[LoRA] feat: lora support for SANA. (#10234)

* feat: lora support for SANA.

* make fix-copies

* rename test class.

* attention_kwargs -> cross_attention_kwargs.

* Revert "attention_kwargs -> cross_attention_kwargs."

This reverts commit 23433bf9bccc12e0f2f55df26bae58a894e8b43b.

* exhaust 119 max line limit

* sana lora fine-tuning script.

* readme

* add a note about the supported models.

* Apply suggestions from code review
Co-authored-by: Aryan <aryan@huggingface.co>

* style

* docs for attention_kwargs.

* remove lora_scale from pag pipeline.

* copy fix

---------
Co-authored-by: Aryan <aryan@huggingface.co>

examples/dreambooth/REAMDE_sana.md

+# DreamBooth training example for SANA
+
+[DreamBooth](https://arxiv.org/abs/2208.12242) is a method to personalize text2image models like stable diffusion given just a few (3~5) images of a subject.
+
+The `train_dreambooth_lora_sana.py` script shows how to implement the training procedure with [LoRA](https://huggingface.co/docs/peft/conceptual_guides/adapter#low-rank-adaptation-lora) and adapt it for [SANA](https://arxiv.org/abs/2410.10629). 
+
+
+This will also allow us to push the trained model parameters to the Hugging Face Hub platform.
+
+## Running locally with PyTorch
+
+### Installing the dependencies
+
+Before running the scripts, make sure to install the library's training dependencies:
+
+**Important**
+
+To make sure you can successfully run the latest versions of the example scripts, we highly recommend **installing from source** and keeping the install up to date as we update the example scripts frequently and install some example-specific requirements. To do this, execute the following steps in a new virtual environment:
+
+```bash
+git clone https://github.com/huggingface/diffusers
+cd diffusers
+pip install -e .
+```
+
+Then cd in the `examples/dreambooth` folder and run
+```bash
+pip install -r requirements_sana.txt
+```
+
+And initialize an [🤗Accelerate](https://github.com/huggingface/accelerate/) environment with:
+
+```bash
+accelerate config
+```
+
+Or for a default accelerate configuration without answering questions about your environment
+
+```bash
+accelerate config default
+```
+
+Or if your environment doesn't support an interactive shell (e.g., a notebook)
+
+```python
+from accelerate.utils import write_basic_config
+write_basic_config()
+```
+
+When running `accelerate config`, if we specify torch compile mode to True there can be dramatic speedups.
+Note also that we use PEFT library as backend for LoRA training, make sure to have `peft>=0.14.0` installed in your environment.
+
+
+### Dog toy example
+
+Now let's get our dataset. For this example we will use some dog images: https://huggingface.co/datasets/diffusers/dog-example.
+
+Let's first download it locally:
+
+```python
+from huggingface_hub import snapshot_download
+
+local_dir = "./dog"
+snapshot_download(
+    "diffusers/dog-example",
+    local_dir=local_dir, repo_type="dataset",
+    ignore_patterns=".gitattributes",
+)
+```
+
+This will also allow us to push the trained LoRA parameters to the Hugging Face Hub platform.
+
+Now, we can launch training using:
+
+```bash
+export MODEL_NAME="Efficient-Large-Model/Sana_1600M_1024px_diffusers"
+export INSTANCE_DIR="dog"
+export OUTPUT_DIR="trained-sana-lora"
+
+accelerate launch train_dreambooth_lora_sana.py \
+  --pretrained_model_name_or_path=$MODEL_NAME  \
+  --instance_data_dir=$INSTANCE_DIR \
+  --output_dir=$OUTPUT_DIR \
+  --mixed_precision="bf16" \
+  --instance_prompt="a photo of sks dog" \
+  --resolution=1024 \
+  --train_batch_size=1 \
+  --gradient_accumulation_steps=4 \
+  --use_8bit_adam \
+  --learning_rate=1e-4 \
+  --report_to="wandb" \
+  --lr_scheduler="constant" \
+  --lr_warmup_steps=0 \
+  --max_train_steps=500 \
+  --validation_prompt="A photo of sks dog in a bucket" \
+  --validation_epochs=25 \
+  --seed="0" \
+  --push_to_hub
+```
+
+For using `push_to_hub`, make you're logged into your Hugging Face account:
+
+```bash
+huggingface-cli login
+```
+
+To better track our training experiments, we're using the following flags in the command above:
+
+* `report_to="wandb` will ensure the training runs are tracked on [Weights and Biases](https://wandb.ai/site). To use it, be sure to install `wandb` with `pip install wandb`. Don't forget to call `wandb login <your_api_key>` before training if you haven't done it before.
+* `validation_prompt` and `validation_epochs` to allow the script to do a few validation inference runs. This allows us to qualitatively check if the training is progressing as expected.
+
+## Notes
+
+Additionally, we welcome you to explore the following CLI arguments:
+
+* `--lora_layers`: The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. E.g. - "to_k,to_q,to_v" will result in lora training of attention layers only.
+* `--complex_human_instruction`: Instructions for complex human attention as shown in [here](https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55).
+* `--max_sequence_length`: Maximum sequence length to use for text embeddings.
+
+
+We provide several options for optimizing memory optimization:
+
+* `--offload`: When enabled, we will offload the text encoder and VAE to CPU, when they are not used.
+* `cache_latents`: When enabled, we will pre-compute the latents from the input images with the VAE and remove the VAE from memory once done.
+* `--use_8bit_adam`: When enabled, we will use the 8bit version of AdamW provided by the `bitsandbytes` library.
+
+Refer to the [official documentation](https://huggingface.co/docs/diffusers/main/en/api/pipelines/sana) of the `SanaPipeline` to know more about the models available under the SANA family and their preferred dtypes during inference.
\ No newline at end of file

examples/dreambooth/requirements_sana.txt

+accelerate>=1.0.0
+torchvision
+transformers>=4.47.0
+ftfy
+tensorboard
+Jinja2
+peft>=0.14.0
+sentencepiece
\ No newline at end of file

examples/dreambooth/train_dreambooth_lora_sana.py

+#!/usr/bin/env python
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. 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
+
+import argparse
+import copy
+import itertools
+import logging
+import math
+import os
+import random
+import shutil
+import warnings
+from pathlib import Path
+
+import numpy as np
+import torch
+import torch.utils.checkpoint
+import transformers
+from accelerate import Accelerator
+from accelerate.logging import get_logger
+from accelerate.utils import DistributedDataParallelKwargs, ProjectConfiguration, set_seed
+from huggingface_hub import create_repo, upload_folder
+from huggingface_hub.utils import insecure_hashlib
+from peft import LoraConfig, set_peft_model_state_dict
+from peft.utils import get_peft_model_state_dict
+from PIL import Image
+from PIL.ImageOps import exif_transpose
+from torch.utils.data import Dataset
+from torchvision import transforms
+from torchvision.transforms.functional import crop
+from tqdm.auto import tqdm
+from transformers import AutoTokenizer, Gemma2Model
+
+import diffusers
+from diffusers import (
+    AutoencoderDC,
+    FlowMatchEulerDiscreteScheduler,
+    SanaPipeline,
+    SanaTransformer2DModel,
+)
+from diffusers.optimization import get_scheduler
+from diffusers.training_utils import (
+    cast_training_params,
+    compute_density_for_timestep_sampling,
+    compute_loss_weighting_for_sd3,
+    free_memory,
+)
+from diffusers.utils import (
+    check_min_version,
+    convert_unet_state_dict_to_peft,
+    is_wandb_available,
+)
+from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
+from diffusers.utils.torch_utils import is_compiled_module
+
+
+if is_wandb_available():
+    import wandb
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.32.0.dev0")
+
+logger = get_logger(__name__)
+
+
+def save_model_card(
+    repo_id: str,
+    images=None,
+    base_model: str = None,
+    instance_prompt=None,
+    validation_prompt=None,
+    repo_folder=None,
+):
+    widget_dict = []
+    if images is not None:
+        for i, image in enumerate(images):
+            image.save(os.path.join(repo_folder, f"image_{i}.png"))
+            widget_dict.append(
+                {"text": validation_prompt if validation_prompt else " ", "output": {"url": f"image_{i}.png"}}
+            )
+
+    model_description = f"""
+# Sana DreamBooth LoRA - {repo_id}
+
+<Gallery />
+
+## Model description
+
+These are {repo_id} DreamBooth LoRA weights for {base_model}.
+
+The weights were trained using [DreamBooth](https://dreambooth.github.io/) with the [Sana diffusers trainer](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_sana.md).
+
+
+## Trigger words
+
+You should use `{instance_prompt}` to trigger the image generation.
+
+## Download model
+
+[Download the *.safetensors LoRA]({repo_id}/tree/main) in the Files & versions tab.
+
+## Use it with the [🧨 diffusers library](https://github.com/huggingface/diffusers)
+
+```py
+TODO
+```
+
+For more details, including weighting, merging and fusing LoRAs, check the [documentation on loading LoRAs in diffusers](https://huggingface.co/docs/diffusers/main/en/using-diffusers/loading_adapters)
+
+## License
+
+TODO
+"""
+    model_card = load_or_create_model_card(
+        repo_id_or_path=repo_id,
+        from_training=True,
+        license="other",
+        base_model=base_model,
+        prompt=instance_prompt,
+        model_description=model_description,
+        widget=widget_dict,
+    )
+    tags = [
+        "text-to-image",
+        "diffusers-training",
+        "diffusers",
+        "lora",
+        "sana",
+        "sana-diffusers",
+        "template:sd-lora",
+    ]
+
+    model_card = populate_model_card(model_card, tags=tags)
+    model_card.save(os.path.join(repo_folder, "README.md"))
+
+
+def log_validation(
+    pipeline,
+    args,
+    accelerator,
+    pipeline_args,
+    epoch,
+    is_final_validation=False,
+):
+    logger.info(
+        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
+        f" {args.validation_prompt}."
+    )
+    pipeline.text_encoder = pipeline.text_encoder.to(torch.bfloat16)
+    pipeline = pipeline.to(accelerator.device)
+    pipeline.set_progress_bar_config(disable=True)
+
+    # run inference
+    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
+
+    images = [pipeline(**pipeline_args, generator=generator).images[0] for _ in range(args.num_validation_images)]
+
+    for tracker in accelerator.trackers:
+        phase_name = "test" if is_final_validation else "validation"
+        if tracker.name == "tensorboard":
+            np_images = np.stack([np.asarray(img) for img in images])
+            tracker.writer.add_images(phase_name, np_images, epoch, dataformats="NHWC")
+        if tracker.name == "wandb":
+            tracker.log(
+                {
+                    phase_name: [
+                        wandb.Image(image, caption=f"{i}: {args.validation_prompt}") for i, image in enumerate(images)
+                    ]
+                }
+            )
+
+    del pipeline
+    if torch.cuda.is_available():
+        torch.cuda.empty_cache()
+
+    return images
+
+
+def parse_args(input_args=None):
+    parser = argparse.ArgumentParser(description="Simple example of a training script.")
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default=None,
+        required=True,
+        help="Path to pretrained model or model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--revision",
+        type=str,
+        default=None,
+        required=False,
+        help="Revision of pretrained model identifier from huggingface.co/models.",
+    )
+    parser.add_argument(
+        "--variant",
+        type=str,
+        default=None,
+        help="Variant of the model files of the pretrained model identifier from huggingface.co/models, 'e.g.' fp16",
+    )
+    parser.add_argument(
+        "--dataset_name",
+        type=str,
+        default=None,
+        help=(
+            "The name of the Dataset (from the HuggingFace hub) containing the training data of instance images (could be your own, possibly private,"
+            " dataset). It can also be a path pointing to a local copy of a dataset in your filesystem,"
+            " or to a folder containing files that 🤗 Datasets can understand."
+        ),
+    )
+    parser.add_argument(
+        "--dataset_config_name",
+        type=str,
+        default=None,
+        help="The config of the Dataset, leave as None if there's only one config.",
+    )
+    parser.add_argument(
+        "--instance_data_dir",
+        type=str,
+        default=None,
+        help=("A folder containing the training data. "),
+    )
+
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default=None,
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+
+    parser.add_argument(
+        "--image_column",
+        type=str,
+        default="image",
+        help="The column of the dataset containing the target image. By "
+        "default, the standard Image Dataset maps out 'file_name' "
+        "to 'image'.",
+    )
+    parser.add_argument(
+        "--caption_column",
+        type=str,
+        default=None,
+        help="The column of the dataset containing the instance prompt for each image",
+    )
+
+    parser.add_argument("--repeats", type=int, default=1, help="How many times to repeat the training data.")
+
+    parser.add_argument(
+        "--class_data_dir",
+        type=str,
+        default=None,
+        required=False,
+        help="A folder containing the training data of class images.",
+    )
+    parser.add_argument(
+        "--instance_prompt",
+        type=str,
+        default=None,
+        required=True,
+        help="The prompt with identifier specifying the instance, e.g. 'photo of a TOK dog', 'in the style of TOK'",
+    )
+    parser.add_argument(
+        "--class_prompt",
+        type=str,
+        default=None,
+        help="The prompt to specify images in the same class as provided instance images.",
+    )
+    parser.add_argument(
+        "--max_sequence_length",
+        type=int,
+        default=300,
+        help="Maximum sequence length to use with with the Gemma model",
+    )
+    parser.add_argument(
+        "--complex_human_instruction",
+        type=str,
+        default=None,
+        help="Instructions for complex human attention: https://github.com/NVlabs/Sana/blob/main/configs/sana_app_config/Sana_1600M_app.yaml#L55.",
+    )
+    parser.add_argument(
+        "--validation_prompt",
+        type=str,
+        default=None,
+        help="A prompt that is used during validation to verify that the model is learning.",
+    )
+    parser.add_argument(
+        "--num_validation_images",
+        type=int,
+        default=4,
+        help="Number of images that should be generated during validation with `validation_prompt`.",
+    )
+    parser.add_argument(
+        "--validation_epochs",
+        type=int,
+        default=50,
+        help=(
+            "Run dreambooth validation every X epochs. Dreambooth validation consists of running the prompt"
+            " `args.validation_prompt` multiple times: `args.num_validation_images`."
+        ),
+    )
+    parser.add_argument(
+        "--rank",
+        type=int,
+        default=4,
+        help=("The dimension of the LoRA update matrices."),
+    )
+    parser.add_argument(
+        "--with_prior_preservation",
+        default=False,
+        action="store_true",
+        help="Flag to add prior preservation loss.",
+    )
+    parser.add_argument("--prior_loss_weight", type=float, default=1.0, help="The weight of prior preservation loss.")
+    parser.add_argument(
+        "--num_class_images",
+        type=int,
+        default=100,
+        help=(
+            "Minimal class images for prior preservation loss. If there are not enough images already present in"
+            " class_data_dir, additional images will be sampled with class_prompt."
+        ),
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="sana-dreambooth-lora",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--center_crop",
+        default=False,
+        action="store_true",
+        help=(
+            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
+            " cropped. The images will be resized to the resolution first before cropping."
+        ),
+    )
+    parser.add_argument(
+        "--random_flip",
+        action="store_true",
+        help="whether to randomly flip images horizontally",
+    )
+    parser.add_argument(
+        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
+    )
+    parser.add_argument(
+        "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
+    )
+    parser.add_argument("--num_train_epochs", type=int, default=1)
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=None,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=int,
+        default=500,
+        help=(
+            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
+            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
+            " training using `--resume_from_checkpoint`."
+        ),
+    )
+    parser.add_argument(
+        "--checkpoints_total_limit",
+        type=int,
+        default=None,
+        help=("Max number of checkpoints to store."),
+    )
+    parser.add_argument(
+        "--resume_from_checkpoint",
+        type=str,
+        default=None,
+        help=(
+            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
+            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
+        ),
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate",
+        type=float,
+        default=1e-4,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_true",
+        default=False,
+        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
+    )
+    parser.add_argument(
+        "--lr_scheduler",
+        type=str,
+        default="constant",
+        help=(
+            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
+            ' "constant", "constant_with_warmup"]'
+        ),
+    )
+    parser.add_argument(
+        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
+    )
+    parser.add_argument(
+        "--lr_num_cycles",
+        type=int,
+        default=1,
+        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
+    )
+    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
+    parser.add_argument(
+        "--dataloader_num_workers",
+        type=int,
+        default=0,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    parser.add_argument(
+        "--weighting_scheme",
+        type=str,
+        default="none",
+        choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"],
+        help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'),
+    )
+    parser.add_argument(
+        "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme."
+    )
+    parser.add_argument(
+        "--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme."
+    )
+    parser.add_argument(
+        "--mode_scale",
+        type=float,
+        default=1.29,
+        help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
+    )
+    parser.add_argument(
+        "--optimizer",
+        type=str,
+        default="AdamW",
+        help=('The optimizer type to use. Choose between ["AdamW", "prodigy"]'),
+    )
+
+    parser.add_argument(
+        "--use_8bit_adam",
+        action="store_true",
+        help="Whether or not to use 8-bit Adam from bitsandbytes. Ignored if optimizer is not set to AdamW",
+    )
+
+    parser.add_argument(
+        "--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam and Prodigy optimizers."
+    )
+    parser.add_argument(
+        "--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam and Prodigy optimizers."
+    )
+    parser.add_argument(
+        "--prodigy_beta3",
+        type=float,
+        default=None,
+        help="coefficients for computing the Prodigy stepsize using running averages. If set to None, "
+        "uses the value of square root of beta2. Ignored if optimizer is adamW",
+    )
+    parser.add_argument("--prodigy_decouple", type=bool, default=True, help="Use AdamW style decoupled weight decay")
+    parser.add_argument("--adam_weight_decay", type=float, default=1e-04, help="Weight decay to use for unet params")
+    parser.add_argument(
+        "--adam_weight_decay_text_encoder", type=float, default=1e-03, help="Weight decay to use for text_encoder"
+    )
+
+    parser.add_argument(
+        "--lora_layers",
+        type=str,
+        default=None,
+        help=(
+            'The transformer modules to apply LoRA training on. Please specify the layers in a comma seperated. E.g. - "to_k,to_q,to_v" will result in lora training of attention layers only'
+        ),
+    )
+
+    parser.add_argument(
+        "--adam_epsilon",
+        type=float,
+        default=1e-08,
+        help="Epsilon value for the Adam optimizer and Prodigy optimizers.",
+    )
+
+    parser.add_argument(
+        "--prodigy_use_bias_correction",
+        type=bool,
+        default=True,
+        help="Turn on Adam's bias correction. True by default. Ignored if optimizer is adamW",
+    )
+    parser.add_argument(
+        "--prodigy_safeguard_warmup",
+        type=bool,
+        default=True,
+        help="Remove lr from the denominator of D estimate to avoid issues during warm-up stage. True by default. "
+        "Ignored if optimizer is adamW",
+    )
+    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
+    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
+    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--allow_tf32",
+        action="store_true",
+        help=(
+            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
+            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
+        ),
+    )
+    parser.add_argument(
+        "--cache_latents",
+        action="store_true",
+        default=False,
+        help="Cache the VAE latents",
+    )
+    parser.add_argument(
+        "--report_to",
+        type=str,
+        default="tensorboard",
+        help=(
+            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
+            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default=None,
+        choices=["no", "fp16", "bf16"],
+        help=(
+            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
+            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
+            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
+        ),
+    )
+    parser.add_argument(
+        "--upcast_before_saving",
+        action="store_true",
+        default=False,
+        help=(
+            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
+            "Defaults to precision dtype used for training to save memory"
+        ),
+    )
+    parser.add_argument(
+        "--offload",
+        action="store_true",
+        help="Whether to offload the VAE and the text encoder to CPU when they are not used.",
+    )
+    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+
+    if input_args is not None:
+        args = parser.parse_args(input_args)
+    else:
+        args = parser.parse_args()
+
+    if args.dataset_name is None and args.instance_data_dir is None:
+        raise ValueError("Specify either `--dataset_name` or `--instance_data_dir`")
+
+    if args.dataset_name is not None and args.instance_data_dir is not None:
+        raise ValueError("Specify only one of `--dataset_name` or `--instance_data_dir`")
+
+    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
+    if env_local_rank != -1 and env_local_rank != args.local_rank:
+        args.local_rank = env_local_rank
+
+    if args.with_prior_preservation:
+        if args.class_data_dir is None:
+            raise ValueError("You must specify a data directory for class images.")
+        if args.class_prompt is None:
+            raise ValueError("You must specify prompt for class images.")
+    else:
+        # logger is not available yet
+        if args.class_data_dir is not None:
+            warnings.warn("You need not use --class_data_dir without --with_prior_preservation.")
+        if args.class_prompt is not None:
+            warnings.warn("You need not use --class_prompt without --with_prior_preservation.")
+
+    return args
+
+
+class DreamBoothDataset(Dataset):
+    """
+    A dataset to prepare the instance and class images with the prompts for fine-tuning the model.
+    It pre-processes the images.
+    """
+
+    def __init__(
+        self,
+        instance_data_root,
+        instance_prompt,
+        class_prompt,
+        class_data_root=None,
+        class_num=None,
+        size=1024,
+        repeats=1,
+        center_crop=False,
+    ):
+        self.size = size
+        self.center_crop = center_crop
+
+        self.instance_prompt = instance_prompt
+        self.custom_instance_prompts = None
+        self.class_prompt = class_prompt
+
+        # if --dataset_name is provided or a metadata jsonl file is provided in the local --instance_data directory,
+        # we load the training data using load_dataset
+        if args.dataset_name is not None:
+            try:
+                from datasets import load_dataset
+            except ImportError:
+                raise ImportError(
+                    "You are trying to load your data using the datasets library. If you wish to train using custom "
+                    "captions please install the datasets library: `pip install datasets`. If you wish to load a "
+                    "local folder containing images only, specify --instance_data_dir instead."
+                )
+            # Downloading and loading a dataset from the hub.
+            # See more about loading custom images at
+            # https://huggingface.co/docs/datasets/v2.0.0/en/dataset_script
+            dataset = load_dataset(
+                args.dataset_name,
+                args.dataset_config_name,
+                cache_dir=args.cache_dir,
+            )
+            # Preprocessing the datasets.
+            column_names = dataset["train"].column_names
+
+            # 6. Get the column names for input/target.
+            if args.image_column is None:
+                image_column = column_names[0]
+                logger.info(f"image column defaulting to {image_column}")
+            else:
+                image_column = args.image_column
+                if image_column not in column_names:
+                    raise ValueError(
+                        f"`--image_column` value '{args.image_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+                    )
+            instance_images = dataset["train"][image_column]
+
+            if args.caption_column is None:
+                logger.info(
+                    "No caption column provided, defaulting to instance_prompt for all images. If your dataset "
+                    "contains captions/prompts for the images, make sure to specify the "
+                    "column as --caption_column"
+                )
+                self.custom_instance_prompts = None
+            else:
+                if args.caption_column not in column_names:
+                    raise ValueError(
+                        f"`--caption_column` value '{args.caption_column}' not found in dataset columns. Dataset columns are: {', '.join(column_names)}"
+                    )
+                custom_instance_prompts = dataset["train"][args.caption_column]
+                # create final list of captions according to --repeats
+                self.custom_instance_prompts = []
+                for caption in custom_instance_prompts:
+                    self.custom_instance_prompts.extend(itertools.repeat(caption, repeats))
+        else:
+            self.instance_data_root = Path(instance_data_root)
+            if not self.instance_data_root.exists():
+                raise ValueError("Instance images root doesn't exists.")
+
+            instance_images = [Image.open(path) for path in list(Path(instance_data_root).iterdir())]
+            self.custom_instance_prompts = None
+
+        self.instance_images = []
+        for img in instance_images:
+            self.instance_images.extend(itertools.repeat(img, repeats))
+
+        self.pixel_values = []
+        train_resize = transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR)
+        train_crop = transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size)
+        train_flip = transforms.RandomHorizontalFlip(p=1.0)
+        train_transforms = transforms.Compose(
+            [
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+        for image in self.instance_images:
+            image = exif_transpose(image)
+            if not image.mode == "RGB":
+                image = image.convert("RGB")
+            image = train_resize(image)
+            if args.random_flip and random.random() < 0.5:
+                # flip
+                image = train_flip(image)
+            if args.center_crop:
+                y1 = max(0, int(round((image.height - args.resolution) / 2.0)))
+                x1 = max(0, int(round((image.width - args.resolution) / 2.0)))
+                image = train_crop(image)
+            else:
+                y1, x1, h, w = train_crop.get_params(image, (args.resolution, args.resolution))
+                image = crop(image, y1, x1, h, w)
+            image = train_transforms(image)
+            self.pixel_values.append(image)
+
+        self.num_instance_images = len(self.instance_images)
+        self._length = self.num_instance_images
+
+        if class_data_root is not None:
+            self.class_data_root = Path(class_data_root)
+            self.class_data_root.mkdir(parents=True, exist_ok=True)
+            self.class_images_path = list(self.class_data_root.iterdir())
+            if class_num is not None:
+                self.num_class_images = min(len(self.class_images_path), class_num)
+            else:
+                self.num_class_images = len(self.class_images_path)
+            self._length = max(self.num_class_images, self.num_instance_images)
+        else:
+            self.class_data_root = None
+
+        self.image_transforms = transforms.Compose(
+            [
+                transforms.Resize(size, interpolation=transforms.InterpolationMode.BILINEAR),
+                transforms.CenterCrop(size) if center_crop else transforms.RandomCrop(size),
+                transforms.ToTensor(),
+                transforms.Normalize([0.5], [0.5]),
+            ]
+        )
+
+    def __len__(self):
+        return self._length
+
+    def __getitem__(self, index):
+        example = {}
+        instance_image = self.pixel_values[index % self.num_instance_images]
+        example["instance_images"] = instance_image
+
+        if self.custom_instance_prompts:
+            caption = self.custom_instance_prompts[index % self.num_instance_images]
+            if caption:
+                example["instance_prompt"] = caption
+            else:
+                example["instance_prompt"] = self.instance_prompt
+
+        else:  # custom prompts were provided, but length does not match size of image dataset
+            example["instance_prompt"] = self.instance_prompt
+
+        if self.class_data_root:
+            class_image = Image.open(self.class_images_path[index % self.num_class_images])
+            class_image = exif_transpose(class_image)
+
+            if not class_image.mode == "RGB":
+                class_image = class_image.convert("RGB")
+            example["class_images"] = self.image_transforms(class_image)
+            example["class_prompt"] = self.class_prompt
+
+        return example
+
+
+def collate_fn(examples, with_prior_preservation=False):
+    pixel_values = [example["instance_images"] for example in examples]
+    prompts = [example["instance_prompt"] for example in examples]
+
+    # Concat class and instance examples for prior preservation.
+    # We do this to avoid doing two forward passes.
+    if with_prior_preservation:
+        pixel_values += [example["class_images"] for example in examples]
+        prompts += [example["class_prompt"] for example in examples]
+
+    pixel_values = torch.stack(pixel_values)
+    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+
+    batch = {"pixel_values": pixel_values, "prompts": prompts}
+    return batch
+
+
+class PromptDataset(Dataset):
+    "A simple dataset to prepare the prompts to generate class images on multiple GPUs."
+
+    def __init__(self, prompt, num_samples):
+        self.prompt = prompt
+        self.num_samples = num_samples
+
+    def __len__(self):
+        return self.num_samples
+
+    def __getitem__(self, index):
+        example = {}
+        example["prompt"] = self.prompt
+        example["index"] = index
+        return example
+
+
+def main(args):
+    if args.report_to == "wandb" and args.hub_token is not None:
+        raise ValueError(
+            "You cannot use both --report_to=wandb and --hub_token due to a security risk of exposing your token."
+            " Please use `huggingface-cli login` to authenticate with the Hub."
+        )
+
+    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
+    logging_dir = Path(args.output_dir, args.logging_dir)
+
+    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
+    kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
+    accelerator = Accelerator(
+        gradient_accumulation_steps=args.gradient_accumulation_steps,
+        mixed_precision=args.mixed_precision,
+        log_with=args.report_to,
+        project_config=accelerator_project_config,
+        kwargs_handlers=[kwargs],
+    )
+
+    # Disable AMP for MPS.
+    if torch.backends.mps.is_available():
+        accelerator.native_amp = False
+
+    if args.report_to == "wandb":
+        if not is_wandb_available():
+            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")
+
+    # Make one log on every process with the configuration for debugging.
+    logging.basicConfig(
+        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
+        datefmt="%m/%d/%Y %H:%M:%S",
+        level=logging.INFO,
+    )
+    logger.info(accelerator.state, main_process_only=False)
+    if accelerator.is_local_main_process:
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        transformers.utils.logging.set_verbosity_error()
+        diffusers.utils.logging.set_verbosity_error()
+
+    # If passed along, set the training seed now.
+    if args.seed is not None:
+        set_seed(args.seed)
+
+    # Generate class images if prior preservation is enabled.
+    if args.with_prior_preservation:
+        class_images_dir = Path(args.class_data_dir)
+        if not class_images_dir.exists():
+            class_images_dir.mkdir(parents=True)
+        cur_class_images = len(list(class_images_dir.iterdir()))
+
+        if cur_class_images < args.num_class_images:
+            pipeline = SanaPipeline.from_pretrained(
+                args.pretrained_model_name_or_path,
+                torch_dtype=torch.float32,
+                revision=args.revision,
+                variant=args.variant,
+            )
+            pipeline.text_encoder = pipeline.text_encoder.to(torch.bfloat16)
+            pipeline.transformer = pipeline.transformer.to(torch.float16)
+            pipeline.set_progress_bar_config(disable=True)
+
+            num_new_images = args.num_class_images - cur_class_images
+            logger.info(f"Number of class images to sample: {num_new_images}.")
+
+            sample_dataset = PromptDataset(args.class_prompt, num_new_images)
+            sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)
+
+            sample_dataloader = accelerator.prepare(sample_dataloader)
+            pipeline.to(accelerator.device)
+
+            for example in tqdm(
+                sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
+            ):
+                images = pipeline(example["prompt"]).images
+
+                for i, image in enumerate(images):
+                    hash_image = insecure_hashlib.sha1(image.tobytes()).hexdigest()
+                    image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
+                    image.save(image_filename)
+
+            del pipeline
+            if torch.cuda.is_available():
+                torch.cuda.empty_cache()
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+
+        if args.push_to_hub:
+            repo_id = create_repo(
+                repo_id=args.hub_model_id or Path(args.output_dir).name,
+                exist_ok=True,
+            ).repo_id
+
+    # Load the tokenizer
+    tokenizer = AutoTokenizer.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="tokenizer",
+        revision=args.revision,
+    )
+
+    # Load scheduler and models
+    noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="scheduler"
+    )
+    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
+    text_encoder = Gemma2Model.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
+    )
+    vae = AutoencoderDC.from_pretrained(
+        args.pretrained_model_name_or_path,
+        subfolder="vae",
+        revision=args.revision,
+        variant=args.variant,
+    )
+    transformer = SanaTransformer2DModel.from_pretrained(
+        args.pretrained_model_name_or_path, subfolder="transformer", revision=args.revision, variant=args.variant
+    )
+
+    # We only train the additional adapter LoRA layers
+    transformer.requires_grad_(False)
+    vae.requires_grad_(False)
+    text_encoder.requires_grad_(False)
+
+    # Initialize a text encoding pipeline and keep it to CPU for now.
+    text_encoding_pipeline = SanaPipeline.from_pretrained(
+        args.pretrained_model_name_or_path,
+        vae=None,
+        transformer=None,
+        text_encoder=text_encoder,
+        tokenizer=tokenizer,
+    )
+
+    # For mixed precision training we cast all non-trainable weights (vae, text_encoder and transformer) to half-precision
+    # as these weights are only used for inference, keeping weights in full precision is not required.
+    weight_dtype = torch.float32
+    if accelerator.mixed_precision == "fp16":
+        weight_dtype = torch.float16
+    elif accelerator.mixed_precision == "bf16":
+        weight_dtype = torch.bfloat16
+
+    if torch.backends.mps.is_available() and weight_dtype == torch.bfloat16:
+        # due to pytorch#99272, MPS does not yet support bfloat16.
+        raise ValueError(
+            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
+        )
+
+    # VAE should always be kept in fp32 for SANA (?)
+    vae.to(dtype=torch.float32)
+    transformer.to(accelerator.device, dtype=weight_dtype)
+    # because Gemma2 is particularly suited for bfloat16.
+    text_encoder.to(dtype=torch.bfloat16)
+
+    if args.gradient_checkpointing:
+        transformer.enable_gradient_checkpointing()
+
+    if args.lora_layers is not None:
+        target_modules = [layer.strip() for layer in args.lora_layers.split(",")]
+    else:
+        target_modules = ["to_k", "to_q", "to_v"]
+
+    # now we will add new LoRA weights the transformer layers
+    transformer_lora_config = LoraConfig(
+        r=args.rank,
+        lora_alpha=args.rank,
+        init_lora_weights="gaussian",
+        target_modules=target_modules,
+    )
+    transformer.add_adapter(transformer_lora_config)
+
+    def unwrap_model(model):
+        model = accelerator.unwrap_model(model)
+        model = model._orig_mod if is_compiled_module(model) else model
+        return model
+
+    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
+    def save_model_hook(models, weights, output_dir):
+        if accelerator.is_main_process:
+            transformer_lora_layers_to_save = None
+
+            for model in models:
+                if isinstance(model, type(unwrap_model(transformer))):
+                    transformer_lora_layers_to_save = get_peft_model_state_dict(model)
+                else:
+                    raise ValueError(f"unexpected save model: {model.__class__}")
+
+                # make sure to pop weight so that corresponding model is not saved again
+                weights.pop()
+
+            SanaPipeline.save_lora_weights(
+                output_dir,
+                transformer_lora_layers=transformer_lora_layers_to_save,
+            )
+
+    def load_model_hook(models, input_dir):
+        transformer_ = None
+
+        while len(models) > 0:
+            model = models.pop()
+
+            if isinstance(model, type(unwrap_model(transformer))):
+                transformer_ = model
+            else:
+                raise ValueError(f"unexpected save model: {model.__class__}")
+
+        lora_state_dict = SanaPipeline.lora_state_dict(input_dir)
+
+        transformer_state_dict = {
+            f'{k.replace("transformer.", "")}': v for k, v in lora_state_dict.items() if k.startswith("transformer.")
+        }
+        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
+        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
+        if incompatible_keys is not None:
+            # check only for unexpected keys
+            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
+            if unexpected_keys:
+                logger.warning(
+                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
+                    f" {unexpected_keys}. "
+                )
+
+        # Make sure the trainable params are in float32. This is again needed since the base models
+        # are in `weight_dtype`. More details:
+        # https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
+        if args.mixed_precision == "fp16":
+            models = [transformer_]
+            # only upcast trainable parameters (LoRA) into fp32
+            cast_training_params(models)
+
+    accelerator.register_save_state_pre_hook(save_model_hook)
+    accelerator.register_load_state_pre_hook(load_model_hook)
+
+    # Enable TF32 for faster training on Ampere GPUs,
+    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
+    if args.allow_tf32 and torch.cuda.is_available():
+        torch.backends.cuda.matmul.allow_tf32 = True
+
+    if args.scale_lr:
+        args.learning_rate = (
+            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
+        )
+
+    # Make sure the trainable params are in float32.
+    if args.mixed_precision == "fp16":
+        models = [transformer]
+        # only upcast trainable parameters (LoRA) into fp32
+        cast_training_params(models, dtype=torch.float32)
+
+    transformer_lora_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))
+
+    # Optimization parameters
+    transformer_parameters_with_lr = {"params": transformer_lora_parameters, "lr": args.learning_rate}
+    params_to_optimize = [transformer_parameters_with_lr]
+
+    # Optimizer creation
+    if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
+        logger.warning(
+            f"Unsupported choice of optimizer: {args.optimizer}.Supported optimizers include [adamW, prodigy]."
+            "Defaulting to adamW"
+        )
+        args.optimizer = "adamw"
+
+    if args.use_8bit_adam and not args.optimizer.lower() == "adamw":
+        logger.warning(
+            f"use_8bit_adam is ignored when optimizer is not set to 'AdamW'. Optimizer was "
+            f"set to {args.optimizer.lower()}"
+        )
+
+    if args.optimizer.lower() == "adamw":
+        if args.use_8bit_adam:
+            try:
+                import bitsandbytes as bnb
+            except ImportError:
+                raise ImportError(
+                    "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
+                )
+
+            optimizer_class = bnb.optim.AdamW8bit
+        else:
+            optimizer_class = torch.optim.AdamW
+
+        optimizer = optimizer_class(
+            params_to_optimize,
+            betas=(args.adam_beta1, args.adam_beta2),
+            weight_decay=args.adam_weight_decay,
+            eps=args.adam_epsilon,
+        )
+
+    if args.optimizer.lower() == "prodigy":
+        try:
+            import prodigyopt
+        except ImportError:
+            raise ImportError("To use Prodigy, please install the prodigyopt library: `pip install prodigyopt`")
+
+        optimizer_class = prodigyopt.Prodigy
+
+        if args.learning_rate <= 0.1:
+            logger.warning(
+                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
+            )
+
+        optimizer = optimizer_class(
+            params_to_optimize,
+            betas=(args.adam_beta1, args.adam_beta2),
+            beta3=args.prodigy_beta3,
+            weight_decay=args.adam_weight_decay,
+            eps=args.adam_epsilon,
+            decouple=args.prodigy_decouple,
+            use_bias_correction=args.prodigy_use_bias_correction,
+            safeguard_warmup=args.prodigy_safeguard_warmup,
+        )
+
+    # Dataset and DataLoaders creation:
+    train_dataset = DreamBoothDataset(
+        instance_data_root=args.instance_data_dir,
+        instance_prompt=args.instance_prompt,
+        class_prompt=args.class_prompt,
+        class_data_root=args.class_data_dir if args.with_prior_preservation else None,
+        class_num=args.num_class_images,
+        size=args.resolution,
+        repeats=args.repeats,
+        center_crop=args.center_crop,
+    )
+
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        batch_size=args.train_batch_size,
+        shuffle=True,
+        collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
+        num_workers=args.dataloader_num_workers,
+    )
+
+    def compute_text_embeddings(prompt, text_encoding_pipeline):
+        text_encoding_pipeline = text_encoding_pipeline.to(accelerator.device)
+        with torch.no_grad():
+            prompt_embeds, prompt_attention_mask, _, _ = text_encoding_pipeline.encode_prompt(
+                prompt,
+                max_sequence_length=args.max_sequence_length,
+                complex_human_instruction=args.complex_human_instruction,
+            )
+        if args.offload:
+            text_encoding_pipeline = text_encoding_pipeline.to("cpu")
+        return prompt_embeds, prompt_attention_mask
+
+    # If no type of tuning is done on the text_encoder and custom instance prompts are NOT
+    # provided (i.e. the --instance_prompt is used for all images), we encode the instance prompt once to avoid
+    # the redundant encoding.
+    if not train_dataset.custom_instance_prompts:
+        instance_prompt_hidden_states, instance_prompt_attention_mask = compute_text_embeddings(
+            args.instance_prompt, text_encoding_pipeline
+        )
+
+    # Handle class prompt for prior-preservation.
+    if args.with_prior_preservation:
+        class_prompt_hidden_states, class_prompt_attention_mask = compute_text_embeddings(
+            args.class_prompt, text_encoding_pipeline
+        )
+
+    # Clear the memory here
+    if not train_dataset.custom_instance_prompts:
+        del text_encoder, tokenizer
+        free_memory()
+
+    # If custom instance prompts are NOT provided (i.e. the instance prompt is used for all images),
+    # pack the statically computed variables appropriately here. This is so that we don't
+    # have to pass them to the dataloader.
+    if not train_dataset.custom_instance_prompts:
+        prompt_embeds = instance_prompt_hidden_states
+        prompt_attention_mask = instance_prompt_attention_mask
+        if args.with_prior_preservation:
+            prompt_embeds = torch.cat([prompt_embeds, class_prompt_hidden_states], dim=0)
+            prompt_attention_mask = torch.cat([prompt_attention_mask, class_prompt_attention_mask], dim=0)
+
+    vae_config_scaling_factor = vae.config.scaling_factor
+    if args.cache_latents:
+        latents_cache = []
+        vae = vae.to("cuda")
+        for batch in tqdm(train_dataloader, desc="Caching latents"):
+            with torch.no_grad():
+                batch["pixel_values"] = batch["pixel_values"].to(
+                    accelerator.device, non_blocking=True, dtype=vae.dtype
+                )
+                latents_cache.append(vae.encode(batch["pixel_values"]).latent)
+
+        if args.validation_prompt is None:
+            del vae
+            free_memory()
+
+    # Scheduler and math around the number of training steps.
+    overrode_max_train_steps = False
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if args.max_train_steps is None:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+        overrode_max_train_steps = True
+
+    lr_scheduler = get_scheduler(
+        args.lr_scheduler,
+        optimizer=optimizer,
+        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
+        num_training_steps=args.max_train_steps * accelerator.num_processes,
+        num_cycles=args.lr_num_cycles,
+        power=args.lr_power,
+    )
+
+    # Prepare everything with our `accelerator`.
+    transformer, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        transformer, optimizer, train_dataloader, lr_scheduler
+    )
+
+    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
+    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
+    if overrode_max_train_steps:
+        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
+    # Afterwards we recalculate our number of training epochs
+    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)
+
+    # We need to initialize the trackers we use, and also store our configuration.
+    # The trackers initializes automatically on the main process.
+    if accelerator.is_main_process:
+        tracker_name = "dreambooth-sana-lora"
+        accelerator.init_trackers(tracker_name, config=vars(args))
+
+    # Train!
+    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps
+
+    logger.info("***** Running training *****")
+    logger.info(f"  Num examples = {len(train_dataset)}")
+    logger.info(f"  Num batches each epoch = {len(train_dataloader)}")
+    logger.info(f"  Num Epochs = {args.num_train_epochs}")
+    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
+    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
+    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
+    logger.info(f"  Total optimization steps = {args.max_train_steps}")
+    global_step = 0
+    first_epoch = 0
+
+    # Potentially load in the weights and states from a previous save
+    if args.resume_from_checkpoint:
+        if args.resume_from_checkpoint != "latest":
+            path = os.path.basename(args.resume_from_checkpoint)
+        else:
+            # Get the mos recent checkpoint
+            dirs = os.listdir(args.output_dir)
+            dirs = [d for d in dirs if d.startswith("checkpoint")]
+            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
+            path = dirs[-1] if len(dirs) > 0 else None
+
+        if path is None:
+            accelerator.print(
+                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
+            )
+            args.resume_from_checkpoint = None
+            initial_global_step = 0
+        else:
+            accelerator.print(f"Resuming from checkpoint {path}")
+            accelerator.load_state(os.path.join(args.output_dir, path))
+            global_step = int(path.split("-")[1])
+
+            initial_global_step = global_step
+            first_epoch = global_step // num_update_steps_per_epoch
+
+    else:
+        initial_global_step = 0
+
+    progress_bar = tqdm(
+        range(0, args.max_train_steps),
+        initial=initial_global_step,
+        desc="Steps",
+        # Only show the progress bar once on each machine.
+        disable=not accelerator.is_local_main_process,
+    )
+
+    def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
+        sigmas = noise_scheduler_copy.sigmas.to(device=accelerator.device, dtype=dtype)
+        schedule_timesteps = noise_scheduler_copy.timesteps.to(accelerator.device)
+        timesteps = timesteps.to(accelerator.device)
+        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]
+
+        sigma = sigmas[step_indices].flatten()
+        while len(sigma.shape) < n_dim:
+            sigma = sigma.unsqueeze(-1)
+        return sigma
+
+    for epoch in range(first_epoch, args.num_train_epochs):
+        transformer.train()
+
+        for step, batch in enumerate(train_dataloader):
+            models_to_accumulate = [transformer]
+            with accelerator.accumulate(models_to_accumulate):
+                prompts = batch["prompts"]
+
+                # encode batch prompts when custom prompts are provided for each image -
+                if train_dataset.custom_instance_prompts:
+                    prompt_embeds, prompt_attention_mask = compute_text_embeddings(prompts, text_encoding_pipeline)
+
+                # Convert images to latent space
+                if args.cache_latents:
+                    model_input = latents_cache[step]
+                else:
+                    vae = vae.to(accelerator.device)
+                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
+                    model_input = vae.encode(pixel_values).latent
+                    if args.offload:
+                        vae = vae.to("cpu")
+                model_input = model_input * vae_config_scaling_factor
+                model_input = model_input.to(dtype=weight_dtype)
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(model_input)
+                bsz = model_input.shape[0]
+
+                # Sample a random timestep for each image
+                # for weighting schemes where we sample timesteps non-uniformly
+                u = compute_density_for_timestep_sampling(
+                    weighting_scheme=args.weighting_scheme,
+                    batch_size=bsz,
+                    logit_mean=args.logit_mean,
+                    logit_std=args.logit_std,
+                    mode_scale=args.mode_scale,
+                )
+                indices = (u * noise_scheduler_copy.config.num_train_timesteps).long()
+                timesteps = noise_scheduler_copy.timesteps[indices].to(device=model_input.device)
+
+                # Add noise according to flow matching.
+                # zt = (1 - texp) * x + texp * z1
+                sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
+                noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise
+
+                # Predict the noise residual
+                model_pred = transformer(
+                    hidden_states=noisy_model_input,
+                    encoder_hidden_states=prompt_embeds,
+                    encoder_attention_mask=prompt_attention_mask,
+                    timestep=timesteps,
+                    return_dict=False,
+                )[0]
+
+                # these weighting schemes use a uniform timestep sampling
+                # and instead post-weight the loss
+                weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas)
+
+                # flow matching loss
+                target = noise - model_input
+
+                if args.with_prior_preservation:
+                    # Chunk the noise and model_pred into two parts and compute the loss on each part separately.
+                    model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
+                    target, target_prior = torch.chunk(target, 2, dim=0)
+
+                    # Compute prior loss
+                    prior_loss = torch.mean(
+                        (weighting.float() * (model_pred_prior.float() - target_prior.float()) ** 2).reshape(
+                            target_prior.shape[0], -1
+                        ),
+                        1,
+                    )
+                    prior_loss = prior_loss.mean()
+
+                # Compute regular loss.
+                loss = torch.mean(
+                    (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(target.shape[0], -1),
+                    1,
+                )
+                loss = loss.mean()
+
+                if args.with_prior_preservation:
+                    # Add the prior loss to the instance loss.
+                    loss = loss + args.prior_loss_weight * prior_loss
+
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    params_to_clip = transformer.parameters()
+                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
+
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has performed an optimization step behind the scenes
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+
+                if accelerator.is_main_process:
+                    if global_step % args.checkpointing_steps == 0:
+                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
+                        if args.checkpoints_total_limit is not None:
+                            checkpoints = os.listdir(args.output_dir)
+                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
+                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))
+
+                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
+                            if len(checkpoints) >= args.checkpoints_total_limit:
+                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
+                                removing_checkpoints = checkpoints[0:num_to_remove]
+
+                                logger.info(
+                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
+                                )
+                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")
+
+                                for removing_checkpoint in removing_checkpoints:
+                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
+                                    shutil.rmtree(removing_checkpoint)
+
+                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
+                        accelerator.save_state(save_path)
+                        logger.info(f"Saved state to {save_path}")
+
+            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+            progress_bar.set_postfix(**logs)
+            accelerator.log(logs, step=global_step)
+
+            if global_step >= args.max_train_steps:
+                break
+
+        if accelerator.is_main_process:
+            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
+                # create pipeline
+                pipeline = SanaPipeline.from_pretrained(
+                    args.pretrained_model_name_or_path,
+                    transformer=accelerator.unwrap_model(transformer),
+                    revision=args.revision,
+                    variant=args.variant,
+                    torch_dtype=torch.float32,
+                )
+                pipeline_args = {
+                    "prompt": args.validation_prompt,
+                    "complex_human_instruction": args.complex_human_instruction,
+                }
+                images = log_validation(
+                    pipeline=pipeline,
+                    args=args,
+                    accelerator=accelerator,
+                    pipeline_args=pipeline_args,
+                    epoch=epoch,
+                )
+                free_memory()
+
+                images = None
+                del pipeline
+
+    # Save the lora layers
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        transformer = unwrap_model(transformer)
+        if args.upcast_before_saving:
+            transformer.to(torch.float32)
+        else:
+            transformer = transformer.to(weight_dtype)
+        transformer_lora_layers = get_peft_model_state_dict(transformer)
+
+        SanaPipeline.save_lora_weights(
+            save_directory=args.output_dir,
+            transformer_lora_layers=transformer_lora_layers,
+        )
+
+        # Final inference
+        # Load previous pipeline
+        pipeline = SanaPipeline.from_pretrained(
+            args.pretrained_model_name_or_path,
+            revision=args.revision,
+            variant=args.variant,
+            torch_dtype=torch.float32,
+        )
+        pipeline.transformer = pipeline.transformer.to(torch.float16)
+        # load attention processors
+        pipeline.load_lora_weights(args.output_dir)
+
+        # run inference
+        images = []
+        if args.validation_prompt and args.num_validation_images > 0:
+            pipeline_args = {
+                "prompt": args.validation_prompt,
+                "complex_human_instruction": args.complex_human_instruction,
+            }
+            images = log_validation(
+                pipeline=pipeline,
+                args=args,
+                accelerator=accelerator,
+                pipeline_args=pipeline_args,
+                epoch=epoch,
+                is_final_validation=True,
+            )
+
+        if args.push_to_hub:
+            save_model_card(
+                repo_id,
+                images=images,
+                base_model=args.pretrained_model_name_or_path,
+                instance_prompt=args.instance_prompt,
+                validation_prompt=args.validation_prompt,
+                repo_folder=args.output_dir,
+            )
+            upload_folder(
+                repo_id=repo_id,
+                folder_path=args.output_dir,
+                commit_message="End of training",
+                ignore_patterns=["step_*", "epoch_*"],
+            )
+
+        images = None
+        del pipeline
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    args = parse_args()
+    main(args)

src/diffusers/loaders/__init__.py

@@ -70,6 +70,7 @@ if is_torch_available():
             "FluxLoraLoaderMixin",
             "CogVideoXLoraLoaderMixin",
             "Mochi1LoraLoaderMixin",
+            "SanaLoraLoaderMixin",
         ]
         _import_structure["textual_inversion"] = ["TextualInversionLoaderMixin"]
         _import_structure["ip_adapter"] = ["IPAdapterMixin"]
@@ -92,6 +93,7 @@ if TYPE_CHECKING or DIFFUSERS_SLOW_IMPORT:
                 LoraLoaderMixin,
                 LTXVideoLoraLoaderMixin,
                 Mochi1LoraLoaderMixin,
+                SanaLoraLoaderMixin,
                 SD3LoraLoaderMixin,
                 StableDiffusionLoraLoaderMixin,
                 StableDiffusionXLLoraLoaderMixin,

src/diffusers/loaders/lora_pipeline.py

@@ -3562,6 +3562,314 @@ class LTXVideoLoraLoaderMixin(LoraBaseMixin):
         super().unfuse_lora(components=components)
 
 
+class SanaLoraLoaderMixin(LoraBaseMixin):
+    r"""
+    Load LoRA layers into [`SanaTransformer2DModel`]. Specific to [`SanaPipeline`].
+    """
+
+    _lora_loadable_modules = ["transformer"]
+    transformer_name = TRANSFORMER_NAME
+
+    @classmethod
+    @validate_hf_hub_args
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.lora_state_dict
+    def lora_state_dict(
+        cls,
+        pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]],
+        **kwargs,
+    ):
+        r"""
+        Return state dict for lora weights and the network alphas.
+
+        <Tip warning={true}>
+
+        We support loading A1111 formatted LoRA checkpoints in a limited capacity.
+
+        This function is experimental and might change in the future.
+
+        </Tip>
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                Can be either:
+
+                    - A string, the *model id* (for example `google/ddpm-celebahq-256`) of a pretrained model hosted on
+                      the Hub.
+                    - A path to a *directory* (for example `./my_model_directory`) containing the model weights saved
+                      with [`ModelMixin.save_pretrained`].
+                    - A [torch state
+                      dict](https://pytorch.org/tutorials/beginner/saving_loading_models.html#what-is-a-state-dict).
+
+            cache_dir (`Union[str, os.PathLike]`, *optional*):
+                Path to a directory where a downloaded pretrained model configuration is cached if the standard cache
+                is not used.
+            force_download (`bool`, *optional*, defaults to `False`):
+                Whether or not to force the (re-)download of the model weights and configuration files, overriding the
+                cached versions if they exist.
+
+            proxies (`Dict[str, str]`, *optional*):
+                A dictionary of proxy servers to use by protocol or endpoint, for example, `{'http': 'foo.bar:3128',
+                'http://hostname': 'foo.bar:4012'}`. The proxies are used on each request.
+            local_files_only (`bool`, *optional*, defaults to `False`):
+                Whether to only load local model weights and configuration files or not. If set to `True`, the model
+                won't be downloaded from the Hub.
+            token (`str` or *bool*, *optional*):
+                The token to use as HTTP bearer authorization for remote files. If `True`, the token generated from
+                `diffusers-cli login` (stored in `~/.huggingface`) is used.
+            revision (`str`, *optional*, defaults to `"main"`):
+                The specific model version to use. It can be a branch name, a tag name, a commit id, or any identifier
+                allowed by Git.
+            subfolder (`str`, *optional*, defaults to `""`):
+                The subfolder location of a model file within a larger model repository on the Hub or locally.
+
+        """
+        # Load the main state dict first which has the LoRA layers for either of
+        # transformer and text encoder or both.
+        cache_dir = kwargs.pop("cache_dir", None)
+        force_download = kwargs.pop("force_download", False)
+        proxies = kwargs.pop("proxies", None)
+        local_files_only = kwargs.pop("local_files_only", None)
+        token = kwargs.pop("token", None)
+        revision = kwargs.pop("revision", None)
+        subfolder = kwargs.pop("subfolder", None)
+        weight_name = kwargs.pop("weight_name", None)
+        use_safetensors = kwargs.pop("use_safetensors", None)
+
+        allow_pickle = False
+        if use_safetensors is None:
+            use_safetensors = True
+            allow_pickle = True
+
+        user_agent = {
+            "file_type": "attn_procs_weights",
+            "framework": "pytorch",
+        }
+
+        state_dict = _fetch_state_dict(
+            pretrained_model_name_or_path_or_dict=pretrained_model_name_or_path_or_dict,
+            weight_name=weight_name,
+            use_safetensors=use_safetensors,
+            local_files_only=local_files_only,
+            cache_dir=cache_dir,
+            force_download=force_download,
+            proxies=proxies,
+            token=token,
+            revision=revision,
+            subfolder=subfolder,
+            user_agent=user_agent,
+            allow_pickle=allow_pickle,
+        )
+
+        is_dora_scale_present = any("dora_scale" in k for k in state_dict)
+        if is_dora_scale_present:
+            warn_msg = "It seems like you are using a DoRA checkpoint that is not compatible in Diffusers at the moment. So, we are going to filter out the keys associated to 'dora_scale` from the state dict. If you think this is a mistake please open an issue https://github.com/huggingface/diffusers/issues/new."
+            logger.warning(warn_msg)
+            state_dict = {k: v for k, v in state_dict.items() if "dora_scale" not in k}
+
+        return state_dict
+
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.load_lora_weights
+    def load_lora_weights(
+        self, pretrained_model_name_or_path_or_dict: Union[str, Dict[str, torch.Tensor]], adapter_name=None, **kwargs
+    ):
+        """
+        Load LoRA weights specified in `pretrained_model_name_or_path_or_dict` into `self.transformer` and
+        `self.text_encoder`. All kwargs are forwarded to `self.lora_state_dict`. See
+        [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`] for more details on how the state dict is loaded.
+        See [`~loaders.StableDiffusionLoraLoaderMixin.load_lora_into_transformer`] for more details on how the state
+        dict is loaded into `self.transformer`.
+
+        Parameters:
+            pretrained_model_name_or_path_or_dict (`str` or `os.PathLike` or `dict`):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+            kwargs (`dict`, *optional*):
+                See [`~loaders.StableDiffusionLoraLoaderMixin.lora_state_dict`].
+        """
+        if not USE_PEFT_BACKEND:
+            raise ValueError("PEFT backend is required for this method.")
+
+        low_cpu_mem_usage = kwargs.pop("low_cpu_mem_usage", _LOW_CPU_MEM_USAGE_DEFAULT_LORA)
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # if a dict is passed, copy it instead of modifying it inplace
+        if isinstance(pretrained_model_name_or_path_or_dict, dict):
+            pretrained_model_name_or_path_or_dict = pretrained_model_name_or_path_or_dict.copy()
+
+        # First, ensure that the checkpoint is a compatible one and can be successfully loaded.
+        state_dict = self.lora_state_dict(pretrained_model_name_or_path_or_dict, **kwargs)
+
+        is_correct_format = all("lora" in key for key in state_dict.keys())
+        if not is_correct_format:
+            raise ValueError("Invalid LoRA checkpoint.")
+
+        self.load_lora_into_transformer(
+            state_dict,
+            transformer=getattr(self, self.transformer_name) if not hasattr(self, "transformer") else self.transformer,
+            adapter_name=adapter_name,
+            _pipeline=self,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.SD3LoraLoaderMixin.load_lora_into_transformer with SD3Transformer2DModel->SanaTransformer2DModel
+    def load_lora_into_transformer(
+        cls, state_dict, transformer, adapter_name=None, _pipeline=None, low_cpu_mem_usage=False
+    ):
+        """
+        This will load the LoRA layers specified in `state_dict` into `transformer`.
+
+        Parameters:
+            state_dict (`dict`):
+                A standard state dict containing the lora layer parameters. The keys can either be indexed directly
+                into the unet or prefixed with an additional `unet` which can be used to distinguish between text
+                encoder lora layers.
+            transformer (`SanaTransformer2DModel`):
+                The Transformer model to load the LoRA layers into.
+            adapter_name (`str`, *optional*):
+                Adapter name to be used for referencing the loaded adapter model. If not specified, it will use
+                `default_{i}` where i is the total number of adapters being loaded.
+            low_cpu_mem_usage (`bool`, *optional*):
+                Speed up model loading by only loading the pretrained LoRA weights and not initializing the random
+                weights.
+        """
+        if low_cpu_mem_usage and is_peft_version("<", "0.13.0"):
+            raise ValueError(
+                "`low_cpu_mem_usage=True` is not compatible with this `peft` version. Please update it with `pip install -U peft`."
+            )
+
+        # Load the layers corresponding to transformer.
+        logger.info(f"Loading {cls.transformer_name}.")
+        transformer.load_lora_adapter(
+            state_dict,
+            network_alphas=None,
+            adapter_name=adapter_name,
+            _pipeline=_pipeline,
+            low_cpu_mem_usage=low_cpu_mem_usage,
+        )
+
+    @classmethod
+    # Copied from diffusers.loaders.lora_pipeline.CogVideoXLoraLoaderMixin.save_lora_weights
+    def save_lora_weights(
+        cls,
+        save_directory: Union[str, os.PathLike],
+        transformer_lora_layers: Dict[str, Union[torch.nn.Module, torch.Tensor]] = None,
+        is_main_process: bool = True,
+        weight_name: str = None,
+        save_function: Callable = None,
+        safe_serialization: bool = True,
+    ):
+        r"""
+        Save the LoRA parameters corresponding to the UNet and text encoder.
+
+        Arguments:
+            save_directory (`str` or `os.PathLike`):
+                Directory to save LoRA parameters to. Will be created if it doesn't exist.
+            transformer_lora_layers (`Dict[str, torch.nn.Module]` or `Dict[str, torch.Tensor]`):
+                State dict of the LoRA layers corresponding to the `transformer`.
+            is_main_process (`bool`, *optional*, defaults to `True`):
+                Whether the process calling this is the main process or not. Useful during distributed training and you
+                need to call this function on all processes. In this case, set `is_main_process=True` only on the main
+                process to avoid race conditions.
+            save_function (`Callable`):
+                The function to use to save the state dictionary. Useful during distributed training when you need to
+                replace `torch.save` with another method. Can be configured with the environment variable
+                `DIFFUSERS_SAVE_MODE`.
+            safe_serialization (`bool`, *optional*, defaults to `True`):
+                Whether to save the model using `safetensors` or the traditional PyTorch way with `pickle`.
+        """
+        state_dict = {}
+
+        if not transformer_lora_layers:
+            raise ValueError("You must pass `transformer_lora_layers`.")
+
+        if transformer_lora_layers:
+            state_dict.update(cls.pack_weights(transformer_lora_layers, cls.transformer_name))
+
+        # Save the model
+        cls.write_lora_layers(
+            state_dict=state_dict,
+            save_directory=save_directory,
+            is_main_process=is_main_process,
+            weight_name=weight_name,
+            save_function=save_function,
+            safe_serialization=safe_serialization,
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.fuse_lora with unet->transformer
+    def fuse_lora(
+        self,
+        components: List[str] = ["transformer", "text_encoder"],
+        lora_scale: float = 1.0,
+        safe_fusing: bool = False,
+        adapter_names: Optional[List[str]] = None,
+        **kwargs,
+    ):
+        r"""
+        Fuses the LoRA parameters into the original parameters of the corresponding blocks.
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components: (`List[str]`): List of LoRA-injectable components to fuse the LoRAs into.
+            lora_scale (`float`, defaults to 1.0):
+                Controls how much to influence the outputs with the LoRA parameters.
+            safe_fusing (`bool`, defaults to `False`):
+                Whether to check fused weights for NaN values before fusing and if values are NaN not fusing them.
+            adapter_names (`List[str]`, *optional*):
+                Adapter names to be used for fusing. If nothing is passed, all active adapters will be fused.
+
+        Example:
+
+        ```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("nerijs/pixel-art-xl", weight_name="pixel-art-xl.safetensors", adapter_name="pixel")
+        pipeline.fuse_lora(lora_scale=0.7)
+        ```
+        """
+        super().fuse_lora(
+            components=components, lora_scale=lora_scale, safe_fusing=safe_fusing, adapter_names=adapter_names
+        )
+
+    # Copied from diffusers.loaders.lora_pipeline.StableDiffusionLoraLoaderMixin.unfuse_lora with unet->transformer
+    def unfuse_lora(self, components: List[str] = ["transformer", "text_encoder"], **kwargs):
+        r"""
+        Reverses the effect of
+        [`pipe.fuse_lora()`](https://huggingface.co/docs/diffusers/main/en/api/loaders#diffusers.loaders.LoraBaseMixin.fuse_lora).
+
+        <Tip warning={true}>
+
+        This is an experimental API.
+
+        </Tip>
+
+        Args:
+            components (`List[str]`): List of LoRA-injectable components to unfuse LoRA from.
+            unfuse_transformer (`bool`, defaults to `True`): Whether to unfuse the UNet LoRA parameters.
+            unfuse_text_encoder (`bool`, defaults to `True`):
+                Whether to unfuse the text encoder LoRA parameters. If the text encoder wasn't monkey-patched with the
+                LoRA parameters then it won't have any effect.
+        """
+        super().unfuse_lora(components=components)
+
+
 class LoraLoaderMixin(StableDiffusionLoraLoaderMixin):
     def __init__(self, *args, **kwargs):
         deprecation_message = "LoraLoaderMixin is deprecated and this will be removed in a future version. Please use `StableDiffusionLoraLoaderMixin`, instead."

src/diffusers/loaders/peft.py

@@ -54,6 +54,7 @@ _SET_ADAPTER_SCALE_FN_MAPPING = {
     "CogVideoXTransformer3DModel": lambda model_cls, weights: weights,
     "MochiTransformer3DModel": lambda model_cls, weights: weights,
     "LTXVideoTransformer3DModel": lambda model_cls, weights: weights,
+    "SanaTransformer2DModel": lambda model_cls, weights: weights,
 }
 
 

src/diffusers/models/transformers/sana_transformer.py

@@ -18,7 +18,8 @@ import torch
 from torch import nn
 
 from ...configuration_utils import ConfigMixin, register_to_config
-from ...utils import is_torch_version, logging
+from ...loaders import PeftAdapterMixin
+from ...utils import USE_PEFT_BACKEND, is_torch_version, logging, scale_lora_layers, unscale_lora_layers
 from ..attention_processor import (
     Attention,
     AttentionProcessor,
@@ -180,7 +181,7 @@ class SanaTransformerBlock(nn.Module):
         return hidden_states
 
 
-class SanaTransformer2DModel(ModelMixin, ConfigMixin):
+class SanaTransformer2DModel(ModelMixin, ConfigMixin, PeftAdapterMixin):
     r"""
     A 2D Transformer model introduced in [Sana](https://huggingface.co/papers/2410.10629) family of models.
 
@@ -363,8 +364,24 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin):
         timestep: torch.LongTensor,
         encoder_attention_mask: Optional[torch.Tensor] = None,
         attention_mask: Optional[torch.Tensor] = None,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
         return_dict: bool = True,
     ) -> Union[Tuple[torch.Tensor, ...], Transformer2DModelOutput]:
+        if attention_kwargs is not None:
+            attention_kwargs = attention_kwargs.copy()
+            lora_scale = attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+        else:
+            if attention_kwargs is not None and attention_kwargs.get("scale", None) is not None:
+                logger.warning(
+                    "Passing `scale` via `attention_kwargs` when not using the PEFT backend is ineffective."
+                )
+
         # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
         #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
         #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
@@ -460,6 +477,11 @@ class SanaTransformer2DModel(ModelMixin, ConfigMixin):
         hidden_states = hidden_states.permute(0, 5, 1, 3, 2, 4)
         output = hidden_states.reshape(batch_size, -1, post_patch_height * p, post_patch_width * p)
 
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
         if not return_dict:
             return (output,)
+
         return Transformer2DModelOutput(sample=output)

src/diffusers/pipelines/pag/pipeline_pag_sana.py

@@ -170,7 +170,6 @@ class SanaPAGPipeline(DiffusionPipeline, PAGMixin):
             pag_attn_processors=(PAGCFGSanaLinearAttnProcessor2_0(), PAGIdentitySanaLinearAttnProcessor2_0()),
         )
 
-    # Copied from diffusers.pipelines.sana.pipeline_sana.SanaPipeline.encode_prompt
     def encode_prompt(
         self,
         prompt: Union[str, List[str]],

src/diffusers/pipelines/sana/pipeline_sana.py

@@ -16,21 +16,25 @@ import html
 import inspect
 import re
 import urllib.parse as ul
-from typing import Callable, Dict, List, Optional, Tuple, Union
+from typing import Any, Callable, Dict, List, Optional, Tuple, Union
 
 import torch
 from transformers import AutoModelForCausalLM, AutoTokenizer
 
 from ...callbacks import MultiPipelineCallbacks, PipelineCallback
 from ...image_processor import PixArtImageProcessor
+from ...loaders import SanaLoraLoaderMixin
 from ...models import AutoencoderDC, SanaTransformer2DModel
 from ...schedulers import DPMSolverMultistepScheduler
 from ...utils import (
     BACKENDS_MAPPING,
+    USE_PEFT_BACKEND,
     is_bs4_available,
     is_ftfy_available,
     logging,
     replace_example_docstring,
+    scale_lora_layers,
+    unscale_lora_layers,
 )
 from ...utils.torch_utils import randn_tensor
 from ..pipeline_utils import DiffusionPipeline
@@ -130,7 +134,7 @@ def retrieve_timesteps(
     return timesteps, num_inference_steps
 
 
-class SanaPipeline(DiffusionPipeline):
+class SanaPipeline(DiffusionPipeline, SanaLoraLoaderMixin):
     r"""
     Pipeline for text-to-image generation using [Sana](https://huggingface.co/papers/2410.10629).
     """
@@ -177,6 +181,7 @@ class SanaPipeline(DiffusionPipeline):
         clean_caption: bool = False,
         max_sequence_length: int = 300,
         complex_human_instruction: Optional[List[str]] = None,
+        lora_scale: Optional[float] = None,
     ):
         r"""
         Encodes the prompt into text encoder hidden states.
@@ -210,6 +215,15 @@ class SanaPipeline(DiffusionPipeline):
         if device is None:
             device = self._execution_device
 
+        # set lora scale so that monkey patched LoRA
+        # function of text encoder can correctly access it
+        if lora_scale is not None and isinstance(self, SanaLoraLoaderMixin):
+            self._lora_scale = lora_scale
+
+            # dynamically adjust the LoRA scale
+            if self.text_encoder is not None and USE_PEFT_BACKEND:
+                scale_lora_layers(self.text_encoder, lora_scale)
+
         if prompt is not None and isinstance(prompt, str):
             batch_size = 1
         elif prompt is not None and isinstance(prompt, list):
@@ -305,6 +319,11 @@ class SanaPipeline(DiffusionPipeline):
             negative_prompt_embeds = None
             negative_prompt_attention_mask = None
 
+        if self.text_encoder is not None:
+            if isinstance(self, SanaLoraLoaderMixin) and USE_PEFT_BACKEND:
+                # Retrieve the original scale by scaling back the LoRA layers
+                unscale_lora_layers(self.text_encoder, lora_scale)
+
         return prompt_embeds, prompt_attention_mask, negative_prompt_embeds, negative_prompt_attention_mask
 
     # Copied from diffusers.pipelines.stable_diffusion.pipeline_stable_diffusion.StableDiffusionPipeline.prepare_extra_step_kwargs
@@ -554,6 +573,10 @@ class SanaPipeline(DiffusionPipeline):
     def guidance_scale(self):
         return self._guidance_scale
 
+    @property
+    def attention_kwargs(self):
+        return self._attention_kwargs
+
     @property
     def do_classifier_free_guidance(self):
         return self._guidance_scale > 1.0
@@ -590,6 +613,7 @@ class SanaPipeline(DiffusionPipeline):
         return_dict: bool = True,
         clean_caption: bool = True,
         use_resolution_binning: bool = True,
+        attention_kwargs: Optional[Dict[str, Any]] = None,
         callback_on_step_end: Optional[Callable[[int, int, Dict], None]] = None,
         callback_on_step_end_tensor_inputs: List[str] = ["latents"],
         max_sequence_length: int = 300,
@@ -662,6 +686,10 @@ class SanaPipeline(DiffusionPipeline):
                 [PIL](https://pillow.readthedocs.io/en/stable/): `PIL.Image.Image` or `np.array`.
             return_dict (`bool`, *optional*, defaults to `True`):
                 Whether or not to return a [`~pipelines.stable_diffusion.IFPipelineOutput`] instead of a plain tuple.
+            attention_kwargs:
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
             clean_caption (`bool`, *optional*, defaults to `True`):
                 Whether or not to clean the caption before creating embeddings. Requires `beautifulsoup4` and `ftfy` to
                 be installed. If the dependencies are not installed, the embeddings will be created from the raw
@@ -722,6 +750,7 @@ class SanaPipeline(DiffusionPipeline):
         )
 
         self._guidance_scale = guidance_scale
+        self._attention_kwargs = attention_kwargs
         self._interrupt = False
 
         # 2. Default height and width to transformer
@@ -733,6 +762,7 @@ class SanaPipeline(DiffusionPipeline):
             batch_size = prompt_embeds.shape[0]
 
         device = self._execution_device
+        lora_scale = self.attention_kwargs.get("scale", None) if self.attention_kwargs is not None else None
 
         # 3. Encode input prompt
         (
@@ -753,6 +783,7 @@ class SanaPipeline(DiffusionPipeline):
             clean_caption=clean_caption,
             max_sequence_length=max_sequence_length,
             complex_human_instruction=complex_human_instruction,
+            lora_scale=lora_scale,
         )
         if self.do_classifier_free_guidance:
             prompt_embeds = torch.cat([negative_prompt_embeds, prompt_embeds], dim=0)
@@ -801,6 +832,7 @@ class SanaPipeline(DiffusionPipeline):
                     encoder_attention_mask=prompt_attention_mask,
                     timestep=timestep,
                     return_dict=False,
+                    attention_kwargs=self.attention_kwargs,
                 )[0]
                 noise_pred = noise_pred.float()
 

tests/lora/test_lora_layers_sana.py

+# coding=utf-8
+# Copyright 2024 HuggingFace Inc.
+#
+# 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.
+import sys
+import unittest
+
+import torch
+from transformers import Gemma2ForCausalLM, GemmaTokenizer
+
+from diffusers import AutoencoderDC, FlowMatchEulerDiscreteScheduler, SanaPipeline, SanaTransformer2DModel
+from diffusers.utils.testing_utils import floats_tensor, require_peft_backend
+
+
+sys.path.append(".")
+
+from utils import PeftLoraLoaderMixinTests  # noqa: E402
+
+
+@require_peft_backend
+class SanaLoRATests(unittest.TestCase, PeftLoraLoaderMixinTests):
+    pipeline_class = SanaPipeline
+    scheduler_cls = FlowMatchEulerDiscreteScheduler(shift=7.0)
+    scheduler_kwargs = {}
+    scheduler_classes = [FlowMatchEulerDiscreteScheduler]
+    transformer_kwargs = {
+        "patch_size": 1,
+        "in_channels": 4,
+        "out_channels": 4,
+        "num_layers": 1,
+        "num_attention_heads": 2,
+        "attention_head_dim": 4,
+        "num_cross_attention_heads": 2,
+        "cross_attention_head_dim": 4,
+        "cross_attention_dim": 8,
+        "caption_channels": 8,
+        "sample_size": 32,
+    }
+    transformer_cls = SanaTransformer2DModel
+    vae_kwargs = {
+        "in_channels": 3,
+        "latent_channels": 4,
+        "attention_head_dim": 2,
+        "encoder_block_types": (
+            "ResBlock",
+            "EfficientViTBlock",
+        ),
+        "decoder_block_types": (
+            "ResBlock",
+            "EfficientViTBlock",
+        ),
+        "encoder_block_out_channels": (8, 8),
+        "decoder_block_out_channels": (8, 8),
+        "encoder_qkv_multiscales": ((), (5,)),
+        "decoder_qkv_multiscales": ((), (5,)),
+        "encoder_layers_per_block": (1, 1),
+        "decoder_layers_per_block": [1, 1],
+        "downsample_block_type": "conv",
+        "upsample_block_type": "interpolate",
+        "decoder_norm_types": "rms_norm",
+        "decoder_act_fns": "silu",
+        "scaling_factor": 0.41407,
+    }
+    vae_cls = AutoencoderDC
+    tokenizer_cls, tokenizer_id = GemmaTokenizer, "hf-internal-testing/dummy-gemma"
+    text_encoder_cls, text_encoder_id = Gemma2ForCausalLM, "hf-internal-testing/dummy-gemma-for-diffusers"
+
+    @property
+    def output_shape(self):
+        return (1, 32, 32, 3)
+
+    def get_dummy_inputs(self, with_generator=True):
+        batch_size = 1
+        sequence_length = 16
+        num_channels = 4
+        sizes = (32, 32)
+
+        generator = torch.manual_seed(0)
+        noise = floats_tensor((batch_size, num_channels) + sizes)
+        input_ids = torch.randint(1, sequence_length, size=(batch_size, sequence_length), generator=generator)
+
+        pipeline_inputs = {
+            "prompt": "",
+            "negative_prompt": "",
+            "num_inference_steps": 4,
+            "guidance_scale": 4.5,
+            "height": 32,
+            "width": 32,
+            "max_sequence_length": sequence_length,
+            "output_type": "np",
+            "complex_human_instruction": None,
+        }
+        if with_generator:
+            pipeline_inputs.update({"generator": generator})
+
+        return noise, input_ids, pipeline_inputs
+
+    @unittest.skip("Not supported in Sana.")
+    def test_modify_padding_mode(self):
+        pass
+
+    @unittest.skip("Not supported in Mochi.")
+    def test_simple_inference_with_text_denoiser_block_scale(self):
+        pass
+
+    @unittest.skip("Not supported in Mochi.")
+    def test_simple_inference_with_text_denoiser_block_scale_for_all_dict_options(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in Mochi.")
+    def test_simple_inference_with_partial_text_lora(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in Mochi.")
+    def test_simple_inference_with_text_lora(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in Mochi.")
+    def test_simple_inference_with_text_lora_and_scale(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in Mochi.")
+    def test_simple_inference_with_text_lora_fused(self):
+        pass
+
+    @unittest.skip("Text encoder LoRA is not supported in Mochi.")
+    def test_simple_inference_with_text_lora_save_load(self):
+        pass

tests/lora/utils.py

@@ -1545,7 +1545,12 @@ class PeftLoraLoaderMixinTests:
                         "adapter-1"
                     ].weight += float("inf")
                 else:
-                    pipe.transformer.transformer_blocks[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
+                    named_modules = [name for name, _ in pipe.transformer.named_modules()]
+                    has_attn1 = any("attn1" in name for name in named_modules)
+                    if has_attn1:
+                        pipe.transformer.transformer_blocks[0].attn1.to_q.lora_A["adapter-1"].weight += float("inf")
+                    else:
+                        pipe.transformer.transformer_blocks[0].attn.to_q.lora_A["adapter-1"].weight += float("inf")
 
             # with `safe_fusing=True` we should see an Error
             with self.assertRaises(ValueError):