Upload folder using huggingface_hub

MMP_Diffusion_Lora_config.py

@@ -0,0 +1,368 @@
+import os
+import argparse
+from transformers import PretrainedConfig
+
+
+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.")
+    parser.add_argument(
+        "--input_perturbation", type=float, default=0, help="The scale of input perturbation. Recommended 0.1."
+    )
+    parser.add_argument(
+        "--pretrained_model_name_or_path",
+        type=str,
+        default='/cpfs04/user/liudawei/jgl/projects/download_models/stable-diffusion/SDXL',
+        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(
+        "--dataset_name",
+        type=str,
+        default='custom',
+        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_path",
+        type=str,
+        default='/cpfs02/shared/llmit6/liudawei/jgl/EmotionDPO/data/ETI_emotion.parquet',
+        help=(
+            "The path of ETI dataset"
+        ),
+    )
+    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(
+        "--visual_prompts_dir",
+        type=str,
+        default='features/origin',
+        help="Path to initial visual prompts",
+    )
+    parser.add_argument(
+        "--prompt_len",
+        type=int,
+        default=16,
+        help="visual prompts length",
+    )
+    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="caption",
+        help="The column of the dataset containing a caption or a list of captions.",
+    )
+    parser.add_argument(
+        "--max_train_samples",
+        type=int,
+        default=None,
+        # default=256,
+        help=(
+            "For debugging purposes or quicker training, truncate the number of training examples to this "
+            "value if set."
+        ),
+    )
+    parser.add_argument(
+        "--cache_dir",
+        type=str,
+        default='/cpfs04/user/liudawei/jgl/projects/download_models',
+        help="The directory where the downloaded models and datasets will be stored.",
+    )
+    parser.add_argument("--seed", 
+        type=int, 
+        # default=None,
+        default=42,
+        # was random for submission, need to test that not distributing same noise etc across devices
+        help="A seed for reproducible training."
+    )
+    parser.add_argument(
+        "--resolution",
+        type=int,
+        default=512,
+        help=(
+            "The resolution for input images, all the images in the dataset will be resized to this"
+            " resolution"
+        ),
+    )
+    parser.add_argument(
+        "--random_crop",
+        default=False,
+        action="store_true",
+        help=(
+            "If set the images will be randomly"
+            " cropped (instead of center). The images will be resized to the resolution first before cropping."
+        ),
+    )
+    parser.add_argument(
+        "--no_hflip",
+        action="store_true",
+        help="whether to supress horizontal flipping",
+    )
+    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=3
+    )
+    parser.add_argument(
+        "--max_train_steps",
+        type=int,
+        default=2000,
+        help="Total number of training steps to perform.  If provided, overrides num_train_epochs.",
+    )
+    parser.add_argument(
+        "--gradient_accumulation_steps",
+        type=int,
+        default=1,
+        help="Number of updates steps to accumulate before performing a backward/update pass.",
+    )
+    parser.add_argument(
+        "--gradient_checkpointing",
+        action="store_true",
+        help="Whether or not to use gradient checkpointing to save memory at the expense of slower backward pass.",
+    )
+    parser.add_argument(
+        "--learning_rate_unet",
+        type=float,
+        default=1e-6,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--learning_rate_lora",
+        type=float,
+        default=1e-5,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--learning_rate_prompts",
+        type=float,
+        default=1e-5,
+        help="Initial learning rate (after the potential warmup period) to use.",
+    )
+    parser.add_argument(
+        "--scale_lr",
+        action="store_false",
+        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_with_warmup",
+        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(
+        "--use_adafactor", action="store_true", help="Whether or not to use adafactor (should save mem)"
+    )
+    # Bram Note: Haven't looked @ this yet
+    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(
+        "--dataloader_num_workers",
+        type=int,
+        default=16,
+        help=(
+            "Number of subprocesses to use for data loading. 0 means that the data will be loaded in the main process."
+        ),
+    )
+    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.")
+    parser.add_argument(
+        "--hub_model_id",
+        type=str,
+        default=None,
+        help="The name of the repository to keep in sync with the local `output_dir`.",
+    )
+    parser.add_argument(
+        "--logging_dir",
+        type=str,
+        default="logs",
+        help=(
+            "[TensorBoard](https://www.tensorflow.org/tensorboard) log directory. Will default to"
+            " *output_dir/runs/**CURRENT_DATETIME_HOSTNAME***."
+        ),
+    )
+    parser.add_argument(
+        "--mixed_precision",
+        type=str,
+        default="no",
+        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(
+        "--report_to",
+        type=str,
+        default="wandb",
+        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("--local_rank", type=int, default=-1, help="For distributed training: local_rank")
+    parser.add_argument(
+        "--checkpointing_steps",
+        type=int,
+        default=100,
+        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(
+        "--resume_from_checkpoint",
+        type=str,
+        default='latest',
+        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("--noise_offset", type=float, default=0, help="The scale of noise offset.")
+    parser.add_argument(
+        "--tracker_project_name",
+        type=str,
+        default="EmotionDPO",
+        help="exp group name",
+    )
+    parser.add_argument(
+        "--tracker_run_name",
+        type=str,
+        default="emotion_dpo_lora_v0_0_3_sdxl_2",
+        help="exp name",
+    )
+    parser.add_argument(
+        "--output_dir",
+        type=str,
+        default="log_training/emotion_dpo_lora_v0_0_3_sdxl_2",
+        help="The output directory where the model predictions and checkpoints will be written.",
+    )
+
+    ## SDXL
+    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("--sdxl", action='store_false', help="Train sdxl")
+    
+    ## DPO
+    parser.add_argument("--sft", action='store_true', help="Run Supervised Fine-Tuning instead of Direct Preference Optimization")
+    parser.add_argument("--beta_dpo", type=float, default=5000, help="The beta DPO temperature controlling strength of KL penalty")
+    parser.add_argument(
+        "--hard_skip_resume", action="store_true", help="Load weights etc. but don't iter through loader for loader resume, useful b/c resume takes forever"
+    )
+    parser.add_argument(
+        "--unet_init", type=str, default='', help="Initialize start of run from unet (not compatible w/ checkpoint load)"
+    )
+    parser.add_argument(
+        "--proportion_empty_prompts",
+        type=float,
+        default=0.2,
+        help="Proportion of image prompts to be replaced with empty strings. Defaults to 0 (no prompt replacement).",
+    )
+    parser.add_argument(
+        "--split", type=str, default='train', help="Datasplit"
+    )
+    parser.add_argument(
+        "--choice_model", type=str, default='', help="Model to use for ranking (override dataset PS label_0/1). choices: aes, clip, hps, pickscore"
+    )
+    parser.add_argument(
+        "--dreamlike_pairs_only", action="store_true", help="Only train on pairs where both generations are from dreamlike"
+    )
+    parser.add_argument(
+        "--use_lora",
+        action="store_true",
+        default=True,
+        help="Whether or not to use LoRA (Low-Rank Adaptation).",
+    )
+    parser.add_argument(
+        "--lora_rank",
+        type=int,
+        default=64,
+        help="Rank parameter for LoRA (Low-Rank Adaptation).",
+    )
+    
+    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
+
+    # Sanity checks
+    if args.dataset_name is None and args.train_data_dir is None:
+        raise ValueError("Need either a dataset name or a training folder.")
+
+    ## SDXL
+    if args.sdxl:
+        print("Running SDXL")
+    if args.resolution is None:
+        if args.sdxl:
+            args.resolution = 512
+        else:
+            args.resolution = 512
+            
+    args.train_method = 'sft' if args.sft else 'dpo'
+    return args

MMP_Diffusion_Lora_train.py

@@ -0,0 +1,861 @@
+import io
+import logging
+import math
+import os
+os.environ["CUDA_VISIBLE_DEVICES"] = "1" 
+import random
+import shutil
+import sys
+sys.path.append('./')
+from pathlib import Path
+
+import accelerate
+import datasets
+import numpy as np
+from PIL import Image
+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.state import AcceleratorState
+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 torchvision import transforms
+from tqdm.auto import tqdm
+from transformers import CLIPTextModel, CLIPTextModelWithProjection, CLIPTokenizer
+from transformers.utils import ContextManagers
+
+import diffusers
+from diffusers import AutoencoderKL, DDPMScheduler, StableDiffusionPipeline, StableDiffusionXLPipeline, UNet2DConditionModel
+from models.unet_2d_condition import UNet2DLoRAConditionModel
+from models.lora import add_lora_to_model
+from diffusers.optimization import get_scheduler
+from diffusers.utils import check_min_version, deprecate, is_wandb_available, make_image_grid
+from diffusers.utils.import_utils import is_xformers_available
+from MMP_Diffusion_Lora_config import parse_args, import_model_class_from_model_name_or_path
+
+from peft.utils import get_peft_model_state_dict
+from diffusers.utils import convert_state_dict_to_diffusers
+from models.visual_prompts import EmotionEmbedding, EmotionEmbedding2
+import copy
+
+if is_wandb_available():
+    import wandb
+
+
+## SDXL
+import functools
+import gc
+from torchvision.transforms.functional import crop
+from transformers import AutoTokenizer
+
+# Will error if the minimal version of diffusers is not installed. Remove at your own risks.
+check_min_version("0.20.0")
+
+logger = get_logger(__name__, log_level="INFO")
+
+# Adapted from pipelines.StableDiffusionXLPipeline.encode_prompt
+def encode_prompt_sdxl(batch, text_encoders, tokenizers, proportion_empty_prompts, caption_column, is_train=True):
+    prompt_embeds_list = []
+    prompt_batch = batch[caption_column]
+
+    captions = []
+    for caption in prompt_batch:
+        if random.random() < proportion_empty_prompts:
+            captions.append("")
+        elif 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])
+
+    for tokenizer, text_encoder in zip(tokenizers, text_encoders):
+        
+        text_input_ids = tokenizer(
+            captions,
+            padding="max_length",
+            max_length=tokenizer.model_max_length,
+            truncation=True,
+            return_tensors="pt",
+        ).input_ids
+        
+        with torch.no_grad():
+            prompt_embeds = text_encoder(
+                text_input_ids.to('cuda'),
+                output_hidden_states=True,
+            )
+
+        # We are only ALWAYS interested in the pooled output of the final text encoder
+        # torch.Size([32, 1280]) this
+        # odict_keys(['text_embeds', 'last_hidden_state', 'hidden_states'])
+        if isinstance(text_encoder, CLIPTextModel):
+            pass
+        elif isinstance(text_encoder, CLIPTextModelWithProjection):
+            pooled_prompt_embeds = prompt_embeds[0]
+        
+        # "2" because SDXL always indexes from the penultimate layer.
+        # torch.Size([32, 77, 768/1280])
+        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)
+
+    # torch.Size([32, 77, 768+1280=2048])
+    prompt_embeds = torch.concat(prompt_embeds_list, dim=-1)
+    # torch.Size([32, 1280])
+    pooled_prompt_embeds = pooled_prompt_embeds.view(bs_embed, -1)
+
+    return {
+        "prompt_embeds": prompt_embeds, 
+        "pooled_prompt_embeds": pooled_prompt_embeds,
+    }
+
+
+def init_emotion_prompts(visual_prompts_dir, is_sdxl=True, prompt_len=16):
+    emotions = ["amusement", "anger", "awe", "contentment", 
+                "disgust", "excitement", "fear", "sadness"]
+    if is_sdxl: 
+        output_dim = 2048
+    else: 
+        output_dim = 768
+    feature_names = ["clip", "vgg", "dinov2"]
+    visual_prompts = EmotionEmbedding(emotions, visual_prompts_dir,
+                                      feature_names, output_dim=output_dim, prompt_len=prompt_len)
+    return visual_prompts
+
+def init_emotion_prompts2(is_sdxl=True):
+    emotions = ["amusement", "anger", "awe", "contentment", 
+                "disgust", "excitement", "fear", "sadness"]
+    if is_sdxl: 
+        output_dim = 2048
+    else: 
+        output_dim = 768
+    input_dim = 2048
+    visual_prompts = EmotionEmbedding2(emotions, input_dim, output_dim=output_dim)
+    return visual_prompts
+
+def random_sample_emotions(anchor_emotions):
+    emotions = ["amusement", "anger", "awe", "contentment", "disgust",
+                "excitement", "fear", "sadness"]
+    random_emotions = []
+    for anchor in anchor_emotions:
+        available_emotions = [emotion for emotion in emotions if emotion != anchor]
+        random_choice = random.choice(available_emotions)
+        random_emotions.append(random_choice)
+    return random_emotions
+
+
+def main():
+    args = parse_args()
+    #### START ACCELERATOR BOILERPLATE ###
+    logging_dir = os.path.join(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,
+    )
+
+    # 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:
+        datasets.utils.logging.set_verbosity_warning()
+        transformers.utils.logging.set_verbosity_warning()
+        diffusers.utils.logging.set_verbosity_info()
+    else:
+        datasets.utils.logging.set_verbosity_error()
+        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 + accelerator.process_index) # added in + term, untested
+
+    # Handle the repository creation
+    if accelerator.is_main_process:
+        if args.output_dir is not None:
+            os.makedirs(args.output_dir, exist_ok=True)
+    ### END ACCELERATOR BOILERPLATE
+    
+    
+    ### START DIFFUSION BOILERPLATE ###
+    # Load scheduler, tokenizer and models.
+    noise_scheduler = DDPMScheduler.from_pretrained(args.pretrained_model_name_or_path, 
+                                                    subfolder="scheduler")
+    
+    # SDXL has two text encoders
+    if args.sdxl:
+        tokenizer_and_encoder_name = args.pretrained_model_name_or_path
+        tokenizer_one = AutoTokenizer.from_pretrained(tokenizer_and_encoder_name, subfolder="tokenizer", revision=args.revision, use_fast=False)
+        tokenizer_two = AutoTokenizer.from_pretrained(tokenizer_and_encoder_name, subfolder="tokenizer_2", revision=args.revision, use_fast=False)
+    else:
+        tokenizer = CLIPTokenizer.from_pretrained(args.pretrained_model_name_or_path, subfolder="tokenizer", revision=args.revision)
+
+    # Not sure if we're hitting this at all
+    def deepspeed_zero_init_disabled_context_manager():
+        """
+        returns either a context list that includes one that will disable zero.Init or an empty context list
+        """
+        deepspeed_plugin = AcceleratorState().deepspeed_plugin if accelerate.state.is_initialized() else None
+        if deepspeed_plugin is None:
+            return []
+
+        return [deepspeed_plugin.zero3_init_context_manager(enable=False)]
+
+    with ContextManagers(deepspeed_zero_init_disabled_context_manager()):
+        # SDXL has two text encoders
+        if args.sdxl:
+            # import correct text encoder classes
+            text_encoder_cls_one = import_model_class_from_model_name_or_path(tokenizer_and_encoder_name, args.revision, subfolder="text_encoder")
+            text_encoder_cls_two = import_model_class_from_model_name_or_path(tokenizer_and_encoder_name, args.revision, subfolder="text_encoder_2")
+            text_encoder_one = text_encoder_cls_one.from_pretrained(tokenizer_and_encoder_name, revision=args.revision, subfolder="text_encoder")
+            text_encoder_two = text_encoder_cls_two.from_pretrained(tokenizer_and_encoder_name, revision=args.revision, subfolder="text_encoder_2")
+            text_encoders = [text_encoder_one, text_encoder_two]
+            tokenizers = [tokenizer_one, tokenizer_two]
+        else:
+            text_encoder = CLIPTextModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="text_encoder", revision=args.revision)
+        # Can custom-select VAE (used in original SDXL tuning)
+        vae_path = (
+            args.pretrained_model_name_or_path
+            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.revision
+        )
+        # clone of model
+        ref_unet = UNet2DConditionModel.from_pretrained(
+            args.pretrained_model_name_or_path,
+            subfolder="unet", revision=args.revision
+        )
+
+    unet = UNet2DLoRAConditionModel.from_pretrained(args.pretrained_model_name_or_path, subfolder="unet", revision=args.revision)
+
+    print("======== init_emotion_prompts ================")
+    visual_prompts = init_emotion_prompts(args.visual_prompts_dir, is_sdxl=args.sdxl, prompt_len=args.prompt_len).to(accelerator.device)
+    # visual_prompts = init_emotion_prompts2(is_sdxl=args.sdxl).to(accelerator.device)
+    print("======== init_emotion_prompts done ================")
+
+    # Freeze vae, text_encoder(s), reference unet
+    vae.requires_grad_(False)
+    if args.sdxl:
+        text_encoder_one.requires_grad_(False)
+        text_encoder_two.requires_grad_(False)
+    else:
+        text_encoder.requires_grad_(False)
+    if args.train_method == 'dpo': 
+        ref_unet.requires_grad_(False)
+
+    # if args.use_lora: 
+    #     unet.requires_grad_(False)
+    # args.lora_rank default 32
+    lora_p, negation = add_lora_to_model(unet, dropout=0.1, lora_rank=args.lora_rank, scale=1.0)
+
+    # xformers 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")
+
+    # BRAM NOTE: We're using >=0.16.0. Below was a bit of a bug hive. I hacked around it, but ideally ref_unet wouldn't
+    # be getting passed here
+    # 
+    # `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):
+
+            print("save_model_hook")
+            for i in range(len(models)):
+                print(models[i].__class__.__name__)
+
+            if len(models) > 1:
+                assert args.train_method == 'dpo' # 2nd model is just ref_unet in DPO case
+            
+            if args.sdxl:
+
+                # UNet2DLoRAConditionModel
+                models[0].save_pretrained(os.path.join(output_dir, 'unet_with_lora'))
+                weights.pop()
+
+                # EmotionEmbedding
+                torch.save(models[1].state_dict(), os.path.join(output_dir, "EmotionEmbedding.pth"))
+                weights.pop()
+
+        def load_model_hook(models, input_dir):
+
+            print("load_model_hook")
+            for i in range(len(models)):
+                print(models[i].__class__.__name__)
+
+            if len(models) > 1:
+                assert args.train_method == 'dpo' # 2nd model is just ref_unet in DPO case
+
+            if args.sdxl:
+
+                # UNet2DLoRAConditionModel
+                model = models.pop(0)
+                from safetensors.torch import load_file
+                # 加载两个safetensors文件
+                state_dict_1 = load_file(os.path.join(input_dir, 'unet_with_lora', 'diffusion_pytorch_model-00001-of-00002.safetensors'))
+                state_dict_2 = load_file(os.path.join(input_dir, 'unet_with_lora', 'diffusion_pytorch_model-00002-of-00002.safetensors'))
+                # 合并状态字典
+                state_dict = {**state_dict_1, **state_dict_2}
+                model.load_state_dict(state_dict)
+
+                # EmotionEmbedding
+                model = models.pop(0)
+                state_dict = torch.load(os.path.join(input_dir, "EmotionEmbedding.pth"), weights_only=True)
+                model.load_state_dict(state_dict)
+
+        accelerator.register_save_state_pre_hook(save_model_hook)
+        accelerator.register_load_state_pre_hook(load_model_hook)
+
+    if args.gradient_checkpointing or args.sdxl: #  (args.sdxl and ('turbo' not in args.pretrained_model_name_or_path) ):
+        print("Enabling gradient checkpointing, either because you asked for this or because you're using SDXL")
+        unet.enable_gradient_checkpointing()
+
+    # Bram Note: haven't touched
+    # 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
+        )
+
+    unet_params = []
+    lora_params = [] 
+    for name, param in unet.named_parameters():
+        if 'lora' in name.lower():
+            lora_params.append(param)
+        else:
+            if param.requires_grad:
+                unet_params.append(param)
+
+    # if args.use_adafactor or args.sdxl:
+    print("Using Adafactor either because you asked for it or you're using SDXL")
+    param_groups = [
+        {
+            "params": unet_params, 
+            "lr": args.learning_rate_unet, 
+        },
+        {
+            "params": lora_params, 
+            "lr": args.learning_rate_lora, 
+        },
+        {
+            "params": visual_prompts.parameters(), 
+            "lr": args.learning_rate_prompts, 
+        }
+    ]
+    optimizer = transformers.Adafactor(
+        param_groups,
+        weight_decay=args.adam_weight_decay,
+        clip_threshold=1.0,
+        scale_parameter=False,
+        relative_step=False
+    )
+    
+    # else:
+    #     optimizer = torch.optim.AdamW([
+    #         {"params": unet_params, "lr": args.learning_rate,
+    #          "beta": (args.adam_beta1, args.adam_beta2), "weight_decay": args.adam_weight_decay,
+    #          "eps": args.adam_epsilon},
+    #         {"params": lora_params, "lr": args.learning_rate*5,
+    #          "beta": (args.adam_beta1, args.adam_beta2), "weight_decay": args.adam_weight_decay,
+    #          "eps": args.adam_epsilon},
+    #         {"params": visual_prompts.parameters(), "lr": args.learning_rate*5,
+    #          "beta": (args.adam_beta1, args.adam_beta2), "weight_decay": args.adam_weight_decay,
+    #          "eps": args.adam_epsilon}
+    #     ])
+        
+        
+    # In distributed training, the load_dataset function guarantees that only one local process can concurrently
+    # download the dataset.
+    dataset = load_dataset(path='parquet', data_files=args.dataset_path)
+    caption_column = args.caption_column
+
+    def tokenize_captions(examples, is_train=True):
+        captions = []
+        for caption in examples[caption_column]:
+            if random.random() < args.proportion_empty_prompts:
+                captions.append("")
+            elif 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])
+            else:
+                raise ValueError(
+                    f"Caption column `{caption_column}` should contain either strings or lists of strings."
+                )
+        inputs = tokenizer(
+            captions, max_length=tokenizer.model_max_length, padding="max_length", truncation=True, return_tensors="pt"
+        )
+        return inputs.input_ids
+
+    # Preprocessing the datasets.
+    train_transforms = transforms.Compose(
+        [
+            transforms.Resize(args.resolution, interpolation=transforms.InterpolationMode.BILINEAR),
+            transforms.RandomCrop(args.resolution) if args.random_crop else transforms.CenterCrop(args.resolution),
+            transforms.Lambda(lambda x: x) if args.no_hflip else transforms.RandomHorizontalFlip(),
+            transforms.ToTensor(),
+            transforms.Normalize([0.5], [0.5]),
+        ]
+    )
+    
+    #### START PREPROCESSING/COLLATION ####
+    if args.train_method == 'dpo':
+        print("Ignoring image_column variable, reading from jpg_0 and jpg_1")
+        def preprocess_train(examples):
+            all_pixel_values = []
+            for col_name in ['jpg_0', 'jpg_1']:
+                images = [Image.open(io.BytesIO(im_bytes)).convert("RGB")
+                            for im_bytes in examples[col_name]]
+                pixel_values = [train_transforms(image) for image in images]
+                all_pixel_values.append(pixel_values)
+            # DOUBLE win images for visual prompts optimization
+            # all_pixel_values 
+            # [[jpg_0,...],[jpg_1,...]]
+            # => [[jpg_0,...],[jpg_1,...],[jpg_0,...]]
+            all_pixel_values.append(copy.deepcopy(all_pixel_values[0]))
+            
+            # Triple on channel dim, jpg_y then jpg_w and jpg_y
+            # im_tup_iterator = [(jpg_0,jpg_1,jpg_0),...]
+            im_tup_iterator = zip(*all_pixel_values)
+            combined_pixel_values = []
+            # item = (jpg_0,jpg_1,jpg_0), label
+            for im_tup, label_0 in zip(im_tup_iterator, examples['label_0']):
+                # print(len(im_tup), im_tup[0].shape)
+                # 3 torch.Size([3, 512, 512])
+                if label_0==0 and (not args.choice_model): # don't want to flip things if using choice_model for AI feedback
+                    im_tup = im_tup[::-1]
+                # [3+3+3, 512, 512]
+                combined_im = torch.cat(im_tup, dim=0) # no batch dim
+                combined_pixel_values.append(combined_im)
+            # [[9, 512, 512],...]
+            examples["pixel_values"] = combined_pixel_values
+            # SDXL takes raw prompts
+            if not args.sdxl:
+                examples["input_ids"] = tokenize_captions(examples)
+            return examples
+
+        def collate_fn(examples):
+            # [bs, 9, 512, 512]
+            pixel_values = torch.stack([example["pixel_values"] for example in examples])
+            pixel_values = pixel_values.to(memory_format=torch.contiguous_format).float()
+            return_d = {"pixel_values": pixel_values}
+            return_d["emotions"] = [example["emotion"] for example in examples]
+            
+            # SDXL takes raw prompts
+            if args.sdxl:
+                return_d["caption"] = [example["caption"] for example in examples]
+            else:
+                return_d["input_ids"] = torch.stack([example["input_ids"] for example in examples])
+                
+            if args.choice_model:
+                # If using AIF then deliver image data for choice model to determine if should flip pixel values
+                for k in ['jpg_0', 'jpg_1']:
+                    return_d[k] = [Image.open(io.BytesIO( example[k])).convert("RGB")
+                                   for example in examples]
+                return_d["caption"] = [example["caption"] for example in examples] 
+            return return_d
+    
+    ### DATASET #####
+    with accelerator.main_process_first():
+        if args.max_train_samples is not None:
+            dataset[args.split] = dataset[args.split].shuffle(seed=args.seed).select(range(args.max_train_samples))
+        train_dataset = dataset[args.split].with_transform(preprocess_train)
+
+    # DataLoaders creation:
+    train_dataloader = torch.utils.data.DataLoader(
+        train_dataset,
+        shuffle=(args.split=='train'),
+        collate_fn=collate_fn,
+        batch_size=args.train_batch_size,
+        num_workers=args.dataloader_num_workers,
+        drop_last=True
+    )
+    ##### END BIG OLD DATASET BLOCK #####
+    
+    # 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,
+    )
+
+    #### START ACCELERATOR PREP ####
+    unet, visual_prompts, optimizer, train_dataloader, lr_scheduler = accelerator.prepare(
+        unet, visual_prompts, optimizer, train_dataloader, lr_scheduler
+    )
+
+    weight_dtype = torch.float32
+        
+    # Move text_encode and vae to gpu and cast to weight_dtype
+    vae.to(accelerator.device, dtype=weight_dtype)
+    if args.sdxl:
+        text_encoder_one.to(accelerator.device, dtype=weight_dtype)
+        text_encoder_two.to(accelerator.device, dtype=weight_dtype)
+        # print("offload vae (this actually stays as CPU)")
+        # vae = accelerate.cpu_offload(vae)
+        # print("Offloading text encoders to cpu")
+        text_encoder_one = accelerate.cpu_offload(text_encoder_one)
+        text_encoder_two = accelerate.cpu_offload(text_encoder_two)
+        if args.train_method == 'dpo':
+            ref_unet.to(accelerator.device, dtype=weight_dtype)
+            # print("offload ref_unet")
+            # ref_unet = accelerate.cpu_offload(ref_unet)
+    else:
+        text_encoder.to(accelerator.device, dtype=weight_dtype)
+        if args.train_method == 'dpo':
+            ref_unet.to(accelerator.device, dtype=weight_dtype)
+    ### END ACCELERATOR PREP ###
+    
+    
+    # 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))
+        init_kwargs = {
+            "wandb": {
+                "name": args.tracker_run_name,
+            }
+        }
+        accelerator.init_trackers(args.tracker_project_name, tracker_config, init_kwargs)
+
+    # Training initialization
+    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
+        else:
+            accelerator.print(f"Resuming from checkpoint {path}")
+            accelerator.load_state(os.path.join(args.output_dir, path))
+            global_step = int(path.split("-")[1])
+
+            resume_global_step = global_step * args.gradient_accumulation_steps
+            first_epoch = global_step // num_update_steps_per_epoch
+            resume_step = resume_global_step % (num_update_steps_per_epoch * args.gradient_accumulation_steps)
+        
+
+    # Bram Note: This was pretty janky to wrangle to look proper but works to my liking now
+    progress_bar = tqdm(range(global_step, args.max_train_steps), disable=not accelerator.is_local_main_process)
+    progress_bar.set_description("Steps")
+
+        
+    #### START MAIN TRAINING LOOP #####
+    for epoch in range(first_epoch, args.num_train_epochs):
+        unet.train()
+        train_loss = 0.0
+        implicit_acc_accumulated_d, implicit_acc_accumulated_c = 0.0, 0.0
+        for step, batch in enumerate(train_dataloader):
+            # Skip steps until we reach the resumed step
+            if args.resume_from_checkpoint and epoch == first_epoch and step < resume_step and (not args.hard_skip_resume):
+                if step % args.gradient_accumulation_steps == 0:
+                    print(f"Dummy processing step {step}, will start training at {resume_step}")
+                continue
+            with accelerator.accumulate(unet):
+                # Convert images to latent space
+                if args.train_method == 'dpo':
+                    # [bs, 6, 512, 512] => 
+                    # [[bs, 3, 512, 512]*3] =>
+                    # [bs*3, 3, 512, 512]
+                    feed_pixel_values = torch.cat(batch["pixel_values"].chunk(3, dim=1))
+                elif args.train_method == 'sft':
+                    feed_pixel_values = batch["pixel_values"]
+                
+                #### Diffusion Stuff ####
+                # encode pixels --> latents
+                with torch.no_grad():
+                    latents = vae.encode(feed_pixel_values.to(weight_dtype)).latent_dist.sample()
+                    latents = latents * vae.config.scaling_factor
+
+                # Sample noise that we'll add to the latents
+                noise = torch.randn_like(latents)
+
+                bsz = latents.shape[0]
+                # Sample a random timestep for each image
+                timesteps = torch.randint(0, noise_scheduler.config.num_train_timesteps, (bsz,), device=latents.device)
+                timesteps = timesteps.long()
+                
+                if args.train_method == 'dpo': 
+                    # make timesteps and noise same for pairs in DPO
+                    # [bs] => [1/3bs, 1/3bs, 1/3bs] => [1/3bs] => [bs]
+                    timesteps = timesteps.chunk(3)[0].repeat(3)
+                    noise = noise.chunk(3)[0].repeat(3, 1, 1, 1)
+
+                # Add noise to the latents according to the noise magnitude at each timestep
+                # (this is the forward diffusion process)
+                
+                noisy_latents = noise_scheduler.add_noise(latents, noise, timesteps)
+                ### START PREP BATCH ###
+                if args.sdxl:
+                    # Get the text embedding for conditioning
+                    with torch.no_grad():
+                        # Need to compute "time_ids" https://github.com/huggingface/diffusers/blob/v0.20.0-release/examples/text_to_image/train_text_to_image_sdxl.py#L969
+                        # for SDXL-base these are torch.tensor([args.resolution, args.resolution, *crop_coords_top_left, *target_size))
+                        add_time_ids = torch.tensor([args.resolution, 
+                                                    args.resolution,
+                                                    0,
+                                                    0,
+                                                    args.resolution, 
+                                                    args.resolution],
+                                                    dtype=weight_dtype,
+                                                    device=accelerator.device)[None, :].repeat(timesteps.size(0), 1)
+                        prompt_batch = encode_prompt_sdxl(batch, 
+                                                          text_encoders,
+                                                           tokenizers,
+                                                           args.proportion_empty_prompts, 
+                                                           caption_column,
+                                                           is_train=True,
+                                                          )
+                    if args.train_method == 'dpo':
+                        prompt_batch["prompt_embeds"] = prompt_batch["prompt_embeds"].repeat(3, 1, 1)
+                        prompt_batch["pooled_prompt_embeds"] = prompt_batch["pooled_prompt_embeds"].repeat(3, 1)
+                    unet_added_conditions = {"time_ids": add_time_ids,
+                                            "text_embeds": prompt_batch["pooled_prompt_embeds"]}
+                else: # sd1.5
+                    # Get the text embedding for conditioning
+                    encoder_hidden_states = text_encoder(batch["input_ids"])[0]
+                    if args.train_method == 'dpo':
+                        encoder_hidden_states = encoder_hidden_states.repeat(2, 1, 1)
+                
+                emotion_visual_prompts = visual_prompts(batch['emotions'])
+
+                if args.train_method == 'dpo':
+                    random_emotions = random_sample_emotions(batch['emotions'])
+                    random_emotion_visual_prompts = visual_prompts(random_emotions)
+                    emotion_visual_prompts = torch.cat([emotion_visual_prompts, emotion_visual_prompts, random_emotion_visual_prompts], dim=0)
+                
+                #### END PREP BATCH ####
+
+                assert noise_scheduler.config.prediction_type == "epsilon"
+                target = noise
+               
+                # Make the prediction from the model we're learning
+                model_batch_args = (
+                    noisy_latents,
+                    timesteps, 
+                    prompt_batch["prompt_embeds"] if args.sdxl else encoder_hidden_states
+                )
+                lora_model_batch_args = (
+                    noisy_latents,
+                    timesteps, 
+                    prompt_batch["prompt_embeds"] if args.sdxl else encoder_hidden_states,
+                    emotion_visual_prompts
+                )
+                added_cond_kwargs = unet_added_conditions if args.sdxl else None
+                
+                model_pred = unet(
+                    *lora_model_batch_args,
+                    added_cond_kwargs = added_cond_kwargs
+                ).sample
+                #### START LOSS COMPUTATION ####
+                if args.train_method == 'sft': # SFT, casting for F.mse_loss
+                    loss = F.mse_loss(model_pred.float(), target.float(), reduction="mean")
+                elif args.train_method == 'dpo':
+                    # model_pred and ref_pred will be (2 * LBS) x 4 x latent_spatial_dim x latent_spatial_dim
+                    # losses are both 2 * LBS
+                    # 1st half of tensors is preferred (y_w), second half is unpreferred
+                    model_losses = (model_pred - target).pow(2).mean(dim=[1,2,3])
+                    model_losses_w, model_losses_l_d, model_losses_l_c = model_losses.chunk(3)
+                    # below for logging purposes
+                    raw_model_loss = (model_losses_w.mean() + model_losses_l_d.mean() + model_losses_l_c.mean()) / 3
+                    
+                    model_diff_d = model_losses_w - model_losses_l_d # These are both LBS (as is t)
+                    model_diff_c = model_losses_w - model_losses_l_c
+                    
+                    with torch.no_grad(): # Get the reference policy (unet) prediction
+                        ref_pred = ref_unet(
+                                    *model_batch_args,
+                                      added_cond_kwargs = added_cond_kwargs
+                                     ).sample.detach()
+                        ref_losses = (ref_pred - target).pow(2).mean(dim=[1,2,3])
+                        ref_losses_w, ref_losses_l_d, ref_losses_l_c = ref_losses.chunk(3)
+                        ref_diff = ref_losses_w - ref_losses_l_d
+                        raw_ref_loss = ref_losses.mean()
+
+                    scale_term = -0.5 * args.beta_dpo # beta_dpo = 5000
+                    inside_term_d = scale_term * (model_diff_d - ref_diff)
+                    implicit_acc_d = (inside_term_d > 0).sum().float() / inside_term_d.size(0)
+                    # the scale_term may need to be adjust
+                    # inside_term_c = -1 * model_diff_c
+                    inside_term_c = scale_term * model_diff_c
+                    implicit_acc_c = (inside_term_c > 0).sum().float() / inside_term_c.size(0)
+                    loss = -1 * 0.5 * (F.logsigmoid(inside_term_d).mean() + F.logsigmoid(inside_term_c).mean())
+                #### END LOSS COMPUTATION ###
+                    
+                # Gather the losses across all processes for logging 
+                avg_loss = accelerator.gather(loss.repeat(args.train_batch_size)).mean()
+                train_loss += avg_loss.item() / args.gradient_accumulation_steps
+                # Also gather:
+                # - model MSE vs reference MSE (useful to observe divergent behavior)
+                # - Implicit accuracy
+                if args.train_method == 'dpo':
+                    avg_model_mse = accelerator.gather(raw_model_loss.repeat(args.train_batch_size)).mean().item()
+                    avg_ref_mse = accelerator.gather(raw_ref_loss.repeat(args.train_batch_size)).mean().item()
+                    avg_acc_d = accelerator.gather(implicit_acc_d).mean().item()
+                    avg_acc_c = accelerator.gather(implicit_acc_c).mean().item()
+                    implicit_acc_accumulated_d += avg_acc_d / args.gradient_accumulation_steps
+                    implicit_acc_accumulated_c += avg_acc_c / args.gradient_accumulation_steps
+
+                # Backpropagate
+                accelerator.backward(loss)
+                if accelerator.sync_gradients:
+                    if not args.use_adafactor: # Adafactor does itself, maybe could do here to cut down on code
+                        accelerator.clip_grad_norm_(unet.parameters(), args.max_grad_norm)
+
+                # # 打印看看梯度
+                # for name, param in unet.named_parameters():
+                #     # if "mid_block.attentions.0.transformer_blocks" in name and "lora" in name:
+                #     if param.grad is not None:
+                #         print(f"{name} has gradient ✅, grad mean: {param.grad.mean().item()}")
+                #     else:
+                #         print(f"{name} has NO gradient ❌")
+                # for name, param in visual_prompts.named_parameters():
+                #     if param.grad is not None:
+                #         print(f"{name} has gradient ✅, grad mean: {param.grad.mean().item()}")
+                #     else:
+                #         print(f"{name} has NO gradient ❌")
+
+                optimizer.step()
+                lr_scheduler.step()
+                optimizer.zero_grad()
+
+            # Checks if the accelerator has just performed an optimization step, if so do "end of batch" logging
+            if accelerator.sync_gradients:
+                progress_bar.update(1)
+                global_step += 1
+                accelerator.log({"train_loss": train_loss}, step=global_step)
+                if args.train_method == 'dpo':
+                    accelerator.log({"model_mse_unaccumulated": avg_model_mse}, step=global_step)
+                    accelerator.log({"ref_mse_unaccumulated": avg_ref_mse}, step=global_step)
+                    accelerator.log({"avg_acc_d": implicit_acc_accumulated_d}, step=global_step)
+                    accelerator.log({"avg_acc_c": implicit_acc_accumulated_c}, step=global_step)
+                train_loss = 0.0
+                implicit_acc_accumulated_d, implicit_acc_accumulated_c = 0.0, 0.0
+
+                if global_step % args.checkpointing_steps == 0:
+                    if accelerator.is_main_process:
+                        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}")
+                        logger.info("Pretty sure saving/loading is fixed but proceed cautiously")
+
+            logs = {"step_loss": loss.detach().item(), "lr": lr_scheduler.get_last_lr()[0]}
+            if args.train_method == 'dpo':
+                logs["implicit_acc_d"] = avg_acc_d
+                logs["implicit_acc_c"] = avg_acc_c
+            progress_bar.set_postfix(**logs)
+
+            if global_step >= args.max_train_steps:
+                break
+
+
+    # Create the pipeline using the trained modules and save it.
+    # This will save to top level of output_dir instead of a checkpoint directory
+    accelerator.wait_for_everyone()
+    if accelerator.is_main_process:
+        unet = accelerator.unwrap_model(unet)
+        if args.sdxl:
+            # Serialize pipeline.
+            if args.use_lora:
+                unet_lora_state_dict = convert_state_dict_to_diffusers(
+                    get_peft_model_state_dict(unet)
+                )
+                StableDiffusionXLPipeline.save_lora_weights(
+                    save_directory=os.path.join(args.output_dir, 'lora_weights_64'),
+                    unet_lora_layers=unet_lora_state_dict,
+                    safe_serialization=True,
+                )
+                logger.info("Saved LoRA Model to {}".format(os.path.join(args.output_dir, 'lora_weights_64')))
+            else:
+                vae = AutoencoderKL.from_pretrained(
+                    vae_path,
+                    subfolder="vae" if args.pretrained_vae_model_name_or_path is None else None,
+                    revision=args.revision,
+                    torch_dtype=weight_dtype,
+                )
+                pipeline = StableDiffusionXLPipeline.from_pretrained(
+                    args.pretrained_model_name_or_path, unet=unet, vae=vae, revision=args.revision, torch_dtype=weight_dtype
+                )
+                pipeline.save_pretrained(args.output_dir)
+                logger.info("Saved Model to {}".format(args.output_dir))
+        else:
+            pipeline = StableDiffusionPipeline.from_pretrained(
+                args.pretrained_model_name_or_path,
+                text_encoder=text_encoder,
+                vae=vae,
+                unet=unet,
+                revision=args.revision,
+            )
+        if not args.use_lora: pipeline.save_pretrained(args.output_dir)
+
+
+    accelerator.end_training()
+
+
+if __name__ == "__main__":
+    main()

models/lora.py

@@ -0,0 +1,246 @@
+import math
+import torch
+import torch.nn as nn
+from typing import Set, List, Optional, Type
+
+UNET_DEFAULT_TARGET_REPLACE = {"CrossAttention", "Attention", "GEGLU"}
+DEFAULT_TARGET_REPLACE = UNET_DEFAULT_TARGET_REPLACE
+
+
+class LoraInjectedLinear(nn.Module):
+    def __init__(
+        self, in_features, out_features, bias=False, r=4, dropout_p=0.1, scale=1.0
+    ):
+        super().__init__()
+
+        if r > min(in_features, out_features):
+            #raise ValueError(
+            #    f"LoRA rank {r} must be less or equal than {min(in_features, out_features)}"
+            #)
+            print(f"LoRA rank {r} is too large. setting to: {min(in_features, out_features)}")
+            r = min(in_features, out_features)
+
+        self.r = r
+        self.linear = nn.Linear(in_features, out_features, bias)
+        self.lora_down = nn.Linear(in_features, r, bias=False)
+        self.dropout = nn.Dropout(dropout_p)
+        self.lora_up = nn.Linear(r, out_features, bias=False)
+        self.scale = scale
+        self.selector = nn.Identity()
+
+        nn.init.normal_(self.lora_down.weight, std=1 / r)
+        nn.init.zeros_(self.lora_up.weight)
+        
+    def update_step(self, cur_step):
+        self.cur_step = cur_step
+        
+    def forward(self, input, return_format='linear'):
+        assert return_format in ['linear', 'lora', 'added', 'full']
+        if return_format == 'linear': return self.linear(input)
+        elif return_format == 'lora': return self.dropout(self.lora_up(self.selector(self.lora_down(input))))
+        elif return_format == 'added':
+            return (
+                self.linear(input)
+                + self.dropout(self.lora_up(self.selector(self.lora_down(input))))
+                * self.scale
+            )
+        linear_res = self.linear(input)
+        lora_res = self.dropout(self.lora_up(self.selector(self.lora_down(input))))
+        return linear_res, lora_res, self.scale
+
+    def realize_as_lora(self):
+        return self.lora_up.weight.data * self.scale, self.lora_down.weight.data
+
+    def set_selector_from_diag(self, diag: torch.Tensor):
+        # diag is a 1D tensor of size (r,)
+        assert diag.shape == (self.r,)
+        self.selector = nn.Linear(self.r, self.r, bias=False)
+        self.selector.weight.data = torch.diag(diag)
+        self.selector.weight.data = self.selector.weight.data.to(
+            self.lora_up.weight.device
+        ).to(self.lora_up.weight.dtype)
+
+
+def _find_modules_v2(
+    model,
+    ancestor_class: Optional[Set[str]] = None,
+    search_name = 'attn2',
+    include_names = ['to_q', 'to_k', 'to_v'],
+    search_class: List[Type[nn.Module]] = [nn.Linear],
+    exclude_children_of: Optional[List[Type[nn.Module]]] = [
+        LoraInjectedLinear,
+    ],
+):
+    """
+    Find all modules of a certain class (or union of classes) that are direct or
+    indirect descendants of other modules of a certain class (or union of classes).
+
+    Returns all matching modules, along with the parent of those moduless and the
+    names they are referenced by.
+    """
+
+    # Get the targets we should replace all linears under
+    if ancestor_class is not None:
+        ancestors = (
+            module
+            for module in model.modules()
+            if module.__class__.__name__ in ancestor_class
+        )
+    else:
+        # this, incase you want to naively iterate over all modules.
+        ancestors = [module for module in model.modules()]
+
+    # For each target find every linear_class module that isn't a child of a LoraInjectedLinear
+    for ancestor in ancestors:
+        for fullname, module in ancestor.named_modules():
+            if search_name in fullname:
+                *path, base_name = fullname.split('.')
+                parent = ancestor
+                while path:
+                    parent = parent.get_submodule(path.pop(0))
+                if base_name in include_names:
+                    assert isinstance(module, search_class[0])
+                    yield parent, base_name, module
+            # if any([isinstance(module, _class) for _class in search_class]):
+            #     # Find the direct parent if this is a descendant, not a child, of target
+            #     *path, name = fullname.split(".")
+            #     parent = ancestor
+            #     while path:
+            #         parent = parent.get_submodule(path.pop(0))
+            #     # Skip this linear if it's a child of a LoraInjectedLinear
+            #     if exclude_children_of and any(
+            #         [isinstance(parent, _class) for _class in exclude_children_of]
+            #     ):
+            #         continue
+            #     # Otherwise, yield it
+            #     yield parent, name, module
+
+
+def inject_trainable_lora(
+    model: nn.Module,
+    target_replace_module: Set[str] = DEFAULT_TARGET_REPLACE,
+    lora_rank: int = 4,
+    loras=None,  # path to lora .pt
+    verbose: bool = False,
+    dropout: float = 0.0,
+    scale: float = 1.0,
+):
+    """
+    inject lora into model, and returns lora parameter groups.
+    """
+
+    require_grad_params = []
+    names = []
+
+    if loras != None:
+        loras = torch.load(loras)
+
+    for _module, name, _child_module in _find_modules_v2(
+        model, target_replace_module, search_class=[nn.Linear]
+    ):
+        weight = _child_module.weight
+        bias = _child_module.bias
+        if verbose:
+            print("LoRA Injection : injecting lora into ", name)
+            print("LoRA Injection : weight shape", weight.shape)
+        _tmp = LoraInjectedLinear(
+            _child_module.in_features,
+            _child_module.out_features,
+            _child_module.bias is not None,
+            r=lora_rank,
+            dropout_p=dropout,
+            scale=scale,
+        )
+        _tmp.linear.weight = weight
+        if bias is not None:
+            _tmp.linear.bias = bias
+
+        # switch the module
+        _tmp.to(_child_module.weight.device).to(_child_module.weight.dtype)
+        _module._modules[name] = _tmp
+
+        require_grad_params.append(_module._modules[name].lora_up.parameters())
+        require_grad_params.append(_module._modules[name].lora_down.parameters())
+
+        if loras != None:
+            _module._modules[name].lora_up.weight = loras.pop(0)
+            _module._modules[name].lora_down.weight = loras.pop(0)
+
+        _module._modules[name].lora_up.weight.requires_grad = True
+        _module._modules[name].lora_down.weight.requires_grad = True
+        names.append(name)
+
+    return require_grad_params, names
+
+
+lora_args = dict(
+    model = None,
+    loras = None,
+    target_replace_module = [],
+    lora_rank = 4,
+    dropout = 0,
+    scale = 0
+)
+
+
+def extract_lora_ups_down(model, target_replace_module=DEFAULT_TARGET_REPLACE):
+    loras = []
+    for _m, _n, _child_module in _find_modules_v2(
+        model,
+        target_replace_module,
+        search_class=[LoraInjectedLinear],
+    ):
+        loras.append((_child_module.lora_up, _child_module.lora_down))
+
+    if len(loras) == 0:
+        raise ValueError("No lora injected.")
+
+    return loras
+
+
+def do_lora_injection(model, replace_modules, lora_loader_args=None):  
+    REPLACE_MODULES = replace_modules
+
+    params = None
+    negation = None
+    injector_args = lora_loader_args
+
+    params, negation = inject_trainable_lora(**injector_args)
+    
+    success_inject = True
+    for _up, _down in extract_lora_ups_down(model, target_replace_module=REPLACE_MODULES):
+
+        if not all(x is not None for x in [_up, _down]): success_inject = False
+
+    if success_inject:
+        print(f"Lora successfully injected into {model.__class__.__name__}.")
+    else: 
+        print(f'Fail to inject Lora into {model.__class__.__name__}')
+        exit(-1)
+
+    return params, negation
+
+
+def add_lora_to_model(model, dropout=0.0, lora_rank=16, 
+        scale=0, replace_modules: str = ["Transformer2DModel"]):
+    '''
+    replace_modules needs to be fixed to the proper block
+    '''
+    params = None
+    negation = None
+
+    lora_loader_args = lora_args.copy()
+    lora_loader_args.update({
+        "model": model,
+        "loras": None,
+        "target_replace_module": replace_modules,
+        "lora_rank": lora_rank,
+        "dropout": dropout,
+        "scale": scale
+    })
+
+    params, negation = do_lora_injection(model, replace_modules, lora_loader_args=lora_loader_args)
+
+    params = model if params is None else params
+    return params, negation
+

models/mm_attention.py

@@ -0,0 +1,1254 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, List, Optional, Tuple
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.utils import deprecate, logging
+from diffusers.utils.torch_utils import maybe_allow_in_graph
+from diffusers.models.activations import GEGLU, GELU, ApproximateGELU, FP32SiLU, LinearActivation, SwiGLU
+from models.attention_processor import Attention, JointAttnProcessor2_0
+from diffusers.models.embeddings import SinusoidalPositionalEmbedding
+from diffusers.models.normalization import AdaLayerNorm, AdaLayerNormContinuous, AdaLayerNormZero, RMSNorm, SD35AdaLayerNormZeroX
+
+
+logger = logging.get_logger(__name__)
+
+
+def _chunked_feed_forward(ff: nn.Module, hidden_states: torch.Tensor, chunk_dim: int, chunk_size: int):
+    # "feed_forward_chunk_size" can be used to save memory
+    if hidden_states.shape[chunk_dim] % chunk_size != 0:
+        raise ValueError(
+            f"`hidden_states` dimension to be chunked: {hidden_states.shape[chunk_dim]} has to be divisible by chunk size: {chunk_size}. Make sure to set an appropriate `chunk_size` when calling `unet.enable_forward_chunking`."
+        )
+
+    num_chunks = hidden_states.shape[chunk_dim] // chunk_size
+    ff_output = torch.cat(
+        [ff(hid_slice) for hid_slice in hidden_states.chunk(num_chunks, dim=chunk_dim)],
+        dim=chunk_dim,
+    )
+    return ff_output
+
+
+@maybe_allow_in_graph
+class GatedSelfAttentionDense(nn.Module):
+    r"""
+    A gated self-attention dense layer that combines visual features and object features.
+
+    Parameters:
+        query_dim (`int`): The number of channels in the query.
+        context_dim (`int`): The number of channels in the context.
+        n_heads (`int`): The number of heads to use for attention.
+        d_head (`int`): The number of channels in each head.
+    """
+
+    def __init__(self, query_dim: int, context_dim: int, n_heads: int, d_head: int):
+        super().__init__()
+
+        # we need a linear projection since we need cat visual feature and obj feature
+        self.linear = nn.Linear(context_dim, query_dim)
+
+        self.attn = Attention(query_dim=query_dim, heads=n_heads, dim_head=d_head)
+        self.ff = FeedForward(query_dim, activation_fn="geglu")
+
+        self.norm1 = nn.LayerNorm(query_dim)
+        self.norm2 = nn.LayerNorm(query_dim)
+
+        self.register_parameter("alpha_attn", nn.Parameter(torch.tensor(0.0)))
+        self.register_parameter("alpha_dense", nn.Parameter(torch.tensor(0.0)))
+
+        self.enabled = True
+
+    def forward(self, x: torch.Tensor, objs: torch.Tensor) -> torch.Tensor:
+        if not self.enabled:
+            return x
+
+        n_visual = x.shape[1]
+        objs = self.linear(objs)
+
+        x = x + self.alpha_attn.tanh() * self.attn(self.norm1(torch.cat([x, objs], dim=1)))[:, :n_visual, :]
+        x = x + self.alpha_dense.tanh() * self.ff(self.norm2(x))
+
+        return x
+
+
+@maybe_allow_in_graph
+class JointTransformerBlock(nn.Module):
+    r"""
+    A Transformer block following the MMDiT architecture, introduced in Stable Diffusion 3.
+
+    Reference: https://arxiv.org/abs/2403.03206
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        context_pre_only (`bool`): Boolean to determine if we should add some blocks associated with the
+            processing of `context` conditions.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        context_pre_only: bool = False,
+        qk_norm: Optional[str] = None,
+        use_dual_attention: bool = False,
+    ):
+        super().__init__()
+
+        self.use_dual_attention = use_dual_attention
+        self.context_pre_only = context_pre_only
+        context_norm_type = "ada_norm_continous" if context_pre_only else "ada_norm_zero"
+
+        if use_dual_attention:
+            self.norm1 = SD35AdaLayerNormZeroX(dim)
+        else:
+            self.norm1 = AdaLayerNormZero(dim)
+
+        if context_norm_type == "ada_norm_continous":
+            self.norm1_context = AdaLayerNormContinuous(
+                dim, dim, elementwise_affine=False, eps=1e-6, bias=True, norm_type="layer_norm"
+            )
+        elif context_norm_type == "ada_norm_zero":
+            self.norm1_context = AdaLayerNormZero(dim)
+        else:
+            raise ValueError(
+                f"Unknown context_norm_type: {context_norm_type}, currently only support `ada_norm_continous`, `ada_norm_zero`"
+            )
+
+        if hasattr(F, "scaled_dot_product_attention"):
+            processor = JointAttnProcessor2_0()
+        else:
+            raise ValueError(
+                "The current PyTorch version does not support the `scaled_dot_product_attention` function."
+            )
+
+        self.attn = Attention(
+            query_dim=dim,
+            cross_attention_dim=None,
+            added_kv_proj_dim=dim,
+            dim_head=attention_head_dim,
+            heads=num_attention_heads,
+            out_dim=dim,
+            context_pre_only=context_pre_only,
+            bias=True,
+            processor=processor,
+            qk_norm=qk_norm,
+            eps=1e-6,
+        )
+
+        if use_dual_attention:
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=None,
+                dim_head=attention_head_dim,
+                heads=num_attention_heads,
+                out_dim=dim,
+                bias=True,
+                processor=processor,
+                qk_norm=qk_norm,
+                eps=1e-6,
+            )
+        else:
+            self.attn2 = None
+
+        self.norm2 = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+        self.ff = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+
+        if not context_pre_only:
+            self.norm2_context = nn.LayerNorm(dim, elementwise_affine=False, eps=1e-6)
+            self.ff_context = FeedForward(dim=dim, dim_out=dim, activation_fn="gelu-approximate")
+        else:
+            self.norm2_context = None
+            self.ff_context = None
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    # Copied from diffusers.models.attention.BasicTransformerBlock.set_chunk_feed_forward
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.FloatTensor,
+        encoder_hidden_states: torch.FloatTensor,
+        temb: torch.FloatTensor,
+        joint_attention_kwargs: Optional[Dict[str, Any]] = None,
+    ):
+        joint_attention_kwargs = joint_attention_kwargs or {}
+        if self.use_dual_attention:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp, norm_hidden_states2, gate_msa2 = self.norm1(
+                hidden_states, emb=temb
+            )
+        else:
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(hidden_states, emb=temb)
+
+        if self.context_pre_only:
+            norm_encoder_hidden_states = self.norm1_context(encoder_hidden_states, temb)
+        else:
+            norm_encoder_hidden_states, c_gate_msa, c_shift_mlp, c_scale_mlp, c_gate_mlp = self.norm1_context(
+                encoder_hidden_states, emb=temb
+            )
+
+        # Attention.
+        attn_output, context_attn_output = self.attn(
+            hidden_states=norm_hidden_states,
+            encoder_hidden_states=norm_encoder_hidden_states,
+            **joint_attention_kwargs,
+        )
+
+        # Process attention outputs for the `hidden_states`.
+        attn_output = gate_msa.unsqueeze(1) * attn_output
+        hidden_states = hidden_states + attn_output
+
+        if self.use_dual_attention:
+            attn_output2 = self.attn2(hidden_states=norm_hidden_states2, **joint_attention_kwargs)
+            attn_output2 = gate_msa2.unsqueeze(1) * attn_output2
+            hidden_states = hidden_states + attn_output2
+
+        norm_hidden_states = self.norm2(hidden_states)
+        norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+        ff_output = gate_mlp.unsqueeze(1) * ff_output
+
+        hidden_states = hidden_states + ff_output
+
+        # Process attention outputs for the `encoder_hidden_states`.
+        if self.context_pre_only:
+            encoder_hidden_states = None
+        else:
+            context_attn_output = c_gate_msa.unsqueeze(1) * context_attn_output
+            encoder_hidden_states = encoder_hidden_states + context_attn_output
+
+            norm_encoder_hidden_states = self.norm2_context(encoder_hidden_states)
+            norm_encoder_hidden_states = norm_encoder_hidden_states * (1 + c_scale_mlp[:, None]) + c_shift_mlp[:, None]
+            if self._chunk_size is not None:
+                # "feed_forward_chunk_size" can be used to save memory
+                context_ff_output = _chunked_feed_forward(
+                    self.ff_context, norm_encoder_hidden_states, self._chunk_dim, self._chunk_size
+                )
+            else:
+                context_ff_output = self.ff_context(norm_encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states + c_gate_mlp.unsqueeze(1) * context_ff_output
+
+        return encoder_hidden_states, hidden_states
+
+
+@maybe_allow_in_graph
+class BasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        num_embeds_ada_norm (:
+            obj: `int`, *optional*): The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (:
+            obj: `bool`, *optional*, defaults to `False`): Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, *optional*):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, *optional*):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, *optional*):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, *optional*, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, *optional*, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` *optional*, defaults to False):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, *optional*, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*, defaults to `None`):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        attention_type: str = "default",
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ada_norm_continous_conditioning_embedding_dim: Optional[int] = None,
+        ada_norm_bias: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.double_self_attention = double_self_attention
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.only_cross_attention = only_cross_attention
+
+        # We keep these boolean flags for backward-compatibility.
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        if norm_type == "ada_norm":
+            self.norm1 = AdaLayerNorm(dim, num_embeds_ada_norm)
+        elif norm_type == "ada_norm_zero":
+            self.norm1 = AdaLayerNormZero(dim, num_embeds_ada_norm)
+        elif norm_type == "ada_norm_continuous":
+            self.norm1 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "rms_norm",
+            )
+        else:
+            self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            if norm_type == "ada_norm":
+                self.norm2 = AdaLayerNorm(dim, num_embeds_ada_norm)
+            elif norm_type == "ada_norm_continuous":
+                self.norm2 = AdaLayerNormContinuous(
+                    dim,
+                    ada_norm_continous_conditioning_embedding_dim,
+                    norm_elementwise_affine,
+                    norm_eps,
+                    ada_norm_bias,
+                    "rms_norm",
+                )
+            else:
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            if norm_type == "ada_norm_single":  # For Latte
+                self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+            else:
+                self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        if norm_type == "ada_norm_continuous":
+            self.norm3 = AdaLayerNormContinuous(
+                dim,
+                ada_norm_continous_conditioning_embedding_dim,
+                norm_elementwise_affine,
+                norm_eps,
+                ada_norm_bias,
+                "layer_norm",
+            )
+
+        elif norm_type in ["ada_norm_zero", "ada_norm", "layer_norm"]:
+            self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+        elif norm_type == "layer_norm_i2vgen":
+            self.norm3 = None
+
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        # 4. Fuser
+        if attention_type == "gated" or attention_type == "gated-text-image":
+            self.fuser = GatedSelfAttentionDense(dim, cross_attention_dim, num_attention_heads, attention_head_dim)
+
+        # 5. Scale-shift for PixArt-Alpha.
+        if norm_type == "ada_norm_single":
+            self.scale_shift_table = nn.Parameter(torch.randn(6, dim) / dim**0.5)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_lora_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+    ) -> torch.Tensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        if self.norm_type == "ada_norm":
+            norm_hidden_states = self.norm1(hidden_states, timestep)
+        elif self.norm_type == "ada_norm_zero":
+            norm_hidden_states, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.norm1(
+                hidden_states, timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+        elif self.norm_type in ["layer_norm", "layer_norm_i2vgen"]:
+            norm_hidden_states = self.norm1(hidden_states)
+        elif self.norm_type == "ada_norm_continuous":
+            norm_hidden_states = self.norm1(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif self.norm_type == "ada_norm_single":
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                self.scale_shift_table[None] + timestep.reshape(batch_size, 6, -1)
+            ).chunk(6, dim=1)
+            norm_hidden_states = self.norm1(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_msa) + shift_msa
+        else:
+            raise ValueError("Incorrect norm used")
+
+        if self.pos_embed is not None:
+            norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+        # 1. Prepare GLIGEN inputs
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+        gligen_kwargs = cross_attention_kwargs.pop("gligen", None)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+            attention_mask=attention_mask,
+            **cross_attention_kwargs,
+        )
+
+        if self.norm_type == "ada_norm_zero":
+            attn_output = gate_msa.unsqueeze(1) * attn_output
+        elif self.norm_type == "ada_norm_single":
+            attn_output = gate_msa * attn_output
+
+        hidden_states = attn_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        # 1.2 GLIGEN Control
+        if gligen_kwargs is not None:
+            hidden_states = self.fuser(hidden_states, gligen_kwargs["objs"])
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            if self.norm_type == "ada_norm":
+                norm_hidden_states = self.norm2(hidden_states, timestep)
+            elif self.norm_type in ["ada_norm_zero", "layer_norm", "layer_norm_i2vgen"]:
+                norm_hidden_states = self.norm2(hidden_states)
+            elif self.norm_type == "ada_norm_single":
+                # For PixArt norm2 isn't applied here:
+                # https://github.com/PixArt-alpha/PixArt-alpha/blob/0f55e922376d8b797edd44d25d0e7464b260dcab/diffusion/model/nets/PixArtMS.py#L70C1-L76C103
+                norm_hidden_states = hidden_states
+            elif self.norm_type == "ada_norm_continuous":
+                norm_hidden_states = self.norm2(hidden_states, added_cond_kwargs["pooled_text_emb"])
+            else:
+                raise ValueError("Incorrect norm")
+
+            if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn2(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states,
+                encoder_lora_states=encoder_lora_states,
+                attention_mask=encoder_attention_mask,
+                **cross_attention_kwargs,
+            )
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        # i2vgen doesn't have this norm 🤷‍♂️
+        if self.norm_type == "ada_norm_continuous":
+            norm_hidden_states = self.norm3(hidden_states, added_cond_kwargs["pooled_text_emb"])
+        elif not self.norm_type == "ada_norm_single":
+            norm_hidden_states = self.norm3(hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp[:, None]) + shift_mlp[:, None]
+
+        if self.norm_type == "ada_norm_single":
+            norm_hidden_states = self.norm2(hidden_states)
+            norm_hidden_states = norm_hidden_states * (1 + scale_mlp) + shift_mlp
+
+        if self._chunk_size is not None:
+            # "feed_forward_chunk_size" can be used to save memory
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.norm_type == "ada_norm_zero":
+            ff_output = gate_mlp.unsqueeze(1) * ff_output
+        elif self.norm_type == "ada_norm_single":
+            ff_output = gate_mlp * ff_output
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class LuminaFeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        hidden_size (`int`):
+            The dimensionality of the hidden layers in the model. This parameter determines the width of the model's
+            hidden representations.
+        intermediate_size (`int`): The intermediate dimension of the feedforward layer.
+        multiple_of (`int`, *optional*): Value to ensure hidden dimension is a multiple
+            of this value.
+        ffn_dim_multiplier (float, *optional*): Custom multiplier for hidden
+            dimension. Defaults to None.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        inner_dim: int,
+        multiple_of: Optional[int] = 256,
+        ffn_dim_multiplier: Optional[float] = None,
+    ):
+        super().__init__()
+        inner_dim = int(2 * inner_dim / 3)
+        # custom hidden_size factor multiplier
+        if ffn_dim_multiplier is not None:
+            inner_dim = int(ffn_dim_multiplier * inner_dim)
+        inner_dim = multiple_of * ((inner_dim + multiple_of - 1) // multiple_of)
+
+        self.linear_1 = nn.Linear(
+            dim,
+            inner_dim,
+            bias=False,
+        )
+        self.linear_2 = nn.Linear(
+            inner_dim,
+            dim,
+            bias=False,
+        )
+        self.linear_3 = nn.Linear(
+            dim,
+            inner_dim,
+            bias=False,
+        )
+        self.silu = FP32SiLU()
+
+    def forward(self, x):
+        return self.linear_2(self.silu(self.linear_1(x)) * self.linear_3(x))
+
+
+@maybe_allow_in_graph
+class TemporalBasicTransformerBlock(nn.Module):
+    r"""
+    A basic Transformer block for video like data.
+
+    Parameters:
+        dim (`int`): The number of channels in the input and output.
+        time_mix_inner_dim (`int`): The number of channels for temporal attention.
+        num_attention_heads (`int`): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`): The number of channels in each head.
+        cross_attention_dim (`int`, *optional*): The size of the encoder_hidden_states vector for cross attention.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        time_mix_inner_dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        cross_attention_dim: Optional[int] = None,
+    ):
+        super().__init__()
+        self.is_res = dim == time_mix_inner_dim
+
+        self.norm_in = nn.LayerNorm(dim)
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.ff_in = FeedForward(
+            dim,
+            dim_out=time_mix_inner_dim,
+            activation_fn="geglu",
+        )
+
+        self.norm1 = nn.LayerNorm(time_mix_inner_dim)
+        self.attn1 = Attention(
+            query_dim=time_mix_inner_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            cross_attention_dim=None,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None:
+            # We currently only use AdaLayerNormZero for self attention where there will only be one attention block.
+            # I.e. the number of returned modulation chunks from AdaLayerZero would not make sense if returned during
+            # the second cross attention block.
+            self.norm2 = nn.LayerNorm(time_mix_inner_dim)
+            self.attn2 = Attention(
+                query_dim=time_mix_inner_dim,
+                cross_attention_dim=cross_attention_dim,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+            )  # is self-attn if encoder_hidden_states is none
+        else:
+            self.norm2 = None
+            self.attn2 = None
+
+        # 3. Feed-forward
+        self.norm3 = nn.LayerNorm(time_mix_inner_dim)
+        self.ff = FeedForward(time_mix_inner_dim, activation_fn="geglu")
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = None
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], **kwargs):
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        # chunk dim should be hardcoded to 1 to have better speed vs. memory trade-off
+        self._chunk_dim = 1
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        num_frames: int,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+    ) -> torch.Tensor:
+        # Notice that normalization is always applied before the real computation in the following blocks.
+        # 0. Self-Attention
+        batch_size = hidden_states.shape[0]
+
+        batch_frames, seq_length, channels = hidden_states.shape
+        batch_size = batch_frames // num_frames
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, num_frames, seq_length, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * seq_length, num_frames, channels)
+
+        residual = hidden_states
+        hidden_states = self.norm_in(hidden_states)
+
+        if self._chunk_size is not None:
+            hidden_states = _chunked_feed_forward(self.ff_in, hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            hidden_states = self.ff_in(hidden_states)
+
+        if self.is_res:
+            hidden_states = hidden_states + residual
+
+        norm_hidden_states = self.norm1(hidden_states)
+        attn_output = self.attn1(norm_hidden_states, encoder_hidden_states=None)
+        hidden_states = attn_output + hidden_states
+
+        # 3. Cross-Attention
+        if self.attn2 is not None:
+            norm_hidden_states = self.norm2(hidden_states)
+            attn_output = self.attn2(norm_hidden_states, encoder_hidden_states=encoder_hidden_states)
+            hidden_states = attn_output + hidden_states
+
+        # 4. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        if self.is_res:
+            hidden_states = ff_output + hidden_states
+        else:
+            hidden_states = ff_output
+
+        hidden_states = hidden_states[None, :].reshape(batch_size, seq_length, num_frames, channels)
+        hidden_states = hidden_states.permute(0, 2, 1, 3)
+        hidden_states = hidden_states.reshape(batch_size * num_frames, seq_length, channels)
+
+        return hidden_states
+
+
+class SkipFFTransformerBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        kv_input_dim: int,
+        kv_input_dim_proj_use_bias: bool,
+        dropout=0.0,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        attention_out_bias: bool = True,
+    ):
+        super().__init__()
+        if kv_input_dim != dim:
+            self.kv_mapper = nn.Linear(kv_input_dim, dim, kv_input_dim_proj_use_bias)
+        else:
+            self.kv_mapper = None
+
+        self.norm1 = RMSNorm(dim, 1e-06)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim,
+            out_bias=attention_out_bias,
+        )
+
+        self.norm2 = RMSNorm(dim, 1e-06)
+
+        self.attn2 = Attention(
+            query_dim=dim,
+            cross_attention_dim=cross_attention_dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            out_bias=attention_out_bias,
+        )
+
+    def forward(self, hidden_states, encoder_hidden_states, cross_attention_kwargs):
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        if self.kv_mapper is not None:
+            encoder_hidden_states = self.kv_mapper(F.silu(encoder_hidden_states))
+
+        norm_hidden_states = self.norm1(hidden_states)
+
+        attn_output = self.attn1(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        norm_hidden_states = self.norm2(hidden_states)
+
+        attn_output = self.attn2(
+            norm_hidden_states,
+            encoder_hidden_states=encoder_hidden_states,
+            **cross_attention_kwargs,
+        )
+
+        hidden_states = attn_output + hidden_states
+
+        return hidden_states
+
+
+@maybe_allow_in_graph
+class FreeNoiseTransformerBlock(nn.Module):
+    r"""
+    A FreeNoise Transformer block.
+
+    Parameters:
+        dim (`int`):
+            The number of channels in the input and output.
+        num_attention_heads (`int`):
+            The number of heads to use for multi-head attention.
+        attention_head_dim (`int`):
+            The number of channels in each head.
+        dropout (`float`, *optional*, defaults to 0.0):
+            The dropout probability to use.
+        cross_attention_dim (`int`, *optional*):
+            The size of the encoder_hidden_states vector for cross attention.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`):
+            Activation function to be used in feed-forward.
+        num_embeds_ada_norm (`int`, *optional*):
+            The number of diffusion steps used during training. See `Transformer2DModel`.
+        attention_bias (`bool`, defaults to `False`):
+            Configure if the attentions should contain a bias parameter.
+        only_cross_attention (`bool`, defaults to `False`):
+            Whether to use only cross-attention layers. In this case two cross attention layers are used.
+        double_self_attention (`bool`, defaults to `False`):
+            Whether to use two self-attention layers. In this case no cross attention layers are used.
+        upcast_attention (`bool`, defaults to `False`):
+            Whether to upcast the attention computation to float32. This is useful for mixed precision training.
+        norm_elementwise_affine (`bool`, defaults to `True`):
+            Whether to use learnable elementwise affine parameters for normalization.
+        norm_type (`str`, defaults to `"layer_norm"`):
+            The normalization layer to use. Can be `"layer_norm"`, `"ada_norm"` or `"ada_norm_zero"`.
+        final_dropout (`bool` defaults to `False`):
+            Whether to apply a final dropout after the last feed-forward layer.
+        attention_type (`str`, defaults to `"default"`):
+            The type of attention to use. Can be `"default"` or `"gated"` or `"gated-text-image"`.
+        positional_embeddings (`str`, *optional*):
+            The type of positional embeddings to apply to.
+        num_positional_embeddings (`int`, *optional*, defaults to `None`):
+            The maximum number of positional embeddings to apply.
+        ff_inner_dim (`int`, *optional*):
+            Hidden dimension of feed-forward MLP.
+        ff_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in feed-forward MLP.
+        attention_out_bias (`bool`, defaults to `True`):
+            Whether or not to use bias in attention output project layer.
+        context_length (`int`, defaults to `16`):
+            The maximum number of frames that the FreeNoise block processes at once.
+        context_stride (`int`, defaults to `4`):
+            The number of frames to be skipped before starting to process a new batch of `context_length` frames.
+        weighting_scheme (`str`, defaults to `"pyramid"`):
+            The weighting scheme to use for weighting averaging of processed latent frames. As described in the
+            Equation 9. of the [FreeNoise](https://arxiv.org/abs/2310.15169) paper, "pyramid" is the default setting
+            used.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        num_attention_heads: int,
+        attention_head_dim: int,
+        dropout: float = 0.0,
+        cross_attention_dim: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        attention_bias: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_elementwise_affine: bool = True,
+        norm_type: str = "layer_norm",
+        norm_eps: float = 1e-5,
+        final_dropout: bool = False,
+        positional_embeddings: Optional[str] = None,
+        num_positional_embeddings: Optional[int] = None,
+        ff_inner_dim: Optional[int] = None,
+        ff_bias: bool = True,
+        attention_out_bias: bool = True,
+        context_length: int = 16,
+        context_stride: int = 4,
+        weighting_scheme: str = "pyramid",
+    ):
+        super().__init__()
+        self.dim = dim
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.dropout = dropout
+        self.cross_attention_dim = cross_attention_dim
+        self.activation_fn = activation_fn
+        self.attention_bias = attention_bias
+        self.double_self_attention = double_self_attention
+        self.norm_elementwise_affine = norm_elementwise_affine
+        self.positional_embeddings = positional_embeddings
+        self.num_positional_embeddings = num_positional_embeddings
+        self.only_cross_attention = only_cross_attention
+
+        self.set_free_noise_properties(context_length, context_stride, weighting_scheme)
+
+        # We keep these boolean flags for backward-compatibility.
+        self.use_ada_layer_norm_zero = (num_embeds_ada_norm is not None) and norm_type == "ada_norm_zero"
+        self.use_ada_layer_norm = (num_embeds_ada_norm is not None) and norm_type == "ada_norm"
+        self.use_ada_layer_norm_single = norm_type == "ada_norm_single"
+        self.use_layer_norm = norm_type == "layer_norm"
+        self.use_ada_layer_norm_continuous = norm_type == "ada_norm_continuous"
+
+        if norm_type in ("ada_norm", "ada_norm_zero") and num_embeds_ada_norm is None:
+            raise ValueError(
+                f"`norm_type` is set to {norm_type}, but `num_embeds_ada_norm` is not defined. Please make sure to"
+                f" define `num_embeds_ada_norm` if setting `norm_type` to {norm_type}."
+            )
+
+        self.norm_type = norm_type
+        self.num_embeds_ada_norm = num_embeds_ada_norm
+
+        if positional_embeddings and (num_positional_embeddings is None):
+            raise ValueError(
+                "If `positional_embedding` type is defined, `num_positition_embeddings` must also be defined."
+            )
+
+        if positional_embeddings == "sinusoidal":
+            self.pos_embed = SinusoidalPositionalEmbedding(dim, max_seq_length=num_positional_embeddings)
+        else:
+            self.pos_embed = None
+
+        # Define 3 blocks. Each block has its own normalization layer.
+        # 1. Self-Attn
+        self.norm1 = nn.LayerNorm(dim, elementwise_affine=norm_elementwise_affine, eps=norm_eps)
+
+        self.attn1 = Attention(
+            query_dim=dim,
+            heads=num_attention_heads,
+            dim_head=attention_head_dim,
+            dropout=dropout,
+            bias=attention_bias,
+            cross_attention_dim=cross_attention_dim if only_cross_attention else None,
+            upcast_attention=upcast_attention,
+            out_bias=attention_out_bias,
+        )
+
+        # 2. Cross-Attn
+        if cross_attention_dim is not None or double_self_attention:
+            self.norm2 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+            self.attn2 = Attention(
+                query_dim=dim,
+                cross_attention_dim=cross_attention_dim if not double_self_attention else None,
+                heads=num_attention_heads,
+                dim_head=attention_head_dim,
+                dropout=dropout,
+                bias=attention_bias,
+                upcast_attention=upcast_attention,
+                out_bias=attention_out_bias,
+            )  # is self-attn if encoder_hidden_states is none
+
+        # 3. Feed-forward
+        self.ff = FeedForward(
+            dim,
+            dropout=dropout,
+            activation_fn=activation_fn,
+            final_dropout=final_dropout,
+            inner_dim=ff_inner_dim,
+            bias=ff_bias,
+        )
+
+        self.norm3 = nn.LayerNorm(dim, norm_eps, norm_elementwise_affine)
+
+        # let chunk size default to None
+        self._chunk_size = None
+        self._chunk_dim = 0
+
+    def _get_frame_indices(self, num_frames: int) -> List[Tuple[int, int]]:
+        frame_indices = []
+        for i in range(0, num_frames - self.context_length + 1, self.context_stride):
+            window_start = i
+            window_end = min(num_frames, i + self.context_length)
+            frame_indices.append((window_start, window_end))
+        return frame_indices
+
+    def _get_frame_weights(self, num_frames: int, weighting_scheme: str = "pyramid") -> List[float]:
+        if weighting_scheme == "flat":
+            weights = [1.0] * num_frames
+
+        elif weighting_scheme == "pyramid":
+            if num_frames % 2 == 0:
+                # num_frames = 4 => [1, 2, 2, 1]
+                mid = num_frames // 2
+                weights = list(range(1, mid + 1))
+                weights = weights + weights[::-1]
+            else:
+                # num_frames = 5 => [1, 2, 3, 2, 1]
+                mid = (num_frames + 1) // 2
+                weights = list(range(1, mid))
+                weights = weights + [mid] + weights[::-1]
+
+        elif weighting_scheme == "delayed_reverse_sawtooth":
+            if num_frames % 2 == 0:
+                # num_frames = 4 => [0.01, 2, 2, 1]
+                mid = num_frames // 2
+                weights = [0.01] * (mid - 1) + [mid]
+                weights = weights + list(range(mid, 0, -1))
+            else:
+                # num_frames = 5 => [0.01, 0.01, 3, 2, 1]
+                mid = (num_frames + 1) // 2
+                weights = [0.01] * mid
+                weights = weights + list(range(mid, 0, -1))
+        else:
+            raise ValueError(f"Unsupported value for weighting_scheme={weighting_scheme}")
+
+        return weights
+
+    def set_free_noise_properties(
+        self, context_length: int, context_stride: int, weighting_scheme: str = "pyramid"
+    ) -> None:
+        self.context_length = context_length
+        self.context_stride = context_stride
+        self.weighting_scheme = weighting_scheme
+
+    def set_chunk_feed_forward(self, chunk_size: Optional[int], dim: int = 0) -> None:
+        # Sets chunk feed-forward
+        self._chunk_size = chunk_size
+        self._chunk_dim = dim
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        *args,
+        **kwargs,
+    ) -> torch.Tensor:
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+
+        cross_attention_kwargs = cross_attention_kwargs.copy() if cross_attention_kwargs is not None else {}
+
+        # hidden_states: [B x H x W, F, C]
+        device = hidden_states.device
+        dtype = hidden_states.dtype
+
+        num_frames = hidden_states.size(1)
+        frame_indices = self._get_frame_indices(num_frames)
+        frame_weights = self._get_frame_weights(self.context_length, self.weighting_scheme)
+        frame_weights = torch.tensor(frame_weights, device=device, dtype=dtype).unsqueeze(0).unsqueeze(-1)
+        is_last_frame_batch_complete = frame_indices[-1][1] == num_frames
+
+        # Handle out-of-bounds case if num_frames isn't perfectly divisible by context_length
+        # For example, num_frames=25, context_length=16, context_stride=4, then we expect the ranges:
+        #    [(0, 16), (4, 20), (8, 24), (10, 26)]
+        if not is_last_frame_batch_complete:
+            if num_frames < self.context_length:
+                raise ValueError(f"Expected {num_frames=} to be greater or equal than {self.context_length=}")
+            last_frame_batch_length = num_frames - frame_indices[-1][1]
+            frame_indices.append((num_frames - self.context_length, num_frames))
+
+        num_times_accumulated = torch.zeros((1, num_frames, 1), device=device)
+        accumulated_values = torch.zeros_like(hidden_states)
+
+        for i, (frame_start, frame_end) in enumerate(frame_indices):
+            # The reason for slicing here is to ensure that if (frame_end - frame_start) is to handle
+            # cases like frame_indices=[(0, 16), (16, 20)], if the user provided a video with 19 frames, or
+            # essentially a non-multiple of `context_length`.
+            weights = torch.ones_like(num_times_accumulated[:, frame_start:frame_end])
+            weights *= frame_weights
+
+            hidden_states_chunk = hidden_states[:, frame_start:frame_end]
+
+            # Notice that normalization is always applied before the real computation in the following blocks.
+            # 1. Self-Attention
+            norm_hidden_states = self.norm1(hidden_states_chunk)
+
+            if self.pos_embed is not None:
+                norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+            attn_output = self.attn1(
+                norm_hidden_states,
+                encoder_hidden_states=encoder_hidden_states if self.only_cross_attention else None,
+                attention_mask=attention_mask,
+                **cross_attention_kwargs,
+            )
+
+            hidden_states_chunk = attn_output + hidden_states_chunk
+            if hidden_states_chunk.ndim == 4:
+                hidden_states_chunk = hidden_states_chunk.squeeze(1)
+
+            # 2. Cross-Attention
+            if self.attn2 is not None:
+                norm_hidden_states = self.norm2(hidden_states_chunk)
+
+                if self.pos_embed is not None and self.norm_type != "ada_norm_single":
+                    norm_hidden_states = self.pos_embed(norm_hidden_states)
+
+                attn_output = self.attn2(
+                    norm_hidden_states,
+                    encoder_hidden_states=encoder_hidden_states,
+                    attention_mask=encoder_attention_mask,
+                    **cross_attention_kwargs,
+                )
+                hidden_states_chunk = attn_output + hidden_states_chunk
+
+            if i == len(frame_indices) - 1 and not is_last_frame_batch_complete:
+                accumulated_values[:, -last_frame_batch_length:] += (
+                    hidden_states_chunk[:, -last_frame_batch_length:] * weights[:, -last_frame_batch_length:]
+                )
+                num_times_accumulated[:, -last_frame_batch_length:] += weights[:, -last_frame_batch_length]
+            else:
+                accumulated_values[:, frame_start:frame_end] += hidden_states_chunk * weights
+                num_times_accumulated[:, frame_start:frame_end] += weights
+
+        # TODO(aryan): Maybe this could be done in a better way.
+        #
+        # Previously, this was:
+        # hidden_states = torch.where(
+        #    num_times_accumulated > 0, accumulated_values / num_times_accumulated, accumulated_values
+        # )
+        #
+        # The reasoning for the change here is `torch.where` became a bottleneck at some point when golfing memory
+        # spikes. It is particularly noticeable when the number of frames is high. My understanding is that this comes
+        # from tensors being copied - which is why we resort to spliting and concatenating here. I've not particularly
+        # looked into this deeply because other memory optimizations led to more pronounced reductions.
+        hidden_states = torch.cat(
+            [
+                torch.where(num_times_split > 0, accumulated_split / num_times_split, accumulated_split)
+                for accumulated_split, num_times_split in zip(
+                    accumulated_values.split(self.context_length, dim=1),
+                    num_times_accumulated.split(self.context_length, dim=1),
+                )
+            ],
+            dim=1,
+        ).to(dtype)
+
+        # 3. Feed-forward
+        norm_hidden_states = self.norm3(hidden_states)
+
+        if self._chunk_size is not None:
+            ff_output = _chunked_feed_forward(self.ff, norm_hidden_states, self._chunk_dim, self._chunk_size)
+        else:
+            ff_output = self.ff(norm_hidden_states)
+
+        hidden_states = ff_output + hidden_states
+        if hidden_states.ndim == 4:
+            hidden_states = hidden_states.squeeze(1)
+
+        return hidden_states
+
+
+class FeedForward(nn.Module):
+    r"""
+    A feed-forward layer.
+
+    Parameters:
+        dim (`int`): The number of channels in the input.
+        dim_out (`int`, *optional*): The number of channels in the output. If not given, defaults to `dim`.
+        mult (`int`, *optional*, defaults to 4): The multiplier to use for the hidden dimension.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to be used in feed-forward.
+        final_dropout (`bool` *optional*, defaults to False): Apply a final dropout.
+        bias (`bool`, defaults to True): Whether to use a bias in the linear layer.
+    """
+
+    def __init__(
+        self,
+        dim: int,
+        dim_out: Optional[int] = None,
+        mult: int = 4,
+        dropout: float = 0.0,
+        activation_fn: str = "geglu",
+        final_dropout: bool = False,
+        inner_dim=None,
+        bias: bool = True,
+    ):
+        super().__init__()
+        if inner_dim is None:
+            inner_dim = int(dim * mult)
+        dim_out = dim_out if dim_out is not None else dim
+
+        if activation_fn == "gelu":
+            act_fn = GELU(dim, inner_dim, bias=bias)
+        if activation_fn == "gelu-approximate":
+            act_fn = GELU(dim, inner_dim, approximate="tanh", bias=bias)
+        elif activation_fn == "geglu":
+            act_fn = GEGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "geglu-approximate":
+            act_fn = ApproximateGELU(dim, inner_dim, bias=bias)
+        elif activation_fn == "swiglu":
+            act_fn = SwiGLU(dim, inner_dim, bias=bias)
+        elif activation_fn == "linear-silu":
+            act_fn = LinearActivation(dim, inner_dim, bias=bias, activation="silu")
+
+        self.net = nn.ModuleList([])
+        # project in
+        self.net.append(act_fn)
+        # project dropout
+        self.net.append(nn.Dropout(dropout))
+        # project out
+        self.net.append(nn.Linear(inner_dim, dim_out, bias=bias))
+        # FF as used in Vision Transformer, MLP-Mixer, etc. have a final dropout
+        if final_dropout:
+            self.net.append(nn.Dropout(dropout))
+
+    def forward(self, hidden_states: torch.Tensor, *args, **kwargs) -> torch.Tensor:
+        if len(args) > 0 or kwargs.get("scale", None) is not None:
+            deprecation_message = "The `scale` argument is deprecated and will be ignored. Please remove it, as passing it will raise an error in the future. `scale` should directly be passed while calling the underlying pipeline component i.e., via `cross_attention_kwargs`."
+            deprecate("scale", "1.0.0", deprecation_message)
+        for module in self.net:
+            hidden_states = module(hidden_states)
+        return hidden_states

models/transformers_2d.py

@@ -0,0 +1,569 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from typing import Any, Dict, Optional
+
+import torch
+import torch.nn.functional as F
+from torch import nn
+
+from diffusers.configuration_utils import LegacyConfigMixin, register_to_config
+from diffusers.utils import deprecate, is_torch_version, logging
+from models.mm_attention import BasicTransformerBlock
+from diffusers.models.embeddings import ImagePositionalEmbeddings, PatchEmbed, PixArtAlphaTextProjection
+from diffusers.models.modeling_outputs import Transformer2DModelOutput
+from diffusers.models.modeling_utils import LegacyModelMixin
+from diffusers.models.normalization import AdaLayerNormSingle
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+class Transformer2DModelOutput(Transformer2DModelOutput):
+    def __init__(self, *args, **kwargs):
+        deprecation_message = "Importing `Transformer2DModelOutput` from `diffusers.models.transformer_2d` is deprecated and this will be removed in a future version. Please use `from diffusers.models.modeling_outputs import Transformer2DModelOutput`, instead."
+        deprecate("Transformer2DModelOutput", "1.0.0", deprecation_message)
+        super().__init__(*args, **kwargs)
+
+
+class Transformer2DModel(LegacyModelMixin, LegacyConfigMixin):
+    """
+    A 2D Transformer model for image-like data.
+
+    Parameters:
+        num_attention_heads (`int`, *optional*, defaults to 16): The number of heads to use for multi-head attention.
+        attention_head_dim (`int`, *optional*, defaults to 88): The number of channels in each head.
+        in_channels (`int`, *optional*):
+            The number of channels in the input and output (specify if the input is **continuous**).
+        num_layers (`int`, *optional*, defaults to 1): The number of layers of Transformer blocks to use.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        cross_attention_dim (`int`, *optional*): The number of `encoder_hidden_states` dimensions to use.
+        sample_size (`int`, *optional*): The width of the latent images (specify if the input is **discrete**).
+            This is fixed during training since it is used to learn a number of position embeddings.
+        num_vector_embeds (`int`, *optional*):
+            The number of classes of the vector embeddings of the latent pixels (specify if the input is **discrete**).
+            Includes the class for the masked latent pixel.
+        activation_fn (`str`, *optional*, defaults to `"geglu"`): Activation function to use in feed-forward.
+        num_embeds_ada_norm ( `int`, *optional*):
+            The number of diffusion steps used during training. Pass if at least one of the norm_layers is
+            `AdaLayerNorm`. This is fixed during training since it is used to learn a number of embeddings that are
+            added to the hidden states.
+
+            During inference, you can denoise for up to but not more steps than `num_embeds_ada_norm`.
+        attention_bias (`bool`, *optional*):
+            Configure if the `TransformerBlocks` attention should contain a bias parameter.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock"]
+
+    @register_to_config
+    def __init__(
+        self,
+        num_attention_heads: int = 16,
+        attention_head_dim: int = 88,
+        in_channels: Optional[int] = None,
+        out_channels: Optional[int] = None,
+        num_layers: int = 1,
+        dropout: float = 0.0,
+        norm_num_groups: int = 32,
+        cross_attention_dim: Optional[int] = None,
+        attention_bias: bool = False,
+        sample_size: Optional[int] = None,
+        num_vector_embeds: Optional[int] = None,
+        patch_size: Optional[int] = None,
+        activation_fn: str = "geglu",
+        num_embeds_ada_norm: Optional[int] = None,
+        use_linear_projection: bool = False,
+        only_cross_attention: bool = False,
+        double_self_attention: bool = False,
+        upcast_attention: bool = False,
+        norm_type: str = "layer_norm",  # 'layer_norm', 'ada_norm', 'ada_norm_zero', 'ada_norm_single', 'ada_norm_continuous', 'layer_norm_i2vgen'
+        norm_elementwise_affine: bool = True,
+        norm_eps: float = 1e-5,
+        attention_type: str = "default",
+        caption_channels: int = None,
+        interpolation_scale: float = None,
+        use_additional_conditions: Optional[bool] = None,
+    ):
+        super().__init__()
+
+        # Validate inputs.
+        if patch_size is not None:
+            if norm_type not in ["ada_norm", "ada_norm_zero", "ada_norm_single"]:
+                raise NotImplementedError(
+                    f"Forward pass is not implemented when `patch_size` is not None and `norm_type` is '{norm_type}'."
+                )
+            elif norm_type in ["ada_norm", "ada_norm_zero"] and num_embeds_ada_norm is None:
+                raise ValueError(
+                    f"When using a `patch_size` and this `norm_type` ({norm_type}), `num_embeds_ada_norm` cannot be None."
+                )
+
+        # 1. Transformer2DModel can process both standard continuous images of shape `(batch_size, num_channels, width, height)` as well as quantized image embeddings of shape `(batch_size, num_image_vectors)`
+        # Define whether input is continuous or discrete depending on configuration
+        self.is_input_continuous = (in_channels is not None) and (patch_size is None)
+        self.is_input_vectorized = num_vector_embeds is not None
+        self.is_input_patches = in_channels is not None and patch_size is not None
+
+        if self.is_input_continuous and self.is_input_vectorized:
+            raise ValueError(
+                f"Cannot define both `in_channels`: {in_channels} and `num_vector_embeds`: {num_vector_embeds}. Make"
+                " sure that either `in_channels` or `num_vector_embeds` is None."
+            )
+        elif self.is_input_vectorized and self.is_input_patches:
+            raise ValueError(
+                f"Cannot define both `num_vector_embeds`: {num_vector_embeds} and `patch_size`: {patch_size}. Make"
+                " sure that either `num_vector_embeds` or `num_patches` is None."
+            )
+        elif not self.is_input_continuous and not self.is_input_vectorized and not self.is_input_patches:
+            raise ValueError(
+                f"Has to define `in_channels`: {in_channels}, `num_vector_embeds`: {num_vector_embeds}, or patch_size:"
+                f" {patch_size}. Make sure that `in_channels`, `num_vector_embeds` or `num_patches` is not None."
+            )
+
+        if norm_type == "layer_norm" and num_embeds_ada_norm is not None:
+            deprecation_message = (
+                f"The configuration file of this model: {self.__class__} is outdated. `norm_type` is either not set or"
+                " incorrectly set to `'layer_norm'`. Make sure to set `norm_type` to `'ada_norm'` in the config."
+                " Please make sure to update the config accordingly as leaving `norm_type` might led to incorrect"
+                " results in future versions. If you have downloaded this checkpoint from the Hugging Face Hub, it"
+                " would be very nice if you could open a Pull request for the `transformer/config.json` file"
+            )
+            deprecate("norm_type!=num_embeds_ada_norm", "1.0.0", deprecation_message, standard_warn=False)
+            norm_type = "ada_norm"
+
+        # Set some common variables used across the board.
+        self.use_linear_projection = use_linear_projection
+        self.interpolation_scale = interpolation_scale
+        self.caption_channels = caption_channels
+        self.num_attention_heads = num_attention_heads
+        self.attention_head_dim = attention_head_dim
+        self.inner_dim = self.config.num_attention_heads * self.config.attention_head_dim
+        self.in_channels = in_channels
+        self.out_channels = in_channels if out_channels is None else out_channels
+        self.gradient_checkpointing = False
+
+        if use_additional_conditions is None:
+            if norm_type == "ada_norm_single" and sample_size == 128:
+                use_additional_conditions = True
+            else:
+                use_additional_conditions = False
+        self.use_additional_conditions = use_additional_conditions
+
+        # 2. Initialize the right blocks.
+        # These functions follow a common structure:
+        # a. Initialize the input blocks. b. Initialize the transformer blocks.
+        # c. Initialize the output blocks and other projection blocks when necessary.
+        if self.is_input_continuous:
+            self._init_continuous_input(norm_type=norm_type)
+        elif self.is_input_vectorized:
+            self._init_vectorized_inputs(norm_type=norm_type)
+        elif self.is_input_patches:
+            self._init_patched_inputs(norm_type=norm_type)
+
+    def _init_continuous_input(self, norm_type):
+        self.norm = torch.nn.GroupNorm(
+            num_groups=self.config.norm_num_groups, num_channels=self.in_channels, eps=1e-6, affine=True
+        )
+        if self.use_linear_projection:
+            self.proj_in = torch.nn.Linear(self.in_channels, self.inner_dim)
+        else:
+            self.proj_in = torch.nn.Conv2d(self.in_channels, self.inner_dim, kernel_size=1, stride=1, padding=0)
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    only_cross_attention=self.config.only_cross_attention,
+                    double_self_attention=self.config.double_self_attention,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        if self.use_linear_projection:
+            self.proj_out = torch.nn.Linear(self.inner_dim, self.out_channels)
+        else:
+            self.proj_out = torch.nn.Conv2d(self.inner_dim, self.out_channels, kernel_size=1, stride=1, padding=0)
+
+    def _init_vectorized_inputs(self, norm_type):
+        assert self.config.sample_size is not None, "Transformer2DModel over discrete input must provide sample_size"
+        assert (
+            self.config.num_vector_embeds is not None
+        ), "Transformer2DModel over discrete input must provide num_embed"
+
+        self.height = self.config.sample_size
+        self.width = self.config.sample_size
+        self.num_latent_pixels = self.height * self.width
+
+        self.latent_image_embedding = ImagePositionalEmbeddings(
+            num_embed=self.config.num_vector_embeds, embed_dim=self.inner_dim, height=self.height, width=self.width
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    only_cross_attention=self.config.only_cross_attention,
+                    double_self_attention=self.config.double_self_attention,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        self.norm_out = nn.LayerNorm(self.inner_dim)
+        self.out = nn.Linear(self.inner_dim, self.config.num_vector_embeds - 1)
+
+    def _init_patched_inputs(self, norm_type):
+        assert self.config.sample_size is not None, "Transformer2DModel over patched input must provide sample_size"
+
+        self.height = self.config.sample_size
+        self.width = self.config.sample_size
+
+        self.patch_size = self.config.patch_size
+        interpolation_scale = (
+            self.config.interpolation_scale
+            if self.config.interpolation_scale is not None
+            else max(self.config.sample_size // 64, 1)
+        )
+        self.pos_embed = PatchEmbed(
+            height=self.config.sample_size,
+            width=self.config.sample_size,
+            patch_size=self.config.patch_size,
+            in_channels=self.in_channels,
+            embed_dim=self.inner_dim,
+            interpolation_scale=interpolation_scale,
+        )
+
+        self.transformer_blocks = nn.ModuleList(
+            [
+                BasicTransformerBlock(
+                    self.inner_dim,
+                    self.config.num_attention_heads,
+                    self.config.attention_head_dim,
+                    dropout=self.config.dropout,
+                    cross_attention_dim=self.config.cross_attention_dim,
+                    activation_fn=self.config.activation_fn,
+                    num_embeds_ada_norm=self.config.num_embeds_ada_norm,
+                    attention_bias=self.config.attention_bias,
+                    only_cross_attention=self.config.only_cross_attention,
+                    double_self_attention=self.config.double_self_attention,
+                    upcast_attention=self.config.upcast_attention,
+                    norm_type=norm_type,
+                    norm_elementwise_affine=self.config.norm_elementwise_affine,
+                    norm_eps=self.config.norm_eps,
+                    attention_type=self.config.attention_type,
+                )
+                for _ in range(self.config.num_layers)
+            ]
+        )
+
+        if self.config.norm_type != "ada_norm_single":
+            self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+            self.proj_out_1 = nn.Linear(self.inner_dim, 2 * self.inner_dim)
+            self.proj_out_2 = nn.Linear(
+                self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels
+            )
+        elif self.config.norm_type == "ada_norm_single":
+            self.norm_out = nn.LayerNorm(self.inner_dim, elementwise_affine=False, eps=1e-6)
+            self.scale_shift_table = nn.Parameter(torch.randn(2, self.inner_dim) / self.inner_dim**0.5)
+            self.proj_out = nn.Linear(
+                self.inner_dim, self.config.patch_size * self.config.patch_size * self.out_channels
+            )
+
+        # PixArt-Alpha blocks.
+        self.adaln_single = None
+        if self.config.norm_type == "ada_norm_single":
+            # TODO(Sayak, PVP) clean this, for now we use sample size to determine whether to use
+            # additional conditions until we find better name
+            self.adaln_single = AdaLayerNormSingle(
+                self.inner_dim, use_additional_conditions=self.use_additional_conditions
+            )
+
+        self.caption_projection = None
+        if self.caption_channels is not None:
+            self.caption_projection = PixArtAlphaTextProjection(
+                in_features=self.caption_channels, hidden_size=self.inner_dim
+            )
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        encoder_hidden_states: Optional[torch.Tensor] = None,
+        encoder_lora_states: Optional[torch.Tensor] = None,
+        timestep: Optional[torch.LongTensor] = None,
+        added_cond_kwargs: Dict[str, torch.Tensor] = None,
+        class_labels: Optional[torch.LongTensor] = None,
+        cross_attention_kwargs: Dict[str, Any] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ):
+        """
+        The [`Transformer2DModel`] forward method.
+
+        Args:
+            hidden_states (`torch.LongTensor` of shape `(batch size, num latent pixels)` if discrete, `torch.Tensor` of shape `(batch size, channel, height, width)` if continuous):
+                Input `hidden_states`.
+            encoder_hidden_states ( `torch.Tensor` of shape `(batch size, sequence len, embed dims)`, *optional*):
+                Conditional embeddings for cross attention layer. If not given, cross-attention defaults to
+                self-attention.
+            timestep ( `torch.LongTensor`, *optional*):
+                Used to indicate denoising step. Optional timestep to be applied as an embedding in `AdaLayerNorm`.
+            class_labels ( `torch.LongTensor` of shape `(batch size, num classes)`, *optional*):
+                Used to indicate class labels conditioning. Optional class labels to be applied as an embedding in
+                `AdaLayerZeroNorm`.
+            cross_attention_kwargs ( `Dict[str, Any]`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            attention_mask ( `torch.Tensor`, *optional*):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            encoder_attention_mask ( `torch.Tensor`, *optional*):
+                Cross-attention mask applied to `encoder_hidden_states`. Two formats supported:
+
+                    * Mask `(batch, sequence_length)` True = keep, False = discard.
+                    * Bias `(batch, 1, sequence_length)` 0 = keep, -10000 = discard.
+
+                If `ndim == 2`: will be interpreted as a mask, then converted into a bias consistent with the format
+                above. This bias will be added to the cross-attention scores.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            If `return_dict` is True, an [`~models.transformers.transformer_2d.Transformer2DModelOutput`] is returned,
+            otherwise a `tuple` where the first element is the sample tensor.
+        """
+        if cross_attention_kwargs is not None:
+            if cross_attention_kwargs.get("scale", None) is not None:
+                logger.warning("Passing `scale` to `cross_attention_kwargs` is deprecated. `scale` will be ignored.")
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension.
+        #   we may have done this conversion already, e.g. if we came here via UNet2DConditionModel#forward.
+        #   we can tell by counting dims; if ndim == 2: it's a mask rather than a bias.
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None and attention_mask.ndim == 2:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(hidden_states.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None and encoder_attention_mask.ndim == 2:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(hidden_states.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 1. Input
+        if self.is_input_continuous:
+            batch_size, _, height, width = hidden_states.shape
+            residual = hidden_states
+            hidden_states, inner_dim = self._operate_on_continuous_inputs(hidden_states)
+        elif self.is_input_vectorized:
+            hidden_states = self.latent_image_embedding(hidden_states)
+        elif self.is_input_patches:
+            height, width = hidden_states.shape[-2] // self.patch_size, hidden_states.shape[-1] // self.patch_size
+            hidden_states, encoder_hidden_states, timestep, embedded_timestep = self._operate_on_patched_inputs(
+                hidden_states, encoder_hidden_states, timestep, added_cond_kwargs
+            )
+
+        # 2. Blocks
+        for block in self.transformer_blocks:
+            if torch.is_grad_enabled() and self.gradient_checkpointing:
+
+                def create_custom_forward(module, return_dict=None):
+                    def custom_forward(*inputs):
+                        if return_dict is not None:
+                            return module(*inputs, return_dict=return_dict)
+                        else:
+                            return module(*inputs)
+
+                    return custom_forward
+
+                ckpt_kwargs: Dict[str, Any] = {"use_reentrant": False} if is_torch_version(">=", "1.11.0") else {}
+                hidden_states = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    hidden_states,
+                    attention_mask,
+                    encoder_hidden_states,
+                    encoder_lora_states,
+                    encoder_attention_mask,
+                    timestep,
+                    cross_attention_kwargs,
+                    class_labels,
+                    **ckpt_kwargs,
+                )
+            else:
+                hidden_states = block(
+                    hidden_states,
+                    attention_mask=attention_mask,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_lora_states=encoder_lora_states,
+                    encoder_attention_mask=encoder_attention_mask,
+                    timestep=timestep,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    class_labels=class_labels,
+                )
+
+        # 3. Output
+        if self.is_input_continuous:
+            output = self._get_output_for_continuous_inputs(
+                hidden_states=hidden_states,
+                residual=residual,
+                batch_size=batch_size,
+                height=height,
+                width=width,
+                inner_dim=inner_dim,
+            )
+        elif self.is_input_vectorized:
+            output = self._get_output_for_vectorized_inputs(hidden_states)
+        elif self.is_input_patches:
+            output = self._get_output_for_patched_inputs(
+                hidden_states=hidden_states,
+                timestep=timestep,
+                class_labels=class_labels,
+                embedded_timestep=embedded_timestep,
+                height=height,
+                width=width,
+            )
+
+        if not return_dict:
+            return (output,)
+
+        return Transformer2DModelOutput(sample=output)
+
+    def _operate_on_continuous_inputs(self, hidden_states):
+        batch, _, height, width = hidden_states.shape
+        hidden_states = self.norm(hidden_states)
+
+        if not self.use_linear_projection:
+            hidden_states = self.proj_in(hidden_states)
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+        else:
+            inner_dim = hidden_states.shape[1]
+            hidden_states = hidden_states.permute(0, 2, 3, 1).reshape(batch, height * width, inner_dim)
+            hidden_states = self.proj_in(hidden_states)
+
+        return hidden_states, inner_dim
+
+    def _operate_on_patched_inputs(self, hidden_states, encoder_hidden_states, timestep, added_cond_kwargs):
+        batch_size = hidden_states.shape[0]
+        hidden_states = self.pos_embed(hidden_states)
+        embedded_timestep = None
+
+        if self.adaln_single is not None:
+            if self.use_additional_conditions and added_cond_kwargs is None:
+                raise ValueError(
+                    "`added_cond_kwargs` cannot be None when using additional conditions for `adaln_single`."
+                )
+            timestep, embedded_timestep = self.adaln_single(
+                timestep, added_cond_kwargs, batch_size=batch_size, hidden_dtype=hidden_states.dtype
+            )
+
+        if self.caption_projection is not None:
+            encoder_hidden_states = self.caption_projection(encoder_hidden_states)
+            encoder_hidden_states = encoder_hidden_states.view(batch_size, -1, hidden_states.shape[-1])
+
+        return hidden_states, encoder_hidden_states, timestep, embedded_timestep
+
+    def _get_output_for_continuous_inputs(self, hidden_states, residual, batch_size, height, width, inner_dim):
+        if not self.use_linear_projection:
+            hidden_states = (
+                hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+            hidden_states = self.proj_out(hidden_states)
+        else:
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = (
+                hidden_states.reshape(batch_size, height, width, inner_dim).permute(0, 3, 1, 2).contiguous()
+            )
+
+        output = hidden_states + residual
+        return output
+
+    def _get_output_for_vectorized_inputs(self, hidden_states):
+        hidden_states = self.norm_out(hidden_states)
+        logits = self.out(hidden_states)
+        # (batch, self.num_vector_embeds - 1, self.num_latent_pixels)
+        logits = logits.permute(0, 2, 1)
+        # log(p(x_0))
+        output = F.log_softmax(logits.double(), dim=1).float()
+        return output
+
+    def _get_output_for_patched_inputs(
+        self, hidden_states, timestep, class_labels, embedded_timestep, height=None, width=None
+    ):
+        if self.config.norm_type != "ada_norm_single":
+            conditioning = self.transformer_blocks[0].norm1.emb(
+                timestep, class_labels, hidden_dtype=hidden_states.dtype
+            )
+            shift, scale = self.proj_out_1(F.silu(conditioning)).chunk(2, dim=1)
+            hidden_states = self.norm_out(hidden_states) * (1 + scale[:, None]) + shift[:, None]
+            hidden_states = self.proj_out_2(hidden_states)
+        elif self.config.norm_type == "ada_norm_single":
+            shift, scale = (self.scale_shift_table[None] + embedded_timestep[:, None]).chunk(2, dim=1)
+            hidden_states = self.norm_out(hidden_states)
+            # Modulation
+            hidden_states = hidden_states * (1 + scale) + shift
+            hidden_states = self.proj_out(hidden_states)
+            hidden_states = hidden_states.squeeze(1)
+
+        # unpatchify
+        if self.adaln_single is None:
+            height = width = int(hidden_states.shape[1] ** 0.5)
+        hidden_states = hidden_states.reshape(
+            shape=(-1, height, width, self.patch_size, self.patch_size, self.out_channels)
+        )
+        hidden_states = torch.einsum("nhwpqc->nchpwq", hidden_states)
+        output = hidden_states.reshape(
+            shape=(-1, self.out_channels, height * self.patch_size, width * self.patch_size)
+        )
+        return output

models/unet_2d_condition.py

@@ -0,0 +1,1316 @@
+# Copyright 2024 The HuggingFace Team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from dataclasses import dataclass
+from typing import Any, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.utils.checkpoint
+
+from diffusers.configuration_utils import ConfigMixin, register_to_config
+from diffusers.loaders import PeftAdapterMixin, UNet2DConditionLoadersMixin
+from diffusers.loaders.single_file_model import FromOriginalModelMixin
+from diffusers.utils import USE_PEFT_BACKEND, BaseOutput, deprecate, logging, scale_lora_layers, unscale_lora_layers
+from diffusers.models.activations import get_activation
+from diffusers.models.attention_processor import (
+    ADDED_KV_ATTENTION_PROCESSORS,
+    CROSS_ATTENTION_PROCESSORS,
+    Attention,
+    AttentionProcessor,
+    AttnAddedKVProcessor,
+    AttnProcessor,
+    FusedAttnProcessor2_0,
+)
+from diffusers.models.embeddings import (
+    GaussianFourierProjection,
+    GLIGENTextBoundingboxProjection,
+    ImageHintTimeEmbedding,
+    ImageProjection,
+    ImageTimeEmbedding,
+    TextImageProjection,
+    TextImageTimeEmbedding,
+    TextTimeEmbedding,
+    TimestepEmbedding,
+    Timesteps,
+)
+from diffusers.models.modeling_utils import ModelMixin
+from models.unet_2d_blocks import (
+    get_down_block,
+    get_mid_block,
+    get_up_block,
+)
+
+
+logger = logging.get_logger(__name__)  # pylint: disable=invalid-name
+
+
+@dataclass
+class UNet2DConditionOutput(BaseOutput):
+    """
+    The output of [`UNet2DConditionModel`].
+
+    Args:
+        sample (`torch.Tensor` of shape `(batch_size, num_channels, height, width)`):
+            The hidden states output conditioned on `encoder_hidden_states` input. Output of last layer of model.
+    """
+
+    sample: torch.Tensor = None
+
+
+class UNet2DLoRAConditionModel(
+    ModelMixin, ConfigMixin, FromOriginalModelMixin, UNet2DConditionLoadersMixin, PeftAdapterMixin
+):
+    r"""
+    A conditional 2D UNet model that takes a noisy sample, conditional state, and a timestep and returns a sample
+    shaped output.
+
+    This model inherits from [`ModelMixin`]. Check the superclass documentation for it's generic methods implemented
+    for all models (such as downloading or saving).
+
+    Parameters:
+        sample_size (`int` or `Tuple[int, int]`, *optional*, defaults to `None`):
+            Height and width of input/output sample.
+        in_channels (`int`, *optional*, defaults to 4): Number of channels in the input sample.
+        out_channels (`int`, *optional*, defaults to 4): Number of channels in the output.
+        center_input_sample (`bool`, *optional*, defaults to `False`): Whether to center the input sample.
+        flip_sin_to_cos (`bool`, *optional*, defaults to `True`):
+            Whether to flip the sin to cos in the time embedding.
+        freq_shift (`int`, *optional*, defaults to 0): The frequency shift to apply to the time embedding.
+        down_block_types (`Tuple[str]`, *optional*, defaults to `("CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "CrossAttnDownBlock2D", "DownBlock2D")`):
+            The tuple of downsample blocks to use.
+        mid_block_type (`str`, *optional*, defaults to `"UNetMidBlock2DCrossAttn"`):
+            Block type for middle of UNet, it can be one of `UNetMidBlock2DCrossAttn`, `UNetMidBlock2D`, or
+            `UNetMidBlock2DSimpleCrossAttn`. If `None`, the mid block layer is skipped.
+        up_block_types (`Tuple[str]`, *optional*, defaults to `("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D")`):
+            The tuple of upsample blocks to use.
+        only_cross_attention(`bool` or `Tuple[bool]`, *optional*, default to `False`):
+            Whether to include self-attention in the basic transformer blocks, see
+            [`~models.attention.BasicTransformerBlock`].
+        block_out_channels (`Tuple[int]`, *optional*, defaults to `(320, 640, 1280, 1280)`):
+            The tuple of output channels for each block.
+        layers_per_block (`int`, *optional*, defaults to 2): The number of layers per block.
+        downsample_padding (`int`, *optional*, defaults to 1): The padding to use for the downsampling convolution.
+        mid_block_scale_factor (`float`, *optional*, defaults to 1.0): The scale factor to use for the mid block.
+        dropout (`float`, *optional*, defaults to 0.0): The dropout probability to use.
+        act_fn (`str`, *optional*, defaults to `"silu"`): The activation function to use.
+        norm_num_groups (`int`, *optional*, defaults to 32): The number of groups to use for the normalization.
+            If `None`, normalization and activation layers is skipped in post-processing.
+        norm_eps (`float`, *optional*, defaults to 1e-5): The epsilon to use for the normalization.
+        cross_attention_dim (`int` or `Tuple[int]`, *optional*, defaults to 1280):
+            The dimension of the cross attention features.
+        transformer_layers_per_block (`int`, `Tuple[int]`, or `Tuple[Tuple]` , *optional*, defaults to 1):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`]. Only relevant for
+            [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        reverse_transformer_layers_per_block : (`Tuple[Tuple]`, *optional*, defaults to None):
+            The number of transformer blocks of type [`~models.attention.BasicTransformerBlock`], in the upsampling
+            blocks of the U-Net. Only relevant if `transformer_layers_per_block` is of type `Tuple[Tuple]` and for
+            [`~models.unets.unet_2d_blocks.CrossAttnDownBlock2D`], [`~models.unets.unet_2d_blocks.CrossAttnUpBlock2D`],
+            [`~models.unets.unet_2d_blocks.UNetMidBlock2DCrossAttn`].
+        encoder_hid_dim (`int`, *optional*, defaults to None):
+            If `encoder_hid_dim_type` is defined, `encoder_hidden_states` will be projected from `encoder_hid_dim`
+            dimension to `cross_attention_dim`.
+        encoder_hid_dim_type (`str`, *optional*, defaults to `None`):
+            If given, the `encoder_hidden_states` and potentially other embeddings are down-projected to text
+            embeddings of dimension `cross_attention` according to `encoder_hid_dim_type`.
+        attention_head_dim (`int`, *optional*, defaults to 8): The dimension of the attention heads.
+        num_attention_heads (`int`, *optional*):
+            The number of attention heads. If not defined, defaults to `attention_head_dim`
+        resnet_time_scale_shift (`str`, *optional*, defaults to `"default"`): Time scale shift config
+            for ResNet blocks (see [`~models.resnet.ResnetBlock2D`]). Choose from `default` or `scale_shift`.
+        class_embed_type (`str`, *optional*, defaults to `None`):
+            The type of class embedding to use which is ultimately summed with the time embeddings. Choose from `None`,
+            `"timestep"`, `"identity"`, `"projection"`, or `"simple_projection"`.
+        addition_embed_type (`str`, *optional*, defaults to `None`):
+            Configures an optional embedding which will be summed with the time embeddings. Choose from `None` or
+            "text". "text" will use the `TextTimeEmbedding` layer.
+        addition_time_embed_dim: (`int`, *optional*, defaults to `None`):
+            Dimension for the timestep embeddings.
+        num_class_embeds (`int`, *optional*, defaults to `None`):
+            Input dimension of the learnable embedding matrix to be projected to `time_embed_dim`, when performing
+            class conditioning with `class_embed_type` equal to `None`.
+        time_embedding_type (`str`, *optional*, defaults to `positional`):
+            The type of position embedding to use for timesteps. Choose from `positional` or `fourier`.
+        time_embedding_dim (`int`, *optional*, defaults to `None`):
+            An optional override for the dimension of the projected time embedding.
+        time_embedding_act_fn (`str`, *optional*, defaults to `None`):
+            Optional activation function to use only once on the time embeddings before they are passed to the rest of
+            the UNet. Choose from `silu`, `mish`, `gelu`, and `swish`.
+        timestep_post_act (`str`, *optional*, defaults to `None`):
+            The second activation function to use in timestep embedding. Choose from `silu`, `mish` and `gelu`.
+        time_cond_proj_dim (`int`, *optional*, defaults to `None`):
+            The dimension of `cond_proj` layer in the timestep embedding.
+        conv_in_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_in` layer.
+        conv_out_kernel (`int`, *optional*, default to `3`): The kernel size of `conv_out` layer.
+        projection_class_embeddings_input_dim (`int`, *optional*): The dimension of the `class_labels` input when
+            `class_embed_type="projection"`. Required when `class_embed_type="projection"`.
+        class_embeddings_concat (`bool`, *optional*, defaults to `False`): Whether to concatenate the time
+            embeddings with the class embeddings.
+        mid_block_only_cross_attention (`bool`, *optional*, defaults to `None`):
+            Whether to use cross attention with the mid block when using the `UNetMidBlock2DSimpleCrossAttn`. If
+            `only_cross_attention` is given as a single boolean and `mid_block_only_cross_attention` is `None`, the
+            `only_cross_attention` value is used as the value for `mid_block_only_cross_attention`. Default to `False`
+            otherwise.
+    """
+
+    _supports_gradient_checkpointing = True
+    _no_split_modules = ["BasicTransformerBlock", "ResnetBlock2D", "CrossAttnUpBlock2D"]
+
+    @register_to_config
+    def __init__(
+        self,
+        sample_size: Optional[Union[int, Tuple[int, int]]] = None,
+        in_channels: int = 4,
+        out_channels: int = 4,
+        center_input_sample: bool = False,
+        flip_sin_to_cos: bool = True,
+        freq_shift: int = 0,
+        down_block_types: Tuple[str] = (
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "CrossAttnDownBlock2D",
+            "DownBlock2D",
+        ),
+        mid_block_type: Optional[str] = "UNetMidBlock2DCrossAttn",
+        up_block_types: Tuple[str] = ("UpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D", "CrossAttnUpBlock2D"),
+        only_cross_attention: Union[bool, Tuple[bool]] = False,
+        block_out_channels: Tuple[int] = (320, 640, 1280, 1280),
+        layers_per_block: Union[int, Tuple[int]] = 2,
+        downsample_padding: int = 1,
+        mid_block_scale_factor: float = 1,
+        dropout: float = 0.0,
+        act_fn: str = "silu",
+        norm_num_groups: Optional[int] = 32,
+        norm_eps: float = 1e-5,
+        cross_attention_dim: Union[int, Tuple[int]] = 1280,
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple]] = 1,
+        reverse_transformer_layers_per_block: Optional[Tuple[Tuple[int]]] = None,
+        encoder_hid_dim: Optional[int] = None,
+        encoder_hid_dim_type: Optional[str] = None,
+        attention_head_dim: Union[int, Tuple[int]] = 8,
+        num_attention_heads: Optional[Union[int, Tuple[int]]] = None,
+        dual_cross_attention: bool = False,
+        use_linear_projection: bool = False,
+        class_embed_type: Optional[str] = None,
+        addition_embed_type: Optional[str] = None,
+        addition_time_embed_dim: Optional[int] = None,
+        num_class_embeds: Optional[int] = None,
+        upcast_attention: bool = False,
+        resnet_time_scale_shift: str = "default",
+        resnet_skip_time_act: bool = False,
+        resnet_out_scale_factor: float = 1.0,
+        time_embedding_type: str = "positional",
+        time_embedding_dim: Optional[int] = None,
+        time_embedding_act_fn: Optional[str] = None,
+        timestep_post_act: Optional[str] = None,
+        time_cond_proj_dim: Optional[int] = None,
+        conv_in_kernel: int = 3,
+        conv_out_kernel: int = 3,
+        projection_class_embeddings_input_dim: Optional[int] = None,
+        attention_type: str = "default",
+        class_embeddings_concat: bool = False,
+        mid_block_only_cross_attention: Optional[bool] = None,
+        cross_attention_norm: Optional[str] = None,
+        addition_embed_type_num_heads: int = 64,
+    ):
+        super().__init__()
+
+        self.sample_size = sample_size
+
+        if num_attention_heads is not None:
+            raise ValueError(
+                "At the moment it is not possible to define the number of attention heads via `num_attention_heads` because of a naming issue as described in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131. Passing `num_attention_heads` will only be supported in diffusers v0.19."
+            )
+
+        # If `num_attention_heads` is not defined (which is the case for most models)
+        # it will default to `attention_head_dim`. This looks weird upon first reading it and it is.
+        # The reason for this behavior is to correct for incorrectly named variables that were introduced
+        # when this library was created. The incorrect naming was only discovered much later in https://github.com/huggingface/diffusers/issues/2011#issuecomment-1547958131
+        # Changing `attention_head_dim` to `num_attention_heads` for 40,000+ configurations is too backwards breaking
+        # which is why we correct for the naming here.
+        num_attention_heads = num_attention_heads or attention_head_dim
+
+        # Check inputs
+        self._check_config(
+            down_block_types=down_block_types,
+            up_block_types=up_block_types,
+            only_cross_attention=only_cross_attention,
+            block_out_channels=block_out_channels,
+            layers_per_block=layers_per_block,
+            cross_attention_dim=cross_attention_dim,
+            transformer_layers_per_block=transformer_layers_per_block,
+            reverse_transformer_layers_per_block=reverse_transformer_layers_per_block,
+            attention_head_dim=attention_head_dim,
+            num_attention_heads=num_attention_heads,
+        )
+
+        # input
+        conv_in_padding = (conv_in_kernel - 1) // 2
+        self.conv_in = nn.Conv2d(
+            in_channels, block_out_channels[0], kernel_size=conv_in_kernel, padding=conv_in_padding
+        )
+
+        # time
+        time_embed_dim, timestep_input_dim = self._set_time_proj(
+            time_embedding_type,
+            block_out_channels=block_out_channels,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            time_embedding_dim=time_embedding_dim,
+        )
+
+        self.time_embedding = TimestepEmbedding(
+            timestep_input_dim,
+            time_embed_dim,
+            act_fn=act_fn,
+            post_act_fn=timestep_post_act,
+            cond_proj_dim=time_cond_proj_dim,
+        )
+
+        self._set_encoder_hid_proj(
+            encoder_hid_dim_type,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+        )
+
+        # class embedding
+        self._set_class_embedding(
+            class_embed_type,
+            act_fn=act_fn,
+            num_class_embeds=num_class_embeds,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+            timestep_input_dim=timestep_input_dim,
+        )
+
+        self._set_add_embedding(
+            addition_embed_type,
+            addition_embed_type_num_heads=addition_embed_type_num_heads,
+            addition_time_embed_dim=addition_time_embed_dim,
+            cross_attention_dim=cross_attention_dim,
+            encoder_hid_dim=encoder_hid_dim,
+            flip_sin_to_cos=flip_sin_to_cos,
+            freq_shift=freq_shift,
+            projection_class_embeddings_input_dim=projection_class_embeddings_input_dim,
+            time_embed_dim=time_embed_dim,
+        )
+
+        if time_embedding_act_fn is None:
+            self.time_embed_act = None
+        else:
+            self.time_embed_act = get_activation(time_embedding_act_fn)
+
+        self.down_blocks = nn.ModuleList([])
+        self.up_blocks = nn.ModuleList([])
+
+        if isinstance(only_cross_attention, bool):
+            if mid_block_only_cross_attention is None:
+                mid_block_only_cross_attention = only_cross_attention
+
+            only_cross_attention = [only_cross_attention] * len(down_block_types)
+
+        if mid_block_only_cross_attention is None:
+            mid_block_only_cross_attention = False
+
+        if isinstance(num_attention_heads, int):
+            num_attention_heads = (num_attention_heads,) * len(down_block_types)
+
+        if isinstance(attention_head_dim, int):
+            attention_head_dim = (attention_head_dim,) * len(down_block_types)
+
+        if isinstance(cross_attention_dim, int):
+            cross_attention_dim = (cross_attention_dim,) * len(down_block_types)
+
+        if isinstance(layers_per_block, int):
+            layers_per_block = [layers_per_block] * len(down_block_types)
+
+        if isinstance(transformer_layers_per_block, int):
+            transformer_layers_per_block = [transformer_layers_per_block] * len(down_block_types)
+
+        if class_embeddings_concat:
+            # The time embeddings are concatenated with the class embeddings. The dimension of the
+            # time embeddings passed to the down, middle, and up blocks is twice the dimension of the
+            # regular time embeddings
+            blocks_time_embed_dim = time_embed_dim * 2
+        else:
+            blocks_time_embed_dim = time_embed_dim
+
+        # down
+        output_channel = block_out_channels[0]
+        for i, down_block_type in enumerate(down_block_types):
+            input_channel = output_channel
+            output_channel = block_out_channels[i]
+            is_final_block = i == len(block_out_channels) - 1
+
+            down_block = get_down_block(
+                down_block_type,
+                num_layers=layers_per_block[i],
+                transformer_layers_per_block=transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_downsample=not is_final_block,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=cross_attention_dim[i],
+                num_attention_heads=num_attention_heads[i],
+                downsample_padding=downsample_padding,
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.down_blocks.append(down_block)
+
+        # mid
+        self.mid_block = get_mid_block(
+            mid_block_type,
+            temb_channels=blocks_time_embed_dim,
+            in_channels=block_out_channels[-1],
+            resnet_eps=norm_eps,
+            resnet_act_fn=act_fn,
+            resnet_groups=norm_num_groups,
+            output_scale_factor=mid_block_scale_factor,
+            transformer_layers_per_block=transformer_layers_per_block[-1],
+            num_attention_heads=num_attention_heads[-1],
+            cross_attention_dim=cross_attention_dim[-1],
+            dual_cross_attention=dual_cross_attention,
+            use_linear_projection=use_linear_projection,
+            mid_block_only_cross_attention=mid_block_only_cross_attention,
+            upcast_attention=upcast_attention,
+            resnet_time_scale_shift=resnet_time_scale_shift,
+            attention_type=attention_type,
+            resnet_skip_time_act=resnet_skip_time_act,
+            cross_attention_norm=cross_attention_norm,
+            attention_head_dim=attention_head_dim[-1],
+            dropout=dropout,
+        )
+
+        # count how many layers upsample the images
+        self.num_upsamplers = 0
+
+        # up
+        reversed_block_out_channels = list(reversed(block_out_channels))
+        reversed_num_attention_heads = list(reversed(num_attention_heads))
+        reversed_layers_per_block = list(reversed(layers_per_block))
+        reversed_cross_attention_dim = list(reversed(cross_attention_dim))
+        reversed_transformer_layers_per_block = (
+            list(reversed(transformer_layers_per_block))
+            if reverse_transformer_layers_per_block is None
+            else reverse_transformer_layers_per_block
+        )
+        only_cross_attention = list(reversed(only_cross_attention))
+
+        output_channel = reversed_block_out_channels[0]
+        for i, up_block_type in enumerate(up_block_types):
+            is_final_block = i == len(block_out_channels) - 1
+
+            prev_output_channel = output_channel
+            output_channel = reversed_block_out_channels[i]
+            input_channel = reversed_block_out_channels[min(i + 1, len(block_out_channels) - 1)]
+
+            # add upsample block for all BUT final layer
+            if not is_final_block:
+                add_upsample = True
+                self.num_upsamplers += 1
+            else:
+                add_upsample = False
+
+            up_block = get_up_block(
+                up_block_type,
+                num_layers=reversed_layers_per_block[i] + 1,
+                transformer_layers_per_block=reversed_transformer_layers_per_block[i],
+                in_channels=input_channel,
+                out_channels=output_channel,
+                prev_output_channel=prev_output_channel,
+                temb_channels=blocks_time_embed_dim,
+                add_upsample=add_upsample,
+                resnet_eps=norm_eps,
+                resnet_act_fn=act_fn,
+                resolution_idx=i,
+                resnet_groups=norm_num_groups,
+                cross_attention_dim=reversed_cross_attention_dim[i],
+                num_attention_heads=reversed_num_attention_heads[i],
+                dual_cross_attention=dual_cross_attention,
+                use_linear_projection=use_linear_projection,
+                only_cross_attention=only_cross_attention[i],
+                upcast_attention=upcast_attention,
+                resnet_time_scale_shift=resnet_time_scale_shift,
+                attention_type=attention_type,
+                resnet_skip_time_act=resnet_skip_time_act,
+                resnet_out_scale_factor=resnet_out_scale_factor,
+                cross_attention_norm=cross_attention_norm,
+                attention_head_dim=attention_head_dim[i] if attention_head_dim[i] is not None else output_channel,
+                dropout=dropout,
+            )
+            self.up_blocks.append(up_block)
+
+        # out
+        if norm_num_groups is not None:
+            self.conv_norm_out = nn.GroupNorm(
+                num_channels=block_out_channels[0], num_groups=norm_num_groups, eps=norm_eps
+            )
+
+            self.conv_act = get_activation(act_fn)
+
+        else:
+            self.conv_norm_out = None
+            self.conv_act = None
+
+        conv_out_padding = (conv_out_kernel - 1) // 2
+        self.conv_out = nn.Conv2d(
+            block_out_channels[0], out_channels, kernel_size=conv_out_kernel, padding=conv_out_padding
+        )
+
+        self._set_pos_net_if_use_gligen(attention_type=attention_type, cross_attention_dim=cross_attention_dim)
+
+    def _check_config(
+        self,
+        down_block_types: Tuple[str],
+        up_block_types: Tuple[str],
+        only_cross_attention: Union[bool, Tuple[bool]],
+        block_out_channels: Tuple[int],
+        layers_per_block: Union[int, Tuple[int]],
+        cross_attention_dim: Union[int, Tuple[int]],
+        transformer_layers_per_block: Union[int, Tuple[int], Tuple[Tuple[int]]],
+        reverse_transformer_layers_per_block: bool,
+        attention_head_dim: int,
+        num_attention_heads: Optional[Union[int, Tuple[int]]],
+    ):
+        if len(down_block_types) != len(up_block_types):
+            raise ValueError(
+                f"Must provide the same number of `down_block_types` as `up_block_types`. `down_block_types`: {down_block_types}. `up_block_types`: {up_block_types}."
+            )
+
+        if len(block_out_channels) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `block_out_channels` as `down_block_types`. `block_out_channels`: {block_out_channels}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(only_cross_attention, bool) and len(only_cross_attention) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `only_cross_attention` as `down_block_types`. `only_cross_attention`: {only_cross_attention}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(num_attention_heads, int) and len(num_attention_heads) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `num_attention_heads` as `down_block_types`. `num_attention_heads`: {num_attention_heads}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(attention_head_dim, int) and len(attention_head_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `attention_head_dim` as `down_block_types`. `attention_head_dim`: {attention_head_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if isinstance(cross_attention_dim, list) and len(cross_attention_dim) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `cross_attention_dim` as `down_block_types`. `cross_attention_dim`: {cross_attention_dim}. `down_block_types`: {down_block_types}."
+            )
+
+        if not isinstance(layers_per_block, int) and len(layers_per_block) != len(down_block_types):
+            raise ValueError(
+                f"Must provide the same number of `layers_per_block` as `down_block_types`. `layers_per_block`: {layers_per_block}. `down_block_types`: {down_block_types}."
+            )
+        if isinstance(transformer_layers_per_block, list) and reverse_transformer_layers_per_block is None:
+            for layer_number_per_block in transformer_layers_per_block:
+                if isinstance(layer_number_per_block, list):
+                    raise ValueError("Must provide 'reverse_transformer_layers_per_block` if using asymmetrical UNet.")
+
+    def _set_time_proj(
+        self,
+        time_embedding_type: str,
+        block_out_channels: int,
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        time_embedding_dim: int,
+    ) -> Tuple[int, int]:
+        if time_embedding_type == "fourier":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 2
+            if time_embed_dim % 2 != 0:
+                raise ValueError(f"`time_embed_dim` should be divisible by 2, but is {time_embed_dim}.")
+            self.time_proj = GaussianFourierProjection(
+                time_embed_dim // 2, set_W_to_weight=False, log=False, flip_sin_to_cos=flip_sin_to_cos
+            )
+            timestep_input_dim = time_embed_dim
+        elif time_embedding_type == "positional":
+            time_embed_dim = time_embedding_dim or block_out_channels[0] * 4
+
+            self.time_proj = Timesteps(block_out_channels[0], flip_sin_to_cos, freq_shift)
+            timestep_input_dim = block_out_channels[0]
+        else:
+            raise ValueError(
+                f"{time_embedding_type} does not exist. Please make sure to use one of `fourier` or `positional`."
+            )
+
+        return time_embed_dim, timestep_input_dim
+
+    def _set_encoder_hid_proj(
+        self,
+        encoder_hid_dim_type: Optional[str],
+        cross_attention_dim: Union[int, Tuple[int]],
+        encoder_hid_dim: Optional[int],
+    ):
+        if encoder_hid_dim_type is None and encoder_hid_dim is not None:
+            encoder_hid_dim_type = "text_proj"
+            self.register_to_config(encoder_hid_dim_type=encoder_hid_dim_type)
+            logger.info("encoder_hid_dim_type defaults to 'text_proj' as `encoder_hid_dim` is defined.")
+
+        if encoder_hid_dim is None and encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim` has to be defined when `encoder_hid_dim_type` is set to {encoder_hid_dim_type}."
+            )
+
+        if encoder_hid_dim_type == "text_proj":
+            self.encoder_hid_proj = nn.Linear(encoder_hid_dim, cross_attention_dim)
+        elif encoder_hid_dim_type == "text_image_proj":
+            # image_embed_dim DOESN'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image_proj"` (Kandinsky 2.1)`
+            self.encoder_hid_proj = TextImageProjection(
+                text_embed_dim=encoder_hid_dim,
+                image_embed_dim=cross_attention_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2
+            self.encoder_hid_proj = ImageProjection(
+                image_embed_dim=encoder_hid_dim,
+                cross_attention_dim=cross_attention_dim,
+            )
+        elif encoder_hid_dim_type is not None:
+            raise ValueError(
+                f"`encoder_hid_dim_type`: {encoder_hid_dim_type} must be None, 'text_proj', 'text_image_proj', or 'image_proj'."
+            )
+        else:
+            self.encoder_hid_proj = None
+
+    def _set_class_embedding(
+        self,
+        class_embed_type: Optional[str],
+        act_fn: str,
+        num_class_embeds: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+        timestep_input_dim: int,
+    ):
+        if class_embed_type is None and num_class_embeds is not None:
+            self.class_embedding = nn.Embedding(num_class_embeds, time_embed_dim)
+        elif class_embed_type == "timestep":
+            self.class_embedding = TimestepEmbedding(timestep_input_dim, time_embed_dim, act_fn=act_fn)
+        elif class_embed_type == "identity":
+            self.class_embedding = nn.Identity(time_embed_dim, time_embed_dim)
+        elif class_embed_type == "projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            # The projection `class_embed_type` is the same as the timestep `class_embed_type` except
+            # 1. the `class_labels` inputs are not first converted to sinusoidal embeddings
+            # 2. it projects from an arbitrary input dimension.
+            #
+            # Note that `TimestepEmbedding` is quite general, being mainly linear layers and activations.
+            # When used for embedding actual timesteps, the timesteps are first converted to sinusoidal embeddings.
+            # As a result, `TimestepEmbedding` can be passed arbitrary vectors.
+            self.class_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif class_embed_type == "simple_projection":
+            if projection_class_embeddings_input_dim is None:
+                raise ValueError(
+                    "`class_embed_type`: 'simple_projection' requires `projection_class_embeddings_input_dim` be set"
+                )
+            self.class_embedding = nn.Linear(projection_class_embeddings_input_dim, time_embed_dim)
+        else:
+            self.class_embedding = None
+
+    def _set_add_embedding(
+        self,
+        addition_embed_type: str,
+        addition_embed_type_num_heads: int,
+        addition_time_embed_dim: Optional[int],
+        flip_sin_to_cos: bool,
+        freq_shift: float,
+        cross_attention_dim: Optional[int],
+        encoder_hid_dim: Optional[int],
+        projection_class_embeddings_input_dim: Optional[int],
+        time_embed_dim: int,
+    ):
+        if addition_embed_type == "text":
+            if encoder_hid_dim is not None:
+                text_time_embedding_from_dim = encoder_hid_dim
+            else:
+                text_time_embedding_from_dim = cross_attention_dim
+
+            self.add_embedding = TextTimeEmbedding(
+                text_time_embedding_from_dim, time_embed_dim, num_heads=addition_embed_type_num_heads
+            )
+        elif addition_embed_type == "text_image":
+            # text_embed_dim and image_embed_dim DON'T have to be `cross_attention_dim`. To not clutter the __init__ too much
+            # they are set to `cross_attention_dim` here as this is exactly the required dimension for the currently only use
+            # case when `addition_embed_type == "text_image"` (Kandinsky 2.1)`
+            self.add_embedding = TextImageTimeEmbedding(
+                text_embed_dim=cross_attention_dim, image_embed_dim=cross_attention_dim, time_embed_dim=time_embed_dim
+            )
+        elif addition_embed_type == "text_time":
+            self.add_time_proj = Timesteps(addition_time_embed_dim, flip_sin_to_cos, freq_shift)
+            self.add_embedding = TimestepEmbedding(projection_class_embeddings_input_dim, time_embed_dim)
+        elif addition_embed_type == "image":
+            # Kandinsky 2.2
+            self.add_embedding = ImageTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet
+            self.add_embedding = ImageHintTimeEmbedding(image_embed_dim=encoder_hid_dim, time_embed_dim=time_embed_dim)
+        elif addition_embed_type is not None:
+            raise ValueError(
+                f"`addition_embed_type`: {addition_embed_type} must be None, 'text', 'text_image', 'text_time', 'image', or 'image_hint'."
+            )
+
+    def _set_pos_net_if_use_gligen(self, attention_type: str, cross_attention_dim: int):
+        if attention_type in ["gated", "gated-text-image"]:
+            positive_len = 768
+            if isinstance(cross_attention_dim, int):
+                positive_len = cross_attention_dim
+            elif isinstance(cross_attention_dim, (list, tuple)):
+                positive_len = cross_attention_dim[0]
+
+            feature_type = "text-only" if attention_type == "gated" else "text-image"
+            self.position_net = GLIGENTextBoundingboxProjection(
+                positive_len=positive_len, out_dim=cross_attention_dim, feature_type=feature_type
+            )
+
+    @property
+    def attn_processors(self) -> Dict[str, AttentionProcessor]:
+        r"""
+        Returns:
+            `dict` of attention processors: A dictionary containing all attention processors used in the model with
+            indexed by its weight name.
+        """
+        # set recursively
+        processors = {}
+
+        def fn_recursive_add_processors(name: str, module: torch.nn.Module, processors: Dict[str, AttentionProcessor]):
+            if hasattr(module, "get_processor"):
+                processors[f"{name}.processor"] = module.get_processor()
+
+            for sub_name, child in module.named_children():
+                fn_recursive_add_processors(f"{name}.{sub_name}", child, processors)
+
+            return processors
+
+        for name, module in self.named_children():
+            fn_recursive_add_processors(name, module, processors)
+
+        return processors
+
+    def set_attn_processor(self, processor: Union[AttentionProcessor, Dict[str, AttentionProcessor]]):
+        r"""
+        Sets the attention processor to use to compute attention.
+
+        Parameters:
+            processor (`dict` of `AttentionProcessor` or only `AttentionProcessor`):
+                The instantiated processor class or a dictionary of processor classes that will be set as the processor
+                for **all** `Attention` layers.
+
+                If `processor` is a dict, the key needs to define the path to the corresponding cross attention
+                processor. This is strongly recommended when setting trainable attention processors.
+
+        """
+        count = len(self.attn_processors.keys())
+
+        if isinstance(processor, dict) and len(processor) != count:
+            raise ValueError(
+                f"A dict of processors was passed, but the number of processors {len(processor)} does not match the"
+                f" number of attention layers: {count}. Please make sure to pass {count} processor classes."
+            )
+
+        def fn_recursive_attn_processor(name: str, module: torch.nn.Module, processor):
+            if hasattr(module, "set_processor"):
+                if not isinstance(processor, dict):
+                    module.set_processor(processor)
+                else:
+                    module.set_processor(processor.pop(f"{name}.processor"))
+
+            for sub_name, child in module.named_children():
+                fn_recursive_attn_processor(f"{name}.{sub_name}", child, processor)
+
+        for name, module in self.named_children():
+            fn_recursive_attn_processor(name, module, processor)
+
+    def set_default_attn_processor(self):
+        """
+        Disables custom attention processors and sets the default attention implementation.
+        """
+        if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnAddedKVProcessor()
+        elif all(proc.__class__ in CROSS_ATTENTION_PROCESSORS for proc in self.attn_processors.values()):
+            processor = AttnProcessor()
+        else:
+            raise ValueError(
+                f"Cannot call `set_default_attn_processor` when attention processors are of type {next(iter(self.attn_processors.values()))}"
+            )
+
+        self.set_attn_processor(processor)
+
+    def set_attention_slice(self, slice_size: Union[str, int, List[int]] = "auto"):
+        r"""
+        Enable sliced attention computation.
+
+        When this option is enabled, the attention module splits the input tensor in slices to compute attention in
+        several steps. This is useful for saving some memory in exchange for a small decrease in speed.
+
+        Args:
+            slice_size (`str` or `int` or `list(int)`, *optional*, defaults to `"auto"`):
+                When `"auto"`, input to the attention heads is halved, so attention is computed in two steps. If
+                `"max"`, maximum amount of memory is saved by running only one slice at a time. If a number is
+                provided, uses as many slices as `attention_head_dim // slice_size`. In this case, `attention_head_dim`
+                must be a multiple of `slice_size`.
+        """
+        sliceable_head_dims = []
+
+        def fn_recursive_retrieve_sliceable_dims(module: torch.nn.Module):
+            if hasattr(module, "set_attention_slice"):
+                sliceable_head_dims.append(module.sliceable_head_dim)
+
+            for child in module.children():
+                fn_recursive_retrieve_sliceable_dims(child)
+
+        # retrieve number of attention layers
+        for module in self.children():
+            fn_recursive_retrieve_sliceable_dims(module)
+
+        num_sliceable_layers = len(sliceable_head_dims)
+
+        if slice_size == "auto":
+            # half the attention head size is usually a good trade-off between
+            # speed and memory
+            slice_size = [dim // 2 for dim in sliceable_head_dims]
+        elif slice_size == "max":
+            # make smallest slice possible
+            slice_size = num_sliceable_layers * [1]
+
+        slice_size = num_sliceable_layers * [slice_size] if not isinstance(slice_size, list) else slice_size
+
+        if len(slice_size) != len(sliceable_head_dims):
+            raise ValueError(
+                f"You have provided {len(slice_size)}, but {self.config} has {len(sliceable_head_dims)} different"
+                f" attention layers. Make sure to match `len(slice_size)` to be {len(sliceable_head_dims)}."
+            )
+
+        for i in range(len(slice_size)):
+            size = slice_size[i]
+            dim = sliceable_head_dims[i]
+            if size is not None and size > dim:
+                raise ValueError(f"size {size} has to be smaller or equal to {dim}.")
+
+        # Recursively walk through all the children.
+        # Any children which exposes the set_attention_slice method
+        # gets the message
+        def fn_recursive_set_attention_slice(module: torch.nn.Module, slice_size: List[int]):
+            if hasattr(module, "set_attention_slice"):
+                module.set_attention_slice(slice_size.pop())
+
+            for child in module.children():
+                fn_recursive_set_attention_slice(child, slice_size)
+
+        reversed_slice_size = list(reversed(slice_size))
+        for module in self.children():
+            fn_recursive_set_attention_slice(module, reversed_slice_size)
+
+    def _set_gradient_checkpointing(self, module, value=False):
+        if hasattr(module, "gradient_checkpointing"):
+            module.gradient_checkpointing = value
+
+    def enable_freeu(self, s1: float, s2: float, b1: float, b2: float):
+        r"""Enables the FreeU mechanism from https://arxiv.org/abs/2309.11497.
+
+        The suffixes after the scaling factors represent the stage blocks where they are being applied.
+
+        Please refer to the [official repository](https://github.com/ChenyangSi/FreeU) for combinations of values that
+        are known to work well for different pipelines such as Stable Diffusion v1, v2, and Stable Diffusion XL.
+
+        Args:
+            s1 (`float`):
+                Scaling factor for stage 1 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            s2 (`float`):
+                Scaling factor for stage 2 to attenuate the contributions of the skip features. This is done to
+                mitigate the "oversmoothing effect" in the enhanced denoising process.
+            b1 (`float`): Scaling factor for stage 1 to amplify the contributions of backbone features.
+            b2 (`float`): Scaling factor for stage 2 to amplify the contributions of backbone features.
+        """
+        for i, upsample_block in enumerate(self.up_blocks):
+            setattr(upsample_block, "s1", s1)
+            setattr(upsample_block, "s2", s2)
+            setattr(upsample_block, "b1", b1)
+            setattr(upsample_block, "b2", b2)
+
+    def disable_freeu(self):
+        """Disables the FreeU mechanism."""
+        freeu_keys = {"s1", "s2", "b1", "b2"}
+        for i, upsample_block in enumerate(self.up_blocks):
+            for k in freeu_keys:
+                if hasattr(upsample_block, k) or getattr(upsample_block, k, None) is not None:
+                    setattr(upsample_block, k, None)
+
+    def fuse_qkv_projections(self):
+        """
+        Enables fused QKV projections. For self-attention modules, all projection matrices (i.e., query, key, value)
+        are fused. For cross-attention modules, key and value projection matrices are fused.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+        """
+        self.original_attn_processors = None
+
+        for _, attn_processor in self.attn_processors.items():
+            if "Added" in str(attn_processor.__class__.__name__):
+                raise ValueError("`fuse_qkv_projections()` is not supported for models having added KV projections.")
+
+        self.original_attn_processors = self.attn_processors
+
+        for module in self.modules():
+            if isinstance(module, Attention):
+                module.fuse_projections(fuse=True)
+
+        self.set_attn_processor(FusedAttnProcessor2_0())
+
+    def unfuse_qkv_projections(self):
+        """Disables the fused QKV projection if enabled.
+
+        <Tip warning={true}>
+
+        This API is 🧪 experimental.
+
+        </Tip>
+
+        """
+        if self.original_attn_processors is not None:
+            self.set_attn_processor(self.original_attn_processors)
+
+    def get_time_embed(
+        self, sample: torch.Tensor, timestep: Union[torch.Tensor, float, int]
+    ) -> Optional[torch.Tensor]:
+        timesteps = timestep
+        if not torch.is_tensor(timesteps):
+            # TODO: this requires sync between CPU and GPU. So try to pass timesteps as tensors if you can
+            # This would be a good case for the `match` statement (Python 3.10+)
+            is_mps = sample.device.type == "mps"
+            if isinstance(timestep, float):
+                dtype = torch.float32 if is_mps else torch.float64
+            else:
+                dtype = torch.int32 if is_mps else torch.int64
+            timesteps = torch.tensor([timesteps], dtype=dtype, device=sample.device)
+        elif len(timesteps.shape) == 0:
+            timesteps = timesteps[None].to(sample.device)
+
+        # broadcast to batch dimension in a way that's compatible with ONNX/Core ML
+        timesteps = timesteps.expand(sample.shape[0])
+
+        t_emb = self.time_proj(timesteps)
+        # `Timesteps` does not contain any weights and will always return f32 tensors
+        # but time_embedding might actually be running in fp16. so we need to cast here.
+        # there might be better ways to encapsulate this.
+        t_emb = t_emb.to(dtype=sample.dtype)
+        return t_emb
+
+    def get_class_embed(self, sample: torch.Tensor, class_labels: Optional[torch.Tensor]) -> Optional[torch.Tensor]:
+        class_emb = None
+        if self.class_embedding is not None:
+            if class_labels is None:
+                raise ValueError("class_labels should be provided when num_class_embeds > 0")
+
+            if self.config.class_embed_type == "timestep":
+                class_labels = self.time_proj(class_labels)
+
+                # `Timesteps` does not contain any weights and will always return f32 tensors
+                # there might be better ways to encapsulate this.
+                class_labels = class_labels.to(dtype=sample.dtype)
+
+            class_emb = self.class_embedding(class_labels).to(dtype=sample.dtype)
+        return class_emb
+
+    def get_aug_embed(
+        self, emb: torch.Tensor, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> Optional[torch.Tensor]:
+        aug_emb = None
+        if self.config.addition_embed_type == "text":
+            aug_emb = self.add_embedding(encoder_hidden_states)
+        elif self.config.addition_embed_type == "text_image":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embs = added_cond_kwargs.get("image_embeds")
+            text_embs = added_cond_kwargs.get("text_embeds", encoder_hidden_states)
+            aug_emb = self.add_embedding(text_embs, image_embs)
+        elif self.config.addition_embed_type == "text_time":
+            # SDXL - style
+            if "text_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `text_embeds` to be passed in `added_cond_kwargs`"
+                )
+            text_embeds = added_cond_kwargs.get("text_embeds")
+            if "time_ids" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'text_time' which requires the keyword argument `time_ids` to be passed in `added_cond_kwargs`"
+                )
+            time_ids = added_cond_kwargs.get("time_ids")
+            time_embeds = self.add_time_proj(time_ids.flatten())
+            time_embeds = time_embeds.reshape((text_embeds.shape[0], -1))
+            add_embeds = torch.concat([text_embeds, time_embeds], dim=-1)
+            add_embeds = add_embeds.to(emb.dtype)
+            aug_emb = self.add_embedding(add_embeds)
+        elif self.config.addition_embed_type == "image":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            aug_emb = self.add_embedding(image_embs)
+        elif self.config.addition_embed_type == "image_hint":
+            # Kandinsky 2.2 ControlNet - style
+            if "image_embeds" not in added_cond_kwargs or "hint" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `addition_embed_type` set to 'image_hint' which requires the keyword arguments `image_embeds` and `hint` to be passed in `added_cond_kwargs`"
+                )
+            image_embs = added_cond_kwargs.get("image_embeds")
+            hint = added_cond_kwargs.get("hint")
+            aug_emb = self.add_embedding(image_embs, hint)
+        return aug_emb
+
+    def process_encoder_hidden_states(
+        self, encoder_hidden_states: torch.Tensor, added_cond_kwargs: Dict[str, Any]
+    ) -> torch.Tensor:
+        if self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_proj":
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "text_image_proj":
+            # Kandinsky 2.1 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'text_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(encoder_hidden_states, image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "image_proj":
+            # Kandinsky 2.2 - style
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'image_proj' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            encoder_hidden_states = self.encoder_hid_proj(image_embeds)
+        elif self.encoder_hid_proj is not None and self.config.encoder_hid_dim_type == "ip_image_proj":
+            if "image_embeds" not in added_cond_kwargs:
+                raise ValueError(
+                    f"{self.__class__} has the config param `encoder_hid_dim_type` set to 'ip_image_proj' which requires the keyword argument `image_embeds` to be passed in `added_cond_kwargs`"
+                )
+
+            if hasattr(self, "text_encoder_hid_proj") and self.text_encoder_hid_proj is not None:
+                encoder_hidden_states = self.text_encoder_hid_proj(encoder_hidden_states)
+
+            image_embeds = added_cond_kwargs.get("image_embeds")
+            image_embeds = self.encoder_hid_proj(image_embeds)
+            encoder_hidden_states = (encoder_hidden_states, image_embeds)
+        return encoder_hidden_states
+
+    def forward(
+        self,
+        sample: torch.Tensor,
+        timestep: Union[torch.Tensor, float, int],
+        encoder_hidden_states: torch.Tensor,
+        encoder_lora_states: torch.Tensor,
+        class_labels: Optional[torch.Tensor] = None,
+        timestep_cond: Optional[torch.Tensor] = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        cross_attention_kwargs: Optional[Dict[str, Any]] = None,
+        added_cond_kwargs: Optional[Dict[str, torch.Tensor]] = None,
+        down_block_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        mid_block_additional_residual: Optional[torch.Tensor] = None,
+        down_intrablock_additional_residuals: Optional[Tuple[torch.Tensor]] = None,
+        encoder_attention_mask: Optional[torch.Tensor] = None,
+        return_dict: bool = True,
+    ) -> Union[UNet2DConditionOutput, Tuple]:
+        r"""
+        The [`UNet2DConditionModel`] forward method.
+
+        Args:
+            sample (`torch.Tensor`):
+                The noisy input tensor with the following shape `(batch, channel, height, width)`.
+            timestep (`torch.Tensor` or `float` or `int`): The number of timesteps to denoise an input.
+            encoder_hidden_states (`torch.Tensor`):
+                The encoder hidden states with shape `(batch, sequence_length, feature_dim)`.
+            class_labels (`torch.Tensor`, *optional*, defaults to `None`):
+                Optional class labels for conditioning. Their embeddings will be summed with the timestep embeddings.
+            timestep_cond: (`torch.Tensor`, *optional*, defaults to `None`):
+                Conditional embeddings for timestep. If provided, the embeddings will be summed with the samples passed
+                through the `self.time_embedding` layer to obtain the timestep embeddings.
+            attention_mask (`torch.Tensor`, *optional*, defaults to `None`):
+                An attention mask of shape `(batch, key_tokens)` is applied to `encoder_hidden_states`. If `1` the mask
+                is kept, otherwise if `0` it is discarded. Mask will be converted into a bias, which adds large
+                negative values to the attention scores corresponding to "discard" tokens.
+            cross_attention_kwargs (`dict`, *optional*):
+                A kwargs dictionary that if specified is passed along to the `AttentionProcessor` as defined under
+                `self.processor` in
+                [diffusers.models.attention_processor](https://github.com/huggingface/diffusers/blob/main/src/diffusers/models/attention_processor.py).
+            added_cond_kwargs: (`dict`, *optional*):
+                A kwargs dictionary containing additional embeddings that if specified are added to the embeddings that
+                are passed along to the UNet blocks.
+            down_block_additional_residuals: (`tuple` of `torch.Tensor`, *optional*):
+                A tuple of tensors that if specified are added to the residuals of down unet blocks.
+            mid_block_additional_residual: (`torch.Tensor`, *optional*):
+                A tensor that if specified is added to the residual of the middle unet block.
+            down_intrablock_additional_residuals (`tuple` of `torch.Tensor`, *optional*):
+                additional residuals to be added within UNet down blocks, for example from T2I-Adapter side model(s)
+            encoder_attention_mask (`torch.Tensor`):
+                A cross-attention mask of shape `(batch, sequence_length)` is applied to `encoder_hidden_states`. If
+                `True` the mask is kept, otherwise if `False` it is discarded. Mask will be converted into a bias,
+                which adds large negative values to the attention scores corresponding to "discard" tokens.
+            return_dict (`bool`, *optional*, defaults to `True`):
+                Whether or not to return a [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] instead of a plain
+                tuple.
+
+        Returns:
+            [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] or `tuple`:
+                If `return_dict` is True, an [`~models.unets.unet_2d_condition.UNet2DConditionOutput`] is returned,
+                otherwise a `tuple` is returned where the first element is the sample tensor.
+        """
+        # By default samples have to be AT least a multiple of the overall upsampling factor.
+        # The overall upsampling factor is equal to 2 ** (# num of upsampling layers).
+        # However, the upsampling interpolation output size can be forced to fit any upsampling size
+        # on the fly if necessary.
+        default_overall_up_factor = 2**self.num_upsamplers
+
+        # upsample size should be forwarded when sample is not a multiple of `default_overall_up_factor`
+        forward_upsample_size = False
+        upsample_size = None
+
+        for dim in sample.shape[-2:]:
+            if dim % default_overall_up_factor != 0:
+                # Forward upsample size to force interpolation output size.
+                forward_upsample_size = True
+                break
+
+        # ensure attention_mask is a bias, and give it a singleton query_tokens dimension
+        # expects mask of shape:
+        #   [batch, key_tokens]
+        # adds singleton query_tokens dimension:
+        #   [batch,                    1, key_tokens]
+        # this helps to broadcast it as a bias over attention scores, which will be in one of the following shapes:
+        #   [batch,  heads, query_tokens, key_tokens] (e.g. torch sdp attn)
+        #   [batch * heads, query_tokens, key_tokens] (e.g. xformers or classic attn)
+        if attention_mask is not None:
+            # assume that mask is expressed as:
+            #   (1 = keep,      0 = discard)
+            # convert mask into a bias that can be added to attention scores:
+            #       (keep = +0,     discard = -10000.0)
+            attention_mask = (1 - attention_mask.to(sample.dtype)) * -10000.0
+            attention_mask = attention_mask.unsqueeze(1)
+
+        # convert encoder_attention_mask to a bias the same way we do for attention_mask
+        if encoder_attention_mask is not None:
+            encoder_attention_mask = (1 - encoder_attention_mask.to(sample.dtype)) * -10000.0
+            encoder_attention_mask = encoder_attention_mask.unsqueeze(1)
+
+        # 0. center input if necessary
+        if self.config.center_input_sample:
+            sample = 2 * sample - 1.0
+
+        # 1. time
+        t_emb = self.get_time_embed(sample=sample, timestep=timestep)
+        emb = self.time_embedding(t_emb, timestep_cond)
+
+        class_emb = self.get_class_embed(sample=sample, class_labels=class_labels)
+        if class_emb is not None:
+            if self.config.class_embeddings_concat:
+                emb = torch.cat([emb, class_emb], dim=-1)
+            else:
+                emb = emb + class_emb
+
+        aug_emb = self.get_aug_embed(
+            emb=emb, encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+        if self.config.addition_embed_type == "image_hint":
+            aug_emb, hint = aug_emb
+            sample = torch.cat([sample, hint], dim=1)
+
+        emb = emb + aug_emb if aug_emb is not None else emb
+
+        if self.time_embed_act is not None:
+            emb = self.time_embed_act(emb)
+
+        encoder_hidden_states = self.process_encoder_hidden_states(
+            encoder_hidden_states=encoder_hidden_states, added_cond_kwargs=added_cond_kwargs
+        )
+
+        # 2. pre-process
+        sample = self.conv_in(sample)
+
+        # 2.5 GLIGEN position net
+        if cross_attention_kwargs is not None and cross_attention_kwargs.get("gligen", None) is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            gligen_args = cross_attention_kwargs.pop("gligen")
+            cross_attention_kwargs["gligen"] = {"objs": self.position_net(**gligen_args)}
+
+        # 3. down
+        # we're popping the `scale` instead of getting it because otherwise `scale` will be propagated
+        # to the internal blocks and will raise deprecation warnings. this will be confusing for our users.
+        if cross_attention_kwargs is not None:
+            cross_attention_kwargs = cross_attention_kwargs.copy()
+            lora_scale = cross_attention_kwargs.pop("scale", 1.0)
+        else:
+            lora_scale = 1.0
+
+        if USE_PEFT_BACKEND:
+            # weight the lora layers by setting `lora_scale` for each PEFT layer
+            scale_lora_layers(self, lora_scale)
+
+        is_controlnet = mid_block_additional_residual is not None and down_block_additional_residuals is not None
+        # using new arg down_intrablock_additional_residuals for T2I-Adapters, to distinguish from controlnets
+        is_adapter = down_intrablock_additional_residuals is not None
+        # maintain backward compatibility for legacy usage, where
+        #       T2I-Adapter and ControlNet both use down_block_additional_residuals arg
+        #       but can only use one or the other
+        if not is_adapter and mid_block_additional_residual is None and down_block_additional_residuals is not None:
+            deprecate(
+                "T2I should not use down_block_additional_residuals",
+                "1.3.0",
+                "Passing intrablock residual connections with `down_block_additional_residuals` is deprecated \
+                       and will be removed in diffusers 1.3.0.  `down_block_additional_residuals` should only be used \
+                       for ControlNet. Please make sure use `down_intrablock_additional_residuals` instead. ",
+                standard_warn=False,
+            )
+            down_intrablock_additional_residuals = down_block_additional_residuals
+            is_adapter = True
+
+        down_block_res_samples = (sample,)
+        for downsample_block in self.down_blocks:
+            if hasattr(downsample_block, "has_cross_attention") and downsample_block.has_cross_attention:
+                # For t2i-adapter CrossAttnDownBlock2D
+                additional_residuals = {}
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    additional_residuals["additional_residuals"] = down_intrablock_additional_residuals.pop(0)
+
+                sample, res_samples = downsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_lora_states=encoder_lora_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                    **additional_residuals,
+                )
+            else:
+                sample, res_samples = downsample_block(hidden_states=sample, temb=emb)
+                if is_adapter and len(down_intrablock_additional_residuals) > 0:
+                    sample += down_intrablock_additional_residuals.pop(0)
+
+            down_block_res_samples += res_samples
+
+        if is_controlnet:
+            new_down_block_res_samples = ()
+
+            for down_block_res_sample, down_block_additional_residual in zip(
+                down_block_res_samples, down_block_additional_residuals
+            ):
+                down_block_res_sample = down_block_res_sample + down_block_additional_residual
+                new_down_block_res_samples = new_down_block_res_samples + (down_block_res_sample,)
+
+            down_block_res_samples = new_down_block_res_samples
+
+        # 4. mid
+        if self.mid_block is not None:
+            if hasattr(self.mid_block, "has_cross_attention") and self.mid_block.has_cross_attention:
+                sample = self.mid_block(
+                    sample,
+                    emb,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_lora_states=encoder_lora_states,
+                    attention_mask=attention_mask,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = self.mid_block(sample, emb)
+
+            # To support T2I-Adapter-XL
+            if (
+                is_adapter
+                and len(down_intrablock_additional_residuals) > 0
+                and sample.shape == down_intrablock_additional_residuals[0].shape
+            ):
+                sample += down_intrablock_additional_residuals.pop(0)
+
+        if is_controlnet:
+            sample = sample + mid_block_additional_residual
+
+        # 5. up
+        for i, upsample_block in enumerate(self.up_blocks):
+            is_final_block = i == len(self.up_blocks) - 1
+
+            res_samples = down_block_res_samples[-len(upsample_block.resnets) :]
+            down_block_res_samples = down_block_res_samples[: -len(upsample_block.resnets)]
+
+            # if we have not reached the final block and need to forward the
+            # upsample size, we do it here
+            if not is_final_block and forward_upsample_size:
+                upsample_size = down_block_res_samples[-1].shape[2:]
+
+            if hasattr(upsample_block, "has_cross_attention") and upsample_block.has_cross_attention:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    encoder_hidden_states=encoder_hidden_states,
+                    encoder_lora_states=encoder_lora_states,
+                    cross_attention_kwargs=cross_attention_kwargs,
+                    upsample_size=upsample_size,
+                    attention_mask=attention_mask,
+                    encoder_attention_mask=encoder_attention_mask,
+                )
+            else:
+                sample = upsample_block(
+                    hidden_states=sample,
+                    temb=emb,
+                    res_hidden_states_tuple=res_samples,
+                    upsample_size=upsample_size,
+                )
+
+        # 6. post-process
+        if self.conv_norm_out:
+            sample = self.conv_norm_out(sample)
+            sample = self.conv_act(sample)
+        sample = self.conv_out(sample)
+
+        if USE_PEFT_BACKEND:
+            # remove `lora_scale` from each PEFT layer
+            unscale_lora_layers(self, lora_scale)
+
+        if not return_dict:
+            return (sample,)
+
+        return UNet2DConditionOutput(sample=sample)

models/visual_prompts.py

@@ -0,0 +1,166 @@
+import torch
+import torch.nn as nn
+import os
+import numpy as np
+from diffusers.models.attention_processor import Attention
+
+class VisualTokenSelfAttn(torch.nn.Module):
+    def __init__(self, in_dim=2792, out_dim=768, num_heads=8):
+        super().__init__()
+
+        self.meta_token_trans = nn.Sequential(
+            nn.Linear(in_dim, out_dim * 4),
+            nn.LayerNorm(out_dim * 4),
+            nn.GELU(),
+            nn.Linear(out_dim * 4, out_dim),
+            nn.LayerNorm(out_dim)
+        )
+
+        self.norm1 = nn.LayerNorm(out_dim, eps=1e-6) # important to avoid attention collapsing
+        self.attn = Attention(query_dim=out_dim, heads=num_heads)
+        self.norm2 = nn.LayerNorm(out_dim, eps=1e-6)
+        self.mlp = nn.Sequential(
+            nn.Linear(out_dim, out_dim * 4),
+            nn.GELU(),
+            nn.Linear(out_dim * 4, out_dim)
+        )
+
+    def forward(self, x):
+        x = self.meta_token_trans(x)
+        x = x + self.attn(self.norm1(x))
+        x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class EmotionEmbedding(nn.Module):
+    def __init__(self, emotions, prompts_dir, feature_names, output_dim, prompt_len=16):
+        super().__init__()
+        
+        input_dim = self.get_input_dim(feature_names=feature_names)
+        self.self_attn = VisualTokenSelfAttn(in_dim=input_dim, out_dim=output_dim)
+        
+        self.emotions = emotions
+        self.emotion2idx = {emotion: idx for idx, emotion in enumerate(emotions)}
+        self.emotion_params = nn.ParameterList()
+        
+        self.emotion_init_features = self.get_features(emotions, prompts_dir, feature_names, prompt_len)
+
+        for emotion in self.emotions:
+            init_params = self.emotion_init_features[emotion]
+            # init_params = torch.from_numpy(init_params).float()
+            param = nn.Parameter(init_params)
+            self.emotion_params.append(param)
+    
+    def get_features(self, emotions, prompts_dir, feature_names, prompt_len):
+        emotion_init_features = {}
+        for emotion in emotions:
+            emotion_features = []
+            for feature_name in feature_names:
+                features = np.load(os.path.join(prompts_dir, f'{emotion}_{feature_name}.npy'), allow_pickle=True)
+                emotion_features.append(features)
+            emotion_features = np.concatenate(emotion_features, axis=1)
+            
+            from sklearn.cluster import KMeans
+            kmeans = KMeans(n_clusters=prompt_len, random_state=42)
+            kmeans.fit_predict(emotion_features)
+            token = torch.tensor(kmeans.cluster_centers_).unsqueeze(0)
+            # print(token.shape)
+            emotion_init_features[emotion] = token
+        return emotion_init_features
+            
+    def get_input_dim(self, feature_names):
+        if feature_names == ["clip"]:
+            in_dim = 768
+        elif feature_names == ["vgg"]:
+            in_dim = 1000
+        elif feature_names == ["dinov2"]:
+            in_dim = 1024
+        elif feature_names == ["clip", "vgg"]:
+            in_dim = 1768
+        elif feature_names == ["clip", "dinov2"]:
+            in_dim = 1768
+        elif feature_names == ["vgg", "dinov2"]:
+            in_dim = 2024
+        elif feature_names == ["clip", "vgg", "dinov2"]:
+            in_dim = 2792
+        else:
+            raise ValueError("Invalid feature names")
+        return in_dim
+    
+    def params_to_prompts(self):
+        self.emotion_prompts = {}
+        for emotion in self.emotions:
+            prompt = self.self_attn(self.emotion_params[self.emotion2idx[emotion]])
+            prompt = prompt.squeeze(0)
+            self.emotion_prompts[emotion] = prompt
+
+    def forward(self, emotion):
+        if isinstance(emotion, str): 
+            emotions = [emotion]
+        else: 
+            emotions = emotion
+        
+        self.params_to_prompts()
+        selected_prompts = [self.emotion_prompts[emotion] for emotion in emotions]
+        prompts = torch.stack(selected_prompts, dim=0)
+        del self.emotion_prompts
+        
+        return prompts
+
+class EmotionEmbedding2(nn.Module):
+    def __init__(self, emotions, input_dim, output_dim):
+        super().__init__()
+        self.self_attn = VisualTokenSelfAttn(in_dim=input_dim, out_dim=output_dim)
+        self.emotions = emotions
+        self.emotion2idx = {emotion: idx for idx, emotion in enumerate(emotions)}
+        self.emotion_params = nn.Embedding(len(emotions), input_dim)
+    
+    def forward(self, emotion):
+        if isinstance(emotion, str): 
+            emotions = [emotion]
+        else: 
+            emotions = emotion
+
+        emotions = [self.emotion2idx[emotion] for emotion in emotions]
+        emotions = torch.tensor(emotions, device=self.emotion_params.weight.device)
+        prompts = self.emotion_params(emotions).unsqueeze(1)
+        prompts = self.self_attn(prompts)
+        return prompts
+    
+if __name__ == "__main__":
+    # emotions = ["amusement", "anger", "awe", "contentment", 
+    #             "disgust", "excitement", "fear", "sadness"]
+    # feature_names = ["clip", "vgg", "dinov2"]
+    # prompts_dir = "features/origin"
+    # model = EmotionEmbedding(emotions, prompts_dir, feature_names, output_dim=2048, prompt_len=16).to("cuda")
+    # optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
+    # output = model('awe')
+    # target = torch.ones_like(output)
+    # loss = ((output - target) ** 2).mean()
+    # print(output)
+
+    emotions = ["amusement", "anger", "awe", "contentment", 
+            "disgust", "excitement", "fear", "sadness"]
+    prompts_dir = "features/origin"
+    model = EmotionEmbedding2(emotions, input_dim=2048, output_dim=2048, prompt_len=16).to("cuda")
+    optimizer = torch.optim.Adam(model.parameters(), lr=1e-4)
+    output = model('awe')
+    target = torch.ones_like(output)
+    loss = ((output - target) ** 2).mean()
+    print(output)
+
+    # 反向传播
+    loss.backward()
+
+    # 打印看看梯度
+    for name, param in model.named_parameters():
+        if param.grad is not None:
+            print(f"{name} has gradient ✅, grad mean: {param.grad.mean().item()}")
+            if name == "emotion_params.weight":
+                print(param.grad)
+        else:
+            print(f"{name} has NO gradient ❌")
+
+    # 更新一下参数
+    optimizer.step()
+    print(output)