train_dreambooth_ziplora_sdxl.py

#!/usr/bin/env python
# coding=utf-8
# Copyright 2023 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 gc
import itertools
import logging
import math
import os
import shutil
import warnings
from pathlib import Path

import numpy as np
import torch
import torch.nn.functional as F
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 packaging import version
from PIL import Image
from PIL.ImageOps import exif_transpose
from torch.utils.data import Dataset
from torchvision import transforms
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig

import diffusers
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    DPMSolverMultistepScheduler,
    StableDiffusionXLPipeline,
    UNet2DConditionModel,
)
from diffusers.loaders import LoraLoaderMixin
from diffusers.models.lora import LoRALinearLayer
from diffusers.optimization import get_scheduler
from diffusers.training_utils import compute_snr, unet_lora_state_dict
from diffusers.utils import check_min_version, is_wandb_available
from diffusers.utils.import_utils import is_xformers_available
from ziplora_pytorch import ZipLoRALinearLayer
from ziplora_pytorch.utils import (
    get_lora_weights,
    merge_lora_weights,
    initialize_ziplora_layer,
    unet_ziplora_state_dict,
    ziplora_set_forward_type,
    ziplora_compute_mergers_similarity,
    insert_ziplora_to_unet,
)


# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.24.0.dev0")

logger = get_logger(__name__)


# TODO: This function should be removed once training scripts are rewritten in PEFT
def text_encoder_lora_state_dict(text_encoder):
    state_dict = {}

    def text_encoder_attn_modules(text_encoder):
        from transformers import CLIPTextModel, CLIPTextModelWithProjection

        attn_modules = []

        if isinstance(text_encoder, (CLIPTextModel, CLIPTextModelWithProjection)):
            for i, layer in enumerate(text_encoder.text_model.encoder.layers):
                name = f"text_model.encoder.layers.{i}.self_attn"
                mod = layer.self_attn
                attn_modules.append((name, mod))

        return attn_modules

    for name, module in text_encoder_attn_modules(text_encoder):
        for k, v in module.q_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.q_proj.lora_linear_layer.{k}"] = v

        for k, v in module.k_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.k_proj.lora_linear_layer.{k}"] = v

        for k, v in module.v_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.v_proj.lora_linear_layer.{k}"] = v

        for k, v in module.out_proj.lora_linear_layer.state_dict().items():
            state_dict[f"{name}.out_proj.lora_linear_layer.{k}"] = v

    return state_dict


def save_model_card(
    repo_id: str,
    images=None,
    base_model=str,
    train_text_encoder=False,
    instance_prompt=str,
    instance_prompt2=str,
    validation_prompt=str,
    repo_folder=None,
    vae_path=None,
):
    img_str = "widget:\n" if images else ""
    for i, image in enumerate(images):
        image.save(os.path.join(repo_folder, f"image_{i}.png"))
        img_str += f"""
        - text: '{validation_prompt if validation_prompt else ' ' }'
          output:
            url:
                "image_{i}.png"
        """

    yaml = f"""
---
tags:
- stable-diffusion-xl
- stable-diffusion-xl-diffusers
- text-to-image
- diffusers
- lora
- template:sd-lora
{img_str}
base_model: {base_model}
instance_prompt: {instance_prompt}
license: openrail++
---
    """

    model_card = f"""
# SDXL ZipLoRA DreamBooth - {repo_id}

<Gallery />

## Model description

These are {repo_id} ZipLoRA adaption weights for {base_model}.

The weights were trained  using [DreamBooth](https://dreambooth.github.io/) and [ZipLoRA](https://ziplora.github.io/).

LoRA for the text encoder was enabled: {train_text_encoder}.

Special VAE used for training: {vae_path}.

## Trigger words

You should use a combination of  {instance_prompt} and {instance_prompt_2} to trigger the image generation 
using the two trained concepts.

## Download model

Weights for this model are available in Safetensors format.

[Download]({repo_id}/tree/main) them in the Files & versions tab.

"""
    with open(os.path.join(repo_folder, "README.md"), "w") as f:
        f.write(yaml + model_card)


def import_model_class_from_model_name_or_path(
    pretrained_model_name_or_path: str, revision: str, subfolder: str = "text_encoder"
):
    text_encoder_config = PretrainedConfig.from_pretrained(
        pretrained_model_name_or_path, subfolder=subfolder, revision=revision
    )
    model_class = text_encoder_config.architectures[0]

    if model_class == "CLIPTextModel":
        from transformers import CLIPTextModel

        return CLIPTextModel
    elif model_class == "CLIPTextModelWithProjection":
        from transformers import CLIPTextModelWithProjection

        return CLIPTextModelWithProjection
    else:
        raise ValueError(f"{model_class} is not supported.")


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(
        "--pretrained_vae_model_name_or_path",
        type=str,
        default=None,
        help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
    )
    parser.add_argument(
        "--lora_name_or_path",
        type=str,
        default=None,
        required=True,
        help="Path to 1st lora weights",
    )
    parser.add_argument(
        "--lora_name_or_path_2",
        type=str,
        default=None,
        required=True,
        help="Path to 2nd lora weights",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained model identifier from huggingface.co/models.",
    )
    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(
        "--instance_data_dir_2",
        type=str,
        default=None,
        help=("A folder containing the training data for 2nd lora"),
    )
    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(
        "--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(
        "--instance_prompt_2",
        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(
        "--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(
        "--output_dir",
        type=str,
        default="lora-dreambooth-model",
        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=1024,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--crops_coords_top_left_h",
        type=int,
        default=0,
        help=(
            "Coordinate for (the height) to be included in the crop coordinate embeddings needed by SDXL UNet."
        ),
    )
    parser.add_argument(
        "--crops_coords_top_left_w",
        type=int,
        default=0,
        help=(
            "Coordinate for (the height) to be included in the crop coordinate embeddings needed by SDXL UNet."
        ),
    )
    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(
        "--train_text_encoder",
        action="store_true",
        help="Whether to train the text encoder. If set, the text encoder should be float32 precision.",
    )
    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(
        "--init_merger_value",
        type=float,
        default=1.0,
        help="initial value of merger coefficient vectors",
    )
    parser.add_argument(
        "--init_merger_value_2",
        type=float,
        default=1.0,
        help="initial value of merger coefficient vectors",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--similarity_lambda",
        type=float,
        default=0.01,
        help="an appropriate multiplier for the cosine similarity loss term",
    )
    parser.add_argument(
        "--text_encoder_lr",
        type=float,
        default=5e-6,
        help="Text encoder learning rate 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(
        "--snr_gamma",
        type=float,
        default=None,
        help="SNR weighting gamma to be used if rebalancing the loss. Recommended value is 5.0. "
        "More details here: https://arxiv.org/abs/2303.09556.",
    )
    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(
        "--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 Prodidy 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(
        "--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(
        "--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(
        "--prior_generation_precision",
        type=str,
        default=None,
        choices=["no", "fp32", "fp16", "bf16"],
        help=(
            "Choose prior generation precision between fp32, fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to  fp16 if a GPU is available else fp32."
        ),
    )
    parser.add_argument(
        "--local_rank",
        type=int,
        default=-1,
        help="For distributed training: local_rank",
    )
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention",
        action="store_true",
        help="Whether or not to use xformers.",
    )
    parser.add_argument(
        "--quick_release",
        action="store_true",
        help="Releases VRAM immediately after processing each layer, conserving it."
    )

    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

    return args


def prepare_instance_images(instance_data_root: str, repeats: int):
    instance_data_root = Path(instance_data_root)
    if not 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())
    ]

    res = []
    for img in instance_images:
        res.extend(itertools.repeat(img, repeats))
    return res


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,
        instance_data_root_2,
        instance_prompt_2,
        size=1024,
        repeats=1,
        center_crop=False,
    ):
        self.size = size
        self.center_crop = center_crop

        self.instance_prompt = instance_prompt
        self.instance_prompt_2 = instance_prompt_2

        # 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:
            raise NotImplementedError
        self.instance_images = prepare_instance_images(instance_data_root, repeats)
        self.instance_images_2 = prepare_instance_images(instance_data_root_2, repeats)
        self.num_instance_images = max(
            len(self.instance_images), len(self.instance_images_2)
        )
        if len(self.instance_images) != self.num_instance_images:
            # repeat
            self.instance_images = self.instance_images * math.ceil(
                self.num_instance_images / len(self.instance_images)
            )
            self.instance_images = self.instance_images[: self.num_instance_images]

        if len(self.instance_images_2) != self.num_instance_images:
            # repeat
            self.instance_images_2 = self.instance_images_2 * math.ceil(
                self.num_instance_images / len(self.instance_images_2)
            )
            self.instance_images_2 = self.instance_images_2[: self.num_instance_images]
        self._length = self.num_instance_images
        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 _transform_image(self, image):
        image = exif_transpose(image)
        if not image.mode == "RGB":
            image = image.convert("RGB")
        return self.image_transforms(image)

    def __getitem__(self, index):
        example = {}
        instance_image = self.instance_images[index % self.num_instance_images]
        example["instance_images"] = self._transform_image(instance_image)
        example["instance_prompt"] = self.instance_prompt
        instance_image_2 = self.instance_images_2[index % self.num_instance_images]
        example["instance_images_2"] = self._transform_image(instance_image_2)
        example["instance_prompt_2"] = self.instance_prompt_2
        return example


def collate_fn(examples):
    pixel_values = [example["instance_images"] for example in examples]
    prompts = [example["instance_prompt"] for example in examples]
    pixel_values_2 = [example["instance_images_2"] for example in examples]
    prompts_2 = [example["instance_prompt_2"] for example in examples]

    pixel_values = torch.stack(pixel_values)
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
    pixel_values_2 = torch.stack(pixel_values_2)
    pixel_values_2 = pixel_values_2.to(memory_format=torch.contiguous_format).float()

    batch = {
        "pixel_values": pixel_values,
        "prompts": prompts,
        "pixel_values_2": pixel_values_2,
        "prompts_2": prompts_2,
    }
    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 tokenize_prompt(tokenizer, prompt):
    text_inputs = tokenizer(
        prompt,
        padding="max_length",
        max_length=tokenizer.model_max_length,
        truncation=True,
        return_tensors="pt",
    )
    text_input_ids = text_inputs.input_ids
    return text_input_ids


# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(text_encoders, tokenizers, prompt, text_input_ids_list=None):
    prompt_embeds_list = []

    for i, text_encoder in enumerate(text_encoders):
        if tokenizers is not None:
            tokenizer = tokenizers[i]
            text_input_ids = tokenize_prompt(tokenizer, prompt)
        else:
            assert text_input_ids_list is not None
            text_input_ids = text_input_ids_list[i]

        prompt_embeds = text_encoder(
            text_input_ids.to(text_encoder.device),
            output_hidden_states=True,
        )

        # We are only ALWAYS interested in the pooled output of the final text encoder
        pooled_prompt_embeds = prompt_embeds[0]
        prompt_embeds = prompt_embeds.hidden_states[-2]
        bs_embed, seq_len, _ = prompt_embeds.shape
        prompt_embeds = prompt_embeds.view(bs_embed, seq_len, -1)
        prompt_embeds_list.append(prompt_embeds)

    prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
    pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
    return prompt_embeds, pooled_prompt_embeds


def main(args):
    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],
    )

    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."
            )
        import wandb

    # 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)

    # 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,
                token=args.hub_token,
            ).repo_id

    # Load the tokenizers
    tokenizer_one = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
        use_fast=False,
    )
    tokenizer_two = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer_2",
        revision=args.revision,
        use_fast=False,
    )

    # import correct text encoder classes
    text_encoder_cls_one = import_model_class_from_model_name_or_path(
        args.pretrained_model_name_or_path, args.revision
    )
    text_encoder_cls_two = import_model_class_from_model_name_or_path(
        args.pretrained_model_name_or_path, args.revision, subfolder="text_encoder_2"
    )

    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler"
    )
    text_encoder_one = text_encoder_cls_one.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="text_encoder",
        revision=args.revision,
    )
    text_encoder_two = text_encoder_cls_two.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="text_encoder_2",
        revision=args.revision,
    )
    vae_path = (
        args.pretrained_model_name_or_path
        if args.pretrained_vae_model_name_or_path is None
        else args.pretrained_vae_model_name_or_path
    )
    vae = AutoencoderKL.from_pretrained(
        vae_path,
        subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
        revision=args.revision,
    )
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision
    )

    # We only train the additional adapter LoRA layers
    vae.requires_grad_(False)
    text_encoder_one.requires_grad_(False)
    text_encoder_two.requires_grad_(False)
    unet.requires_grad_(False)

    # For mixed precision training we cast all non-trainable weights (vae, non-lora text_encoder and non-lora unet) 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

    # Move unet, vae and text_encoder to device and cast to weight_dtype
    unet.to(accelerator.device, dtype=weight_dtype)

    # The VAE is always in float32 to avoid NaN losses.
    vae.to(accelerator.device, dtype=torch.float32)

    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
    text_encoder_two.to(accelerator.device, dtype=weight_dtype)

    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

            xformers_version = version.parse(xformers.__version__)
            if xformers_version == version.parse("0.0.16"):
                logger.warn(
                    "xFormers 0.0.16 cannot be used for training in some GPUs. If you observe problems during training, "
                    "please update xFormers to at least 0.0.17. See https://huggingface.co/docs/diffusers/main/en/optimization/xformers for more details."
                )
            unet.enable_xformers_memory_efficient_attention()
        else:
            raise ValueError(
                "xformers is not available. Make sure it is installed correctly"
            )

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
        if args.train_text_encoder:
            text_encoder_one.gradient_checkpointing_enable()
            text_encoder_two.gradient_checkpointing_enable()

    # Load 2 pre-trained LoRAs
    lora_weights = get_lora_weights(args.lora_name_or_path)
    lora_weights_2 = get_lora_weights(args.lora_name_or_path_2)

    # now we will add new LoRA weights to the attention layers
    # Set correct lora layers
    unet_lora_parameters = []
    for attn_processor_name, attn_processor in unet.attn_processors.items():
        # Parse the attention module.
        attn_module = unet
        for n in attn_processor_name.split(".")[:-1]:
            attn_module = getattr(attn_module, n)
        # Get prepared for ziplora
        attn_name = ".".join(attn_processor_name.split(".")[:-1])
        merged_lora_weights_dict = merge_lora_weights(lora_weights, attn_name)
        merged_lora_weights_dict_2 = merge_lora_weights(lora_weights_2, attn_name)
        kwargs = {
            "state_dict": merged_lora_weights_dict,
            "state_dict_2": merged_lora_weights_dict_2,
        }
        # Set the `lora_layer` attribute of the attention-related matrices.
        attn_module.to_q.set_lora_layer(
            initialize_ziplora_layer(
                part="to_q",
                in_features=attn_module.to_q.in_features,
                out_features=attn_module.to_q.out_features,
                init_merger_value=args.init_merger_value,
                init_merger_value_2=args.init_merger_value_2,
                **kwargs,
            )
        )
        attn_module.to_k.set_lora_layer(
            initialize_ziplora_layer(
                part="to_k",
                in_features=attn_module.to_k.in_features,
                out_features=attn_module.to_k.out_features,
                init_merger_value=args.init_merger_value,
                init_merger_value_2=args.init_merger_value_2,
                **kwargs,
            )
        )
        attn_module.to_v.set_lora_layer(
            initialize_ziplora_layer(
                part="to_v",
                in_features=attn_module.to_v.in_features,
                out_features=attn_module.to_v.out_features,
                init_merger_value=args.init_merger_value,
                init_merger_value_2=args.init_merger_value_2,
                **kwargs,
            )
        )
        attn_module.to_out[0].set_lora_layer(
            initialize_ziplora_layer(
                part="to_out.0",
                in_features=attn_module.to_out[0].in_features,
                out_features=attn_module.to_out[0].out_features,
                init_merger_value=args.init_merger_value,
                init_merger_value_2=args.init_merger_value_2,
                **kwargs,
            )
        )

        # Accumulate the LoRA params to optimize.
        unet_lora_parameters.extend(
            [p for p in attn_module.to_q.lora_layer.parameters() if p.requires_grad]
        )
        unet_lora_parameters.extend(
            [p for p in attn_module.to_k.lora_layer.parameters() if p.requires_grad]
        )
        unet_lora_parameters.extend(
            [p for p in attn_module.to_v.lora_layer.parameters() if p.requires_grad]
        )
        unet_lora_parameters.extend(
            [
                p
                for p in attn_module.to_out[0].lora_layer.parameters()
                if p.requires_grad
            ]
        )

    # The text encoder comes from 🤗 transformers, so we cannot directly modify it.
    # So, instead, we monkey-patch the forward calls of its attention-blocks.
    if args.train_text_encoder:
        raise NotImplementedError
        # ensure that dtype is float32, even if rest of the model that isn't trained is loaded in fp16
        text_lora_parameters_one = LoraLoaderMixin._modify_text_encoder(
            text_encoder_one, dtype=torch.float32, rank=args.rank
        )
        text_lora_parameters_two = LoraLoaderMixin._modify_text_encoder(
            text_encoder_two, dtype=torch.float32, rank=args.rank
        )

    # 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:
            # there are only two options here. Either are just the unet attn processor layers
            # or there are the unet and text encoder atten layers
            unet_lora_layers_to_save = None
            text_encoder_one_lora_layers_to_save = None
            text_encoder_two_lora_layers_to_save = None

            for model in models:
                if isinstance(model, type(accelerator.unwrap_model(unet))):
                    unet_lora_layers_to_save = unet_ziplora_state_dict(model)
                elif isinstance(
                    model, type(accelerator.unwrap_model(text_encoder_one))
                ):
                    text_encoder_one_lora_layers_to_save = text_encoder_lora_state_dict(
                        model
                    )
                elif isinstance(
                    model, type(accelerator.unwrap_model(text_encoder_two))
                ):
                    text_encoder_two_lora_layers_to_save = text_encoder_lora_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()

            StableDiffusionXLPipeline.save_lora_weights(
                output_dir,
                unet_lora_layers=unet_lora_layers_to_save,
                text_encoder_lora_layers=text_encoder_one_lora_layers_to_save,
                text_encoder_2_lora_layers=text_encoder_two_lora_layers_to_save,
            )

    accelerator.register_save_state_pre_hook(save_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:
        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
        )

    # Optimization parameters

    unet_lora_parameters_with_lr = {
        "params": unet_lora_parameters,
        "lr": args.learning_rate,
    }
    total_params = sum(p.numel() for p in unet_lora_parameters)
    logger.info(f"Number of Trainable Parameters: {total_params * 1.e-6:.2f} M")

    if args.train_text_encoder:
        raise NotImplementedError
        # different learning rate for text encoder and unet
        text_lora_parameters_one_with_lr = {
            "params": text_lora_parameters_one,
            "weight_decay": args.adam_weight_decay_text_encoder,
            "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
        }
        text_lora_parameters_two_with_lr = {
            "params": text_lora_parameters_two,
            "weight_decay": args.adam_weight_decay_text_encoder,
            "lr": args.text_encoder_lr if args.text_encoder_lr else args.learning_rate,
        }
        params_to_optimize = [
            unet_lora_parameters_with_lr,
            text_lora_parameters_one_with_lr,
            text_lora_parameters_two_with_lr,
        ]
    else:
        params_to_optimize = [unet_lora_parameters_with_lr]

    # Optimizer creation
    if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
        logger.warn(
            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.warn(
            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.warn(
                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
            )
        if args.train_text_encoder and args.text_encoder_lr:
            logger.warn(
                f"Learning rates were provided both for the unet and the text encoder- e.g. text_encoder_lr:"
                f" {args.text_encoder_lr} and learning_rate: {args.learning_rate}. "
                f"When using prodigy only learning_rate is used as the initial learning rate."
            )
            # changes the learning rate of text_encoder_parameters_one and text_encoder_parameters_two to be
            # --learning_rate
            params_to_optimize[1]["lr"] = args.learning_rate
            params_to_optimize[2]["lr"] = args.learning_rate

        optimizer = optimizer_class(
            params_to_optimize,
            lr=args.learning_rate,
            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,
        instance_data_root_2=args.instance_data_dir_2,
        instance_prompt_2=args.instance_prompt_2,
        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=collate_fn,
        num_workers=args.dataloader_num_workers,
    )

    # Computes additional embeddings/ids required by the SDXL UNet.
    # regular text embeddings (when `train_text_encoder` is not True)
    # pooled text embeddings
    # time ids

    def compute_time_ids():
        # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
        original_size = (args.resolution, args.resolution)
        target_size = (args.resolution, args.resolution)
        crops_coords_top_left = (
            args.crops_coords_top_left_h,
            args.crops_coords_top_left_w,
        )
        add_time_ids = list(original_size + crops_coords_top_left + target_size)
        add_time_ids = torch.tensor([add_time_ids])
        add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
        return add_time_ids

    if not args.train_text_encoder:
        tokenizers = [tokenizer_one, tokenizer_two]
        text_encoders = [text_encoder_one, text_encoder_two]

        def compute_text_embeddings(prompt, text_encoders, tokenizers):
            with torch.no_grad():
                prompt_embeds, pooled_prompt_embeds = encode_prompt(
                    text_encoders, tokenizers, prompt
                )
                prompt_embeds = prompt_embeds.to(accelerator.device)
                pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
            return prompt_embeds, pooled_prompt_embeds

    # Handle instance prompt.
    instance_time_ids = compute_time_ids()

    # 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 args.train_text_encoder:
        (
            instance_prompt_hidden_states,
            instance_pooled_prompt_embeds,
        ) = compute_text_embeddings(args.instance_prompt, text_encoders, tokenizers)
        (
            instance_prompt_hidden_states_2,
            instance_pooled_prompt_embeds_2,
        ) = compute_text_embeddings(args.instance_prompt_2, text_encoders, tokenizers)

    # Clear the memory here
    if not args.train_text_encoder:
        del tokenizers, text_encoders
        gc.collect()
        torch.cuda.empty_cache()

    # 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.
    add_time_ids = instance_time_ids

    if not args.train_text_encoder:
        prompt_embeds = instance_prompt_hidden_states
        unet_add_text_embeds = instance_pooled_prompt_embeds
        prompt_embeds_2 = instance_prompt_hidden_states_2
        unet_add_text_embeds_2 = instance_pooled_prompt_embeds_2

    # if we're optimizing the text encoder (both if instance prompt is used for all images or custom prompts) we need to tokenize and encode the
    # batch prompts on all training steps
    else:
        tokens_one = tokenize_prompt(tokenizer_one, args.instance_prompt)
        tokens_two = tokenize_prompt(tokenizer_two, args.instance_prompt)

    # 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`.
    if args.train_text_encoder:
        (
            unet,
            text_encoder_one,
            text_encoder_two,
            optimizer,
            train_dataloader,
            lr_scheduler,
        ) = accelerator.prepare(
            unet,
            text_encoder_one,
            text_encoder_two,
            optimizer,
            train_dataloader,
            lr_scheduler,
        )
    else:
        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, 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:
        accelerator.init_trackers("dreambooth-ziplora-sd-xl", 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:
        raise NotImplementedError
        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,
    )

    for epoch in range(first_epoch, args.num_train_epochs):
        unet.train()
        if args.train_text_encoder:
            text_encoder_one.train()
            text_encoder_two.train()

            # set top parameter requires_grad = True for gradient checkpointing works
            text_encoder_one.text_model.embeddings.requires_grad_(True)
            text_encoder_two.text_model.embeddings.requires_grad_(True)

        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                model_inputs = []
                for i in range(2):
                    pixel_values_key = "pixel_values" if i == 0 else "pixel_values_2"
                    prompts_key = "prompts" if i == 0 else "prompts_2"
                    pixel_values = batch[pixel_values_key].to(dtype=vae.dtype)
                    prompts = batch[prompts_key]
                    # Convert images to latent space
                    model_input = vae.encode(pixel_values).latent_dist.sample()
                    model_input = model_input * vae.config.scaling_factor
                    if args.pretrained_vae_model_name_or_path is None:
                        model_input = model_input.to(weight_dtype)
                    model_inputs.append(model_input)

                # 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
                timesteps = torch.randint(
                    0,
                    noise_scheduler.config.num_train_timesteps,
                    (bsz,),
                    device=model_input.device,
                )
                timesteps = timesteps.long()
                # Predict the noise residual
                unet_added_conditions = {
                    "time_ids": add_time_ids.repeat(bsz, 1),
                    "text_embeds": unet_add_text_embeds.repeat(bsz, 1),
                }
                prompt_embeds_input = prompt_embeds.repeat(bsz, 1, 1)

                unet_added_conditions_2 = {
                    "time_ids": unet_added_conditions["time_ids"],
                    "text_embeds": unet_add_text_embeds_2.repeat(bsz, 1),
                }
                prompt_embeds_input_2 = prompt_embeds_2.repeat(bsz, 1, 1)

                # Add noise to the model input according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                for i in range(2):
                    model_inputs[i] = noise_scheduler.add_noise(
                        model_inputs[i], noise, timesteps
                    )

                # 1. merged weights + concept
                model_pred_mc = unet(
                    model_inputs[0],
                    timesteps,
                    prompt_embeds_input,
                    added_cond_kwargs=unet_added_conditions,
                ).sample
                # 2. merged weights + style
                model_pred_ms = unet(
                    model_inputs[1],
                    timesteps,
                    prompt_embeds_input_2,
                    added_cond_kwargs=unet_added_conditions_2,
                ).sample

                # 3. concept weights + concept
                ziplora_set_forward_type(unet, type="weight_1")
                with torch.no_grad():
                    model_pred_cc = unet(
                        model_inputs[0],
                        timesteps,
                        prompt_embeds_input,
                        added_cond_kwargs=unet_added_conditions,
                    ).sample
                # 4. style weights + style
                ziplora_set_forward_type(unet, type="weight_2")
                with torch.no_grad():
                    model_pred_ss = unet(
                        model_inputs[1],
                        timesteps,
                        prompt_embeds_input_2,
                        added_cond_kwargs=unet_added_conditions_2,
                    ).sample
                # compute losses
                ziplora_set_forward_type(unet, type="merge")
                loss_1 = F.mse_loss(
                    model_pred_mc.float(), model_pred_cc.float(), reduction="mean"
                )
                loss_2 = F.mse_loss(
                    model_pred_ms.float(), model_pred_ss.float(), reduction="mean"
                )
                loss_3 = args.similarity_lambda * ziplora_compute_mergers_similarity(
                    unet
                )
                loss = loss_1 + loss_2 + loss_3

                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    accelerator.clip_grad_norm_(
                        unet_lora_parameters, 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(),
                "loss_1": loss_1.detach().item(),
                "loss_2": loss_2.detach().item(),
                "loss_3": loss_3.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 or loss_3.detach().item() == 0.0:
                break

        if accelerator.is_main_process:
            if (
                args.validation_prompt is not None
                and epoch % args.validation_epochs == 0
            ):
                logger.info(
                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
                    f" {args.validation_prompt}."
                )
                # create pipeline
                if not args.train_text_encoder:
                    text_encoder_one = text_encoder_cls_one.from_pretrained(
                        args.pretrained_model_name_or_path,
                        subfolder="text_encoder",
                        revision=args.revision,
                    )
                    text_encoder_two = text_encoder_cls_two.from_pretrained(
                        args.pretrained_model_name_or_path,
                        subfolder="text_encoder_2",
                        revision=args.revision,
                    )
                pipeline = StableDiffusionXLPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
                    text_encoder=accelerator.unwrap_model(text_encoder_one),
                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
                    unet=accelerator.unwrap_model(unet),
                    revision=args.revision,
                    torch_dtype=weight_dtype,
                )

                # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
                scheduler_args = {}

                if "variance_type" in pipeline.scheduler.config:
                    variance_type = pipeline.scheduler.config.variance_type

                    if variance_type in ["learned", "learned_range"]:
                        variance_type = "fixed_small"

                    scheduler_args["variance_type"] = variance_type

                pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
                    pipeline.scheduler.config, **scheduler_args
                )

                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
                )
                pipeline_args = {"prompt": args.validation_prompt}

                images = [
                    pipeline(**pipeline_args, generator=generator).images[0]
                    for _ in range(args.num_validation_images)
                ]

                for tracker in accelerator.trackers:
                    if tracker.name == "tensorboard":
                        np_images = np.stack([np.asarray(img) for img in images])
                        tracker.writer.add_images(
                            "validation", np_images, epoch, dataformats="NHWC"
                        )
                    if tracker.name == "wandb":
                        tracker.log(
                            {
                                "validation": [
                                    wandb.Image(
                                        image, caption=f"{i}: {args.validation_prompt}"
                                    )
                                    for i, image in enumerate(images)
                                ]
                            }
                        )

                del pipeline
                torch.cuda.empty_cache()

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(unet)
        unet = unet.to(torch.float32)
        unet_lora_layers = unet_ziplora_state_dict(unet, args.quick_release)

        if args.train_text_encoder:
            text_encoder_one = accelerator.unwrap_model(text_encoder_one)
            text_encoder_lora_layers = text_encoder_lora_state_dict(
                text_encoder_one.to(torch.float32)
            )
            text_encoder_two = accelerator.unwrap_model(text_encoder_two)
            text_encoder_2_lora_layers = text_encoder_lora_state_dict(
                text_encoder_two.to(torch.float32)
            )
        else:
            text_encoder_lora_layers = None
            text_encoder_2_lora_layers = None

        StableDiffusionXLPipeline.save_lora_weights(
            save_directory=args.output_dir,
            unet_lora_layers=unet_lora_layers,
            text_encoder_lora_layers=text_encoder_lora_layers,
            text_encoder_2_lora_layers=text_encoder_2_lora_layers,
        )

        # Final inference
        # Load previous pipeline
        vae = AutoencoderKL.from_pretrained(
            vae_path,
            subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
            revision=args.revision,
            torch_dtype=weight_dtype,
        )
        pipeline = StableDiffusionXLPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            vae=vae,
            revision=args.revision,
            torch_dtype=weight_dtype,
        )

        # We train on the simplified learning objective. If we were previously predicting a variance, we need the scheduler to ignore it
        scheduler_args = {}

        if "variance_type" in pipeline.scheduler.config:
            variance_type = pipeline.scheduler.config.variance_type

            if variance_type in ["learned", "learned_range"]:
                variance_type = "fixed_small"

            scheduler_args["variance_type"] = variance_type

        pipeline.scheduler = DPMSolverMultistepScheduler.from_config(
            pipeline.scheduler.config, **scheduler_args
        )

        # load attention processors
        # pipeline.load_lora_weights(args.output_dir)
        pipeline.unet = insert_ziplora_to_unet(pipeline.unet, args.output_dir)
        # run inference
        images = []
        if args.validation_prompt and args.num_validation_images > 0:
            pipeline = pipeline.to(accelerator.device, dtype=weight_dtype)
            generator = (
                torch.Generator(device=accelerator.device).manual_seed(args.seed)
                if args.seed
                else None
            )
            images = [
                pipeline(
                    args.validation_prompt, num_inference_steps=25, generator=generator
                ).images[0]
                for _ in range(args.num_validation_images)
            ]

            for tracker in accelerator.trackers:
                if tracker.name == "tensorboard":
                    np_images = np.stack([np.asarray(img) for img in images])
                    tracker.writer.add_images(
                        "test", np_images, epoch, dataformats="NHWC"
                    )
                if tracker.name == "wandb":
                    tracker.log(
                        {
                            "test": [
                                wandb.Image(
                                    image, caption=f"{i}: {args.validation_prompt}"
                                )
                                for i, image in enumerate(images)
                            ]
                        }
                    )

        if args.push_to_hub:
            save_model_card(
                repo_id,
                images=images,
                base_model=args.pretrained_model_name_or_path,
                train_text_encoder=args.train_text_encoder,
                instance_prompt=args.instance_prompt,
                validation_prompt=args.validation_prompt,
                repo_folder=args.output_dir,
                vae_path=args.pretrained_vae_model_name_or_path,
            )
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )

    accelerator.end_training()


if __name__ == "__main__":
    args = parse_args()
    main(args)