip_adapter) guide.
+
+> [!TIP]
+> Diffusers currently only supports IP-Adapter for some of the most popular pipelines. Feel free to open a feature request if you have a cool use case and want to integrate IP-Adapter with an unsupported pipeline!
+> Official IP-Adapter checkpoints are available from [h94/IP-Adapter](https://huggingface.co/h94/IP-Adapter).
+
+To start, load a Stable Diffusion checkpoint.
+
+```py
+from diffusers import AutoPipelineForText2Image
+import torch
+from diffusers.utils import load_image
+
+pipeline = AutoPipelineForText2Image.from_pretrained("stable-diffusion-v1-5/stable-diffusion-v1-5", torch_dtype=torch.float16).to("cuda")
+```
+
+Then load the IP-Adapter weights and add it to the pipeline with the [`~loaders.IPAdapterMixin.load_ip_adapter`] method.
+
+```py
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="models", weight_name="ip-adapter_sd15.bin")
+```
+
+Once loaded, you can use the pipeline with an image and text prompt to guide the image generation process.
+
+```py
+image = load_image("https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/load_neg_embed.png")
+generator = torch.Generator(device="cpu").manual_seed(33)
+images = pipeline(
+    prompt='best quality, high quality, wearing sunglasses',
+    ip_adapter_image=image,
+    negative_prompt="monochrome, lowres, bad anatomy, worst quality, low quality",
+    num_inference_steps=50,
+    generator=generator,
+).images[0]
+images
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/YiYiXu/testing-images/resolve/main/ip-bear.png" />
+</div>
+
+### IP-Adapter Plus
+
+IP-Adapter relies on an image encoder to generate image features. If the IP-Adapter repository contains an `image_encoder` subfolder, the image encoder is automatically loaded and registered to the pipeline. Otherwise, you'll need to explicitly load the image encoder with a [`~transformers.CLIPVisionModelWithProjection`] model and pass it to the pipeline.
+
+This is the case for *IP-Adapter Plus* checkpoints which use the ViT-H image encoder.
+
+```py
+from transformers import CLIPVisionModelWithProjection
+
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+    "h94/IP-Adapter",
+    subfolder="models/image_encoder",
+    torch_dtype=torch.float16
+)
+
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    image_encoder=image_encoder,
+    torch_dtype=torch.float16
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter", subfolder="sdxl_models", weight_name="ip-adapter-plus_sdxl_vit-h.safetensors")
+```
+
+### IP-Adapter Face ID models
+
+The IP-Adapter FaceID models are experimental IP Adapters that use image embeddings generated by `insightface` instead of CLIP image embeddings. Some of these models also use LoRA to improve ID consistency.
+You need to install `insightface` and all its requirements to use these models.
+
+<Tip warning={true}>
+As InsightFace pretrained models are available for non-commercial research purposes, IP-Adapter-FaceID models are released exclusively for research purposes and are not intended for commercial use.
+</Tip>
+
+```py
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter-FaceID", subfolder=None, weight_name="ip-adapter-faceid_sdxl.bin", image_encoder_folder=None)
+```
+
+If you want to use one of the two IP-Adapter FaceID Plus models, you must also load the CLIP image encoder, as this models use both `insightface` and CLIP image embeddings to achieve better photorealism.
+
+```py
+from transformers import CLIPVisionModelWithProjection
+
+image_encoder = CLIPVisionModelWithProjection.from_pretrained(
+    "laion/CLIP-ViT-H-14-laion2B-s32B-b79K",
+    torch_dtype=torch.float16,
+)
+
+pipeline = AutoPipelineForText2Image.from_pretrained(
+    "stable-diffusion-v1-5/stable-diffusion-v1-5",
+    image_encoder=image_encoder,
+    torch_dtype=torch.float16
+).to("cuda")
+
+pipeline.load_ip_adapter("h94/IP-Adapter-FaceID", subfolder=None, weight_name="ip-adapter-faceid-plus_sd15.bin")
+```

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

@@ -0,0 +1,266 @@
+<!--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.
+-->
+
+# Merge LoRAs
+
+It can be fun and creative to use multiple [LoRAs]((https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora)) together to generate something entirely new and unique. This works by merging multiple LoRA weights together to produce images that are a blend of different styles. Diffusers provides a few methods to merge LoRAs depending on *how* you want to merge their weights, which can affect image quality.
+
+This guide will show you how to merge LoRAs using the [`~loaders.PeftAdapterMixin.set_adapters`] and [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) methods. To improve inference speed and reduce memory-usage of merged LoRAs, you'll also see how to use the [`~loaders.StableDiffusionLoraLoaderMixin.fuse_lora`] method to fuse the LoRA weights with the original weights of the underlying model.
+
+For this guide, load a Stable Diffusion XL (SDXL) checkpoint and the [KappaNeuro/studio-ghibli-style](https://huggingface.co/KappaNeuro/studio-ghibli-style) and [Norod78/sdxl-chalkboarddrawing-lora](https://huggingface.co/Norod78/sdxl-chalkboarddrawing-lora) LoRAs with the [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_weights`] method. You'll need to assign each LoRA an `adapter_name` to combine them later.
+
+```py
+from diffusers import DiffusionPipeline
+import torch
+
+pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl", weight_name="ikea_instructions_xl_v1_5.safetensors", adapter_name="ikea")
+pipeline.load_lora_weights("lordjia/by-feng-zikai", weight_name="fengzikai_v1.0_XL.safetensors", adapter_name="feng")
+```
+
+## set_adapters
+
+The [`~loaders.PeftAdapterMixin.set_adapters`] method merges LoRA adapters by concatenating their weighted matrices. Use the adapter name to specify which LoRAs to merge, and the `adapter_weights` parameter to control the scaling for each LoRA. For example, if `adapter_weights=[0.5, 0.5]`, then the merged LoRA output is an average of both LoRAs. Try adjusting the adapter weights to see how it affects the generated image!
+
+```py
+pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])
+
+generator = torch.manual_seed(0)
+prompt = "A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai"
+image = pipeline(prompt, generator=generator, cross_attention_kwargs={"scale": 1.0}).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/lora_merge_set_adapters.png"/>
+</div>
+
+## add_weighted_adapter
+
+> [!WARNING]
+> This is an experimental method that adds PEFTs [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method to Diffusers to enable more efficient merging methods. Check out this [issue](https://github.com/huggingface/diffusers/issues/6892) if you're interested in learning more about the motivation and design behind this integration.
+
+The [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method provides access to more efficient merging method such as [TIES and DARE](https://huggingface.co/docs/peft/developer_guides/model_merging). To use these merging methods, make sure you have the latest stable version of Diffusers and PEFT installed.
+
+```bash
+pip install -U diffusers peft
+```
+
+There are three steps to merge LoRAs with the [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method:
+
+1. Create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the underlying model and LoRA checkpoint.
+2. Load a base UNet model and the LoRA adapters.
+3. Merge the adapters using the [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method and the merging method of your choice.
+
+Let's dive deeper into what these steps entail.
+
+1. Load a UNet that corresponds to the UNet in the LoRA checkpoint. In this case, both LoRAs use the SDXL UNet as their base model.
+
+```python
+from diffusers import UNet2DConditionModel
+import torch
+
+unet = UNet2DConditionModel.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+    variant="fp16",
+    subfolder="unet",
+).to("cuda")
+```
+
+Load the SDXL pipeline and the LoRA checkpoints, starting with the [ostris/ikea-instructions-lora-sdxl](https://huggingface.co/ostris/ikea-instructions-lora-sdxl) LoRA.
+
+```python
+from diffusers import DiffusionPipeline
+
+pipeline = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    variant="fp16",
+    torch_dtype=torch.float16,
+    unet=unet
+).to("cuda")
+pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl", weight_name="ikea_instructions_xl_v1_5.safetensors", adapter_name="ikea")
+```
+
+Now you'll create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the loaded LoRA checkpoint by combining the SDXL UNet and the LoRA UNet from the pipeline.
+
+```python
+from peft import get_peft_model, LoraConfig
+import copy
+
+sdxl_unet = copy.deepcopy(unet)
+ikea_peft_model = get_peft_model(
+    sdxl_unet,
+    pipeline.unet.peft_config["ikea"],
+    adapter_name="ikea"
+)
+
+original_state_dict = {f"base_model.model.{k}": v for k, v in pipeline.unet.state_dict().items()}
+ikea_peft_model.load_state_dict(original_state_dict, strict=True)
+```
+
+> [!TIP]
+> You can optionally push the ikea_peft_model to the Hub by calling `ikea_peft_model.push_to_hub("ikea_peft_model", token=TOKEN)`.
+
+Repeat this process to create a [PeftModel](https://huggingface.co/docs/peft/package_reference/peft_model#peft.PeftModel) from the [lordjia/by-feng-zikai](https://huggingface.co/lordjia/by-feng-zikai) LoRA.
+
+```python
+pipeline.delete_adapters("ikea")
+sdxl_unet.delete_adapters("ikea")
+
+pipeline.load_lora_weights("lordjia/by-feng-zikai", weight_name="fengzikai_v1.0_XL.safetensors", adapter_name="feng")
+pipeline.set_adapters(adapter_names="feng")
+
+feng_peft_model = get_peft_model(
+    sdxl_unet,
+    pipeline.unet.peft_config["feng"],
+    adapter_name="feng"
+)
+
+original_state_dict = {f"base_model.model.{k}": v for k, v in pipe.unet.state_dict().items()}
+feng_peft_model.load_state_dict(original_state_dict, strict=True)
+```
+
+2. Load a base UNet model and then load the adapters onto it.
+
+```python
+from peft import PeftModel
+
+base_unet = UNet2DConditionModel.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0",
+    torch_dtype=torch.float16,
+    use_safetensors=True,
+    variant="fp16",
+    subfolder="unet",
+).to("cuda")
+
+model = PeftModel.from_pretrained(base_unet, "stevhliu/ikea_peft_model", use_safetensors=True, subfolder="ikea", adapter_name="ikea")
+model.load_adapter("stevhliu/feng_peft_model", use_safetensors=True, subfolder="feng", adapter_name="feng")
+```
+
+3. Merge the adapters using the [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) method and the merging method of your choice (learn more about other merging methods in this [blog post](https://huggingface.co/blog/peft_merging)). For this example, let's use the `"dare_linear"` method to merge the LoRAs.
+
+> [!WARNING]
+> Keep in mind the LoRAs need to have the same rank to be merged!
+
+```python
+model.add_weighted_adapter(
+    adapters=["ikea", "feng"],
+    weights=[1.0, 1.0],
+    combination_type="dare_linear",
+    adapter_name="ikea-feng"
+)
+model.set_adapters("ikea-feng")
+```
+
+Now you can generate an image with the merged LoRA.
+
+```python
+model = model.to(dtype=torch.float16, device="cuda")
+
+pipeline = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-xl-base-1.0", unet=model, variant="fp16", torch_dtype=torch.float16,
+).to("cuda")
+
+image = pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai", generator=torch.manual_seed(0)).images[0]
+image
+```
+
+<div class="flex justify-center">
+    <img src="https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/ikea-feng-dare-linear.png"/>
+</div>
+
+## fuse_lora
+
+Both the [`~loaders.PeftAdapterMixin.set_adapters`] and [add_weighted_adapter](https://huggingface.co/docs/peft/package_reference/lora#peft.LoraModel.add_weighted_adapter) methods require loading the base model and the LoRA adapters separately which incurs some overhead. The [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method allows you to fuse the LoRA weights directly with the original weights of the underlying model. This way, you're only loading the model once which can increase inference and lower memory-usage.
+
+You can use PEFT to easily fuse/unfuse multiple adapters directly into the model weights (both UNet and text encoder) using the [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method, which can lead to a speed-up in inference and lower VRAM usage.
+
+For example, if you have a base model and adapters loaded and set as active with the following adapter weights:
+
+```py
+from diffusers import DiffusionPipeline
+import torch
+
+pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl", weight_name="ikea_instructions_xl_v1_5.safetensors", adapter_name="ikea")
+pipeline.load_lora_weights("lordjia/by-feng-zikai", weight_name="fengzikai_v1.0_XL.safetensors", adapter_name="feng")
+
+pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])
+```
+
+Fuse these LoRAs into the UNet with the [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method. The `lora_scale` parameter controls how much to scale the output by with the LoRA weights. It is important to make the `lora_scale` adjustments in the [`~loaders.lora_base.LoraBaseMixin.fuse_lora`] method because it won’t work if you try to pass `scale` to the `cross_attention_kwargs` in the pipeline.
+
+```py
+pipeline.fuse_lora(adapter_names=["ikea", "feng"], lora_scale=1.0)
+```
+
+Then you should use [`~loaders.StableDiffusionLoraLoaderMixin.unload_lora_weights`] to unload the LoRA weights since they've already been fused with the underlying base model. Finally, call [`~DiffusionPipeline.save_pretrained`] to save the fused pipeline locally or you could call [`~DiffusionPipeline.push_to_hub`] to push the fused pipeline to the Hub.
+
+```py
+pipeline.unload_lora_weights()
+# save locally
+pipeline.save_pretrained("path/to/fused-pipeline")
+# save to the Hub
+pipeline.push_to_hub("fused-ikea-feng")
+```
+
+Now you can quickly load the fused pipeline and use it for inference without needing to separately load the LoRA adapters.
+
+```py
+pipeline = DiffusionPipeline.from_pretrained(
+    "username/fused-ikea-feng", torch_dtype=torch.float16,
+).to("cuda")
+
+image = pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai", generator=torch.manual_seed(0)).images[0]
+image
+```
+
+You can call [`~~loaders.lora_base.LoraBaseMixin.unfuse_lora`] to restore the original model's weights (for example, if you want to use a different `lora_scale` value). However, this only works if you've only fused one LoRA adapter to the original model. If you've fused multiple LoRAs, you'll need to reload the model.
+
+```py
+pipeline.unfuse_lora()
+```
+
+### torch.compile
+
+[torch.compile](../optimization/torch2.0#torchcompile) can speed up your pipeline even more, but the LoRA weights must be fused first and then unloaded. Typically, the UNet is compiled because it is such a computationally intensive component of the pipeline.
+
+```py
+from diffusers import DiffusionPipeline
+import torch
+
+# load base model and LoRAs
+pipeline = DiffusionPipeline.from_pretrained("stabilityai/stable-diffusion-xl-base-1.0", torch_dtype=torch.float16).to("cuda")
+pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl", weight_name="ikea_instructions_xl_v1_5.safetensors", adapter_name="ikea")
+pipeline.load_lora_weights("lordjia/by-feng-zikai", weight_name="fengzikai_v1.0_XL.safetensors", adapter_name="feng")
+
+# activate both LoRAs and set adapter weights
+pipeline.set_adapters(["ikea", "feng"], adapter_weights=[0.7, 0.8])
+
+# fuse LoRAs and unload weights
+pipeline.fuse_lora(adapter_names=["ikea", "feng"], lora_scale=1.0)
+pipeline.unload_lora_weights()
+
+# torch.compile
+pipeline.unet.to(memory_format=torch.channels_last)
+pipeline.unet = torch.compile(pipeline.unet, mode="reduce-overhead", fullgraph=True)
+
+image = pipeline("A bowl of ramen shaped like a cute kawaii bear, by Feng Zikai", generator=torch.manual_seed(0)).images[0]
+```
+
+Learn more about torch.compile in the [Accelerate inference of text-to-image diffusion models](../tutorials/fast_diffusion#torchcompile) guide.
+
+## Next steps
+
+For more conceptual details about how each merging method works, take a look at the [🤗 PEFT welcomes new merging methods](https://huggingface.co/blog/peft_merging#concatenation-cat) blog post!

src/diffusers/hooks/group_offloading.py

@@ -135,9 +135,58 @@ class ModuleGroup:
         finally:
             pinned_dict = None
 
+    def _transfer_tensor_to_device(self, tensor, source_tensor, current_stream=None):
+        tensor.data = source_tensor.to(self.onload_device, non_blocking=self.non_blocking)
+        if self.record_stream and current_stream is not None:
+            tensor.data.record_stream(current_stream)
+
+    def _process_tensors_from_modules(self, pinned_memory=None, current_stream=None):
+        for group_module in self.modules:
+            for param in group_module.parameters():
+                source = pinned_memory[param] if pinned_memory else param.data
+                self._transfer_tensor_to_device(param, source, current_stream)
+            for buffer in group_module.buffers():
+                source = pinned_memory[buffer] if pinned_memory else buffer.data
+                self._transfer_tensor_to_device(buffer, source, current_stream)
+
+        for param in self.parameters:
+            source = pinned_memory[param] if pinned_memory else param.data
+            self._transfer_tensor_to_device(param, source, current_stream)
+
+        for buffer in self.buffers:
+            source = pinned_memory[buffer] if pinned_memory else buffer.data
+            self._transfer_tensor_to_device(buffer, source, current_stream)
+
+    def _onload_from_disk(self, current_stream):
+        if self.stream is not None:
+            loaded_cpu_tensors = safetensors.torch.load_file(self.safetensors_file_path, device="cpu")
+
+            for key, tensor_obj in self.key_to_tensor.items():
+                self.cpu_param_dict[tensor_obj] = loaded_cpu_tensors[key]
+
+            with self._pinned_memory_tensors() as pinned_memory:
+                for key, tensor_obj in self.key_to_tensor.items():
+                    self._transfer_tensor_to_device(tensor_obj, pinned_memory[tensor_obj], current_stream)
+
+            self.cpu_param_dict.clear()
+
+        else:
+            onload_device = (
+                self.onload_device.type if isinstance(self.onload_device, torch.device) else self.onload_device
+            )
+            loaded_tensors = safetensors.torch.load_file(self.safetensors_file_path, device=onload_device)
+            for key, tensor_obj in self.key_to_tensor.items():
+                tensor_obj.data = loaded_tensors[key]
+
+    def _onload_from_memory(self, current_stream):
+        if self.stream is not None:
+            with self._pinned_memory_tensors() as pinned_memory:
+                self._process_tensors_from_modules(pinned_memory, current_stream)
+        else:
+            self._process_tensors_from_modules(None, current_stream)
+
     @torch.compiler.disable()
     def onload_(self):
-        r"""Onloads the group of modules to the onload_device."""
         torch_accelerator_module = (
             getattr(torch, torch.accelerator.current_accelerator().type)
             if hasattr(torch, "accelerator")
@@ -175,67 +224,32 @@ class ModuleGroup:
             self.stream.synchronize()
 
         with context:
-            if self.stream is not None:
-                with self._pinned_memory_tensors() as pinned_memory:
-                    for group_module in self.modules:
-                        for param in group_module.parameters():
-                            param.data = pinned_memory[param].to(self.onload_device, non_blocking=self.non_blocking)
-                            if self.record_stream:
-                                param.data.record_stream(current_stream)
-                        for buffer in group_module.buffers():
-                            buffer.data = pinned_memory[buffer].to(self.onload_device, non_blocking=self.non_blocking)
-                            if self.record_stream:
-                                buffer.data.record_stream(current_stream)
-
-                    for param in self.parameters:
-                        param.data = pinned_memory[param].to(self.onload_device, non_blocking=self.non_blocking)
-                        if self.record_stream:
-                            param.data.record_stream(current_stream)
-
-                    for buffer in self.buffers:
-                        buffer.data = pinned_memory[buffer].to(self.onload_device, non_blocking=self.non_blocking)
-                        if self.record_stream:
-                            buffer.data.record_stream(current_stream)
-
+            if self.offload_to_disk_path:
+                self._onload_from_disk(current_stream)
             else:
-                for group_module in self.modules:
-                    for param in group_module.parameters():
-                        param.data = param.data.to(self.onload_device, non_blocking=self.non_blocking)
-                    for buffer in group_module.buffers():
-                        buffer.data = buffer.data.to(self.onload_device, non_blocking=self.non_blocking)
-
-                for param in self.parameters:
-                    param.data = param.data.to(self.onload_device, non_blocking=self.non_blocking)
-
-                for buffer in self.buffers:
-                    buffer.data = buffer.data.to(self.onload_device, non_blocking=self.non_blocking)
-                    if self.record_stream:
-                        buffer.data.record_stream(current_stream)
+                self._onload_from_memory(current_stream)
 
     @torch.compiler.disable()
-    def offload_(self):
-        r"""Offloads the group of modules to the offload_device."""
-        if self.offload_to_disk_path:
-            # TODO: we can potentially optimize this code path by checking if the _all_ the desired
-            # safetensor files exist on the disk and if so, skip this step entirely, reducing IO
-            # overhead. Currently, we just check if the given `safetensors_file_path` exists and if not
-            # we perform a write.
-            # Check if the file has been saved in this session or if it already exists on disk.
-            if not self._is_offloaded_to_disk and not os.path.exists(self.safetensors_file_path):
-                os.makedirs(os.path.dirname(self.safetensors_file_path), exist_ok=True)
-                tensors_to_save = {
-                    key: tensor.data.to(self.offload_device) for tensor, key in self.tensor_to_key.items()
-                }
-                safetensors.torch.save_file(tensors_to_save, self.safetensors_file_path)
+    def _offload_to_disk(self):
+        # TODO: we can potentially optimize this code path by checking if the _all_ the desired
+        # safetensor files exist on the disk and if so, skip this step entirely, reducing IO
+        # overhead. Currently, we just check if the given `safetensors_file_path` exists and if not
+        # we perform a write.
+        # Check if the file has been saved in this session or if it already exists on disk.
+        if not self._is_offloaded_to_disk and not os.path.exists(self.safetensors_file_path):
+            os.makedirs(os.path.dirname(self.safetensors_file_path), exist_ok=True)
+            tensors_to_save = {key: tensor.data.to(self.offload_device) for tensor, key in self.tensor_to_key.items()}
+            safetensors.torch.save_file(tensors_to_save, self.safetensors_file_path)
 
-            # The group is now considered offloaded to disk for the rest of the session.
-            self._is_offloaded_to_disk = True
+        # The group is now considered offloaded to disk for the rest of the session.
+        self._is_offloaded_to_disk = True
 
-            # We do this to free up the RAM which is still holding the up tensor data.
-            for tensor_obj in self.tensor_to_key.keys():
-                tensor_obj.data = torch.empty_like(tensor_obj.data, device=self.offload_device)
-            return
+        # We do this to free up the RAM which is still holding the up tensor data.
+        for tensor_obj in self.tensor_to_key.keys():
+            tensor_obj.data = torch.empty_like(tensor_obj.data, device=self.offload_device)
 
+    @torch.compiler.disable()
+    def _offload_to_memory(self):
         torch_accelerator_module = (
             getattr(torch, torch.accelerator.current_accelerator().type)
             if hasattr(torch, "accelerator")
@@ -260,6 +274,14 @@ class ModuleGroup:
             for buffer in self.buffers:
                 buffer.data = buffer.data.to(self.offload_device, non_blocking=self.non_blocking)
 
+    @torch.compiler.disable()
+    def offload_(self):
+        r"""Offloads the group of modules to the offload_device."""
+        if self.offload_to_disk_path:
+            self._offload_to_disk()
+        else:
+            self._offload_to_memory()
+
 
 class GroupOffloadingHook(ModelHook):
     r"""
@@ -484,6 +506,8 @@ def apply_group_offloading(
             option only matters when using streamed CPU offloading (i.e. `use_stream=True`). This can be useful when
             the CPU memory is a bottleneck but may counteract the benefits of using streams.
 
+    (TODO: include example with `offload_to_disk_path`)
+
     Example:
         ```python
         >>> from diffusers import CogVideoXTransformer3DModel