def main()

in threestudio/scripts/train_dreambooth_lora.py [0:0]


def main(args):
    logging_dir = Path(args.output_dir, args.logging_dir)

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

    accelerator = Accelerator(
        gradient_accumulation_steps=args.gradient_accumulation_steps,
        mixed_precision=args.mixed_precision,
        log_with=args.report_to,
        project_config=accelerator_project_config,
    )

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

    # Currently, it's not possible to do gradient accumulation when training two models with accelerate.accumulate
    # This will be enabled soon in accelerate. For now, we don't allow gradient accumulation when training two models.
    # TODO (sayakpaul): Remove this check when gradient accumulation with two models is enabled in accelerate.
    if args.train_text_encoder and args.gradient_accumulation_steps > 1 and accelerator.num_processes > 1:
        raise ValueError(
            "Gradient accumulation is not supported when training the text encoder in distributed training. "
            "Please set gradient_accumulation_steps to 1. This feature will be supported in the future."
        )

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

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

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

        if cur_class_images < args.num_class_images:
            torch_dtype = torch.float16 if accelerator.device.type == "cuda" else torch.float32
            if args.prior_generation_precision == "fp32":
                torch_dtype = torch.float32
            elif args.prior_generation_precision == "fp16":
                torch_dtype = torch.float16
            elif args.prior_generation_precision == "bf16":
                torch_dtype = torch.bfloat16
            pipeline = DiffusionPipeline.from_pretrained(
                args.pretrained_model_name_or_path,
                torch_dtype=torch_dtype,
                safety_checker=None,
                revision=args.revision,
                variant=args.variant,
            )
            pipeline.set_progress_bar_config(disable=True)

            num_new_images = args.num_class_images - cur_class_images
            logger.info(f"Number of class images to sample: {num_new_images}.")

            sample_dataset = PromptDataset(args.class_prompt, num_new_images)
            sample_dataloader = torch.utils.data.DataLoader(sample_dataset, batch_size=args.sample_batch_size)

            sample_dataloader = accelerator.prepare(sample_dataloader)
            pipeline.to(accelerator.device)

            for example in tqdm(
                sample_dataloader, desc="Generating class images", disable=not accelerator.is_local_main_process
            ):
                images = pipeline(example["prompt"]).images

                for i, image in enumerate(images):
                    hash_image = hashlib.sha1(image.tobytes()).hexdigest()
                    image_filename = class_images_dir / f"{example['index'][i] + cur_class_images}-{hash_image}.jpg"
                    image.save(image_filename)

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

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

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

    # Load the tokenizer
    if args.tokenizer_name:
        tokenizer = AutoTokenizer.from_pretrained(args.tokenizer_name, revision=args.revision, use_fast=False)
    elif args.pretrained_model_name_or_path:
        tokenizer = AutoTokenizer.from_pretrained(
            args.pretrained_model_name_or_path,
            subfolder="tokenizer",
            revision=args.revision,
            use_fast=False,
            cache_dir=CACHE_DIR,
            local_files_only=True,
        )

    # import correct text encoder class
    text_encoder_cls = import_model_class_from_model_name_or_path(args.pretrained_model_name_or_path, args.revision)

    # Load scheduler and models
    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, subfolder="scheduler", cache_dir=CACHE_DIR, local_files_only=True)
    text_encoder = text_encoder_cls.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision, variant=args.variant, cache_dir=CACHE_DIR, local_files_only=True,
    )
    try:
        vae = AutoencoderKL.from_pretrained(
            args.pretrained_model_name_or_path, subfolder="vae", revision=args.revision, variant=args.variant, cache_dir=CACHE_DIR, local_files_only=True,
        )
    except OSError:
        # IF does not have a VAE so let's just set it to None
        # We don't have to error out here
        vae = None

    unet = UNet2DConditionModel.from_pretrained(
        args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision, variant=args.variant, cache_dir=CACHE_DIR, local_files_only=True,
    )

    # We only train the additional adapter LoRA layers
    if vae is not None:
        vae.requires_grad_(False)
    text_encoder.requires_grad_(False)
    unet.requires_grad_(False)

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

    # Move unet, vae and text_encoder to device and cast to weight_dtype
    unet.to(accelerator.device, dtype=weight_dtype)
    if vae is not None:
        vae.to(accelerator.device, dtype=weight_dtype)
    text_encoder.to(accelerator.device, dtype=weight_dtype)

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

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

    if args.gradient_checkpointing:
        unet.enable_gradient_checkpointing()
        if args.train_text_encoder:
            text_encoder.gradient_checkpointing_enable()

    # now we will add new LoRA weights to the attention layers
    # It's important to realize here how many attention weights will be added and of which sizes
    # The sizes of the attention layers consist only of two different variables:
    # 1) - the "hidden_size", which is increased according to `unet.config.block_out_channels`.
    # 2) - the "cross attention size", which is set to `unet.config.cross_attention_dim`.

    # Let's first see how many attention processors we will have to set.
    # For Stable Diffusion, it should be equal to:
    # - down blocks (2x attention layers) * (2x transformer layers) * (3x down blocks) = 12
    # - mid blocks (2x attention layers) * (1x transformer layers) * (1x mid blocks) = 2
    # - up blocks (2x attention layers) * (3x transformer layers) * (3x up blocks) = 18
    # => 32 layers

    # Set correct lora layers
    unet_lora_parameters = []
    for attn_processor_name, attn_processor in unet.attn_processors.items():
        # Parse the attention module.
        attn_module = unet
        for n in attn_processor_name.split(".")[:-1]:
            attn_module = getattr(attn_module, n)

        # Set the `lora_layer` attribute of the attention-related matrices.
        attn_module.to_q.set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_q.in_features, out_features=attn_module.to_q.out_features, rank=args.rank
            )
        )
        attn_module.to_k.set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_k.in_features, out_features=attn_module.to_k.out_features, rank=args.rank
            )
        )
        attn_module.to_v.set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_v.in_features, out_features=attn_module.to_v.out_features, rank=args.rank
            )
        )
        attn_module.to_out[0].set_lora_layer(
            LoRALinearLayer(
                in_features=attn_module.to_out[0].in_features,
                out_features=attn_module.to_out[0].out_features,
                rank=args.rank,
            )
        )

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

        if isinstance(attn_processor, (AttnAddedKVProcessor, SlicedAttnAddedKVProcessor, AttnAddedKVProcessor2_0)):
            attn_module.add_k_proj.set_lora_layer(
                LoRALinearLayer(
                    in_features=attn_module.add_k_proj.in_features,
                    out_features=attn_module.add_k_proj.out_features,
                    rank=args.rank,
                )
            )
            attn_module.add_v_proj.set_lora_layer(
                LoRALinearLayer(
                    in_features=attn_module.add_v_proj.in_features,
                    out_features=attn_module.add_v_proj.out_features,
                    rank=args.rank,
                )
            )
            unet_lora_parameters.extend(attn_module.add_k_proj.lora_layer.parameters())
            unet_lora_parameters.extend(attn_module.add_v_proj.lora_layer.parameters())

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

    # create custom saving & loading hooks so that `accelerator.save_state(...)` serializes in a nice format
    def save_model_hook(models, weights, output_dir):
        if accelerator.is_main_process:
            # there are only two options here. Either are just the unet attn processor layers
            # or there are the unet and text encoder atten layers
            unet_lora_layers_to_save = None
            text_encoder_lora_layers_to_save = None

            for model in models:
                if isinstance(model, type(accelerator.unwrap_model(unet))):
                    unet_lora_layers_to_save = unet_lora_state_dict(model)
                elif isinstance(model, type(accelerator.unwrap_model(text_encoder))):
                    text_encoder_lora_layers_to_save = text_encoder_lora_state_dict(model)
                else:
                    raise ValueError(f"unexpected save model: {model.__class__}")

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

            LoraLoaderMixin.save_lora_weights(
                output_dir,
                unet_lora_layers=unet_lora_layers_to_save,
                text_encoder_lora_layers=text_encoder_lora_layers_to_save,
            )

    def load_model_hook(models, input_dir):
        unet_ = None
        text_encoder_ = None

        while len(models) > 0:
            model = models.pop()

            if isinstance(model, type(accelerator.unwrap_model(unet))):
                unet_ = model
            elif isinstance(model, type(accelerator.unwrap_model(text_encoder))):
                text_encoder_ = model
            else:
                raise ValueError(f"unexpected save model: {model.__class__}")

        lora_state_dict, network_alphas = LoraLoaderMixin.lora_state_dict(input_dir)
        LoraLoaderMixin.load_lora_into_unet(lora_state_dict, network_alphas=network_alphas, unet=unet_)
        LoraLoaderMixin.load_lora_into_text_encoder(
            lora_state_dict, network_alphas=network_alphas, text_encoder=text_encoder_
        )

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

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

    # 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

    # Optimizer creation
    params_to_optimize = (
        itertools.chain(unet_lora_parameters, text_lora_parameters)
        if args.train_text_encoder
        else unet_lora_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,
    )

    if args.pre_compute_text_embeddings:

        def compute_text_embeddings(prompt):
            with torch.no_grad():
                text_inputs = tokenize_prompt(tokenizer, prompt, tokenizer_max_length=args.tokenizer_max_length)
                prompt_embeds = encode_prompt(
                    text_encoder,
                    text_inputs.input_ids,
                    text_inputs.attention_mask,
                    text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
                )

            return prompt_embeds

        pre_computed_encoder_hidden_states = compute_text_embeddings(args.instance_prompt)
        validation_prompt_negative_prompt_embeds = compute_text_embeddings("")

        if args.validation_prompt is not None:
            validation_prompt_encoder_hidden_states = compute_text_embeddings(args.validation_prompt)
        else:
            validation_prompt_encoder_hidden_states = None

        if args.class_prompt is not None:
            pre_computed_class_prompt_encoder_hidden_states = compute_text_embeddings(args.class_prompt)
        else:
            pre_computed_class_prompt_encoder_hidden_states = None

        text_encoder = None
        tokenizer = None

        gc.collect()
        torch.cuda.empty_cache()
    else:
        pre_computed_encoder_hidden_states = None
        validation_prompt_encoder_hidden_states = None
        validation_prompt_negative_prompt_embeds = None
        pre_computed_class_prompt_encoder_hidden_states = None

    # Dataset and DataLoaders creation:
    train_dataset = DreamBoothDataset(
        instance_data_root=args.instance_data_dir,
        instance_prompt=args.instance_prompt,
        class_data_root=args.class_data_dir if args.with_prior_preservation else None,
        class_prompt=args.class_prompt,
        class_num=args.num_class_images,
        tokenizer=tokenizer,
        size=args.resolution,
        center_crop=args.center_crop,
        encoder_hidden_states=pre_computed_encoder_hidden_states,
        class_prompt_encoder_hidden_states=pre_computed_class_prompt_encoder_hidden_states,
        tokenizer_max_length=args.tokenizer_max_length,
        use_view_dependent_prompt=args.use_view_dependent_prompt,
    )

    train_dataloader = torch.utils.data.DataLoader(
        train_dataset,
        batch_size=args.train_batch_size,
        shuffle=True,
        collate_fn=lambda examples: collate_fn(examples, args.with_prior_preservation),
        num_workers=args.dataloader_num_workers,
    )

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

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

    # Prepare everything with our `accelerator`.
    if args.train_text_encoder:
        unet, text_encoder, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, text_encoder, optimizer, train_dataloader, lr_scheduler
        )
    else:
        unet, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
            unet, optimizer, train_dataloader, lr_scheduler
        )

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

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

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

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

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

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

            initial_global_step = global_step
            first_epoch = global_step // num_update_steps_per_epoch
    else:
        initial_global_step = 0

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

    for epoch in range(first_epoch, args.num_train_epochs):
        unet.train()
        if args.train_text_encoder:
            text_encoder.train()
        for step, batch in enumerate(train_dataloader):
            with accelerator.accumulate(unet):
                pixel_values = batch["pixel_values"].to(dtype=weight_dtype)

                if vae is not None:
                    # Convert images to latent space
                    model_input = vae.encode(pixel_values).latent_dist.sample()
                    model_input = model_input * vae.config.scaling_factor
                else:
                    model_input = pixel_values

                # Sample noise that we'll add to the latents
                noise = torch.randn_like(model_input)
                bsz, channels, height, width = model_input.shape
                # Sample a random timestep for each image
                timesteps = torch.randint(
                    0, noise_scheduler.config.num_train_timesteps, (bsz,), device=model_input.device
                )
                timesteps = timesteps.long()

                # Add noise to the model input according to the noise magnitude at each timestep
                # (this is the forward diffusion process)
                noisy_model_input = noise_scheduler.add_noise(model_input, noise, timesteps)

                # Get the text embedding for conditioning
                if args.pre_compute_text_embeddings:
                    encoder_hidden_states = batch["input_ids"]
                else:
                    encoder_hidden_states = encode_prompt(
                        text_encoder,
                        batch["input_ids"],
                        batch["attention_mask"],
                        text_encoder_use_attention_mask=args.text_encoder_use_attention_mask,
                    )

                if accelerator.unwrap_model(unet).config.in_channels == channels * 2:
                    noisy_model_input = torch.cat([noisy_model_input, noisy_model_input], dim=1)

                if args.class_labels_conditioning == "timesteps":
                    class_labels = timesteps
                else:
                    class_labels = None

                # Predict the noise residual
                model_pred = unet(
                    noisy_model_input, timesteps, encoder_hidden_states, class_labels=class_labels
                ).sample

                # if model predicts variance, throw away the prediction. we will only train on the
                # simplified training objective. This means that all schedulers using the fine tuned
                # model must be configured to use one of the fixed variance variance types.
                if model_pred.shape[1] == 6:
                    model_pred, _ = torch.chunk(model_pred, 2, dim=1)

                # Get the target for loss depending on the prediction type
                if noise_scheduler.config.prediction_type == "epsilon":
                    target = noise
                elif noise_scheduler.config.prediction_type == "v_prediction":
                    target = noise_scheduler.get_velocity(model_input, noise, timesteps)
                else:
                    raise ValueError(f"Unknown prediction type {noise_scheduler.config.prediction_type}")

                if args.with_prior_preservation:
                    # Chunk the noise and model_pred into two parts and compute the loss on each part separately.
                    model_pred, model_pred_prior = torch.chunk(model_pred, 2, dim=0)
                    target, target_prior = torch.chunk(target, 2, dim=0)

                    # Compute instance loss
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")

                    # Compute prior loss
                    prior_loss = F.mse_loss(model_pred_prior.float(), target_prior.float(), reduction="mean")

                    # Add the prior loss to the instance loss.
                    loss = loss + args.prior_loss_weight * prior_loss
                else:
                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")

                accelerator.backward(loss)
                if accelerator.sync_gradients:
                    params_to_clip = (
                        itertools.chain(unet_lora_parameters, text_lora_parameters)
                        if args.train_text_encoder
                        else unet_lora_parameters
                    )
                    accelerator.clip_grad_norm_(params_to_clip, args.max_grad_norm)
                optimizer.step()
                lr_scheduler.step()
                optimizer.zero_grad()

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

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

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

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

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

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

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

            if global_step >= args.max_train_steps:
                break

        if accelerator.is_main_process:
            if args.validation_prompt is not None and epoch % args.validation_epochs == 0:
                logger.info(
                    f"Running validation... \n Generating {args.num_validation_images} images with prompt:"
                    f" {args.validation_prompt}."
                )
                # create pipeline
                pipeline = DiffusionPipeline.from_pretrained(
                    args.pretrained_model_name_or_path,
                    unet=accelerator.unwrap_model(unet),
                    text_encoder=None if args.pre_compute_text_embeddings else accelerator.unwrap_model(text_encoder),
                    revision=args.revision,
                    variant=args.variant,
                    torch_dtype=weight_dtype,
                    safety_checker=None,
                    cache_dir=CACHE_DIR,
                    local_files_only=True,
                )

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

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

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

                    scheduler_args["variance_type"] = variance_type

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

                pipeline = pipeline.to(accelerator.device)
                pipeline.set_progress_bar_config(disable=True)

                # run inference
                generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
                if args.pre_compute_text_embeddings:
                    pipeline_args = {
                        "prompt_embeds": validation_prompt_encoder_hidden_states,
                        "negative_prompt_embeds": validation_prompt_negative_prompt_embeds,
                    }
                else:
                    pipeline_args = {"prompt": args.validation_prompt}

                if args.validation_images is None:
                    images = []
                    for _ in range(args.num_validation_images):
                        with torch.cuda.amp.autocast():
                            image = pipeline(**pipeline_args, generator=generator).images[0]
                            images.append(image)
                else:
                    images = []
                    for image in args.validation_images:
                        image = Image.open(image)
                        with torch.cuda.amp.autocast():
                            image = pipeline(**pipeline_args, image=image, generator=generator).images[0]
                        images.append(image)

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

                del pipeline
                torch.cuda.empty_cache()

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

        if text_encoder is not None and args.train_text_encoder:
            text_encoder = accelerator.unwrap_model(text_encoder)
            text_encoder = text_encoder.to(torch.float32)
            text_encoder_lora_layers = text_encoder_lora_state_dict(text_encoder)
        else:
            text_encoder_lora_layers = None

        LoraLoaderMixin.save_lora_weights(
            save_directory=args.output_dir,
            unet_lora_layers=unet_lora_layers,
            text_encoder_lora_layers=text_encoder_lora_layers,
        )

        # Final inference
        # Load previous pipeline
        pipeline = DiffusionPipeline.from_pretrained(
            args.pretrained_model_name_or_path, revision=args.revision, variant=args.variant, torch_dtype=weight_dtype,
            cache_dir=CACHE_DIR, local_files_only=True, safety_checker=None,
        )

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

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

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

            scheduler_args["variance_type"] = variance_type

        pipeline.scheduler = DPMSolverMultistepScheduler.from_config(pipeline.scheduler.config, **scheduler_args, cache_dir=CACHE_DIR, local_files_only=True,)

        pipeline = pipeline.to(accelerator.device)

        # load attention processors
        pipeline.load_lora_weights(args.output_dir, weight_name="pytorch_lora_weights.safetensors")

        # run inference
        images = []
        if args.validation_prompt and args.num_validation_images > 0:
            generator = torch.Generator(device=accelerator.device).manual_seed(args.seed) if args.seed else None
            images = [
                pipeline(args.validation_prompt, num_inference_steps=25, generator=generator).images[0]
                for _ in range(args.num_validation_images)
            ]

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

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

    accelerator.end_training()