#!/usr/bin/env python
# coding=utf-8
# Copyright 2025 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

# /// script
# dependencies = [
#     "diffusers @ git+https://github.com/huggingface/diffusers.git",
#     "torch>=2.0.0",
#     "accelerate>=0.31.0",
#     "transformers>=4.41.2",
#     "ftfy",
#     "tensorboard",
#     "Jinja2",
#     "peft>=0.11.1",
#     "sentencepiece",
#     "torchvision",
#     "datasets",
#     "bitsandbytes",
#     "prodigyopt",
# ]
# ///

import argparse
import copy
import itertools
import json
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 transformers
from accelerate import Accelerator, DistributedType
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, prepare_model_for_kbit_training, set_peft_model_state_dict
from peft.utils import get_peft_model_state_dict
from PIL import Image
from PIL.ImageOps import exif_transpose
from torch.utils.data import BatchSampler, Dataset
from torchvision import transforms
from torchvision.transforms import functional as TF
from tqdm.auto import tqdm
from transformers import AutoTokenizer, Qwen3VLModel

import diffusers
from diffusers import (
    AutoencoderKLQwenImage,
    BitsAndBytesConfig,
    FlowMatchEulerDiscreteScheduler,
    Krea2Pipeline,
    Krea2Transformer2DModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import (
    _collate_lora_metadata,
    cast_training_params,
    compute_density_for_timestep_sampling,
    compute_loss_weighting_for_sd3,
    find_nearest_bucket,
    free_memory,
    generate_aspect_ratio_buckets,
    offload_models,
    parse_buckets_string,
)
from diffusers.utils import (
    check_min_version,
    convert_unet_state_dict_to_peft,
    is_wandb_available,
)
from diffusers.utils.hub_utils import load_or_create_model_card, populate_model_card
from diffusers.utils.import_utils import is_torch_npu_available
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("0.40.0.dev0")

logger = get_logger(__name__)

if is_torch_npu_available():
    torch.npu.config.allow_internal_format = False


def save_model_card(
    repo_id: str,
    images=None,
    base_model: str = None,
    instance_prompt=None,
    validation_prompt=None,
    repo_folder=None,
    inference_model: str = "krea/Krea-2-Turbo",
):
    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"}}
            )

    # Only put `base_model` in the card metadata when it's a Hub id — a local training path is not a
    # valid model id and the Hub rejects it. RAW is the (non-distilled) training base.
    def _is_hub_id(s):
        return bool(s) and "/" in s and not os.path.exists(s)

    # A local training path is not a valid Hub model id (the Hub rejects it in card metadata). Krea 2
    # LoRAs are trained on RAW, so fall back to the canonical RAW id when given a local path.
    card_base_model = base_model if _is_hub_id(base_model) else "krea/Krea-2-Raw"
    base_display = card_base_model
    # The inference snippet always targets the distilled Turbo model; fall back to the canonical id
    # if a local path (or nothing) was passed.
    if not _is_hub_id(inference_model):
        inference_model = "krea/Krea-2-Turbo"

    # List Turbo first in the card's `base_model` metadata: the Hub keys the Inference Providers
    # widget off the first entry, and these LoRAs are served on Turbo — so leading with it enables
    # the widget on the pushed LoRA. RAW (the training base) stays linked as the second entry.
    card_base_models = [inference_model] + ([card_base_model] if card_base_model != inference_model else [])

    model_description = f"""
# Krea 2 DreamBooth LoRA - {repo_id}

<Gallery />

## Model description

These are {repo_id} DreamBooth LoRA weights, trained on {base_display}.

The weights were trained using [DreamBooth](https://dreambooth.github.io/) with the [Krea 2 diffusers trainer](https://github.com/huggingface/diffusers/blob/main/examples/dreambooth/README_krea2.md).

Krea 2 ships as two checkpoints: **RAW** (the non-distilled base you fine-tune on) and **Turbo** (an 8-step distilled checkpoint for fast, high-quality inference). Train your LoRA on RAW and run it on Turbo — LoRAs trained on RAW express strongly on Turbo.

## 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
>>> import torch
>>> from diffusers import Krea2Pipeline

>>> # Load the LoRA onto Krea 2 Turbo (the distilled inference model)
>>> pipe = Krea2Pipeline.from_pretrained("{inference_model}", torch_dtype=torch.bfloat16).to("cuda")
>>> pipe.load_lora_weights("{repo_id}")

>>> # Turbo recipe: 8 steps, no classifier-free guidance
>>> image = pipe("{instance_prompt}", num_inference_steps=8, guidance_scale=0.0).images[0]
>>> image.save("output.png")
```

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)
"""
    model_card = load_or_create_model_card(
        repo_id_or_path=repo_id,
        from_training=True,
        license="apache-2.0",
        base_model=card_base_models,
        prompt=instance_prompt,
        model_description=model_description,
        widget=widget_dict,
    )
    tags = [
        "text-to-image",
        "diffusers-training",
        "diffusers",
        "lora",
        "krea2",
        "krea2-diffusers",
        "template:sd-lora",
    ]

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


def log_validation(
    pipeline,
    args,
    accelerator,
    pipeline_args,
    epoch,
    torch_dtype,
    is_final_validation=False,
    pipeline_call_kwargs=None,
):
    args.num_validation_images = args.num_validation_images if args.num_validation_images else 1
    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 is not None else None
    autocast_ctx = torch.autocast(accelerator.device.type) if not is_final_validation else nullcontext()

    images = []
    for _ in range(args.num_validation_images):
        with autocast_ctx:
            image = pipeline(
                prompt_embeds=pipeline_args["prompt_embeds"],
                prompt_embeds_mask=pipeline_args["prompt_embeds_mask"],
                negative_prompt_embeds=pipeline_args["negative_prompt_embeds"],
                negative_prompt_embeds_mask=pipeline_args["negative_prompt_embeds_mask"],
                generator=generator,
                **(pipeline_call_kwargs or {}),
            ).images[0]
            images.append(image)

    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
    free_memory()

    return images


def _validation_call_kwargs(args):
    # When validating on a dedicated inference checkpoint (e.g. Krea 2 Turbo), use its recipe
    # (few-step, no CFG). When validating on the training checkpoint, use pipeline defaults.
    if args.validation_model_path is None:
        return {}
    return {
        "num_inference_steps": args.validation_num_inference_steps,
        "guidance_scale": args.validation_guidance_scale,
    }


def build_validation_pipeline(args, accelerator, transformer, weight_dtype):
    # Krea 2 RAW is a non-distilled base not meant for inference. If --validation_model_path is set
    # (e.g. Krea 2 Turbo), build the pipeline from THAT checkpoint and transplant the adapter trained
    # on RAW onto it (LoRAs trained on RAW express strongly on Turbo). Otherwise reuse the in-training
    # transformer. Either way the text encoder is skipped — validation reuses precomputed embeddings.
    if args.validation_model_path is not None:
        tmp_lora = os.path.join(args.output_dir, "_val_lora")
        Krea2Pipeline.save_lora_weights(
            tmp_lora,
            transformer_lora_layers=get_peft_model_state_dict(accelerator.unwrap_model(transformer)),
        )
        pipeline = Krea2Pipeline.from_pretrained(
            args.validation_model_path,
            tokenizer=None,
            text_encoder=None,
            revision=args.revision,
            variant=args.variant,
            torch_dtype=weight_dtype,
        )
        pipeline.load_lora_weights(tmp_lora)
        return pipeline
    return Krea2Pipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        tokenizer=None,
        text_encoder=None,
        transformer=accelerator.unwrap_model(transformer),
        revision=args.revision,
        variant=args.variant,
        torch_dtype=weight_dtype,
    )


def module_filter_fn(mod: torch.nn.Module, fqn: str):
    # Keep precision-sensitive modules in higher precision: the final output projection and the
    # patterns Krea2Transformer2DModel flags in `_skip_layerwise_casting_patterns` (time embedding,
    # norms), plus the timestep modulation projection.
    skip_patterns = ("final_layer.linear", "time_embed", "time_mod_proj", "norm")
    if any(pattern in fqn for pattern in skip_patterns):
        return False
    # don't convert linear modules with weight dimensions not divisible by 16
    if isinstance(mod, torch.nn.Linear):
        if mod.in_features % 16 != 0 or mod.out_features % 16 != 0:
            return False
    return True


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(
        "--bnb_quantization_config_path",
        type=str,
        default=None,
        help="Quantization config in a JSON file that will be used to define the bitsandbytes quant config of the DiT.",
    )
    parser.add_argument(
        "--do_fp8_training",
        action="store_true",
        help="if we are doing FP8 training (torchao float8 scaled-mm on a bf16-loaded transformer).",
    )
    parser.add_argument(
        "--validation_model_path",
        type=str,
        default=None,
        help=(
            "Path to the checkpoint validation and final inference run on. Krea 2 RAW is a non-distilled"
            " base not meant for inference, so validation should run on the distilled Krea 2 Turbo"
            " checkpoint: pass its path here and the adapter trained on RAW is transplanted onto Turbo for"
            " every validation. If unset, validation falls back to the (RAW) training checkpoint."
        ),
    )
    parser.add_argument(
        "--validation_num_inference_steps",
        type=int,
        default=8,
        help="num_inference_steps for validation on --validation_model_path (Krea 2 Turbo is an 8-step model).",
    )
    parser.add_argument(
        "--validation_guidance_scale",
        type=float,
        default=0.0,
        help="guidance_scale for validation on --validation_model_path (Krea 2 Turbo runs without CFG).",
    )
    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 the Qwen3-VL 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(
        "--skip_final_inference",
        default=False,
        action="store_true",
        help="Whether to skip the final inference step with loaded lora weights upon training completion. This will run intermediate validation inference if `validation_prompt` is provided. Specify to reduce memory.",
    )

    parser.add_argument(
        "--final_validation_prompt",
        type=str,
        default=None,
        help="A prompt that is used during a final validation to verify that the model is learning. Ignored if `--validation_prompt` is provided.",
    )
    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=32,
        help=(
            "The dimension of the LoRA update matrices. The Krea 2 authors recommend rank 32 for most styles; "
            "increase it (and focus on the attention layers) for long runs or high-frequency styles."
        ),
    )
    parser.add_argument(
        "--lora_alpha",
        type=int,
        default=32,
        help="LoRA alpha to be used for additional scaling. The Krea 2 authors recommend alpha == rank (scale 1.0).",
    )
    parser.add_argument("--lora_dropout", type=float, default=0.0, help="Dropout probability for LoRA layers")

    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="krea2-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(
        "--aspect_ratio_buckets",
        type=str,
        default=None,
        help=(
            "Aspect ratio buckets to use for training. Define as a string of 'h1,w1;h2,w2;...'. "
            "e.g. '1024,1024;768,1360;1360,768;880,1168;1168,880;1248,832;832,1248'. "
            "Requires --use_aspect_ratio_buckets. Images are resized to cover and cropped to the nearest "
            "listed bucket (smaller images are upscaled). When set, --resolution is ignored."
        ),
    )
    parser.add_argument(
        "--use_aspect_ratio_buckets",
        action="store_true",
        help=(
            "Enable aspect-ratio bucketing. Without --aspect_ratio_buckets, the buckets are computed on the "
            "fly from --resolution and capped to each image's own resolution, so smaller images are assigned "
            "to a smaller bucket instead of being upscaled. Provide --aspect_ratio_buckets to use an explicit list."
        ),
    )
    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(
        "--caption_dropout",
        type=float,
        default=0.0,
        help=(
            "Probability of replacing an instance image's caption with an empty string during training, so that"
            " fraction of samples is trained unconditionally. Improves classifier-free guidance. A common value is"
            " 0.1. Class/prior-preservation captions are never dropped."
        ),
    )
    parser.add_argument(
        "--train_batch_size", type=int, default=4, help="Batch size (per device) for the training dataloader."
    )
    parser.add_argument(
        "--sample_batch_size", type=int, default=4, help="Batch size (per device) for sampling images."
    )
    parser.add_argument("--num_train_epochs", type=int, default=1)
    parser.add_argument(
        "--max_train_steps",
        type=int,
        default=None,
        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
    )
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
    parser.add_argument(
        "--checkpoints_total_limit",
        type=int,
        default=None,
        help=("Max number of checkpoints to store."),
    )
    parser.add_argument(
        "--resume_from_checkpoint",
        type=str,
        default=None,
        help=(
            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
        ),
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--gradient_checkpointing",
        action="store_true",
        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=3e-4,
        help=(
            "Initial learning rate (after the potential warmup period) to use. The Krea 2 authors recommend "
            "3e-4 - 7e-4 with a constant schedule (lower end for a constant schedule; higher is fine with cosine)."
        ),
    )
    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(
        "--lora_layers",
        type=str,
        default=None,
        help=(
            "The transformer modules to apply LoRA training on, comma separated (matched as module-name suffixes). "
            'E.g. "to_q,to_k,to_v,to_out.0,to_gate" trains the attention layers only (the authors\' suggestion for '
            "long runs). If omitted, the Krea 2 authors' recommended default layer set is used."
        ),
    )

    parser.add_argument(
        "--adam_epsilon",
        type=float,
        default=1e-08,
        help="Epsilon value for the Adam optimizer and Prodigy optimizers.",
    )

    parser.add_argument(
        "--prodigy_use_bias_correction",
        type=bool,
        default=True,
        help="Turn on Adam's bias correction. True by default. Ignored if optimizer is adamW",
    )
    parser.add_argument(
        "--prodigy_safeguard_warmup",
        type=bool,
        default=True,
        help="Remove lr from the denominator of D estimate to avoid issues during warm-up stage. True by default. "
        "Ignored if optimizer is adamW",
    )
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument("--push_to_hub", action="store_true", help="Whether or not to push the model to the Hub.")
    parser.add_argument("--hub_token", type=str, default=None, help="The token to use to push to the Model Hub.")
    parser.add_argument(
        "--hub_model_id",
        type=str,
        default=None,
        help="The name of the repository to keep in sync with the local `output_dir`.",
    )
    parser.add_argument(
        "--logging_dir",
        type=str,
        default="logs",
        help=(
            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
        ),
    )
    parser.add_argument(
        "--allow_tf32",
        action="store_true",
        help=(
            "Whether or not to allow TF32 on Ampere GPUs. Can be used to speed up training. For more information, see"
            " https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices"
        ),
    )
    parser.add_argument(
        "--cache_latents",
        action="store_true",
        default=False,
        help="Cache the VAE latents",
    )
    parser.add_argument(
        "--report_to",
        type=str,
        default="tensorboard",
        help=(
            'The integration to report the results and logs to. Supported platforms are `"tensorboard"`'
            ' (default), `"wandb"` and `"comet_ml"`. Use `"all"` to report to all integrations.'
        ),
    )
    parser.add_argument(
        "--mixed_precision",
        type=str,
        default=None,
        choices=["no", "fp16", "bf16"],
        help=(
            "Whether to use mixed precision. Choose between fp16 and bf16 (bfloat16). Bf16 requires PyTorch >="
            " 1.10.and an Nvidia Ampere GPU.  Default to the value of accelerate config of the current system or the"
            " flag passed with the `accelerate.launch` command. Use this argument to override the accelerate config."
        ),
    )
    parser.add_argument(
        "--upcast_before_saving",
        action="store_true",
        default=False,
        help=(
            "Whether to upcast the trained transformer layers to float32 before saving (at the end of training). "
            "Defaults to precision dtype used for training to save memory"
        ),
    )
    parser.add_argument(
        "--offload",
        action="store_true",
        help="Whether to offload the VAE and the text encoder to CPU when they are not used.",
    )
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")

    if input_args is not None:
        args = parser.parse_args(input_args)
    else:
        args = parser.parse_args()

    if args.dataset_name is None and args.instance_data_dir is None:
        raise ValueError("Specify either `--dataset_name` or `--instance_data_dir`")

    if args.dataset_name is not None and args.instance_data_dir is not None:
        raise ValueError("Specify only one of `--dataset_name` or `--instance_data_dir`")

    if args.do_fp8_training and args.bnb_quantization_config_path:
        raise ValueError("Both `do_fp8_training` and `bnb_quantization_config_path` cannot be passed.")

    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,
        buckets=None,
        use_aspect_ratio_buckets=False,
        bucket_divisibility=16,
        bucket_base_resolutions=None,
    ):
        self.size = size
        self.resolution = size
        self.center_crop = center_crop

        self.instance_prompt = instance_prompt
        self.custom_instance_prompts = None
        self.class_prompt = class_prompt

        # Explicit user-provided bucket list (or None). The concrete list of buckets actually used is
        # built from the data in `self.buckets` during preprocessing below.
        self._explicit_buckets = buckets
        self.use_aspect_ratio_buckets = use_aspect_ratio_buckets
        self.bucket_divisibility = bucket_divisibility
        self.bucket_base_resolutions = bucket_base_resolutions

        # 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 = []
        self.buckets = []
        bucket_to_idx = {}
        for image in self.instance_images:
            image = exif_transpose(image)
            if not image.mode == "RGB":
                image = image.convert("RGB")

            width, height = image.size

            # Assign the image to a bucket.
            target = self._bucket_for_image(height, width)
            if target not in bucket_to_idx:
                bucket_to_idx[target] = len(self.buckets)
                self.buckets.append(target)
            bucket_idx = bucket_to_idx[target]

            # based on the bucket assignment, define the transformations
            image = self.train_transform(
                image,
                size=target,
                center_crop=args.center_crop,
                random_flip=args.random_flip,
            )
            self.pixel_values.append((image, bucket_idx))

        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

    def __len__(self):
        return self._length

    def __getitem__(self, index):
        example = {}
        instance_image, bucket_idx = self.pixel_values[index % self.num_instance_images]
        example["index"] = index
        example["instance_images"] = instance_image
        example["bucket_idx"] = bucket_idx
        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")
            # Match the class image to the paired instance image's bucket so they can be stacked into one batch.
            example["class_images"] = self.train_transform(
                class_image, size=self.buckets[bucket_idx], center_crop=self.center_crop
            )
            example["class_prompt"] = self.class_prompt

        return example

    def _bucket_for_image(self, height, width):
        # An explicit bucket list takes priority: pick the nearest, upscaling smaller images to cover it.
        if self._explicit_buckets is not None:
            return self._explicit_buckets[find_nearest_bucket(height, width, self._explicit_buckets)]
        # On-the-fly bucketing: cap the ladder to the image's own resolution so smaller images are
        # assigned to a smaller bucket rather than being upscaled (mirrors ostris' bucketing).
        if self.use_aspect_ratio_buckets:
            resolution = min(self.resolution, round((height * width) ** 0.5))
            ladder = generate_aspect_ratio_buckets(
                resolution,
                divisibility=self.bucket_divisibility,
                base_resolutions=self.bucket_base_resolutions,
            )
            return ladder[find_nearest_bucket(height, width, ladder)]
        # No bucketing: a single square bucket reproduces the fixed-size resize + crop.
        return (self.resolution, self.resolution)

    def train_transform(self, image, size, center_crop=False, random_flip=False):
        # Resize preserving aspect ratio so the image covers the bucket, then crop to the bucket size.
        target_height, target_width = size
        width, height = image.size
        scale = max(target_height / height, target_width / width)
        new_height, new_width = round(height * scale), round(width * scale)
        image = TF.resize(image, [new_height, new_width], interpolation=transforms.InterpolationMode.BILINEAR)
        if center_crop:
            image = TF.center_crop(image, size)
        else:
            i, j, h, w = transforms.RandomCrop.get_params(image, output_size=size)
            image = TF.crop(image, i, j, h, w)
        if random_flip and random.random() < 0.5:
            image = TF.hflip(image)
        return TF.normalize(TF.to_tensor(image), [0.5], [0.5])


def collate_fn(examples, with_prior_preservation=False):
    indices = [example["index"] for example in examples]
    pixel_values = [example["instance_images"] for example in examples]
    # Keep instance_prompts unchanged for prompt cache precompute; prompts may be extended with class prompts below.
    instance_prompts = [example["instance_prompt"] 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)
    # Qwen expects a `num_frames` dimension too.
    if pixel_values.ndim == 4:
        pixel_values = pixel_values.unsqueeze(2)
    pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()

    batch = {
        "indices": indices,
        "pixel_values": pixel_values,
        "instance_prompts": instance_prompts,
        "prompts": prompts,
    }
    return batch


class BucketBatchSampler(BatchSampler):
    def __init__(self, dataset: DreamBoothDataset, batch_size: int, drop_last: bool = False, seed: int = None):
        if not isinstance(batch_size, int) or batch_size <= 0:
            raise ValueError("batch_size should be a positive integer value, but got batch_size={}".format(batch_size))
        if not isinstance(drop_last, bool):
            raise ValueError("drop_last should be a boolean value, but got drop_last={}".format(drop_last))

        self.dataset = dataset
        self.batch_size = batch_size
        self.drop_last = drop_last
        self.generator = random.Random(seed) if seed is not None else random

        # Group indices by bucket
        self.bucket_indices = [[] for _ in range(len(self.dataset.buckets))]
        for idx, (_, bucket_idx) in enumerate(self.dataset.pixel_values):
            self.bucket_indices[bucket_idx].append(idx)

        self.sampler_len = 0
        for indices_in_bucket in self.bucket_indices:
            num_batches, remainder = divmod(len(indices_in_bucket), self.batch_size)
            self.sampler_len += num_batches
            if remainder > 0 and not self.drop_last:
                self.sampler_len += 1

    def __iter__(self):
        batches = []
        for indices_in_bucket in self.bucket_indices:
            shuffled_indices = indices_in_bucket.copy()
            self.generator.shuffle(shuffled_indices)
            for i in range(0, len(shuffled_indices), self.batch_size):
                batch = shuffled_indices[i : i + self.batch_size]
                if len(batch) < self.batch_size and self.drop_last:
                    continue
                batches.append(batch)

        self.generator.shuffle(batches)
        for batch in batches:
            yield batch

    def __len__(self):
        return self.sampler_len


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 concat_prompt_embedding_batches(
    *prompt_embedding_pairs: tuple[torch.Tensor, torch.Tensor],
) -> tuple[torch.Tensor, torch.Tensor]:
    """Concatenate prompt embedding batches along the batch dimension for prior preservation.

    Krea 2 tokenizes every prompt to the same fixed sequence length, so the `(B, seq, num_text_layers,
    dim)` embeddings and their `(B, seq)` masks already share a sequence length and can be concatenated
    directly.
    """
    merged_prompt_embeds = torch.cat([prompt_embeds for prompt_embeds, _ in prompt_embedding_pairs], dim=0)
    merged_mask = torch.cat([prompt_embeds_mask for _, prompt_embeds_mask in prompt_embedding_pairs], dim=0)
    return merged_prompt_embeds, merged_mask


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 `hf auth login` to authenticate with the Hub."
        )

    if torch.backends.mps.is_available() and args.mixed_precision == "bf16":
        # due to pytorch#99272, MPS does not yet support bfloat16.
        raise ValueError(
            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
        )

    logging_dir = Path(args.output_dir, args.logging_dir)

    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)
    kwargs = DistributedDataParallelKwargs(find_unused_parameters=True)
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
        kwargs_handlers=[kwargs],
    )

    # Disable AMP for MPS.
    if torch.backends.mps.is_available():
        accelerator.native_amp = False

    if args.report_to == "wandb":
        if not is_wandb_available():
            raise ImportError("Make sure to install wandb if you want to use it for logging during training.")

    # Make one log on every process with the configuration for debugging.
    logging.basicConfig(
        format="%(asctime)s - %(levelname)s - %(name)s - %(message)s",
        datefmt="%m/%d/%Y %H:%M:%S",
        level=logging.INFO,
    )
    logger.info(accelerator.state, main_process_only=False)
    if accelerator.is_local_main_process:
        transformers.utils.logging.set_verbosity_warning()
        diffusers.utils.logging.set_verbosity_info()
    else:
        transformers.utils.logging.set_verbosity_error()
        diffusers.utils.logging.set_verbosity_error()

    # If passed along, set the training seed now.
    if args.seed is not None:
        set_seed(args.seed)

    # Generate class images if prior preservation is enabled.
    if args.with_prior_preservation:
        class_images_dir = Path(args.class_data_dir)
        if not class_images_dir.exists():
            class_images_dir.mkdir(parents=True)
        cur_class_images = len(list(class_images_dir.iterdir()))

        if cur_class_images < args.num_class_images:
            pipeline = Krea2Pipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                torch_dtype=torch.bfloat16 if args.mixed_precision == "bf16" else torch.float16,
                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)

            pipeline.to("cpu")
            del pipeline
            free_memory()

    # 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 = AutoTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
    )

    # 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

    # Load scheduler and models
    # Krea 2's scheduler uses resolution-aware dynamic shifting, so the static `shift` is ignored for the training
    # sigma grid; load it straight from the checkpoint config.
    noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler", revision=args.revision
    )
    noise_scheduler_copy = copy.deepcopy(noise_scheduler)
    vae = AutoencoderKLQwenImage.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="vae",
        revision=args.revision,
        variant=args.variant,
    )
    latents_mean = (torch.tensor(vae.config.latents_mean).view(1, vae.config.z_dim, 1, 1, 1)).to(accelerator.device)
    latents_std = 1.0 / torch.tensor(vae.config.latents_std).view(1, vae.config.z_dim, 1, 1, 1).to(accelerator.device)
    text_encoder = Qwen3VLModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, torch_dtype=weight_dtype
    )
    quantization_config = None
    if args.bnb_quantization_config_path is not None:
        with open(args.bnb_quantization_config_path, "r") as f:
            config_kwargs = json.load(f)
            if "load_in_4bit" in config_kwargs and config_kwargs["load_in_4bit"]:
                config_kwargs["bnb_4bit_compute_dtype"] = weight_dtype
        quantization_config = BitsAndBytesConfig(**config_kwargs)

    transformer = Krea2Transformer2DModel.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="transformer",
        revision=args.revision,
        variant=args.variant,
        quantization_config=quantization_config,
        torch_dtype=weight_dtype,
    )
    if args.bnb_quantization_config_path is not None:
        transformer = prepare_model_for_kbit_training(transformer, use_gradient_checkpointing=False)

    if args.do_fp8_training:
        from torchao.float8 import Float8LinearConfig, convert_to_float8_training

        convert_to_float8_training(
            transformer, module_filter_fn=module_filter_fn, config=Float8LinearConfig(pad_inner_dim=True)
        )

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

    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."
        )

    to_kwargs = {"dtype": weight_dtype, "device": accelerator.device} if not args.offload else {"dtype": weight_dtype}
    # flux vae is stable in bf16 so load it in weight_dtype to reduce memory
    vae.to(**to_kwargs)
    text_encoder.to(**to_kwargs)
    # we never offload the transformer to CPU, so we can just use the accelerator device
    transformer_to_kwargs = (
        {"device": accelerator.device}
        if args.bnb_quantization_config_path is not None
        else {"device": accelerator.device, "dtype": weight_dtype}
    )
    transformer.to(**transformer_to_kwargs)

    # Initialize a text encoding pipeline and keep it to CPU for now. `text_encoder_select_layers` (which
    # decoder layers to tap) is restored from the pipeline config by `from_pretrained`.
    text_encoding_pipeline = Krea2Pipeline.from_pretrained(
        args.pretrained_model_name_or_path,
        vae=None,
        transformer=None,
        tokenizer=tokenizer,
        text_encoder=text_encoder,
        scheduler=None,
    )

    if args.gradient_checkpointing:
        transformer.enable_gradient_checkpointing()

    if args.lora_layers is not None:
        target_modules = [layer.strip() for layer in args.lora_layers.split(",")]
    else:
        # The Krea 2 authors' recommended default config (fits most styles, including high-frequency detail):
        # rank/alpha 32 on the layers below. Names map to their reference layer list as:
        #   first -> img_in, last.linear -> final_layer.linear, wq/wk/wv/wo -> to_q/to_k/to_v/to_out.0,
        #   gate -> to_gate, mlp.up/mlp.down -> ff.up/ff.down, txtfusion.projector -> text_fusion.projector,
        #   txtmlp.1/txtmlp.3 -> txt_in.linear_1/txt_in.linear_2, tmlp.0/tmlp.2 -> time_embed.linear_1/linear_2,
        #   tproj.1 -> time_mod_proj.
        # For long runs, the authors suggest raising the rank and narrowing to the attention layers
        # ("to_q,to_k,to_v,to_out.0,to_gate") via --lora_layers so prompt adherence doesn't drop.
        target_modules = [
            "img_in",
            "final_layer.linear",
            "to_q",
            "to_k",
            "to_v",
            "to_out.0",
            "to_gate",
            "ff.up",
            "ff.down",
            "text_fusion.projector",
            "txt_in.linear_1",
            "txt_in.linear_2",
            "time_embed.linear_1",
            "time_embed.linear_2",
            "time_mod_proj",
        ]

    # now we will add new LoRA weights the transformer layers
    transformer_lora_config = LoraConfig(
        r=args.rank,
        lora_alpha=args.lora_alpha,
        lora_dropout=args.lora_dropout,
        init_lora_weights="gaussian",
        target_modules=target_modules,
    )
    transformer.add_adapter(transformer_lora_config)

    def unwrap_model(model):
        model = accelerator.unwrap_model(model)
        model = model._orig_mod if is_compiled_module(model) else model
        return model

    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
        if accelerator.is_main_process:
            transformer_lora_layers_to_save = None
            modules_to_save = {}

            for model in models:
                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
                    model = unwrap_model(model)
                    transformer_lora_layers_to_save = get_peft_model_state_dict(model)
                    modules_to_save["transformer"] = model
                else:
                    raise ValueError(f"unexpected save model: {model.__class__}")

                # make sure to pop weight so that corresponding model is not saved again
                if weights:
                    weights.pop()

            Krea2Pipeline.save_lora_weights(
                output_dir,
                transformer_lora_layers=transformer_lora_layers_to_save,
                **_collate_lora_metadata(modules_to_save),
            )

    def load_model_hook(models, input_dir):
        transformer_ = None

        if not accelerator.distributed_type == DistributedType.DEEPSPEED:
            while len(models) > 0:
                model = models.pop()

                if isinstance(unwrap_model(model), type(unwrap_model(transformer))):
                    model = unwrap_model(model)
                    transformer_ = model
                else:
                    raise ValueError(f"unexpected save model: {model.__class__}")
        else:
            transformer_ = Krea2Transformer2DModel.from_pretrained(
                args.pretrained_model_name_or_path, subfolder="transformer"
            )
            transformer_.add_adapter(transformer_lora_config)

        lora_state_dict = Krea2Pipeline.lora_state_dict(input_dir)

        transformer_state_dict = {
            f"{k.replace('transformer.', '')}": v for k, v in lora_state_dict.items() if k.startswith("transformer.")
        }
        transformer_state_dict = convert_unet_state_dict_to_peft(transformer_state_dict)
        incompatible_keys = set_peft_model_state_dict(transformer_, transformer_state_dict, adapter_name="default")
        if incompatible_keys is not None:
            # check only for unexpected keys
            unexpected_keys = getattr(incompatible_keys, "unexpected_keys", None)
            if unexpected_keys:
                logger.warning(
                    f"Loading adapter weights from state_dict led to unexpected keys not found in the model: "
                    f" {unexpected_keys}. "
                )

        # Make sure the trainable params are in float32. This is again needed since the base models
        # are in `weight_dtype`. More details:
        # https://github.com/huggingface/diffusers/pull/6514#discussion_r1449796804
        if args.mixed_precision == "fp16":
            models = [transformer_]
            # only upcast trainable parameters (LoRA) into fp32
            cast_training_params(models)

    accelerator.register_save_state_pre_hook(save_model_hook)
    accelerator.register_load_state_pre_hook(load_model_hook)

    # Enable TF32 for faster training on Ampere GPUs,
    # cf https://pytorch.org/docs/stable/notes/cuda.html#tensorfloat-32-tf32-on-ampere-devices
    if args.allow_tf32 and torch.cuda.is_available():
        torch.backends.cuda.matmul.allow_tf32 = True

    if args.scale_lr:
        args.learning_rate = (
            args.learning_rate * args.gradient_accumulation_steps * args.train_batch_size * accelerator.num_processes
        )

    # Make sure the trainable params are in float32.
    if args.mixed_precision == "fp16":
        models = [transformer]
        # only upcast trainable parameters (LoRA) into fp32
        cast_training_params(models, dtype=torch.float32)

    transformer_lora_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))

    # Optimization parameters
    transformer_parameters_with_lr = {"params": transformer_lora_parameters, "lr": args.learning_rate}
    params_to_optimize = [transformer_parameters_with_lr]

    # Optimizer creation
    if not (args.optimizer.lower() == "prodigy" or args.optimizer.lower() == "adamw"):
        logger.warning(
            f"Unsupported choice of optimizer: {args.optimizer}.Supported optimizers include [adamW, prodigy]."
            "Defaulting to adamW"
        )
        args.optimizer = "adamw"

    if args.use_8bit_adam and not args.optimizer.lower() == "adamw":
        logger.warning(
            f"use_8bit_adam is ignored when optimizer is not set to 'AdamW'. Optimizer was "
            f"set to {args.optimizer.lower()}"
        )

    if args.optimizer.lower() == "adamw":
        if args.use_8bit_adam:
            try:
                import bitsandbytes as bnb
            except ImportError:
                raise ImportError(
                    "To use 8-bit Adam, please install the bitsandbytes library: `pip install bitsandbytes`."
                )

            optimizer_class = bnb.optim.AdamW8bit
        else:
            optimizer_class = torch.optim.AdamW

        optimizer = optimizer_class(
            params_to_optimize,
            betas=(args.adam_beta1, args.adam_beta2),
            weight_decay=args.adam_weight_decay,
            eps=args.adam_epsilon,
        )

    if args.optimizer.lower() == "prodigy":
        try:
            import prodigyopt
        except ImportError:
            raise ImportError("To use Prodigy, please install the prodigyopt library: `pip install prodigyopt`")

        optimizer_class = prodigyopt.Prodigy

        if args.learning_rate <= 0.1:
            logger.warning(
                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
            )

        optimizer = optimizer_class(
            params_to_optimize,
            betas=(args.adam_beta1, args.adam_beta2),
            beta3=args.prodigy_beta3,
            weight_decay=args.adam_weight_decay,
            eps=args.adam_epsilon,
            decouple=args.prodigy_decouple,
            use_bias_correction=args.prodigy_use_bias_correction,
            safeguard_warmup=args.prodigy_safeguard_warmup,
        )

    # Resolve the bucketing mode. Bucketing must be enabled explicitly with --use_aspect_ratio_buckets;
    # a bucket list without that flag is an error. With the flag, an explicit --aspect_ratio_buckets list
    # drives assignment, otherwise buckets are computed on the fly inside the dataset. Without the flag a
    # single square bucket reproduces the fixed-size resize + crop.
    if args.aspect_ratio_buckets is not None and not args.use_aspect_ratio_buckets:
        raise ValueError("--aspect_ratio_buckets requires --use_aspect_ratio_buckets to be set.")
    if args.aspect_ratio_buckets is not None:
        buckets = parse_buckets_string(args.aspect_ratio_buckets)
        use_aspect_ratio_buckets = False
        logger.info(f"Using explicit aspect ratio buckets: {buckets}")
    elif args.use_aspect_ratio_buckets:
        buckets = None
        use_aspect_ratio_buckets = True
        logger.info(
            "No --aspect_ratio_buckets provided; auto-computing aspect ratio buckets on the fly from --resolution."
        )
    else:
        buckets = [(args.resolution, args.resolution)]
        use_aspect_ratio_buckets = False

    # 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,
        buckets=buckets,
        use_aspect_ratio_buckets=use_aspect_ratio_buckets,
    )
    precompute_latents = args.cache_latents or train_dataset.custom_instance_prompts
    batch_sampler = BucketBatchSampler(train_dataset, batch_size=args.train_batch_size, drop_last=True, seed=args.seed)
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        batch_sampler=batch_sampler,
        collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
        num_workers=args.dataloader_num_workers,
    )

    def compute_text_embeddings(prompt, text_encoding_pipeline):
        with torch.no_grad():
            prompt_embeds, prompt_embeds_mask = text_encoding_pipeline.encode_prompt(
                prompt=prompt, max_sequence_length=args.max_sequence_length
            )
        return prompt_embeds, prompt_embeds_mask

    # If no type of tuning is done on the text_encoder and custom instance prompts are NOT
    # provided (i.e. the --instance_prompt is used for all images), we encode the instance prompt once to avoid
    # the redundant encoding.
    if not train_dataset.custom_instance_prompts:
        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
            instance_prompt_embeds, instance_prompt_embeds_mask = compute_text_embeddings(
                args.instance_prompt, text_encoding_pipeline
            )

    # Handle class prompt for prior-preservation.
    if args.with_prior_preservation:
        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
            class_prompt_embeds, class_prompt_embeds_mask = compute_text_embeddings(
                args.class_prompt, text_encoding_pipeline
            )

    # When caption dropout is enabled, we precompute the empty ("") prompt embedding once and swap it in
    # for randomly selected instance samples at training time (see the training loop below).
    if args.caption_dropout > 0:
        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
            empty_prompt_embeds, empty_prompt_embeds_mask = compute_text_embeddings("", text_encoding_pipeline)

    validation_embeddings = {}
    if args.validation_prompt is not None:
        with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
            (validation_embeddings["prompt_embeds"], validation_embeddings["prompt_embeds_mask"]) = (
                compute_text_embeddings(args.validation_prompt, text_encoding_pipeline)
            )
            # Krea 2 enables classifier-free guidance whenever `guidance_scale > 0` and then encodes the
            # negative prompt. The validation pipeline drops the text encoder to save memory, so precompute
            # the (empty) negative-prompt embeddings here and pass them through to inference.
            (
                validation_embeddings["negative_prompt_embeds"],
                validation_embeddings["negative_prompt_embeds_mask"],
            ) = compute_text_embeddings("", text_encoding_pipeline)

    # if cache_latents is set to True, we encode images to latents and store them.
    # Similar to pre-encoding in the case of a single instance prompt, if custom prompts are provided
    # we encode them in advance as well. Caches are keyed by dataset index so they stay correct under
    # aspect-ratio bucketing, where the batch composition differs between the caching pass and training.
    if args.cache_latents:
        instance_latents_cache = [None] * train_dataset.num_instance_images
        class_latents_cache = [None] * train_dataset.num_instance_images if args.with_prior_preservation else None
    if train_dataset.custom_instance_prompts:
        prompt_embeds_cache = [None] * train_dataset.num_instance_images
        prompt_embeds_mask_cache = [None] * train_dataset.num_instance_images
    if precompute_latents:
        cache_batch_sampler = BucketBatchSampler(
            train_dataset, batch_size=args.train_batch_size, drop_last=False, seed=args.seed
        )
        cache_dataloader = torch.utils.data.DataLoader(
            train_dataset,
            batch_sampler=cache_batch_sampler,
            collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
            num_workers=args.dataloader_num_workers,
        )
        for batch in tqdm(cache_dataloader, desc="Caching latents"):
            with torch.no_grad():
                sample_indices = batch["indices"]
                if args.cache_latents:
                    with offload_models(vae, device=accelerator.device, offload=args.offload):
                        batch["pixel_values"] = batch["pixel_values"].to(
                            accelerator.device, non_blocking=True, dtype=vae.dtype
                        )
                        latents = vae.encode(batch["pixel_values"]).latent_dist.sample()
                    if args.with_prior_preservation:
                        instance_latents, class_latents = torch.chunk(latents, 2, dim=0)
                    else:
                        instance_latents = latents
                    for i, idx in enumerate(sample_indices):
                        instance_latents_cache[idx] = instance_latents[i : i + 1]
                        if args.with_prior_preservation:
                            class_latents_cache[idx] = class_latents[i : i + 1]
                if train_dataset.custom_instance_prompts:
                    with offload_models(text_encoding_pipeline, device=accelerator.device, offload=args.offload):
                        prompt_embeds, prompt_embeds_mask = compute_text_embeddings(
                            batch["instance_prompts"], text_encoding_pipeline
                        )
                    for i, idx in enumerate(sample_indices):
                        prompt_embeds_cache[idx] = prompt_embeds[i : i + 1]
                        prompt_embeds_mask_cache[idx] = prompt_embeds_mask[i : i + 1]

        if args.cache_latents:
            assert all(latents is not None for latents in instance_latents_cache), "Latent cache has unfilled entries."
            if args.with_prior_preservation:
                assert all(latents is not None for latents in class_latents_cache), (
                    "Class latent cache has unfilled entries."
                )
        if train_dataset.custom_instance_prompts:
            assert all(embeds is not None for embeds in prompt_embeds_cache), (
                "Prompt embedding cache has unfilled entries."
            )

    # move back to cpu before deleting to ensure memory is freed see: https://github.com/huggingface/diffusers/issues/11376#issue-3008144624
    if args.cache_latents:
        vae = vae.to("cpu")
        del vae

    # move back to cpu before deleting to ensure memory is freed see: https://github.com/huggingface/diffusers/issues/11376#issue-3008144624
    text_encoding_pipeline = text_encoding_pipeline.to("cpu")
    del text_encoder, tokenizer
    free_memory()

    # Scheduler and math around the number of training steps.
    overrode_max_train_steps = False
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if args.max_train_steps is None:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
        overrode_max_train_steps = True

    lr_scheduler = get_scheduler(
        args.lr_scheduler,
        optimizer=optimizer,
        num_warmup_steps=args.lr_warmup_steps * accelerator.num_processes,
        num_training_steps=args.max_train_steps * accelerator.num_processes,
        num_cycles=args.lr_num_cycles,
        power=args.lr_power,
    )

    # Prepare everything with our `accelerator`.
    transformer, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        transformer, optimizer, train_dataloader, lr_scheduler
    )

    # We need to recalculate our total training steps as the size of the training dataloader may have changed.
    num_update_steps_per_epoch = math.ceil(len(train_dataloader) / args.gradient_accumulation_steps)
    if overrode_max_train_steps:
        args.max_train_steps = args.num_train_epochs * num_update_steps_per_epoch
    # Afterwards we recalculate our number of training epochs
    args.num_train_epochs = math.ceil(args.max_train_steps / num_update_steps_per_epoch)

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        tracker_name = "dreambooth-krea2-lora"
        accelerator.init_trackers(tracker_name, config=vars(args))

    # Train!
    total_batch_size = args.train_batch_size * accelerator.num_processes * args.gradient_accumulation_steps

    logger.info("***** Running training *****")
    logger.info(f"  Num examples = {len(train_dataset)}")
    logger.info(f"  Num batches each epoch = {len(train_dataloader)}")
    logger.info(f"  Num Epochs = {args.num_train_epochs}")
    logger.info(f"  Instantaneous batch size per device = {args.train_batch_size}")
    logger.info(f"  Total train batch size (w. parallel, distributed & accumulation) = {total_batch_size}")
    logger.info(f"  Gradient Accumulation steps = {args.gradient_accumulation_steps}")
    logger.info(f"  Total optimization steps = {args.max_train_steps}")
    global_step = 0
    first_epoch = 0

    # Potentially load in the weights and states from a previous save
    if args.resume_from_checkpoint:
        if args.resume_from_checkpoint != "latest":
            path = os.path.basename(args.resume_from_checkpoint)
        else:
            # Get the mos recent checkpoint
            dirs = os.listdir(args.output_dir)
            dirs = [d for d in dirs if d.startswith("checkpoint")]
            dirs = sorted(dirs, key=lambda x: int(x.split("-")[1]))
            path = dirs[-1] if len(dirs) > 0 else None

        if path is None:
            accelerator.print(
                f"Checkpoint '{args.resume_from_checkpoint}' does not exist. Starting a new training run."
            )
            args.resume_from_checkpoint = None
            initial_global_step = 0
        else:
            accelerator.print(f"Resuming from checkpoint {path}")
            accelerator.load_state(os.path.join(args.output_dir, path))
            global_step = int(path.split("-")[1])

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch

    else:
        initial_global_step = 0

    progress_bar = tqdm(
        range(0, args.max_train_steps),
        initial=initial_global_step,
        desc="Steps",
        # Only show the progress bar once on each machine.
        disable=not accelerator.is_local_main_process,
    )

    def get_sigmas(timesteps, n_dim=4, dtype=torch.float32):
        sigmas = noise_scheduler_copy.sigmas.to(device=accelerator.device, dtype=dtype)
        schedule_timesteps = noise_scheduler_copy.timesteps.to(accelerator.device)
        timesteps = timesteps.to(accelerator.device)
        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]

        sigma = sigmas[step_indices].flatten()
        while len(sigma.shape) < n_dim:
            sigma = sigma.unsqueeze(-1)
        return sigma

    # Keep the most recent validation batch around so the model card gallery is populated even when
    # `--skip_final_inference` is set (we fall back to the last interim images).
    images = []

    for epoch in range(first_epoch, args.num_train_epochs):
        transformer.train()

        for batch in train_dataloader:
            models_to_accumulate = [transformer]
            sample_indices = batch["indices"]
            n_inst = len(sample_indices)

            with accelerator.accumulate(models_to_accumulate):
                # Assemble this batch's instance prompt embeddings as one (embeds, mask) pair per sample,
                # gathered by dataset index so they stay aligned with the latents under aspect-ratio bucketing.
                if train_dataset.custom_instance_prompts:
                    instance_pairs = [
                        (prompt_embeds_cache[idx], prompt_embeds_mask_cache[idx]) for idx in sample_indices
                    ]
                else:
                    instance_pairs = [(instance_prompt_embeds, instance_prompt_embeds_mask)] * n_inst

                # Caption dropout: replace a sample's caption embedding with the empty-prompt embedding so it
                # trains unconditionally. Only instance captions are dropped, never class/prior captions.
                if args.caption_dropout > 0:
                    instance_pairs = [
                        (empty_prompt_embeds, empty_prompt_embeds_mask)
                        if random.random() < args.caption_dropout
                        else pair
                        for pair in instance_pairs
                    ]

                # collate_fn orders batches as [instance..., class...]; keep the prompt embeddings in the same order.
                prompt_pairs = instance_pairs
                if args.with_prior_preservation:
                    prompt_pairs = prompt_pairs + [(class_prompt_embeds, class_prompt_embeds_mask)] * n_inst
                prompt_embeds, prompt_embeds_mask = concat_prompt_embedding_batches(*prompt_pairs)

                # Convert images to latent space
                if args.cache_latents:
                    model_input = torch.cat([instance_latents_cache[idx] for idx in sample_indices], dim=0)
                    if args.with_prior_preservation:
                        model_input = torch.cat(
                            [model_input, torch.cat([class_latents_cache[idx] for idx in sample_indices], dim=0)],
                            dim=0,
                        )
                else:
                    with offload_models(vae, device=accelerator.device, offload=args.offload):
                        pixel_values = batch["pixel_values"].to(dtype=vae.dtype)
                    model_input = vae.encode(pixel_values).latent_dist.sample()

                model_input = (model_input - latents_mean) * latents_std
                model_input = model_input.to(dtype=weight_dtype)

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(model_input)
                bsz = model_input.shape[0]

                # Sample a random timestep for each image
                # for weighting schemes where we sample timesteps non-uniformly
                u = compute_density_for_timestep_sampling(
                    weighting_scheme=args.weighting_scheme,
                    batch_size=bsz,
                    logit_mean=args.logit_mean,
                    logit_std=args.logit_std,
                    mode_scale=args.mode_scale,
                )
                indices = (u * noise_scheduler_copy.config.num_train_timesteps).long()
                timesteps = noise_scheduler_copy.timesteps[indices].to(device=model_input.device)

                # Add noise according to flow matching.
                # zt = (1 - texp) * x + texp * z1
                sigmas = get_sigmas(timesteps, n_dim=model_input.ndim, dtype=model_input.dtype)
                noisy_model_input = (1.0 - sigmas) * model_input + sigmas * noise

                # Predict the noise residual.
                # Pack the latents into 2x2 patches: (B, C, 1, H, W) -> (B, (H/2)*(W/2), C*4).
                # Inlined from `Krea2Pipeline._pack_latents` (patch_size=2): that pipeline method is an
                # instance method (uses `self.patch_size`), so it can't be invoked at the class level here.
                noisy_model_input = noisy_model_input.permute(0, 2, 1, 3, 4)
                bsz_pack, c_pack = model_input.shape[0], model_input.shape[1]
                h_pack, w_pack, p_pack = model_input.shape[3], model_input.shape[4], 2
                packed_noisy_model_input = noisy_model_input.view(
                    bsz_pack, c_pack, h_pack // p_pack, p_pack, w_pack // p_pack, p_pack
                )
                packed_noisy_model_input = packed_noisy_model_input.permute(0, 2, 4, 1, 3, 5)
                packed_noisy_model_input = packed_noisy_model_input.reshape(
                    bsz_pack, (h_pack // p_pack) * (w_pack // p_pack), c_pack * p_pack * p_pack
                )
                # Rotary coordinates for the combined [text, image] sequence. A batch is single-bucket, so all
                # images share a resolution; derive the latent grid from the latents to support aspect-ratio
                # buckets and reuse a single set of position ids for the whole batch.
                latent_height, latent_width = model_input.shape[3], model_input.shape[4]
                grid_height = latent_height // 2
                grid_width = latent_width // 2
                position_ids = Krea2Pipeline.prepare_position_ids(
                    prompt_embeds.shape[1], grid_height, grid_width, accelerator.device
                )
                model_pred = transformer(
                    hidden_states=packed_noisy_model_input,
                    encoder_hidden_states=prompt_embeds,
                    timestep=timesteps / 1000,
                    position_ids=position_ids,
                    encoder_attention_mask=prompt_embeds_mask,
                    return_dict=False,
                )[0]
                # Unpack the predicted patches back to a latent grid. Inlined from
                # `Krea2Pipeline._unpack_latents` (patch_size=2): that pipeline method is an instance method
                # (uses `self.patch_size`/`self.vae_scale_factor`), so it can't be invoked at the class level here.
                p_un = 2
                bsz_un, _, ch_un = model_pred.shape
                h_un = latent_height
                w_un = latent_width
                model_pred = model_pred.view(bsz_un, h_un // p_un, w_un // p_un, ch_un // (p_un * p_un), p_un, p_un)
                model_pred = model_pred.permute(0, 3, 1, 4, 2, 5)
                model_pred = model_pred.reshape(bsz_un, ch_un // (p_un * p_un), 1, h_un, w_un)

                # 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)

                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)
                    weighting, weighting_prior = torch.chunk(weighting, 2, dim=0)

                    # Compute prior loss
                    prior_loss = torch.mean(
                        (weighting_prior.float() * (model_pred_prior.float() - target_prior.float()) ** 2).reshape(
                            target_prior.shape[0], -1
                        ),
                        1,
                    )
                    prior_loss = prior_loss.mean()

                # Compute regular loss.
                loss = torch.mean(
                    (weighting.float() * (model_pred.float() - target.float()) ** 2).reshape(target.shape[0], -1),
                    1,
                )
                loss = loss.mean()

                if args.with_prior_preservation:
                    # Add the prior loss to the instance loss.
                    loss = loss + args.prior_loss_weight * prior_loss

                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    params_to_clip = transformer.parameters()
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)

                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

            # Checks if the accelerator has performed an optimization step behind the scenes
            if accelerator.sync_gradients:
                progress_bar.update(1)
                global_step += 1

                if accelerator.is_main_process or accelerator.distributed_type == DistributedType.DEEPSPEED:
                    if global_step % args.checkpointing_steps == 0:
                        # _before_ saving state, check if this save would set us over the `checkpoints_total_limit`
                        if args.checkpoints_total_limit is not None:
                            checkpoints = os.listdir(args.output_dir)
                            checkpoints = [d for d in checkpoints if d.startswith("checkpoint")]
                            checkpoints = sorted(checkpoints, key=lambda x: int(x.split("-")[1]))

                            # before we save the new checkpoint, we need to have at _most_ `checkpoints_total_limit - 1` checkpoints
                            if len(checkpoints) >= args.checkpoints_total_limit:
                                num_to_remove = len(checkpoints) - args.checkpoints_total_limit + 1
                                removing_checkpoints = checkpoints[0:num_to_remove]

                                logger.info(
                                    f"{len(checkpoints)} checkpoints already exist, removing {len(removing_checkpoints)} checkpoints"
                                )
                                logger.info(f"removing checkpoints: {', '.join(removing_checkpoints)}")

                                for removing_checkpoint in removing_checkpoints:
                                    removing_checkpoint = os.path.join(args.output_dir, removing_checkpoint)
                                    shutil.rmtree(removing_checkpoint)

                        save_path = os.path.join(args.output_dir, f"checkpoint-{global_step}")
                        accelerator.save_state(save_path)
                        logger.info(f"Saved state to {save_path}")

            logs = {"loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
            progress_bar.set_postfix(**logs)
            accelerator.log(logs, step=global_step)

            if global_step >= args.max_train_steps:
                break

        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # Validation runs on --validation_model_path (e.g. Krea 2 Turbo) when set, since RAW
                # is not meant for inference; otherwise it falls back to the training checkpoint.
                pipeline = build_validation_pipeline(args, accelerator, transformer, weight_dtype)
                images = log_validation(
                    pipeline=pipeline,
                    args=args,
                    accelerator=accelerator,
                    pipeline_args=validation_embeddings,
                    torch_dtype=weight_dtype,
                    epoch=epoch,
                    pipeline_call_kwargs=_validation_call_kwargs(args),
                )
                del pipeline
                free_memory()

    # Save the lora layers
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        modules_to_save = {}
        transformer = unwrap_model(transformer)
        if args.bnb_quantization_config_path is None:
            if args.upcast_before_saving:
                transformer.to(torch.float32)
            else:
                transformer = transformer.to(weight_dtype)
        transformer_lora_layers = get_peft_model_state_dict(transformer)
        modules_to_save["transformer"] = transformer

        Krea2Pipeline.save_lora_weights(
            save_directory=args.output_dir,
            transformer_lora_layers=transformer_lora_layers,
            **_collate_lora_metadata(modules_to_save),
        )

        # `images` keeps the last interim validation batch (if any) as the gallery fallback; final
        # inference below overwrites it with freshly generated images when it runs.
        run_validation = (args.validation_prompt and args.num_validation_images > 0) or (args.final_validation_prompt)
        should_run_final_inference = not args.skip_final_inference and run_validation
        if should_run_final_inference:
            # Final inference. Like interim validation, run on --validation_model_path (e.g. Krea 2
            # Turbo) when set, since RAW is not meant for inference; else the training checkpoint.
            pipeline = Krea2Pipeline.from_pretrained(
                args.validation_model_path or args.pretrained_model_name_or_path,
                tokenizer=None,
                text_encoder=None,
                revision=args.revision,
                variant=args.variant,
                torch_dtype=weight_dtype,
            )
            # load attention processors
            pipeline.load_lora_weights(args.output_dir)

            # run inference
            images = log_validation(
                pipeline=pipeline,
                args=args,
                accelerator=accelerator,
                pipeline_args=validation_embeddings,
                epoch=epoch,
                is_final_validation=True,
                torch_dtype=weight_dtype,
                pipeline_call_kwargs=_validation_call_kwargs(args),
            )
            del pipeline
            free_memory()

        validation_prompt = args.validation_prompt if args.validation_prompt else args.final_validation_prompt
        save_model_card(
            (args.hub_model_id or Path(args.output_dir).name) if not args.push_to_hub else repo_id,
            images=images,
            base_model=args.pretrained_model_name_or_path,
            instance_prompt=args.instance_prompt,
            validation_prompt=validation_prompt,
            repo_folder=args.output_dir,
            inference_model=args.validation_model_path,
        )

        if args.push_to_hub:
            upload_folder(
                repo_id=repo_id,
                folder_path=args.output_dir,
                commit_message="End of training",
                ignore_patterns=["step_*", "epoch_*"],
            )

        images = None

    accelerator.end_training()


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