(Not so rigrously tested) example showing how to use `bitsandbytes`, `peft`, etc. to LoRA fine-tune Flux.1 Dev.

inference.md

When loading the LoRA params (that were obtained on a quantized base model) and merging them into the base model, it is recommended to first dequantize the base model, merge the LoRA params into it, and then quantize the model again. This is because merging into 4bit quantized models can lead to some rounding errors. Below, we provide an end-to-end example:

1. First, load the original model and merge the LoRA params into it:

from diffusers import FluxPipeline 
import torch 

ckpt_id = "black-forest-labs/FLUX.1-dev"
pipeline = FluxPipeline.from_pretrained(
    ckpt_id, text_encoder=None, text_encoder_2=None, torch_dtype=torch.float16
)
pipeline.load_lora_weights("yarn_art_lora_flux_nf4", weight_name="pytorch_lora_weights.safetensors")
pipeline.fuse_lora()
pipeline.unload_lora_weights()

pipeline.transformer.save_pretrained("fused_transformer")

2. Quantize the model and run inference

from diffusers import AutoPipelineForText2Image, FluxTransformer2DModel, BitsAndBytesConfig
import torch

ckpt_id = "black-forest-labs/FLUX.1-dev"
bnb_4bit_compute_dtype = torch.float16
nf4_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
)
transformer = FluxTransformer2DModel.from_pretrained(
    "fused_transformer",
    quantization_config=nf4_config,
    torch_dtype=bnb_4bit_compute_dtype,
)
pipeline = AutoPipelineForText2Image.from_pretrained(
    ckpt_id, transformer=transformer, torch_dtype=bnb_4bit_compute_dtype
)
pipeline.enable_model_cpu_offload()

image = pipeline(
    "a puppy in a pond, yarn art style", num_inference_steps=28, guidance_scale=3.5, height=768
).images[0]
image.save("yarn_merged.png")

Dequantize, merge, quantize	Merging directly into quantized model

As we can notice the first column result follows the style more closely.

instructions.md

Make sure to install the following libraries and ensure an NVIDIA GPU environment (with at least 24GB VRAM):

pip install -U bitsandbytes peft
pip install git+https://github.com/huggingface/diffusers
pip install git+https://github.com/huggingface/transformers

I didn't test the underlying scripts on a 16GB card but they might work. I used a 4090 card for testing.

The training script train_dreambooth_lora_flux_nf4.py is hacked up version of the original diffusers training script: https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/train_dreambooth_lora_flux.py.

Note

Training takes at least 1.5 hours on the settings I tried out.

notes.md

In case the following code OOMs:

nf4_config = BitsAndBytesConfig(
    load_in_4bit=True,
    bnb_4bit_quant_type="nf4",
    bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
)
transformer = FluxTransformer2DModel.from_pretrained(
    ckpt_id, 
    subfolder="transformer", 
    quantization_config=nf4_config,
    torch_dtype=bnb_4bit_compute_dtype,
)

Consider directly loading a quantized checkpoint such as: https://huggingface.co/sayakpaul/flux.1-dev-nf4-pkg. You can specify --quantized_model_path to either supply a local path or a valid repo id on the Hub.

Since we're fine-tuning a quantized checkpoint, it might need additional hyperparameter-tuning. The results shown above are not optimal but it at least captures the subject well.

Also, note that the training script currently DOES NOT support resuming from an intermediate checkpoint.

train_dreambooth_lora_flux_nf4.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 contextlib import nullcontext
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 peft import prepare_model_for_kbit_training
from diffusers import BitsAndBytesConfig
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 CLIPTokenizer, PretrainedConfig, T5TokenizerFast

import diffusers
from diffusers import (
    AutoencoderKL,
    FlowMatchEulerDiscreteScheduler,
    FluxPipeline,
    FluxTransformer2DModel,
)
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, 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("0.31.0.dev0")

logger = get_logger(__name__)


def save_model_card(
    repo_id: str,
    images=None,
    base_model: str = None,
    train_text_encoder=False,
    instance_prompt=None,
    validation_prompt=None,
    repo_folder=None,
    quantization_config=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"""
# Flux 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 [Flux diffusers trainer](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_flux.md).

Was LoRA for the text encoder enabled? {train_text_encoder}.

Trained with quantization: Yes.

Quantization config:

```json
{quantization_config}
```

## 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
from diffusers import AutoPipelineForText2Image
import torch
pipeline = AutoPipelineForText2Image.from_pretrained("black-forest-labs/FLUX.1-dev", torch_dtype=torch.bfloat16).to('cuda')
pipeline.load_lora_weights('{repo_id}', weight_name='pytorch_lora_weights.safetensors')
image = pipeline('{validation_prompt if validation_prompt else instance_prompt}').images[0]
```

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

Please adhere to the licensing terms as described [here](https://huggingface.co/black-forest-labs/FLUX.1-dev/blob/main/LICENSE.md).
"""
    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",
        "flux",
        "flux-diffusers",
        "template:sd-lora",
    ]

    model_card = populate_model_card(model_card, tags=tags)
    model_card.save(os.path.join(repo_folder, "README.md"))


def load_text_encoders(class_one, class_two):
    text_encoder_one = class_one.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant
    )
    text_encoder_two = class_two.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder_2", revision=args.revision, variant=args.variant
    )
    return text_encoder_one, text_encoder_two


def log_validation(
    pipeline,
    args,
    accelerator,
    pipeline_args,
    epoch,
    torch_dtype,
    is_final_validation=False,
):
    logger.info(
        f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
        f" {args.validation_prompt}."
    )
    pipeline = pipeline.to(accelerator.device, dtype=torch_dtype)
    pipeline.set_progress_bar_config(disable=True)

    # run inference
    generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
    # autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()
    autocast_ctx = nullcontext()

    with autocast_ctx:
        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 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 == "T5EncoderModel":
        from transformers import T5EncoderModel

        return T5EncoderModel
    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(
        "--quantized_model_path",
        type=str,
        default=None,
        help="Path to pretrained model or that is pre-quantized.",
    )
    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=512,
        help="Maximum sequence length to use with with the T5 text encoder",
    )
    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="flux-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(
        "--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(
        "--guidance_scale",
        type=float,
        default=3.5,
        help="the FLUX.1 dev variant is a guidance distilled model",
    )

    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_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(
        "--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")

    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 tokenize_prompt(tokenizer, prompt, max_sequence_length):
    text_inputs = tokenizer(
        prompt,
        padding="max_length",
        max_length=max_sequence_length,
        truncation=True,
        return_length=False,
        return_overflowing_tokens=False,
        return_tensors="pt",
    )
    text_input_ids = text_inputs.input_ids
    return text_input_ids


def _encode_prompt_with_t5(
    text_encoder,
    tokenizer,
    max_sequence_length=512,
    prompt=None,
    num_images_per_prompt=1,
    device=None,
    text_input_ids=None,
):
    prompt = [prompt] if isinstance(prompt, str) else prompt
    batch_size = len(prompt)

    if tokenizer is not None:
        text_inputs = tokenizer(
            prompt,
            padding="max_length",
            max_length=max_sequence_length,
            truncation=True,
            return_length=False,
            return_overflowing_tokens=False,
            return_tensors="pt",
        )
        text_input_ids = text_inputs.input_ids
    else:
        if text_input_ids is None:
            raise ValueError("text_input_ids must be provided when the tokenizer is not specified")

    prompt_embeds = text_encoder(text_input_ids.to(device))[0]

    dtype = text_encoder.dtype
    prompt_embeds = prompt_embeds.to(dtype=dtype, device=device)

    _, seq_len, _ = prompt_embeds.shape

    # duplicate text embeddings and attention mask for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
    prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, seq_len, -1)

    return prompt_embeds


def _encode_prompt_with_clip(
    text_encoder,
    tokenizer,
    prompt: str,
    device=None,
    text_input_ids=None,
    num_images_per_prompt: int = 1,
):
    prompt = [prompt] if isinstance(prompt, str) else prompt
    batch_size = len(prompt)

    if tokenizer is not None:
        text_inputs = tokenizer(
            prompt,
            padding="max_length",
            max_length=77,
            truncation=True,
            return_overflowing_tokens=False,
            return_length=False,
            return_tensors="pt",
        )

        text_input_ids = text_inputs.input_ids
    else:
        if text_input_ids is None:
            raise ValueError("text_input_ids must be provided when the tokenizer is not specified")

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

    # Use pooled output of CLIPTextModel
    prompt_embeds = prompt_embeds.pooler_output
    prompt_embeds = prompt_embeds.to(dtype=text_encoder.dtype, device=device)

    # duplicate text embeddings for each generation per prompt, using mps friendly method
    prompt_embeds = prompt_embeds.repeat(1, num_images_per_prompt, 1)
    prompt_embeds = prompt_embeds.view(batch_size * num_images_per_prompt, -1)

    return prompt_embeds


def encode_prompt(
    text_encoders,
    tokenizers,
    prompt: str,
    max_sequence_length,
    device=None,
    num_images_per_prompt: int = 1,
    text_input_ids_list=None,
):
    prompt = [prompt] if isinstance(prompt, str) else prompt
    dtype = text_encoders[0].dtype

    pooled_prompt_embeds = _encode_prompt_with_clip(
        text_encoder=text_encoders[0],
        tokenizer=tokenizers[0],
        prompt=prompt,
        device=device if device is not None else text_encoders[0].device,
        num_images_per_prompt=num_images_per_prompt,
        text_input_ids=text_input_ids_list[0] if text_input_ids_list else None,
    )

    prompt_embeds = _encode_prompt_with_t5(
        text_encoder=text_encoders[1],
        tokenizer=tokenizers[1],
        max_sequence_length=max_sequence_length,
        prompt=prompt,
        num_images_per_prompt=num_images_per_prompt,
        device=device if device is not None else text_encoders[1].device,
        text_input_ids=text_input_ids_list[1] if text_input_ids_list else None,
    )

    text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=device, dtype=dtype)

    return prompt_embeds, pooled_prompt_embeds, text_ids


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:
            has_supported_fp16_accelerator = torch.cuda.is_available() or torch.backends.mps.is_available()
            torch_dtype = torch.float16 if has_supported_fp16_accelerator else torch.float32
            if args.prior_generation_precision == "fp32":
                torch_dtype = torch.float32
            elif args.prior_generation_precision == "fp16":
                torch_dtype = torch.float16
            elif args.prior_generation_precision == "bf16":
                torch_dtype = torch.bfloat16
            pipeline = FluxPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                torch_dtype=torch_dtype,
                revision=args.revision,
                variant=args.variant,
            )
            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 tokenizers
    tokenizer_one = CLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
    )
    tokenizer_two = T5TokenizerFast.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer_2",
        revision=args.revision,
    )

    # 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 = FlowMatchEulerDiscreteScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler"
    )
    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
    text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="vae",
        revision=args.revision,
        variant=args.variant,
    )
    if args.quantized_model_path is not None:
        transformer = FluxTransformer2DModel.from_pretrained(
            args.quantized_model_path, subfolder="transformer", revision=args.revision, variant=args.variant
        )
    else:
        bnb_4bit_compute_dtype =  torch.float32 
        if args.mixed_precision == "fp16":
            bnb_4bit_compute_dtype = torch.float16
        elif args.mixed_precision == "bf16":
            bnb_4bit_compute_dtype = torch.bfloat16
        nf4_config = BitsAndBytesConfig(
            load_in_4bit=True,
            bnb_4bit_quant_type="nf4",
            bnb_4bit_compute_dtype=bnb_4bit_compute_dtype,
        )
        transformer = FluxTransformer2DModel.from_pretrained(
            args.pretrained_model_name_or_path, 
            subfolder="transformer", 
            revision=args.revision, 
            variant=args.variant,
            quantization_config=nf4_config,
            torch_dtype=bnb_4bit_compute_dtype,
        )
    transformer = prepare_model_for_kbit_training(transformer, use_gradient_checkpointing=False)

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

    # 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."
        )

    # CPU for now.
    vae.to("cpu", dtype=weight_dtype)
    transformer.to("cpu") 
    text_encoder_one.to("cpu", dtype=weight_dtype)
    text_encoder_two.to("cpu", dtype=weight_dtype)

    if args.gradient_checkpointing:
        transformer.enable_gradient_checkpointing()

    # now we will add new LoRA weights to the attention layers
    # since we're quantizing, only linear layers should be targeted, no conv layers here.
    transformer_lora_config = LoraConfig(
        r=args.rank,
        lora_alpha=args.rank,
        init_lora_weights="gaussian",
        target_modules=["to_k", "to_q", "to_v", "to_out.0"],
    )
    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

    # 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,
            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,
        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,
    )

    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, text_ids = encode_prompt(
                text_encoders, tokenizers, prompt, args.max_sequence_length
            )
            prompt_embeds = prompt_embeds.to(accelerator.device)
            pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
            text_ids = text_ids.to(accelerator.device)
        return prompt_embeds, pooled_prompt_embeds, text_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.
    text_encoders = [t.to(device=accelerator.device) for t in text_encoders]
    instance_prompt_hidden_states, instance_pooled_prompt_embeds, instance_text_ids = compute_text_embeddings(
        args.instance_prompt, text_encoders, tokenizers
    )

    # Handle class prompt for prior-preservation.
    if args.with_prior_preservation:
        class_prompt_hidden_states, class_pooled_prompt_embeds, class_text_ids = compute_text_embeddings(
            args.class_prompt, text_encoders, tokenizers
        )

    # Clear the memory here
    if not train_dataset.custom_instance_prompts:
        del text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two
        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
        pooled_prompt_embeds = instance_pooled_prompt_embeds
        text_ids = instance_text_ids
        if args.with_prior_preservation:
            prompt_embeds = torch.cat([prompt_embeds, class_prompt_hidden_states], dim=0)
            pooled_prompt_embeds = torch.cat([pooled_prompt_embeds, class_pooled_prompt_embeds], dim=0)
            text_ids = torch.cat([text_ids, class_text_ids], dim=0)

    # move to CUDA
    vae.to(accelerator.device)

    vae_config_shift_factor = vae.config.shift_factor
    vae_config_scaling_factor = vae.config.scaling_factor
    vae_config_block_out_channels = vae.config.block_out_channels
    if args.cache_latents:
        latents_cache = []
        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=weight_dtype
                )
                latents_cache.append(vae.encode(batch["pixel_values"]).latent_dist)

        if args.validation_prompt is None:
            del vae
            free_memory()

    # Finally, move to CUDA.
    transformer.to(accelerator.device) 

    # 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-flux-dev-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
    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, pooled_prompt_embeds, text_ids = compute_text_embeddings(
                        prompts, text_encoders, tokenizers
                    )

                # Convert images to latent space
                if args.cache_latents:
                    model_input = latents_cache[step].sample()
                else:
                    pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                    model_input = vae.encode(pixel_values).latent_dist.sample()
                model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                model_input = model_input.to(dtype=weight_dtype)

                vae_scale_factor = 2 ** (len(vae_config_block_out_channels))

                latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                    model_input.shape[0],
                    model_input.shape[2],
                    model_input.shape[3],
                    accelerator.device,
                    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

                packed_noisy_model_input = FluxPipeline._pack_latents(
                    noisy_model_input,
                    batch_size=model_input.shape[0],
                    num_channels_latents=model_input.shape[1],
                    height=model_input.shape[2],
                    width=model_input.shape[3],
                )

                # handle guidance
                if transformer.config.guidance_embeds:
                    guidance = torch.tensor([args.guidance_scale], device=accelerator.device)
                    guidance = guidance.expand(model_input.shape[0])
                else:
                    guidance = None

                # Predict the noise residual
                model_pred = transformer(
                    hidden_states=packed_noisy_model_input,
                    # YiYi notes: divide it by 1000 for now because we scale it by 1000 in the transforme rmodel (we should not keep it but I want to keep the inputs same for the model for testing)
                    timestep=timesteps / 1000,
                    guidance=guidance,
                    pooled_projections=pooled_prompt_embeds,
                    encoder_hidden_states=prompt_embeds,
                    txt_ids=text_ids,
                    img_ids=latent_image_ids,
                    return_dict=False,
                )[0]
                model_pred = FluxPipeline._unpack_latents(
                    model_pred,
                    height=int(model_input.shape[2] * vae_scale_factor / 2),
                    width=int(model_input.shape[3] * vae_scale_factor / 2),
                    vae_scale_factor=vae_scale_factor,
                )

                # 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

            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
                text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
                pipeline = FluxPipeline.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),
                    transformer=accelerator.unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
                )
                pipeline_args = {"prompt": args.validation_prompt}
                images = log_validation(
                    pipeline=pipeline,
                    args=args,
                    accelerator=accelerator,
                    pipeline_args=pipeline_args,
                    epoch=epoch,
                    torch_dtype=weight_dtype,
                )
                del text_encoder_one, text_encoder_two
                free_memory()

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        transformer = unwrap_model(transformer)
        transformer_lora_layers = get_peft_model_state_dict(transformer)
        text_encoder_lora_layers = None

        FluxPipeline.save_lora_weights(
            save_directory=args.output_dir,
            transformer_lora_layers=transformer_lora_layers,
            text_encoder_lora_layers=text_encoder_lora_layers,
        )

        # Final inference
        # Load previous pipeline
        pipeline = FluxPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            revision=args.revision,
            variant=args.variant,
            torch_dtype=weight_dtype,
        )
        # 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}
            images = log_validation(
                pipeline=pipeline,
                args=args,
                accelerator=accelerator,
                pipeline_args=pipeline_args,
                epoch=epoch,
                is_final_validation=True,
                torch_dtype=weight_dtype,
            )

        if args.push_to_hub:
            save_model_card(
                repo_id,
                images=images,
                base_model=args.pretrained_model_name_or_path,
                train_text_encoder=False,
                instance_prompt=args.instance_prompt,
                validation_prompt=args.validation_prompt,
                repo_folder=args.output_dir,
                quantization_config=transformer.config["quantization_config"]
            )
            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)

training_command.md

Accelerate config:

compute_environment: LOCAL_MACHINE
debug: false
distributed_type: NO
downcast_bf16: 'no'
enable_cpu_affinity: true
gpu_ids: all
machine_rank: 0
main_training_function: main
mixed_precision: bf16
num_machines: 1
num_processes: 1
rdzv_backend: static
same_network: true
tpu_env: []
tpu_use_cluster: false
tpu_use_sudo: false
use_cpu: false

Training command:

CUDA_VISIBLE_DEVICES=0 accelerate launch --config_file=accelerate.yaml train_dreambooth_lora_flux_nf4.py \
  --pretrained_model_name_or_path="black-forest-labs/FLUX.1-dev" \
  --dataset_name="Norod78/Yarn-art-style" \
  --instance_prompt="a puppy, yarn art style" \
  --output_dir="yarn_art_lora_flux_nf4" \
  --mixed_precision="bf16" \
  --use_8bit_adam \
  --weighting_scheme="none" \
  --resolution=1024 \
  --train_batch_size=1 \
  --repeats=1 \
  --learning_rate=1e-4 \
  --guidance_scale=1 \
  --report_to="wandb" \
  --gradient_accumulation_steps=4 \
  --gradient_checkpointing \
  --lr_scheduler="constant" \
  --lr_warmup_steps=0 \
  --guidance_scale=1 \
  --cache_latents \
  --rank=4 \
  --max_train_steps=700 \
  --seed="0" \
  --push_to_hub

Hi I am trying to load flux and a LoRA adapter onto my 16gb vram gpu. Both came unquantized.
I am looking for a way to quantize both separately and load them into the pipeline together.
I didn't find bitsandbytes documentation for this particular scenario. Do you happen to know how to do that?
Thanks in advance!

@ariG23498's https://github.com/ariG23498/quantized-diffusion-inference might very useful here.

Hi @sayakpaul Thanks a lot for the comprehensive scripts, But when I run it (exactly as it is) I run into the following issue

out = (x.float() * cos + x_rotated.float() * sin).to(x.dtype)
           ~~~~~~~~~~^~~~~
RuntimeError: The size of tensor a (4608) must match the size of tensor b (16896) at non-singleton dimension 2

Seems there is a shape mismatch, so I printed the shapes of the embeddings that were being sent to the transformer. Namely

print(packed_noisy_model_input.shape)
print(pooled_prompt_embeds.shape)
print(prompt_embeds.shape)
print(text_ids.shape)
print(latent_image_ids.shape)

And the output I got was

torch.Size([1, 4096, 64])
torch.Size([1, 768])
torch.Size([1, 512, 4096])
torch.Size([512, 3])
torch.Size([16384, 3])

I am unable to figure out where the matrix multiplication is going wrong. I would much appreciate it if you could guide me a little on this matter.

Refer to this guide:
https://github.com/huggingface/diffusers/tree/main/examples/research_projects/flux_lora_quantization

Thank you!!!