train_cm.py

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

import argparse
import copy
import functools
import gc
import logging
import math
import os
import random
import shutil
from contextlib import nullcontext
from pathlib import Path

import accelerate
import numpy as np
import torch
import torch.nn.functional as F
import torch.utils.checkpoint
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed
from datasets import load_dataset
from huggingface_hub import create_repo, upload_folder
from packaging import version
from peft import LoraConfig, get_peft_model_state_dict, set_peft_model_state_dict
from torchvision import transforms
from torchvision.transforms.functional import crop
from tqdm.auto import tqdm
from transformers import AutoTokenizer, PretrainedConfig

import diffusers
from diffusers import (
    AutoencoderKL,
    DDPMScheduler,
    StableDiffusionXLPipeline,
    UNet2DConditionModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import cast_training_params, resolve_interpolation_mode
from diffusers.utils import (
    check_min_version,
    convert_state_dict_to_diffusers,
    convert_unet_state_dict_to_peft,
    is_wandb_available,
)
from diffusers.utils.import_utils import is_xformers_available

from scheduler_lcm import LCMScheduler
from instance_metrics import InstanceMetrics


if is_wandb_available():
    import wandb  # type: ignore

# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.27.0.dev0")

logger = get_logger(__name__)

DATASET_NAME_MAPPING = {
    "lambdalabs/pokemon-blip-captions": ("image", "text"),
}


class DDIMSolver:
    def __init__(self, alphas, sigmas, num_train_timesteps=1000, num_discr_timesteps=100):
        assert num_train_timesteps % num_discr_timesteps == 0
        step_ratio = num_train_timesteps // num_discr_timesteps  # 10
        self.num_train_timesteps = num_train_timesteps
        self.num_discr_timesteps = num_discr_timesteps
        # timesteps in Stable Diffusion are from 1 to T=1000
        self.start_timesteps = (np.arange(1, num_discr_timesteps + 1) * step_ratio).astype(np.int64)  # 10,20,30,...,990,1000
        self.start_timesteps = self.start_timesteps[::-1].copy()  # 1000,990,...,30,20,10
        start_timesteps_indices = self.start_timesteps - 1  # 999,989,...,29,19,9
        self.next_timesteps = (np.arange(num_discr_timesteps) * step_ratio).astype(np.int64)  # 0,10,20,...,980,990
        self.next_timesteps = self.next_timesteps[::-1].copy()  # 990,980,...,20,10,0
        next_timesteps_indices = self.next_timesteps - 1  # 989,979,...,19,9,-1
        self.ddim_alphas_start = alphas[start_timesteps_indices]
        self.ddim_alphas_next = np.asarray(alphas[next_timesteps_indices[:-1]].tolist() + [1.])
        self.ddim_sigmas_start = sigmas[start_timesteps_indices]
        self.ddim_sigmas_next = np.asarray(sigmas[next_timesteps_indices[:-1]].tolist() + [0.])
        # Convert to torch tensors
        self.start_timesteps = torch.from_numpy(self.start_timesteps).long()
        self.next_timesteps = torch.from_numpy(self.next_timesteps).long()
        self.ddim_alphas_start = torch.from_numpy(self.ddim_alphas_start)
        self.ddim_alphas_next = torch.from_numpy(self.ddim_alphas_next)
        self.ddim_sigmas_start = torch.from_numpy(self.ddim_sigmas_start)
        self.ddim_sigmas_next = torch.from_numpy(self.ddim_sigmas_next)

    def to(self, device):
        self.start_timesteps = self.start_timesteps.to(device)
        self.next_timesteps = self.next_timesteps.to(device)
        self.ddim_alphas_start = self.ddim_alphas_start.to(device)
        self.ddim_alphas_next = self.ddim_alphas_next.to(device)
        self.ddim_sigmas_start = self.ddim_sigmas_start.to(device)
        self.ddim_sigmas_next = self.ddim_sigmas_next.to(device)
        return self

    def step(self, x_t, pred_noise, timestep_indices):
        alpha_start = self.ddim_alphas_start[timestep_indices]
        alpha_start = reshape_tensor_like(alpha_start, x_t.shape)
        alpha_next = self.ddim_alphas_next[timestep_indices]
        alpha_next = reshape_tensor_like(alpha_next, x_t.shape)
        sigma_start = self.ddim_sigmas_start[timestep_indices]
        sigma_start = reshape_tensor_like(sigma_start, x_t.shape)
        sigma_next = self.ddim_sigmas_next[timestep_indices]
        sigma_next = reshape_tensor_like(sigma_next, x_t.shape)
        pred_x_0 = (x_t - sigma_start * pred_noise) / alpha_start
        dir_x_t = sigma_next * pred_noise
        x_next = alpha_next * pred_x_0 + dir_x_t
        return None, x_next


class EulerSolver:
    def __init__(self, betas, sigmas, num_train_timesteps=1000, num_discr_timesteps=100):
        assert num_train_timesteps % num_discr_timesteps == 0
        step_ratio = num_train_timesteps // num_discr_timesteps  # 10
        self.num_train_timesteps = num_train_timesteps
        self.num_discr_timesteps = num_discr_timesteps
        # timesteps in Stable Diffusion are from 1 to T=1000
        self.start_timesteps = (np.arange(1, num_discr_timesteps + 1) * step_ratio).astype(np.int64)  # 10,20,30,...,990,1000
        self.start_timesteps = self.start_timesteps[::-1].copy()  # 1000,990,...,30,20,10
        start_timesteps_indices = self.start_timesteps - 1  # 999,989,...,29,19,9
        self.next_timesteps = (np.arange(num_discr_timesteps) * step_ratio).astype(np.int64)  # 0,10,20,...,980,990
        self.next_timesteps = self.next_timesteps[::-1].copy()  # 990,980,...,20,10,0
        self.euler_betas = betas[start_timesteps_indices]
        self.euler_sigmas = sigmas[start_timesteps_indices]
        # Convert to torch tensors
        self.start_timesteps = torch.from_numpy(self.start_timesteps).long()
        self.next_timesteps = torch.from_numpy(self.next_timesteps).long()
        self.euler_betas = torch.from_numpy(self.euler_betas)
        self.euler_sigmas = torch.from_numpy(self.euler_sigmas)

    def to(self, device):
        self.start_timesteps = self.start_timesteps.to(device)
        self.next_timesteps = self.next_timesteps.to(device)
        self.euler_betas = self.euler_betas.to(device)
        self.euler_sigmas = self.euler_sigmas.to(device)
        return self

    def step(self, x_t, pred_noise, timestep_indices):
        start_timestep = self.start_timesteps[timestep_indices].float()
        next_timestep = self.next_timesteps[timestep_indices].float()
        d_t = (next_timestep - start_timestep) / self.num_train_timesteps
        d_t = reshape_tensor_like(d_t, x_t.shape)
        beta = self.euler_betas[timestep_indices]
        beta = reshape_tensor_like(beta, x_t.shape)
        sigma = self.euler_sigmas[timestep_indices]
        sigma = reshape_tensor_like(sigma, x_t.shape)
        x_t_coeff = -0.5 * beta
        pred_noise_coeff = 0.5 * (1 / sigma) * beta
        derivative = x_t_coeff * x_t + pred_noise_coeff * pred_noise
        x_next = x_t + d_t * derivative
        return derivative, x_next


class HeunSolver:
    def __init__(self, betas, sigmas, num_train_timesteps=1000, num_discr_timesteps=100):
        assert num_train_timesteps % num_discr_timesteps == 0
        step_ratio = num_train_timesteps // num_discr_timesteps  # 10
        self.num_train_timesteps = num_train_timesteps
        self.num_discr_timesteps = num_discr_timesteps
        # timesteps in Stable Diffusion are from 1 to T=1000
        self.start_timesteps = (np.arange(1, num_discr_timesteps + 1) * step_ratio).astype(np.int64)  # 10,20,30,...,990,1000
        self.start_timesteps = self.start_timesteps[::-1].copy()  # 1000,990,...,30,20,10
        self.next_timesteps = (np.arange(num_discr_timesteps) * step_ratio).astype(np.int64)  # 0,10,20,...,980,990
        self.next_timesteps = self.next_timesteps[::-1].copy()  # 990,980,...,20,10,0
        next_timesteps_indices = self.next_timesteps - 1  # 989,979,...,19,9,-1
        self.heun_betas = np.asarray(betas[next_timesteps_indices[:-1]].tolist() + [betas[0]])
        self.heun_sigmas = np.asarray(sigmas[next_timesteps_indices[:-1]].tolist() + [sigmas[0]])
        # Convert to torch tensors
        self.start_timesteps = torch.from_numpy(self.start_timesteps).long()
        self.next_timesteps = torch.from_numpy(self.next_timesteps).long()
        self.heun_betas = torch.from_numpy(self.heun_betas)
        self.heun_sigmas = torch.from_numpy(self.heun_sigmas)

    def to(self, device):
        self.start_timesteps = self.start_timesteps.to(device)
        self.next_timesteps = self.next_timesteps.to(device)
        self.heun_betas = self.heun_betas.to(device)
        self.heun_sigmas = self.heun_sigmas.to(device)
        return self

    def step(self, x_t, derivative_t, pred_x_next, pred_noise_next, timestep_indices):
        start_timestep = self.start_timesteps[timestep_indices].float()
        next_timestep = self.next_timesteps[timestep_indices].float()
        d_t = (next_timestep - start_timestep) / self.num_train_timesteps
        d_t = reshape_tensor_like(d_t, x_t.shape)
        beta = self.heun_betas[timestep_indices]
        beta = reshape_tensor_like(beta, x_t.shape)
        sigma = self.heun_sigmas[timestep_indices]
        sigma = reshape_tensor_like(sigma, x_t.shape)
        pred_x_next_coeff = -0.5 * beta
        pred_noise_next_coeff = 0.5 * (1 / sigma) * beta
        derivative_next = pred_x_next_coeff * pred_x_next + pred_noise_next_coeff * pred_noise_next
        x_next = x_t + 0.5 * d_t * (derivative_t + derivative_next)
        return x_next


def log_validation(vae, args, accelerator, weight_dtype, step, unet=None, is_final_validation=False):
    logger.info("Running validation... ")

    pipeline = StableDiffusionXLPipeline.from_pretrained(
        args.pretrained_teacher_model,
        vae=vae,
        scheduler=LCMScheduler.from_pretrained(args.pretrained_teacher_model, subfolder="scheduler"),
        revision=args.revision,
        torch_dtype=weight_dtype,
    ).to(accelerator.device)
    pipeline.set_progress_bar_config(disable=True)
    pipeline.scheduler.config.num_original_inference_steps = args.num_discr_timesteps

    to_load = None
    if not is_final_validation:
        if unet is None:
            raise ValueError("Must provide a `unet` when doing intermediate validation.")
        unet = accelerator.unwrap_model(unet)
        state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
        to_load = state_dict
    else:
        to_load = args.output_dir

    pipeline.load_lora_weights(to_load)
    pipeline.fuse_lora()

    if args.enable_xformers_memory_efficient_attention:
        pipeline.enable_xformers_memory_efficient_attention()

    pipeline.unet.eval()

    instance_metrics = InstanceMetrics(accelerator.device)

    if args.seed is None:
        generator = None
    else:
        generator = torch.Generator(device=accelerator.device).manual_seed(args.seed)

    validation_prompts = [
        "cute sundar pichai character",
        "robotic cat with wings",
        "a photo of yoda",
        "a cute creature with blue eyes",
    ]

    image_logs, clip_scores, aesthetic_scores = [], [], []
    for _, prompt in enumerate(validation_prompts):
        images = []
        if torch.backends.mps.is_available():
            autocast_ctx = nullcontext()
        else:
            autocast_ctx = torch.autocast(accelerator.device.type, dtype=weight_dtype)

        with autocast_ctx:
            images = pipeline(
                prompt=prompt,
                num_inference_steps=args.num_inference_steps,
                num_images_per_prompt=args.num_images_per_prompt,
                generator=generator,
                guidance_scale=0.0,
            ).images
        image_logs.append({"validation_prompt": prompt, "images": images})

        clip_score, aesthetic_score = instance_metrics.compute_instance_metrics(images[0], prompt)
        clip_scores.append(clip_score)
        aesthetic_scores.append(aesthetic_score)

    avg_clip_score = sum(clip_scores) / len(clip_scores)
    avg_aesthetic_score = sum(aesthetic_scores) / len(aesthetic_scores)

    for tracker in accelerator.trackers:
        if tracker.name == "tensorboard":
            for log in image_logs:
                images = log["images"]
                validation_prompt = log["validation_prompt"]
                formatted_images = []
                for image in images:
                    formatted_images.append(np.asarray(image))

                formatted_images = np.stack(formatted_images)

                tracker.writer.add_images(validation_prompt, formatted_images, step, dataformats="NHWC")
            tracker.writer.add_scalar("clip score", avg_clip_score, step)
            tracker.writer.add_scalar("aesthetic score", avg_aesthetic_score, step)
        elif tracker.name == "wandb":
            formatted_images = []

            for log in image_logs:
                images = log["images"]
                validation_prompt = log["validation_prompt"]
                for image in images:
                    image = wandb.Image(image, caption=validation_prompt)
                    formatted_images.append(image)
            logger_name = "test" if is_final_validation else "validation"
            tracker.log({logger_name: formatted_images})
        else:
            logger.warning(f"image logging not implemented for {tracker.name}")

        del pipeline
        gc.collect()
        torch.cuda.empty_cache()

        return image_logs


def append_dims(x, target_dims):
    """Appends dimensions to the end of a tensor until it has target_dims dimensions."""
    dims_to_append = target_dims - x.ndim
    if dims_to_append < 0:
        raise ValueError(f"input has {x.ndim} dims but target_dims is {target_dims}, which is less")
    return x[(...,) + (None,) * dims_to_append]


def scalings_for_boundary_conditions(timestep, sigma_data=0.5, timestep_scaling=10.0):
    scaled_timestep = timestep_scaling * timestep
    c_skip = sigma_data**2 / (scaled_timestep**2 + sigma_data**2)
    c_out = scaled_timestep / (scaled_timestep**2 + sigma_data**2) ** 0.5
    return c_skip, c_out


def extract_into_tensor(a, t, x_shape):
    b, *_ = t.shape
    t = torch.nn.functional.relu(t)  # Zero negative timesteps
    out = a.gather(-1, t)
    return out.reshape(b, *((1,) * (len(x_shape) - 1)))


def reshape_tensor_like(a, x_shape):
    b, *_ = x_shape
    return a.reshape(b, *((1,) * (len(x_shape) - 1)))


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

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

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

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


def parse_args():
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    # ----------Model Checkpoint Loading Arguments----------
    parser.add_argument(
        "--pretrained_teacher_model",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained LDM teacher model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--pretrained_vae_model_name_or_path",
        type=str,
        default=None,
        help="Path to pretrained VAE model with better numerical stability. More details: https://github.com/huggingface/diffusers/pull/4038.",
    )
    parser.add_argument(
        "--teacher_revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained LDM teacher model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--revision",
        type=str,
        default=None,
        required=False,
        help="Revision of pretrained LDM model identifier from huggingface.co/models.",
    )
    # ----------Training Arguments----------
    # ----General Training Arguments----
    parser.add_argument(
        "--output_dir",
        type=str,
        default="lcm-xl-distilled",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    # ----Logging----
    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(
        "--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.'
        ),
    )
    # ----Checkpointing----
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=500,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints are only 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.'
        ),
    )
    # ----Image Processing----
    parser.add_argument(
        "--dataset_name",
        type=str,
        default=None,
        help=(
            "The name of the Dataset (from the HuggingFace hub) to train on (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(
        "--train_data_dir",
        type=str,
        default=None,
        help=(
            "A folder containing the training data. Folder contents must follow the structure described in"
            " https://huggingface.co/docs/datasets/image_dataset#imagefolder. In particular, a `metadata.jsonl` file"
            " must exist to provide the captions for the images. Ignored if `dataset_name` is specified."
        ),
    )
    parser.add_argument(
        "--image_column", type=str, default="image", help="The column of the dataset containing an image."
    )
    parser.add_argument(
        "--caption_column",
        type=str,
        default="text",
        help="The column of the dataset containing a caption or a list of captions.",
    )
    parser.add_argument(
        "--resolution",
        type=int,
        default=1024,
        help=(
            "The resolution for input images, all the images in the train/validation dataset will be resized to this"
            " resolution"
        ),
    )
    parser.add_argument(
        "--interpolation_type",
        type=str,
        default="bilinear",
        help=(
            "The interpolation function used when resizing images to the desired resolution. Choose between `bilinear`,"
            " `bicubic`, `box`, `nearest`, `nearest_exact`, `hamming`, and `lanczos`."
        ),
    )
    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",
    )
    # ----Dataloader----
    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."
        ),
    )
    # ----Batch Size and Training Steps----
    parser.add_argument(
        "--train_batch_size", type=int, default=16, help="Batch size (per device) for the training dataloader."
    )
    parser.add_argument("--num_train_epochs", type=int, default=100)
    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(
        "--max_train_samples",
        type=int,
        default=None,
        help=(
            "For debugging purposes or quicker training, truncate the number of training examples to this "
            "value if set."
        ),
    )
    # ----Learning Rate----
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-6,
        help="Initial learning rate (after the potential warmup period) to use.",
    )
    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=False,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="constant",
        help=(
            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
            ' "constant", "constant_with_warmup"]'
        ),
    )
    parser.add_argument(
        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    # ----Optimizer (Adam)----
    parser.add_argument(
        "--use_8bit_adam", action="store_true", help="Whether or not to use 8-bit Adam from bitsandbytes."
    )
    parser.add_argument("--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam optimizer.")
    parser.add_argument("--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam optimizer.")
    parser.add_argument("--adam_weight_decay", type=float, default=1e-2, help="Weight decay to use.")
    parser.add_argument("--adam_epsilon", type=float, default=1e-08, help="Epsilon value for the Adam optimizer")
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    # ----Diffusion Training Arguments----
    # ----Latent Consistency Distillation (LCD) Specific Arguments----
    parser.add_argument(
        "--w",
        type=float,
        default=8.5,
        required=False,
        help=(
            "The guidance scale value for guidance scale sampling of the teacher. Note that we are using the Imagen CFG"
            " formulation rather than the LCM formulation, which means all guidance scales have 1 added to them as"
            " compared to the original paper."
        ),
    )
    parser.add_argument(
        "--solver",
        type=str,
        default="ddim",
        choices=["ddim", "euler", "heun"],
        help="The ODE solver to use for the teacher model",
    )
    parser.add_argument(
        "--num_discr_timesteps",
        type=int,
        default=1000,
        help="The number of ODE discretization steps",
    )
    parser.add_argument(
        "--loss_type",
        type=str,
        default="l2",
        choices=["l2", "huber"],
        help="The type of loss to use for the LCD loss.",
    )
    parser.add_argument(
        "--huber_c",
        type=float,
        default=0.001,
        help="The huber loss parameter. Only used if `--loss_type=huber`.",
    )
    parser.add_argument(
        "--lora_rank",
        type=int,
        default=64,
        help="The rank of the LoRA projection matrix.",
    )
    parser.add_argument(
        "--lora_alpha",
        type=int,
        default=64,
        help=(
            "The value of the LoRA alpha parameter, which controls the scaling factor in front of the LoRA weight"
            " update delta_W. No scaling will be performed if this value is equal to `lora_rank`."
        ),
    )
    parser.add_argument(
        "--lora_dropout",
        type=float,
        default=0.0,
        help="The dropout probability for the dropout layer added before applying the LoRA to each layer input.",
    )
    parser.add_argument(
        "--lora_target_modules",
        type=str,
        default=None,
        help=(
            "A comma-separated string of target module keys to add LoRA to. If not set, a default list of modules will"
            " be used. By default, LoRA will be applied to all conv and linear layers."
        ),
    )
    parser.add_argument(
        "--vae_encode_batch_size",
        type=int,
        default=8,
        required=False,
        help=(
            "The batch size used when encoding (and decoding) images to latents (and vice versa) using the VAE."
            " Encoding or decoding the whole batch at once may run into OOM issues."
        ),
    )
    parser.add_argument(
        "--timestep_scaling_factor",
        type=float,
        default=10.0,
        help=(
            "The multiplicative timestep scaling factor used when calculating the boundary scalings for LCM. The"
            " higher the scaling is, the lower the approximation error, but the default value of 10.0 should typically"
            " suffice."
        ),
    )
    # ----Mixed Precision----
    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(
        "--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"
        ),
    )
    # ----Training Optimizations----
    parser.add_argument(
        "--enable_xformers_memory_efficient_attention", action="store_true", help="Whether or not to use xformers."
    )
    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.",
    )
    # ----Distributed Training----
    parser.add_argument("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
    # ----------Validation Arguments----------
    parser.add_argument(
        "--validation_steps",
        type=int,
        default=200,
        help="Run validation every X steps.",
    )
    parser.add_argument("--num_inference_steps", type=int, default=4, help="Number of inference steps during validation")
    parser.add_argument("--num_images_per_prompt", type=int, default=4, help="Number of images to generate per prompt")
    # ----------Huggingface Hub Arguments-----------
    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`.",
    )
    # ----------Accelerate Arguments----------
    parser.add_argument(
        "--tracker_project_name",
        type=str,
        default="text2image-fine-tune",
        help=(
            "The `project_name` argument passed to Accelerator.init_trackers for"
            " more information see https://huggingface.co/docs/accelerate/v0.17.0/en/package_reference/accelerator#accelerate.Accelerator"
        ),
    )

    args = parser.parse_args()
    env_local_rank = int(os.environ.get("LOCAL_RANK", -1))
    if env_local_rank != -1 and env_local_rank != args.local_rank:
        args.local_rank = env_local_rank

    return args


# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
def encode_prompt(prompt_batch, text_encoders, tokenizers, is_train=True):
    prompt_embeds_list = []

    captions = []
    for caption in prompt_batch:
        if isinstance(caption, str):
            captions.append(caption)
        elif isinstance(caption, (list, np.ndarray)):
            # take a random caption if there are multiple
            captions.append(random.choice(caption) if is_train else caption[0])

    with torch.no_grad():
        for tokenizer, text_encoder in zip(tokenizers, text_encoders):
            text_inputs = tokenizer(
                captions,
                padding="max_length",
                max_length=tokenizer.model_max_length,
                truncation=True,
                return_tensors="pt",
            )
            text_input_ids = text_inputs.input_ids
            prompt_embeds = text_encoder(
                text_input_ids.to(text_encoder.device),
                output_hidden_states=True,
            )

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

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


def main(args):
    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."
        )

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

    accelerator_project_config = ProjectConfiguration(project_dir=args.output_dir, logging_dir=logging_dir)

    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
        split_batches=True,  # It's important to set this to True when using webdataset to get the right number of steps for lr scheduling. If set to False, the number of steps will be devide by the number of processes assuming batches are multiplied by the number of processes
    )

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

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

    # Handle the repository creation
    if accelerator.is_main_process:
        if args.output_dir is not None:
            os.makedirs(args.output_dir, exist_ok=True)

        if args.push_to_hub:
            repo_id = create_repo(
                repo_id=args.hub_model_id or Path(args.output_dir).name,
                exist_ok=True,
                token=args.hub_token,
                private=True,
            ).repo_id

    # 1. Create the noise scheduler and the desired noise schedule.
    #### REPLACE WITH EulerDISCREte
    noise_scheduler = DDPMScheduler.from_pretrained(
        args.pretrained_teacher_model, subfolder="scheduler", revision=args.teacher_revision
    )

    # DDPMScheduler calculates the alpha and sigma noise schedules (based on the alpha bars) for us
    alpha_schedule = torch.sqrt(noise_scheduler.alphas_cumprod)
    sigma_schedule = torch.sqrt(1 - noise_scheduler.alphas_cumprod)

    # Initialize the DDIM ODE solver for distillation.
    heun_solver = None
    if args.solver == "ddim":
        solver = DDIMSolver(
            alphas=alpha_schedule.numpy(),
            sigmas=sigma_schedule.numpy(),
            num_train_timesteps=noise_scheduler.config.num_train_timesteps,
            num_discr_timesteps=args.num_discr_timesteps,
        )
    # Initialize the Euler ODE solver for distillation.
    elif args.solver == "euler":
        solver = EulerSolver(
            betas=noise_scheduler.betas.numpy() * noise_scheduler.config.num_train_timesteps,
            sigmas=sigma_schedule.numpy(),
            num_train_timesteps=noise_scheduler.config.num_train_timesteps,
            num_discr_timesteps=args.num_discr_timesteps,
        )
    # Initialize the Heun Solver for distillation as the corrector for Euler
    elif args.solver == "heun":
        solver = EulerSolver(
            betas=noise_scheduler.betas.numpy() * noise_scheduler.config.num_train_timesteps,
            sigmas=sigma_schedule.numpy(),
            num_train_timesteps=noise_scheduler.config.num_train_timesteps,
            num_discr_timesteps=args.num_discr_timesteps,
        )

        heun_solver = HeunSolver(
            betas=noise_scheduler.betas.numpy() * noise_scheduler.config.num_train_timesteps,
            sigmas=sigma_schedule.numpy(),
            num_train_timesteps=noise_scheduler.config.num_train_timesteps,
            num_discr_timesteps=args.num_discr_timesteps,
        )

    # 2. Load tokenizers from SDXL checkpoint.
    tokenizer_one = AutoTokenizer.from_pretrained(
        args.pretrained_teacher_model, subfolder="tokenizer", revision=args.teacher_revision, use_fast=False
    )
    tokenizer_two = AutoTokenizer.from_pretrained(
        args.pretrained_teacher_model, subfolder="tokenizer_2", revision=args.teacher_revision, use_fast=False
    )

    # 3. Load text encoders from SDXL checkpoint.
    # import correct text encoder classes
    text_encoder_cls_one = import_model_class_from_model_name_or_path(
        args.pretrained_teacher_model, args.teacher_revision
    )
    text_encoder_cls_two = import_model_class_from_model_name_or_path(
        args.pretrained_teacher_model, args.teacher_revision, subfolder="text_encoder_2"
    )

    text_encoder_one = text_encoder_cls_one.from_pretrained(
        args.pretrained_teacher_model, subfolder="text_encoder", revision=args.teacher_revision
    )
    text_encoder_two = text_encoder_cls_two.from_pretrained(
        args.pretrained_teacher_model, subfolder="text_encoder_2", revision=args.teacher_revision
    )

    # 4. Load VAE from SDXL checkpoint (or more stable VAE)
    vae_path = (
        args.pretrained_teacher_model
        if args.pretrained_vae_model_name_or_path is None
        else args.pretrained_vae_model_name_or_path
    )
    vae = AutoencoderKL.from_pretrained(
        vae_path,
        subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
        revision=args.teacher_revision,
    )

    # 6. Freeze teacher vae, text_encoders.
    vae.requires_grad_(False)
    text_encoder_one.requires_grad_(False)
    text_encoder_two.requires_grad_(False)

    # 7. Create online student U-Net.
    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_teacher_model, subfolder="unet", revision=args.teacher_revision
    )
    unet.requires_grad_(False)

    # Check that all trainable models are in full precision
    low_precision_error_string = (
        " Please make sure to always have all model weights in full float32 precision when starting training - even if"
        " doing mixed precision training, copy of the weights should still be float32."
    )

    if accelerator.unwrap_model(unet).dtype != torch.float32:
        raise ValueError(
            f"Unet loaded as datatype {accelerator.unwrap_model(unet).dtype}. {low_precision_error_string}"
        )

    # 8. Handle mixed precision and device placement
    # For mixed precision training we cast all non-trainable weigths to half-precision
    # as these weights are only used for inference, keeping weights in full precision is not required.
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16

    # Move unet, vae and text_encoder to device and cast to weight_dtype
    # The VAE is in float32 to avoid NaN losses.
    unet.to(accelerator.device, dtype=weight_dtype)
    if args.pretrained_vae_model_name_or_path is None:
        vae.to(accelerator.device, dtype=torch.float32)
    else:
        vae.to(accelerator.device, dtype=weight_dtype)
    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
    text_encoder_two.to(accelerator.device, dtype=weight_dtype)

    # 9. Add LoRA to the student U-Net, only the LoRA projection matrix will be updated by the optimizer.
    if args.lora_target_modules is not None:
        lora_target_modules = [module_key.strip() for module_key in args.lora_target_modules.split(",")]
    else:
        lora_target_modules = [
            "to_q",
            "to_k",
            "to_v",
            "to_out.0",
            "proj_in",
            "proj_out",
            "ff.net.0.proj",
            "ff.net.2",
            "conv1",
            "conv2",
            "conv_shortcut",
            "downsamplers.0.conv",
            "upsamplers.0.conv",
            "time_emb_proj",
        ]
    lora_config = LoraConfig(
        r=args.lora_rank,
        target_modules=lora_target_modules,
        lora_alpha=args.lora_alpha,
        lora_dropout=args.lora_dropout,
    )
    unet.add_adapter(lora_config)

    # Also move the alpha and sigma noise schedules to accelerator.device.
    alpha_schedule = alpha_schedule.to(accelerator.device)
    sigma_schedule = sigma_schedule.to(accelerator.device)
    solver = solver.to(accelerator.device)
    if heun_solver is not None:
        heun_solver = heun_solver.to(accelerator.device)

    # 10. Handle saving and loading of checkpoints
    # `accelerate` 0.16.0 will have better support for customized saving
    if version.parse(accelerate.__version__) >= version.parse("0.16.0"):
        # 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:
                unet_ = accelerator.unwrap_model(unet)
                # also save the checkpoints in native `diffusers` format so that it can be easily
                # be independently loaded via `load_lora_weights()`.
                state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet_))
                StableDiffusionXLPipeline.save_lora_weights(output_dir, unet_lora_layers=state_dict)

                for _, model in enumerate(models):
                    # make sure to pop weight so that corresponding model is not saved again
                    weights.pop()

        def load_model_hook(models, input_dir):
            # load the LoRA into the model
            unet_ = accelerator.unwrap_model(unet)
            lora_state_dict, _ = StableDiffusionXLPipeline.lora_state_dict(input_dir)
            unet_state_dict = {
                f'{k.replace("unet.", "")}': v for k, v in lora_state_dict.items() if k.startswith("unet.")
            }
            unet_state_dict = convert_unet_state_dict_to_peft(unet_state_dict)
            incompatible_keys = set_peft_model_state_dict(unet_, unet_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}. "
                    )

            for _ in range(len(models)):
                # pop models so that they are not loaded again
                models.pop()

            # 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":
                cast_training_params(unet_, dtype=torch.float32)

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

    # 11. Enable optimizations
    if args.enable_xformers_memory_efficient_attention:
        if is_xformers_available():
            import xformers

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

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

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()

    # Use 8-bit Adam for lower memory usage or to fine-tune the model in 16GB GPUs
    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

    # 12. Optimizer creation
    params_to_optimize = filter(lambda p: p.requires_grad, unet.parameters())
    optimizer = optimizer_class(
        params_to_optimize,
        lr=args.learning_rate,
        betas=(args.adam_beta1, args.adam_beta2),
        weight_decay=args.adam_weight_decay,
        eps=args.adam_epsilon,
    )

    # 13. Dataset creation and data processing
    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
    # download the dataset.
    if args.dataset_name is not None:
        # Downloading and loading a dataset from the hub.
        dataset = load_dataset(
            args.dataset_name,
            args.dataset_config_name,
            cache_dir=args.cache_dir,
        )
    else:
        data_files = {}
        if args.train_data_dir is not None:
            data_files["train"] = os.path.join(args.train_data_dir, "**")
        dataset = load_dataset(
            "imagefolder",
            data_files=data_files,
            cache_dir=args.cache_dir,
        )
        # See more about loading custom images at
        # https://huggingface.co/docs/datasets/v2.4.0/en/image_load#imagefolder

    # Preprocessing the datasets.
    column_names = dataset["train"].column_names

    # Get the column names for input/target.
    dataset_columns = DATASET_NAME_MAPPING.get(args.dataset_name, None)
    if args.image_column is None:
        image_column = dataset_columns[0] if dataset_columns is not None else column_names[0]
    else:
        image_column = args.image_column
        if image_column not in column_names:
            raise ValueError(
                f"--image_column' value '{args.image_column}' needs to be one of: {', '.join(column_names)}"
            )
    if args.caption_column is None:
        caption_column = dataset_columns[1] if dataset_columns is not None else column_names[1]
    else:
        caption_column = args.caption_column
        if caption_column not in column_names:
            raise ValueError(
                f"--caption_column' value '{args.caption_column}' needs to be one of: {', '.join(column_names)}"
            )

    # Preprocessing the datasets.
    interpolation_mode = resolve_interpolation_mode(args.interpolation_type)
    train_resize = transforms.Resize(args.resolution, interpolation=interpolation_mode)
    train_crop = transforms.CenterCrop(args.resolution) if args.center_crop else transforms.RandomCrop(args.resolution)
    train_flip = transforms.RandomHorizontalFlip(p=1.0)
    train_transforms = transforms.Compose([transforms.ToTensor(), transforms.Normalize([0.5], [0.5])])

    def preprocess_train(examples):
        images = [image.convert("RGB") for image in examples[image_column]]
        # image aug
        original_sizes = []
        all_images = []
        crop_top_lefts = []
        for image in images:
            original_sizes.append((image.height, image.width))
            image = train_resize(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)
            if args.random_flip and random.random() < 0.5:
                # flip
                x1 = image.width - x1
                image = train_flip(image)
            crop_top_left = (y1, x1)
            crop_top_lefts.append(crop_top_left)
            image = train_transforms(image)
            all_images.append(image)

        examples["original_sizes"] = original_sizes
        examples["crop_top_lefts"] = crop_top_lefts
        examples["pixel_values"] = all_images
        examples["captions"] = list(examples[caption_column])
        return examples

    with accelerator.main_process_first():
        if args.max_train_samples is not None:
            dataset["train"] = dataset["train"].shuffle(seed=args.seed).select(range(args.max_train_samples))
        # Set the training transforms
        train_dataset = dataset["train"].with_transform(preprocess_train)

    def collate_fn(examples):
        pixel_values = torch.stack([example["pixel_values"] for example in examples])
        pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
        original_sizes = [example["original_sizes"] for example in examples]
        crop_top_lefts = [example["crop_top_lefts"] for example in examples]
        captions = [example["captions"] for example in examples]

        return {
            "pixel_values": pixel_values,
            "captions": captions,
            "original_sizes": original_sizes,
            "crop_top_lefts": crop_top_lefts,
        }

    # DataLoaders creation:
    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        shuffle=True,
        collate_fn=collate_fn,
        batch_size=args.train_batch_size,
        num_workers=args.dataloader_num_workers,
    )

    # 14. Embeddings for the UNet.
    # Adapted from pipeline.StableDiffusionXLPipeline._get_add_time_ids
    def compute_embeddings(prompt_batch, original_sizes, crop_coords, text_encoders, tokenizers, is_train=True):
        def compute_time_ids(original_size, crops_coords_top_left):
            target_size = (args.resolution, args.resolution)
            add_time_ids = list(original_size + crops_coords_top_left + target_size)
            add_time_ids = torch.tensor([add_time_ids])
            add_time_ids = add_time_ids.to(accelerator.device, dtype=weight_dtype)
            return add_time_ids

        prompt_embeds, pooled_prompt_embeds = encode_prompt(prompt_batch, text_encoders, tokenizers, is_train)
        add_text_embeds = pooled_prompt_embeds

        add_time_ids = torch.cat([compute_time_ids(s, c) for s, c in zip(original_sizes, crop_coords)])

        prompt_embeds = prompt_embeds.to(accelerator.device)
        add_text_embeds = add_text_embeds.to(accelerator.device)
        unet_added_cond_kwargs = {"text_embeds": add_text_embeds, "time_ids": add_time_ids}

        return {"prompt_embeds": prompt_embeds, **unet_added_cond_kwargs}

    text_encoders = [text_encoder_one, text_encoder_two]
    tokenizers = [tokenizer_one, tokenizer_two]

    compute_embeddings_fn = functools.partial(compute_embeddings, text_encoders=text_encoders, tokenizers=tokenizers)

    # 15. LR Scheduler creation
    # 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

    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":
        # only upcast trainable parameters (LoRA) into fp32
        cast_training_params(unet, dtype=torch.float32)

    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,
    )

    # 16. Prepare for training
    # Prepare everything with our `accelerator`.
    unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
        unet, optimizer, train_dataloader, lr_scheduler
    )

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

    # We need to initialize the trackers we use, and also store our configuration.
    # The trackers initializes automatically on the main process.
    if accelerator.is_main_process:
        tracker_config = dict(vars(args))
        # remove huggingface token from config so it is not exposed
        tracker_config.pop("hub_token", None)
        accelerator.init_trackers(args.tracker_project_name, config=tracker_config)

    # 17. 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 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 most 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,
    )

    unet.train()
    for epoch in range(first_epoch, args.num_train_epochs):
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                # 1. Load and process the image and text conditioning
                pixel_values, text, orig_size, crop_coords = (
                    batch["pixel_values"],
                    batch["captions"],
                    batch["original_sizes"],
                    batch["crop_top_lefts"],
                )

                encoded_text = compute_embeddings_fn(text, orig_size, crop_coords)

                # encode pixel values with batch size of at most args.vae_encode_batch_size
                pixel_values = pixel_values.to(dtype=vae.dtype)
                latents = []
                for i in range(0, pixel_values.shape[0], args.vae_encode_batch_size):
                    latents.append(vae.encode(pixel_values[i : i + args.vae_encode_batch_size]).latent_dist.sample())
                latents = torch.cat(latents, dim=0)

                latents = latents * vae.config.scaling_factor
                if args.pretrained_vae_model_name_or_path is None:
                    latents = latents.to(weight_dtype)

                # 2. Sample a random timestep for each image t_n from the ODE solver timesteps without bias.
                bsz = latents.shape[0]
                index = torch.randint(0, solver.num_discr_timesteps, (bsz,), device=latents.device).long()
                start_timesteps = solver.start_timesteps[index]
                next_timesteps = solver.next_timesteps[index]

                # 3. Sample noise from the prior and add it to the latents according to the noise magnitude at each
                # timestep (this is the forward diffusion process)
                noise = torch.randn_like(latents)
                noisy_model_input = noise_scheduler.add_noise(latents, noise, start_timesteps - 1)

                # 4. Set guidance scale w for the teacher
                w = args.w * torch.ones((bsz,))
                w = w.reshape(bsz, 1, 1, 1)
                w = w.to(device=latents.device, dtype=latents.dtype)

                # 5. Prepare conditional and unconditional prompts
                prompt_embeds = encoded_text.pop("prompt_embeds")
                uncond_prompt_embeds = torch.zeros_like(prompt_embeds)
                uncond_pooled_prompt_embeds = torch.zeros_like(encoded_text["text_embeds"])
                uncond_added_conditions = copy.deepcopy(encoded_text)
                uncond_added_conditions["text_embeds"] = uncond_pooled_prompt_embeds

                # 6. Solve one step of the augmented PF-ODE using the teacher model. Compute
                # conditional and unconditional teacher model samples to get CFG estimates of the predicted
                # noise, then run the ODE solver using these estimates to predict the data point in the
                # augmented PF-ODE trajectory corresponding to the next ODE solver timestep

                # With the adapters disabled, the `unet` is the regular teacher model.
                accelerator.unwrap_model(unet).disable_adapters()
                with torch.no_grad():
                    unet.eval()
                    # 7. Get teacher model prediction on noisy_model_input and conditional embedding c
                    cond_pred_noise = unet(
                        noisy_model_input,
                        start_timesteps,
                        encoder_hidden_states=prompt_embeds,
                        added_cond_kwargs={k: v.to(weight_dtype) for k, v in encoded_text.items()},
                    ).sample

                    # 8. Get teacher model prediction on noisy_model_input and unconditional embedding 0
                    uncond_pred_noise = unet(
                        noisy_model_input,
                        start_timesteps,
                        encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
                        added_cond_kwargs={k: v.to(weight_dtype) for k, v in uncond_added_conditions.items()},
                    ).sample

                    # 9. Calculate the CFG estimate of pred_noise
                    # Note that this uses the Imagen CFG formulation
                    pred_noise = uncond_pred_noise + w * (cond_pred_noise - uncond_pred_noise)

                    # 10. Run one step of the ODE solver to estimate the next point on the augmented
                    # PF-ODE trajectory
                    derivative, noisy_model_input_next = solver.step(noisy_model_input, pred_noise, index)

                    if args.solver == "ddim" or args.solver == "euler":
                        noisy_model_input_prev = noisy_model_input_next.to(accelerator.unwrap_model(unet).dtype)
                    elif args.solver == "heun":
                        # In a Heun step, we treat a Euler step as a predictor step and then run a corrector step
                        # 11. Get teacher model prediction on Euler predicted noisy_model_input and conditional embedding c
                        cond_pred_noise_next = unet(
                            noisy_model_input_next,
                            next_timesteps,
                            encoder_hidden_states=prompt_embeds,
                            added_cond_kwargs={k: v.to(weight_dtype) for k, v in encoded_text.items()},
                        ).sample

                        # 12. Get teacher model prediction on Euler predicted noisy_model_input and unconditional embedding 0
                        uncond_pred_noise_next = unet(
                            noisy_model_input_next,
                            next_timesteps,
                            encoder_hidden_states=uncond_prompt_embeds.to(weight_dtype),
                            added_cond_kwargs={k: v.to(weight_dtype) for k, v in uncond_added_conditions.items()},
                        ).sample

                        # 13. Calculate the CFG estimate of pred_noise_next
                        # Note that this uses the Imagen CFG formulation
                        pred_noise_next = uncond_pred_noise_next + w * (cond_pred_noise_next - uncond_pred_noise_next)

                        # 14. Run one step of the Heun ODE solver to estimate the next point on the augmented
                        # PF-ODE trajectory by correcting the previous Euler step
                        noisy_model_input_prev = heun_solver.step(noisy_model_input, derivative, noisy_model_input_next, pred_noise_next, index).to(accelerator.unwrap_model(unet).dtype)

                    unet.train()

                # Re-enable unet adapters to turn the `unet` into a student unet.
                accelerator.unwrap_model(unet).enable_adapters()

                # 15. Get online prediction on noisy_model_input
                noise_pred = unet(
                    noisy_model_input,
                    start_timesteps,
                    encoder_hidden_states=prompt_embeds,
                    added_cond_kwargs=encoded_text,
                ).sample

                alphas = extract_into_tensor(alpha_schedule, start_timesteps - 1, noisy_model_input.shape)
                sigmas = extract_into_tensor(sigma_schedule, start_timesteps - 1, noisy_model_input.shape)
                pred_x_0 = (noisy_model_input - sigmas * noise_pred) / alphas

                # 16. Get boundary scalings for start timesteps and apply to predicted x_0
                c_skip, c_out = scalings_for_boundary_conditions(
                    start_timesteps, timestep_scaling=args.timestep_scaling_factor
                )
                c_skip, c_out = [append_dims(x, noisy_model_input.ndim) for x in [c_skip, c_out]]

                pred_x_0 = c_skip * noisy_model_input + c_out * pred_x_0

                with torch.no_grad():
                    # 17. Get target prediction on noisy_model_input_prev from the teacher
                    target_noise_pred = unet(
                        noisy_model_input_prev,
                        next_timesteps,
                        encoder_hidden_states=prompt_embeds,
                        added_cond_kwargs=encoded_text,
                    ).sample

                    alphas = extract_into_tensor(alpha_schedule, next_timesteps - 1, noisy_model_input_prev.shape)
                    sigmas = extract_into_tensor(sigma_schedule, next_timesteps - 1, noisy_model_input_prev.shape)
                    target_pred_x_0 = (noisy_model_input_prev - sigmas * target_noise_pred) / alphas

                    # 18. Get boundary scalings for start timesteps and apply to predicted x_0
                    c_skip, c_out = scalings_for_boundary_conditions(
                        next_timesteps, timestep_scaling=args.timestep_scaling_factor
                    )
                    c_skip, c_out = [append_dims(x, noisy_model_input_prev.ndim) for x in [c_skip, c_out]]

                    target_pred_x_0 = c_skip * noisy_model_input_prev + c_out * target_pred_x_0

                # 19. Calculate loss
                if args.loss_type == "l2":
                    loss = F.mse_loss(pred_x_0.float(), target_pred_x_0.float(), reduction="mean")
                elif args.loss_type == "huber":
                    loss = torch.mean(
                        torch.sqrt((pred_x_0.float() - target_pred_x_0.float()) ** 2 + args.huber_c**2) - args.huber_c
                    )

                # 20. Backpropagate on the online student model (`unet`) (only LoRA)
                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    grad_norm = accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad(set_to_none=True)

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

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

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

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

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

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

                    if global_step % args.validation_steps == 0:
                        log_validation(
                            vae, args, accelerator, weight_dtype, global_step, unet=unet, is_final_validation=False
                        )

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

            if global_step >= args.max_train_steps:
                break

    # Create the pipeline using using the trained modules and save it.
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        unet = accelerator.unwrap_model(unet)
        unet_lora_state_dict = convert_state_dict_to_diffusers(get_peft_model_state_dict(unet))
        StableDiffusionXLPipeline.save_lora_weights(args.output_dir, unet_lora_layers=unet_lora_state_dict)

        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_*"],
                token=args.hub_token,
            )

        del unet
        torch.cuda.empty_cache()

        # Final inference.
        if args.validation_steps is not None:
            log_validation(vae, args, accelerator, weight_dtype, step=global_step, unet=None, is_final_validation=True)

    accelerator.end_training()


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