Initial commit

.gitattributes

@@ -0,0 +1,38 @@
+*.7z filter=lfs diff=lfs merge=lfs -text
+*.arrow filter=lfs diff=lfs merge=lfs -text
+*.bin filter=lfs diff=lfs merge=lfs -text
+*.bz2 filter=lfs diff=lfs merge=lfs -text
+*.ckpt filter=lfs diff=lfs merge=lfs -text
+*.ftz filter=lfs diff=lfs merge=lfs -text
+*.gz filter=lfs diff=lfs merge=lfs -text
+*.h5 filter=lfs diff=lfs merge=lfs -text
+*.joblib filter=lfs diff=lfs merge=lfs -text
+*.lfs.* filter=lfs diff=lfs merge=lfs -text
+*.mlmodel filter=lfs diff=lfs merge=lfs -text
+*.model filter=lfs diff=lfs merge=lfs -text
+*.msgpack filter=lfs diff=lfs merge=lfs -text
+*.npy filter=lfs diff=lfs merge=lfs -text
+*.npz filter=lfs diff=lfs merge=lfs -text
+*.onnx filter=lfs diff=lfs merge=lfs -text
+*.ot filter=lfs diff=lfs merge=lfs -text
+*.parquet filter=lfs diff=lfs merge=lfs -text
+*.pb filter=lfs diff=lfs merge=lfs -text
+*.pickle filter=lfs diff=lfs merge=lfs -text
+*.pkl filter=lfs diff=lfs merge=lfs -text
+*.pt filter=lfs diff=lfs merge=lfs -text
+*.pth filter=lfs diff=lfs merge=lfs -text
+*.rar filter=lfs diff=lfs merge=lfs -text
+*.safetensors filter=lfs diff=lfs merge=lfs -text
+saved_model/**/* filter=lfs diff=lfs merge=lfs -text
+*.tar.* filter=lfs diff=lfs merge=lfs -text
+*.tar filter=lfs diff=lfs merge=lfs -text
+*.tflite filter=lfs diff=lfs merge=lfs -text
+*.tgz filter=lfs diff=lfs merge=lfs -text
+*.wasm filter=lfs diff=lfs merge=lfs -text
+*.xz filter=lfs diff=lfs merge=lfs -text
+*.zip filter=lfs diff=lfs merge=lfs -text
+*.zst filter=lfs diff=lfs merge=lfs -text
+*tfevents* filter=lfs diff=lfs merge=lfs -text
+*.egg filter=lfs diff=lfs merge=lfs -text
+**/tokenizer.json filter=lfs diff=lfs merge=lfs -text
+*.mp4 filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,81 @@
+
+
+
+
+*mp4
+!examples/*.mp4
+data/*
+logs/*
+
+
+
+*pyc
+
+checkpoints/*
+
+
+
+
+
+*egg-info
+
+frames
+
+
+
+
+*png
+
+*gif
+
+
+*ipynb
+daniel_tools
+daniel_tools/*
+
+
+*jpg
+
+
+build
+
+
+run*sh
+
+
+.m*
+
+
+
+
+
+
+
+scripts/*
+
+
+
+*.sh
+wandb
+benchmark
+*jsonl
+*json
+*npz
+DKT_models
+trash
+gradio
+tmp*
+*.webp
+*.ico
+*.model
+__pycache__/
+*.pyc
+**/tokenizer_configs/**/vocab.txt
+**/tokenizer_configs/**/spiece.model
+**/tokenizer_configs/**/tokenizer.model
+
+
+
+dist
+build
+

README.md

@@ -0,0 +1,14 @@
+---
+title: DKT 1
+emoji: 🏃
+colorFrom: green
+colorTo: gray
+sdk: gradio
+sdk_version: 6.0.2
+app_file: app.py
+pinned: false
+license: apache-2.0
+short_description: DKT-1
+---
+
+Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference

app.py

@@ -0,0 +1,690 @@
+
+import os
+
+import os
+import subprocess
+
+def install(package):
+    subprocess.check_call([os.sys.executable, "-m", "pip", "install", package])
+
+install("torchvision")
+install("loguru")
+install("imageio")
+install("modelscope")
+install("einops")
+install("safetensors")
+install("transformers")
+install("ftfy")
+install("accelerate")
+install("sentencepiece")
+install("spaces")
+install("opencv-python")
+install("trimesh")
+install("gradio_litmodel3d")
+install("open3d")
+
+
+
+
+
+import gradio as gr
+import numpy as np
+import torch
+from PIL import Image
+from loguru import logger
+from tqdm import tqdm
+from tools.common_utils import save_video
+from dkt.pipelines.wan_video_new import WanVideoPipeline, ModelConfig
+try:
+    import gradio_client.utils as _gc_utils
+    if hasattr(_gc_utils, "get_type"):
+        _orig_get_type = _gc_utils.get_type
+        def _get_type_safe(schema):
+            if not isinstance(schema, dict):
+                return "Any"
+            return _orig_get_type(schema)
+        _gc_utils.get_type = _get_type_safe
+except Exception:
+    pass
+
+# Additional guard: handle boolean JSON Schemas and parsing errors
+try:
+    import gradio_client.utils as _gc_utils
+    # Wrap the internal _json_schema_to_python_type if present
+    if hasattr(_gc_utils, "_json_schema_to_python_type"):
+        _orig_internal = _gc_utils._json_schema_to_python_type
+        def _json_schema_to_python_type_safe(schema, defs=None):
+            if isinstance(schema, bool):
+                return "Any"
+            try:
+                return _orig_internal(schema, defs)
+            except Exception:
+                return "Any"
+        _gc_utils._json_schema_to_python_type = _json_schema_to_python_type_safe
+
+    # Also wrap the public json_schema_to_python_type to be extra defensive
+    if hasattr(_gc_utils, "json_schema_to_python_type"):
+        _orig_public = _gc_utils.json_schema_to_python_type
+        def json_schema_to_python_type_safe(schema):
+            try:
+                return _orig_public(schema)
+            except Exception:
+                return "Any"
+        _gc_utils.json_schema_to_python_type = json_schema_to_python_type_safe
+except Exception:
+    pass
+
+import cv2
+import copy
+import trimesh
+from gradio_litmodel3d import LitModel3D
+from os.path import join
+from tools.depth2pcd import depth2pcd
+
+try:
+    from moge.model.v2 import MoGeModel
+
+except:
+    os.system('pip install git+https://github.com/microsoft/MoGe.git -i https://pypi.org/simple/  --trusted-host pypi.org --trusted-host pypi.python.org --trusted-host files.pythonhosted.org')
+    from moge.model.v2 import MoGeModel
+
+
+from tools.eval_utils import transfer_pred_disp2depth, colorize_depth_map
+import glob
+import datetime
+import shutil
+import tempfile
+import spaces
+
+
+PIPE_1_3B = None
+MOGE_MODULE =  None
+#* better for bg:  logs/outs/train/remote/sft-T2SQNet_glassverse_cleargrasp_HISS_DREDS_DREDS_glassverse_interiorverse-4gpus-origin-lora128-1.3B-rgb_depth-w832-h480-Wan2.1-Fun-Control-2025-10-28-23:26:41/epoch-0-20000.safetensors
+PROMPT = 'depth'
+NEGATIVE_PROMPT = ''
+
+
+
+def resize_frame(frame, height, width):
+    frame = np.array(frame)
+    frame = torch.from_numpy(frame).permute(2, 0, 1).unsqueeze(0).float() / 255.0
+    frame = torch.nn.functional.interpolate(frame, (height, width), mode="bicubic", align_corners=False, antialias=True)
+    frame = (frame.squeeze(0).permute(1, 2, 0).clamp(0, 1) * 255).byte().numpy()
+    frame = Image.fromarray(frame)
+    return frame
+
+
+
+def pmap_to_glb(point_map, valid_mask, frame) -> trimesh.Scene:
+    pts_3d = point_map[valid_mask] * np.array([-1, -1, 1])
+    pts_rgb = frame[valid_mask] 
+
+    # Initialize a 3D scene
+    scene_3d = trimesh.Scene()
+
+    # Add point cloud data to the scene
+    point_cloud_data = trimesh.PointCloud(
+        vertices=pts_3d, colors=pts_rgb
+    )
+    
+    scene_3d.add_geometry(point_cloud_data)
+    return scene_3d
+
+
+
+def create_simple_glb_from_pointcloud(points, colors, glb_filename):
+    try:
+        if len(points) == 0:
+            logger.warning(f"No valid points to create GLB for {glb_filename}")
+            return False
+        
+        if colors is not None:
+            # logger.info(f"Adding colors to GLB: shape={colors.shape}, range=[{colors.min():.3f}, {colors.max():.3f}]")
+            pts_rgb = colors
+        else:
+            logger.info("No colors provided, adding default white colors")
+            pts_rgb = np.ones((len(points), 3))
+        
+        valid_mask = np.ones(len(points), dtype=bool)
+        
+        scene_3d = pmap_to_glb(points, valid_mask, pts_rgb)
+        
+        scene_3d.export(glb_filename)
+        # logger.info(f"Saved GLB file using trimesh: {glb_filename}")
+        
+        return True
+        
+    except Exception as e:
+        logger.error(f"Error creating GLB from pointcloud using trimesh: {str(e)}")
+        return False
+
+
+
+
+
+def extract_frames_from_video_file(video_path):
+    try:
+        cap = cv2.VideoCapture(video_path)
+        frames = []
+        
+        fps = cap.get(cv2.CAP_PROP_FPS)
+        if fps <= 0:
+            fps = 15.0
+        
+        while True:
+            ret, frame = cap.read()
+            if not ret:
+                break
+            frame_rgb = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+            frame_rgb = Image.fromarray(frame_rgb)
+            frames.append(frame_rgb)
+        
+        cap.release()
+        return frames, fps
+    except Exception as e:
+        logger.error(f"Error extracting frames from {video_path}: {str(e)}")
+        return [], 15.0
+
+
+
+def load_moge_model(device="cuda:0"):
+    global MOGE_MODULE
+    if MOGE_MODULE is not None:
+        return MOGE_MODULE
+    logger.info(f"Loading MoGe model on {device}...")
+    MOGE_MODULE = MoGeModel.from_pretrained('Ruicheng/moge-2-vitl-normal').to(device)
+    return MOGE_MODULE
+
+
+def load_model_1_3b(device="cuda:0"):
+    global PIPE_1_3B
+    
+    if PIPE_1_3B is not None:
+        return PIPE_1_3B
+    
+    logger.info(f"Loading 1.3B model on {device}...")
+    
+    pipe = WanVideoPipeline.from_pretrained(
+        torch_dtype=torch.bfloat16,
+        device=device,
+        model_configs=[
+            ModelConfig(
+                model_id="PAI/Wan2.1-Fun-1.3B-Control",
+                origin_file_pattern="diffusion_pytorch_model*.safetensors",
+                offload_device="cpu",
+            ),
+            ModelConfig(
+                model_id="PAI/Wan2.1-Fun-1.3B-Control",
+                origin_file_pattern="models_t5_umt5-xxl-enc-bf16.pth",
+                offload_device="cpu",
+            ),
+            ModelConfig(
+                model_id="PAI/Wan2.1-Fun-1.3B-Control",
+                origin_file_pattern="Wan2.1_VAE.pth",
+                offload_device="cpu",
+            ),
+            ModelConfig(
+                model_id="PAI/Wan2.1-Fun-1.3B-Control",
+                origin_file_pattern="models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth",
+                offload_device="cpu",
+            ),
+        ],
+        training_strategy="origin",
+    )
+    
+    
+    lora_config = ModelConfig(
+        model_id="Daniellesry/DKT-Depth-1-3B",
+        origin_file_pattern="dkt-1-3B.safetensors",
+        offload_device="cpu",
+    )
+
+    lora_config.download_if_necessary(use_usp=False)
+    
+    pipe.load_lora(pipe.dit, lora_config.path, alpha=1.0)#todo is it work?
+    pipe.enable_vram_management()
+
+    
+    PIPE_1_3B = pipe
+    
+    return pipe
+
+
+
+
+
+    
+
+def get_model(model_size):
+    if model_size == "1.3B":
+        assert PIPE_1_3B is not None, "1.3B model not initialized"
+        return PIPE_1_3B
+    else:
+        raise ValueError(f"Unsupported model size: {model_size}")
+
+
+
+def process_video(
+    video_file,
+    model_size,
+    height,
+    width,
+    num_inference_steps,
+    window_size,
+    overlap
+):
+    try:
+        pipe = get_model(model_size)
+        if pipe is None:
+            return None, f"Model {model_size} not initialized. Please restart the application."
+        
+        tmp_video_path = video_file
+        timestamp = datetime.datetime.now().strftime("%Y%m%d_%H%M%S")
+
+        # 使用临时目录存储所有文件
+        cur_save_dir = tempfile.mkdtemp(prefix=f'dkt_{timestamp}_{model_size}_')
+        
+        
+        original_filename = f"input_{timestamp}.mp4"
+        dst_path = os.path.join(cur_save_dir, original_filename)
+        shutil.copy2(tmp_video_path, dst_path)
+        origin_frames, input_fps = extract_frames_from_video_file(tmp_video_path)
+        
+        if not origin_frames:
+            return None, "Failed to extract frames from video"
+        
+        logger.info(f"Extracted {len(origin_frames)} frames from video")
+
+        
+        original_width, original_height = origin_frames[0].size
+        ROTATE = False 
+        if original_width <  original_height:
+            ROTATE = True
+            origin_frames = [x.transpose(Image.ROTATE_90) for x in origin_frames]
+            tmp = original_width
+            original_width = original_height
+            original_height = tmp
+
+        
+                
+        frames = [resize_frame(frame, height, width) for frame in origin_frames]
+        frame_length = len(frames)
+        if (frame_length - 1) % 4 != 0:
+            new_len = ((frame_length - 1) // 4 + 1) * 4 + 1
+            frames = frames + [copy.deepcopy(frames[-1]) for _ in range(new_len - frame_length)]
+
+        
+        control_video = frames
+        video, vae_outs = pipe(
+            prompt=PROMPT,
+            negative_prompt=NEGATIVE_PROMPT,
+            control_video=control_video,
+            height=height,
+            width=width,
+            num_frames=len(control_video),
+            seed=1,
+            tiled=False,
+            num_inference_steps=num_inference_steps,
+            sliding_window_size=window_size,
+            sliding_window_stride=window_size - overlap,
+            cfg_scale=1.0,
+        )
+
+        #* moge process 
+        torch.cuda.empty_cache()
+        processed_video = video[:frame_length]
+
+
+        processed_video = [resize_frame(frame, original_height, original_width) for frame in processed_video]
+        if ROTATE:
+            processed_video = [x.transpose(Image.ROTATE_270) for x in processed_video]
+            origin_frames = [x.transpose(Image.ROTATE_270) for x in origin_frames]
+            
+
+        output_filename = f"output_{timestamp}.mp4"
+        output_path = os.path.join(cur_save_dir, output_filename)
+        color_predictions = []
+        if PROMPT == 'depth':
+            prediced_depth_map_np = [np.array(item).astype(np.float32).mean(-1)  for item in processed_video]
+            prediced_depth_map_np = np.stack(prediced_depth_map_np)
+            prediced_depth_map_np = prediced_depth_map_np/ 255.0
+            __min = prediced_depth_map_np.min()
+            __max = prediced_depth_map_np.max()
+            prediced_depth_map_np = (prediced_depth_map_np - __min) / (__max - __min)
+            color_predictions = [colorize_depth_map(item) for item in prediced_depth_map_np]
+        else:
+            color_predictions = processed_video
+        save_video(color_predictions, output_path, fps=input_fps, quality=5)
+
+
+        
+        frame_num = len(origin_frames)
+        resize_W,resize_H = origin_frames[0].size
+        
+        vis_pc_num = 4
+        indices = np.linspace(0, frame_num-1, vis_pc_num)
+        indices = np.round(indices).astype(np.int32)
+        pc_save_dir = os.path.join(cur_save_dir, 'pointclouds')
+        os.makedirs(pc_save_dir, exist_ok=True)
+
+        glb_files = []
+        moge_device = MOGE_MODULE.device if MOGE_MODULE is not None else torch.device("cuda:0")
+        for idx in tqdm(indices):
+            orgin_rgb_frame = origin_frames[idx]
+            predicted_depth = processed_video[idx]
+
+            # Read the input image and convert to tensor (3, H, W) with RGB values normalized to [0, 1]
+            input_image_np = np.array(orgin_rgb_frame)  # Convert PIL Image to numpy array
+            input_image = torch.tensor(input_image_np / 255, dtype=torch.float32, device=moge_device).permute(2, 0, 1) 
+
+            output = MOGE_MODULE.infer(input_image)
+            #* "dict_keys(['points', 'intrinsics', 'depth', 'mask', 'normal'])"
+            moge_intrinsics = output['intrinsics'].cpu().numpy()
+            moge_mask = output['mask'].cpu().numpy()
+            moge_depth = output['depth'].cpu().numpy()
+        
+            predicted_depth = np.array(predicted_depth)
+            predicted_depth = predicted_depth.mean(-1) / 255.0
+    
+            metric_depth = transfer_pred_disp2depth(predicted_depth, moge_depth, moge_mask)
+            
+            moge_intrinsics[0, 0] *= resize_W 
+            moge_intrinsics[1, 1] *= resize_H
+            moge_intrinsics[0, 2] *= resize_W
+            moge_intrinsics[1, 2] *= resize_H
+            
+            # pcd = depth2pcd(metric_depth, moge_intrinsics, color=cv2.cvtColor(input_image_np, cv2.COLOR_BGR2RGB), input_mask=moge_mask, ret_pcd=True)
+            pcd = depth2pcd(metric_depth, moge_intrinsics, color=input_image_np, input_mask=moge_mask, ret_pcd=True)
+            
+            # pcd.points = o3d.utility.Vector3dVector(np.asarray(pcd.points) * np.array([1, -1, -1], dtype=np.float32))
+            
+            apply_filter = True
+            if apply_filter:
+                cl, ind = pcd.remove_statistical_outlier(nb_neighbors=20, std_ratio=3.0)
+                pcd = pcd.select_by_index(ind)
+
+            #* save pcd:  o3d.io.write_point_cloud(f'{pc_save_dir}/{timestamp}_{idx:02d}.ply', pcd)
+            points = np.asarray(pcd.points)
+            colors = np.asarray(pcd.colors) if pcd.has_colors() else None
+            glb_filename = os.path.join(pc_save_dir, f'{timestamp}_{idx:02d}.glb')
+            success = create_simple_glb_from_pointcloud(points, colors, glb_filename)
+            if not success:
+                logger.warning(f"Failed to save GLB file: {glb_filename}")
+
+            glb_files.append(glb_filename)
+
+        return output_path, glb_files
+        
+    except Exception as e:
+        logger.error(f"Error processing video: {str(e)}")
+        return None, f"Error: {str(e)}"
+
+
+
+
+def main():
+
+
+
+    #* gradio creation and initialization
+
+
+    css = """
+    #video-display-container {
+        max-height: 100vh;
+    }
+    #video-display-input {
+        max-height: 80vh;
+    }
+    #video-display-output {
+        max-height: 80vh;
+    }
+    #download {
+        height: 62px;
+    }
+    .title {
+        text-align: center;
+    }
+    .description {
+        text-align: center;
+    }
+    .gradio-examples {
+        max-height: 400px;
+        overflow-y: auto;
+    }
+    .gradio-examples .examples-container {
+        display: grid;
+        grid-template-columns: repeat(auto-fit, minmax(200px, 1fr));
+        gap: 10px;
+        padding: 10px;
+    }
+    .gradio-container .gradio-examples .pagination,
+    .gradio-container .gradio-examples .pagination button,
+    div[data-testid="examples"] .pagination,
+    div[data-testid="examples"] .pagination button {
+        font-size: 28px !important;
+        font-weight: bold !important;
+        padding: 15px 20px !important;
+        min-width: 60px !important;
+        height: 60px !important;
+        border-radius: 10px !important;
+        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
+        color: white !important;
+        border: none !important;
+        cursor: pointer !important;
+        margin: 8px !important;
+        display: inline-block !important;
+        box-shadow: 0 4px 8px rgba(0,0,0,0.2) !important;
+        transition: all 0.3s ease !important;
+    }
+
+    div[data-testid="examples"] .pagination button:not(.active),
+    .gradio-container .gradio-examples .pagination button:not(.active) {
+        font-size: 32px !important;
+        font-weight: bold !important;
+        padding: 15px 20px !important;
+        min-width: 60px !important;
+        height: 60px !important;
+        background: linear-gradient(135deg, #8a9cf0 0%, #9a6bb2 100%) !important;
+        opacity: 0.8 !important;
+    }
+
+    div[data-testid="examples"] .pagination button:hover,
+    .gradio-container .gradio-examples .pagination button:hover {
+        background: linear-gradient(135deg, #5a6fd8 0%, #6a4190 100%) !important;
+        transform: translateY(-2px) !important;
+        box-shadow: 0 6px 12px rgba(0,0,0,0.3) !important;
+        opacity: 1 !important;
+    }
+
+    div[data-testid="examples"] .pagination button.active,
+    .gradio-container .gradio-examples .pagination button.active {
+        background: linear-gradient(135deg, #11998e 0%, #38ef7d 100%) !important;
+        box-shadow: 0 4px 8px rgba(17,153,142,0.4) !important;
+        opacity: 1 !important;
+    }
+
+    button[class*="pagination"],
+    button[class*="page"] {
+        font-size: 28px !important;
+        font-weight: bold !important;
+        padding: 15px 20px !important;
+        min-width: 60px !important;
+        height: 60px !important;
+        border-radius: 10px !important;
+        background: linear-gradient(135deg, #667eea 0%, #764ba2 100%) !important;
+        color: white !important;
+        border: none !important;
+        cursor: pointer !important;
+        margin: 8px !important;
+        box-shadow: 0 4px 8px rgba(0,0,0,0.2) !important;
+        transition: all 0.3s ease !important;
+    }
+    """
+
+
+
+    head_html = """
+    <link rel="icon" type="image/svg+xml" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'%3E%3Ctext y='.9em' font-size='90'%3E🦾%3C/text%3E%3C/svg%3E">
+    <link rel="shortcut icon" type="image/svg+xml" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'%3E%3Ctext y='.9em' font-size='90'%3E🦾%3C/text%3E%3C/svg%3E">
+    <link rel="icon" type="image/png" href="data:image/svg+xml,%3Csvg xmlns='http://www.w3.org/2000/svg' viewBox='0 0 100 100'%3E%3Ctext y='.9em' font-size='90'%3E🦾%3C/text%3E%3C/svg%3E">
+    <meta name="viewport" content="width=device-width, initial-scale=1.0">
+    """
+
+
+    # title = "# Diffusion Knows Transparency: Repurposing Video Diffusion for Transparent Object Depth and Normal Estimation "
+    # description = """Official demo for **DKT **."""
+
+    # with gr.Blocks(css=css, title="DKT - Diffusion Knows Transparency", favicon_path="favicon.ico") as demo:
+
+    height = 480
+    width = 832
+    window_size = 21
+    with gr.Blocks(css=css, title="DKT", head=head_html) as demo:
+        # gr.Markdown(title, elem_classes=["title"])
+        # gr.Markdown(description, elem_classes=["description"])
+        # gr.Markdown("### Video Processing Demo", elem_classes=["description"])
+
+        with gr.Row():
+            with gr.Column():
+                input_video = gr.Video(label="Input Video", elem_id='video-display-input')
+                
+                model_size = gr.Radio(
+                    choices=["1.3B", "14B"],
+                    value="1.3B",
+                    label="Model Size"
+                )
+
+
+                with gr.Accordion("Advanced Parameters", open=False):
+                    num_inference_steps = gr.Slider(
+                        minimum=1, maximum=50, value=5, step=1,
+                        label="Number of Inference Steps"
+                    )
+                    overlap = gr.Slider(
+                        minimum=1, maximum=20, value=3, step=1,
+                        label="Overlap"
+                    )
+                    
+                submit = gr.Button(value="Compute Depth", variant="primary")
+            
+            with gr.Column():
+                output_video = gr.Video(
+                    label="Depth Outputs", 
+                    elem_id='video-display-output',
+                    autoplay=True
+                )
+                vis_video = gr.Video(
+                    label="Visualization Video", 
+                    visible=False,
+                    autoplay=True
+                )
+
+        with gr.Row():
+            gr.Markdown("### 3D Point Cloud Visualization", elem_classes=["title"])
+        
+        with gr.Row(equal_height=True):
+            with gr.Column(scale=1):
+                output_point_map0 = LitModel3D(
+                    label="Point Cloud Key Frame 1",
+                    clear_color=[1.0, 1.0, 1.0, 1.0],
+                    interactive=False,
+                    # height=400,
+                    
+                )
+            with gr.Column(scale=1):
+                output_point_map1 = LitModel3D(
+                    label="Point Cloud Key Frame 2",
+                    clear_color=[1.0, 1.0, 1.0, 1.0],
+                    interactive=False
+                )
+            
+        
+        with gr.Row(equal_height=True):
+            
+            with gr.Column(scale=1):
+                output_point_map2 = LitModel3D(
+                    label="Point Cloud Key Frame 3",
+                    clear_color=[1.0, 1.0, 1.0, 1.0],
+                    interactive=False
+                )
+            with gr.Column(scale=1):
+                output_point_map3 = LitModel3D(
+                    label="Point Cloud Key Frame 4",
+                    clear_color=[1.0, 1.0, 1.0, 1.0],
+                    interactive=False
+                )
+
+        def on_submit(video_file, model_size, num_inference_steps, overlap):
+            if video_file is None:
+                return None, None, None, None, None, None, "Please upload a video file"
+            
+            try:
+
+                output_path, glb_files = process_video(
+                    video_file, model_size, height, width, num_inference_steps, window_size, overlap
+                )
+
+                
+                
+                if output_path is None:
+                    return None, None, None, None, None, None, glb_files
+                
+                model3d_outputs = [None] * 4
+                if glb_files:
+                    for i, glb_file in enumerate(glb_files[:4]):
+                        if os.path.exists(glb_file):
+                            model3d_outputs[i] = glb_file
+                
+
+
+                return output_path, None, *model3d_outputs
+                    
+            except Exception as e:
+                return None, None, None, None, None, None, f"Error: {str(e)}"
+
+        
+        submit.click(
+            on_submit,
+            inputs=[
+                input_video, model_size, num_inference_steps, overlap
+            ],
+            outputs=[
+                output_video, vis_video, 
+                output_point_map0, output_point_map1, output_point_map2, output_point_map3
+            ]
+        )
+        
+        
+        
+        example_files = glob.glob('examples/*')
+        if example_files:
+            example_inputs = []
+            for file_path in example_files:
+                example_inputs.append([file_path, "1.3B"])
+            
+            examples = gr.Examples(
+                examples=example_inputs, 
+                inputs=[input_video, model_size], 
+                outputs=[
+                    output_video, vis_video, 
+                    output_point_map0, output_point_map1, output_point_map2, output_point_map3
+                ], 
+                fn=on_submit,
+                examples_per_page=6
+            )
+
+
+    #* main code, model and moge model initialization
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    load_model_1_3b(device=device)
+    load_moge_model(device=device)
+    torch.cuda.empty_cache()
+
+    demo.queue().launch(share = True,server_name="0.0.0.0", server_port=7860)
+
+
+if __name__ == '__main__':
+    process_video = spaces.GPU(process_video)
+    
+    main()

dkt/__init__.py

@@ -0,0 +1,4 @@
+from .models import *
+from .prompters import *
+from .schedulers import *
+from .pipelines import *

dkt/configs/model_config.py

@@ -0,0 +1,68 @@
+from typing_extensions import Literal, TypeAlias
+
+
+from ..models.wan_video_dit import WanModel
+from ..models.wan_video_text_encoder import WanTextEncoder
+from ..models.wan_video_image_encoder import WanImageEncoder
+from ..models.wan_video_vae import WanVideoVAE, WanVideoVAE38
+from ..models.wan_video_motion_controller import WanMotionControllerModel
+from ..models.wan_video_vace import VaceWanModel
+model_loader_configs = [
+    # These configs are provided for detecting model type automatically.
+    # The format is (state_dict_keys_hash, state_dict_keys_hash_with_shape, model_names, model_classes, model_resource)
+    (None, "9269f8db9040a9d860eaca435be61814", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "aafcfd9672c3a2456dc46e1cb6e52c70", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "6bfcfb3b342cb286ce886889d519a77e", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "6d6ccde6845b95ad9114ab993d917893", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "6bfcfb3b342cb286ce886889d519a77e", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "349723183fc063b2bfc10bb2835cf677", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "efa44cddf936c70abd0ea28b6cbe946c", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "3ef3b1f8e1dab83d5b71fd7b617f859f", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "70ddad9d3a133785da5ea371aae09504", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "26bde73488a92e64cc20b0a7485b9e5b", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "ac6a5aa74f4a0aab6f64eb9a72f19901", ["wan_video_dit"], [WanModel], "civitai"), 
+    (None, "b61c605c2adbd23124d152ed28e049ae", ["wan_video_dit"], [WanModel], "civitai"), 
+    (None, "1f5ab7703c6fc803fdded85ff040c316", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "5b013604280dd715f8457c6ed6d6a626", ["wan_video_dit"], [WanModel], "civitai"),
+    (None, "a61453409b67cd3246cf0c3bebad47ba", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
+    (None, "7a513e1f257a861512b1afd387a8ecd9", ["wan_video_dit", "wan_video_vace"], [WanModel, VaceWanModel], "civitai"),
+    (None, "cb104773c6c2cb6df4f9529ad5c60d0b", ["wan_video_dit"], [WanModel], "diffusers"),
+    (None, "9c8818c2cbea55eca56c7b447df170da", ["wan_video_text_encoder"], [WanTextEncoder], "civitai"),
+    (None, "5941c53e207d62f20f9025686193c40b", ["wan_video_image_encoder"], [WanImageEncoder], "civitai"),
+    (None, "1378ea763357eea97acdef78e65d6d96", ["wan_video_vae"], [WanVideoVAE], "civitai"),
+    (None, "ccc42284ea13e1ad04693284c7a09be6", ["wan_video_vae"], [WanVideoVAE], "civitai"),
+    (None, "e1de6c02cdac79f8b739f4d3698cd216", ["wan_video_vae"], [WanVideoVAE38], "civitai"),
+    (None, "dbd5ec76bbf977983f972c151d545389", ["wan_video_motion_controller"], [WanMotionControllerModel], "civitai"),
+]
+huggingface_model_loader_configs = [
+    # These configs are provided for detecting model type automatically.
+    # The format is (architecture_in_huggingface_config, huggingface_lib, model_name, redirected_architecture)
+    ("ChatGLMModel", "dkt.models.kolors_text_encoder", "kolors_text_encoder", None),
+    ("MarianMTModel", "transformers.models.marian.modeling_marian", "translator", None),
+    ("BloomForCausalLM", "transformers.models.bloom.modeling_bloom", "beautiful_prompt", None),
+    ("Qwen2ForCausalLM", "transformers.models.qwen2.modeling_qwen2", "qwen_prompt", None),
+    # ("LlamaForCausalLM", "transformers.models.llama.modeling_llama", "omost_prompt", None),
+    ("T5EncoderModel", "dkt.models.flux_text_encoder", "flux_text_encoder_2", "FluxTextEncoder2"),
+    ("CogVideoXTransformer3DModel", "dkt.models.cog_dit", "cog_dit", "CogDiT"),
+    ("SiglipModel", "transformers.models.siglip.modeling_siglip", "siglip_vision_model", "SiglipVisionModel"),
+    ("LlamaForCausalLM", "dkt.models.hunyuan_video_text_encoder", "hunyuan_video_text_encoder_2", "HunyuanVideoLLMEncoder"),
+    ("LlavaForConditionalGeneration", "dkt.models.hunyuan_video_text_encoder", "hunyuan_video_text_encoder_2", "HunyuanVideoMLLMEncoder"),
+    ("Step1Model", "dkt.models.stepvideo_text_encoder", "stepvideo_text_encoder_2", "STEP1TextEncoder"),
+    ("Qwen2_5_VLForConditionalGeneration", "dkt.models.qwenvl", "qwenvl", "Qwen25VL_7b_Embedder"),
+]
+patch_model_loader_configs = [
+    # These configs are provided for detecting model type automatically.
+    # The format is (state_dict_keys_hash_with_shape, model_name, model_class, extra_kwargs)
+    # ("9a4ab6869ac9b7d6e31f9854e397c867", ["svd_unet"], [SVDUNet], {"add_positional_conv": 128}),
+]
+
+preset_models_on_huggingface = {
+    
+}
+preset_models_on_modelscope = {
+
+}
+Preset_model_id: TypeAlias = Literal[
+    ...
+    
+]

dkt/lora/__init__.py

@@ -0,0 +1,45 @@
+import torch
+
+
+
+class GeneralLoRALoader:
+    def __init__(self, device="cpu", torch_dtype=torch.float32):
+        self.device = device
+        self.torch_dtype = torch_dtype
+    
+    
+    def get_name_dict(self, lora_state_dict):
+        lora_name_dict = {}
+        for key in lora_state_dict:
+            if ".lora_B." not in key:
+                continue
+            keys = key.split(".")
+            if len(keys) > keys.index("lora_B") + 2:
+                keys.pop(keys.index("lora_B") + 1)
+            keys.pop(keys.index("lora_B"))
+            if keys[0] == "diffusion_model":
+                keys.pop(0)
+            keys.pop(-1)
+            target_name = ".".join(keys)
+            lora_name_dict[target_name] = (key, key.replace(".lora_B.", ".lora_A."))
+        return lora_name_dict
+
+
+    def load(self, model: torch.nn.Module, state_dict_lora, alpha=1.0):
+        updated_num = 0
+        lora_name_dict = self.get_name_dict(state_dict_lora)
+        for name, module in model.named_modules():
+            if name in lora_name_dict:
+                weight_up = state_dict_lora[lora_name_dict[name][0]].to(device=self.device, dtype=self.torch_dtype)
+                weight_down = state_dict_lora[lora_name_dict[name][1]].to(device=self.device, dtype=self.torch_dtype)
+                if len(weight_up.shape) == 4:
+                    weight_up = weight_up.squeeze(3).squeeze(2)
+                    weight_down = weight_down.squeeze(3).squeeze(2)
+                    weight_lora = alpha * torch.mm(weight_up, weight_down).unsqueeze(2).unsqueeze(3)
+                else:
+                    weight_lora = alpha * torch.mm(weight_up, weight_down)
+                state_dict = module.state_dict()
+                state_dict["weight"] = state_dict["weight"].to(device=self.device, dtype=self.torch_dtype) + weight_lora
+                module.load_state_dict(state_dict)
+                updated_num += 1
+        print(f"{updated_num} tensors are updated by LoRA.")

dkt/models/__init__.py

@@ -0,0 +1 @@
+from .model_manager import *

dkt/models/attention.py

@@ -0,0 +1,89 @@
+import torch
+from einops import rearrange
+
+
+def low_version_attention(query, key, value, attn_bias=None):
+    scale = 1 / query.shape[-1] ** 0.5
+    query = query * scale
+    attn = torch.matmul(query, key.transpose(-2, -1))
+    if attn_bias is not None:
+        attn = attn + attn_bias
+    attn = attn.softmax(-1)
+    return attn @ value
+
+
+class Attention(torch.nn.Module):
+
+    def __init__(self, q_dim, num_heads, head_dim, kv_dim=None, bias_q=False, bias_kv=False, bias_out=False):
+        super().__init__()
+        dim_inner = head_dim * num_heads
+        kv_dim = kv_dim if kv_dim is not None else q_dim
+        self.num_heads = num_heads
+        self.head_dim = head_dim
+
+        self.to_q = torch.nn.Linear(q_dim, dim_inner, bias=bias_q)
+        self.to_k = torch.nn.Linear(kv_dim, dim_inner, bias=bias_kv)
+        self.to_v = torch.nn.Linear(kv_dim, dim_inner, bias=bias_kv)
+        self.to_out = torch.nn.Linear(dim_inner, q_dim, bias=bias_out)
+
+    def interact_with_ipadapter(self, hidden_states, q, ip_k, ip_v, scale=1.0):
+        batch_size = q.shape[0]
+        ip_k = ip_k.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
+        ip_v = ip_v.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
+        ip_hidden_states = torch.nn.functional.scaled_dot_product_attention(q, ip_k, ip_v)
+        hidden_states = hidden_states + scale * ip_hidden_states
+        return hidden_states
+
+    def torch_forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None, ipadapter_kwargs=None, qkv_preprocessor=None):
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        batch_size = encoder_hidden_states.shape[0]
+
+        q = self.to_q(hidden_states)
+        k = self.to_k(encoder_hidden_states)
+        v = self.to_v(encoder_hidden_states)
+
+        q = q.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
+        k = k.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
+        v = v.view(batch_size, -1, self.num_heads, self.head_dim).transpose(1, 2)
+
+        if qkv_preprocessor is not None:
+            q, k, v = qkv_preprocessor(q, k, v)
+
+        hidden_states = torch.nn.functional.scaled_dot_product_attention(q, k, v, attn_mask=attn_mask)
+        if ipadapter_kwargs is not None:
+            hidden_states = self.interact_with_ipadapter(hidden_states, q, **ipadapter_kwargs)
+        hidden_states = hidden_states.transpose(1, 2).reshape(batch_size, -1, self.num_heads * self.head_dim)
+        hidden_states = hidden_states.to(q.dtype)
+
+        hidden_states = self.to_out(hidden_states)
+
+        return hidden_states
+    
+    def xformers_forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None):
+        if encoder_hidden_states is None:
+            encoder_hidden_states = hidden_states
+
+        q = self.to_q(hidden_states)
+        k = self.to_k(encoder_hidden_states)
+        v = self.to_v(encoder_hidden_states)
+
+        q = rearrange(q, "b f (n d) -> (b n) f d", n=self.num_heads)
+        k = rearrange(k, "b f (n d) -> (b n) f d", n=self.num_heads)
+        v = rearrange(v, "b f (n d) -> (b n) f d", n=self.num_heads)
+
+        if attn_mask is not None:
+            hidden_states = low_version_attention(q, k, v, attn_bias=attn_mask)
+        else:
+            import xformers.ops as xops
+            hidden_states = xops.memory_efficient_attention(q, k, v)
+        hidden_states = rearrange(hidden_states, "(b n) f d -> b f (n d)", n=self.num_heads)
+
+        hidden_states = hidden_states.to(q.dtype)
+        hidden_states = self.to_out(hidden_states)
+
+        return hidden_states
+
+    def forward(self, hidden_states, encoder_hidden_states=None, attn_mask=None, ipadapter_kwargs=None, qkv_preprocessor=None):
+        return self.torch_forward(hidden_states, encoder_hidden_states=encoder_hidden_states, attn_mask=attn_mask, ipadapter_kwargs=ipadapter_kwargs, qkv_preprocessor=qkv_preprocessor)

dkt/models/downloader.py

@@ -0,0 +1,111 @@
+from huggingface_hub import hf_hub_download
+from modelscope import snapshot_download
+import os, shutil
+from typing_extensions import Literal, TypeAlias
+from typing import List
+from ..configs.model_config import preset_models_on_huggingface, preset_models_on_modelscope, Preset_model_id
+
+
+def download_from_modelscope(model_id, origin_file_path, local_dir):
+    os.makedirs(local_dir, exist_ok=True)
+    file_name = os.path.basename(origin_file_path)
+    if file_name in os.listdir(local_dir):
+        print(f"    {file_name} has been already in {local_dir}.")
+    else:
+        print(f"    Start downloading {os.path.join(local_dir, file_name)}")
+        snapshot_download(model_id, allow_file_pattern=origin_file_path, local_dir=local_dir)
+        downloaded_file_path = os.path.join(local_dir, origin_file_path)
+        target_file_path = os.path.join(local_dir, os.path.split(origin_file_path)[-1])
+        if downloaded_file_path != target_file_path:
+            shutil.move(downloaded_file_path, target_file_path)
+            shutil.rmtree(os.path.join(local_dir, origin_file_path.split("/")[0]))
+
+
+def download_from_huggingface(model_id, origin_file_path, local_dir):
+    os.makedirs(local_dir, exist_ok=True)
+    file_name = os.path.basename(origin_file_path)
+    if file_name in os.listdir(local_dir):
+        print(f"    {file_name} has been already in {local_dir}.")
+    else:
+        print(f"    Start downloading {os.path.join(local_dir, file_name)}")
+        hf_hub_download(model_id, origin_file_path, local_dir=local_dir)
+        downloaded_file_path = os.path.join(local_dir, origin_file_path)
+        target_file_path = os.path.join(local_dir, file_name)
+        if downloaded_file_path != target_file_path:
+            shutil.move(downloaded_file_path, target_file_path)
+            shutil.rmtree(os.path.join(local_dir, origin_file_path.split("/")[0]))
+
+
+Preset_model_website: TypeAlias = Literal[
+    "HuggingFace",
+    "ModelScope",
+]
+website_to_preset_models = {
+    "HuggingFace": preset_models_on_huggingface,
+    "ModelScope": preset_models_on_modelscope,
+}
+website_to_download_fn = {
+    "HuggingFace": download_from_huggingface,
+    "ModelScope": download_from_modelscope,
+}
+
+
+def download_customized_models(
+    model_id,
+    origin_file_path,
+    local_dir,
+    downloading_priority: List[Preset_model_website] = ["ModelScope", "HuggingFace"],
+):
+    downloaded_files = []
+    for website in downloading_priority:
+        # Check if the file is downloaded.
+        file_to_download = os.path.join(local_dir, os.path.basename(origin_file_path))
+        if file_to_download in downloaded_files:
+            continue
+        # Download
+        website_to_download_fn[website](model_id, origin_file_path, local_dir)
+        if os.path.basename(origin_file_path) in os.listdir(local_dir):
+            downloaded_files.append(file_to_download)
+    return downloaded_files
+
+
+def download_models(
+    model_id_list: List[Preset_model_id] = [],
+    downloading_priority: List[Preset_model_website] = ["ModelScope", "HuggingFace"],
+):
+    print(f"Downloading models: {model_id_list}")
+    downloaded_files = []
+    load_files = []
+
+    for model_id in model_id_list:
+        for website in downloading_priority:
+            if model_id in website_to_preset_models[website]:
+                
+                # Parse model metadata
+                model_metadata = website_to_preset_models[website][model_id]
+                if isinstance(model_metadata, list):
+                    file_data = model_metadata
+                else:
+                    file_data = model_metadata.get("file_list", [])
+
+                # Try downloading the model from this website.
+                model_files = []
+                for model_id, origin_file_path, local_dir in file_data:
+                    # Check if the file is downloaded.
+                    file_to_download = os.path.join(local_dir, os.path.basename(origin_file_path))
+                    if file_to_download in downloaded_files:
+                        continue
+                    # Download
+                    website_to_download_fn[website](model_id, origin_file_path, local_dir)
+                    if os.path.basename(origin_file_path) in os.listdir(local_dir):
+                        downloaded_files.append(file_to_download)
+                        model_files.append(file_to_download)
+                
+                # If the model is successfully downloaded, break.
+                if len(model_files) > 0:
+                    if isinstance(model_metadata, dict) and "load_path" in model_metadata:
+                        model_files = model_metadata["load_path"]
+                    load_files.extend(model_files)
+                    break
+                
+    return load_files

dkt/models/lora.py

@@ -0,0 +1,197 @@
+import torch
+from .wan_video_dit import WanModel
+
+
+
+class LoRAFromCivitai:
+    def __init__(self):
+        self.supported_model_classes = []
+        self.lora_prefix = []
+        self.renamed_lora_prefix = {}
+        self.special_keys = {}
+
+
+    def convert_state_dict(self, state_dict, lora_prefix="lora_unet_", alpha=1.0):
+        for key in state_dict:
+            if ".lora_up" in key:
+                return self.convert_state_dict_up_down(state_dict, lora_prefix, alpha)
+        return self.convert_state_dict_AB(state_dict, lora_prefix, alpha)
+
+
+    def convert_state_dict_up_down(self, state_dict, lora_prefix="lora_unet_", alpha=1.0):
+        renamed_lora_prefix = self.renamed_lora_prefix.get(lora_prefix, "")
+        state_dict_ = {}
+        for key in state_dict:
+            if ".lora_up" not in key:
+                continue
+            if not key.startswith(lora_prefix):
+                continue
+            weight_up = state_dict[key].to(device="cuda", dtype=torch.float16)
+            weight_down = state_dict[key.replace(".lora_up", ".lora_down")].to(device="cuda", dtype=torch.float16)
+            if len(weight_up.shape) == 4:
+                weight_up = weight_up.squeeze(3).squeeze(2).to(torch.float32)
+                weight_down = weight_down.squeeze(3).squeeze(2).to(torch.float32)
+                lora_weight = alpha * torch.mm(weight_up, weight_down).unsqueeze(2).unsqueeze(3)
+            else:
+                lora_weight = alpha * torch.mm(weight_up, weight_down)
+            target_name = key.split(".")[0].replace(lora_prefix, renamed_lora_prefix).replace("_", ".") + ".weight"
+            for special_key in self.special_keys:
+                target_name = target_name.replace(special_key, self.special_keys[special_key])
+            state_dict_[target_name] = lora_weight.cpu()
+        return state_dict_
+    
+
+    def convert_state_dict_AB(self, state_dict, lora_prefix="", alpha=1.0, device="cuda", torch_dtype=torch.float16):
+        state_dict_ = {}
+        for key in state_dict:
+            if ".lora_B." not in key:
+                continue
+            if not key.startswith(lora_prefix):
+                continue
+            weight_up = state_dict[key].to(device=device, dtype=torch_dtype)
+            weight_down = state_dict[key.replace(".lora_B.", ".lora_A.")].to(device=device, dtype=torch_dtype)
+            if len(weight_up.shape) == 4:
+                weight_up = weight_up.squeeze(3).squeeze(2)
+                weight_down = weight_down.squeeze(3).squeeze(2)
+                lora_weight = alpha * torch.mm(weight_up, weight_down).unsqueeze(2).unsqueeze(3)
+            else:
+                lora_weight = alpha * torch.mm(weight_up, weight_down)
+            keys = key.split(".")
+            keys.pop(keys.index("lora_B"))
+            target_name = ".".join(keys)
+            target_name = target_name[len(lora_prefix):]
+            state_dict_[target_name] = lora_weight.cpu()
+        return state_dict_
+    
+
+    def load(self, model, state_dict_lora, lora_prefix, alpha=1.0, model_resource=None):
+        state_dict_model = model.state_dict()
+        state_dict_lora = self.convert_state_dict(state_dict_lora, lora_prefix=lora_prefix, alpha=alpha)
+        if model_resource == "diffusers":
+            state_dict_lora = model.__class__.state_dict_converter().from_diffusers(state_dict_lora)
+        elif model_resource == "civitai":
+            state_dict_lora = model.__class__.state_dict_converter().from_civitai(state_dict_lora)
+        if isinstance(state_dict_lora, tuple):
+            state_dict_lora = state_dict_lora[0]
+        if len(state_dict_lora) > 0:
+            print(f"    {len(state_dict_lora)} tensors are updated.")
+            for name in state_dict_lora:
+                fp8=False
+                if state_dict_model[name].dtype == torch.float8_e4m3fn:
+                    state_dict_model[name]= state_dict_model[name].to(state_dict_lora[name].dtype)
+                    fp8=True
+                state_dict_model[name] += state_dict_lora[name].to(
+                    dtype=state_dict_model[name].dtype, device=state_dict_model[name].device)
+                if fp8:
+                    state_dict_model[name] = state_dict_model[name].to(torch.float8_e4m3fn)
+            model.load_state_dict(state_dict_model)
+    
+
+    def match(self, model, state_dict_lora):
+        for lora_prefix, model_class in zip(self.lora_prefix, self.supported_model_classes):
+            if not isinstance(model, model_class):
+                continue
+            state_dict_model = model.state_dict()
+            for model_resource in ["diffusers", "civitai"]:
+                try:
+                    state_dict_lora_ = self.convert_state_dict(state_dict_lora, lora_prefix=lora_prefix, alpha=1.0)
+                    converter_fn = model.__class__.state_dict_converter().from_diffusers if model_resource == "diffusers" \
+                        else model.__class__.state_dict_converter().from_civitai
+                    state_dict_lora_ = converter_fn(state_dict_lora_)
+                    if isinstance(state_dict_lora_, tuple):
+                        state_dict_lora_ = state_dict_lora_[0]
+                    if len(state_dict_lora_) == 0:
+                        continue
+                    for name in state_dict_lora_:
+                        if name not in state_dict_model:
+                            break
+                    else:
+                        return lora_prefix, model_resource
+                except:
+                    pass
+        return None
+
+
+class GeneralLoRAFromPeft:
+    def __init__(self):
+        self.supported_model_classes = [ WanModel]
+    
+    
+    def get_name_dict(self, lora_state_dict):
+        lora_name_dict = {}
+        for key in lora_state_dict:
+            if ".lora_B." not in key:
+                continue
+            keys = key.split(".")
+            if len(keys) > keys.index("lora_B") + 2:
+                keys.pop(keys.index("lora_B") + 1)
+            keys.pop(keys.index("lora_B"))
+            if keys[0] == "diffusion_model":
+                keys.pop(0)
+            target_name = ".".join(keys)
+            lora_name_dict[target_name] = (key, key.replace(".lora_B.", ".lora_A."))
+        return lora_name_dict
+    
+    
+    def match(self, model: torch.nn.Module, state_dict_lora):
+        lora_name_dict = self.get_name_dict(state_dict_lora)
+        model_name_dict = {name: None for name, _ in model.named_parameters()}
+        matched_num = sum([i in model_name_dict for i in lora_name_dict])
+        if matched_num == len(lora_name_dict):
+            return "", ""
+        else:
+            return None
+    
+    
+    def fetch_device_and_dtype(self, state_dict):
+        device, dtype = None, None
+        for name, param in state_dict.items():
+            device, dtype = param.device, param.dtype
+            break
+        computation_device = device
+        computation_dtype = dtype
+        if computation_device == torch.device("cpu"):
+            if torch.cuda.is_available():
+                computation_device = torch.device("cuda")
+        if computation_dtype == torch.float8_e4m3fn:
+            computation_dtype = torch.float32
+        return device, dtype, computation_device, computation_dtype
+
+
+    def load(self, model, state_dict_lora, lora_prefix="", alpha=1.0, model_resource=""):
+        state_dict_model = model.state_dict()
+        device, dtype, computation_device, computation_dtype = self.fetch_device_and_dtype(state_dict_model)
+        lora_name_dict = self.get_name_dict(state_dict_lora)
+        for name in lora_name_dict:
+            weight_up = state_dict_lora[lora_name_dict[name][0]].to(device=computation_device, dtype=computation_dtype)
+            weight_down = state_dict_lora[lora_name_dict[name][1]].to(device=computation_device, dtype=computation_dtype)
+            if len(weight_up.shape) == 4:
+                weight_up = weight_up.squeeze(3).squeeze(2)
+                weight_down = weight_down.squeeze(3).squeeze(2)
+                weight_lora = alpha * torch.mm(weight_up, weight_down).unsqueeze(2).unsqueeze(3)
+            else:
+                weight_lora = alpha * torch.mm(weight_up, weight_down)
+            weight_model = state_dict_model[name].to(device=computation_device, dtype=computation_dtype)
+            weight_patched = weight_model + weight_lora
+            state_dict_model[name] = weight_patched.to(device=device, dtype=dtype)
+        print(f"    {len(lora_name_dict)} tensors are updated.")
+        model.load_state_dict(state_dict_model)
+    
+    
+
+class WanLoRAConverter:
+    def __init__(self):
+        pass
+
+    @staticmethod
+    def align_to_opensource_format(state_dict, **kwargs):
+        state_dict = {"diffusion_model." + name.replace(".default.", "."): param for name, param in state_dict.items()}
+        return state_dict
+    
+    @staticmethod
+    def align_to_dkt_format(state_dict, **kwargs):
+        state_dict = {name.replace("diffusion_model.", "").replace(".lora_A.weight", ".lora_A.default.weight").replace(".lora_B.weight", ".lora_B.default.weight"): param for name, param in state_dict.items()}
+        return state_dict
+
+def get_lora_loaders():
+    return [GeneralLoRAFromPeft()]

dkt/models/model_manager.py

@@ -0,0 +1,421 @@
+import os, torch, json, importlib
+from typing import List
+
+from .downloader import download_models
+from .lora import get_lora_loaders
+
+from ..configs.model_config import model_loader_configs
+from .utils import load_state_dict, init_weights_on_device, hash_state_dict_keys, split_state_dict_with_prefix
+
+from .downloader import Preset_model_id, Preset_model_website
+
+def load_model_from_single_file(state_dict, model_names, model_classes, model_resource, torch_dtype, device):
+    loaded_model_names, loaded_models = [], []
+    for model_name, model_class in zip(model_names, model_classes):
+        print(f"    model_name: {model_name} model_class: {model_class.__name__}")
+        state_dict_converter = model_class.state_dict_converter()
+        if model_resource == "civitai":
+            state_dict_results = state_dict_converter.from_civitai(state_dict)
+        elif model_resource == "diffusers":
+            state_dict_results = state_dict_converter.from_diffusers(state_dict)
+        if isinstance(state_dict_results, tuple):
+            model_state_dict, extra_kwargs = state_dict_results
+            print(f"        This model is initialized with extra kwargs: {extra_kwargs}")
+        else:
+            model_state_dict, extra_kwargs = state_dict_results, {}
+        torch_dtype = torch.float32 if extra_kwargs.get("upcast_to_float32", False) else torch_dtype
+        with init_weights_on_device():
+            model = model_class(**extra_kwargs)
+        if hasattr(model, "eval"):
+            model = model.eval()
+        model.load_state_dict(model_state_dict, assign=True)
+        model = model.to(dtype=torch_dtype, device=device)
+        loaded_model_names.append(model_name)
+        loaded_models.append(model)
+    return loaded_model_names, loaded_models
+
+
+def load_model_from_huggingface_folder(file_path, model_names, model_classes, torch_dtype, device):
+    loaded_model_names, loaded_models = [], []
+    for model_name, model_class in zip(model_names, model_classes):
+        if torch_dtype in [torch.float32, torch.float16, torch.bfloat16]:
+            model = model_class.from_pretrained(file_path, torch_dtype=torch_dtype).eval()
+        else:
+            model = model_class.from_pretrained(file_path).eval().to(dtype=torch_dtype)
+        if torch_dtype == torch.float16 and hasattr(model, "half"):
+            model = model.half()
+        try:
+            model = model.to(device=device)
+        except:
+            pass
+        loaded_model_names.append(model_name)
+        loaded_models.append(model)
+    return loaded_model_names, loaded_models
+
+
+def load_single_patch_model_from_single_file(state_dict, model_name, model_class, base_model, extra_kwargs, torch_dtype, device):
+    print(f"    model_name: {model_name} model_class: {model_class.__name__} extra_kwargs: {extra_kwargs}")
+    base_state_dict = base_model.state_dict()
+    base_model.to("cpu")
+    del base_model
+    model = model_class(**extra_kwargs)
+    model.load_state_dict(base_state_dict, strict=False)
+    model.load_state_dict(state_dict, strict=False)
+    model.to(dtype=torch_dtype, device=device)
+    return model
+
+
+def load_patch_model_from_single_file(state_dict, model_names, model_classes, extra_kwargs, model_manager, torch_dtype, device):
+    loaded_model_names, loaded_models = [], []
+    for model_name, model_class in zip(model_names, model_classes):
+        while True:
+            for model_id in range(len(model_manager.model)):
+                base_model_name = model_manager.model_name[model_id]
+                if base_model_name == model_name:
+                    base_model_path = model_manager.model_path[model_id]
+                    base_model = model_manager.model[model_id]
+                    print(f"    Adding patch model to {base_model_name} ({base_model_path})")
+                    patched_model = load_single_patch_model_from_single_file(
+                        state_dict, model_name, model_class, base_model, extra_kwargs, torch_dtype, device)
+                    loaded_model_names.append(base_model_name)
+                    loaded_models.append(patched_model)
+                    model_manager.model.pop(model_id)
+                    model_manager.model_path.pop(model_id)
+                    model_manager.model_name.pop(model_id)
+                    break
+            else:
+                break
+    return loaded_model_names, loaded_models
+
+
+
+class ModelDetectorTemplate:
+    def __init__(self):
+        pass
+
+    def match(self, file_path="", state_dict={}):
+        return False
+    
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        return [], []
+    
+
+
+class ModelDetectorFromSingleFile:
+    def __init__(self, model_loader_configs=[]):
+        self.keys_hash_with_shape_dict = {}
+        self.keys_hash_dict = {}
+        for metadata in model_loader_configs:
+            self.add_model_metadata(*metadata)
+
+
+    def add_model_metadata(self, keys_hash, keys_hash_with_shape, model_names, model_classes, model_resource):
+        self.keys_hash_with_shape_dict[keys_hash_with_shape] = (model_names, model_classes, model_resource)
+        if keys_hash is not None:
+            self.keys_hash_dict[keys_hash] = (model_names, model_classes, model_resource)
+
+
+    def match(self, file_path="", state_dict={}):
+        if isinstance(file_path, str) and os.path.isdir(file_path):
+            return False
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            return True
+        keys_hash = hash_state_dict_keys(state_dict, with_shape=False)
+        if keys_hash in self.keys_hash_dict:
+            return True
+        return False
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+
+        # Load models with strict matching
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            model_names, model_classes, model_resource = self.keys_hash_with_shape_dict[keys_hash_with_shape]
+            loaded_model_names, loaded_models = load_model_from_single_file(state_dict, model_names, model_classes, model_resource, torch_dtype, device)
+            return loaded_model_names, loaded_models
+
+        # Load models without strict matching
+        # (the shape of parameters may be inconsistent, and the state_dict_converter will modify the model architecture)
+        keys_hash = hash_state_dict_keys(state_dict, with_shape=False)
+        if keys_hash in self.keys_hash_dict:
+            model_names, model_classes, model_resource = self.keys_hash_dict[keys_hash]
+            loaded_model_names, loaded_models = load_model_from_single_file(state_dict, model_names, model_classes, model_resource, torch_dtype, device)
+            return loaded_model_names, loaded_models
+
+        return loaded_model_names, loaded_models
+
+
+
+class ModelDetectorFromSplitedSingleFile(ModelDetectorFromSingleFile):
+    def __init__(self, model_loader_configs=[]):
+        super().__init__(model_loader_configs)
+
+
+    def match(self, file_path="", state_dict={}):
+        if isinstance(file_path, str) and os.path.isdir(file_path):
+            return False
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        splited_state_dict = split_state_dict_with_prefix(state_dict)
+        for sub_state_dict in splited_state_dict:
+            if super().match(file_path, sub_state_dict):
+                return True
+        return False
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        # Split the state_dict and load from each component
+        splited_state_dict = split_state_dict_with_prefix(state_dict)
+        valid_state_dict = {}
+        for sub_state_dict in splited_state_dict:
+            if super().match(file_path, sub_state_dict):
+                valid_state_dict.update(sub_state_dict)
+        if super().match(file_path, valid_state_dict):
+            loaded_model_names, loaded_models = super().load(file_path, valid_state_dict, device, torch_dtype)
+        else:
+            loaded_model_names, loaded_models = [], []
+            for sub_state_dict in splited_state_dict:
+                if super().match(file_path, sub_state_dict):
+                    loaded_model_names_, loaded_models_ = super().load(file_path, valid_state_dict, device, torch_dtype)
+                    loaded_model_names += loaded_model_names_
+                    loaded_models += loaded_models_
+        return loaded_model_names, loaded_models
+    
+
+
+class ModelDetectorFromHuggingfaceFolder:
+    def __init__(self, model_loader_configs=[]):
+        self.architecture_dict = {}
+        for metadata in model_loader_configs:
+            self.add_model_metadata(*metadata)
+
+
+    def add_model_metadata(self, architecture, huggingface_lib, model_name, redirected_architecture):
+        self.architecture_dict[architecture] = (huggingface_lib, model_name, redirected_architecture)
+
+
+    def match(self, file_path="", state_dict={}):
+        if not isinstance(file_path, str) or os.path.isfile(file_path):
+            return False
+        file_list = os.listdir(file_path)
+        if "config.json" not in file_list:
+            return False
+        with open(os.path.join(file_path, "config.json"), "r") as f:
+            config = json.load(f)
+        if "architectures" not in config and "_class_name" not in config:
+            return False
+        return True
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, **kwargs):
+        with open(os.path.join(file_path, "config.json"), "r") as f:
+            config = json.load(f)
+        loaded_model_names, loaded_models = [], []
+        architectures = config["architectures"] if "architectures" in config else [config["_class_name"]]
+        for architecture in architectures:
+            huggingface_lib, model_name, redirected_architecture = self.architecture_dict[architecture]
+            if redirected_architecture is not None:
+                architecture = redirected_architecture
+            model_class = importlib.import_module(huggingface_lib).__getattribute__(architecture)
+            loaded_model_names_, loaded_models_ = load_model_from_huggingface_folder(file_path, [model_name], [model_class], torch_dtype, device)
+            loaded_model_names += loaded_model_names_
+            loaded_models += loaded_models_
+        return loaded_model_names, loaded_models
+    
+
+
+class ModelDetectorFromPatchedSingleFile:
+    def __init__(self, model_loader_configs=[]):
+        self.keys_hash_with_shape_dict = {}
+        for metadata in model_loader_configs:
+            self.add_model_metadata(*metadata)
+
+
+    def add_model_metadata(self, keys_hash_with_shape, model_name, model_class, extra_kwargs):
+        self.keys_hash_with_shape_dict[keys_hash_with_shape] = (model_name, model_class, extra_kwargs)
+
+
+    def match(self, file_path="", state_dict={}):
+        if not isinstance(file_path, str) or os.path.isdir(file_path):
+            return False
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            return True
+        return False
+
+
+    def load(self, file_path="", state_dict={}, device="cuda", torch_dtype=torch.float16, model_manager=None, **kwargs):
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+
+        # Load models with strict matching
+        loaded_model_names, loaded_models = [], []
+        keys_hash_with_shape = hash_state_dict_keys(state_dict, with_shape=True)
+        if keys_hash_with_shape in self.keys_hash_with_shape_dict:
+            model_names, model_classes, extra_kwargs = self.keys_hash_with_shape_dict[keys_hash_with_shape]
+            loaded_model_names_, loaded_models_ = load_patch_model_from_single_file(
+                state_dict, model_names, model_classes, extra_kwargs, model_manager, torch_dtype, device)
+            loaded_model_names += loaded_model_names_
+            loaded_models += loaded_models_
+        return loaded_model_names, loaded_models
+
+
+
+class ModelManager:
+    def __init__(
+        self,
+        torch_dtype=torch.float16,
+        device="cuda",
+        model_id_list: List[Preset_model_id] = [],
+        downloading_priority: List[Preset_model_website] = ["ModelScope", "HuggingFace"],
+        file_path_list: List[str] = [],
+    ):
+        self.torch_dtype = torch_dtype
+        self.device = device
+        self.model = []
+        self.model_path = []
+        self.model_name = []
+        downloaded_files = download_models(model_id_list, downloading_priority) if len(model_id_list) > 0 else []
+        self.model_detector = [
+            ModelDetectorFromSingleFile(model_loader_configs),
+            ModelDetectorFromSplitedSingleFile(model_loader_configs),
+        ]
+        self.load_models(downloaded_files + file_path_list)
+
+
+    def load_model_from_single_file(self, file_path="", state_dict={}, model_names=[], model_classes=[], model_resource=None):
+        print(f"Loading models from file: {file_path}")
+        if len(state_dict) == 0:
+            state_dict = load_state_dict(file_path)
+        model_names, models = load_model_from_single_file(state_dict, model_names, model_classes, model_resource, self.torch_dtype, self.device)
+        for model_name, model in zip(model_names, models):
+            self.model.append(model)
+            self.model_path.append(file_path)
+            self.model_name.append(model_name)
+        print(f"    The following models are loaded: {model_names}.")
+
+
+    def load_model_from_huggingface_folder(self, file_path="", model_names=[], model_classes=[]):
+        print(f"Loading models from folder: {file_path}")
+        model_names, models = load_model_from_huggingface_folder(file_path, model_names, model_classes, self.torch_dtype, self.device)
+        for model_name, model in zip(model_names, models):
+            self.model.append(model)
+            self.model_path.append(file_path)
+            self.model_name.append(model_name)
+        print(f"    The following models are loaded: {model_names}.")
+
+
+    def load_patch_model_from_single_file(self, file_path="", state_dict={}, model_names=[], model_classes=[], extra_kwargs={}):
+        print(f"Loading patch models from file: {file_path}")
+        model_names, models = load_patch_model_from_single_file(
+            state_dict, model_names, model_classes, extra_kwargs, self, self.torch_dtype, self.device)
+        for model_name, model in zip(model_names, models):
+            self.model.append(model)
+            self.model_path.append(file_path)
+            self.model_name.append(model_name)
+        print(f"    The following patched models are loaded: {model_names}.")
+
+
+    def load_lora(self, file_path="", state_dict={}, lora_alpha=1.0):
+        if isinstance(file_path, list):
+            for file_path_ in file_path:
+                self.load_lora(file_path_, state_dict=state_dict, lora_alpha=lora_alpha)
+        else:
+            print(f"Loading LoRA models from file: {file_path}")
+            is_loaded = False
+            if len(state_dict) == 0:
+                state_dict = load_state_dict(file_path)
+            for model_name, model, model_path in zip(self.model_name, self.model, self.model_path):
+                for lora in get_lora_loaders():
+                    match_results = lora.match(model, state_dict)
+                    if match_results is not None:
+                        print(f"    Adding LoRA to {model_name} ({model_path}).")
+                        lora_prefix, model_resource = match_results
+                        lora.load(model, state_dict, lora_prefix, alpha=lora_alpha, model_resource=model_resource)
+                        is_loaded = True
+                        break
+            if not is_loaded:
+                print(f"    Cannot load LoRA: {file_path}")
+
+
+    def load_model(self, file_path, model_names=None, device=None, torch_dtype=None):
+        print(f"Loading models from: {file_path}")
+        if device is None: device = self.device
+        if torch_dtype is None: torch_dtype = self.torch_dtype
+        if isinstance(file_path, list):
+            state_dict = {}
+            for path in file_path:
+                state_dict.update(load_state_dict(path))
+        elif os.path.isfile(file_path):
+            state_dict = load_state_dict(file_path)
+        else:
+            state_dict = None
+        for model_detector in self.model_detector:
+            if model_detector.match(file_path, state_dict):
+                model_names, models = model_detector.load(
+                    file_path, state_dict,
+                    device=device, torch_dtype=torch_dtype,
+                    allowed_model_names=model_names, model_manager=self
+                )
+                for model_name, model in zip(model_names, models):
+                    self.model.append(model)
+                    self.model_path.append(file_path)
+                    self.model_name.append(model_name)
+                print(f"    The following models are loaded: {model_names}.")
+                break
+        else:
+            print(f"    We cannot detect the model type. No models are loaded.")
+        
+
+    def load_models(self, file_path_list, model_names=None, device=None, torch_dtype=None):
+        for file_path in file_path_list:
+            self.load_model(file_path, model_names, device=device, torch_dtype=torch_dtype)
+
+    
+    def fetch_model(self, model_name, file_path=None, require_model_path=False, index=None):
+        fetched_models = []
+        fetched_model_paths = []
+        for model, model_path, model_name_ in zip(self.model, self.model_path, self.model_name):
+            if file_path is not None and file_path != model_path:
+                continue
+            if model_name == model_name_:
+                fetched_models.append(model)
+                fetched_model_paths.append(model_path)
+        if len(fetched_models) == 0:
+            print(f"No {model_name} models available.")
+            return None
+        if len(fetched_models) == 1:
+            print(f"Using {model_name} from {fetched_model_paths[0]}.")
+            model = fetched_models[0]
+            path = fetched_model_paths[0]
+        else:
+            if index is None:
+                model = fetched_models[0]
+                path = fetched_model_paths[0]
+                print(f"More than one {model_name} models are loaded in model manager: {fetched_model_paths}. Using {model_name} from {fetched_model_paths[0]}.")
+            elif isinstance(index, int):
+                model = fetched_models[:index]
+                path = fetched_model_paths[:index]
+                print(f"More than one {model_name} models are loaded in model manager: {fetched_model_paths}. Using {model_name} from {fetched_model_paths[:index]}.")
+            else:
+                model = fetched_models
+                path = fetched_model_paths
+                print(f"More than one {model_name} models are loaded in model manager: {fetched_model_paths}. Using {model_name} from {fetched_model_paths}.")
+        if require_model_path:
+            return model, path
+        else:
+            return model
+        
+
+    def to(self, device):
+        for model in self.model:
+            model.to(device)
+

dkt/models/tiler.py

@@ -0,0 +1,234 @@
+import torch
+from einops import rearrange, repeat
+
+
+class TileWorker:
+    def __init__(self):
+        pass
+
+
+    def mask(self, height, width, border_width):
+        # Create a mask with shape (height, width).
+        # The centre area is filled with 1, and the border line is filled with values in range (0, 1].
+        x = torch.arange(height).repeat(width, 1).T
+        y = torch.arange(width).repeat(height, 1)
+        mask = torch.stack([x + 1, height - x, y + 1, width - y]).min(dim=0).values
+        mask = (mask / border_width).clip(0, 1)
+        return mask
+
+
+    def tile(self, model_input, tile_size, tile_stride, tile_device, tile_dtype):
+        # Convert a tensor (b, c, h, w) to (b, c, tile_size, tile_size, tile_num)
+        batch_size, channel, _, _ = model_input.shape
+        model_input = model_input.to(device=tile_device, dtype=tile_dtype)
+        unfold_operator = torch.nn.Unfold(
+            kernel_size=(tile_size, tile_size),
+            stride=(tile_stride, tile_stride)
+        )
+        model_input = unfold_operator(model_input)
+        model_input = model_input.view((batch_size, channel, tile_size, tile_size, -1))
+
+        return model_input
+
+
+    def tiled_inference(self, forward_fn, model_input, tile_batch_size, inference_device, inference_dtype, tile_device, tile_dtype):
+        # Call y=forward_fn(x) for each tile
+        tile_num = model_input.shape[-1]
+        model_output_stack = []
+
+        for tile_id in range(0, tile_num, tile_batch_size):
+
+            # process input
+            tile_id_ = min(tile_id + tile_batch_size, tile_num)
+            x = model_input[:, :, :, :, tile_id: tile_id_]
+            x = x.to(device=inference_device, dtype=inference_dtype)
+            x = rearrange(x, "b c h w n -> (n b) c h w")
+
+            # process output
+            y = forward_fn(x)
+            y = rearrange(y, "(n b) c h w -> b c h w n", n=tile_id_-tile_id)
+            y = y.to(device=tile_device, dtype=tile_dtype)
+            model_output_stack.append(y)
+
+        model_output = torch.concat(model_output_stack, dim=-1)
+        return model_output
+
+
+    def io_scale(self, model_output, tile_size):
+        # Determine the size modification happened in forward_fn
+        # We only consider the same scale on height and width.
+        io_scale = model_output.shape[2] / tile_size
+        return io_scale
+    
+
+    def untile(self, model_output, height, width, tile_size, tile_stride, border_width, tile_device, tile_dtype):
+        # The reversed function of tile
+        mask = self.mask(tile_size, tile_size, border_width)
+        mask = mask.to(device=tile_device, dtype=tile_dtype)
+        mask = rearrange(mask, "h w -> 1 1 h w 1")
+        model_output = model_output * mask
+
+        fold_operator = torch.nn.Fold(
+            output_size=(height, width),
+            kernel_size=(tile_size, tile_size),
+            stride=(tile_stride, tile_stride)
+        )
+        mask = repeat(mask[0, 0, :, :, 0], "h w -> 1 (h w) n", n=model_output.shape[-1])
+        model_output = rearrange(model_output, "b c h w n -> b (c h w) n")
+        model_output = fold_operator(model_output) / fold_operator(mask)
+
+        return model_output
+
+
+    def tiled_forward(self, forward_fn, model_input, tile_size, tile_stride, tile_batch_size=1, tile_device="cpu", tile_dtype=torch.float32, border_width=None):
+        # Prepare
+        inference_device, inference_dtype = model_input.device, model_input.dtype
+        height, width = model_input.shape[2], model_input.shape[3]
+        border_width = int(tile_stride*0.5) if border_width is None else border_width
+
+        # tile
+        model_input = self.tile(model_input, tile_size, tile_stride, tile_device, tile_dtype)
+
+        # inference
+        model_output = self.tiled_inference(forward_fn, model_input, tile_batch_size, inference_device, inference_dtype, tile_device, tile_dtype)
+
+        # resize
+        io_scale = self.io_scale(model_output, tile_size)
+        height, width = int(height*io_scale), int(width*io_scale)
+        tile_size, tile_stride = int(tile_size*io_scale), int(tile_stride*io_scale)
+        border_width = int(border_width*io_scale)
+
+        # untile
+        model_output = self.untile(model_output, height, width, tile_size, tile_stride, border_width, tile_device, tile_dtype)
+        
+        # Done!
+        model_output = model_output.to(device=inference_device, dtype=inference_dtype)
+        return model_output
+    
+
+
+class FastTileWorker:
+    def __init__(self):
+        pass
+
+
+    def build_mask(self, data, is_bound):
+        _, _, H, W = data.shape
+        h = repeat(torch.arange(H), "H -> H W", H=H, W=W)
+        w = repeat(torch.arange(W), "W -> H W", H=H, W=W)
+        border_width = (H + W) // 4
+        pad = torch.ones_like(h) * border_width
+        mask = torch.stack([
+            pad if is_bound[0] else h + 1,
+            pad if is_bound[1] else H - h,
+            pad if is_bound[2] else w + 1,
+            pad if is_bound[3] else W - w
+        ]).min(dim=0).values
+        mask = mask.clip(1, border_width)
+        mask = (mask / border_width).to(dtype=data.dtype, device=data.device)
+        mask = rearrange(mask, "H W -> 1 H W")
+        return mask
+
+
+    def tiled_forward(self, forward_fn, model_input, tile_size, tile_stride, tile_device="cpu", tile_dtype=torch.float32, border_width=None):
+        # Prepare
+        B, C, H, W = model_input.shape
+        border_width = int(tile_stride*0.5) if border_width is None else border_width
+        weight = torch.zeros((1, 1, H, W), dtype=tile_dtype, device=tile_device)
+        values = torch.zeros((B, C, H, W), dtype=tile_dtype, device=tile_device)
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, tile_stride):
+            for w in range(0, W, tile_stride):
+                if (h-tile_stride >= 0 and h-tile_stride+tile_size >= H) or (w-tile_stride >= 0 and w-tile_stride+tile_size >= W):
+                    continue
+                h_, w_ = h + tile_size, w + tile_size
+                if h_ > H: h, h_ = H - tile_size, H
+                if w_ > W: w, w_ = W - tile_size, W
+                tasks.append((h, h_, w, w_))
+        
+        # Run
+        for hl, hr, wl, wr in tasks:
+            # Forward
+            hidden_states_batch = forward_fn(hl, hr, wl, wr).to(dtype=tile_dtype, device=tile_device)
+
+            mask = self.build_mask(hidden_states_batch, is_bound=(hl==0, hr>=H, wl==0, wr>=W))
+            values[:, :, hl:hr, wl:wr] += hidden_states_batch * mask
+            weight[:, :, hl:hr, wl:wr] += mask
+        values /= weight
+        return values
+
+
+
+class TileWorker2Dto3D:
+    """
+    Process 3D tensors, but only enable TileWorker on 2D.
+    """
+    def __init__(self):
+        pass
+
+
+    def build_mask(self, T, H, W, dtype, device, is_bound, border_width):
+        t = repeat(torch.arange(T), "T -> T H W", T=T, H=H, W=W)
+        h = repeat(torch.arange(H), "H -> T H W", T=T, H=H, W=W)
+        w = repeat(torch.arange(W), "W -> T H W", T=T, H=H, W=W)
+        border_width = (H + W) // 4 if border_width is None else border_width
+        pad = torch.ones_like(h) * border_width
+        mask = torch.stack([
+            pad if is_bound[0] else t + 1,
+            pad if is_bound[1] else T - t,
+            pad if is_bound[2] else h + 1,
+            pad if is_bound[3] else H - h,
+            pad if is_bound[4] else w + 1,
+            pad if is_bound[5] else W - w
+        ]).min(dim=0).values
+        mask = mask.clip(1, border_width)
+        mask = (mask / border_width).to(dtype=dtype, device=device)
+        mask = rearrange(mask, "T H W -> 1 1 T H W")
+        return mask
+
+
+    def tiled_forward(
+        self,
+        forward_fn,
+        model_input,
+        tile_size, tile_stride,
+        tile_device="cpu", tile_dtype=torch.float32,
+        computation_device="cuda", computation_dtype=torch.float32,
+        border_width=None, scales=[1, 1, 1, 1],
+        progress_bar=lambda x:x
+    ):
+        B, C, T, H, W = model_input.shape
+        scale_C, scale_T, scale_H, scale_W = scales
+        tile_size_H, tile_size_W = tile_size
+        tile_stride_H, tile_stride_W = tile_stride
+
+        value = torch.zeros((B, int(C*scale_C), int(T*scale_T), int(H*scale_H), int(W*scale_W)), dtype=tile_dtype, device=tile_device)
+        weight = torch.zeros((1, 1, int(T*scale_T), int(H*scale_H), int(W*scale_W)), dtype=tile_dtype, device=tile_device)
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, tile_stride_H):
+            for w in range(0, W, tile_stride_W):
+                if (h-tile_stride_H >= 0 and h-tile_stride_H+tile_size_H >= H) or (w-tile_stride_W >= 0 and w-tile_stride_W+tile_size_W >= W):
+                    continue
+                h_, w_ = h + tile_size_H, w + tile_size_W
+                if h_ > H: h, h_ = max(H - tile_size_H, 0), H
+                if w_ > W: w, w_ = max(W - tile_size_W, 0), W
+                tasks.append((h, h_, w, w_))
+
+        # Run
+        for hl, hr, wl, wr in progress_bar(tasks):
+            mask = self.build_mask(
+                int(T*scale_T), int((hr-hl)*scale_H), int((wr-wl)*scale_W),
+                tile_dtype, tile_device,
+                is_bound=(True, True, hl==0, hr>=H, wl==0, wr>=W),
+                border_width=border_width
+            )
+            grid_input = model_input[:, :, :, hl:hr, wl:wr].to(dtype=computation_dtype, device=computation_device)
+            grid_output = forward_fn(grid_input).to(dtype=tile_dtype, device=tile_device)
+            value[:, :, :, int(hl*scale_H):int(hr*scale_H), int(wl*scale_W):int(wr*scale_W)] += grid_output * mask
+            weight[:, :, :, int(hl*scale_H):int(hr*scale_H), int(wl*scale_W):int(wr*scale_W)] += mask
+        value = value / weight
+        return value

dkt/models/utils.py

@@ -0,0 +1,182 @@
+import torch, os
+from safetensors import safe_open
+from contextlib import contextmanager
+import hashlib
+
+@contextmanager
+def init_weights_on_device(device = torch.device("meta"), include_buffers :bool = False):
+    
+    old_register_parameter = torch.nn.Module.register_parameter
+    if include_buffers:
+        old_register_buffer = torch.nn.Module.register_buffer
+    
+    def register_empty_parameter(module, name, param):
+        old_register_parameter(module, name, param)
+        if param is not None:
+            param_cls = type(module._parameters[name])
+            kwargs = module._parameters[name].__dict__
+            kwargs["requires_grad"] = param.requires_grad
+            module._parameters[name] = param_cls(module._parameters[name].to(device), **kwargs)
+
+    def register_empty_buffer(module, name, buffer, persistent=True):
+        old_register_buffer(module, name, buffer, persistent=persistent)
+        if buffer is not None:
+            module._buffers[name] = module._buffers[name].to(device)
+            
+    def patch_tensor_constructor(fn):
+        def wrapper(*args, **kwargs):
+            kwargs["device"] = device
+            return fn(*args, **kwargs)
+
+        return wrapper
+    
+    if include_buffers:
+        tensor_constructors_to_patch = {
+            torch_function_name: getattr(torch, torch_function_name)
+            for torch_function_name in ["empty", "zeros", "ones", "full"]
+        }
+    else:
+        tensor_constructors_to_patch = {}
+    
+    try:
+        torch.nn.Module.register_parameter = register_empty_parameter
+        if include_buffers:
+            torch.nn.Module.register_buffer = register_empty_buffer
+        for torch_function_name in tensor_constructors_to_patch.keys():
+            setattr(torch, torch_function_name, patch_tensor_constructor(getattr(torch, torch_function_name)))
+        yield
+    finally:
+        torch.nn.Module.register_parameter = old_register_parameter
+        if include_buffers:
+            torch.nn.Module.register_buffer = old_register_buffer
+        for torch_function_name, old_torch_function in tensor_constructors_to_patch.items():
+            setattr(torch, torch_function_name, old_torch_function)
+
+def load_state_dict_from_folder(file_path, torch_dtype=None):
+    state_dict = {}
+    for file_name in os.listdir(file_path):
+        if "." in file_name and file_name.split(".")[-1] in [
+            "safetensors", "bin", "ckpt", "pth", "pt"
+        ]:
+            state_dict.update(load_state_dict(os.path.join(file_path, file_name), torch_dtype=torch_dtype))
+    return state_dict
+
+
+def load_state_dict(file_path, torch_dtype=None, device="cpu"):
+    if file_path.endswith(".safetensors"):
+        return load_state_dict_from_safetensors(file_path, torch_dtype=torch_dtype, device=device)
+    else:
+        return load_state_dict_from_bin(file_path, torch_dtype=torch_dtype, device=device)
+
+
+def load_state_dict_from_safetensors(file_path, torch_dtype=None, device="cpu"):
+    state_dict = {}
+    with safe_open(file_path, framework="pt", device=str(device)) as f:
+        for k in f.keys():
+            state_dict[k] = f.get_tensor(k)
+            if torch_dtype is not None:
+                state_dict[k] = state_dict[k].to(torch_dtype)
+    return state_dict
+
+
+def load_state_dict_from_bin(file_path, torch_dtype=None, device="cpu"):
+    state_dict = torch.load(file_path, map_location=device, weights_only=True)
+    if torch_dtype is not None:
+        for i in state_dict:
+            if isinstance(state_dict[i], torch.Tensor):
+                state_dict[i] = state_dict[i].to(torch_dtype)
+    return state_dict
+
+
+def search_for_embeddings(state_dict):
+    embeddings = []
+    for k in state_dict:
+        if isinstance(state_dict[k], torch.Tensor):
+            embeddings.append(state_dict[k])
+        elif isinstance(state_dict[k], dict):
+            embeddings += search_for_embeddings(state_dict[k])
+    return embeddings
+
+
+def search_parameter(param, state_dict):
+    for name, param_ in state_dict.items():
+        if param.numel() == param_.numel():
+            if param.shape == param_.shape:
+                if torch.dist(param, param_) < 1e-3:
+                    return name
+            else:
+                if torch.dist(param.flatten(), param_.flatten()) < 1e-3:
+                    return name
+    return None
+
+
+def build_rename_dict(source_state_dict, target_state_dict, split_qkv=False):
+    matched_keys = set()
+    with torch.no_grad():
+        for name in source_state_dict:
+            rename = search_parameter(source_state_dict[name], target_state_dict)
+            if rename is not None:
+                print(f'"{name}": "{rename}",')
+                matched_keys.add(rename)
+            elif split_qkv and len(source_state_dict[name].shape)>=1 and source_state_dict[name].shape[0]%3==0:
+                length = source_state_dict[name].shape[0] // 3
+                rename = []
+                for i in range(3):
+                    rename.append(search_parameter(source_state_dict[name][i*length: i*length+length], target_state_dict))
+                if None not in rename:
+                    print(f'"{name}": {rename},')
+                    for rename_ in rename:
+                        matched_keys.add(rename_)
+    for name in target_state_dict:
+        if name not in matched_keys:
+            print("Cannot find", name, target_state_dict[name].shape)
+
+
+def search_for_files(folder, extensions):
+    files = []
+    if os.path.isdir(folder):
+        for file in sorted(os.listdir(folder)):
+            files += search_for_files(os.path.join(folder, file), extensions)
+    elif os.path.isfile(folder):
+        for extension in extensions:
+            if folder.endswith(extension):
+                files.append(folder)
+                break
+    return files
+
+
+def convert_state_dict_keys_to_single_str(state_dict, with_shape=True):
+    keys = []
+    for key, value in state_dict.items():
+        if isinstance(key, str):
+            if isinstance(value, torch.Tensor):
+                if with_shape:
+                    shape = "_".join(map(str, list(value.shape)))
+                    keys.append(key + ":" + shape)
+                keys.append(key)
+            elif isinstance(value, dict):
+                keys.append(key + "|" + convert_state_dict_keys_to_single_str(value, with_shape=with_shape))
+    keys.sort()
+    keys_str = ",".join(keys)
+    return keys_str
+
+
+def split_state_dict_with_prefix(state_dict):
+    keys = sorted([key for key in state_dict if isinstance(key, str)])
+    prefix_dict = {}
+    for key in  keys:
+        prefix = key if "." not in key else key.split(".")[0]
+        if prefix not in prefix_dict:
+            prefix_dict[prefix] = []
+        prefix_dict[prefix].append(key)
+    state_dicts = []
+    for prefix, keys in prefix_dict.items():
+        sub_state_dict = {key: state_dict[key] for key in keys}
+        state_dicts.append(sub_state_dict)
+    return state_dicts
+
+
+def hash_state_dict_keys(state_dict, with_shape=True):
+    keys_str = convert_state_dict_keys_to_single_str(state_dict, with_shape=with_shape)
+    keys_str = keys_str.encode(encoding="UTF-8")
+    return hashlib.md5(keys_str).hexdigest()

dkt/models/wan_video_camera_controller.py

@@ -0,0 +1,202 @@
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+import os
+from typing_extensions import Literal
+
+class SimpleAdapter(nn.Module):
+    def __init__(self, in_dim, out_dim, kernel_size, stride, num_residual_blocks=1):
+        super(SimpleAdapter, self).__init__()
+
+        # Pixel Unshuffle: reduce spatial dimensions by a factor of 8
+        self.pixel_unshuffle = nn.PixelUnshuffle(downscale_factor=8)
+
+        # Convolution: reduce spatial dimensions by a factor
+        #  of 2 (without overlap)
+        self.conv = nn.Conv2d(in_dim * 64, out_dim, kernel_size=kernel_size, stride=stride, padding=0)
+
+        # Residual blocks for feature extraction
+        self.residual_blocks = nn.Sequential(
+            *[ResidualBlock(out_dim) for _ in range(num_residual_blocks)]
+        )
+
+    def forward(self, x):
+        # Reshape to merge the frame dimension into batch
+        bs, c, f, h, w = x.size()
+        x = x.permute(0, 2, 1, 3, 4).contiguous().view(bs * f, c, h, w)
+
+        # Pixel Unshuffle operation
+        x_unshuffled = self.pixel_unshuffle(x)
+
+        # Convolution operation
+        x_conv = self.conv(x_unshuffled)
+
+        # Feature extraction with residual blocks
+        out = self.residual_blocks(x_conv)
+
+        # Reshape to restore original bf dimension
+        out = out.view(bs, f, out.size(1), out.size(2), out.size(3))
+
+        # Permute dimensions to reorder (if needed), e.g., swap channels and feature frames
+        out = out.permute(0, 2, 1, 3, 4)
+
+        return out
+    
+    def process_camera_coordinates(
+        self,
+        direction: Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown"],
+        length: int,
+        height: int,
+        width: int,
+        speed: float = 1/54,
+        origin=(0, 0.532139961, 0.946026558, 0.5, 0.5, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
+    ):
+        if origin is None:
+            origin = (0, 0.532139961, 0.946026558, 0.5, 0.5, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
+        coordinates = generate_camera_coordinates(direction, length, speed, origin)
+        plucker_embedding = process_pose_file(coordinates, width, height)
+        return plucker_embedding
+        
+    
+
+class ResidualBlock(nn.Module):
+    def __init__(self, dim):
+        super(ResidualBlock, self).__init__()
+        self.conv1 = nn.Conv2d(dim, dim, kernel_size=3, padding=1)
+        self.relu = nn.ReLU(inplace=True)
+        self.conv2 = nn.Conv2d(dim, dim, kernel_size=3, padding=1)
+
+    def forward(self, x):
+        residual = x
+        out = self.relu(self.conv1(x))
+        out = self.conv2(out)
+        out += residual
+        return out
+    
+class Camera(object):
+    """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py
+    """
+    def __init__(self, entry):
+        fx, fy, cx, cy = entry[1:5]
+        self.fx = fx
+        self.fy = fy
+        self.cx = cx
+        self.cy = cy
+        w2c_mat = np.array(entry[7:]).reshape(3, 4)
+        w2c_mat_4x4 = np.eye(4)
+        w2c_mat_4x4[:3, :] = w2c_mat
+        self.w2c_mat = w2c_mat_4x4
+        self.c2w_mat = np.linalg.inv(w2c_mat_4x4)
+
+def get_relative_pose(cam_params):
+    """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py
+    """
+    abs_w2cs = [cam_param.w2c_mat for cam_param in cam_params]
+    abs_c2ws = [cam_param.c2w_mat for cam_param in cam_params]
+    cam_to_origin = 0
+    target_cam_c2w = np.array([
+        [1, 0, 0, 0],
+        [0, 1, 0, -cam_to_origin],
+        [0, 0, 1, 0],
+        [0, 0, 0, 1]
+    ])
+    abs2rel = target_cam_c2w @ abs_w2cs[0]
+    ret_poses = [target_cam_c2w, ] + [abs2rel @ abs_c2w for abs_c2w in abs_c2ws[1:]]
+    ret_poses = np.array(ret_poses, dtype=np.float32)
+    return ret_poses
+
+def custom_meshgrid(*args):
+    # torch>=2.0.0 only
+    return torch.meshgrid(*args, indexing='ij')
+
+
+def ray_condition(K, c2w, H, W, device):
+    """Copied from https://github.com/hehao13/CameraCtrl/blob/main/inference.py
+    """
+    # c2w: B, V, 4, 4
+    # K: B, V, 4
+
+    B = K.shape[0]
+
+    j, i = custom_meshgrid(
+        torch.linspace(0, H - 1, H, device=device, dtype=c2w.dtype),
+        torch.linspace(0, W - 1, W, device=device, dtype=c2w.dtype),
+    )
+    i = i.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5  # [B, HxW]
+    j = j.reshape([1, 1, H * W]).expand([B, 1, H * W]) + 0.5  # [B, HxW]
+
+    fx, fy, cx, cy = K.chunk(4, dim=-1)  # B,V, 1
+
+    zs = torch.ones_like(i)  # [B, HxW]
+    xs = (i - cx) / fx * zs
+    ys = (j - cy) / fy * zs
+    zs = zs.expand_as(ys)
+
+    directions = torch.stack((xs, ys, zs), dim=-1)  # B, V, HW, 3
+    directions = directions / directions.norm(dim=-1, keepdim=True)  # B, V, HW, 3
+
+    rays_d = directions @ c2w[..., :3, :3].transpose(-1, -2)  # B, V, 3, HW
+    rays_o = c2w[..., :3, 3]  # B, V, 3
+    rays_o = rays_o[:, :, None].expand_as(rays_d)  # B, V, 3, HW
+    # c2w @ dirctions
+    rays_dxo = torch.linalg.cross(rays_o, rays_d)
+    plucker = torch.cat([rays_dxo, rays_d], dim=-1)
+    plucker = plucker.reshape(B, c2w.shape[1], H, W, 6)  # B, V, H, W, 6
+    # plucker = plucker.permute(0, 1, 4, 2, 3)
+    return plucker
+
+
+def process_pose_file(cam_params, width=672, height=384, original_pose_width=1280, original_pose_height=720, device='cpu', return_poses=False):
+    if return_poses:
+        return cam_params
+    else:
+        cam_params = [Camera(cam_param) for cam_param in cam_params]
+
+        sample_wh_ratio = width / height
+        pose_wh_ratio = original_pose_width / original_pose_height  # Assuming placeholder ratios, change as needed
+
+        if pose_wh_ratio > sample_wh_ratio:
+            resized_ori_w = height * pose_wh_ratio
+            for cam_param in cam_params:
+                cam_param.fx = resized_ori_w * cam_param.fx / width
+        else:
+            resized_ori_h = width / pose_wh_ratio
+            for cam_param in cam_params:
+                cam_param.fy = resized_ori_h * cam_param.fy / height
+
+        intrinsic = np.asarray([[cam_param.fx * width,
+                                cam_param.fy * height,
+                                cam_param.cx * width,
+                                cam_param.cy * height]
+                                for cam_param in cam_params], dtype=np.float32)
+
+        K = torch.as_tensor(intrinsic)[None]  # [1, 1, 4]
+        c2ws = get_relative_pose(cam_params)  # Assuming this function is defined elsewhere
+        c2ws = torch.as_tensor(c2ws)[None]  # [1, n_frame, 4, 4]
+        plucker_embedding = ray_condition(K, c2ws, height, width, device=device)[0].permute(0, 3, 1, 2).contiguous()  # V, 6, H, W
+        plucker_embedding = plucker_embedding[None]
+        plucker_embedding = rearrange(plucker_embedding, "b f c h w -> b f h w c")[0]
+        return plucker_embedding
+
+
+
+def generate_camera_coordinates(
+    direction: Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown"],
+    length: int,
+    speed: float = 1/54,
+    origin=(0, 0.532139961, 0.946026558, 0.5, 0.5, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0)
+):
+    coordinates = [list(origin)]
+    while len(coordinates) < length:
+        coor = coordinates[-1].copy()
+        if "Left" in direction:
+            coor[9] += speed
+        if "Right" in direction:
+            coor[9] -= speed
+        if "Up" in direction:
+            coor[13] += speed
+        if "Down" in direction:
+            coor[13] -= speed
+        coordinates.append(coor)
+    return coordinates

dkt/models/wan_video_dit.py

@@ -0,0 +1,719 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import math
+from typing import Tuple, Optional
+from einops import rearrange
+from .utils import hash_state_dict_keys
+from .wan_video_camera_controller import SimpleAdapter
+try:
+    import flash_attn_interface
+    FLASH_ATTN_3_AVAILABLE = True
+except ModuleNotFoundError:
+    FLASH_ATTN_3_AVAILABLE = False
+
+try:
+    import flash_attn
+    FLASH_ATTN_2_AVAILABLE = True
+except ModuleNotFoundError:
+    FLASH_ATTN_2_AVAILABLE = False
+
+try:
+    from sageattention import sageattn
+    SAGE_ATTN_AVAILABLE = True
+except ModuleNotFoundError:
+    SAGE_ATTN_AVAILABLE = False
+    
+    
+def flash_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor, num_heads: int, compatibility_mode=False):
+    if compatibility_mode:
+        q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
+        x = F.scaled_dot_product_attention(q, k, v)
+        x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
+    elif FLASH_ATTN_3_AVAILABLE:
+        q = rearrange(q, "b s (n d) -> b s n d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b s n d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b s n d", n=num_heads)
+        x = flash_attn_interface.flash_attn_func(q, k, v)
+        if isinstance(x,tuple):
+            x = x[0]
+        x = rearrange(x, "b s n d -> b s (n d)", n=num_heads)
+    elif FLASH_ATTN_2_AVAILABLE:
+        q = rearrange(q, "b s (n d) -> b s n d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b s n d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b s n d", n=num_heads)
+        x = flash_attn.flash_attn_func(q, k, v)
+        x = rearrange(x, "b s n d -> b s (n d)", n=num_heads)
+    elif SAGE_ATTN_AVAILABLE:
+        q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
+        x = sageattn(q, k, v)
+        x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
+    else:
+        q = rearrange(q, "b s (n d) -> b n s d", n=num_heads)
+        k = rearrange(k, "b s (n d) -> b n s d", n=num_heads)
+        v = rearrange(v, "b s (n d) -> b n s d", n=num_heads)
+        x = F.scaled_dot_product_attention(q, k, v)
+        x = rearrange(x, "b n s d -> b s (n d)", n=num_heads)
+    return x
+
+
+def modulate(x: torch.Tensor, shift: torch.Tensor, scale: torch.Tensor):
+    return (x * (1 + scale) + shift)
+
+
+def sinusoidal_embedding_1d(dim, position):
+    sinusoid = torch.outer(position.type(torch.float64), torch.pow(
+        10000, -torch.arange(dim//2, dtype=torch.float64, device=position.device).div(dim//2)))
+    x = torch.cat([torch.cos(sinusoid), torch.sin(sinusoid)], dim=1)
+    return x.to(position.dtype)
+
+
+def precompute_freqs_cis_3d(dim: int, end: int = 1024, theta: float = 10000.0):
+    # 3d rope precompute
+    f_freqs_cis = precompute_freqs_cis(dim - 2 * (dim // 3), end, theta)
+    h_freqs_cis = precompute_freqs_cis(dim // 3, end, theta)
+    w_freqs_cis = precompute_freqs_cis(dim // 3, end, theta)
+    return f_freqs_cis, h_freqs_cis, w_freqs_cis
+
+
+def precompute_freqs_cis(dim: int, end: int = 1024, theta: float = 10000.0):
+    # 1d rope precompute
+    freqs = 1.0 / (theta ** (torch.arange(0, dim, 2)
+                   [: (dim // 2)].double() / dim))
+    freqs = torch.outer(torch.arange(end, device=freqs.device), freqs)
+    freqs_cis = torch.polar(torch.ones_like(freqs), freqs)  # complex64
+    return freqs_cis
+
+
+def rope_apply(x, freqs, num_heads):
+    x = rearrange(x, "b s (n d) -> b s n d", n=num_heads)
+    x_out = torch.view_as_complex(x.to(torch.float64).reshape(
+        x.shape[0], x.shape[1], x.shape[2], -1, 2))
+    x_out = torch.view_as_real(x_out * freqs).flatten(2)
+    return x_out.to(x.dtype)
+
+
+class RMSNorm(nn.Module):
+    def __init__(self, dim, eps=1e-5):
+        super().__init__()
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def norm(self, x):
+        return x * torch.rsqrt(x.pow(2).mean(dim=-1, keepdim=True) + self.eps)
+
+    def forward(self, x):
+        dtype = x.dtype
+        return self.norm(x.float()).to(dtype) * self.weight
+
+
+class AttentionModule(nn.Module):
+    def __init__(self, num_heads):
+        super().__init__()
+        self.num_heads = num_heads
+        
+    def forward(self, q, k, v):
+        x = flash_attention(q=q, k=k, v=v, num_heads=self.num_heads)
+        return x
+
+
+class SelfAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int, eps: float = 1e-6):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+        
+        self.attn = AttentionModule(self.num_heads)
+
+    def forward(self, x, freqs):
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(x))
+        v = self.v(x)
+        q = rope_apply(q, freqs, self.num_heads)
+        k = rope_apply(k, freqs, self.num_heads)
+        x = self.attn(q, k, v)
+        return self.o(x)
+
+
+class CrossAttention(nn.Module):
+    def __init__(self, dim: int, num_heads: int, eps: float = 1e-6, has_image_input: bool = False):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.norm_q = RMSNorm(dim, eps=eps)
+        self.norm_k = RMSNorm(dim, eps=eps)
+        self.has_image_input = has_image_input
+        if has_image_input:
+            self.k_img = nn.Linear(dim, dim)
+            self.v_img = nn.Linear(dim, dim)
+            self.norm_k_img = RMSNorm(dim, eps=eps)
+            
+        self.attn = AttentionModule(self.num_heads)
+
+    def forward(self, x: torch.Tensor, y: torch.Tensor):
+        if self.has_image_input:
+            img = y[:, :257]
+            ctx = y[:, 257:]
+        else:
+            ctx = y
+        q = self.norm_q(self.q(x))
+        k = self.norm_k(self.k(ctx))
+        v = self.v(ctx)
+        x = self.attn(q, k, v)
+        if self.has_image_input:
+            k_img = self.norm_k_img(self.k_img(img))
+            v_img = self.v_img(img)
+            y = flash_attention(q, k_img, v_img, num_heads=self.num_heads)
+            x = x + y
+        return self.o(x)
+
+
+class GateModule(nn.Module):
+    def __init__(self,):
+        super().__init__()
+
+    def forward(self, x, gate, residual):
+        return x + gate * residual
+
+class DiTBlock(nn.Module):
+    def __init__(self, has_image_input: bool, dim: int, num_heads: int, ffn_dim: int, eps: float = 1e-6):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.ffn_dim = ffn_dim
+
+        self.self_attn = SelfAttention(dim, num_heads, eps)
+        self.cross_attn = CrossAttention(
+            dim, num_heads, eps, has_image_input=has_image_input)
+        self.norm1 = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm2 = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.norm3 = nn.LayerNorm(dim, eps=eps)
+        self.ffn = nn.Sequential(nn.Linear(dim, ffn_dim), nn.GELU(
+            approximate='tanh'), nn.Linear(ffn_dim, dim))
+        self.modulation = nn.Parameter(torch.randn(1, 6, dim) / dim**0.5)
+        self.gate = GateModule()
+
+    def forward(self, x, context, t_mod, freqs):
+        has_seq = len(t_mod.shape) == 4
+        chunk_dim = 2 if has_seq else 1
+        # msa: multi-head self-attention  mlp: multi-layer perceptron
+        shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+            self.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + t_mod).chunk(6, dim=chunk_dim)
+        if has_seq:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = (
+                shift_msa.squeeze(2), scale_msa.squeeze(2), gate_msa.squeeze(2),
+                shift_mlp.squeeze(2), scale_mlp.squeeze(2), gate_mlp.squeeze(2),
+            )
+        input_x = modulate(self.norm1(x), shift_msa, scale_msa)
+        x = self.gate(x, gate_msa, self.self_attn(input_x, freqs))
+        x = x + self.cross_attn(self.norm3(x), context)
+        input_x = modulate(self.norm2(x), shift_mlp, scale_mlp)
+        x = self.gate(x, gate_mlp, self.ffn(input_x))
+        return x
+
+
+class MLP(torch.nn.Module):
+    def __init__(self, in_dim, out_dim, has_pos_emb=False):
+        super().__init__()
+        self.proj = torch.nn.Sequential(
+            nn.LayerNorm(in_dim),
+            nn.Linear(in_dim, in_dim),
+            nn.GELU(),
+            nn.Linear(in_dim, out_dim),
+            nn.LayerNorm(out_dim)
+        )
+        self.has_pos_emb = has_pos_emb
+        if has_pos_emb:
+            self.emb_pos = torch.nn.Parameter(torch.zeros((1, 514, 1280)))
+
+    def forward(self, x):
+        if self.has_pos_emb:
+            x = x + self.emb_pos.to(dtype=x.dtype, device=x.device)
+        return self.proj(x)
+
+
+class Head(nn.Module):
+    def __init__(self, dim: int, out_dim: int, patch_size: Tuple[int, int, int], eps: float):
+        super().__init__()
+        self.dim = dim
+        self.patch_size = patch_size
+        self.norm = nn.LayerNorm(dim, eps=eps, elementwise_affine=False)
+        self.head = nn.Linear(dim, out_dim * math.prod(patch_size))
+        self.modulation = nn.Parameter(torch.randn(1, 2, dim) / dim**0.5)
+
+    def forward(self, x, t_mod):
+        if len(t_mod.shape) == 3:
+            shift, scale = (self.modulation.unsqueeze(0).to(dtype=t_mod.dtype, device=t_mod.device) + t_mod.unsqueeze(2)).chunk(2, dim=2)
+            x = (self.head(self.norm(x) * (1 + scale.squeeze(2)) + shift.squeeze(2)))
+        else:
+            shift, scale = (self.modulation.to(dtype=t_mod.dtype, device=t_mod.device) + t_mod).chunk(2, dim=1)
+            x = (self.head(self.norm(x) * (1 + scale) + shift))
+        return x
+
+
+class WanModel(torch.nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        in_dim: int,
+        ffn_dim: int,
+        out_dim: int,
+        text_dim: int,
+        freq_dim: int,
+        eps: float,
+        patch_size: Tuple[int, int, int],
+        num_heads: int,
+        num_layers: int,
+        has_image_input: bool,
+        has_image_pos_emb: bool = False,
+        has_ref_conv: bool = False,
+        add_control_adapter: bool = False,
+        in_dim_control_adapter: int = 24,
+        seperated_timestep: bool = False,
+        require_vae_embedding: bool = True,
+        require_clip_embedding: bool = True,
+        fuse_vae_embedding_in_latents: bool = False,
+    ):
+        super().__init__()
+        self.dim = dim
+        self.freq_dim = freq_dim
+        self.has_image_input = has_image_input
+        self.patch_size = patch_size
+        self.seperated_timestep = seperated_timestep
+        self.require_vae_embedding = require_vae_embedding
+        self.require_clip_embedding = require_clip_embedding
+        self.fuse_vae_embedding_in_latents = fuse_vae_embedding_in_latents
+
+        self.patch_embedding = nn.Conv3d(
+            in_dim, dim, kernel_size=patch_size, stride=patch_size)
+        self.text_embedding = nn.Sequential(
+            nn.Linear(text_dim, dim),
+            nn.GELU(approximate='tanh'),
+            nn.Linear(dim, dim)
+        )
+        self.time_embedding = nn.Sequential(
+            nn.Linear(freq_dim, dim),
+            nn.SiLU(),
+            nn.Linear(dim, dim)
+        )
+        self.time_projection = nn.Sequential(
+            nn.SiLU(), nn.Linear(dim, dim * 6))
+        self.blocks = nn.ModuleList([
+            DiTBlock(has_image_input, dim, num_heads, ffn_dim, eps)
+            for _ in range(num_layers)
+        ])
+        self.head = Head(dim, out_dim, patch_size, eps)
+        head_dim = dim // num_heads
+        self.freqs = precompute_freqs_cis_3d(head_dim)
+
+        if has_image_input:
+            self.img_emb = MLP(1280, dim, has_pos_emb=has_image_pos_emb)  # clip_feature_dim = 1280
+        if has_ref_conv:
+            self.ref_conv = nn.Conv2d(16, dim, kernel_size=(2, 2), stride=(2, 2))
+        self.has_image_pos_emb = has_image_pos_emb
+        self.has_ref_conv = has_ref_conv
+        if add_control_adapter:
+            self.control_adapter = SimpleAdapter(in_dim_control_adapter, dim, kernel_size=patch_size[1:], stride=patch_size[1:])
+        else:
+            self.control_adapter = None
+
+    def patchify(self, x: torch.Tensor,control_camera_latents_input: torch.Tensor = None):
+        
+        x = self.patch_embedding(x) #* from ([1, 48, 21, 30, 40]) to [1, 1536, 21, 15, 20]), 
+        if self.control_adapter is not None and control_camera_latents_input is not None:
+            y_camera = self.control_adapter(control_camera_latents_input)
+            x = [u + v for u, v in zip(x, y_camera)]
+            x = x[0].unsqueeze(0)
+        grid_size = x.shape[2:]#* get the (F,H,W)
+        x = rearrange(x, 'b c f h w -> b (f h w) c').contiguous()
+        return x, grid_size  # x, grid_size: (f, h, w)
+
+    def unpatchify(self, x: torch.Tensor, grid_size: torch.Tensor):
+        return rearrange(
+            x, 'b (f h w) (x y z c) -> b c (f x) (h y) (w z)',
+            f=grid_size[0], h=grid_size[1], w=grid_size[2], 
+            x=self.patch_size[0], y=self.patch_size[1], z=self.patch_size[2]
+        )
+
+    def forward(self,
+                x: torch.Tensor,
+                timestep: torch.Tensor,
+                context: torch.Tensor,
+                clip_feature: Optional[torch.Tensor] = None,
+                y: Optional[torch.Tensor] = None,
+                use_gradient_checkpointing: bool = False,
+                use_gradient_checkpointing_offload: bool = False,
+                **kwargs,
+                ):
+        t = self.time_embedding(
+            sinusoidal_embedding_1d(self.freq_dim, timestep))
+        t_mod = self.time_projection(t).unflatten(1, (6, self.dim))
+        context = self.text_embedding(context)
+        
+        if self.has_image_input:
+            x = torch.cat([x, y], dim=1)  # (b, c_x + c_y, f, h, w)
+            clip_embdding = self.img_emb(clip_feature)
+            context = torch.cat([clip_embdding, context], dim=1)
+        
+        x, (f, h, w) = self.patchify(x)
+        
+        freqs = torch.cat([
+            self.freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
+            self.freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+            self.freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+        ], dim=-1).reshape(f * h * w, 1, -1).to(x.device)
+        
+        def create_custom_forward(module):
+            def custom_forward(*inputs):
+                return module(*inputs)
+            return custom_forward
+
+        for block in self.blocks:
+            if self.training and use_gradient_checkpointing:
+                if use_gradient_checkpointing_offload:
+                    with torch.autograd.graph.save_on_cpu():
+                        x = torch.utils.checkpoint.checkpoint(
+                            create_custom_forward(block),
+                            x, context, t_mod, freqs,
+                            use_reentrant=False,
+                        )
+                else:
+                    x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        x, context, t_mod, freqs,
+                        use_reentrant=False,
+                    )
+            else:
+                x = block(x, context, t_mod, freqs)
+
+        x = self.head(x, t)
+        x = self.unpatchify(x, (f, h, w))
+        return x
+
+    @staticmethod
+    def state_dict_converter():
+        return WanModelStateDictConverter()
+    
+    
+class WanModelStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        rename_dict = {
+            "blocks.0.attn1.norm_k.weight": "blocks.0.self_attn.norm_k.weight",
+            "blocks.0.attn1.norm_q.weight": "blocks.0.self_attn.norm_q.weight",
+            "blocks.0.attn1.to_k.bias": "blocks.0.self_attn.k.bias",
+            "blocks.0.attn1.to_k.weight": "blocks.0.self_attn.k.weight",
+            "blocks.0.attn1.to_out.0.bias": "blocks.0.self_attn.o.bias",
+            "blocks.0.attn1.to_out.0.weight": "blocks.0.self_attn.o.weight",
+            "blocks.0.attn1.to_q.bias": "blocks.0.self_attn.q.bias",
+            "blocks.0.attn1.to_q.weight": "blocks.0.self_attn.q.weight",
+            "blocks.0.attn1.to_v.bias": "blocks.0.self_attn.v.bias",
+            "blocks.0.attn1.to_v.weight": "blocks.0.self_attn.v.weight",
+            "blocks.0.attn2.norm_k.weight": "blocks.0.cross_attn.norm_k.weight",
+            "blocks.0.attn2.norm_q.weight": "blocks.0.cross_attn.norm_q.weight",
+            "blocks.0.attn2.to_k.bias": "blocks.0.cross_attn.k.bias",
+            "blocks.0.attn2.to_k.weight": "blocks.0.cross_attn.k.weight",
+            "blocks.0.attn2.to_out.0.bias": "blocks.0.cross_attn.o.bias",
+            "blocks.0.attn2.to_out.0.weight": "blocks.0.cross_attn.o.weight",
+            "blocks.0.attn2.to_q.bias": "blocks.0.cross_attn.q.bias",
+            "blocks.0.attn2.to_q.weight": "blocks.0.cross_attn.q.weight",
+            "blocks.0.attn2.to_v.bias": "blocks.0.cross_attn.v.bias",
+            "blocks.0.attn2.to_v.weight": "blocks.0.cross_attn.v.weight",
+            "blocks.0.ffn.net.0.proj.bias": "blocks.0.ffn.0.bias",
+            "blocks.0.ffn.net.0.proj.weight": "blocks.0.ffn.0.weight",
+            "blocks.0.ffn.net.2.bias": "blocks.0.ffn.2.bias",
+            "blocks.0.ffn.net.2.weight": "blocks.0.ffn.2.weight",
+            "blocks.0.norm2.bias": "blocks.0.norm3.bias",
+            "blocks.0.norm2.weight": "blocks.0.norm3.weight",
+            "blocks.0.scale_shift_table": "blocks.0.modulation",
+            "condition_embedder.text_embedder.linear_1.bias": "text_embedding.0.bias",
+            "condition_embedder.text_embedder.linear_1.weight": "text_embedding.0.weight",
+            "condition_embedder.text_embedder.linear_2.bias": "text_embedding.2.bias",
+            "condition_embedder.text_embedder.linear_2.weight": "text_embedding.2.weight",
+            "condition_embedder.time_embedder.linear_1.bias": "time_embedding.0.bias",
+            "condition_embedder.time_embedder.linear_1.weight": "time_embedding.0.weight",
+            "condition_embedder.time_embedder.linear_2.bias": "time_embedding.2.bias",
+            "condition_embedder.time_embedder.linear_2.weight": "time_embedding.2.weight",
+            "condition_embedder.time_proj.bias": "time_projection.1.bias",
+            "condition_embedder.time_proj.weight": "time_projection.1.weight",
+            "patch_embedding.bias": "patch_embedding.bias",
+            "patch_embedding.weight": "patch_embedding.weight",
+            "scale_shift_table": "head.modulation",
+            "proj_out.bias": "head.head.bias",
+            "proj_out.weight": "head.head.weight",
+        }
+        state_dict_ = {}
+        for name, param in state_dict.items():
+            if name in rename_dict:
+                state_dict_[rename_dict[name]] = param
+            else:
+                name_ = ".".join(name.split(".")[:1] + ["0"] + name.split(".")[2:])
+                if name_ in rename_dict:
+                    name_ = rename_dict[name_]
+                    name_ = ".".join(name_.split(".")[:1] + [name.split(".")[1]] + name_.split(".")[2:])
+                    state_dict_[name_] = param
+        if hash_state_dict_keys(state_dict) == "cb104773c6c2cb6df4f9529ad5c60d0b":
+            config = {
+                "model_type": "t2v",
+                "patch_size": (1, 2, 2),
+                "text_len": 512,
+                "in_dim": 16,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "window_size": (-1, -1),
+                "qk_norm": True,
+                "cross_attn_norm": True,
+                "eps": 1e-6,
+            }
+        else:
+            config = {}
+        return state_dict_, config
+    
+    def from_civitai(self, state_dict):
+        state_dict = {name: param for name, param in state_dict.items() if not name.startswith("vace")}
+        if hash_state_dict_keys(state_dict) == "9269f8db9040a9d860eaca435be61814":
+            config = {
+                "has_image_input": False,
+                "patch_size": [1, 2, 2],
+                "in_dim": 16,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "aafcfd9672c3a2456dc46e1cb6e52c70":
+            config = {
+                "has_image_input": False,
+                "patch_size": [1, 2, 2],
+                "in_dim": 16,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "6bfcfb3b342cb286ce886889d519a77e":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "6d6ccde6845b95ad9114ab993d917893":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "6bfcfb3b342cb286ce886889d519a77e":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "349723183fc063b2bfc10bb2835cf677":
+            # 1.3B PAI control
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "efa44cddf936c70abd0ea28b6cbe946c":
+            # 14B PAI control
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6
+            }
+        elif hash_state_dict_keys(state_dict) == "3ef3b1f8e1dab83d5b71fd7b617f859f":
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "has_image_pos_emb": True
+            }
+        elif hash_state_dict_keys(state_dict) == "70ddad9d3a133785da5ea371aae09504":
+            # 1.3B PAI control v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+                "has_ref_conv": True
+            }
+        elif hash_state_dict_keys(state_dict) == "26bde73488a92e64cc20b0a7485b9e5b":
+            # 14B PAI control v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "has_ref_conv": True
+            }
+        elif hash_state_dict_keys(state_dict) == "ac6a5aa74f4a0aab6f64eb9a72f19901":
+            # 1.3B PAI control-camera v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 32,
+                "dim": 1536,
+                "ffn_dim": 8960,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 12,
+                "num_layers": 30,
+                "eps": 1e-6,
+                "has_ref_conv": False,
+                "add_control_adapter": True,
+                "in_dim_control_adapter": 24,
+            }
+        elif hash_state_dict_keys(state_dict) == "b61c605c2adbd23124d152ed28e049ae":
+            # 14B PAI control-camera v1.1
+            config = {
+                "has_image_input": True,
+                "patch_size": [1, 2, 2],
+                "in_dim": 32,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "has_ref_conv": False,
+                "add_control_adapter": True,
+                "in_dim_control_adapter": 24,
+            }
+        elif hash_state_dict_keys(state_dict) == "1f5ab7703c6fc803fdded85ff040c316":
+            # Wan-AI/Wan2.2-TI2V-5B
+            config = {
+                "has_image_input": False,
+                "patch_size": [1, 2, 2],
+                "in_dim": 48,
+                "dim": 3072,
+                "ffn_dim": 14336,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 48,
+                "num_heads": 24,
+                "num_layers": 30,
+                "eps": 1e-6,
+                "seperated_timestep": True,
+                "require_clip_embedding": False,
+                "require_vae_embedding": False,
+                "fuse_vae_embedding_in_latents": True,
+            }
+        elif hash_state_dict_keys(state_dict) == "5b013604280dd715f8457c6ed6d6a626":
+            # Wan-AI/Wan2.2-I2V-A14B
+            config = {
+                "has_image_input": False,
+                "patch_size": [1, 2, 2],
+                "in_dim": 36,
+                "dim": 5120,
+                "ffn_dim": 13824,
+                "freq_dim": 256,
+                "text_dim": 4096,
+                "out_dim": 16,
+                "num_heads": 40,
+                "num_layers": 40,
+                "eps": 1e-6,
+                "require_clip_embedding": False,
+            }
+        else:
+            config = {}
+        return state_dict, config

dkt/models/wan_video_image_encoder.py

@@ -0,0 +1,902 @@
+"""
+Concise re-implementation of
+``https://github.com/openai/CLIP'' and
+``https://github.com/mlfoundations/open_clip''.
+"""
+import math
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torchvision.transforms as T
+from .wan_video_dit import flash_attention
+
+
+class SelfAttention(nn.Module):
+
+    def __init__(self, dim, num_heads, dropout=0.1, eps=1e-5):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.eps = eps
+
+        # layers
+        self.q = nn.Linear(dim, dim)
+        self.k = nn.Linear(dim, dim)
+        self.v = nn.Linear(dim, dim)
+        self.o = nn.Linear(dim, dim)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x, mask):
+        """
+        x:   [B, L, C].
+        """
+        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.q(x).reshape(b, s, n, d).permute(0, 2, 1, 3)
+        k = self.k(x).reshape(b, s, n, d).permute(0, 2, 1, 3)
+        v = self.v(x).reshape(b, s, n, d).permute(0, 2, 1, 3)
+
+        # compute attention
+        p = self.dropout.p if self.training else 0.0
+        x = F.scaled_dot_product_attention(q, k, v, mask, p)
+        x = x.permute(0, 2, 1, 3).reshape(b, s, c)
+
+        # output
+        x = self.o(x)
+        x = self.dropout(x)
+        return x
+
+
+class AttentionBlock(nn.Module):
+
+    def __init__(self, dim, num_heads, post_norm, dropout=0.1, eps=1e-5):
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.post_norm = post_norm
+        self.eps = eps
+
+        # layers
+        self.attn = SelfAttention(dim, num_heads, dropout, eps)
+        self.norm1 = nn.LayerNorm(dim, eps=eps)
+        self.ffn = nn.Sequential(
+            nn.Linear(dim, dim * 4), nn.GELU(), nn.Linear(dim * 4, dim),
+            nn.Dropout(dropout))
+        self.norm2 = nn.LayerNorm(dim, eps=eps)
+
+    def forward(self, x, mask):
+        if self.post_norm:
+            x = self.norm1(x + self.attn(x, mask))
+            x = self.norm2(x + self.ffn(x))
+        else:
+            x = x + self.attn(self.norm1(x), mask)
+            x = x + self.ffn(self.norm2(x))
+        return x
+
+
+class XLMRoberta(nn.Module):
+    """
+    XLMRobertaModel with no pooler and no LM head.
+    """
+
+    def __init__(self,
+                 vocab_size=250002,
+                 max_seq_len=514,
+                 type_size=1,
+                 pad_id=1,
+                 dim=1024,
+                 num_heads=16,
+                 num_layers=24,
+                 post_norm=True,
+                 dropout=0.1,
+                 eps=1e-5):
+        super().__init__()
+        self.vocab_size = vocab_size
+        self.max_seq_len = max_seq_len
+        self.type_size = type_size
+        self.pad_id = pad_id
+        self.dim = dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.post_norm = post_norm
+        self.eps = eps
+
+        # embeddings
+        self.token_embedding = nn.Embedding(vocab_size, dim, padding_idx=pad_id)
+        self.type_embedding = nn.Embedding(type_size, dim)
+        self.pos_embedding = nn.Embedding(max_seq_len, dim, padding_idx=pad_id)
+        self.dropout = nn.Dropout(dropout)
+
+        # blocks
+        self.blocks = nn.ModuleList([
+            AttentionBlock(dim, num_heads, post_norm, dropout, eps)
+            for _ in range(num_layers)
+        ])
+
+        # norm layer
+        self.norm = nn.LayerNorm(dim, eps=eps)
+
+    def forward(self, ids):
+        """
+        ids: [B, L] of torch.LongTensor.
+        """
+        b, s = ids.shape
+        mask = ids.ne(self.pad_id).long()
+
+        # embeddings
+        x = self.token_embedding(ids) + \
+            self.type_embedding(torch.zeros_like(ids)) + \
+            self.pos_embedding(self.pad_id + torch.cumsum(mask, dim=1) * mask)
+        if self.post_norm:
+            x = self.norm(x)
+        x = self.dropout(x)
+
+        # blocks
+        mask = torch.where(
+            mask.view(b, 1, 1, s).gt(0), 0.0,
+            torch.finfo(x.dtype).min)
+        for block in self.blocks:
+            x = block(x, mask)
+
+        # output
+        if not self.post_norm:
+            x = self.norm(x)
+        return x
+
+
+def xlm_roberta_large(pretrained=False,
+                      return_tokenizer=False,
+                      device='cpu',
+                      **kwargs):
+    """
+    XLMRobertaLarge adapted from Huggingface.
+    """
+    # params
+    cfg = dict(
+        vocab_size=250002,
+        max_seq_len=514,
+        type_size=1,
+        pad_id=1,
+        dim=1024,
+        num_heads=16,
+        num_layers=24,
+        post_norm=True,
+        dropout=0.1,
+        eps=1e-5)
+    cfg.update(**kwargs)
+
+    # init model
+    if pretrained:
+        from sora import DOWNLOAD_TO_CACHE
+
+        # init a meta model
+        with torch.device('meta'):
+            model = XLMRoberta(**cfg)
+
+        # load checkpoint
+        model.load_state_dict(
+            torch.load(
+                DOWNLOAD_TO_CACHE('models/xlm_roberta/xlm_roberta_large.pth'),
+                map_location=device),
+            assign=True)
+    else:
+        # init a model on device
+        with torch.device(device):
+            model = XLMRoberta(**cfg)
+
+    # init tokenizer
+    if return_tokenizer:
+        from sora.data import HuggingfaceTokenizer
+        tokenizer = HuggingfaceTokenizer(
+            name='xlm-roberta-large',
+            seq_len=model.text_len,
+            clean='whitespace')
+        return model, tokenizer
+    else:
+        return model
+
+
+
+def pos_interpolate(pos, seq_len):
+    if pos.size(1) == seq_len:
+        return pos
+    else:
+        src_grid = int(math.sqrt(pos.size(1)))
+        tar_grid = int(math.sqrt(seq_len))
+        n = pos.size(1) - src_grid * src_grid
+        return torch.cat([
+            pos[:, :n],
+            F.interpolate(
+                pos[:, n:].float().reshape(1, src_grid, src_grid, -1).permute(
+                    0, 3, 1, 2),
+                size=(tar_grid, tar_grid),
+                mode='bicubic',
+                align_corners=False).flatten(2).transpose(1, 2)
+        ],
+                         dim=1)
+
+
+class QuickGELU(nn.Module):
+
+    def forward(self, x):
+        return x * torch.sigmoid(1.702 * x)
+
+
+class LayerNorm(nn.LayerNorm):
+
+    def forward(self, x):
+        return super().forward(x).type_as(x)
+
+
+class SelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 num_heads,
+                 causal=False,
+                 attn_dropout=0.0,
+                 proj_dropout=0.0):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.causal = causal
+        self.attn_dropout = attn_dropout
+        self.proj_dropout = proj_dropout
+
+        # layers
+        self.to_qkv = nn.Linear(dim, dim * 3)
+        self.proj = nn.Linear(dim, dim)
+
+    def forward(self, x):
+        """
+        x:   [B, L, C].
+        """
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).chunk(3, dim=-1)
+
+        # compute attention
+        x = flash_attention(q, k, v, num_heads=self.num_heads, compatibility_mode=True)
+
+        # output
+        x = self.proj(x)
+        x = F.dropout(x, self.proj_dropout, self.training)
+        return x
+
+
+class SwiGLU(nn.Module):
+
+    def __init__(self, dim, mid_dim):
+        super().__init__()
+        self.dim = dim
+        self.mid_dim = mid_dim
+
+        # layers
+        self.fc1 = nn.Linear(dim, mid_dim)
+        self.fc2 = nn.Linear(dim, mid_dim)
+        self.fc3 = nn.Linear(mid_dim, dim)
+
+    def forward(self, x):
+        x = F.silu(self.fc1(x)) * self.fc2(x)
+        x = self.fc3(x)
+        return x
+
+
+class AttentionBlock(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 mlp_ratio,
+                 num_heads,
+                 post_norm=False,
+                 causal=False,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 norm_eps=1e-5):
+        assert activation in ['quick_gelu', 'gelu', 'swi_glu']
+        super().__init__()
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.post_norm = post_norm
+        self.causal = causal
+        self.norm_eps = norm_eps
+
+        # layers
+        self.norm1 = LayerNorm(dim, eps=norm_eps)
+        self.attn = SelfAttention(dim, num_heads, causal, attn_dropout,
+                                  proj_dropout)
+        self.norm2 = LayerNorm(dim, eps=norm_eps)
+        if activation == 'swi_glu':
+            self.mlp = SwiGLU(dim, int(dim * mlp_ratio))
+        else:
+            self.mlp = nn.Sequential(
+                nn.Linear(dim, int(dim * mlp_ratio)),
+                QuickGELU() if activation == 'quick_gelu' else nn.GELU(),
+                nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))
+
+    def forward(self, x):
+        if self.post_norm:
+            x = x + self.norm1(self.attn(x))
+            x = x + self.norm2(self.mlp(x))
+        else:
+            x = x + self.attn(self.norm1(x))
+            x = x + self.mlp(self.norm2(x))
+        return x
+
+
+class AttentionPool(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 mlp_ratio,
+                 num_heads,
+                 activation='gelu',
+                 proj_dropout=0.0,
+                 norm_eps=1e-5):
+        assert dim % num_heads == 0
+        super().__init__()
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.num_heads = num_heads
+        self.head_dim = dim // num_heads
+        self.proj_dropout = proj_dropout
+        self.norm_eps = norm_eps
+
+        # layers
+        gain = 1.0 / math.sqrt(dim)
+        self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim))
+        self.to_q = nn.Linear(dim, dim)
+        self.to_kv = nn.Linear(dim, dim * 2)
+        self.proj = nn.Linear(dim, dim)
+        self.norm = LayerNorm(dim, eps=norm_eps)
+        self.mlp = nn.Sequential(
+            nn.Linear(dim, int(dim * mlp_ratio)),
+            QuickGELU() if activation == 'quick_gelu' else nn.GELU(),
+            nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))
+
+    def forward(self, x):
+        """
+        x:  [B, L, C].
+        """
+        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.to_q(self.cls_embedding).view(1, 1, n*d).expand(b, -1, -1)
+        k, v = self.to_kv(x).chunk(2, dim=-1)
+
+        # compute attention
+        x = flash_attention(q, k, v, num_heads=self.num_heads, compatibility_mode=True)
+        x = x.reshape(b, 1, c)
+
+        # output
+        x = self.proj(x)
+        x = F.dropout(x, self.proj_dropout, self.training)
+
+        # mlp
+        x = x + self.mlp(self.norm(x))
+        return x[:, 0]
+
+
+class VisionTransformer(nn.Module):
+
+    def __init__(self,
+                 image_size=224,
+                 patch_size=16,
+                 dim=768,
+                 mlp_ratio=4,
+                 out_dim=512,
+                 num_heads=12,
+                 num_layers=12,
+                 pool_type='token',
+                 pre_norm=True,
+                 post_norm=False,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        if image_size % patch_size != 0:
+            print(
+                '[WARNING] image_size is not divisible by patch_size',
+                flush=True)
+        assert pool_type in ('token', 'token_fc', 'attn_pool')
+        out_dim = out_dim or dim
+        super().__init__()
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.num_patches = (image_size // patch_size)**2
+        self.dim = dim
+        self.mlp_ratio = mlp_ratio
+        self.out_dim = out_dim
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.pool_type = pool_type
+        self.post_norm = post_norm
+        self.norm_eps = norm_eps
+
+        # embeddings
+        gain = 1.0 / math.sqrt(dim)
+        self.patch_embedding = nn.Conv2d(
+            3,
+            dim,
+            kernel_size=patch_size,
+            stride=patch_size,
+            bias=not pre_norm)
+        if pool_type in ('token', 'token_fc'):
+            self.cls_embedding = nn.Parameter(gain * torch.randn(1, 1, dim))
+        self.pos_embedding = nn.Parameter(gain * torch.randn(
+            1, self.num_patches +
+            (1 if pool_type in ('token', 'token_fc') else 0), dim))
+        self.dropout = nn.Dropout(embedding_dropout)
+
+        # transformer
+        self.pre_norm = LayerNorm(dim, eps=norm_eps) if pre_norm else None
+        self.transformer = nn.Sequential(*[
+            AttentionBlock(dim, mlp_ratio, num_heads, post_norm, False,
+                           activation, attn_dropout, proj_dropout, norm_eps)
+            for _ in range(num_layers)
+        ])
+        self.post_norm = LayerNorm(dim, eps=norm_eps)
+
+        # head
+        if pool_type == 'token':
+            self.head = nn.Parameter(gain * torch.randn(dim, out_dim))
+        elif pool_type == 'token_fc':
+            self.head = nn.Linear(dim, out_dim)
+        elif pool_type == 'attn_pool':
+            self.head = AttentionPool(dim, mlp_ratio, num_heads, activation,
+                                      proj_dropout, norm_eps)
+
+    def forward(self, x, interpolation=False, use_31_block=False):
+        b = x.size(0)
+
+        # embeddings
+        x = self.patch_embedding(x).flatten(2).permute(0, 2, 1)
+        if self.pool_type in ('token', 'token_fc'):
+            x = torch.cat([self.cls_embedding.expand(b, -1, -1).to(dtype=x.dtype, device=x.device), x], dim=1)
+        if interpolation:
+            e = pos_interpolate(self.pos_embedding, x.size(1))
+        else:
+            e = self.pos_embedding
+        e = e.to(dtype=x.dtype, device=x.device)
+        x = self.dropout(x + e)
+        if self.pre_norm is not None:
+            x = self.pre_norm(x)
+
+        # transformer
+        if use_31_block:
+            x = self.transformer[:-1](x)
+            return x
+        else:
+            x = self.transformer(x)
+            return x
+
+
+class CLIP(nn.Module):
+
+    def __init__(self,
+                 embed_dim=512,
+                 image_size=224,
+                 patch_size=16,
+                 vision_dim=768,
+                 vision_mlp_ratio=4,
+                 vision_heads=12,
+                 vision_layers=12,
+                 vision_pool='token',
+                 vision_pre_norm=True,
+                 vision_post_norm=False,
+                 vocab_size=49408,
+                 text_len=77,
+                 text_dim=512,
+                 text_mlp_ratio=4,
+                 text_heads=8,
+                 text_layers=12,
+                 text_causal=True,
+                 text_pool='argmax',
+                 text_head_bias=False,
+                 logit_bias=None,
+                 activation='quick_gelu',
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.vision_dim = vision_dim
+        self.vision_mlp_ratio = vision_mlp_ratio
+        self.vision_heads = vision_heads
+        self.vision_layers = vision_layers
+        self.vision_pool = vision_pool
+        self.vision_pre_norm = vision_pre_norm
+        self.vision_post_norm = vision_post_norm
+        self.vocab_size = vocab_size
+        self.text_len = text_len
+        self.text_dim = text_dim
+        self.text_mlp_ratio = text_mlp_ratio
+        self.text_heads = text_heads
+        self.text_layers = text_layers
+        self.text_causal = text_causal
+        self.text_pool = text_pool
+        self.text_head_bias = text_head_bias
+        self.norm_eps = norm_eps
+
+        # models
+        self.visual = VisionTransformer(
+            image_size=image_size,
+            patch_size=patch_size,
+            dim=vision_dim,
+            mlp_ratio=vision_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=vision_heads,
+            num_layers=vision_layers,
+            pool_type=vision_pool,
+            pre_norm=vision_pre_norm,
+            post_norm=vision_post_norm,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.textual = TextTransformer(
+            vocab_size=vocab_size,
+            text_len=text_len,
+            dim=text_dim,
+            mlp_ratio=text_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=text_heads,
+            num_layers=text_layers,
+            causal=text_causal,
+            pool_type=text_pool,
+            head_bias=text_head_bias,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([]))
+        if logit_bias is not None:
+            self.logit_bias = nn.Parameter(logit_bias * torch.ones([]))
+
+        # initialize weights
+        self.init_weights()
+
+    def forward(self, imgs, txt_ids):
+        """
+        imgs:       [B, 3, H, W] of torch.float32.
+        - mean:     [0.48145466, 0.4578275, 0.40821073]
+        - std:      [0.26862954, 0.26130258, 0.27577711]
+        txt_ids:    [B, L] of torch.long. Encoded by data.CLIPTokenizer.
+        """
+        xi = self.visual(imgs)
+        xt = self.textual(txt_ids)
+        return xi, xt
+
+    def init_weights(self):
+        # embeddings
+        nn.init.normal_(self.textual.token_embedding.weight, std=0.02)
+        nn.init.normal_(self.visual.patch_embedding.weight, std=0.1)
+
+        # attentions
+        for modality in ['visual', 'textual']:
+            dim = self.vision_dim if modality == 'visual' else self.text_dim
+            transformer = getattr(self, modality).transformer
+            proj_gain = (1.0 / math.sqrt(dim)) * (
+                1.0 / math.sqrt(2 * len(transformer)))
+            attn_gain = 1.0 / math.sqrt(dim)
+            mlp_gain = 1.0 / math.sqrt(2.0 * dim)
+            for block in transformer:
+                nn.init.normal_(block.attn.to_qkv.weight, std=attn_gain)
+                nn.init.normal_(block.attn.proj.weight, std=proj_gain)
+                nn.init.normal_(block.mlp[0].weight, std=mlp_gain)
+                nn.init.normal_(block.mlp[2].weight, std=proj_gain)
+
+    def param_groups(self):
+        groups = [{
+            'params': [
+                p for n, p in self.named_parameters()
+                if 'norm' in n or n.endswith('bias')
+            ],
+            'weight_decay': 0.0
+        }, {
+            'params': [
+                p for n, p in self.named_parameters()
+                if not ('norm' in n or n.endswith('bias'))
+            ]
+        }]
+        return groups
+
+
+class XLMRobertaWithHead(XLMRoberta):
+
+    def __init__(self, **kwargs):
+        self.out_dim = kwargs.pop('out_dim')
+        super().__init__(**kwargs)
+
+        # head
+        mid_dim = (self.dim + self.out_dim) // 2
+        self.head = nn.Sequential(
+            nn.Linear(self.dim, mid_dim, bias=False), nn.GELU(),
+            nn.Linear(mid_dim, self.out_dim, bias=False))
+
+    def forward(self, ids):
+        # xlm-roberta
+        x = super().forward(ids)
+
+        # average pooling
+        mask = ids.ne(self.pad_id).unsqueeze(-1).to(x)
+        x = (x * mask).sum(dim=1) / mask.sum(dim=1)
+
+        # head
+        x = self.head(x)
+        return x
+
+
+class XLMRobertaCLIP(nn.Module):
+
+    def __init__(self,
+                 embed_dim=1024,
+                 image_size=224,
+                 patch_size=14,
+                 vision_dim=1280,
+                 vision_mlp_ratio=4,
+                 vision_heads=16,
+                 vision_layers=32,
+                 vision_pool='token',
+                 vision_pre_norm=True,
+                 vision_post_norm=False,
+                 activation='gelu',
+                 vocab_size=250002,
+                 max_text_len=514,
+                 type_size=1,
+                 pad_id=1,
+                 text_dim=1024,
+                 text_heads=16,
+                 text_layers=24,
+                 text_post_norm=True,
+                 text_dropout=0.1,
+                 attn_dropout=0.0,
+                 proj_dropout=0.0,
+                 embedding_dropout=0.0,
+                 norm_eps=1e-5):
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.image_size = image_size
+        self.patch_size = patch_size
+        self.vision_dim = vision_dim
+        self.vision_mlp_ratio = vision_mlp_ratio
+        self.vision_heads = vision_heads
+        self.vision_layers = vision_layers
+        self.vision_pre_norm = vision_pre_norm
+        self.vision_post_norm = vision_post_norm
+        self.activation = activation
+        self.vocab_size = vocab_size
+        self.max_text_len = max_text_len
+        self.type_size = type_size
+        self.pad_id = pad_id
+        self.text_dim = text_dim
+        self.text_heads = text_heads
+        self.text_layers = text_layers
+        self.text_post_norm = text_post_norm
+        self.norm_eps = norm_eps
+
+        # models
+        self.visual = VisionTransformer(
+            image_size=image_size,
+            patch_size=patch_size,
+            dim=vision_dim,
+            mlp_ratio=vision_mlp_ratio,
+            out_dim=embed_dim,
+            num_heads=vision_heads,
+            num_layers=vision_layers,
+            pool_type=vision_pool,
+            pre_norm=vision_pre_norm,
+            post_norm=vision_post_norm,
+            activation=activation,
+            attn_dropout=attn_dropout,
+            proj_dropout=proj_dropout,
+            embedding_dropout=embedding_dropout,
+            norm_eps=norm_eps)
+        self.textual = None
+        self.log_scale = nn.Parameter(math.log(1 / 0.07) * torch.ones([]))
+
+    def forward(self, imgs, txt_ids):
+        """
+        imgs:       [B, 3, H, W] of torch.float32.
+        - mean:     [0.48145466, 0.4578275, 0.40821073]
+        - std:      [0.26862954, 0.26130258, 0.27577711]
+        txt_ids:    [B, L] of torch.long.
+                    Encoded by data.CLIPTokenizer.
+        """
+        xi = self.visual(imgs)
+        xt = self.textual(txt_ids)
+        return xi, xt
+
+    def param_groups(self):
+        groups = [{
+            'params': [
+                p for n, p in self.named_parameters()
+                if 'norm' in n or n.endswith('bias')
+            ],
+            'weight_decay': 0.0
+        }, {
+            'params': [
+                p for n, p in self.named_parameters()
+                if not ('norm' in n or n.endswith('bias'))
+            ]
+        }]
+        return groups
+
+
+def _clip(pretrained=False,
+          pretrained_name=None,
+          model_cls=CLIP,
+          return_transforms=False,
+          return_tokenizer=False,
+          tokenizer_padding='eos',
+          dtype=torch.float32,
+          device='cpu',
+          **kwargs):
+    # init model
+    if pretrained and pretrained_name:
+        from sora import BUCKET, DOWNLOAD_TO_CACHE
+
+        # init a meta model
+        with torch.device('meta'):
+            model = model_cls(**kwargs)
+
+        # checkpoint path
+        checkpoint = f'models/clip/{pretrained_name}'
+        if dtype in (torch.float16, torch.bfloat16):
+            suffix = '-' + {
+                torch.float16: 'fp16',
+                torch.bfloat16: 'bf16'
+            }[dtype]
+            if object_exists(BUCKET, f'{checkpoint}{suffix}.pth'):
+                checkpoint = f'{checkpoint}{suffix}'
+        checkpoint += '.pth'
+
+        # load
+        model.load_state_dict(
+            torch.load(DOWNLOAD_TO_CACHE(checkpoint), map_location=device),
+            assign=True,
+            strict=False)
+    else:
+        # init a model on device
+        with torch.device(device):
+            model = model_cls(**kwargs)
+
+    # set device
+    output = (model,)
+
+    # init transforms
+    if return_transforms:
+        # mean and std
+        if 'siglip' in pretrained_name.lower():
+            mean, std = [0.5, 0.5, 0.5], [0.5, 0.5, 0.5]
+        else:
+            mean = [0.48145466, 0.4578275, 0.40821073]
+            std = [0.26862954, 0.26130258, 0.27577711]
+
+        # transforms
+        transforms = T.Compose([
+            T.Resize((model.image_size, model.image_size),
+                     interpolation=T.InterpolationMode.BICUBIC),
+            T.ToTensor(),
+            T.Normalize(mean=mean, std=std)
+        ])
+        output += (transforms,)
+
+    # init tokenizer
+    if return_tokenizer:
+        from sora import data
+        if 'siglip' in pretrained_name.lower():
+            tokenizer = data.HuggingfaceTokenizer(
+                name=f'timm/{pretrained_name}',
+                seq_len=model.text_len,
+                clean='canonicalize')
+        elif 'xlm' in pretrained_name.lower():
+            tokenizer = data.HuggingfaceTokenizer(
+                name='xlm-roberta-large',
+                seq_len=model.max_text_len - 2,
+                clean='whitespace')
+        elif 'mba' in pretrained_name.lower():
+            tokenizer = data.HuggingfaceTokenizer(
+                name='facebook/xlm-roberta-xl',
+                seq_len=model.max_text_len - 2,
+                clean='whitespace')
+        else:
+            tokenizer = data.CLIPTokenizer(
+                seq_len=model.text_len, padding=tokenizer_padding)
+        output += (tokenizer,)
+    return output[0] if len(output) == 1 else output
+
+
+def clip_xlm_roberta_vit_h_14(
+        pretrained=False,
+        pretrained_name='open-clip-xlm-roberta-large-vit-huge-14',
+        **kwargs):
+    cfg = dict(
+        embed_dim=1024,
+        image_size=224,
+        patch_size=14,
+        vision_dim=1280,
+        vision_mlp_ratio=4,
+        vision_heads=16,
+        vision_layers=32,
+        vision_pool='token',
+        activation='gelu',
+        vocab_size=250002,
+        max_text_len=514,
+        type_size=1,
+        pad_id=1,
+        text_dim=1024,
+        text_heads=16,
+        text_layers=24,
+        text_post_norm=True,
+        text_dropout=0.1,
+        attn_dropout=0.0,
+        proj_dropout=0.0,
+        embedding_dropout=0.0)
+    cfg.update(**kwargs)
+    return _clip(pretrained, pretrained_name, XLMRobertaCLIP, **cfg)
+
+
+class WanImageEncoder(torch.nn.Module):
+
+    def __init__(self):
+        super().__init__()
+        # init model
+        self.model, self.transforms = clip_xlm_roberta_vit_h_14(
+            pretrained=False,
+            return_transforms=True,
+            return_tokenizer=False,
+            dtype=torch.float32,
+            device="cpu")
+
+    def encode_image(self, videos):
+        # preprocess
+        size = (self.model.image_size,) * 2
+        videos = torch.cat([
+            F.interpolate(
+                u,
+                size=size,
+                mode='bicubic',
+                align_corners=False) for u in videos
+        ])
+        videos = self.transforms.transforms[-1](videos.mul_(0.5).add_(0.5))
+
+        # forward
+        dtype = next(iter(self.model.visual.parameters())).dtype
+        videos = videos.to(dtype)
+        out = self.model.visual(videos, use_31_block=True)
+        return out
+        
+    @staticmethod
+    def state_dict_converter():
+        return WanImageEncoderStateDictConverter()
+    
+    
+class WanImageEncoderStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        state_dict_ = {}
+        for name, param in state_dict.items():
+            if name.startswith("textual."):
+                continue
+            name = "model." + name
+            state_dict_[name] = param
+        return state_dict_
+

dkt/models/wan_video_motion_controller.py

@@ -0,0 +1,44 @@
+import torch
+import torch.nn as nn
+from .wan_video_dit import sinusoidal_embedding_1d
+
+
+
+class WanMotionControllerModel(torch.nn.Module):
+    def __init__(self, freq_dim=256, dim=1536):
+        super().__init__()
+        self.freq_dim = freq_dim
+        self.linear = nn.Sequential(
+            nn.Linear(freq_dim, dim),
+            nn.SiLU(),
+            nn.Linear(dim, dim),
+            nn.SiLU(),
+            nn.Linear(dim, dim * 6),
+        )
+
+    def forward(self, motion_bucket_id):
+        emb = sinusoidal_embedding_1d(self.freq_dim, motion_bucket_id * 10)
+        emb = self.linear(emb)
+        return emb
+
+    def init(self):
+        state_dict = self.linear[-1].state_dict()
+        state_dict = {i: state_dict[i] * 0 for i in state_dict}
+        self.linear[-1].load_state_dict(state_dict)
+
+    @staticmethod
+    def state_dict_converter():
+        return WanMotionControllerModelDictConverter()
+    
+    
+
+class WanMotionControllerModelDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        return state_dict
+

dkt/models/wan_video_text_encoder.py

@@ -0,0 +1,269 @@
+import math
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def fp16_clamp(x):
+    if x.dtype == torch.float16 and torch.isinf(x).any():
+        clamp = torch.finfo(x.dtype).max - 1000
+        x = torch.clamp(x, min=-clamp, max=clamp)
+    return x
+
+
+class GELU(nn.Module):
+
+    def forward(self, x):
+        return 0.5 * x * (1.0 + torch.tanh(
+            math.sqrt(2.0 / math.pi) * (x + 0.044715 * torch.pow(x, 3.0))))
+
+
+class T5LayerNorm(nn.Module):
+
+    def __init__(self, dim, eps=1e-6):
+        super(T5LayerNorm, self).__init__()
+        self.dim = dim
+        self.eps = eps
+        self.weight = nn.Parameter(torch.ones(dim))
+
+    def forward(self, x):
+        x = x * torch.rsqrt(x.float().pow(2).mean(dim=-1, keepdim=True) +
+                            self.eps)
+        if self.weight.dtype in [torch.float16, torch.bfloat16]:
+            x = x.type_as(self.weight)
+        return self.weight * x
+
+
+class T5Attention(nn.Module):
+
+    def __init__(self, dim, dim_attn, num_heads, dropout=0.1):
+        assert dim_attn % num_heads == 0
+        super(T5Attention, self).__init__()
+        self.dim = dim
+        self.dim_attn = dim_attn
+        self.num_heads = num_heads
+        self.head_dim = dim_attn // num_heads
+
+        # layers
+        self.q = nn.Linear(dim, dim_attn, bias=False)
+        self.k = nn.Linear(dim, dim_attn, bias=False)
+        self.v = nn.Linear(dim, dim_attn, bias=False)
+        self.o = nn.Linear(dim_attn, dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x, context=None, mask=None, pos_bias=None):
+        """
+        x:          [B, L1, C].
+        context:    [B, L2, C] or None.
+        mask:       [B, L2] or [B, L1, L2] or None.
+        """
+        # check inputs
+        context = x if context is None else context
+        b, n, c = x.size(0), self.num_heads, self.head_dim
+
+        # compute query, key, value
+        q = self.q(x).view(b, -1, n, c)
+        k = self.k(context).view(b, -1, n, c)
+        v = self.v(context).view(b, -1, n, c)
+
+        # attention bias
+        attn_bias = x.new_zeros(b, n, q.size(1), k.size(1))
+        if pos_bias is not None:
+            attn_bias += pos_bias
+        if mask is not None:
+            assert mask.ndim in [2, 3]
+            mask = mask.view(b, 1, 1,
+                             -1) if mask.ndim == 2 else mask.unsqueeze(1)
+            attn_bias.masked_fill_(mask == 0, torch.finfo(x.dtype).min)
+
+        # compute attention (T5 does not use scaling)
+        attn = torch.einsum('binc,bjnc->bnij', q, k) + attn_bias
+        attn = F.softmax(attn.float(), dim=-1).type_as(attn)
+        x = torch.einsum('bnij,bjnc->binc', attn, v)
+
+        # output
+        x = x.reshape(b, -1, n * c)
+        x = self.o(x)
+        x = self.dropout(x)
+        return x
+
+
+class T5FeedForward(nn.Module):
+
+    def __init__(self, dim, dim_ffn, dropout=0.1):
+        super(T5FeedForward, self).__init__()
+        self.dim = dim
+        self.dim_ffn = dim_ffn
+
+        # layers
+        self.gate = nn.Sequential(nn.Linear(dim, dim_ffn, bias=False), GELU())
+        self.fc1 = nn.Linear(dim, dim_ffn, bias=False)
+        self.fc2 = nn.Linear(dim_ffn, dim, bias=False)
+        self.dropout = nn.Dropout(dropout)
+
+    def forward(self, x):
+        x = self.fc1(x) * self.gate(x)
+        x = self.dropout(x)
+        x = self.fc2(x)
+        x = self.dropout(x)
+        return x
+
+
+class T5SelfAttention(nn.Module):
+
+    def __init__(self,
+                 dim,
+                 dim_attn,
+                 dim_ffn,
+                 num_heads,
+                 num_buckets,
+                 shared_pos=True,
+                 dropout=0.1):
+        super(T5SelfAttention, self).__init__()
+        self.dim = dim
+        self.dim_attn = dim_attn
+        self.dim_ffn = dim_ffn
+        self.num_heads = num_heads
+        self.num_buckets = num_buckets
+        self.shared_pos = shared_pos
+
+        # layers
+        self.norm1 = T5LayerNorm(dim)
+        self.attn = T5Attention(dim, dim_attn, num_heads, dropout)
+        self.norm2 = T5LayerNorm(dim)
+        self.ffn = T5FeedForward(dim, dim_ffn, dropout)
+        self.pos_embedding = None if shared_pos else T5RelativeEmbedding(
+            num_buckets, num_heads, bidirectional=True)
+
+    def forward(self, x, mask=None, pos_bias=None):
+        e = pos_bias if self.shared_pos else self.pos_embedding(
+            x.size(1), x.size(1))
+        x = fp16_clamp(x + self.attn(self.norm1(x), mask=mask, pos_bias=e))
+        x = fp16_clamp(x + self.ffn(self.norm2(x)))
+        return x
+
+
+class T5RelativeEmbedding(nn.Module):
+
+    def __init__(self, num_buckets, num_heads, bidirectional, max_dist=128):
+        super(T5RelativeEmbedding, self).__init__()
+        self.num_buckets = num_buckets
+        self.num_heads = num_heads
+        self.bidirectional = bidirectional
+        self.max_dist = max_dist
+
+        # layers
+        self.embedding = nn.Embedding(num_buckets, num_heads)
+
+    def forward(self, lq, lk):
+        device = self.embedding.weight.device
+        # rel_pos = torch.arange(lk).unsqueeze(0).to(device) - \
+        #     torch.arange(lq).unsqueeze(1).to(device)
+        rel_pos = torch.arange(lk, device=device).unsqueeze(0) - \
+            torch.arange(lq, device=device).unsqueeze(1)
+        rel_pos = self._relative_position_bucket(rel_pos)
+        rel_pos_embeds = self.embedding(rel_pos)
+        rel_pos_embeds = rel_pos_embeds.permute(2, 0, 1).unsqueeze(
+            0)  # [1, N, Lq, Lk]
+        return rel_pos_embeds.contiguous()
+
+    def _relative_position_bucket(self, rel_pos):
+        # preprocess
+        if self.bidirectional:
+            num_buckets = self.num_buckets // 2
+            rel_buckets = (rel_pos > 0).long() * num_buckets
+            rel_pos = torch.abs(rel_pos)
+        else:
+            num_buckets = self.num_buckets
+            rel_buckets = 0
+            rel_pos = -torch.min(rel_pos, torch.zeros_like(rel_pos))
+
+        # embeddings for small and large positions
+        max_exact = num_buckets // 2
+        rel_pos_large = max_exact + (torch.log(rel_pos.float() / max_exact) /
+                                     math.log(self.max_dist / max_exact) *
+                                     (num_buckets - max_exact)).long()
+        rel_pos_large = torch.min(
+            rel_pos_large, torch.full_like(rel_pos_large, num_buckets - 1))
+        rel_buckets += torch.where(rel_pos < max_exact, rel_pos, rel_pos_large)
+        return rel_buckets
+
+def init_weights(m):
+    if isinstance(m, T5LayerNorm):
+        nn.init.ones_(m.weight)
+    elif isinstance(m, T5FeedForward):
+        nn.init.normal_(m.gate[0].weight, std=m.dim**-0.5)
+        nn.init.normal_(m.fc1.weight, std=m.dim**-0.5)
+        nn.init.normal_(m.fc2.weight, std=m.dim_ffn**-0.5)
+    elif isinstance(m, T5Attention):
+        nn.init.normal_(m.q.weight, std=(m.dim * m.dim_attn)**-0.5)
+        nn.init.normal_(m.k.weight, std=m.dim**-0.5)
+        nn.init.normal_(m.v.weight, std=m.dim**-0.5)
+        nn.init.normal_(m.o.weight, std=(m.num_heads * m.dim_attn)**-0.5)
+    elif isinstance(m, T5RelativeEmbedding):
+        nn.init.normal_(
+            m.embedding.weight, std=(2 * m.num_buckets * m.num_heads)**-0.5)
+
+
+class WanTextEncoder(torch.nn.Module):
+
+    def __init__(self,
+                 vocab=256384,
+                 dim=4096,
+                 dim_attn=4096,
+                 dim_ffn=10240,
+                 num_heads=64,
+                 num_layers=24,
+                 num_buckets=32,
+                 shared_pos=False,
+                 dropout=0.1):
+        super(WanTextEncoder, self).__init__()
+        self.dim = dim
+        self.dim_attn = dim_attn
+        self.dim_ffn = dim_ffn
+        self.num_heads = num_heads
+        self.num_layers = num_layers
+        self.num_buckets = num_buckets
+        self.shared_pos = shared_pos
+
+        # layers
+        self.token_embedding = vocab if isinstance(vocab, nn.Embedding) \
+            else nn.Embedding(vocab, dim)
+        self.pos_embedding = T5RelativeEmbedding(
+            num_buckets, num_heads, bidirectional=True) if shared_pos else None
+        self.dropout = nn.Dropout(dropout)
+        self.blocks = nn.ModuleList([
+            T5SelfAttention(dim, dim_attn, dim_ffn, num_heads, num_buckets,
+                            shared_pos, dropout) for _ in range(num_layers)
+        ])
+        self.norm = T5LayerNorm(dim)
+
+        # initialize weights
+        self.apply(init_weights)
+
+    def forward(self, ids, mask=None):
+        x = self.token_embedding(ids)
+        x = self.dropout(x)
+        e = self.pos_embedding(x.size(1),
+                               x.size(1)) if self.shared_pos else None
+        for block in self.blocks:
+            x = block(x, mask, pos_bias=e)
+        x = self.norm(x)
+        x = self.dropout(x)
+        return x
+    
+    @staticmethod
+    def state_dict_converter():
+        return WanTextEncoderStateDictConverter()
+    
+    
+class WanTextEncoderStateDictConverter:
+    def __init__(self):
+        pass
+
+    def from_diffusers(self, state_dict):
+        return state_dict
+    
+    def from_civitai(self, state_dict):
+        return state_dict

dkt/models/wan_video_vace.py

@@ -0,0 +1,113 @@
+import torch
+from .wan_video_dit import DiTBlock
+from .utils import hash_state_dict_keys
+
+class VaceWanAttentionBlock(DiTBlock):
+    def __init__(self, has_image_input, dim, num_heads, ffn_dim, eps=1e-6, block_id=0):
+        super().__init__(has_image_input, dim, num_heads, ffn_dim, eps=eps)
+        self.block_id = block_id
+        if block_id == 0:
+            self.before_proj = torch.nn.Linear(self.dim, self.dim)
+        self.after_proj = torch.nn.Linear(self.dim, self.dim)
+
+    def forward(self, c, x, context, t_mod, freqs):
+        if self.block_id == 0:
+            c = self.before_proj(c) + x
+            all_c = []
+        else:
+            all_c = list(torch.unbind(c))
+            c = all_c.pop(-1)
+        c = super().forward(c, context, t_mod, freqs)
+        c_skip = self.after_proj(c)
+        all_c += [c_skip, c]
+        c = torch.stack(all_c)
+        return c
+
+
+class VaceWanModel(torch.nn.Module):
+    def __init__(
+        self,
+        vace_layers=(0, 2, 4, 6, 8, 10, 12, 14, 16, 18, 20, 22, 24, 26, 28),
+        vace_in_dim=96,
+        patch_size=(1, 2, 2),
+        has_image_input=False,
+        dim=1536,
+        num_heads=12,
+        ffn_dim=8960,
+        eps=1e-6,
+    ):
+        super().__init__()
+        self.vace_layers = vace_layers
+        self.vace_in_dim = vace_in_dim
+        self.vace_layers_mapping = {i: n for n, i in enumerate(self.vace_layers)}
+
+        # vace blocks
+        self.vace_blocks = torch.nn.ModuleList([
+            VaceWanAttentionBlock(has_image_input, dim, num_heads, ffn_dim, eps, block_id=i)
+            for i in self.vace_layers
+        ])
+
+        # vace patch embeddings
+        self.vace_patch_embedding = torch.nn.Conv3d(vace_in_dim, dim, kernel_size=patch_size, stride=patch_size)
+
+    def forward(
+        self, x, vace_context, context, t_mod, freqs,
+        use_gradient_checkpointing: bool = False,
+        use_gradient_checkpointing_offload: bool = False,
+    ):
+        c = [self.vace_patch_embedding(u.unsqueeze(0)) for u in vace_context]
+        c = [u.flatten(2).transpose(1, 2) for u in c]
+        c = torch.cat([
+            torch.cat([u, u.new_zeros(1, x.shape[1] - u.size(1), u.size(2))],
+                      dim=1) for u in c
+        ])
+        
+        def create_custom_forward(module):
+            def custom_forward(*inputs):
+                return module(*inputs)
+            return custom_forward
+        
+        for block in self.vace_blocks:
+            if use_gradient_checkpointing_offload:
+                with torch.autograd.graph.save_on_cpu():
+                    c = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        c, x, context, t_mod, freqs,
+                        use_reentrant=False,
+                    )
+            elif use_gradient_checkpointing:
+                c = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    c, x, context, t_mod, freqs,
+                    use_reentrant=False,
+                )
+            else:
+                c = block(c, x, context, t_mod, freqs)
+        hints = torch.unbind(c)[:-1]
+        return hints
+    
+    @staticmethod
+    def state_dict_converter():
+        return VaceWanModelDictConverter()
+    
+    
+class VaceWanModelDictConverter:
+    def __init__(self):
+        pass
+    
+    def from_civitai(self, state_dict):
+        state_dict_ = {name: param for name, param in state_dict.items() if name.startswith("vace")}
+        if hash_state_dict_keys(state_dict_) == '3b2726384e4f64837bdf216eea3f310d': # vace 14B
+            config = {
+                "vace_layers": (0, 5, 10, 15, 20, 25, 30, 35),
+                "vace_in_dim": 96,
+                "patch_size": (1, 2, 2),
+                "has_image_input": False,
+                "dim": 5120,
+                "num_heads": 40,
+                "ffn_dim": 13824,
+                "eps": 1e-06,                
+            }
+        else:
+            config = {}
+        return state_dict_, config

dkt/models/wan_video_vae.py

@@ -0,0 +1,1376 @@
+from einops import rearrange, repeat
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from tqdm import tqdm
+
+CACHE_T = 2
+
+
+def check_is_instance(model, module_class):
+    if isinstance(model, module_class):
+        return True
+    if hasattr(model, "module") and isinstance(model.module, module_class):
+        return True
+    return False
+
+
+def block_causal_mask(x, block_size):
+    # params
+    b, n, s, _, device = *x.size(), x.device
+    assert s % block_size == 0
+    num_blocks = s // block_size
+
+    # build mask
+    mask = torch.zeros(b, n, s, s, dtype=torch.bool, device=device)
+    for i in range(num_blocks):
+        mask[:, :,
+             i * block_size:(i + 1) * block_size, :(i + 1) * block_size] = 1
+    return mask
+
+
+class CausalConv3d(nn.Conv3d):
+    """
+    Causal 3d convolusion.
+    """
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+        self._padding = (self.padding[2], self.padding[2], self.padding[1],
+                         self.padding[1], 2 * self.padding[0], 0)
+        self.padding = (0, 0, 0)
+
+    def forward(self, x, cache_x=None):
+        padding = list(self._padding)
+        if cache_x is not None and self._padding[4] > 0:
+            cache_x = cache_x.to(x.device)
+            x = torch.cat([cache_x, x], dim=2)
+            padding[4] -= cache_x.shape[2]
+        x = F.pad(x, padding)
+
+        return super().forward(x)
+
+
+class RMS_norm(nn.Module):
+
+    def __init__(self, dim, channel_first=True, images=True, bias=False):
+        super().__init__()
+        broadcastable_dims = (1, 1, 1) if not images else (1, 1)
+        shape = (dim, *broadcastable_dims) if channel_first else (dim,)
+
+        self.channel_first = channel_first
+        self.scale = dim**0.5
+        self.gamma = nn.Parameter(torch.ones(shape))
+        self.bias = nn.Parameter(torch.zeros(shape)) if bias else 0.
+
+    def forward(self, x):
+        return F.normalize(
+            x, dim=(1 if self.channel_first else
+                    -1)) * self.scale * self.gamma + self.bias
+
+
+class Upsample(nn.Upsample):
+
+    def forward(self, x):
+        """
+        Fix bfloat16 support for nearest neighbor interpolation.
+        """
+        return super().forward(x.float()).type_as(x)
+
+
+class Resample(nn.Module):
+
+    def __init__(self, dim, mode):
+        assert mode in ('none', 'upsample2d', 'upsample3d', 'downsample2d',
+                        'downsample3d')
+        super().__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == 'upsample2d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+        elif mode == 'upsample3d':
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2., 2.), mode='nearest-exact'),
+                nn.Conv2d(dim, dim // 2, 3, padding=1))
+            self.time_conv = CausalConv3d(dim,
+                                          dim * 2, (3, 1, 1),
+                                          padding=(1, 0, 0))
+
+        elif mode == 'downsample2d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+        elif mode == 'downsample3d':
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)),
+                nn.Conv2d(dim, dim, 3, stride=(2, 2)))
+            self.time_conv = CausalConv3d(dim,
+                                          dim, (3, 1, 1),
+                                          stride=(2, 1, 1),
+                                          padding=(0, 0, 0))
+
+        else:
+            self.resample = nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        b, c, t, h, w = x.size()
+        if self.mode == 'upsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = 'Rep'
+                    feat_idx[0] += 1
+                else:
+
+                    cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] != 'Rep':
+                        # cache last frame of last two chunk
+                        cache_x = torch.cat([
+                            feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                                cache_x.device), cache_x
+                        ],
+                                            dim=2)
+                    if cache_x.shape[2] < 2 and feat_cache[
+                            idx] is not None and feat_cache[idx] == 'Rep':
+                        cache_x = torch.cat([
+                            torch.zeros_like(cache_x).to(cache_x.device),
+                            cache_x
+                        ],
+                                            dim=2)
+                    if feat_cache[idx] == 'Rep':
+                        x = self.time_conv(x)
+                    else:
+                        x = self.time_conv(x, feat_cache[idx])
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+
+                    x = x.reshape(b, 2, c, t, h, w)
+                    x = torch.stack((x[:, 0, :, :, :, :], x[:, 1, :, :, :, :]),
+                                    3)
+                    x = x.reshape(b, c, t * 2, h, w)
+        t = x.shape[2]
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.resample(x)
+        x = rearrange(x, '(b t) c h w -> b c t h w', t=t)
+
+        if self.mode == 'downsample3d':
+            if feat_cache is not None:
+                idx = feat_idx[0]
+                if feat_cache[idx] is None:
+                    feat_cache[idx] = x.clone()
+                    feat_idx[0] += 1
+                else:
+                    cache_x = x[:, :, -1:, :, :].clone()
+                    x = self.time_conv(
+                        torch.cat([feat_cache[idx][:, :, -1:, :, :], x], 2))
+                    feat_cache[idx] = cache_x
+                    feat_idx[0] += 1
+        return x
+
+    def init_weight(self, conv):
+        conv_weight = conv.weight
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        one_matrix = torch.eye(c1, c2)
+        init_matrix = one_matrix
+        nn.init.zeros_(conv_weight)
+        conv_weight.data[:, :, 1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+    def init_weight2(self, conv):
+        conv_weight = conv.weight.data
+        nn.init.zeros_(conv_weight)
+        c1, c2, t, h, w = conv_weight.size()
+        init_matrix = torch.eye(c1 // 2, c2)
+        conv_weight[:c1 // 2, :, -1, 0, 0] = init_matrix
+        conv_weight[c1 // 2:, :, -1, 0, 0] = init_matrix
+        conv.weight.data.copy_(conv_weight)
+        nn.init.zeros_(conv.bias.data)
+
+
+
+def patchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(x, "b c (h q) (w r) -> b (c r q) h w", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(x,
+                      "b c f (h q) (w r) -> b (c r q) f h w",
+                      q=patch_size,
+                      r=patch_size)
+    else:
+        raise ValueError(f"Invalid input shape: {x.shape}")
+    return x
+
+
+def unpatchify(x, patch_size):
+    if patch_size == 1:
+        return x
+    if x.dim() == 4:
+        x = rearrange(x, "b (c r q) h w -> b c (h q) (w r)", q=patch_size, r=patch_size)
+    elif x.dim() == 5:
+        x = rearrange(x,
+                      "b (c r q) f h w -> b c f (h q) (w r)",
+                      q=patch_size,
+                      r=patch_size)
+    return x
+
+
+class Resample38(Resample):
+
+    def __init__(self, dim, mode):
+        assert mode in (
+            "none",
+            "upsample2d",
+            "upsample3d",
+            "downsample2d",
+            "downsample3d",
+        )
+        super(Resample, self).__init__()
+        self.dim = dim
+        self.mode = mode
+
+        # layers
+        if mode == "upsample2d":
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim, 3, padding=1),
+            )
+        elif mode == "upsample3d":
+            self.resample = nn.Sequential(
+                Upsample(scale_factor=(2.0, 2.0), mode="nearest-exact"),
+                nn.Conv2d(dim, dim, 3, padding=1),
+            )
+            self.time_conv = CausalConv3d(dim, dim * 2, (3, 1, 1), padding=(1, 0, 0))
+        elif mode == "downsample2d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2))
+            )
+        elif mode == "downsample3d":
+            self.resample = nn.Sequential(
+                nn.ZeroPad2d((0, 1, 0, 1)), nn.Conv2d(dim, dim, 3, stride=(2, 2))
+            )
+            self.time_conv = CausalConv3d(
+                dim, dim, (3, 1, 1), stride=(2, 1, 1), padding=(0, 0, 0)
+            )
+        else:
+            self.resample = nn.Identity()
+
+class ResidualBlock(nn.Module):
+
+    def __init__(self, in_dim, out_dim, dropout=0.0):
+        super().__init__()
+        self.in_dim = in_dim
+        self.out_dim = out_dim
+
+        # layers
+        self.residual = nn.Sequential(
+            RMS_norm(in_dim, images=False), nn.SiLU(),
+            CausalConv3d(in_dim, out_dim, 3, padding=1),
+            RMS_norm(out_dim, images=False), nn.SiLU(), nn.Dropout(dropout),
+            CausalConv3d(out_dim, out_dim, 3, padding=1))
+        self.shortcut = CausalConv3d(in_dim, out_dim, 1) \
+            if in_dim != out_dim else nn.Identity()
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        h = self.shortcut(x)
+        for layer in self.residual:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x + h
+
+
+class AttentionBlock(nn.Module):
+    """
+    Causal self-attention with a single head.
+    """
+
+    def __init__(self, dim):
+        super().__init__()
+        self.dim = dim
+
+        # layers
+        self.norm = RMS_norm(dim)
+        self.to_qkv = nn.Conv2d(dim, dim * 3, 1)
+        self.proj = nn.Conv2d(dim, dim, 1)
+
+        # zero out the last layer params
+        nn.init.zeros_(self.proj.weight)
+
+    def forward(self, x):
+        identity = x
+        b, c, t, h, w = x.size()
+        x = rearrange(x, 'b c t h w -> (b t) c h w')
+        x = self.norm(x)
+        # compute query, key, value
+        q, k, v = self.to_qkv(x).reshape(b * t, 1, c * 3, -1).permute(
+            0, 1, 3, 2).contiguous().chunk(3, dim=-1)
+
+        # apply attention
+        x = F.scaled_dot_product_attention(
+            q,
+            k,
+            v,
+            #attn_mask=block_causal_mask(q, block_size=h * w)
+        )
+        x = x.squeeze(1).permute(0, 2, 1).reshape(b * t, c, h, w)
+
+        # output
+        x = self.proj(x)
+        x = rearrange(x, '(b t) c h w-> b c t h w', t=t)
+        return x + identity
+
+
+class AvgDown3D(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert in_channels * self.factor % out_channels == 0
+        self.group_size = in_channels * self.factor // out_channels
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        pad_t = (self.factor_t - x.shape[2] % self.factor_t) % self.factor_t
+        pad = (0, 0, 0, 0, pad_t, 0)
+        x = F.pad(x, pad)
+        B, C, T, H, W = x.shape
+        x = x.view(
+            B,
+            C,
+            T // self.factor_t,
+            self.factor_t,
+            H // self.factor_s,
+            self.factor_s,
+            W // self.factor_s,
+            self.factor_s,
+        )
+        x = x.permute(0, 1, 3, 5, 7, 2, 4, 6).contiguous()
+        x = x.view(
+            B,
+            C * self.factor,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.view(
+            B,
+            self.out_channels,
+            self.group_size,
+            T // self.factor_t,
+            H // self.factor_s,
+            W // self.factor_s,
+        )
+        x = x.mean(dim=2)
+        return x
+
+
+class DupUp3D(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        factor_t,
+        factor_s=1,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        self.factor_t = factor_t
+        self.factor_s = factor_s
+        self.factor = self.factor_t * self.factor_s * self.factor_s
+
+        assert out_channels * self.factor % in_channels == 0
+        self.repeats = out_channels * self.factor // in_channels
+
+    def forward(self, x: torch.Tensor, first_chunk=False) -> torch.Tensor:
+        x = x.repeat_interleave(self.repeats, dim=1)
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            self.factor_t,
+            self.factor_s,
+            self.factor_s,
+            x.size(2),
+            x.size(3),
+            x.size(4),
+        )
+        x = x.permute(0, 1, 5, 2, 6, 3, 7, 4).contiguous()
+        x = x.view(
+            x.size(0),
+            self.out_channels,
+            x.size(2) * self.factor_t,
+            x.size(4) * self.factor_s,
+            x.size(6) * self.factor_s,
+        )
+        if first_chunk:
+            x = x[:, :, self.factor_t - 1 :, :, :]
+        return x
+
+
+class Down_ResidualBlock(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, dropout, mult, temperal_downsample=False, down_flag=False
+    ):
+        super().__init__()
+
+        # Shortcut path with downsample
+        self.avg_shortcut = AvgDown3D(
+            in_dim,
+            out_dim,
+            factor_t=2 if temperal_downsample else 1,
+            factor_s=2 if down_flag else 1,
+        )
+
+        # Main path with residual blocks and downsample
+        downsamples = []
+        for _ in range(mult):
+            downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final downsample block
+        if down_flag:
+            mode = "downsample3d" if temperal_downsample else "downsample2d"
+            downsamples.append(Resample38(out_dim, mode=mode))
+
+        self.downsamples = nn.Sequential(*downsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        x_copy = x.clone()
+        for module in self.downsamples:
+            x = module(x, feat_cache, feat_idx)
+
+        return x + self.avg_shortcut(x_copy)
+
+
+class Up_ResidualBlock(nn.Module):
+    def __init__(
+        self, in_dim, out_dim, dropout, mult, temperal_upsample=False, up_flag=False
+    ):
+        super().__init__()
+        # Shortcut path with upsample
+        if up_flag:
+            self.avg_shortcut = DupUp3D(
+                in_dim,
+                out_dim,
+                factor_t=2 if temperal_upsample else 1,
+                factor_s=2 if up_flag else 1,
+            )
+        else:
+            self.avg_shortcut = None
+
+        # Main path with residual blocks and upsample
+        upsamples = []
+        for _ in range(mult):
+            upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+            in_dim = out_dim
+
+        # Add the final upsample block
+        if up_flag:
+            mode = "upsample3d" if temperal_upsample else "upsample2d"
+            upsamples.append(Resample38(out_dim, mode=mode))
+
+        self.upsamples = nn.Sequential(*upsamples)
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        x_main = x.clone()
+        for module in self.upsamples:
+            x_main = module(x_main, feat_cache, feat_idx)
+        if self.avg_shortcut is not None:
+            x_shortcut = self.avg_shortcut(x, first_chunk)
+            return x_main + x_shortcut
+        else:
+            return x_main
+
+
+class Encoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[True, True, False],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(3, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            for _ in range(num_res_blocks):
+                downsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    downsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # downsample block
+            if i != len(dim_mult) - 1:
+                mode = 'downsample3d' if temperal_downsample[
+                    i] else 'downsample2d'
+                downsamples.append(Resample(out_dim, mode=mode))
+                scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(out_dim, out_dim, dropout),
+                                    AttentionBlock(out_dim),
+                                    ResidualBlock(out_dim, out_dim, dropout))
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, z_dim, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+class Encoder3d_38(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+
+        # dimensions
+        dims = [dim * u for u in [1] + dim_mult]
+        scale = 1.0
+
+        # init block
+        self.conv1 = CausalConv3d(12, dims[0], 3, padding=1)
+
+        # downsample blocks
+        downsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_down_flag = (
+                temperal_downsample[i] if i < len(temperal_downsample) else False
+            )
+            downsamples.append(
+                Down_ResidualBlock(
+                    in_dim=in_dim,
+                    out_dim=out_dim,
+                    dropout=dropout,
+                    mult=num_res_blocks,
+                    temperal_downsample=t_down_flag,
+                    down_flag=i != len(dim_mult) - 1,
+                )
+            )
+            scale /= 2.0
+        self.downsamples = nn.Sequential(*downsamples)
+
+        # middle blocks
+        self.middle = nn.Sequential(
+            ResidualBlock(out_dim, out_dim, dropout),
+            AttentionBlock(out_dim),
+            ResidualBlock(out_dim, out_dim, dropout),
+        )
+
+        # # output blocks
+        self.head = nn.Sequential(
+            RMS_norm(out_dim, images=False),
+            nn.SiLU(),
+            CausalConv3d(out_dim, z_dim, 3, padding=1),
+        )
+
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## downsamples
+        for layer in self.downsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## middle
+        for layer in self.middle:
+            if isinstance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if isinstance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :]
+                            .unsqueeze(2)
+                            .to(cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+
+        return x
+
+
+class Decoder3d(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2**(len(dim_mult) - 2)
+
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            # residual (+attention) blocks
+            if i == 1 or i == 2 or i == 3:
+                in_dim = in_dim // 2
+            for _ in range(num_res_blocks + 1):
+                upsamples.append(ResidualBlock(in_dim, out_dim, dropout))
+                if scale in attn_scales:
+                    upsamples.append(AttentionBlock(out_dim))
+                in_dim = out_dim
+
+            # upsample block
+            if i != len(dim_mult) - 1:
+                mode = 'upsample3d' if temperal_upsample[i] else 'upsample2d'
+                upsamples.append(Resample(out_dim, mode=mode))
+                scale *= 2.0
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 3, 3, padding=1))
+
+    def forward(self, x, feat_cache=None, feat_idx=[0]):
+        ## conv1
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                # cache last frame of last two chunk
+                cache_x = torch.cat([
+                    feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                        cache_x.device), cache_x
+                ],
+                                    dim=2)
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        ## middle
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    # cache last frame of last two chunk
+                    cache_x = torch.cat([
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(
+                            cache_x.device), cache_x
+                    ],
+                                        dim=2)
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+
+class Decoder3d_38(nn.Module):
+
+    def __init__(self,
+                 dim=128,
+                 z_dim=4,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_upsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_upsample = temperal_upsample
+
+        # dimensions
+        dims = [dim * u for u in [dim_mult[-1]] + dim_mult[::-1]]
+        scale = 1.0 / 2 ** (len(dim_mult) - 2)
+        # init block
+        self.conv1 = CausalConv3d(z_dim, dims[0], 3, padding=1)
+
+        # middle blocks
+        self.middle = nn.Sequential(ResidualBlock(dims[0], dims[0], dropout),
+                                    AttentionBlock(dims[0]),
+                                    ResidualBlock(dims[0], dims[0], dropout))
+
+        # upsample blocks
+        upsamples = []
+        for i, (in_dim, out_dim) in enumerate(zip(dims[:-1], dims[1:])):
+            t_up_flag = temperal_upsample[i] if i < len(temperal_upsample) else False
+            upsamples.append(
+                Up_ResidualBlock(in_dim=in_dim,
+                                 out_dim=out_dim,
+                                 dropout=dropout,
+                                 mult=num_res_blocks + 1,
+                                 temperal_upsample=t_up_flag,
+                                 up_flag=i != len(dim_mult) - 1))
+        self.upsamples = nn.Sequential(*upsamples)
+
+        # output blocks
+        self.head = nn.Sequential(RMS_norm(out_dim, images=False), nn.SiLU(),
+                                  CausalConv3d(out_dim, 12, 3, padding=1))
+
+
+    def forward(self, x, feat_cache=None, feat_idx=[0], first_chunk=False):
+        if feat_cache is not None:
+            idx = feat_idx[0]
+            cache_x = x[:, :, -CACHE_T:, :, :].clone()
+            if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                cache_x = torch.cat(
+                    [
+                        feat_cache[idx][:, :, -1, :, :].unsqueeze(2).to(cache_x.device),
+                        cache_x,
+                    ],
+                    dim=2,
+                )
+            x = self.conv1(x, feat_cache[idx])
+            feat_cache[idx] = cache_x
+            feat_idx[0] += 1
+        else:
+            x = self.conv1(x)
+
+        for layer in self.middle:
+            if check_is_instance(layer, ResidualBlock) and feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx)
+            else:
+                x = layer(x)
+
+        ## upsamples
+        for layer in self.upsamples:
+            if feat_cache is not None:
+                x = layer(x, feat_cache, feat_idx, first_chunk)
+            else:
+                x = layer(x)
+
+        ## head
+        for layer in self.head:
+            if check_is_instance(layer, CausalConv3d) and feat_cache is not None:
+                idx = feat_idx[0]
+                cache_x = x[:, :, -CACHE_T:, :, :].clone()
+                if cache_x.shape[2] < 2 and feat_cache[idx] is not None:
+                    cache_x = torch.cat(
+                        [
+                            feat_cache[idx][:, :, -1, :, :]
+                            .unsqueeze(2)
+                            .to(cache_x.device),
+                            cache_x,
+                        ],
+                        dim=2,
+                    )
+                x = layer(x, feat_cache[idx])
+                feat_cache[idx] = cache_x
+                feat_idx[0] += 1
+            else:
+                x = layer(x)
+        return x
+
+
+def count_conv3d(model):
+    count = 0
+    for m in model.modules():
+        if isinstance(m, CausalConv3d):
+            count += 1
+    return count
+
+
+class VideoVAE_(nn.Module):
+
+    def __init__(self,
+                 dim=96,
+                 z_dim=16,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super().__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d(dim, z_dim, dim_mult, num_res_blocks,
+                                 attn_scales, self.temperal_upsample, dropout)
+
+    def forward(self, x):
+        mu, log_var = self.encode(x)
+        z = self.reparameterize(mu, log_var)
+        x_recon = self.decode(z)
+        return x_recon, mu, log_var
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        ## cache
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        return mu
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        # z: [b,c,t,h,w]
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2) # may add tensor offload
+        return out
+
+    def reparameterize(self, mu, log_var):
+        std = torch.exp(0.5 * log_var)
+        eps = torch.randn_like(std)
+        return eps * std + mu
+
+    def sample(self, imgs, deterministic=False):
+        mu, log_var = self.encode(imgs)
+        if deterministic:
+            return mu
+        std = torch.exp(0.5 * log_var.clamp(-30.0, 20.0))
+        return mu + std * torch.randn_like(std)
+
+    def clear_cache(self):
+        self._conv_num = count_conv3d(self.decoder)
+        self._conv_idx = [0]
+        self._feat_map = [None] * self._conv_num
+        # cache encode
+        self._enc_conv_num = count_conv3d(self.encoder)
+        self._enc_conv_idx = [0]
+        self._enc_feat_map = [None] * self._enc_conv_num
+
+
+class WanVideoVAE(nn.Module):
+
+    def __init__(self, z_dim=16):
+        super().__init__()
+
+        mean = [
+            -0.7571, -0.7089, -0.9113, 0.1075, -0.1745, 0.9653, -0.1517, 1.5508,
+            0.4134, -0.0715, 0.5517, -0.3632, -0.1922, -0.9497, 0.2503, -0.2921
+        ]
+        std = [
+            2.8184, 1.4541, 2.3275, 2.6558, 1.2196, 1.7708, 2.6052, 2.0743,
+            3.2687, 2.1526, 2.8652, 1.5579, 1.6382, 1.1253, 2.8251, 1.9160
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE_(z_dim=z_dim).eval().requires_grad_(False)
+        self.upsampling_factor = 8
+        self.z_dim = z_dim
+
+
+    def build_1d_mask(self, length, left_bound, right_bound, border_width):
+        x = torch.ones((length,))
+        if not left_bound:
+            x[:border_width] = (torch.arange(border_width) + 1) / border_width
+        if not right_bound:
+            x[-border_width:] = torch.flip((torch.arange(border_width) + 1) / border_width, dims=(0,))
+        return x
+
+
+    def build_mask(self, data, is_bound, border_width):
+        _, _, _, H, W = data.shape
+        h = self.build_1d_mask(H, is_bound[0], is_bound[1], border_width[0])
+        w = self.build_1d_mask(W, is_bound[2], is_bound[3], border_width[1])
+
+        h = repeat(h, "H -> H W", H=H, W=W)
+        w = repeat(w, "W -> H W", H=H, W=W)
+
+        mask = torch.stack([h, w]).min(dim=0).values
+        mask = rearrange(mask, "H W -> 1 1 1 H W")
+        return mask
+
+
+    def tiled_decode(self, hidden_states, device, tile_size, tile_stride):
+        _, _, T, H, W = hidden_states.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H): continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W): continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = T * 4 - 3
+        weight = torch.zeros((1, 1, out_T, H * self.upsampling_factor, W * self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+        values = torch.zeros((1, 3, out_T, H * self.upsampling_factor, W * self.upsampling_factor), dtype=hidden_states.dtype, device=data_device)
+
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE decoding"):
+            hidden_states_batch = hidden_states[:, :, :, h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.decode(hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h==0, h_>=H, w==0, w_>=W),
+                border_width=((size_h - stride_h) * self.upsampling_factor, (size_w - stride_w) * self.upsampling_factor)
+            ).to(dtype=hidden_states.dtype, device=data_device)
+
+            target_h = h * self.upsampling_factor
+            target_w = w * self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        values = values.clamp_(-1, 1)
+        return values
+
+
+    def tiled_encode(self, video, device, tile_size, tile_stride):
+        _, _, T, H, W = video.shape
+        size_h, size_w = tile_size
+        stride_h, stride_w = tile_stride
+
+        # Split tasks
+        tasks = []
+        for h in range(0, H, stride_h):
+            if (h-stride_h >= 0 and h-stride_h+size_h >= H): continue
+            for w in range(0, W, stride_w):
+                if (w-stride_w >= 0 and w-stride_w+size_w >= W): continue
+                h_, w_ = h + size_h, w + size_w
+                tasks.append((h, h_, w, w_))
+
+        data_device = "cpu"
+        computation_device = device
+
+        out_T = (T + 3) // 4
+        weight = torch.zeros((1, 1, out_T, H // self.upsampling_factor, W // self.upsampling_factor), dtype=video.dtype, device=data_device)
+        values = torch.zeros((1, self.z_dim, out_T, H // self.upsampling_factor, W // self.upsampling_factor), dtype=video.dtype, device=data_device)
+
+        for h, h_, w, w_ in tqdm(tasks, desc="VAE encoding"):
+            hidden_states_batch = video[:, :, :, h:h_, w:w_].to(computation_device)
+            hidden_states_batch = self.model.encode(hidden_states_batch, self.scale).to(data_device)
+
+            mask = self.build_mask(
+                hidden_states_batch,
+                is_bound=(h==0, h_>=H, w==0, w_>=W),
+                border_width=((size_h - stride_h) // self.upsampling_factor, (size_w - stride_w) // self.upsampling_factor)
+            ).to(dtype=video.dtype, device=data_device)
+
+            target_h = h // self.upsampling_factor
+            target_w = w // self.upsampling_factor
+            values[
+                :,
+                :,
+                :,
+                target_h:target_h + hidden_states_batch.shape[3],
+                target_w:target_w + hidden_states_batch.shape[4],
+            ] += hidden_states_batch * mask
+            weight[
+                :,
+                :,
+                :,
+                target_h: target_h + hidden_states_batch.shape[3],
+                target_w: target_w + hidden_states_batch.shape[4],
+            ] += mask
+        values = values / weight
+        return values
+
+
+    def single_encode(self, video, device):
+        video = video.to(device)
+        x = self.model.encode(video, self.scale)
+        return x
+
+
+    def single_decode(self, hidden_state, device):
+        hidden_state = hidden_state.to(device)
+        video = self.model.decode(hidden_state, self.scale)
+        return video.clamp_(-1, 1)
+
+
+    def encode(self, videos, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+
+        videos = [video.to("cpu") for video in videos]
+        hidden_states = []
+        for video in videos:
+            video = video.unsqueeze(0)
+            if tiled:
+                tile_size = (tile_size[0] * self.upsampling_factor, tile_size[1] * self.upsampling_factor)
+                tile_stride = (tile_stride[0] * self.upsampling_factor, tile_stride[1] * self.upsampling_factor)
+                hidden_state = self.tiled_encode(video, device, tile_size, tile_stride)
+            else:
+                hidden_state = self.single_encode(video, device)
+            hidden_state = hidden_state.squeeze(0)
+            hidden_states.append(hidden_state)
+        hidden_states = torch.stack(hidden_states)
+        return hidden_states
+
+
+    def decode(self, hidden_states, device, tiled=False, tile_size=(34, 34), tile_stride=(18, 16)):
+        if tiled:
+            video = self.tiled_decode(hidden_states, device, tile_size, tile_stride)
+        else:
+            video = self.single_decode(hidden_states, device)
+        return video
+
+
+    @staticmethod
+    def state_dict_converter():
+        return WanVideoVAEStateDictConverter()
+
+
+class WanVideoVAEStateDictConverter:
+
+    def __init__(self):
+        pass
+
+    def from_civitai(self, state_dict):
+        state_dict_ = {}
+        if 'model_state' in state_dict:
+            state_dict = state_dict['model_state']
+        for name in state_dict:
+            state_dict_['model.' + name] = state_dict[name]
+        return state_dict_
+
+
+class VideoVAE38_(VideoVAE_):
+
+    def __init__(self,
+                 dim=160,
+                 z_dim=48,
+                 dec_dim=256,
+                 dim_mult=[1, 2, 4, 4],
+                 num_res_blocks=2,
+                 attn_scales=[],
+                 temperal_downsample=[False, True, True],
+                 dropout=0.0):
+        super(VideoVAE_, self).__init__()
+        self.dim = dim
+        self.z_dim = z_dim
+        self.dim_mult = dim_mult
+        self.num_res_blocks = num_res_blocks
+        self.attn_scales = attn_scales
+        self.temperal_downsample = temperal_downsample
+        self.temperal_upsample = temperal_downsample[::-1]
+
+        # modules
+        self.encoder = Encoder3d_38(dim, z_dim * 2, dim_mult, num_res_blocks,
+                                    attn_scales, self.temperal_downsample, dropout)
+        self.conv1 = CausalConv3d(z_dim * 2, z_dim * 2, 1)
+        self.conv2 = CausalConv3d(z_dim, z_dim, 1)
+        self.decoder = Decoder3d_38(dec_dim, z_dim, dim_mult, num_res_blocks,
+                                    attn_scales, self.temperal_upsample, dropout)
+
+
+    def encode(self, x, scale):
+        self.clear_cache()
+        x = patchify(x, patch_size=2)
+        t = x.shape[2]
+        iter_ = 1 + (t - 1) // 4
+        for i in range(iter_):
+            self._enc_conv_idx = [0]
+            if i == 0:
+                out = self.encoder(x[:, :, :1, :, :],
+                                   feat_cache=self._enc_feat_map,
+                                   feat_idx=self._enc_conv_idx)
+            else:
+                out_ = self.encoder(x[:, :, 1 + 4 * (i - 1):1 + 4 * i, :, :],
+                                    feat_cache=self._enc_feat_map,
+                                    feat_idx=self._enc_conv_idx)
+                out = torch.cat([out, out_], 2)
+        mu, log_var = self.conv1(out).chunk(2, dim=1)
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=mu.dtype, device=mu.device) for s in scale]
+            mu = (mu - scale[0].view(1, self.z_dim, 1, 1, 1)) * scale[1].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=mu.dtype, device=mu.device)
+            mu = (mu - scale[0]) * scale[1]
+        self.clear_cache()
+        return mu
+
+
+    def decode(self, z, scale):
+        self.clear_cache()
+        if isinstance(scale[0], torch.Tensor):
+            scale = [s.to(dtype=z.dtype, device=z.device) for s in scale]
+            z = z / scale[1].view(1, self.z_dim, 1, 1, 1) + scale[0].view(
+                1, self.z_dim, 1, 1, 1)
+        else:
+            scale = scale.to(dtype=z.dtype, device=z.device)
+            z = z / scale[1] + scale[0]
+        iter_ = z.shape[2]
+        x = self.conv2(z)
+        for i in range(iter_):
+            self._conv_idx = [0]
+            if i == 0:
+                out = self.decoder(x[:, :, i:i + 1, :, :],
+                                   feat_cache=self._feat_map,
+                                   feat_idx=self._conv_idx,
+                                   first_chunk=True)
+            else:
+                out_ = self.decoder(x[:, :, i:i + 1, :, :],
+                                    feat_cache=self._feat_map,
+                                    feat_idx=self._conv_idx)
+                out = torch.cat([out, out_], 2)
+        out = unpatchify(out, patch_size=2)
+        self.clear_cache()
+        return out
+
+
+class WanVideoVAE38(WanVideoVAE):
+
+    def __init__(self, z_dim=48, dim=160):
+        super(WanVideoVAE, self).__init__()
+
+        mean = [
+            -0.2289, -0.0052, -0.1323, -0.2339, -0.2799,  0.0174,  0.1838,  0.1557,
+            -0.1382,  0.0542,  0.2813,  0.0891,  0.1570, -0.0098,  0.0375, -0.1825,
+            -0.2246, -0.1207, -0.0698,  0.5109,  0.2665, -0.2108, -0.2158,  0.2502,
+            -0.2055, -0.0322,  0.1109,  0.1567, -0.0729,  0.0899, -0.2799, -0.1230,
+            -0.0313, -0.1649,  0.0117,  0.0723, -0.2839, -0.2083, -0.0520,  0.3748,
+            0.0152,  0.1957,  0.1433, -0.2944,  0.3573, -0.0548, -0.1681, -0.0667
+        ]
+        std = [
+            0.4765, 1.0364, 0.4514, 1.1677, 0.5313, 0.4990, 0.4818, 0.5013,
+            0.8158, 1.0344, 0.5894, 1.0901, 0.6885, 0.6165, 0.8454, 0.4978,
+            0.5759, 0.3523, 0.7135, 0.6804, 0.5833, 1.4146, 0.8986, 0.5659,
+            0.7069, 0.5338, 0.4889, 0.4917, 0.4069, 0.4999, 0.6866, 0.4093,
+            0.5709, 0.6065, 0.6415, 0.4944, 0.5726, 1.2042, 0.5458, 1.6887,
+            0.3971, 1.0600, 0.3943, 0.5537, 0.5444, 0.4089, 0.7468, 0.7744
+        ]
+        self.mean = torch.tensor(mean)
+        self.std = torch.tensor(std)
+        self.scale = [self.mean, 1.0 / self.std]
+
+        # init model
+        self.model = VideoVAE38_(z_dim=z_dim, dim=dim).eval().requires_grad_(False)
+        self.upsampling_factor = 16
+        self.z_dim = z_dim

dkt/pipelines/wan_video_new.py

@@ -0,0 +1,1486 @@
+import torch, warnings, glob, os, types
+import numpy as np
+from PIL import Image
+from einops import repeat, reduce
+from typing import Optional, Union
+from dataclasses import dataclass
+from modelscope import snapshot_download as ms_snap_download
+from huggingface_hub import snapshot_download as hf_snap_download
+
+from einops import rearrange
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+from typing import Optional
+from typing_extensions import Literal
+
+from ..utils import BasePipeline, ModelConfig, PipelineUnit, PipelineUnitRunner
+from ..models import ModelManager, load_state_dict
+from ..models.wan_video_dit import WanModel, RMSNorm, sinusoidal_embedding_1d
+from ..models.wan_video_text_encoder import WanTextEncoder, T5RelativeEmbedding, T5LayerNorm
+from ..models.wan_video_vae import WanVideoVAE, RMS_norm, CausalConv3d, Upsample
+from ..models.wan_video_image_encoder import WanImageEncoder
+from ..models.wan_video_vace import VaceWanModel
+from ..models.wan_video_motion_controller import WanMotionControllerModel
+from ..schedulers.flow_match import FlowMatchScheduler
+from ..prompters import WanPrompter
+from ..vram_management import enable_vram_management, AutoWrappedModule, AutoWrappedLinear, WanAutoCastLayerNorm
+from ..lora import GeneralLoRALoader
+
+from loguru import logger
+
+
+
+class BasePipeline(torch.nn.Module):
+
+    def __init__(
+        self,
+        device="cuda", torch_dtype=torch.float16,
+        height_division_factor=64, width_division_factor=64,
+        time_division_factor=None, time_division_remainder=None,
+    ):
+        super().__init__()
+        # The device and torch_dtype is used for the storage of intermediate variables, not models.
+        self.device = device
+        self.torch_dtype = torch_dtype
+        # The following parameters are used for shape check.
+        self.height_division_factor = height_division_factor
+        self.width_division_factor = width_division_factor
+        self.time_division_factor = time_division_factor
+        self.time_division_remainder = time_division_remainder
+        self.vram_management_enabled = False
+        
+        
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.torch_dtype = dtype
+        super().to(*args, **kwargs)
+        return self
+
+
+    def check_resize_height_width(self, height, width, num_frames=None):
+        # Shape check
+        if height % self.height_division_factor != 0:
+            height = (height + self.height_division_factor - 1) // self.height_division_factor * self.height_division_factor
+            print(f"height % {self.height_division_factor} != 0. We round it up to {height}.")
+        if width % self.width_division_factor != 0:
+            width = (width + self.width_division_factor - 1) // self.width_division_factor * self.width_division_factor
+            print(f"width % {self.width_division_factor} != 0. We round it up to {width}.")
+        if num_frames is None:
+            return height, width
+        else:
+            if num_frames % self.time_division_factor != self.time_division_remainder:
+                num_frames = (num_frames + self.time_division_factor - 1) // self.time_division_factor * self.time_division_factor + self.time_division_remainder
+                print(f"num_frames % {self.time_division_factor} != {self.time_division_remainder}. We round it up to {num_frames}.")
+            return height, width, num_frames
+
+
+    def preprocess_image(self, image, torch_dtype=None, device=None, pattern="B C H W", min_value=-1, max_value=1):
+        # Transform a PIL.Image to torch.Tensor
+        image = torch.Tensor(np.array(image, dtype=np.float32))
+        image = image.to(dtype=torch_dtype or self.torch_dtype, device=device or self.device)
+        image = image * ((max_value - min_value) / 255) + min_value
+        image = repeat(image, f"H W C -> {pattern}", **({"B": 1} if "B" in pattern else {}))
+        return image
+
+
+    def preprocess_video(self, video, torch_dtype=None, device=None, pattern="B C T H W", min_value=-1, max_value=1):
+        # Transform a list of PIL.Image to torch.Tensor
+        
+        
+        if hasattr(video, 'length') and video.length is not None:
+            video = [self.preprocess_image(video[idx], torch_dtype=torch_dtype, device=device, min_value=min_value, max_value=max_value) for idx in range(video.length)]
+        else:
+            video = [self.preprocess_image(image, torch_dtype=torch_dtype, device=device, min_value=min_value, max_value=max_value) for image in video]
+        video = torch.stack(video, dim=pattern.index("T") // 2)
+        return video
+
+
+    def vae_output_to_image(self, vae_output, pattern="B C H W", min_value=-1, max_value=1):
+        # Transform a torch.Tensor to PIL.Image
+        if pattern != "H W C":
+            vae_output = reduce(vae_output, f"{pattern} -> H W C", reduction="mean")
+        image = ((vae_output - min_value) * (255 / (max_value - min_value))).clip(0, 255)
+        image = image.to(device="cpu", dtype=torch.uint8)
+        image = Image.fromarray(image.numpy())
+        return image
+
+
+    def vae_output_to_video(self, vae_output, pattern="B C T H W", min_value=-1, max_value=1):
+        # Transform a torch.Tensor to list of PIL.Image
+        if pattern != "T H W C":
+            vae_output = reduce(vae_output, f"{pattern} -> T H W C", reduction="mean")
+        video = [self.vae_output_to_image(image, pattern="H W C", min_value=min_value, max_value=max_value) for image in vae_output]
+        return video
+
+
+    def load_models_to_device(self, model_names=[]):
+        if self.vram_management_enabled:
+            # offload models
+            for name, model in self.named_children():
+                if name not in model_names:
+                    if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
+                        for module in model.modules():
+                            if hasattr(module, "offload"):
+                                module.offload()
+                    else:
+                        model.cpu()
+            torch.cuda.empty_cache()
+            # onload models
+            for name, model in self.named_children():
+                if name in model_names:
+                    if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
+                        for module in model.modules():
+                            if hasattr(module, "onload"):
+                                module.onload()
+                    else:
+                        model.to(self.device)
+
+
+    def generate_noise(self, shape, seed=None, rand_device="cpu", rand_torch_dtype=torch.float32, device=None, torch_dtype=None):
+        # Initialize Gaussian noise
+        generator = None if seed is None else torch.Generator(rand_device).manual_seed(seed)
+        noise = torch.randn(shape, generator=generator, device=rand_device, dtype=rand_torch_dtype)
+        noise = noise.to(dtype=torch_dtype or self.torch_dtype, device=device or self.device)
+        return noise
+
+
+    def enable_cpu_offload(self):
+        warnings.warn("`enable_cpu_offload` will be deprecated. Please use `enable_vram_management`.")
+        self.vram_management_enabled = True
+        
+        
+    def get_vram(self):
+        return torch.cuda.mem_get_info(self.device)[1] / (1024 ** 3)
+    
+    
+    def freeze_except(self, model_names):
+        for name, model in self.named_children():
+            if name in model_names:
+                model.train()
+                model.requires_grad_(True)
+            else:
+                model.eval()
+                model.requires_grad_(False)
+
+
+@dataclass
+class ModelConfig:
+    path: Union[str, list[str]] = None
+    model_id: str = None
+    origin_file_pattern: Union[str, list[str]] = None
+    download_resource: str = "ModelScope"
+    offload_device: Optional[Union[str, torch.device]] = None
+    offload_dtype: Optional[torch.dtype] = None
+
+    def download_if_necessary(self, local_model_path="./checkpoints", skip_download=False, use_usp=False):
+        if self.path is None:
+            # Check model_id and origin_file_pattern
+            if self.model_id is None:
+                raise ValueError(f"""No valid model files. Please use `ModelConfig(path="xxx")` or `ModelConfig(model_id="xxx/yyy", origin_file_pattern="zzz")`.""")
+            
+            # Skip if not in rank 0
+            if use_usp:
+                import torch.distributed as dist
+                skip_download = dist.get_rank() != 0
+                
+            # Check whether the origin path is a folder
+            if self.origin_file_pattern is None or self.origin_file_pattern == "":
+                self.origin_file_pattern = ""
+                allow_file_pattern = None
+                is_folder = True
+            elif isinstance(self.origin_file_pattern, str) and self.origin_file_pattern.endswith("/"):
+                allow_file_pattern = self.origin_file_pattern + "*"
+                is_folder = True
+            else:
+                allow_file_pattern = self.origin_file_pattern
+                is_folder = False
+            
+            # Download
+            if not skip_download:
+                
+                # downloaded_files = glob.glob(self.origin_file_pattern, root_dir=os.path.join(local_model_path, self.model_id))
+                #!========================================================================================================================
+                downloaded_files = glob.glob(os.path.join(local_model_path, self.model_id, self.origin_file_pattern))
+                #!========================================================================================================================
+                
+                if downloaded_files is None or len(downloaded_files) == 0 or not os.path.exists(downloaded_files[0]) :
+                    if 'Wan2' in self.model_id:
+                        ms_snap_download(
+                            self.model_id,
+                            local_dir=os.path.join(local_model_path, self.model_id),
+                            allow_file_pattern=allow_file_pattern,
+                            ignore_file_pattern=downloaded_files,
+                        )
+                    else:
+                        hf_snap_download(
+                        repo_id=self.model_id,
+                        local_dir=os.path.join(local_model_path, self.model_id),
+                        allow_patterns=allow_file_pattern,
+                        ignore_patterns=downloaded_files if downloaded_files else None
+                        )
+            
+            # Let rank 1, 2, ... wait for rank 0
+            if use_usp:
+                import torch.distributed as dist
+                dist.barrier(device_ids=[dist.get_rank()])
+                
+            # Return downloaded files
+            if is_folder:
+                self.path = os.path.join(local_model_path, self.model_id, self.origin_file_pattern)
+            else:
+                self.path = glob.glob(os.path.join(local_model_path, self.model_id, self.origin_file_pattern))
+            if isinstance(self.path, list) and len(self.path) == 1:
+                self.path = self.path[0]
+
+
+        
+
+class WanVideoPipeline(BasePipeline):
+
+    def __init__(self, device="cuda", torch_dtype=torch.bfloat16, tokenizer_path=None):
+        super().__init__(
+            device=device, torch_dtype=torch_dtype,
+            height_division_factor=16, width_division_factor=16, time_division_factor=4, time_division_remainder=1
+        )
+        self.scheduler = FlowMatchScheduler(shift=5, sigma_min=0.0, extra_one_step=True)
+
+        self.prompter = WanPrompter(tokenizer_path=tokenizer_path)
+        self.text_encoder: WanTextEncoder = None
+        self.image_encoder: WanImageEncoder = None
+        self.dit: WanModel = None
+        self.dit2: WanModel = None
+        self.vae: WanVideoVAE = None
+        self.motion_controller: WanMotionControllerModel = None
+        self.vace: VaceWanModel = None
+        self.in_iteration_models = ("dit", "motion_controller", "vace")
+        self.in_iteration_models_2 = ("dit2", "motion_controller", "vace")
+        self.unit_runner = PipelineUnitRunner()
+        self.units = [
+            WanVideoUnit_ShapeChecker(),
+            WanVideoUnit_NoiseInitializer(),
+            WanVideoUnit_InputVideoEmbedder(),
+            WanVideoUnit_PromptEmbedder(),
+            # WanVideoUnit_ImageEmbedderVAE(),
+            # WanVideoUnit_ImageEmbedderCLIP(),
+            # WanVideoUnit_ImageEmbedderFused(),
+            # WanVideoUnit_FunControl(),
+            WanVideoUnit_FunControl_Mask(),
+            # WanVideoUnit_FunReference(),
+            # WanVideoUnit_FunCameraControl(),
+            # WanVideoUnit_SpeedControl(),
+            # WanVideoUnit_VACE(),
+            # WanVideoUnit_UnifiedSequenceParallel(),
+            # WanVideoUnit_TeaCache(),
+            # WanVideoUnit_CfgMerger(),
+        ]
+        self.model_fn = model_fn_wan_video
+        
+    
+    def load_lora(self, module, path, alpha=1):
+        loader = GeneralLoRALoader(torch_dtype=self.torch_dtype, device=self.device)
+        lora = load_state_dict(path, torch_dtype=self.torch_dtype, device=self.device)
+        loader.load(module, lora, alpha=alpha)
+
+        
+    def training_loss(self, **inputs):
+        max_timestep_boundary = int(inputs.get("max_timestep_boundary", 1) * self.scheduler.num_train_timesteps)
+        min_timestep_boundary = int(inputs.get("min_timestep_boundary", 0) * self.scheduler.num_train_timesteps)
+        timestep_id = torch.randint(min_timestep_boundary, max_timestep_boundary, (1,))
+        timestep = self.scheduler.timesteps[timestep_id].to(dtype=self.torch_dtype, device=self.device)
+        #* 单步去噪的时候,每次返回的都是纯噪声 
+        #? 指的是input_latents 吧? 
+        #* 本来就有inputs["latents"], 只不过是完全等于inputs["noise"], 这里做了更新然后覆盖
+        inputs["latents"] = self.scheduler.add_noise(inputs["input_latents"], inputs["noise"], timestep)
+        training_target = self.scheduler.training_target(inputs["input_latents"], inputs["noise"], timestep)
+        
+        noise_pred = self.model_fn(**inputs, timestep=timestep)#* timestep === 1
+        
+        loss = torch.nn.functional.mse_loss(noise_pred.float(), training_target.float())
+        loss = loss * self.scheduler.training_weight(timestep)
+        return loss
+
+
+    def enable_vram_management(self, num_persistent_param_in_dit=None, vram_limit=None, vram_buffer=0.5):
+        self.vram_management_enabled = True
+        if num_persistent_param_in_dit is not None:
+            vram_limit = None
+        else:
+            if vram_limit is None:
+                vram_limit = self.get_vram()
+            vram_limit = vram_limit - vram_buffer
+        if self.text_encoder is not None:
+            dtype = next(iter(self.text_encoder.parameters())).dtype
+            enable_vram_management(
+                self.text_encoder,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Embedding: AutoWrappedModule,
+                    T5RelativeEmbedding: AutoWrappedModule,
+                    T5LayerNorm: AutoWrappedModule,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+        if self.dit is not None:
+            dtype = next(iter(self.dit.parameters())).dtype
+            device = "cpu" if vram_limit is not None else self.device
+            enable_vram_management(
+                self.dit,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv3d: AutoWrappedModule,
+                    torch.nn.LayerNorm: WanAutoCastLayerNorm,
+                    RMSNorm: AutoWrappedModule,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                max_num_param=num_persistent_param_in_dit,
+                overflow_module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+        if self.dit2 is not None:
+            dtype = next(iter(self.dit2.parameters())).dtype
+            device = "cpu" if vram_limit is not None else self.device
+            enable_vram_management(
+                self.dit2,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv3d: AutoWrappedModule,
+                    torch.nn.LayerNorm: WanAutoCastLayerNorm,
+                    RMSNorm: AutoWrappedModule,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                max_num_param=num_persistent_param_in_dit,
+                overflow_module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+        if self.vae is not None:
+            dtype = next(iter(self.vae.parameters())).dtype
+            enable_vram_management(
+                self.vae,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                    RMS_norm: AutoWrappedModule,
+                    CausalConv3d: AutoWrappedModule,
+                    Upsample: AutoWrappedModule,
+                    torch.nn.SiLU: AutoWrappedModule,
+                    torch.nn.Dropout: AutoWrappedModule,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=self.device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+            )
+        if self.image_encoder is not None:
+            dtype = next(iter(self.image_encoder.parameters())).dtype
+            enable_vram_management(
+                self.image_encoder,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv2d: AutoWrappedModule,
+                    torch.nn.LayerNorm: AutoWrappedModule,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=dtype,
+                    computation_device=self.device,
+                ),
+            )
+        if self.motion_controller is not None:
+            dtype = next(iter(self.motion_controller.parameters())).dtype
+            enable_vram_management(
+                self.motion_controller,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device="cpu",
+                    computation_dtype=dtype,
+                    computation_device=self.device,
+                ),
+            )
+        if self.vace is not None:
+            device = "cpu" if vram_limit is not None else self.device
+            enable_vram_management(
+                self.vace,
+                module_map = {
+                    torch.nn.Linear: AutoWrappedLinear,
+                    torch.nn.Conv3d: AutoWrappedModule,
+                    torch.nn.LayerNorm: AutoWrappedModule,
+                    RMSNorm: AutoWrappedModule,
+                },
+                module_config = dict(
+                    offload_dtype=dtype,
+                    offload_device="cpu",
+                    onload_dtype=dtype,
+                    onload_device=device,
+                    computation_dtype=self.torch_dtype,
+                    computation_device=self.device,
+                ),
+                vram_limit=vram_limit,
+            )
+            
+            
+    def initialize_usp(self):
+        import torch.distributed as dist
+        from xfuser.core.distributed import initialize_model_parallel, init_distributed_environment
+        dist.init_process_group(backend="nccl", init_method="env://")
+        init_distributed_environment(rank=dist.get_rank(), world_size=dist.get_world_size())
+        initialize_model_parallel(
+            sequence_parallel_degree=dist.get_world_size(),
+            ring_degree=1,
+            ulysses_degree=dist.get_world_size(),
+        )
+        torch.cuda.set_device(dist.get_rank())
+            
+            
+    def enable_usp(self):
+        from xfuser.core.distributed import get_sequence_parallel_world_size
+        from ..distributed.xdit_context_parallel import usp_attn_forward, usp_dit_forward
+
+        for block in self.dit.blocks:
+            block.self_attn.forward = types.MethodType(usp_attn_forward, block.self_attn)
+        self.dit.forward = types.MethodType(usp_dit_forward, self.dit)
+        if self.dit2 is not None:
+            for block in self.dit2.blocks:
+                block.self_attn.forward = types.MethodType(usp_attn_forward, block.self_attn)
+            self.dit2.forward = types.MethodType(usp_dit_forward, self.dit2)
+        self.sp_size = get_sequence_parallel_world_size()
+        self.use_unified_sequence_parallel = True
+
+
+    @staticmethod
+    def from_pretrained(
+        torch_dtype: torch.dtype = torch.bfloat16,
+        device: Union[str, torch.device] = "cuda",
+        model_configs: list[ModelConfig] = [],
+        tokenizer_config: ModelConfig = ModelConfig(model_id="Wan-AI/Wan2.1-T2V-1.3B", origin_file_pattern="google/*"),
+        local_model_path: str = "./checkpoints",
+        skip_download: bool = False,
+        redirect_common_files: bool = True,
+        use_usp=False,
+        training_strategy='origin',
+    ):
+    
+        # Redirect model path
+        
+        if redirect_common_files:
+            
+            redirect_dict = {
+                "models_t5_umt5-xxl-enc-bf16.pth": "Wan-AI/Wan2.1-T2V-1.3B",
+                "Wan2.1_VAE.pth": "Wan-AI/Wan2.1-T2V-1.3B",
+                "models_clip_open-clip-xlm-roberta-large-vit-huge-14.pth": "Wan-AI/Wan2.1-I2V-14B-480P",
+            }
+            for model_config in model_configs:
+                if model_config.origin_file_pattern is None or model_config.model_id is None:
+                    continue
+                if model_config.origin_file_pattern in redirect_dict and model_config.model_id != redirect_dict[model_config.origin_file_pattern]:
+                    print(f"To avoid repeatedly downloading model files, ({model_config.model_id}, {model_config.origin_file_pattern}) is redirected to ({redirect_dict[model_config.origin_file_pattern]}, {model_config.origin_file_pattern}). You can use `redirect_common_files=False` to disable file redirection.")
+                    model_config.model_id = redirect_dict[model_config.origin_file_pattern]
+        
+        # Initialize pipeline
+        
+        if training_strategy == 'origin':
+            pipe = WanVideoPipeline(device=device, torch_dtype=torch_dtype)
+            logger.warning("Using origin generative model  training")
+        else:
+            raise ValueError(f"Invalid training strategy: {training_strategy}")
+        
+        if use_usp: pipe.initialize_usp()
+        
+        # Download and load models
+        model_manager = ModelManager()
+        
+        for model_config in model_configs:
+            model_config.download_if_necessary(use_usp=use_usp)
+            model_manager.load_model(
+                model_config.path,
+                device=model_config.offload_device or device,
+                torch_dtype=model_config.offload_dtype or torch_dtype
+            )
+        
+        # Load models
+        pipe.text_encoder = model_manager.fetch_model("wan_video_text_encoder")
+        dit = model_manager.fetch_model("wan_video_dit", index=2)
+        if isinstance(dit, list):
+            pipe.dit, pipe.dit2 = dit
+        else:
+            pipe.dit = dit
+        pipe.vae = model_manager.fetch_model("wan_video_vae")
+        pipe.image_encoder = model_manager.fetch_model("wan_video_image_encoder")
+        pipe.motion_controller = model_manager.fetch_model("wan_video_motion_controller")
+        pipe.vace = model_manager.fetch_model("wan_video_vace")
+        
+        # Size division factor
+        if pipe.vae is not None:
+            pipe.height_division_factor = pipe.vae.upsampling_factor * 2
+            pipe.width_division_factor = pipe.vae.upsampling_factor * 2
+
+        # Initialize tokenizer
+        tokenizer_config.download_if_necessary(use_usp=use_usp)
+        pipe.prompter.fetch_models(pipe.text_encoder)
+        pipe.prompter.fetch_tokenizer(tokenizer_config.path)
+        
+        # Unified Sequence Parallel
+        if use_usp: pipe.enable_usp()
+        return pipe
+
+
+    
+
+    @torch.no_grad()
+    def __call__(
+        self,
+        # Prompt
+        prompt: str,
+        negative_prompt: Optional[str] = "",
+        # Image-to-video
+        input_image: Optional[Image.Image] = None,
+        # First-last-frame-to-video
+        end_image: Optional[Image.Image] = None,
+        # Video-to-video
+        input_video: Optional[list[Image.Image]] = None,
+        denoising_strength: Optional[float] = 1.0,
+        # ControlNet
+        control_video: Optional[list[Image.Image]] = None,
+        reference_image: Optional[Image.Image] = None,
+        # Camera control
+        camera_control_direction: Optional[Literal["Left", "Right", "Up", "Down", "LeftUp", "LeftDown", "RightUp", "RightDown"]] = None,
+        camera_control_speed: Optional[float] = 1/54,
+        camera_control_origin: Optional[tuple] = (0, 0.532139961, 0.946026558, 0.5, 0.5, 0, 0, 1, 0, 0, 0, 0, 1, 0, 0, 0, 0, 1, 0),
+        # VACE
+        vace_video: Optional[list[Image.Image]] = None,
+        vace_video_mask: Optional[Image.Image] = None,
+        vace_reference_image: Optional[Image.Image] = None,
+        vace_scale: Optional[float] = 1.0,
+        # Randomness
+        seed: Optional[int] = None,
+        rand_device: Optional[str] = "cpu",
+        # Shape
+        height: Optional[int] = 480,
+        width: Optional[int] = 832,
+        num_frames=81,
+        # Classifier-free guidance
+        cfg_scale: Optional[float] = 5.0,
+        cfg_merge: Optional[bool] = False,
+        # Boundary
+        switch_DiT_boundary: Optional[float] = 0.875,
+        # Scheduler
+        num_inference_steps: Optional[int] = 50,
+        sigma_shift: Optional[float] = 5.0,
+        # Speed control
+        motion_bucket_id: Optional[int] = None,
+        # VAE tiling
+        tiled: Optional[bool] = True,
+        tile_size: Optional[tuple[int, int]] = (30, 52),
+        tile_stride: Optional[tuple[int, int]] = (15, 26),
+        # Sliding window
+        sliding_window_size: Optional[int] = None,
+        sliding_window_stride: Optional[int] = None,
+        # Teacache
+        tea_cache_l1_thresh: Optional[float] = None,
+        tea_cache_model_id: Optional[str] = "",
+        # progress_bar
+        progress_bar_cmd=tqdm,
+        mask: Optional[Image.Image] = None,
+    ):
+
+
+
+        # Scheduler
+        self.scheduler.set_timesteps(num_inference_steps, denoising_strength=denoising_strength, shift=sigma_shift)
+        
+        # Inputs
+        inputs_posi = {
+            "prompt": prompt,
+            "tea_cache_l1_thresh": tea_cache_l1_thresh, "tea_cache_model_id": tea_cache_model_id, "num_inference_steps": num_inference_steps,
+        }
+        inputs_nega = {
+            "negative_prompt": negative_prompt,
+            "tea_cache_l1_thresh": tea_cache_l1_thresh, "tea_cache_model_id": tea_cache_model_id, "num_inference_steps": num_inference_steps,
+        }
+        inputs_shared = {
+            "input_image": input_image,
+            "end_image": end_image,
+            "input_video": input_video, "denoising_strength": denoising_strength,
+            "control_video": control_video, "reference_image": reference_image,
+            "camera_control_direction": camera_control_direction, "camera_control_speed": camera_control_speed, "camera_control_origin": camera_control_origin,
+            "vace_video": vace_video, "vace_video_mask": vace_video_mask, "vace_reference_image": vace_reference_image, "vace_scale": vace_scale,
+            "seed": seed, "rand_device": rand_device,
+            "height": height, "width": width, "num_frames": num_frames,
+            "cfg_scale": cfg_scale, "cfg_merge": cfg_merge,
+            "sigma_shift": sigma_shift,
+            "motion_bucket_id": motion_bucket_id,
+            "tiled": tiled, "tile_size": tile_size, "tile_stride": tile_stride,
+            "sliding_window_size": sliding_window_size, "sliding_window_stride": sliding_window_stride,
+            "mask":mask,
+        }
+        
+        for unit in self.units:
+            inputs_shared, inputs_posi, inputs_nega = self.unit_runner(unit, self, inputs_shared, inputs_posi, inputs_nega)
+
+        # Denoise
+        self.load_models_to_device(self.in_iteration_models)
+        models = {name: getattr(self, name) for name in self.in_iteration_models}
+        for progress_id, timestep in enumerate(progress_bar_cmd(self.scheduler.timesteps)):
+            # Switch DiT if necessary
+            if timestep.item() < switch_DiT_boundary * self.scheduler.num_train_timesteps and self.dit2 is not None and not models["dit"] is self.dit2:
+                self.load_models_to_device(self.in_iteration_models_2)
+                models["dit"] = self.dit2
+                
+            # Timestep
+            timestep = timestep.unsqueeze(0).to(dtype=self.torch_dtype, device=self.device)
+            
+            # Inference
+            noise_pred_posi = self.model_fn(**models, **inputs_shared, **inputs_posi, timestep=timestep)
+            if cfg_scale != 1.0:
+                if cfg_merge:
+                    noise_pred_posi, noise_pred_nega = noise_pred_posi.chunk(2, dim=0)
+                else:
+                    noise_pred_nega = self.model_fn(**models, **inputs_shared, **inputs_nega, timestep=timestep)
+                noise_pred = noise_pred_nega + cfg_scale * (noise_pred_posi - noise_pred_nega)
+            else:
+                noise_pred = noise_pred_posi
+
+            # Scheduler
+            inputs_shared["latents"] = self.scheduler.step(noise_pred, self.scheduler.timesteps[progress_id], inputs_shared["latents"])
+            if "first_frame_latents" in inputs_shared:
+                inputs_shared["latents"][:, :, 0:1] = inputs_shared["first_frame_latents"]
+        
+        # VACE (TODO: remove it)
+        if vace_reference_image is not None:
+            inputs_shared["latents"] = inputs_shared["latents"][:, :, 1:]
+
+        # Decode
+        self.load_models_to_device(['vae'])
+        vae_outs = self.vae.decode(inputs_shared["latents"], device=self.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
+        # from einops import  reduce
+        # video = reduce(vae_outs, 'b c t h w -> b c t', 'mean')
+
+        video = self.vae_output_to_video(vae_outs)
+        self.load_models_to_device([])
+
+        return video,vae_outs
+
+
+
+
+
+
+
+        
+
+class WanVideoUnit_ShapeChecker(PipelineUnit):
+    def __init__(self):
+        super().__init__(input_params=("height", "width", "num_frames"))
+
+    def process(self, pipe: WanVideoPipeline, height, width, num_frames):
+        height, width, num_frames = pipe.check_resize_height_width(height, width, num_frames)
+        return {"height": height, "width": width, "num_frames": num_frames}
+
+
+
+class WanVideoUnit_NoiseInitializer(PipelineUnit):
+    def __init__(self):
+        super().__init__(input_params=("height", "width", "num_frames", "seed", "rand_device", "vace_reference_image"))
+
+    def process(self, pipe: WanVideoPipeline, height, width, num_frames, seed, rand_device, vace_reference_image):
+        length = (num_frames - 1) // 4 + 1
+        if vace_reference_image is not None:
+            length += 1
+        shape = (1, pipe.vae.model.z_dim, length, height // pipe.vae.upsampling_factor, width // pipe.vae.upsampling_factor)
+        noise = pipe.generate_noise(shape, seed=seed, rand_device=rand_device)
+        if vace_reference_image is not None:
+            noise = torch.concat((noise[:, :, -1:], noise[:, :, :-1]), dim=2)
+        return {"noise": noise}
+    
+
+
+class WanVideoUnit_InputVideoEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_video", "noise", "tiled", "tile_size", "tile_stride", "vace_reference_image"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, input_video, noise, tiled, tile_size, tile_stride, vace_reference_image):
+        if input_video is None:
+            return {"latents": noise}
+        
+        pipe.load_models_to_device(["vae"])#* input_video is the GT 
+        input_video = pipe.preprocess_video(input_video)  #* [B,3,F,W,H]
+        #* [B,3,(F/4) + 1 ,W/8,H/8]
+        input_latents = pipe.vae.encode(input_video, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device) 
+        if vace_reference_image is not None:
+            vace_reference_image = pipe.preprocess_video([vace_reference_image])
+            vace_reference_latents = pipe.vae.encode(vace_reference_image, device=pipe.device).to(dtype=pipe.torch_dtype, device=pipe.device)
+            input_latents = torch.concat([vace_reference_latents, input_latents], dim=2)
+        #?  during training, the input_latents have nothing to do with the noise,
+        #?  but during inference, the input_latents is used to generate the noise
+        if pipe.scheduler.training:
+            return {"latents": noise, "input_latents": input_latents}
+        else:
+            latents = pipe.scheduler.add_noise(input_latents, noise, timestep=pipe.scheduler.timesteps[0])
+            return {"latents": latents}
+
+
+
+class WanVideoUnit_PromptEmbedder(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            seperate_cfg=True,
+            input_params_posi={"prompt": "prompt", "positive": "positive"},
+            input_params_nega={"prompt": "negative_prompt", "positive": "positive"},
+            onload_model_names=("text_encoder",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, prompt, positive) -> dict:
+        pipe.load_models_to_device(self.onload_model_names)
+        prompt_emb = pipe.prompter.encode_prompt(prompt, positive=positive, device=pipe.device)
+        return {"context": prompt_emb}
+
+
+
+class WanVideoUnit_ImageEmbedder(PipelineUnit):
+    """
+    Deprecated
+    """
+    def __init__(self):
+        super().__init__(
+            input_params=("input_image", "end_image", "num_frames", "height", "width", "tiled", "tile_size", "tile_stride"),
+            onload_model_names=("image_encoder", "vae")
+        )
+
+    def process(self, pipe: WanVideoPipeline, input_image, end_image, num_frames, height, width, tiled, tile_size, tile_stride):
+        if input_image is None or pipe.image_encoder is None:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        image = pipe.preprocess_image(input_image.resize((width, height))).to(pipe.device)
+        clip_context = pipe.image_encoder.encode_image([image])
+        msk = torch.ones(1, num_frames, height//8, width//8, device=pipe.device) #* indicate which image is  reference image
+        msk[:, 1:] = 0
+        if end_image is not None:
+            end_image = pipe.preprocess_image(end_image.resize((width, height))).to(pipe.device)
+            vae_input = torch.concat([image.transpose(0,1), torch.zeros(3, num_frames-2, height, width).to(image.device), end_image.transpose(0,1)],dim=1)
+            if pipe.dit.has_image_pos_emb:
+                clip_context = torch.concat([clip_context, pipe.image_encoder.encode_image([end_image])], dim=1)
+            msk[:, -1:] = 1
+        else:
+            vae_input = torch.concat([image.transpose(0, 1), torch.zeros(3, num_frames-1, height, width).to(image.device)], dim=1)
+
+        msk = torch.concat([torch.repeat_interleave(msk[:, 0:1], repeats=4, dim=1), msk[:, 1:]], dim=1)
+        msk = msk.view(1, msk.shape[1] // 4, 4, height//8, width//8)
+        msk = msk.transpose(1, 2)[0]
+        
+        y = pipe.vae.encode([vae_input.to(dtype=pipe.torch_dtype, device=pipe.device)], device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        y = torch.concat([msk, y])
+        y = y.unsqueeze(0)
+        clip_context = clip_context.to(dtype=pipe.torch_dtype, device=pipe.device)
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        return {"clip_feature": clip_context, "y": y}
+
+
+
+class WanVideoUnit_ImageEmbedderCLIP(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_image", "end_image", "height", "width"),
+            onload_model_names=("image_encoder",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, input_image, end_image, height, width):
+        if input_image is None or pipe.image_encoder is None or not pipe.dit.require_clip_embedding:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        image = pipe.preprocess_image(input_image.resize((width, height))).to(pipe.device)
+        clip_context = pipe.image_encoder.encode_image([image])
+        if end_image is not None:
+            end_image = pipe.preprocess_image(end_image.resize((width, height))).to(pipe.device)
+            if pipe.dit.has_image_pos_emb:
+                clip_context = torch.concat([clip_context, pipe.image_encoder.encode_image([end_image])], dim=1)
+        clip_context = clip_context.to(dtype=pipe.torch_dtype, device=pipe.device)
+        return {"clip_feature": clip_context}
+    
+
+
+class WanVideoUnit_ImageEmbedderVAE(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("input_image", "end_image", "num_frames", "height", "width", "tiled", "tile_size", "tile_stride"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, input_image, end_image, num_frames, height, width, tiled, tile_size, tile_stride):
+        if input_image is None or not pipe.dit.require_vae_embedding:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        image = pipe.preprocess_image(input_image.resize((width, height))).to(pipe.device)
+        msk = torch.ones(1, num_frames, height//8, width//8, device=pipe.device)
+        msk[:, 1:] = 0
+        if end_image is not None:
+            end_image = pipe.preprocess_image(end_image.resize((width, height))).to(pipe.device)
+            vae_input = torch.concat([image.transpose(0,1), torch.zeros(3, num_frames-2, height, width).to(image.device), end_image.transpose(0,1)],dim=1)
+            msk[:, -1:] = 1
+        else:
+            vae_input = torch.concat([image.transpose(0, 1), torch.zeros(3, num_frames-1, height, width).to(image.device)], dim=1)
+
+        msk = torch.concat([torch.repeat_interleave(msk[:, 0:1], repeats=4, dim=1), msk[:, 1:]], dim=1)
+        msk = msk.view(1, msk.shape[1] // 4, 4, height//8, width//8)
+        msk = msk.transpose(1, 2)[0]
+        
+        y = pipe.vae.encode([vae_input.to(dtype=pipe.torch_dtype, device=pipe.device)], device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)[0]
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        y = torch.concat([msk, y])
+        y = y.unsqueeze(0)
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        return {"y": y}
+
+
+
+class WanVideoUnit_ImageEmbedderFused(PipelineUnit):
+    """
+    Encode input image to latents using VAE. This unit is for Wan-AI/Wan2.2-TI2V-5B.
+    """
+    def __init__(self):
+        super().__init__(
+            input_params=("input_image", "latents", "height", "width", "tiled", "tile_size", "tile_stride"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, input_image, latents, height, width, tiled, tile_size, tile_stride):
+        if input_image is None or not pipe.dit.fuse_vae_embedding_in_latents:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        image = pipe.preprocess_image(input_image.resize((width, height))).transpose(0, 1)
+        z = pipe.vae.encode([image], device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride)
+        latents[:, :, 0: 1] = z
+        return {"latents": latents, "fuse_vae_embedding_in_latents": True, "first_frame_latents": z}
+
+
+
+class WanVideoUnit_FunControl(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("control_video", "num_frames", "height", "width", "tiled", "tile_size", "tile_stride", "clip_feature", "y"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, control_video, num_frames, height, width, tiled, tile_size, tile_stride, clip_feature, y):
+        
+        if control_video is None:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        #* transfer to torch.tensor from PIL.Image
+        #* result size: [1, 3, F, H, W]
+        control_video = pipe.preprocess_video(control_video) 
+        
+        control_latents = pipe.vae.encode(control_video, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+        #* size of control_latents: [1, 3, (F/4) + 1 , H/8, W/8]
+        control_latents = control_latents.to(dtype=pipe.torch_dtype, device=pipe.device)
+        
+        if clip_feature is None or y is None:
+            #* this branch is used during  training 
+            clip_feature = torch.zeros((1, 257, 1280), dtype=pipe.torch_dtype, device=pipe.device)
+
+            # y = torch.zeros((1, 16, (num_frames - 1) // 4 + 1, height//8, width//8), dtype=pipe.torch_dtype, device=pipe.device)
+
+            #* [1, 16, (F/4) + 1 , H/8, W/8]
+            y = torch.zeros((1, 16, control_latents.shape[-3], height//8, width//8), dtype=pipe.torch_dtype, device=pipe.device) 
+        else:
+            y = y[:, -16:]
+        #* control_latents: [1, 16, 21, 60, 80];  y: [1, 16, 21, 60, 80])
+        
+        #* [1, 32, (F/4) + 1 , H/8, W/8], 前16个通道是control_latents, 后16个通道是y(或者说0 vector)
+        y = torch.concat([control_latents, y], dim=1)
+        return {"clip_feature": clip_feature, "y": y}
+
+
+
+class WanVideoUnit_FunControl_Mask(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("control_video", "mask","num_frames", "height", "width", "tiled", "tile_size", "tile_stride", "clip_feature", "y"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, control_video, mask, num_frames, height, width, tiled, tile_size, tile_stride, clip_feature, y):
+        
+        if control_video is None:
+            return {}
+        pipe.load_models_to_device(self.onload_model_names)
+        #* transfer to torch.tensor from PIL.Image
+        #* result size: [1, 3, F, H, W]
+        
+        control_video = pipe.preprocess_video(control_video) 
+        
+        
+        control_latents = pipe.vae.encode(control_video, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+        
+        #* size of control_latents: [1, 3, (F/4) + 1 , H/8, W/8]
+        control_latents = control_latents.to(dtype=pipe.torch_dtype, device=pipe.device)
+
+        if mask is not None:
+            mask = pipe.preprocess_video(mask) 
+            mask_latents = pipe.vae.encode(mask, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+            mask_latents = mask_latents.to(dtype=pipe.torch_dtype, device=pipe.device)
+
+        
+        if clip_feature is None or y is None:
+            #* this branch is used during  training 
+            clip_feature = torch.zeros((1, 257, 1280), dtype=pipe.torch_dtype, device=pipe.device)
+
+            # y = torch.zeros((1, 16, (num_frames - 1) // 4 + 1, height//8, width//8), dtype=pipe.torch_dtype, device=pipe.device)
+
+            #* [1, 16, (F/4) + 1 , H/8, W/8]
+            y = torch.zeros((1, 16, control_latents.shape[-3], height//8, width//8), dtype=pipe.torch_dtype, device=pipe.device) 
+        else:
+            y = y[:, -16:]
+
+        #* control_latents: [1, 16, 21, 60, 80];  y: [1, 16, 21, 60, 80])
+        
+        #* [1, 32, (F/4) + 1 , H/8, W/8], 前16个通道是control_latents, 后16个通道是y(或者说0 vector)
+        
+        if mask is not None:
+            y = torch.concat([control_latents, mask_latents], dim=1)
+            # logger.warning(f"mask is provided, using mask_latents instead of y")
+        else:
+            y = torch.concat([control_latents, y], dim=1)
+            # logger.warning(f"mask is not provided, using y")
+            
+        return {"clip_feature": clip_feature, "y": y}
+    
+
+
+class WanVideoUnit_FunReference(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("reference_image", "height", "width", "reference_image"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, reference_image, height, width):
+        if reference_image is None:
+            return {}
+        pipe.load_models_to_device(["vae"])
+        reference_image = reference_image.resize((width, height))
+        reference_latents = pipe.preprocess_video([reference_image])
+        reference_latents = pipe.vae.encode(reference_latents, device=pipe.device)
+        clip_feature = pipe.preprocess_image(reference_image)
+        clip_feature = pipe.image_encoder.encode_image([clip_feature])
+        return {"reference_latents": reference_latents, "clip_feature": clip_feature}
+
+
+
+
+class WanVideoUnit_FunCameraControl(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("height", "width", "num_frames", "camera_control_direction", "camera_control_speed", "camera_control_origin", "latents", "input_image"),
+            onload_model_names=("vae",)
+        )
+
+    def process(self, pipe: WanVideoPipeline, height, width, num_frames, camera_control_direction, camera_control_speed, camera_control_origin, latents, input_image):
+        if camera_control_direction is None:
+            return {}
+        camera_control_plucker_embedding = pipe.dit.control_adapter.process_camera_coordinates(
+            camera_control_direction, num_frames, height, width, camera_control_speed, camera_control_origin)
+        
+        control_camera_video = camera_control_plucker_embedding[:num_frames].permute([3, 0, 1, 2]).unsqueeze(0)
+        control_camera_latents = torch.concat(
+            [
+                torch.repeat_interleave(control_camera_video[:, :, 0:1], repeats=4, dim=2),
+                control_camera_video[:, :, 1:]
+            ], dim=2
+        ).transpose(1, 2)
+        b, f, c, h, w = control_camera_latents.shape
+        control_camera_latents = control_camera_latents.contiguous().view(b, f // 4, 4, c, h, w).transpose(2, 3)
+        control_camera_latents = control_camera_latents.contiguous().view(b, f // 4, c * 4, h, w).transpose(1, 2)
+        control_camera_latents_input = control_camera_latents.to(device=pipe.device, dtype=pipe.torch_dtype)
+
+        input_image = input_image.resize((width, height))
+        input_latents = pipe.preprocess_video([input_image])
+        pipe.load_models_to_device(self.onload_model_names)
+        input_latents = pipe.vae.encode(input_latents, device=pipe.device)
+        y = torch.zeros_like(latents).to(pipe.device)
+        y[:, :, :1] = input_latents
+        y = y.to(dtype=pipe.torch_dtype, device=pipe.device)
+        return {"control_camera_latents_input": control_camera_latents_input, "y": y}
+
+
+
+class WanVideoUnit_SpeedControl(PipelineUnit):
+    def __init__(self):
+        super().__init__(input_params=("motion_bucket_id",))
+
+    def process(self, pipe: WanVideoPipeline, motion_bucket_id):
+        if motion_bucket_id is None:
+            return {}
+        motion_bucket_id = torch.Tensor((motion_bucket_id,)).to(dtype=pipe.torch_dtype, device=pipe.device)
+        return {"motion_bucket_id": motion_bucket_id}
+
+
+
+class WanVideoUnit_VACE(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            input_params=("vace_video", "vace_video_mask", "vace_reference_image", "vace_scale", "height", "width", "num_frames", "tiled", "tile_size", "tile_stride"),
+            onload_model_names=("vae",)
+        )
+
+    def process(
+        self,
+        pipe: WanVideoPipeline,
+        vace_video, vace_video_mask, vace_reference_image, vace_scale,
+        height, width, num_frames,
+        tiled, tile_size, tile_stride
+    ):
+        if vace_video is not None or vace_video_mask is not None or vace_reference_image is not None:
+            pipe.load_models_to_device(["vae"])
+            if vace_video is None:
+                vace_video = torch.zeros((1, 3, num_frames, height, width), dtype=pipe.torch_dtype, device=pipe.device)
+            else:
+                vace_video = pipe.preprocess_video(vace_video)
+            
+            if vace_video_mask is None:
+                vace_video_mask = torch.ones_like(vace_video)
+            else:
+                vace_video_mask = pipe.preprocess_video(vace_video_mask, min_value=0, max_value=1)
+            
+            inactive = vace_video * (1 - vace_video_mask) + 0 * vace_video_mask
+            reactive = vace_video * vace_video_mask + 0 * (1 - vace_video_mask)
+            inactive = pipe.vae.encode(inactive, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+            reactive = pipe.vae.encode(reactive, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+            vace_video_latents = torch.concat((inactive, reactive), dim=1)
+            
+            vace_mask_latents = rearrange(vace_video_mask[0,0], "T (H P) (W Q) -> 1 (P Q) T H W", P=8, Q=8)
+            vace_mask_latents = torch.nn.functional.interpolate(vace_mask_latents, size=((vace_mask_latents.shape[2] + 3) // 4, vace_mask_latents.shape[3], vace_mask_latents.shape[4]), mode='nearest-exact')
+            
+            if vace_reference_image is None:
+                pass
+            else:
+                vace_reference_image = pipe.preprocess_video([vace_reference_image])
+                vace_reference_latents = pipe.vae.encode(vace_reference_image, device=pipe.device, tiled=tiled, tile_size=tile_size, tile_stride=tile_stride).to(dtype=pipe.torch_dtype, device=pipe.device)
+                vace_reference_latents = torch.concat((vace_reference_latents, torch.zeros_like(vace_reference_latents)), dim=1)
+                vace_video_latents = torch.concat((vace_reference_latents, vace_video_latents), dim=2)
+                vace_mask_latents = torch.concat((torch.zeros_like(vace_mask_latents[:, :, :1]), vace_mask_latents), dim=2)
+            
+            vace_context = torch.concat((vace_video_latents, vace_mask_latents), dim=1)
+            return {"vace_context": vace_context, "vace_scale": vace_scale}
+        else:
+            return {"vace_context": None, "vace_scale": vace_scale}
+
+
+
+class WanVideoUnit_UnifiedSequenceParallel(PipelineUnit):
+    def __init__(self):
+        super().__init__(input_params=())
+
+    def process(self, pipe: WanVideoPipeline):
+        if hasattr(pipe, "use_unified_sequence_parallel"):
+            if pipe.use_unified_sequence_parallel:
+                return {"use_unified_sequence_parallel": True}
+        return {}
+
+
+
+class WanVideoUnit_TeaCache(PipelineUnit):
+    def __init__(self):
+        super().__init__(
+            seperate_cfg=True,
+            input_params_posi={"num_inference_steps": "num_inference_steps", "tea_cache_l1_thresh": "tea_cache_l1_thresh", "tea_cache_model_id": "tea_cache_model_id"},
+            input_params_nega={"num_inference_steps": "num_inference_steps", "tea_cache_l1_thresh": "tea_cache_l1_thresh", "tea_cache_model_id": "tea_cache_model_id"},
+        )
+
+    def process(self, pipe: WanVideoPipeline, num_inference_steps, tea_cache_l1_thresh, tea_cache_model_id):
+        if tea_cache_l1_thresh is None:
+            return {}
+        return {"tea_cache": TeaCache(num_inference_steps, rel_l1_thresh=tea_cache_l1_thresh, model_id=tea_cache_model_id)}
+
+
+
+class WanVideoUnit_CfgMerger(PipelineUnit):
+    def __init__(self):
+        super().__init__(take_over=True)
+        self.concat_tensor_names = ["context", "clip_feature", "y", "reference_latents"]
+
+    def process(self, pipe: WanVideoPipeline, inputs_shared, inputs_posi, inputs_nega):
+        if not inputs_shared["cfg_merge"]:
+            return inputs_shared, inputs_posi, inputs_nega
+        for name in self.concat_tensor_names:
+            tensor_posi = inputs_posi.get(name)
+            tensor_nega = inputs_nega.get(name)
+            tensor_shared = inputs_shared.get(name)
+            if tensor_posi is not None and tensor_nega is not None:
+                inputs_shared[name] = torch.concat((tensor_posi, tensor_nega), dim=0)
+            elif tensor_shared is not None:
+                inputs_shared[name] = torch.concat((tensor_shared, tensor_shared), dim=0)
+        inputs_posi.clear()
+        inputs_nega.clear()
+        return inputs_shared, inputs_posi, inputs_nega
+
+
+
+class TeaCache:
+    def __init__(self, num_inference_steps, rel_l1_thresh, model_id):
+        self.num_inference_steps = num_inference_steps
+        self.step = 0
+        self.accumulated_rel_l1_distance = 0
+        self.previous_modulated_input = None
+        self.rel_l1_thresh = rel_l1_thresh
+        self.previous_residual = None
+        self.previous_hidden_states = None
+        
+        self.coefficients_dict = {
+            "Wan2.1-T2V-1.3B": [-5.21862437e+04, 9.23041404e+03, -5.28275948e+02, 1.36987616e+01, -4.99875664e-02],
+            "Wan2.1-T2V-14B": [-3.03318725e+05, 4.90537029e+04, -2.65530556e+03, 5.87365115e+01, -3.15583525e-01],
+            "Wan2.1-I2V-14B-480P": [2.57151496e+05, -3.54229917e+04,  1.40286849e+03, -1.35890334e+01, 1.32517977e-01],
+            "Wan2.1-I2V-14B-720P": [ 8.10705460e+03,  2.13393892e+03, -3.72934672e+02,  1.66203073e+01, -4.17769401e-02],
+        }
+        if model_id not in self.coefficients_dict:
+            supported_model_ids = ", ".join([i for i in self.coefficients_dict])
+            raise ValueError(f"{model_id} is not a supported TeaCache model id. Please choose a valid model id in ({supported_model_ids}).")
+        self.coefficients = self.coefficients_dict[model_id]
+
+    def check(self, dit: WanModel, x, t_mod):
+        modulated_inp = t_mod.clone()
+        if self.step == 0 or self.step == self.num_inference_steps - 1:
+            should_calc = True
+            self.accumulated_rel_l1_distance = 0
+        else:
+            coefficients = self.coefficients
+            rescale_func = np.poly1d(coefficients)
+            self.accumulated_rel_l1_distance += rescale_func(((modulated_inp-self.previous_modulated_input).abs().mean() / self.previous_modulated_input.abs().mean()).cpu().item())
+            if self.accumulated_rel_l1_distance < self.rel_l1_thresh:
+                should_calc = False
+            else:
+                should_calc = True
+                self.accumulated_rel_l1_distance = 0
+        self.previous_modulated_input = modulated_inp
+        self.step += 1
+        if self.step == self.num_inference_steps:
+            self.step = 0
+        if should_calc:
+            self.previous_hidden_states = x.clone()
+        return not should_calc
+
+    def store(self, hidden_states):
+        self.previous_residual = hidden_states - self.previous_hidden_states
+        self.previous_hidden_states = None
+
+    def update(self, hidden_states):
+        hidden_states = hidden_states + self.previous_residual
+        return hidden_states
+
+
+
+class TemporalTiler_BCTHW:
+    def __init__(self):
+        pass
+
+    def build_1d_mask(self, length, left_bound, right_bound, border_width):
+        x = torch.ones((length,))
+        if border_width == 0:
+            return x
+        
+        shift = 0.5
+        if not left_bound:
+            x[:border_width] = (torch.arange(border_width) + shift) / border_width
+        if not right_bound:
+            x[-border_width:] = torch.flip((torch.arange(border_width) + shift) / border_width, dims=(0,))
+        return x
+
+    def build_mask(self, data, is_bound, border_width):
+        _, _, T, _, _ = data.shape
+        t = self.build_1d_mask(T, is_bound[0], is_bound[1], border_width[0])
+        mask = repeat(t, "T -> 1 1 T 1 1")
+        return mask
+    
+    def run(self, model_fn, sliding_window_size, sliding_window_stride, computation_device, computation_dtype, model_kwargs, tensor_names, batch_size=None):
+        tensor_names = [tensor_name for tensor_name in tensor_names if model_kwargs.get(tensor_name) is not None]
+        tensor_dict = {tensor_name: model_kwargs[tensor_name] for tensor_name in tensor_names}
+
+        B, C, T, H, W = tensor_dict[tensor_names[0]].shape
+        if batch_size is not None:
+            B *= batch_size
+        data_device, data_dtype = tensor_dict[tensor_names[0]].device, tensor_dict[tensor_names[0]].dtype
+        value = torch.zeros((B, C, T, H, W), device=data_device, dtype=data_dtype)
+        weight = torch.zeros((1, 1, T, 1, 1), device=data_device, dtype=data_dtype)
+        for t in range(0, T, sliding_window_stride):
+            if t - sliding_window_stride >= 0 and t - sliding_window_stride + sliding_window_size >= T: #* 如果上一个窗口已经走到最后一帧了, 那么就continue/break
+                continue
+            t_ = min(t + sliding_window_size, T)
+            
+            model_kwargs.update({
+                tensor_name: tensor_dict[tensor_name][:, :, t: t_:, :].to(device=computation_device, dtype=computation_dtype) \
+                    for tensor_name in tensor_names
+            })
+            model_output = model_fn(**model_kwargs).to(device=data_device, dtype=data_dtype)
+            
+            mask = self.build_mask(
+                model_output,
+                is_bound=(t == 0, t_ == T),
+                border_width=(sliding_window_size - sliding_window_stride,)
+            ).to(device=data_device, dtype=data_dtype)
+
+            # logger.info(f"t: {t}, t_: {t_}, sliding_window_size: {sliding_window_size}, sliding_window_stride: {sliding_window_stride}")
+
+            value[:, :, t: t_, :, :] += model_output * mask
+            weight[:, :, t: t_, :, :] += mask
+        value /= weight
+        model_kwargs.update(tensor_dict)
+        return value
+
+
+
+def model_fn_wan_video(
+    dit: WanModel,
+    motion_controller: WanMotionControllerModel = None,
+    vace: VaceWanModel = None,
+    latents: torch.Tensor = None,
+    timestep: torch.Tensor = None,
+    context: torch.Tensor = None,
+    clip_feature: Optional[torch.Tensor] = None,
+    y: Optional[torch.Tensor] = None,
+    reference_latents = None,
+    vace_context = None,
+    vace_scale = 1.0,
+    tea_cache: TeaCache = None,
+    use_unified_sequence_parallel: bool = False,
+    motion_bucket_id: Optional[torch.Tensor] = None,
+    sliding_window_size: Optional[int] = None,
+    sliding_window_stride: Optional[int] = None,
+    cfg_merge: bool = False,
+    use_gradient_checkpointing: bool = False,
+    use_gradient_checkpointing_offload: bool = False,
+    control_camera_latents_input = None,
+    fuse_vae_embedding_in_latents: bool = False,
+    **kwargs,
+):
+    
+    
+    if sliding_window_size is not None and sliding_window_stride is not None: #* skip for training, 
+        model_kwargs = dict(
+            dit=dit,
+            motion_controller=motion_controller,
+            vace=vace,
+            latents=latents,
+            timestep=timestep,
+            context=context,
+            clip_feature=clip_feature,
+            y=y,
+            reference_latents=reference_latents,
+            vace_context=vace_context,
+            vace_scale=vace_scale,
+            tea_cache=tea_cache,
+            use_unified_sequence_parallel=use_unified_sequence_parallel,
+            motion_bucket_id=motion_bucket_id,
+        )
+        return TemporalTiler_BCTHW().run(
+            model_fn_wan_video,
+            sliding_window_size, sliding_window_stride,
+            latents.device, latents.dtype,
+            model_kwargs=model_kwargs,
+            tensor_names=["latents", "y"],
+            batch_size=2 if cfg_merge else 1
+        )
+    
+    if use_unified_sequence_parallel:#* skip
+        import torch.distributed as dist
+        from xfuser.core.distributed import (get_sequence_parallel_rank,
+                                            get_sequence_parallel_world_size,
+                                            get_sp_group)
+
+    
+    # Timestep
+    if dit.seperated_timestep and fuse_vae_embedding_in_latents: 
+        timestep = torch.concat([
+            torch.zeros((1, latents.shape[3] * latents.shape[4] // 4), dtype=latents.dtype, device=latents.device),
+            torch.ones((latents.shape[2] - 1, latents.shape[3] * latents.shape[4] // 4), dtype=latents.dtype, device=latents.device) * timestep
+        ]).flatten()
+        t = dit.time_embedding(sinusoidal_embedding_1d(dit.freq_dim, timestep).unsqueeze(0))
+        if use_unified_sequence_parallel and dist.is_initialized() and dist.get_world_size() > 1:
+            t_chunks = torch.chunk(t, get_sequence_parallel_world_size(), dim=1)
+            t_chunks = [torch.nn.functional.pad(chunk, (0, 0, 0, t_chunks[0].shape[1]-chunk.shape[1]), value=0) for chunk in t_chunks]
+            t = t_chunks[get_sequence_parallel_rank()]
+        t_mod = dit.time_projection(t).unflatten(2, (6, dit.dim))
+    else:#* this branch 
+        t = dit.time_embedding(sinusoidal_embedding_1d(dit.freq_dim, timestep)) #* out: torch.Size([1, 1536])
+        t_mod = dit.time_projection(t).unflatten(1, (6, dit.dim)) #* out: torch.Size([1, 6, 1536]); dit.dim: 1536
+    
+    
+
+    
+    if motion_bucket_id is not None and motion_controller is not None: #* skip
+        t_mod = t_mod + motion_controller(motion_bucket_id).unflatten(1, (6, dit.dim))
+    
+    context = dit.text_embedding(context)#* text prompt, 比如“depth”, : from torch.Size([1, 512, 4096]) to torch.Size([1, 512, 1536])
+    #todo double check 这个x 
+    #* [1, 16, (F-1)/4, H/8, W/8], 纯高斯噪声 或者 加噪后的gt
+    x = latents 
+    
+    # Merged cfg
+    #* batch 这个维度必须一致, 跟
+    if x.shape[0] != context.shape[0]:
+        x = torch.concat([x] * context.shape[0], dim=0)
+    if timestep.shape[0] != context.shape[0]:
+        timestep = torch.concat([timestep] * context.shape[0], dim=0)
+
+    # Image Embedding
+    """
+        new parameters: 
+            #* require_vae_embedding
+            #* require_clip_embedding
+    """
+    
+    # todo: x 是target video(也就是depth/normal video) 通过噪声调整的结果 / 纯高斯噪声; y是输入的rgb video 
+    #todo , double check 这个y, [1, 32, (F-1)/4, H/8, W/8]
+    if y is not None and dit.require_vae_embedding: 
+        x = torch.cat([x, y], dim=1)# (b, c_x + c_y, f, h, w) #* [1, 48, (F-1)/4, H/8, W/8]
+    if clip_feature is not None and dit.require_clip_embedding:
+        #* clip_feature is initialized by zero, from torch.Size([1, 257, 1280]) to torch.Size([1, 257, 1536])
+        clip_embdding = dit.img_emb(clip_feature) 
+        #*  concat 257 and 512 to form torch.Size([1, 769, 1536])
+        context = torch.cat([clip_embdding, context], dim=1)
+    
+    # Add camera control
+    #* from torch.Size([1, 48, (F-1)/4, H/8, W/8]),
+    #*  to [1, 1536, (F-1)/4, H/16, W/16] (函数内的mlp)
+    #*  to [1, 1536, ( (F-1)/4 * H/16 * W/16)]
+    #* x_out: [1, 1536, ( (F-1)/4 * H/16 * W/16)]
+    x, (f, h, w) = dit.patchify(x, control_camera_latents_input)
+    
+    # Reference image
+    if reference_latents is not None: #* skip
+        if len(reference_latents.shape) == 5:
+            reference_latents = reference_latents[:, :, 0]
+        reference_latents = dit.ref_conv(reference_latents).flatten(2).transpose(1, 2)
+        x = torch.concat([reference_latents, x], dim=1)
+        f += 1
+    
+    #* RoPE position embedding for 3D video,  [ ( (F-1)/4 * H/16 *  W/16), 1, 64]
+    freqs = torch.cat([
+        dit.freqs[0][:f].view(f, 1, 1, -1).expand(f, h, w, -1),
+        dit.freqs[1][:h].view(1, h, 1, -1).expand(f, h, w, -1),
+        dit.freqs[2][:w].view(1, 1, w, -1).expand(f, h, w, -1)
+    ], dim=-1).reshape(f * h * w, 1, -1).to(x.device)
+    
+    # TeaCache
+    if tea_cache is not None:#*skip
+        tea_cache_update = tea_cache.check(dit, x, t_mod)
+    else:
+        tea_cache_update = False
+        
+    if vace_context is not None:#*skip
+        vace_hints = vace(x, vace_context, context, t_mod, freqs)
+    
+    # blocks
+    if use_unified_sequence_parallel:#* skip
+        if dist.is_initialized() and dist.get_world_size() > 1:
+            chunks = torch.chunk(x, get_sequence_parallel_world_size(), dim=1)
+            pad_shape = chunks[0].shape[1] - chunks[-1].shape[1]
+            chunks = [torch.nn.functional.pad(chunk, (0, 0, 0, chunks[0].shape[1]-chunk.shape[1]), value=0) for chunk in chunks]
+            x = chunks[get_sequence_parallel_rank()]
+    if tea_cache_update:
+        x = tea_cache.update(x)
+    else:
+        def create_custom_forward(module):
+            def custom_forward(*inputs):
+                return module(*inputs)
+            return custom_forward
+        #* pass through  dit blocks 30 times
+        for block_id, block in enumerate(dit.blocks):
+            if use_gradient_checkpointing_offload:
+                with torch.autograd.graph.save_on_cpu():
+                    x = torch.utils.checkpoint.checkpoint(
+                        create_custom_forward(block),
+                        x, context, t_mod, freqs,
+                        use_reentrant=False,
+                    )
+            elif use_gradient_checkpointing:
+                x = torch.utils.checkpoint.checkpoint(
+                    create_custom_forward(block),
+                    x, context, t_mod, freqs,
+                    use_reentrant=False,
+                )
+            else:
+                x = block(x, context, t_mod, freqs)#* x_in: [1, ( (F-1)/4 * H/16 *  W/16), 1536], context_in: [1, 769, 1536], t_mod_in: [1, 6, 1536], freqs_in: [ ( (F-1)/4 * H/16 *  W/16), 1, 64], x_out: [1, ( (F-1)/4 * H/16 *  W/16), 1536]
+            if vace_context is not None and block_id in vace.vace_layers_mapping:#* skip
+                current_vace_hint = vace_hints[vace.vace_layers_mapping[block_id]]
+                if use_unified_sequence_parallel and dist.is_initialized() and dist.get_world_size() > 1:
+                    current_vace_hint = torch.chunk(current_vace_hint, get_sequence_parallel_world_size(), dim=1)[get_sequence_parallel_rank()]
+                    current_vace_hint = torch.nn.functional.pad(current_vace_hint, (0, 0, 0, chunks[0].shape[1] - current_vace_hint.shape[1]), value=0)
+                x = x + current_vace_hint * vace_scale
+        if tea_cache is not None:#* skip
+            tea_cache.store(x)
+    
+    #* x_in: [1, ( (F-1)/4 * H/16 *  W/16), 1536], t_in: [1, 1536], 
+    #* x_out: [1, ( (F-1)/4 * H/16 *  W/16), 64]
+    x = dit.head(x, t)
+    if use_unified_sequence_parallel:#* skip
+        if dist.is_initialized() and dist.get_world_size() > 1:
+            x = get_sp_group().all_gather(x, dim=1)
+            x = x[:, :-pad_shape] if pad_shape > 0 else x
+    
+    # Remove reference latents
+    if reference_latents is not None:#* skip
+        x = x[:, reference_latents.shape[1]:]
+        f -= 1
+
+    #* unpatchify, from [1, ( (F-1)/4 * H/16 *  W/16), 64] to [1, 16, (F-1)/4, H/8, W/8]
+    x = dit.unpatchify(x, (f, h, w))
+    return x
+
+
+

dkt/prompters/__init__.py

@@ -0,0 +1,13 @@
+# Only wan_prompter is used by entry files
+# from .prompt_refiners import Translator, BeautifulPrompt, QwenPrompt
+# from .sd_prompter import SDPrompter
+# from .sdxl_prompter import SDXLPrompter
+# from .sd3_prompter import SD3Prompter
+# from .hunyuan_dit_prompter import HunyuanDiTPrompter
+# from .kolors_prompter import KolorsPrompter
+# from .flux_prompter import FluxPrompter
+# from .omost import OmostPromter
+# from .cog_prompter import CogPrompter
+# from .hunyuan_video_prompter import HunyuanVideoPrompter
+# from .stepvideo_prompter import StepVideoPrompter
+from .wan_prompter import WanPrompter

dkt/prompters/base_prompter.py

@@ -0,0 +1,70 @@
+from ..models.model_manager import ModelManager
+import torch
+
+
+
+def tokenize_long_prompt(tokenizer, prompt, max_length=None):
+    # Get model_max_length from self.tokenizer
+    length = tokenizer.model_max_length if max_length is None else max_length
+
+    # To avoid the warning. set self.tokenizer.model_max_length to +oo.
+    tokenizer.model_max_length = 99999999
+
+    # Tokenize it!
+    input_ids = tokenizer(prompt, return_tensors="pt").input_ids
+
+    # Determine the real length.
+    max_length = (input_ids.shape[1] + length - 1) // length * length
+
+    # Restore tokenizer.model_max_length
+    tokenizer.model_max_length = length
+    
+    # Tokenize it again with fixed length.
+    input_ids = tokenizer(
+        prompt,
+        return_tensors="pt",
+        padding="max_length",
+        max_length=max_length,
+        truncation=True
+    ).input_ids
+
+    # Reshape input_ids to fit the text encoder.
+    num_sentence = input_ids.shape[1] // length
+    input_ids = input_ids.reshape((num_sentence, length))
+    
+    return input_ids
+
+
+
+class BasePrompter:
+    def __init__(self):
+        self.refiners = []
+        self.extenders = []
+
+
+    def load_prompt_refiners(self, model_manager: ModelManager, refiner_classes=[]):
+        for refiner_class in refiner_classes:
+            refiner = refiner_class.from_model_manager(model_manager)
+            self.refiners.append(refiner)
+    
+    def load_prompt_extenders(self,model_manager:ModelManager,extender_classes=[]):
+        for extender_class in extender_classes:
+            extender = extender_class.from_model_manager(model_manager)
+            self.extenders.append(extender)
+
+
+    @torch.no_grad()
+    def process_prompt(self, prompt, positive=True):
+        if isinstance(prompt, list):
+            prompt = [self.process_prompt(prompt_, positive=positive) for prompt_ in prompt]
+        else:
+            for refiner in self.refiners:
+                prompt = refiner(prompt, positive=positive)
+        return prompt
+
+    @torch.no_grad()
+    def extend_prompt(self, prompt:str, positive=True):
+        extended_prompt = dict(prompt=prompt)
+        for extender in self.extenders:
+            extended_prompt = extender(extended_prompt)
+        return extended_prompt

dkt/prompters/wan_prompter.py

@@ -0,0 +1,109 @@
+from .base_prompter import BasePrompter
+from ..models.wan_video_text_encoder import WanTextEncoder
+from transformers import AutoTokenizer
+import os, torch
+import ftfy
+import html
+import string
+import regex as re
+
+
+def basic_clean(text):
+    text = ftfy.fix_text(text)
+    text = html.unescape(html.unescape(text))
+    return text.strip()
+
+
+def whitespace_clean(text):
+    text = re.sub(r'\s+', ' ', text)
+    text = text.strip()
+    return text
+
+
+def canonicalize(text, keep_punctuation_exact_string=None):
+    text = text.replace('_', ' ')
+    if keep_punctuation_exact_string:
+        text = keep_punctuation_exact_string.join(
+            part.translate(str.maketrans('', '', string.punctuation))
+            for part in text.split(keep_punctuation_exact_string))
+    else:
+        text = text.translate(str.maketrans('', '', string.punctuation))
+    text = text.lower()
+    text = re.sub(r'\s+', ' ', text)
+    return text.strip()
+
+
+class HuggingfaceTokenizer:
+
+    def __init__(self, name, seq_len=None, clean=None, **kwargs):
+        assert clean in (None, 'whitespace', 'lower', 'canonicalize')
+        self.name = name
+        self.seq_len = seq_len
+        self.clean = clean
+
+        # init tokenizer
+        self.tokenizer = AutoTokenizer.from_pretrained(name, **kwargs)
+        self.vocab_size = self.tokenizer.vocab_size
+
+    def __call__(self, sequence, **kwargs):
+        return_mask = kwargs.pop('return_mask', False)
+
+        # arguments
+        _kwargs = {'return_tensors': 'pt'}
+        if self.seq_len is not None:
+            _kwargs.update({
+                'padding': 'max_length',
+                'truncation': True,
+                'max_length': self.seq_len
+            })
+        _kwargs.update(**kwargs)
+
+        # tokenization
+        if isinstance(sequence, str):
+            sequence = [sequence]
+        if self.clean:
+            sequence = [self._clean(u) for u in sequence]
+        ids = self.tokenizer(sequence, **_kwargs)
+
+        # output
+        if return_mask:
+            return ids.input_ids, ids.attention_mask
+        else:
+            return ids.input_ids
+
+    def _clean(self, text):
+        if self.clean == 'whitespace':
+            text = whitespace_clean(basic_clean(text))
+        elif self.clean == 'lower':
+            text = whitespace_clean(basic_clean(text)).lower()
+        elif self.clean == 'canonicalize':
+            text = canonicalize(basic_clean(text))
+        return text
+
+
+class WanPrompter(BasePrompter):
+
+    def __init__(self, tokenizer_path=None, text_len=512):
+        super().__init__()
+        self.text_len = text_len
+        self.text_encoder = None
+        self.fetch_tokenizer(tokenizer_path)
+        
+    def fetch_tokenizer(self, tokenizer_path=None):
+        if tokenizer_path is not None:
+            self.tokenizer = HuggingfaceTokenizer(name=tokenizer_path, seq_len=self.text_len, clean='whitespace')
+
+    def fetch_models(self, text_encoder: WanTextEncoder = None):
+        self.text_encoder = text_encoder
+
+    def encode_prompt(self, prompt, positive=True, device="cuda"):
+        prompt = self.process_prompt(prompt, positive=positive)
+        
+        ids, mask = self.tokenizer(prompt, return_mask=True, add_special_tokens=True)
+        ids = ids.to(device)
+        mask = mask.to(device)
+        seq_lens = mask.gt(0).sum(dim=1).long()
+        prompt_emb = self.text_encoder(ids, mask)
+        for i, v in enumerate(seq_lens):
+            prompt_emb[:, v:] = 0
+        return prompt_emb

dkt/schedulers/__init__.py

@@ -0,0 +1 @@
+from .flow_match import FlowMatchScheduler

dkt/schedulers/flow_match.py

@@ -0,0 +1,126 @@
+import torch, math
+
+
+
+class FlowMatchScheduler():
+
+    def __init__(
+        self,
+        num_inference_steps=100,
+        num_train_timesteps=1000,
+        shift=3.0,
+        sigma_max=1.0,
+        sigma_min=0.003/1.002,
+        inverse_timesteps=False,
+        extra_one_step=False,
+        reverse_sigmas=False,
+        exponential_shift=False,
+        exponential_shift_mu=None,
+        shift_terminal=None,
+    ):
+        self.num_train_timesteps = num_train_timesteps
+        self.shift = shift
+        self.sigma_max = sigma_max
+        self.sigma_min = sigma_min
+        self.inverse_timesteps = inverse_timesteps
+        self.extra_one_step = extra_one_step
+        self.reverse_sigmas = reverse_sigmas
+        self.exponential_shift = exponential_shift
+        self.exponential_shift_mu = exponential_shift_mu
+        self.shift_terminal = shift_terminal
+        self.set_timesteps(num_inference_steps)
+
+
+    def set_timesteps(self, num_inference_steps=100, denoising_strength=1.0, training=False, shift=None, dynamic_shift_len=None):
+        if shift is not None:
+            self.shift = shift
+        sigma_start = self.sigma_min + (self.sigma_max - self.sigma_min) * denoising_strength
+        if self.extra_one_step:
+            self.sigmas = torch.linspace(sigma_start, self.sigma_min, num_inference_steps + 1)[:-1]
+        else:
+            self.sigmas = torch.linspace(sigma_start, self.sigma_min, num_inference_steps)
+
+        if self.inverse_timesteps:
+            self.sigmas = torch.flip(self.sigmas, dims=[0])
+        if self.exponential_shift:
+            mu = self.calculate_shift(dynamic_shift_len) if dynamic_shift_len is not None else self.exponential_shift_mu
+            self.sigmas = math.exp(mu) / (math.exp(mu) + (1 / self.sigmas - 1))
+        else:
+            self.sigmas = self.shift * self.sigmas / (1 + (self.shift - 1) * self.sigmas)
+        if self.shift_terminal is not None:
+            one_minus_z = 1 - self.sigmas
+            scale_factor = one_minus_z[-1] / (1 - self.shift_terminal)
+            self.sigmas = 1 - (one_minus_z / scale_factor)
+        if self.reverse_sigmas:
+            self.sigmas = 1 - self.sigmas
+        self.timesteps = self.sigmas * self.num_train_timesteps
+        
+        if training:
+            x = self.timesteps
+            y = torch.exp(-2 * ((x - num_inference_steps / 2) / num_inference_steps) ** 2)
+            y_shifted = y - y.min()
+            bsmntw_weighing = y_shifted * (num_inference_steps / y_shifted.sum())
+            self.linear_timesteps_weights = bsmntw_weighing
+            self.training = True
+        else:
+            self.training = False
+
+
+    def step(self, model_output, timestep, sample, to_final=False, **kwargs):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.cpu()
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        
+        sigma = self.sigmas[timestep_id] #* 当前的sigma值, 也就是加了多少比例的噪声
+        if to_final or timestep_id + 1 >= len(self.timesteps): #* 下一步的噪声比例,  
+            sigma_ = 1 if (self.inverse_timesteps or self.reverse_sigmas) else 0
+        else:
+            sigma_ = self.sigmas[timestep_id + 1]
+        prev_sample = sample + model_output * (sigma_ - sigma)
+        return prev_sample
+    
+
+    def return_to_timestep(self, timestep, sample, sample_stablized):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.cpu()
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        model_output = (sample - sample_stablized) / sigma
+        return model_output
+    
+    
+    def add_noise(self, original_samples, noise, timestep):
+        if isinstance(timestep, torch.Tensor):
+            timestep = timestep.cpu()
+        timestep_id = torch.argmin((self.timesteps - timestep).abs())
+        sigma = self.sigmas[timestep_id]
+        sample = (1 - sigma) * original_samples + sigma * noise
+        return sample
+    
+
+    def training_target(self, sample, noise, timestep):
+        #* so: noise - target = sample 
+        #* sample + target = noise
+        target = noise - sample 
+        #* noise: is rgb images
+        return target
+    
+
+    def training_weight(self, timestep):
+        timestep_id = torch.argmin((self.timesteps - timestep.to(self.timesteps.device)).abs())
+        weights = self.linear_timesteps_weights[timestep_id]
+        return weights
+    
+    
+    def calculate_shift(
+        self,
+        image_seq_len,
+        base_seq_len: int = 256,
+        max_seq_len: int = 8192,
+        base_shift: float = 0.5,
+        max_shift: float = 0.9,
+    ):
+        m = (max_shift - base_shift) / (max_seq_len - base_seq_len)
+        b = base_shift - m * base_seq_len
+        mu = image_seq_len * m + b
+        return mu

dkt/utils/__init__.py

@@ -0,0 +1,261 @@
+import torch, warnings, glob, os
+import numpy as np
+from PIL import Image
+from einops import repeat, reduce
+from typing import Optional, Union
+from dataclasses import dataclass
+from modelscope import snapshot_download
+import numpy as np
+from PIL import Image
+from typing import Optional
+
+
+class BasePipeline(torch.nn.Module):
+
+    def __init__(
+        self,
+        device="cuda", torch_dtype=torch.float16,
+        height_division_factor=64, width_division_factor=64,
+        time_division_factor=None, time_division_remainder=None,
+    ):
+        super().__init__()
+        # The device and torch_dtype is used for the storage of intermediate variables, not models.
+        self.device = device
+        self.torch_dtype = torch_dtype
+        # The following parameters are used for shape check.
+        self.height_division_factor = height_division_factor
+        self.width_division_factor = width_division_factor
+        self.time_division_factor = time_division_factor
+        self.time_division_remainder = time_division_remainder
+        self.vram_management_enabled = False
+        
+        
+    def to(self, *args, **kwargs):
+        device, dtype, non_blocking, convert_to_format = torch._C._nn._parse_to(*args, **kwargs)
+        if device is not None:
+            self.device = device
+        if dtype is not None:
+            self.torch_dtype = dtype
+        super().to(*args, **kwargs)
+        return self
+
+
+    def check_resize_height_width(self, height, width, num_frames=None):
+        # Shape check
+        if height % self.height_division_factor != 0:
+            height = (height + self.height_division_factor - 1) // self.height_division_factor * self.height_division_factor
+            print(f"height % {self.height_division_factor} != 0. We round it up to {height}.")
+        if width % self.width_division_factor != 0:
+            width = (width + self.width_division_factor - 1) // self.width_division_factor * self.width_division_factor
+            print(f"width % {self.width_division_factor} != 0. We round it up to {width}.")
+        if num_frames is None:
+            return height, width
+        else:
+            if num_frames % self.time_division_factor != self.time_division_remainder:
+                num_frames = (num_frames + self.time_division_factor - 1) // self.time_division_factor * self.time_division_factor + self.time_division_remainder
+                print(f"num_frames % {self.time_division_factor} != {self.time_division_remainder}. We round it up to {num_frames}.")
+            return height, width, num_frames
+
+
+    def preprocess_image(self, image, torch_dtype=None, device=None, pattern="B C H W", min_value=-1, max_value=1):
+        # Transform a PIL.Image to torch.Tensor
+        image = torch.Tensor(np.array(image, dtype=np.float32))
+        image = image.to(dtype=torch_dtype or self.torch_dtype, device=device or self.device)
+        image = image * ((max_value - min_value) / 255) + min_value
+        image = repeat(image, f"H W C -> {pattern}", **({"B": 1} if "B" in pattern else {}))
+        return image
+
+
+    def preprocess_video(self, video, torch_dtype=None, device=None, pattern="B C T H W", min_value=-1, max_value=1):
+        # Transform a list of PIL.Image to torch.Tensor
+        video = [self.preprocess_image(image, torch_dtype=torch_dtype, device=device, min_value=min_value, max_value=max_value) for image in video]
+        video = torch.stack(video, dim=pattern.index("T") // 2)
+        return video
+
+
+    def vae_output_to_image(self, vae_output, pattern="B C H W", min_value=-1, max_value=1):
+        # Transform a torch.Tensor to PIL.Image
+        if pattern != "H W C":
+            vae_output = reduce(vae_output, f"{pattern} -> H W C", reduction="mean")
+        image = ((vae_output - min_value) * (255 / (max_value - min_value))).clip(0, 255)
+        image = image.to(device="cpu", dtype=torch.uint8)
+        image = Image.fromarray(image.numpy())
+        return image
+
+
+    def vae_output_to_video(self, vae_output, pattern="B C T H W", min_value=-1, max_value=1):
+        # Transform a torch.Tensor to list of PIL.Image
+        if pattern != "T H W C":
+            vae_output = reduce(vae_output, f"{pattern} -> T H W C", reduction="mean")
+        video = [self.vae_output_to_image(image, pattern="H W C", min_value=min_value, max_value=max_value) for image in vae_output]
+        return video
+
+
+    def load_models_to_device(self, model_names=[]):
+        if self.vram_management_enabled:
+            # offload models
+            for name, model in self.named_children():
+                if name not in model_names:
+                    if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
+                        for module in model.modules():
+                            if hasattr(module, "offload"):
+                                module.offload()
+                    else:
+                        model.cpu()
+            torch.cuda.empty_cache()
+            # onload models
+            for name, model in self.named_children():
+                if name in model_names:
+                    if hasattr(model, "vram_management_enabled") and model.vram_management_enabled:
+                        for module in model.modules():
+                            if hasattr(module, "onload"):
+                                module.onload()
+                    else:
+                        model.to(self.device)
+
+
+    def generate_noise(self, shape, seed=None, rand_device="cpu", rand_torch_dtype=torch.float32, device=None, torch_dtype=None):
+        # Initialize Gaussian noise
+        generator = None if seed is None else torch.Generator(rand_device).manual_seed(seed)
+        noise = torch.randn(shape, generator=generator, device=rand_device, dtype=rand_torch_dtype)
+        noise = noise.to(dtype=torch_dtype or self.torch_dtype, device=device or self.device)
+        return noise
+
+
+    def enable_cpu_offload(self):
+        warnings.warn("`enable_cpu_offload` will be deprecated. Please use `enable_vram_management`.")
+        self.vram_management_enabled = True
+        
+        
+    def get_vram(self):
+        return torch.cuda.mem_get_info(self.device)[1] / (1024 ** 3)
+    
+    
+    def freeze_except(self, model_names):
+        for name, model in self.named_children():
+            if name in model_names:
+                model.train()
+                model.requires_grad_(True)
+            else:
+                model.eval()
+                model.requires_grad_(False)
+
+
+@dataclass
+class ModelConfig:
+    path: Union[str, list[str]] = None
+    model_id: str = None
+    origin_file_pattern: Union[str, list[str]] = None
+    download_resource: str = "ModelScope"
+    offload_device: Optional[Union[str, torch.device]] = None
+    offload_dtype: Optional[torch.dtype] = None
+    local_model_path: str = None
+    skip_download: bool = False
+
+    def download_if_necessary(self, use_usp=False):
+        if self.path is None:
+            # Check model_id and origin_file_pattern
+            if self.model_id is None:
+                raise ValueError(f"""No valid model files. Please use `ModelConfig(path="xxx")` or `ModelConfig(model_id="xxx/yyy", origin_file_pattern="zzz")`.""")
+            
+            # Skip if not in rank 0
+            if use_usp:
+                import torch.distributed as dist
+                skip_download = self.skip_download or dist.get_rank() != 0
+            else:
+                skip_download = self.skip_download
+                
+            # Check whether the origin path is a folder
+            if self.origin_file_pattern is None or self.origin_file_pattern == "":
+                self.origin_file_pattern = ""
+                allow_file_pattern = None
+                is_folder = True
+            elif isinstance(self.origin_file_pattern, str) and self.origin_file_pattern.endswith("/"):
+                allow_file_pattern = self.origin_file_pattern + "*"
+                is_folder = True
+            else:
+                allow_file_pattern = self.origin_file_pattern
+                is_folder = False
+            
+            # Download
+            if self.local_model_path is None:
+                self.local_model_path = "./models"
+            if not skip_download:
+                downloaded_files = glob.glob(self.origin_file_pattern, root_dir=os.path.join(self.local_model_path, self.model_id))
+                snapshot_download(
+                    self.model_id,
+                    local_dir=os.path.join(self.local_model_path, self.model_id),
+                    allow_file_pattern=allow_file_pattern,
+                    ignore_file_pattern=downloaded_files,
+                    local_files_only=False
+                )
+            
+            # Let rank 1, 2, ... wait for rank 0
+            if use_usp:
+                import torch.distributed as dist
+                dist.barrier(device_ids=[dist.get_rank()])
+                
+            # Return downloaded files
+            if is_folder:
+                self.path = os.path.join(self.local_model_path, self.model_id, self.origin_file_pattern)
+            else:
+                self.path = glob.glob(os.path.join(self.local_model_path, self.model_id, self.origin_file_pattern))
+            if isinstance(self.path, list) and len(self.path) == 1:
+                self.path = self.path[0]
+
+
+
+class PipelineUnit:
+    def __init__(
+        self,
+        seperate_cfg: bool = False,
+        take_over: bool = False,
+        input_params: tuple[str] = None,
+        input_params_posi: dict[str, str] = None,
+        input_params_nega: dict[str, str] = None,
+        onload_model_names: tuple[str] = None
+    ):
+        self.seperate_cfg = seperate_cfg
+        self.take_over = take_over
+        self.input_params = input_params
+        self.input_params_posi = input_params_posi
+        self.input_params_nega = input_params_nega
+        self.onload_model_names = onload_model_names
+
+
+    def process(self, pipe: BasePipeline, inputs: dict, positive=True, **kwargs) -> dict:
+        raise NotImplementedError("`process` is not implemented.")
+
+
+
+class PipelineUnitRunner:
+    def __init__(self):
+        pass
+
+    def __call__(self, unit: PipelineUnit, pipe: BasePipeline, inputs_shared: dict, inputs_posi: dict, inputs_nega: dict) -> tuple[dict, dict]:
+        if unit.take_over:
+            # Let the pipeline unit take over this function.
+            inputs_shared, inputs_posi, inputs_nega = unit.process(pipe, inputs_shared=inputs_shared, inputs_posi=inputs_posi, inputs_nega=inputs_nega)
+        elif unit.seperate_cfg:
+            # Positive side
+            processor_inputs = {name: inputs_posi.get(name_) for name, name_ in unit.input_params_posi.items()}
+            if unit.input_params is not None:
+                for name in unit.input_params:
+                    processor_inputs[name] = inputs_shared.get(name)
+            processor_outputs = unit.process(pipe, **processor_inputs)
+            inputs_posi.update(processor_outputs)
+            # Negative side
+            if inputs_shared["cfg_scale"] != 1:
+                processor_inputs = {name: inputs_nega.get(name_) for name, name_ in unit.input_params_nega.items()}
+                if unit.input_params is not None:
+                    for name in unit.input_params:
+                        processor_inputs[name] = inputs_shared.get(name)
+                processor_outputs = unit.process(pipe, **processor_inputs)
+                inputs_nega.update(processor_outputs)
+            else:
+                inputs_nega.update(processor_outputs)
+        else:
+            processor_inputs = {name: inputs_shared.get(name) for name in unit.input_params}
+            processor_outputs = unit.process(pipe, **processor_inputs)
+            inputs_shared.update(processor_outputs)
+        return inputs_shared, inputs_posi, inputs_nega

dkt/vram_management/__init__.py

@@ -0,0 +1,2 @@
+from .layers import *
+from .gradient_checkpointing import *

dkt/vram_management/gradient_checkpointing.py

@@ -0,0 +1,34 @@
+import torch
+
+
+def create_custom_forward(module):
+    def custom_forward(*inputs, **kwargs):
+        return module(*inputs, **kwargs)
+    return custom_forward
+
+
+def gradient_checkpoint_forward(
+    model,
+    use_gradient_checkpointing,
+    use_gradient_checkpointing_offload,
+    *args,
+    **kwargs,
+):
+    if use_gradient_checkpointing_offload:
+        with torch.autograd.graph.save_on_cpu():
+            model_output = torch.utils.checkpoint.checkpoint(
+                create_custom_forward(model),
+                *args,
+                **kwargs,
+                use_reentrant=False,
+            )
+    elif use_gradient_checkpointing:
+        model_output = torch.utils.checkpoint.checkpoint(
+            create_custom_forward(model),
+            *args,
+            **kwargs,
+            use_reentrant=False,
+        )
+    else:
+        model_output = model(*args, **kwargs)
+    return model_output

dkt/vram_management/layers.py

@@ -0,0 +1,213 @@
+import torch, copy
+from typing import Union
+from ..models.utils import init_weights_on_device
+
+
+def cast_to(weight, dtype, device):
+    r = torch.empty_like(weight, dtype=dtype, device=device)
+    r.copy_(weight)
+    return r
+
+
+class AutoTorchModule(torch.nn.Module):
+    def __init__(self):
+        super().__init__()
+        
+    def check_free_vram(self):
+        gpu_mem_state = torch.cuda.mem_get_info(self.computation_device)
+        used_memory = (gpu_mem_state[1] - gpu_mem_state[0]) / (1024 ** 3)
+        return used_memory < self.vram_limit
+
+    def offload(self):
+        if self.state != 0:
+            self.to(dtype=self.offload_dtype, device=self.offload_device)
+            self.state = 0
+
+    def onload(self):
+        if self.state != 1:
+            self.to(dtype=self.onload_dtype, device=self.onload_device)
+            self.state = 1
+            
+    def keep(self):
+        if self.state != 2:
+            self.to(dtype=self.computation_dtype, device=self.computation_device)
+            self.state = 2
+
+
+class AutoWrappedModule(AutoTorchModule):
+    def __init__(self, module: torch.nn.Module, offload_dtype, offload_device, onload_dtype, onload_device, computation_dtype, computation_device, vram_limit, **kwargs):
+        super().__init__()
+        self.module = module.to(dtype=offload_dtype, device=offload_device)
+        self.offload_dtype = offload_dtype
+        self.offload_device = offload_device
+        self.onload_dtype = onload_dtype
+        self.onload_device = onload_device
+        self.computation_dtype = computation_dtype
+        self.computation_device = computation_device
+        self.vram_limit = vram_limit
+        self.state = 0
+
+    def forward(self, *args, **kwargs):
+        if self.state == 2:
+            module = self.module
+        else:
+            if self.onload_dtype == self.computation_dtype and self.onload_device == self.computation_device:
+                module = self.module
+            elif self.vram_limit is not None and self.check_free_vram():
+                self.keep()
+                module = self.module
+            else:
+                module = copy.deepcopy(self.module).to(dtype=self.computation_dtype, device=self.computation_device)
+        return module(*args, **kwargs)
+    
+
+class WanAutoCastLayerNorm(torch.nn.LayerNorm, AutoTorchModule):
+    def __init__(self, module: torch.nn.LayerNorm, offload_dtype, offload_device, onload_dtype, onload_device, computation_dtype, computation_device, vram_limit, **kwargs):
+        with init_weights_on_device(device=torch.device("meta")):
+            super().__init__(module.normalized_shape, eps=module.eps, elementwise_affine=module.elementwise_affine, bias=module.bias is not None, dtype=offload_dtype, device=offload_device)
+        self.weight = module.weight
+        self.bias = module.bias
+        self.offload_dtype = offload_dtype
+        self.offload_device = offload_device
+        self.onload_dtype = onload_dtype
+        self.onload_device = onload_device
+        self.computation_dtype = computation_dtype
+        self.computation_device = computation_device
+        self.vram_limit = vram_limit
+        self.state = 0
+
+    def forward(self, x, *args, **kwargs):
+        if self.state == 2:
+            weight, bias = self.weight, self.bias
+        else:
+            if self.onload_dtype == self.computation_dtype and self.onload_device == self.computation_device:
+                weight, bias = self.weight, self.bias
+            elif self.vram_limit is not None and self.check_free_vram():
+                self.keep()
+                weight, bias = self.weight, self.bias
+            else:
+                weight = None if self.weight is None else cast_to(self.weight, self.computation_dtype, self.computation_device)
+                bias = None if self.bias is None else cast_to(self.bias, self.computation_dtype, self.computation_device)
+        with torch.amp.autocast(device_type=x.device.type):
+            x = torch.nn.functional.layer_norm(x.float(), self.normalized_shape, weight, bias, self.eps).type_as(x)
+        return x
+    
+
+class AutoWrappedLinear(torch.nn.Linear, AutoTorchModule):
+    def __init__(self, module: torch.nn.Linear, offload_dtype, offload_device, onload_dtype, onload_device, computation_dtype, computation_device, vram_limit, name="", **kwargs):
+        with init_weights_on_device(device=torch.device("meta")):
+            super().__init__(in_features=module.in_features, out_features=module.out_features, bias=module.bias is not None, dtype=offload_dtype, device=offload_device)
+        self.weight = module.weight
+        self.bias = module.bias
+        self.offload_dtype = offload_dtype
+        self.offload_device = offload_device
+        self.onload_dtype = onload_dtype
+        self.onload_device = onload_device
+        self.computation_dtype = computation_dtype
+        self.computation_device = computation_device
+        self.vram_limit = vram_limit
+        self.state = 0
+        self.name = name
+        self.lora_A_weights = []
+        self.lora_B_weights = []
+        self.lora_merger = None
+        self.enable_fp8 = computation_dtype in [torch.float8_e4m3fn, torch.float8_e4m3fnuz]
+        
+    def fp8_linear(
+        self,
+        input: torch.Tensor,
+        weight: torch.Tensor,
+        bias: Union[torch.Tensor, None] = None):
+        device = input.device
+        origin_dtype = input.dtype
+        origin_shape = input.shape
+        input = input.reshape(-1, origin_shape[-1])
+
+        x_max = torch.max(torch.abs(input), dim=-1, keepdim=True).values
+        fp8_max = 448.0
+        # For float8_e4m3fnuz, the maximum representable value is half of that of e4m3fn.
+        # To avoid overflow and ensure numerical compatibility during FP8 computation,
+        # we scale down the input by 2.0 in advance.
+        # This scaling will be compensated later during the final result scaling.
+        if self.computation_dtype == torch.float8_e4m3fnuz:
+            fp8_max = fp8_max / 2.0
+        scale_a = torch.clamp(x_max / fp8_max, min=1.0).float().to(device=device)
+        scale_b = torch.ones((weight.shape[0], 1)).to(device=device)
+        input = input / (scale_a + 1e-8)
+        input = input.to(self.computation_dtype)
+        weight = weight.to(self.computation_dtype)
+        bias = bias.to(torch.bfloat16)
+
+        result = torch._scaled_mm(
+            input,
+            weight.T,
+            scale_a=scale_a,
+            scale_b=scale_b.T,
+            bias=bias,
+            out_dtype=origin_dtype,
+        )
+        new_shape = origin_shape[:-1] + result.shape[-1:]
+        result = result.reshape(new_shape)
+        return result
+
+    def forward(self, x, *args, **kwargs):
+        # VRAM management
+        if self.state == 2:
+            weight, bias = self.weight, self.bias
+        else:
+            if self.onload_dtype == self.computation_dtype and self.onload_device == self.computation_device:
+                weight, bias = self.weight, self.bias
+            elif self.vram_limit is not None and self.check_free_vram():
+                self.keep()
+                weight, bias = self.weight, self.bias
+            else:
+                weight = cast_to(self.weight, self.computation_dtype, self.computation_device)
+                bias = None if self.bias is None else cast_to(self.bias, self.computation_dtype, self.computation_device)
+        
+        # Linear forward
+        if self.enable_fp8:
+            out = self.fp8_linear(x, weight, bias)
+        else:
+            out = torch.nn.functional.linear(x, weight, bias)
+        
+        # LoRA
+        if len(self.lora_A_weights) == 0:
+            # No LoRA
+            return out
+        elif self.lora_merger is None:
+            # Native LoRA inference
+            for lora_A, lora_B in zip(self.lora_A_weights, self.lora_B_weights):
+                out = out + x @ lora_A.T @ lora_B.T
+        else:
+            # LoRA fusion
+            lora_output = []
+            for lora_A, lora_B in zip(self.lora_A_weights, self.lora_B_weights):
+                lora_output.append(x @ lora_A.T @ lora_B.T)
+            lora_output = torch.stack(lora_output)
+            out = self.lora_merger(out, lora_output)
+        return out
+
+
+def enable_vram_management_recursively(model: torch.nn.Module, module_map: dict, module_config: dict, max_num_param=None, overflow_module_config: dict = None, total_num_param=0, vram_limit=None, name_prefix=""):
+    for name, module in model.named_children():
+        layer_name = name if name_prefix == "" else name_prefix + "." + name
+        for source_module, target_module in module_map.items():
+            if isinstance(module, source_module):
+                num_param = sum(p.numel() for p in module.parameters())
+                if max_num_param is not None and total_num_param + num_param > max_num_param:
+                    module_config_ = overflow_module_config
+                else:
+                    module_config_ = module_config
+                module_ = target_module(module, **module_config_, vram_limit=vram_limit, name=layer_name)
+                setattr(model, name, module_)
+                total_num_param += num_param
+                break
+        else:
+            total_num_param = enable_vram_management_recursively(module, module_map, module_config, max_num_param, overflow_module_config, total_num_param, vram_limit=vram_limit, name_prefix=layer_name)
+    return total_num_param
+
+
+def enable_vram_management(model: torch.nn.Module, module_map: dict, module_config: dict, max_num_param=None, overflow_module_config: dict = None, vram_limit=None):
+    enable_vram_management_recursively(model, module_map, module_config, max_num_param, overflow_module_config, total_num_param=0, vram_limit=vram_limit)
+    model.vram_management_enabled = True
+

examples/1.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:68ae88d1729f13d2eba6b9c3ae265af24ba563c725cd6ec7430fa9cf2a8f3584
+size 695111

examples/10.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:33faeefd4d21ca9ddc6386f0c7a83523632901d7f791fb5bd307b43739235c3d
+size 3886742

examples/178db6e89ab682bfc612a3290fec58dd.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:803bea53274f55f02463af8855585a0e4950ec1ae498ed1e6ef261d83d38b371
+size 1552729

examples/1b0daeb776471c7389b36cee53049417.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:c89828ebf754c762daa916ba764293193986fde8e2ff0e70be66926fbc9a8d07
+size 1447735

examples/2.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:6dd9a4527924fa2c26fd5bcc85237951180dcc136bac88d08bcd78635de58848
+size 883548

examples/3.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:da710ac25e12ef740d0bac04a452d1da6078a7acfe35bd456ab8a659a81401ff
+size 628311

examples/30.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:f0aee9bbb44a52ffd26424b4a0d804ee236e56edddadbcec5b7b6d79b8b2464e
+size 2677102

examples/31.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:9d0060e4564595ecc6397b7516ffc906beab7ff2a11971be2c9cd0e7807e6772
+size 569935

examples/32.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:209d7f97bd9e5c881789d35859c06840a7321db7a1f5cfe285dadf18d84e847c
+size 1593158

examples/33.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:54c8de91a05ceb1c01c3bcc267e527a55adffcf6dee1714629bb5b09850a38fd
+size 918682

examples/35.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:d9fba530d137d1ebeae598f502c4b59f7693d87738038d6f726512d00b54a6e4
+size 952985

examples/36.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:eff2bff79f77dadc26bb74623b50ef6f9e6b37b68247a0779f48b57822e5f74d
+size 1008476

examples/39.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:df7ff152c9835a20d95d3ccf8bd0386ed8fd25ed393b7eb2b0f669f7866124e3
+size 740198

examples/40.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:deb684b15a71152cc7377957946042e9e8fa8efce619f7801de8cb91eb3c1e82
+size 1021691

examples/5.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:631fc39bbde6f098c4f539f87c7c43de16a0d117ec51d16976d4b4e7e7279bc0
+size 7130292

examples/69230f105ad8740e08d743a8ee11c651.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:7d2920ef6ab0be19b5459a51aaacff4dbc773a18ecab150939866804d46b98ce
+size 416952

examples/7.mp4

@@ -0,0 +1,3 @@
+version https://git-lfs.github.com/spec/v1
+oid sha256:2cc7c666bcb736f6c8fca479eb49eb9257ad321db5f0737cea946f1dc831d888
+size 6361240