[docs] Merge LoRAs (huggingface#7213)

stevhliu · web-flow · commit 3ce905c9d0e9 · 2024-03-07T11:28:50.000-08:00
* merge loras

* feedback

* torch.compile

* feedback
diff --git a/docs/source/en/_toctree.yml b/docs/source/en/_toctree.yml
@@ -18,7 +18,7 @@
   - local: tutorials/basic_training
     title: Train a diffusion model
   - local: tutorials/using_peft_for_inference
-    title: Inference with PEFT
+    title: Load LoRAs for inference
   - local: tutorials/fast_diffusion
     title: Accelerate inference of text-to-image diffusion models
   title: Tutorials
@@ -62,6 +62,8 @@
       title: Textual inversion
     - local: using-diffusers/ip_adapter
       title: IP-Adapter
+    - local: using-diffusers/merge_loras
+      title: Merge LoRAs
     - local: training/distributed_inference
       title: Distributed inference with multiple GPUs
     - local: using-diffusers/reusing_seeds
diff --git a/docs/source/en/tutorials/using_peft_for_inference.md b/docs/source/en/tutorials/using_peft_for_inference.md
@@ -14,19 +14,17 @@ specific language governing permissions and limitations under the License.
 
 # Load LoRAs for inference
 
-There are many adapters (with LoRAs being the most common type) trained in different styles to achieve different effects. You can even combine multiple adapters to create new and unique images. With the 🤗 [PEFT](https://huggingface.co/docs/peft/index) integration in 🤗 Diffusers, it is really easy to load and manage adapters for inference. In this guide, you'll learn how to use different adapters with [Stable Diffusion XL (SDXL)](../api/pipelines/stable_diffusion/stable_diffusion_xl) for inference.
+There are many adapter types (with [LoRAs](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) being the most popular) trained in different styles to achieve different effects. You can even combine multiple adapters to create new and unique images.
 
-Throughout this guide, you'll use LoRA as the main adapter technique, so we'll use the terms LoRA and adapter interchangeably. You should have some familiarity with LoRA, and if you don't, we welcome you to check out the [LoRA guide](https://huggingface.co/docs/peft/conceptual_guides/lora).
+In this tutorial, you'll learn how to easily load and manage adapters for inference with the 🤗 [PEFT](https://huggingface.co/docs/peft/index) integration in 🤗 Diffusers. You'll use LoRA as the main adapter technique, so you'll see the terms LoRA and adapter used interchangeably.
 
 Let's first install all the required libraries.
 
 ```bash
-!pip install -q transformers accelerate
-!pip install peft
-!pip install diffusers
+!pip install -q transformers accelerate peft diffusers
 ```
 
-Now, let's load a pipeline with a SDXL checkpoint:
+Now, load a pipeline with a [Stable Diffusion XL (SDXL)](../api/pipelines/stable_diffusion/stable_diffusion_xl) checkpoint:
 
 ```python
 from diffusers import DiffusionPipeline
@@ -36,16 +34,13 @@ pipe_id = "stabilityai/stable-diffusion-xl-base-1.0"
 pipe = DiffusionPipeline.from_pretrained(pipe_id, torch_dtype=torch.float16).to("cuda")
 ```
 
-
-Next, load a LoRA checkpoint with the [`~diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] method.
-
-With the 🤗 PEFT integration, you can assign a specific `adapter_name` to the checkpoint, which let's you easily switch between different LoRA checkpoints. Let's call this adapter `"toy"`.
+Next, load a [CiroN2022/toy-face](https://huggingface.co/CiroN2022/toy-face) adapter with the [`~diffusers.loaders.StableDiffusionXLLoraLoaderMixin.load_lora_weights`] method. With the 🤗 PEFT integration, you can assign a specific `adapter_name` to the checkpoint, which let's you easily switch between different LoRA checkpoints. Let's call this adapter `"toy"`.
 
 ```python
 pipe.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
 ```
 
-And then perform inference:
+Make sure to include the token `toy_face` in the prompt and then you can perform inference:
 
 ```python
 prompt = "toy_face of a hacker with a hoodie"
@@ -59,17 +54,16 @@ image
 
 ![toy-face](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_8_1.png)
 
+With the `adapter_name` parameter, it is really easy to use another adapter for inference! Load the [nerijs/pixel-art-xl](https://huggingface.co/nerijs/pixel-art-xl) adapter that has been fine-tuned to generate pixel art images and call it `"pixel"`.
 
-With the `adapter_name` parameter, it is really easy to use another adapter for inference! Load the [nerijs/pixel-art-xl](https://huggingface.co/nerijs/pixel-art-xl) adapter that has been fine-tuned to generate pixel art images, and let's call it `"pixel"`.
-
-The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter. But you can activate the `"pixel"` adapter with the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method as shown below:
+The pipeline automatically sets the first loaded adapter (`"toy"`) as the active adapter, but you can activate the `"pixel"` adapter with the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method:
 
 ```python
 pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
 pipe.set_adapters("pixel")
 ```
 
-Let's now generate an image with the second adapter and check the result:
+Make sure you include the token `pixel art` in your prompt to generate a pixel art image:
 
 ```python
 prompt = "a hacker with a hoodie, pixel art"
@@ -81,29 +75,25 @@ image
 
 ![pixel-art](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_12_1.png)
 
-## Combine multiple adapters
+## Merge adapters
 
-You can also perform multi-adapter inference where you combine different adapter checkpoints for inference.
+You can also merge different adapter checkpoints for inference to blend their styles together.
 
-Once again, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate two LoRA checkpoints and specify the weight for how the checkpoints should be combined.
+Once again, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `pixel` and `toy` adapters and specify the weights for how they should be merged.
 
 ```python
 pipe.set_adapters(["pixel", "toy"], adapter_weights=[0.5, 1.0])
 ```
 
-Now that we have set these two adapters, let's generate an image from the combined adapters!
-
 <Tip>
 
 LoRA checkpoints in the diffusion community are almost always obtained with [DreamBooth](https://huggingface.co/docs/diffusers/main/en/training/dreambooth). DreamBooth training often relies on "trigger" words in the input text prompts in order for the generation results to look as expected. When you combine multiple LoRA checkpoints, it's important to ensure the trigger words for the corresponding LoRA checkpoints are present in the input text prompts.
 
 </Tip>
 
-The trigger words for [CiroN2022/toy-face](https://hf.co/CiroN2022/toy-face) and [nerijs/pixel-art-xl](https://hf.co/nerijs/pixel-art-xl) are found in their repositories.
-
+Remember to use the trigger words for [CiroN2022/toy-face](https://hf.co/CiroN2022/toy-face) and [nerijs/pixel-art-xl](https://hf.co/nerijs/pixel-art-xl) (these are found in their repositories) in the prompt to generate an image.
 
 ```python
-# Notice how the prompt is constructed.
 prompt = "toy_face of a hacker with a hoodie, pixel art"
 image = pipe(
     prompt, num_inference_steps=30, cross_attention_kwargs={"scale": 1.0}, generator=torch.manual_seed(0)
@@ -113,15 +103,16 @@ image
 
 ![toy-face-pixel-art](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_16_1.png)
 
-Impressive! As you can see, the model was able to generate an image that mixes the characteristics of both adapters.
+Impressive! As you can see, the model generated an image that mixed the characteristics of both adapters.
+
+> [!TIP]
+> Through its PEFT integration, Diffusers also offers more efficient merging methods which you can learn about in the [Merge LoRAs](../using-diffusers/merge_loras) guide!
 
-If you want to go back to using only one adapter, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `"toy"` adapter:
+To return to only using one adapter, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.set_adapters`] method to activate the `"toy"` adapter:
 
 ```python
-# First, set the adapter.
 pipe.set_adapters("toy")
 
-# Then, run inference.
 prompt = "toy_face of a hacker with a hoodie"
 lora_scale= 0.9
 image = pipe(
@@ -130,11 +121,7 @@ image = pipe(
 image
 ```
 
-![toy-face-again](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_18_1.png)
-
-
-If you want to switch to only the base model, disable all LoRAs with the [`~diffusers.loaders.UNet2DConditionLoadersMixin.disable_lora`] method.
-
+Or to disable all adapters entirely, use the [`~diffusers.loaders.UNet2DConditionLoadersMixin.disable_lora`] method to return the base model.
 
 ```python
 pipe.disable_lora()
@@ -145,11 +132,9 @@ image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).ima
 image
 ```
 
-![no-lora](https://huggingface.co/datasets/huggingface/documentation-images/resolve/main/diffusers/peft_integration/diffusers_peft_lora_inference_20_1.png)
+## Manage active adapters
 
-## Monitoring active adapters
-
-You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, you can easily check the list of active adapters using the [`~diffusers.loaders.LoraLoaderMixin.get_active_adapters`] method:
+You have attached multiple adapters in this tutorial, and if you're feeling a bit lost on what adapters have been attached to the pipeline's components, use the [`~diffusers.loaders.LoraLoaderMixin.get_active_adapters`] method to check the list of active adapters:
 
 ```py
 active_adapters = pipe.get_active_adapters()
@@ -164,78 +149,3 @@ list_adapters_component_wise = pipe.get_list_adapters()
 list_adapters_component_wise
 {"text_encoder": ["toy", "pixel"], "unet": ["toy", "pixel"], "text_encoder_2": ["toy", "pixel"]}
 ```
-
-## Compatibility with `torch.compile`
-
-If you want to compile your model with `torch.compile` make sure to first fuse the LoRA weights into the base model and unload them.
-
-```diff
-pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
-pipe.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
-
-pipe.set_adapters(["pixel", "toy"], adapter_weights=[0.5, 1.0])
-# Fuses the LoRAs into the Unet
-pipe.fuse_lora()
-pipe.unload_lora_weights()
-
-+ pipe.unet.to(memory_format=torch.channels_last)
-+ pipe.unet = torch.compile(pipe.unet, mode="reduce-overhead", fullgraph=True)
-
-prompt = "toy_face of a hacker with a hoodie, pixel art"
-image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
-```
-
-> [!TIP]
-> You can refer to the `torch.compile()` section [here](https://huggingface.co/docs/diffusers/main/en/optimization/torch2.0#torchcompile) and [here](https://huggingface.co/docs/diffusers/main/en/tutorials/fast_diffusion#torchcompile) for more elaborate examples.
-
-## Fusing adapters into the model
-
-You can use PEFT to easily fuse/unfuse multiple adapters directly into the model weights (both UNet and text encoder) using the [`~diffusers.loaders.LoraLoaderMixin.fuse_lora`] method, which can lead to a speed-up in inference and lower VRAM usage.
-
-```py
-pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
-pipe.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
-
-pipe.set_adapters(["pixel", "toy"], adapter_weights=[0.5, 1.0])
-# Fuses the LoRAs into the Unet
-pipe.fuse_lora()
-
-prompt = "toy_face of a hacker with a hoodie, pixel art"
-image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
-
-# Gets the Unet back to the original state
-pipe.unfuse_lora()
-```
-
-You can also fuse some adapters using `adapter_names` for faster generation:
-
-```py
-pipe.load_lora_weights("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
-pipe.load_lora_weights("CiroN2022/toy-face", weight_name="toy_face_sdxl.safetensors", adapter_name="toy")
-
-pipe.set_adapters(["pixel"], adapter_weights=[0.5, 1.0])
-# Fuses the LoRAs into the Unet
-pipe.fuse_lora(adapter_names=["pixel"])
-
-prompt = "a hacker with a hoodie, pixel art"
-image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
-
-# Gets the Unet back to the original state
-pipe.unfuse_lora()
-
-# Fuse all adapters
-pipe.fuse_lora(adapter_names=["pixel", "toy"])
-
-prompt = "toy_face of a hacker with a hoodie, pixel art"
-image = pipe(prompt, num_inference_steps=30, generator=torch.manual_seed(0)).images[0]
-```
-
-## Saving a pipeline after fusing the adapters
-
-To properly save a pipeline after it's been loaded with the adapters, it should be serialized like so:
-
-```python
-pipe.fuse_lora(lora_scale=1.0)
-pipe.unload_lora_weights()
-pipe.save_pretrained("path-to-pipeline")
-```
diff --git a/docs/source/en/using-diffusers/loading_adapters.md b/docs/source/en/using-diffusers/loading_adapters.md
@@ -103,7 +103,7 @@ image
 
 <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.
+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>
 
@@ -165,101 +165,14 @@ To unload the LoRA weights, use the [`~loaders.LoraLoaderMixin.unload_lora_weigh
 pipeline.unload_lora_weights()
 ```
 
-### Load multiple LoRAs
-
-It can be fun to use multiple LoRAs together to create something entirely new and unique. The [`~loaders.LoraLoaderMixin.fuse_lora`] method allows you to fuse the LoRA weights with the original weights of the underlying model.
-
-<Tip>
-
-Fusing the weights can lead to a speedup in inference latency because you don't need to separately load the base model and LoRA! You can save your fused pipeline with [`~DiffusionPipeline.save_pretrained`] to avoid loading and fusing the weights every time you want to use the model.
-
-</Tip>
-
-Load an initial model:
-
-```py
-from diffusers import StableDiffusionXLPipeline, AutoencoderKL
-import torch
-
-vae = AutoencoderKL.from_pretrained("madebyollin/sdxl-vae-fp16-fix", torch_dtype=torch.float16)
-pipeline = StableDiffusionXLPipeline.from_pretrained(
-    "stabilityai/stable-diffusion-xl-base-1.0",
-    vae=vae,
-    torch_dtype=torch.float16,
-).to("cuda")
-```
-
-Next, load the LoRA checkpoint and fuse it with the original weights. 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.LoraLoaderMixin.fuse_lora`] method because it won't work if you try to pass `scale` to the `cross_attention_kwargs` in the pipeline.
-
-If you need to reset the original model weights for any reason (use a different `lora_scale`), you should use the [`~loaders.LoraLoaderMixin.unfuse_lora`] method.
-
-```py
-pipeline.load_lora_weights("ostris/ikea-instructions-lora-sdxl")
-pipeline.fuse_lora(lora_scale=0.7)
-
-# to unfuse the LoRA weights
-pipeline.unfuse_lora()
-```
-
-Then fuse this pipeline with the next set of LoRA weights:
-
-```py
-pipeline.load_lora_weights("ostris/super-cereal-sdxl-lora")
-pipeline.fuse_lora(lora_scale=0.7)
-```
-
-<Tip warning={true}>
-
-You can't unfuse multiple LoRA checkpoints, so if you need to reset the model to its original weights, you'll need to reload it.
-
-</Tip>
-
-Now you can generate an image that uses the weights from both LoRAs:
-
-```py
-prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
-image = pipeline(prompt).images[0]
-image
-```
-
-### 🤗 PEFT
-
-<Tip>
-
-Read the [Inference with 🤗 PEFT](../tutorials/using_peft_for_inference) tutorial to learn more about its integration with 🤗 Diffusers and how you can easily work with and juggle multiple adapters. You'll need to install 🤗 Diffusers and PEFT from source to run the example in this section.
-
-</Tip>
-
-Another way you can load and use multiple LoRAs is to specify the `adapter_name` parameter in [`~loaders.LoraLoaderMixin.load_lora_weights`]. This method takes advantage of the 🤗 PEFT integration. For example, load and name both LoRA 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("ostris/super-cereal-sdxl-lora", weight_name="cereal_box_sdxl_v1.safetensors", adapter_name="cereal")
-```
-
-Now use the [`~loaders.UNet2DConditionLoadersMixin.set_adapters`] to activate both LoRAs, and you can configure how much weight each LoRA should have on the output:
-
-```py
-pipeline.set_adapters(["ikea", "cereal"], adapter_weights=[0.7, 0.5])
-```
-
-Then, generate an image:
-
-```py
-prompt = "A cute brown bear eating a slice of pizza, stunning color scheme, masterpiece, illustration"
-image = pipeline(prompt, num_inference_steps=30, cross_attention_kwargs={"scale": 1.0}).images[0]
-image
-```
-
 ### 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.
 
-Let's download the [Blueprintify SD XL 1.0](https://civitai.com/models/150986/blueprintify-sd-xl-10) checkpoint from [Civitai](https://civitai.com/):
+<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
@@ -293,6 +206,9 @@ Some limitations of using Kohya LoRAs with 🤗 Diffusers include:
 
 </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
@@ -308,6 +224,9 @@ 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.
diff --git a/docs/source/en/using-diffusers/merge_loras.md b/docs/source/en/using-diffusers/merge_loras.md
