distill.py

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

import numpy as np
import torch
import torch.utils.checkpoint
from torch.utils.data import RandomSampler
import transformers
from accelerate import Accelerator
from accelerate.logging import get_logger
from accelerate.utils import ProjectConfiguration, set_seed

from tqdm.auto import tqdm
from transformers import CLIPTokenizer, PretrainedConfig, T5TokenizerFast

import diffusers
from diffusers import (
    AutoencoderKL,
    FlowMatchEulerDiscreteScheduler,
    FluxPipeline,
    FluxTransformer2DModel,
)
from diffusers.optimization import get_scheduler
from diffusers.training_utils import (
    cast_training_params,
    compute_density_for_timestep_sampling,
    compute_loss_weighting_for_sd3,
    free_memory,
)
from diffusers.utils import (
    check_min_version,
    is_wandb_available,
)
from diffusers.utils.torch_utils import is_compiled_module
from safetensors.torch import save_file
from safetensors import safe_open
from dataset import loader
from attention_processor import LocalFlexAttnProcessor, LocalDownsampleFlexAttnProcessor, init_local_mask_flex, init_local_downsample_mask_flex
from attention_processor import attn_outputs_teacher, attn_outputs


if is_wandb_available():
    import wandb

# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
check_min_version("0.31.0.dev0")

logger = get_logger(__name__)


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


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

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

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

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

    del pipeline
    if torch.cuda.is_available():
        torch.cuda.empty_cache()

    return images


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

        return CLIPTextModel
    elif model_class == "T5EncoderModel":
        from transformers import T5EncoderModel

        return T5EncoderModel
    else:
        raise ValueError(f"{model_class} is not supported.")


def parse_args(input_args=None):
    parser = argparse.ArgumentParser(description="Simple example of a training script.")
    parser.add_argument(
        "--pretrained_model_name_or_path",
        type=str,
        default=None,
        required=True,
        help="Path to pretrained model or model identifier from huggingface.co/models.",
    )
    parser.add_argument(
        "--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(
        "--data_root",
        type=str,
        default=None
    )
    
    parser.add_argument(
        "--cache_dir",
        type=str,
        default=None,
        help="The directory where the downloaded models and datasets will be stored.",
    )

    parser.add_argument("--repeats", type=int, default=1, help="How many times to repeat the training data.")

    parser.add_argument(
        "--max_sequence_length",
        type=int,
        default=512,
        help="Maximum sequence length to use with with the T5 text encoder",
    )
    parser.add_argument(
        "--validation_prompt",
        type=str,
        default=None,
        help="A prompt that is used during validation to verify that the model is learning.",
    )
    parser.add_argument(
        "--num_validation_images",
        type=int,
        default=4,
        help="Number of images that should be generated during validation with `validation_prompt`.",
    )
    parser.add_argument(
        "--validation_epochs",
        type=int,
        default=1,
        help=(
            "Run validation every X epochs. Validation consists of running the prompt"
            " `args.validation_prompt` multiple times: `args.num_validation_images`."
        ),
    )
    parser.add_argument(
        "--output_dir",
        type=str,
        default="flux-attn",
        help="The output directory where the model predictions and checkpoints will be written.",
    )
    parser.add_argument("--seed", type=int, default=None, help="A seed for reproducible training.")
    parser.add_argument(
        "--resolution",
        type=int,
        default=1024,
        help=(
            "The resolution for input images"
        )
    )
    parser.add_argument(
        "--center_crop",
        default=False,
        action="store_true",
        help=(
            "Whether to center crop the input images to the resolution. If not set, the images will be randomly"
            " cropped. The images will be resized to the resolution first before cropping."
        ),
    )
    parser.add_argument(
        "--train_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(
        "--window_size",
        type=int,
        default=16,
        help=(
            "Size of local window for attention sampling."
        ),
    )
    parser.add_argument(
        "--down_factor",
        type=int,
        default=1,
        help=(
            "Factor of downsampling for key-value tokens."
        ),
    )
    parser.add_argument(
        "--checkpointing_steps",
        type=int,
        default=10000,
        help=(
            "Save a checkpoint of the training state every X updates. These checkpoints can be used both as final"
            " checkpoints in case they are better than the last checkpoint, and are also suitable for resuming"
            " training using `--resume_from_checkpoint`."
        ),
    )
    parser.add_argument(
        "--checkpoints_total_limit",
        type=int,
        default=None,
        help=("Max number of checkpoints to store."),
    )
    parser.add_argument(
        "--resume_from_checkpoint",
        type=str,
        default=None,
        help=(
            "Whether training should be resumed from a previous checkpoint. Use a path saved by"
            ' `--checkpointing_steps`, or `"latest"` to automatically select the last available checkpoint.'
        ),
    )
    parser.add_argument(
        "--gradient_accumulation_steps",
        type=int,
        default=1,
        help="Number of updates steps to accumulate before performing a backward/update pass.",
    )
    parser.add_argument(
        "--gradient_checkpointing",
        action="store_true",
        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
    )
    parser.add_argument(
        "--learning_rate",
        type=float,
        default=1e-4,
        help="Initial learning rate (after the potential warmup period) to use.",
    )

    parser.add_argument(
        "--guidance_scale",
        type=float,
        default=3.5,
        help="the FLUX.1 dev variant is a guidance distilled model",
    )

    parser.add_argument(
        "--scale_lr",
        action="store_true",
        default=False,
        help="Scale the learning rate by the number of GPUs, gradient accumulation steps, and batch size.",
    )
    parser.add_argument(
        "--lr_scheduler",
        type=str,
        default="constant",
        help=(
            'The scheduler type to use. Choose between ["linear", "cosine", "cosine_with_restarts", "polynomial",'
            ' "constant", "constant_with_warmup"]'
        ),
    )
    parser.add_argument(
        "--lr_warmup_steps", type=int, default=500, help="Number of steps for the warmup in the lr scheduler."
    )
    parser.add_argument(
        "--lr_num_cycles",
        type=int,
        default=1,
        help="Number of hard resets of the lr in cosine_with_restarts scheduler.",
    )
    parser.add_argument("--lr_power", type=float, default=1.0, help="Power factor of the polynomial scheduler.")
    parser.add_argument(
        "--dataloader_num_workers",
        type=int,
        default=0,
        help=(
            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
        ),
    )
    parser.add_argument(
        "--weighting_scheme",
        type=str,
        default="none",
        choices=["sigma_sqrt", "logit_normal", "mode", "cosmap", "none"],
        help=('We default to the "none" weighting scheme for uniform sampling and uniform loss'),
    )
    parser.add_argument(
        "--logit_mean", type=float, default=0.0, help="mean to use when using the `'logit_normal'` weighting scheme."
    )
    parser.add_argument(
        "--logit_std", type=float, default=1.0, help="std to use when using the `'logit_normal'` weighting scheme."
    )
    parser.add_argument(
        "--mode_scale",
        type=float,
        default=1.29,
        help="Scale of mode weighting scheme. Only effective when using the `'mode'` as the `weighting_scheme`.",
    )
    parser.add_argument(
        "--optimizer",
        type=str,
        default="AdamW",
        help=('The optimizer type to use. Choose between ["AdamW", "prodigy"]'),
    )

    parser.add_argument(
        "--use_8bit_adam",
        action="store_true",
        help="Whether or not to use 8-bit Adam from bitsandbytes. Ignored if optimizer is not set to AdamW",
    )

    parser.add_argument(
        "--adam_beta1", type=float, default=0.9, help="The beta1 parameter for the Adam and Prodigy optimizers."
    )
    parser.add_argument(
        "--adam_beta2", type=float, default=0.999, help="The beta2 parameter for the Adam and Prodigy optimizers."
    )
    parser.add_argument(
        "--prodigy_beta3",
        type=float,
        default=None,
        help="coefficients for computing the Prodigy stepsize using running averages. If set to None, "
        "uses the value of square root of beta2. Ignored if optimizer is adamW",
    )
    parser.add_argument("--prodigy_decouple", type=bool, default=True, help="Use AdamW style decoupled weight decay")
    parser.add_argument("--adam_weight_decay", type=float, default=1e-04, help="Weight decay to use for unet params")
    
    parser.add_argument(
        "--adam_epsilon",
        type=float,
        default=1e-08,
        help="Epsilon value for the Adam optimizer and Prodigy optimizers.",
    )

    parser.add_argument(
        "--prodigy_use_bias_correction",
        type=bool,
        default=True,
        help="Turn on Adam's bias correction. True by default. Ignored if optimizer is adamW",
    )
    parser.add_argument(
        "--prodigy_safeguard_warmup",
        type=bool,
        default=True,
        help="Remove lr from the denominator of D estimate to avoid issues during warm-up stage. True by default. "
        "Ignored if optimizer is adamW",
    )
    parser.add_argument("--max_grad_norm", default=1.0, type=float, help="Max gradient norm.")
    parser.add_argument(
        "--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(
        "--use_cached_latent",
        action="store_true",
        default=False,
        help="Cache the VAE latents",
    )
    parser.add_argument(
        "--use_cached_prompt_embed",
        action="store_true",
        default=False,
        help="Use cached T5 and CLIP features",
    )
    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("--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()

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

    return args


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


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

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

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

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

    _, seq_len, _ = prompt_embeds.shape

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

    return prompt_embeds


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

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

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

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

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

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

    return prompt_embeds


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

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

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

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


def main(args):
    if 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)
    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config
    )

    # 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)

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

    # Load the tokenizers
    tokenizer_one = CLIPTokenizer.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer",
        revision=args.revision,
    )
    tokenizer_two = T5TokenizerFast.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="tokenizer_2",
        revision=args.revision,
    )

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

    # Load scheduler and models
    noise_scheduler = FlowMatchEulerDiscreteScheduler.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="scheduler"
    )
    text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
    vae = AutoencoderKL.from_pretrained(
        args.pretrained_model_name_or_path,
        subfolder="vae",
        revision=args.revision,
        variant=args.variant,
    )
    transformer = FluxTransformer2DModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="transformer", revision=args.revision, variant=args.variant
    )
    transformer_teacher = FluxTransformer2DModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="transformer", revision=args.revision, variant=args.variant
    )

    if args.down_factor == 1:
        init_local_mask_flex(args.resolution // 16, args.resolution // 16, text_length=args.max_sequence_length, window_size=args.window_size, device=accelerator.device)
        attn_processors = {}
        for idx, name in enumerate(pipe.transformer.attn_processors):
            attn_processors[k] = LocalFlexAttnProcessor(distill='single' in name and idx % 4 == 0)
    else:
        init_local_downsample_mask_flex(args.resolution // 16, args.resolution // 16, text_length=args.max_sequence_length, window_size=args.window_size, 
                                        down_factor=args.down_factor, device=device)
        attn_processors = {}
        for idx, name in enumerate(pipe.transformer.attn_processors):
            attn_processors[k] = LocalDownsampleFlexAttnProcessor(down_factor=args.down_factor, distill='single' in name and idx % 4 == 0)
    attn_processors_teacher = {}
    for idx, name in enumerate(transformer_teacher.attn_processors.keys()):
        attn_processors_teacher[name] = FluxAttnProcessor2_0(distill='single' in name and idx % 4 == 0)
    transformer.set_attn_processor(attn_processors)
    transformer_teacher.set_attn_processor(attn_processors_teacher)

    # We only train the Attn layers
    transformer.requires_grad_(False)
    transformer_teacher.requires_grad_(False)
    vae.requires_grad_(False)
    text_encoder_one.requires_grad_(False)
    text_encoder_two.requires_grad_(False)

    # For mixed precision training we cast all non-trainable weights (vae, text_encoder and transformer) to half-precision
    # as these weights are only used for inference, keeping weights in full precision is not required.
    weight_dtype = torch.float32
    if accelerator.mixed_precision == "fp16":
        weight_dtype = torch.float16
    elif accelerator.mixed_precision == "bf16":
        weight_dtype = torch.bfloat16
        
    if torch.backends.mps.is_available() and weight_dtype == torch.bfloat16:
        # due to pytorch#99272, MPS does not yet support bfloat16.
        raise ValueError(
            "Mixed precision training with bfloat16 is not supported on MPS. Please use fp16 (recommended) or fp32 instead."
        )

    vae.to(accelerator.device, dtype=weight_dtype)
    transformer.to(accelerator.device, dtype=weight_dtype)
    transformer_teacher.to(accelerator.device, dtype=weight_dtype)
    text_encoder_one.to(accelerator.device, dtype=weight_dtype)
    text_encoder_two.to(accelerator.device, dtype=weight_dtype)
    
    if args.gradient_checkpointing:
        transformer.enable_gradient_checkpointing()
        
    for _name, _param in transformer.named_parameters():
        if '.attn.to_q.' in _name or '.attn.to_k.' in _name or '.attn.to_v.' in _name or '.attn.to_out.' in _name or 'spatial_weight' in _name:
            _param.requires_grad = True
    
    def unwrap_model(model):
        model = accelerator.unwrap_model(model)
        model = model._orig_mod if is_compiled_module(model) else model
        return model

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

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

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

    # Optimization parameters
    transformer_attn_parameters = list(filter(lambda p: p.requires_grad, transformer.parameters()))
    transformer_parameters_with_lr = {"params": transformer_attn_parameters, "lr": args.learning_rate}
    
    params_to_optimize = [transformer_parameters_with_lr]

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

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

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

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

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

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

        optimizer_class = prodigyopt.Prodigy

        if args.learning_rate <= 0.1:
            logger.warning(
                "Learning rate is too low. When using prodigy, it's generally better to set learning rate around 1.0"
            )
        
        optimizer = optimizer_class(
            params_to_optimize,
            lr=args.learning_rate,
            betas=(args.adam_beta1, args.adam_beta2),
            beta3=args.prodigy_beta3,
            weight_decay=args.adam_weight_decay,
            eps=args.adam_epsilon,
            decouple=args.prodigy_decouple,
            use_bias_correction=args.prodigy_use_bias_correction,
            safeguard_warmup=args.prodigy_safeguard_warmup,
        )

    # Dataset and DataLoaders creation:
    train_dataloader = loader(train_batch_size=args.train_batch_size, num_workers=args.dataloader_num_workers,
                              img_dir=args.data_root, img_size=args.resolution, 
                              use_cached_prompt_embeds=args.use_cached_prompt_embed,
                              use_cached_latent_codes=args.use_cached_latent)

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

    def compute_text_embeddings(prompt, text_encoders, tokenizers):
        with torch.no_grad():
            prompt_embeds, pooled_prompt_embeds, text_ids = encode_prompt(
                text_encoders, tokenizers, prompt, args.max_sequence_length
            )
            prompt_embeds = prompt_embeds.to(accelerator.device)
            pooled_prompt_embeds = pooled_prompt_embeds.to(accelerator.device)
            text_ids = text_ids.to(accelerator.device)
        return prompt_embeds, pooled_prompt_embeds, text_ids
   
    # Clear the memory here
    if args.use_cached_prompt_embed:
        del text_encoder_one, text_encoder_two, tokenizer_one, tokenizer_two, text_encoders, tokenizers
        free_memory()
        
    vae_config_shift_factor = vae.config.shift_factor
    vae_config_scaling_factor = vae.config.scaling_factor
    vae_config_block_out_channels = vae.config.block_out_channels
    if args.use_cached_latent:
        del vae
        free_memory()

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

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

    # Prepare everything with our `accelerator`.
    guidance_embeds = transformer.config.guidance_embeds
    
    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 = "flux-dev-attn"
        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 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 = 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(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.sigmas.to(device=accelerator.device, dtype=dtype)
        schedule_timesteps = noise_scheduler.timesteps.to(accelerator.device)
        timesteps = timesteps.to(accelerator.device)
        step_indices = [(schedule_timesteps == t).nonzero().item() for t in timesteps]

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

    for epoch in range(first_epoch, args.num_train_epochs):
        transformer.train()
        
        for _, batch in enumerate(train_dataloader):
            with accelerator.accumulate(transformer):

                if not args.use_cached_prompt_embed:
                    prompt_embeds, pooled_prompt_embeds, text_ids = compute_text_embeddings(
                        batch['prompts'], text_encoders, tokenizers
                    )
                else:
                    prompt_embeds = batch['prompt_embeds_t5'].to(dtype=weight_dtype)
                    pooled_prompt_embeds = batch['prompt_embeds_clip'].to(dtype=weight_dtype)
                    text_ids = torch.zeros(prompt_embeds.shape[1], 3).to(device=accelerator.device, dtype=weight_dtype)

                with torch.no_grad():
                    # Convert images to latent space
                    if args.use_cached_latent:
                        mean = batch['latent_codes_mean'].to(dtype=weight_dtype)
                        std = batch['latent_codes_std'].to(dtype=weight_dtype)
                        sample = torch.randn_like(mean)
                        model_input = mean + std * sample
                    else:
                        model_input = vae.encode(batch[0].to(dtype=vae.dtype)).latent_dist.sample()
                    model_input = (model_input - vae_config_shift_factor) * vae_config_scaling_factor
                    
                    vae_scale_factor = 2 ** (len(vae_config_block_out_channels))

                    latent_image_ids = FluxPipeline._prepare_latent_image_ids(
                        model_input.shape[0],
                        model_input.shape[2],
                        model_input.shape[3],
                        accelerator.device,
                        weight_dtype,
                    )
                    # Sample noise that we'll add to the latents
                    noise = torch.randn_like(model_input)
                    bsz = model_input.shape[0]

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

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

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

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

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

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

                # these weighting schemes use a uniform timestep sampling
                # and instead post-weight the loss
                weighting = compute_loss_weighting_for_sd3(weighting_scheme=args.weighting_scheme, sigmas=sigmas)

                # flow matching loss
                target = noise - model_input

                # Compute regular loss.
                loss_fm = (weighting.float() * (model_pred.float() - target.float()) ** 2).mean()
                loss_distill = (weighting.float() * (model_pred.float() - teacher_pred.float()) ** 2).mean()
                loss_attn = sum([(weighting.float().squeeze(-1) * (attn_output.float() - attn_output_teacher.float()) ** 2).mean()
                    for attn_output, attn_output_teacher in zip(attn_outputs, attn_outputs_teacher)]) / len(attn_outputs)
                loss = loss_fm + loss_distill * 0.5 + loss_attn * 0.5
                
                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()
                attn_outputs.clear()
                attn_outputs_teacher.clear()

            # 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)

                if global_step % args.checkpointing_steps == 0:
                    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_fm": loss_fm.detach().item(), 
                    "loss_distill": loss_distill.detach().item(), 
                    "loss_attn": loss_attn.detach().item(), 
                    "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

        free_memory()
        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                # create pipeline
                if args.use_cached_prompt_embed:
                    text_encoder_one, text_encoder_two = load_text_encoders(text_encoder_cls_one, text_encoder_cls_two)
                if args.use_cached_latent:
                    vae = AutoencoderKL.from_pretrained(
                        args.pretrained_model_name_or_path,
                        subfolder="vae",
                        revision=args.revision,
                        variant=args.variant,
                    )
                pipeline = FluxPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    vae=vae,
                    text_encoder=accelerator.unwrap_model(text_encoder_one),
                    text_encoder_2=accelerator.unwrap_model(text_encoder_two),
                    transformer=accelerator.unwrap_model(transformer),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
                )
                pipeline_args = {"prompt": args.validation_prompt}
                images = log_validation(
                    pipeline=pipeline,
                    args=args,
                    accelerator=accelerator,
                    pipeline_args=pipeline_args,
                    epoch=epoch,
                    torch_dtype=weight_dtype,
                )
                if args.use_cached_prompt_embed:
                    del text_encoder_one, text_encoder_two
                if args.use_cached_latent:
                    del vae
                free_memory()

    # Save the attn weights
    accelerator.wait_for_everyone()
    if accelerator.is_main_process:
        transformer = unwrap_model(transformer)
        if args.upcast_before_saving:
            transformer.to(torch.float32)
        else:
            transformer = transformer.to(weight_dtype)

        state_dict = {}
        for _name, _param in transformer.named_parameters():
            if '.attn.to_q.' in _name or '.attn.to_k.' in _name or '.attn.to_v.' in _name or '.attn.to_out.' in _name or 'spatial_weight' in _name:
                state_dict[_name] = _param

        save_file(
            state_dict,
            os.path.join(args.output_dir, 'attn_weights.safetensors')
        )

        # Final inference
        # Load previous pipeline
        free_memory()
        pipeline = FluxPipeline.from_pretrained(
            args.pretrained_model_name_or_path,
            revision=args.revision,
            variant=args.variant,
            torch_dtype=weight_dtype,
        )
        # load attention processors
        state_dict = {}
        with safe_open(os.path.join(args.output_dir, 'attn_weights.safetensor'), framework="pt") as f:
            for k in f.keys():
                state_dict[k] = f.get_tensor(k)

        missing_keys, unexpected_keys = pipeline.transformer.load_state_dict(state_dict)

        missing_keys = list(filter(lambda p: ('.attn.to_q.' in p or 
                                              '.attn.to_k.' in p or 
                                              '.attn.to_v.' in p or 
                                              '.attn.to_out.' in p or 
                                              'spatial_weight' in p), missing_keys))

        if len(missing_keys) != 0 or len(unexpected_keys) != 0:
            logger.warning(
                f"Loading attn weights from state_dict led to unexpected keys: {unexpected_keys}"
                f" and missing keys: {missing_keys}."
            )

        # run inference
        images = []
        if args.validation_prompt and args.num_validation_images > 0:
            pipeline_args = {"prompt": args.validation_prompt}
            images = log_validation(
                pipeline=pipeline,
                args=args,
                accelerator=accelerator,
                pipeline_args=pipeline_args,
                epoch=epoch,
                is_final_validation=True,
                torch_dtype=weight_dtype,
            )

            for idx, image in enumerate(images):
                image.save(os.path.join(args.output_dir, 'validation_idx_%02d.jpg' % idx))

    accelerator.end_training()


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