Merge branch 'main' into memory-optims

Author: sayakpaul

.github/workflows/nightly_tests.yml

@@ -180,6 +180,55 @@ jobs:
         pip install slack_sdk tabulate
         python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
 
+  run_torch_compile_tests:
+    name: PyTorch Compile CUDA tests
+
+    runs-on:
+      group: aws-g4dn-2xlarge
+
+    container:
+      image: diffusers/diffusers-pytorch-compile-cuda
+      options: --gpus 0 --shm-size "16gb" --ipc host
+
+    steps:
+    - name: Checkout diffusers
+      uses: actions/checkout@v3
+      with:
+        fetch-depth: 2
+
+    - name: NVIDIA-SMI
+      run: |
+        nvidia-smi
+    - name: Install dependencies
+      run: |
+        python -m venv /opt/venv && export PATH="/opt/venv/bin:$PATH"
+        python -m uv pip install -e [quality,test,training]
+    - name: Environment
+      run: |
+        python utils/print_env.py
+    - name: Run torch compile tests on GPU
+      env:
+        HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
+        RUN_COMPILE: yes
+      run: |
+        python -m pytest -n 1 --max-worker-restart=0 --dist=loadfile -s -v -k "compile" --make-reports=tests_torch_compile_cuda tests/
+    - name: Failure short reports
+      if: ${{ failure() }}
+      run: cat reports/tests_torch_compile_cuda_failures_short.txt
+
+    - name: Test suite reports artifacts
+      if: ${{ always() }}
+      uses: actions/upload-artifact@v4
+      with:
+        name: torch_compile_test_reports
+        path: reports
+
+    - name: Generate Report and Notify Channel
+      if: always()
+      run: |
+        pip install slack_sdk tabulate
+        python utils/log_reports.py >> $GITHUB_STEP_SUMMARY
+  
   run_big_gpu_torch_tests:
     name: Torch tests on big GPU
     strategy:

.github/workflows/release_tests_fast.yml

@@ -335,7 +335,7 @@ jobs:
     - name: Environment
       run: |
         python utils/print_env.py
-    - name: Run example tests on GPU
+    - name: Run torch compile tests on GPU
       env:
         HF_TOKEN: ${{ secrets.DIFFUSERS_HF_HUB_READ_TOKEN }}
         RUN_COMPILE: yes

docs/source/en/api/pipelines/flux.md

@@ -347,7 +347,7 @@ image = pipe(
     height=1024,
     prompt="wearing sunglasses",
     negative_prompt="",
-    true_cfg=4.0,
+    true_cfg_scale=4.0,
     generator=torch.Generator().manual_seed(4444),
     ip_adapter_image=image,
 ).images[0]

docs/source/en/api/pipelines/wan.md

@@ -24,7 +24,7 @@
 
 ## Generating Videos with Wan 2.1
 
-We will first need to install some addtional dependencies.
+We will first need to install some additional dependencies.
 
 ```shell
 pip install -u ftfy imageio-ffmpeg imageio

docs/source/en/training/cogvideox.md

@@ -216,7 +216,7 @@ Setting the `<ID_TOKEN>` is not necessary. From some limited experimentation, we
 > - The original repository uses a `lora_alpha` of `1`. We found this not suitable in many runs, possibly due to difference in modeling backends and training settings. Our recommendation is to set to the `lora_alpha` to either `rank` or `rank // 2`.
 > - If you're training on data whose captions generate bad results with the original model, a `rank` of 64 and above is good and also the recommendation by the team behind CogVideoX. If the generations are already moderately good on your training captions, a `rank` of 16/32 should work. We found that setting the rank too low, say `4`, is not ideal and doesn't produce promising results.
 > - The authors of CogVideoX recommend 4000 training steps and 100 training videos overall to achieve the best result. While that might yield the best results, we found from our limited experimentation that 2000 steps and 25 videos could also be sufficient.
-> - When using the Prodigy opitimizer for training, one can follow the recommendations from [this](https://huggingface.co/blog/sdxl_lora_advanced_script) blog. Prodigy tends to overfit quickly. From my very limited testing, I found a learning rate of `0.5` to be suitable in addition to `--prodigy_use_bias_correction`, `prodigy_safeguard_warmup` and `--prodigy_decouple`.
+> - When using the Prodigy optimizer for training, one can follow the recommendations from [this](https://huggingface.co/blog/sdxl_lora_advanced_script) blog. Prodigy tends to overfit quickly. From my very limited testing, I found a learning rate of `0.5` to be suitable in addition to `--prodigy_use_bias_correction`, `prodigy_safeguard_warmup` and `--prodigy_decouple`.
 > - The recommended learning rate by the CogVideoX authors and from our experimentation with Adam/AdamW is between `1e-3` and `1e-4` for a dataset of 25+ videos.
 >
 > Note that our testing is not exhaustive due to limited time for exploration. Our recommendation would be to play around with the different knobs and dials to find the best settings for your data.

docs/source/en/training/dreambooth.md

@@ -589,7 +589,7 @@ For stage 2 of DeepFloyd IF with DreamBooth, pay attention to these parameters:
 
 * `--learning_rate=5e-6`, use a lower learning rate with a smaller effective batch size
 * `--resolution=256`, the expected resolution for the upscaler
-* `--train_batch_size=2` and `--gradient_accumulation_steps=6`, to effectively train on images wiht faces requires larger batch sizes
+* `--train_batch_size=2` and `--gradient_accumulation_steps=6`, to effectively train on images with faces requires larger batch sizes
 
 ```bash
 export MODEL_NAME="DeepFloyd/IF-II-L-v1.0"

docs/source/en/training/t2i_adapters.md

@@ -89,7 +89,7 @@ Many of the basic and important parameters are described in the [Text-to-image](
 
 As with the script parameters, a walkthrough of the training script is provided in the [Text-to-image](text2image#training-script) training guide. Instead, this guide takes a look at the T2I-Adapter relevant parts of the script.
 
-The training script begins by preparing the dataset. This incudes [tokenizing](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L674) the prompt and [applying transforms](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L714) to the images and conditioning images.
+The training script begins by preparing the dataset. This includes [tokenizing](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L674) the prompt and [applying transforms](https://github.com/huggingface/diffusers/blob/aab6de22c33cc01fb7bc81c0807d6109e2c998c9/examples/t2i_adapter/train_t2i_adapter_sdxl.py#L714) to the images and conditioning images.
 
 ```py
 conditioning_image_transforms = transforms.Compose(

examples/advanced_diffusion_training/train_dreambooth_lora_flux_advanced.py

@@ -2181,7 +2181,7 @@ def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
                 # Predict the noise residual
                 model_pred = transformer(
                     hidden_states=packed_noisy_model_input,
-                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
+                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transformer model (we should not keep it but I want to keep the inputs same for the model for testing)
                     timestep=timesteps / 1000,
                     guidance=guidance,
                     pooled_projections=pooled_prompt_embeds,

examples/community/README.md

@@ -86,6 +86,7 @@ PIXART-α Controlnet pipeline | Implementation of the controlnet model for pixar
 | Perturbed-Attention Guidance |StableDiffusionPAGPipeline is a modification of StableDiffusionPipeline to support Perturbed-Attention Guidance (PAG).|[Perturbed-Attention Guidance](#perturbed-attention-guidance)|[Notebook](https://github.com/huggingface/notebooks/blob/main/diffusers/perturbed_attention_guidance.ipynb)|[Hyoungwon Cho](https://github.com/HyoungwonCho)|
 | CogVideoX DDIM Inversion Pipeline | Implementation of DDIM inversion and guided attention-based editing denoising process on CogVideoX. | [CogVideoX DDIM Inversion Pipeline](#cogvideox-ddim-inversion-pipeline) | - | [LittleNyima](https://github.com/LittleNyima) |
 | FaithDiff Stable Diffusion XL Pipeline | Implementation of [(CVPR 2025) FaithDiff: Unleashing Diffusion Priors for Faithful Image Super-resolutionUnleashing Diffusion Priors for Faithful Image Super-resolution](https://arxiv.org/abs/2411.18824) - FaithDiff is a faithful image super-resolution method that leverages latent diffusion models by actively adapting the diffusion prior and jointly fine-tuning its components (encoder and diffusion model) with an alignment module to ensure high fidelity and structural consistency. | [FaithDiff Stable Diffusion XL Pipeline](#faithdiff-stable-diffusion-xl-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/jychen9811/FaithDiff) | [Junyang Chen, Jinshan Pan, Jiangxin Dong, IMAG Lab, (Adapted by Eliseu Silva)](https://github.com/JyChen9811/FaithDiff) |
+| Stable Diffusion 3 InstructPix2Pix Pipeline | Implementation of Stable Diffusion 3 InstructPix2Pix Pipeline | [Stable Diffusion 3 InstructPix2Pix Pipeline](#stable-diffusion-3-instructpix2pix-pipeline) | [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/BleachNick/SD3_UltraEdit_freeform) [![Hugging Face Models](https://img.shields.io/badge/%F0%9F%A4%97%20Hugging%20Face-Models-blue)](https://huggingface.co/CaptainZZZ/sd3-instructpix2pix) | [Jiayu Zhang](https://github.com/xduzhangjiayu) and [Haozhe Zhao](https://github.com/HaozheZhao)|
 To load a custom pipeline you just need to pass the `custom_pipeline` argument to `DiffusionPipeline`, as one of the files in `diffusers/examples/community`. Feel free to send a PR with your own pipelines, we will merge them quickly.
 
 ```py
@@ -5381,7 +5382,7 @@ pipe = DiffusionPipeline.from_pretrained(
 # Here we need use pipeline internal unet model
 pipe.unet = pipe.unet_model.from_pretrained(model_id, subfolder="unet", variant="fp16", use_safetensors=True)
 
-# Load aditional layers to the model
+# Load additional layers to the model
 pipe.unet.load_additional_layers(weight_path="proc_data/faithdiff/FaithDiff.bin", dtype=dtype)
 
 # Enable vae tiling
@@ -5432,4 +5433,50 @@ cropped_image = gen_image.crop((0, 0, width_init, height_init))
 cropped_image.save("data/result.png")
 ````
 ### Result
-[<img src="https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/faithdiff_restored.PNG" width="512px" height="512px"/>](https://imgsli.com/MzY1NzE2)
+[<img src="https://huggingface.co/datasets/DEVAIEXP/assets/resolve/main/faithdiff_restored.PNG" width="512px" height="512px"/>](https://imgsli.com/MzY1NzE2)
+
+
+# Stable Diffusion 3 InstructPix2Pix Pipeline
+This the implementation of the Stable Diffusion 3 InstructPix2Pix Pipeline, based on the HuggingFace Diffusers.
+
+## Example Usage
+This pipeline aims to edit image based on user's instruction by using SD3
+````py
+import torch
+from diffusers import SD3Transformer2DModel
+from diffusers import DiffusionPipeline
+from diffusers.utils import load_image
+
+
+resolution = 512
+image = load_image("https://hf.co/datasets/diffusers/diffusers-images-docs/resolve/main/mountain.png").resize(
+    (resolution, resolution)
+)
+edit_instruction = "Turn sky into a sunny one"
+
+
+pipe = DiffusionPipeline.from_pretrained(
+    "stabilityai/stable-diffusion-3-medium-diffusers", custom_pipeline="pipeline_stable_diffusion_3_instruct_pix2pix", torch_dtype=torch.float16).to('cuda')
+
+pipe.transformer = SD3Transformer2DModel.from_pretrained("CaptainZZZ/sd3-instructpix2pix",torch_dtype=torch.float16).to('cuda')
+
+edited_image = pipe(
+    prompt=edit_instruction,
+    image=image,
+    height=resolution,
+    width=resolution,
+    guidance_scale=7.5,
+    image_guidance_scale=1.5,
+    num_inference_steps=30,
+).images[0]
+
+edited_image.save("edited_image.png")
+````
+|Original|Edited|
+|---|---|
+|![Original image](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/StableDiffusion3InstructPix2Pix/mountain.png)|![Edited image](https://huggingface.co/datasets/diffusers/docs-images/resolve/main/StableDiffusion3InstructPix2Pix/edited.png)
+
+### Note
+This model is trained on 512x512, so input size is better on 512x512.
+For better editing performance, please refer to this powerful model https://huggingface.co/BleachNick/SD3_UltraEdit_freeform and Paper "UltraEdit: Instruction-based Fine-Grained Image
+Editing at Scale", many thanks to their contribution!

examples/community/dps_pipeline.py

@@ -312,9 +312,9 @@ def contributions(self, in_length, out_length, scale, kernel, kernel_width, anti
                     # These are the coordinates of the output image
                     out_coordinates = np.arange(1, out_length + 1)
 
-                    # since both scale-factor and output size can be provided simulatneously, perserving the center of the image requires shifting
-                    # the output coordinates. the deviation is because out_length doesn't necesary equal in_length*scale.
-                    # to keep the center we need to subtract half of this deivation so that we get equal margins for boths sides and center is preserved.
+                    # since both scale-factor and output size can be provided simultaneously, preserving the center of the image requires shifting
+                    # the output coordinates. the deviation is because out_length doesn't necessary equal in_length*scale.
+                    # to keep the center we need to subtract half of this deviation so that we get equal margins for both sides and center is preserved.
                     shifted_out_coordinates = out_coordinates - (out_length - in_length * scale) / 2
 
                     # These are the matching positions of the output-coordinates on the input image coordinates.

examples/community/fresco_v2v.py

@@ -351,7 +351,7 @@ def forward(
             cross_attention_kwargs (`dict`, *optional*):
                 A kwargs dictionary that if specified is passed along to the [`AttnProcessor`].
             added_cond_kwargs: (`dict`, *optional*):
-                A kwargs dictionary containin additional embeddings that if specified are added to the embeddings that
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
                 are passed along to the UNet blocks.
 
         Returns:
@@ -864,9 +864,9 @@ def get_flow_and_interframe_paras(flow_model, imgs):
 class AttentionControl:
     """
     Control FRESCO-based attention
-    * enable/diable spatial-guided attention
-    * enable/diable temporal-guided attention
-    * enable/diable cross-frame attention
+    * enable/disable spatial-guided attention
+    * enable/disable temporal-guided attention
+    * enable/disable cross-frame attention
     * collect intermediate attention feature (for spatial-guided attention)
     """
 

examples/community/hd_painter.py

@@ -34,7 +34,7 @@ def __call__(
         temb: Optional[torch.Tensor] = None,
         scale: float = 1.0,
     ) -> torch.Tensor:
-        # Same as the default AttnProcessor up untill the part where similarity matrix gets saved
+        # Same as the default AttnProcessor up until the part where similarity matrix gets saved
         downscale_factor = self.mask_resoltuion // hidden_states.shape[1]
         residual = hidden_states