diff --git a/app.py b/app.py
new file mode 100644
index 0000000000000000000000000000000000000000..2c2eeea76528d765d73b1b34c503831ee7d7d555
--- /dev/null
+++ b/app.py
@@ -0,0 +1,387 @@
+import argparse
+import cv2
+import cv2
+import torch
+import gradio as gr
+from transformers import SamModel, SamProcessor
+
+import spaces
+import numpy as np
+from PIL import Image
+from tqdm import tqdm
+from torchvision.transforms import v2
+
+from rynnec import disable_torch_init, model_init, mm_infer, mm_infer_segmentation
+from rynnec.mm_utils import annToMask, load_video, load_images
+
+from PIL import Image
+from tqdm import tqdm
+import numpy as np
+import colorsys
+import argparse
+
+
+def get_hsv_palette(n_colors):
+    hues = np.linspace(0, 1, int(n_colors) + 1)[1:-1]
+    s = 0.8
+    v = 0.9
+    palette = [(0.0, 0.0, 0.0)] + [
+        colorsys.hsv_to_rgb(h_i, s, v) for h_i in hues
+    ]
+    return (255 * np.asarray(palette)).astype("uint8")
+
+
+def colorize_masks(images, index_masks, fac: float = 0.8, draw_contour=True, edge_thickness=20):
+    max_idx = max([m.max() for m in index_masks])
+    palette = get_hsv_palette(max_idx + 1)
+    color_masks = []
+    out_frames = []
+    for img, mask in tqdm(zip(images, index_masks), desc='Visualize masks ...'):
+        clr_mask = palette[mask.astype("int")]
+        blended_img = img
+
+        blended_img = compose_img_mask(blended_img, clr_mask, fac)
+
+        if draw_contour:
+            blended_img = draw_contours_on_image(blended_img, mask, clr_mask,
+                                                 brightness_factor=1.8,
+                                                 alpha=0.6,
+                                                 thickness=edge_thickness)
+        out_frames.append(blended_img)
+
+    return out_frames, color_masks
+
+
+def compose_img_mask(img, color_mask, fac: float = 0.5):
+    mask_region = (color_mask.sum(axis=-1) > 0)[..., None]
+    out_f = img.copy() / 255
+    out_f[mask_region[:, :, 0]] = fac * img[mask_region[:, :, 0]] / 255 + (1 - fac) * color_mask[mask_region[:, :, 0]] / 255
+    out_u = (255 * out_f).astype("uint8")
+    return out_u
+
+
+def draw_contours_on_image(img, index_mask, color_mask, brightness_factor=1.6, alpha=0.5, thickness=2, ignore_index=0):
+    img = img.astype("float32")
+    overlay = img.copy()
+
+    unique_indices = np.unique(index_mask)
+    if ignore_index is not None:
+        unique_indices = [idx for idx in unique_indices if idx != ignore_index]
+
+    for i in unique_indices:
+        bin_mask = (index_mask == i).astype("uint8") * 255
+        if bin_mask.sum() == 0:
+            continue
+
+        contours, _ = cv2.findContours(bin_mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+
+        color = color_mask[index_mask == i][0].astype("float32")
+        bright_color = np.clip(color * brightness_factor, 0, 255).tolist()
+
+        cv2.drawContours(overlay, contours, -1, bright_color, thickness)
+
+    blended = (1 - alpha) * img + alpha * overlay
+    return np.clip(blended, 0, 255).astype("uint8")
+
+
+def extract_first_frame_from_video(video):
+    cap = cv2.VideoCapture(video)
+    success, frame = cap.read()
+    cap.release()
+    if success:
+        return Image.fromarray(cv2.cvtColor(frame, cv2.COLOR_BGR2RGB))
+    return None
+
+
+def extract_points_from_mask(mask_pil):
+    mask = np.asarray(mask_pil)[..., 0]
+    coords = np.nonzero(mask)
+    coords = np.stack((coords[1], coords[0]), axis=1)
+
+    return coords
+
+def add_contour(img, mask, color=(1., 1., 1.)):
+    img = img.copy()
+
+    mask = mask.astype(np.uint8) * 255
+    contours, _ = cv2.findContours(mask, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
+    cv2.drawContours(img, contours, -1, color, thickness=8)
+
+    return img
+
+
+def load_first_frame(video_path):
+    cap = cv2.VideoCapture(video_path)
+    ret, frame = cap.read()
+    cap.release()
+    if not ret:
+        raise gr.Error("Could not read the video file.")
+    frame = cv2.cvtColor(frame, cv2.COLOR_BGR2RGB)
+    image = Image.fromarray(frame)  
+    return image
+
+
+def clear_masks():
+    return [], [], [], []
+
+def clear_all():
+    return [], [], [], [], None, "", ""
+
+
+@spaces.GPU(duration=120)
+def apply_sam(image, input_points):
+    inputs = sam_processor(image, input_points=input_points, return_tensors="pt").to(device)
+
+    with torch.no_grad():
+        outputs = sam_model(**inputs)
+
+    masks = sam_processor.image_processor.post_process_masks(outputs.pred_masks.cpu(), inputs["original_sizes"].cpu(), inputs["reshaped_input_sizes"].cpu())[0][0]
+    scores = outputs.iou_scores[0, 0]
+
+    mask_selection_index = scores.argmax()
+
+    mask_np = masks[mask_selection_index].numpy()
+
+    return mask_np
+
+
+@spaces.GPU(duration=120)
+def run(mode, images, timestamps, masks, mask_ids, instruction, mask_output_video):
+    if mode == "QA":
+        response = run_text_inference(images, timestamps, masks, mask_ids, instruction)
+    else:
+        response, mask_output_video = run_seg_inference(images, timestamps, instruction)
+    return response, mask_output_video
+
+
+def run_text_inference(images, timestamps, masks, mask_ids, instruction):
+    masks = torch.from_numpy(np.stack(masks, axis=0))
+
+    if "<video>" not in instruction:
+        instruction = "<video>\n" + instruction
+
+    if len(masks) >= 2:
+        obj_str = f"<video>\nThere are {len(masks)} objects in the video: " + ", ".join([f"<object{i}> [<REGION>]" for i in range(len(masks))])
+        instruction = instruction.replace("<video>\n", obj_str)
+    else:
+        instruction = instruction.replace("<object0>", '[<REGION>]')
+
+    output = mm_infer(
+        (images, timestamps),
+        processor,
+        instruction,
+        model=model,
+        tokenizer=processor.tokenizer,
+        do_sample=False,
+        modal='video',
+        masks=masks.cuda() if masks is not None else None,
+        mask_ids=mask_ids
+    )
+
+    return output
+
+
+def run_seg_inference(images, timestamps, instruction):
+    output, masks = mm_infer_segmentation(
+        (images, timestamps),
+        processor,
+        instruction,
+        model=model,
+        tokenizer=processor.tokenizer,
+        do_sample=False,
+        modal='video',
+    )
+
+    w, h = images[0].size
+    masks = v2.Resize([h, w])(masks).cpu().numpy()
+
+    mask_list_video = []
+
+    images = [np.array(image) for image in images]
+    masks = [mask[0] for mask in masks]
+    show_images, _ = colorize_masks(images, masks)
+    for i, image in enumerate(show_images):
+        if masks[i].sum() > 1000:
+            mask_list_video.append((Image.fromarray(image), f"Frame {i}"))        
+
+    return output, mask_list_video
+
+
+def generate_masks_video(image, mask_list_video, mask_raw_list_video, mask_ids, frame_idx):
+    image['image'] = image['background'].convert('RGB')
+    # del image['background'], image['composite']
+    assert len(image['layers']) == 1, f"Expected 1 layer, got {len(image['layers'])}"
+
+    mask = Image.fromarray((np.asarray(image['layers'][0])[..., 3] > 0).astype(np.uint8) * 255).convert('RGB')
+    points = extract_points_from_mask(mask)
+    np.random.seed(0)
+    if points.shape[0] == 0:
+        raise gr.Error("No points selected")
+
+    points_selected_indices = np.random.choice(points.shape[0], size=min(points.shape[0], 8), replace=False)
+    points = points[points_selected_indices]
+    coords = [points.tolist()]
+    mask_np = apply_sam(image['image'], coords)
+
+    mask_raw_list_video.append(mask_np)
+    mask_image = Image.fromarray((mask_np[:,:,np.newaxis] * np.array(image['image'])).astype(np.uint8))
+    
+    mask_list_video.append((mask_image, f"<object{len(mask_list_video)}>"))
+    # Return a list containing the mask image.
+    image['layers'] = []
+    image['composite'] = image['background']
+    mask_ids.append(frame_idx)
+    return mask_list_video, image, mask_list_video, mask_raw_list_video, mask_ids
+
+
+if __name__ == "__main__":
+    parser = argparse.ArgumentParser(description="VideoRefer gradio demo")
+    parser.add_argument("--model-path", type=str, default="Alibaba-DAMO-Academy/RynnEC-2B", help="Path to the model checkpoint")
+    parser.add_argument("--port", type=int, default=4001)
+
+    args_cli = parser.parse_args()
+
+    with gr.Blocks(theme=gr.themes.Soft(primary_hue="amber")) as demo:
+
+        mask_list = gr.State([])  
+        mask_raw_list = gr.State([])  
+        mask_list_video = gr.State([])  
+        mask_raw_list_video = gr.State([])  
+
+
+        HEADER = ("""
+            <div>
+                <h1>RynnEC Demo</h1>
+                <h5 style="margin: 0;">Feel free to click on anything that grabs your interest!</h5>
+                <h5 style="margin: 0;">If this demo please you, please give us a star ⭐ on Github or 💖 on this space.</h5>
+            </div>
+            </div>
+            <div style="display: flex; justify-content: left; margin-top: 10px;">
+            <a href="https://arxiv.org/pdf/2501.00599"><img src="https://img.shields.io/badge/Arxiv-2501.00599-ECA8A7" style="margin-right: 5px;"></a>
+            <a href="https://github.com/DAMO-NLP-SG/VideoRefer"><img src='https://img.shields.io/badge/Github-VideoRefer-F7C97E' style="margin-right: 5px;"></a>
+            <a href="https://github.com/DAMO-NLP-SG/VideoLLaMA3"><img src='https://img.shields.io/badge/Github-VideoLLaMA3-9DC3E6' style="margin-right: 5px;"></a>
+            </div>
+            """)
+
+
+        image_tips = """
+                ### 💡 Tips:
+
+                🧸 Upload an image, and you can use the drawing tool✍️ to highlight the areas you're interested in.
+            
+                🔖 For single-object caption mode, simply select the area and click the 'Generate Caption' button to receive a caption for the object.
+                
+                🔔 In QA mode, you can generate multiple masks by clicking the 'Generate Mask' button multiple times. Afterward, use the corresponding object id to ask questions.
+                
+                📌 Click the button 'Clear Masks' to clear the current generated masks.
+                
+                """
+        
+        video_tips = """
+                ### 💡 Tips:
+                🧸 Upload an video, and you can use the drawing tool✍️ to highlight the areas you're interested in the first frame.
+                
+                🔔 In QA mode, you can generate multiple masks by clicking the 'Generate Mask' button multiple times. Afterward, use the corresponding object id to ask questions.
+                
+                📌 Click the button 'Clear Masks' to clear the current generated masks.
+                
+                """
+
+
+        with gr.TabItem("Video"):
+            with gr.Row():
+                with gr.Column():
+                    video_input = gr.Video(label="Video", interactive=True)
+                    frame_idx = gr.Slider(minimum=0, maximum=0, value=0, step=1, label="Select Frame", interactive=False)
+                    selected_frame = gr.ImageEditor(
+                        label="Annotate Frame",
+                        type="pil", 
+                        sources=[], 
+                        interactive=True,
+                    )
+                    generate_mask_btn_video = gr.Button("1️⃣ Generate Mask", visible=True, variant="primary")
+                    gr.Examples([f"./demo/videos/{i+1}.mp4" for i in range(4)], inputs=video_input, label="Examples")
+
+                with gr.Column():
+                    mode_video = gr.Radio(label="Mode", choices=["QA", "Seg"], value="QA")
+                    mask_output_video = gr.Gallery(label="Referred Masks", object_fit='scale-down')
+
+                    query_video = gr.Textbox(label="Question", value="Please describe <object0>.", interactive=True, visible=True)
+                    response_video = gr.Textbox(label="Answer", interactive=False)
+
+                    submit_btn_video = gr.Button("Generate Caption", variant="primary", visible=False)
+                    submit_btn_video1 = gr.Button("2️⃣ Generate Answer", variant="primary", visible=True)
+                    description_video = gr.Textbox(label="Output", visible=False)
+                    
+                    clear_masks_btn_video = gr.Button("Clear Masks", variant="secondary")
+
+            gr.Markdown(video_tips)
+
+            frames = gr.State(value=[])
+            timestamps = gr.State(value=[])
+            mask_ids = gr.State(value=[])
+
+        def on_video_upload(video_path):
+            frames, timestamps = load_video(video_path, fps=1, max_frames=128)
+            frames = [Image.fromarray(x.transpose(1, 2, 0)) for x in frames]
+            return frames, timestamps, frames[0], gr.update(value=0, maximum=len(frames) - 1, interactive=True)
+
+        def on_frame_idx_change(frame_idx, frames):
+            return frames[frame_idx]
+
+        def to_seg_mode():
+            return (
+                *[gr.update(visible=False) for _ in range(4)],
+                []
+            )
+
+        def to_qa_mode():
+            return (
+                *[gr.update(visible=True) for _ in range(4)],
+                []
+            )
+
+        def on_mode_change(mode):
+            if mode == "QA":
+                return to_qa_mode()
+            return to_seg_mode()
+
+        mode_video.change(on_mode_change, inputs=[mode_video], outputs=[frame_idx, selected_frame, generate_mask_btn_video, response_video, mask_output_video])
+        video_input.change(on_video_upload, inputs=[video_input], outputs=[frames, timestamps, selected_frame, frame_idx])
+        frame_idx.change(on_frame_idx_change, inputs=[frame_idx, frames], outputs=[selected_frame])
+
+        generate_mask_btn_video.click(
+            fn=generate_masks_video,
+            inputs=[selected_frame, mask_list_video, mask_raw_list_video, mask_ids, frame_idx],
+            outputs=[mask_output_video, selected_frame, mask_list_video, mask_raw_list_video, mask_ids]
+        )
+
+        submit_btn_video1.click(
+            fn=run,
+            inputs=[mode_video, frames, timestamps, mask_raw_list_video, mask_ids, query_video, mask_output_video],
+            outputs=[response_video, mask_output_video],
+            api_name="describe_video"
+        )
+
+        video_input.clear(
+            fn=clear_all,
+            outputs=[mask_output_video, mask_list_video, mask_raw_list_video, mask_ids, selected_frame, query_video, response_video]
+        )
+
+        clear_masks_btn_video.click(
+            fn=clear_masks,
+            outputs=[mask_output_video, mask_list_video, mask_raw_list_video, mask_ids]
+        )
+
+    device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
+    sam_model = SamModel.from_pretrained("facebook/sam-vit-huge").to(device)
+    sam_processor = SamProcessor.from_pretrained("facebook/sam-vit-huge")
+    # sam_model = sam_processor = None
+    disable_torch_init()
+    model, processor = model_init(args_cli.model_path)
+    # model = processor = None
+
+    # demo.launch()
+    demo.launch(
+        share=False,
+    )
diff --git a/requirements.txt b/requirements.txt
new file mode 100644
index 0000000000000000000000000000000000000000..7c3d61ec155f9ed3c879c24515ba7b37d5a5668f
--- /dev/null
+++ b/requirements.txt
@@ -0,0 +1,44 @@
+--extra-index-url https://download.pytorch.org/whl/cu124
+# basic dependencies
+torch==2.4.0
+torchvision==0.19.0
+datasets==2.21.0
+transformers==4.46.3
+tokenizers==0.20.3
+deepspeed==0.15.4
+accelerate==1.0.1
+peft==0.4.0
+timm==1.0.3
+numpy==1.24.4
+# data processing
+decord==0.6.0
+imageio==2.34.0
+imageio-ffmpeg==0.4.9
+moviepy==1.0.3
+opencv-python==4.6.0.66
+pyarrow
+pysubs2
+ffmpeg-python
+# misc
+scikit-learn==1.2.2
+huggingface_hub==0.23.4
+sentencepiece==0.1.99
+shortuuid
+einops==0.6.1
+einops-exts==0.0.4
+bitsandbytes==0.43.3 # for cuda 124
+pydantic>=2.0
+markdown2[all]
+gradio==3.50.0
+gradio_client==0.6.1
+httpx==0.24.1
+requests
+openai
+uvicorn
+fastapi
+tensorboard
+wandb
+tabulate
+hydra-core
+pycocotools==2.0.10
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.3/flash_attn-2.7.3+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
diff --git a/rynnec/__init__.py b/rynnec/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..77c53343169dcbaa7dbff7326ad8710667aac5be
--- /dev/null
+++ b/rynnec/__init__.py
@@ -0,0 +1,269 @@
+import os
+import copy
+import math
+import warnings
+import shutil
+from functools import partial
+
+import torch
+import numpy as np
+from .model import load_pretrained_model
+from .mm_utils import load_images, process_images, load_video, process_video, tokenizer_multimodal_token, get_model_name_from_path, KeywordsStoppingCriteria, DirectResize, sam_preprocess_batch
+from .constants import NUM_FRAMES, DEFAULT_IMAGE_TOKEN, DEFAULT_VIDEO_TOKEN, MODAL_INDEX_MAP, STREAM_START_TOKEN, STREAM_END_TOKEN
+from .model.rynnec_qwen2 import Videollama3Qwen2Processor
+
+def disable_torch_init():
+    """
+    Disable the redundant torch default initialization to accelerate model creation.
+    """
+    import torch
+    setattr(torch.nn.Linear, "reset_parameters", lambda self: None)
+    setattr(torch.nn.LayerNorm, "reset_parameters", lambda self: None)
+
+
+def model_init(model_path=None, min_visual_tokens=None, max_visual_tokens=None, **kwargs):
+    model_path = "Alibaba-DAMO-Academy/RynnEC-2B" if model_path is None else model_path
+    model_name = get_model_name_from_path(model_path)
+    tokenizer, model, image_processor, context_len = load_pretrained_model(model_path, None, model_name, **kwargs)
+
+    if max_visual_tokens is not None:
+        image_processor.max_tokens = max_visual_tokens
+    if min_visual_tokens is not None:
+        image_processor.min_tokens = min_visual_tokens
+
+    if tokenizer.pad_token is None and tokenizer.unk_token is not None:
+        tokenizer.pad_token = tokenizer.unk_token
+
+    processor = Videollama3Qwen2Processor(image_processor, tokenizer)
+
+    return model, processor
+
+
+def mm_infer(images_or_videos, vlprocessor, instruct, model, tokenizer, modal='video', **kwargs):
+
+    mask_ids = kwargs.pop('mask_ids', None)
+    masks = kwargs.pop('masks', None)
+    if modal == 'image':
+        modal_token = DEFAULT_IMAGE_TOKEN
+        images = images_or_videos
+        timestamps = None
+    elif modal == 'video':
+        modal_token = DEFAULT_VIDEO_TOKEN
+        images, timestamps = images_or_videos
+    elif modal == 'text':
+        modal_token = ''
+    else:
+        raise ValueError(f"Unsupported modal: {modal}")
+
+
+    # 1. text preprocess (tag process & generate prompt).
+    if isinstance(instruct, str):
+        messages = [{'role': 'user', 'content': instruct}]
+    elif isinstance(instruct, list):
+        messages = copy.deepcopy(instruct)
+    else:
+        raise ValueError(f"Unsupported type of instruct: {type(instruct)}")
+
+    if all(not modal_token in message["content"] for message in messages):
+        warnings.warn(f"Image tag not found in the conversation, add it automatically at the beginning!")
+        messages[0]["content"] = modal_token + messages[0]["content"]
+
+    converted_messages = []
+    for message in messages:
+        chunks = message["content"].split(modal_token)
+        converted_messages.append({
+            "role": "user",
+            "content": []
+        })
+
+        for chunk_idx in range(1, 2 * len(chunks)):
+            if chunk_idx % 2 == 1:
+                chunk = chunks[chunk_idx // 2].strip()
+                converted_messages[-1]["content"].append({"type": "text",  "text": chunk}) if chunk else None
+            else:
+                if modal == 'image':
+                    converted_messages[-1]["content"].append({"type": "image"})
+                elif modal == 'video':
+                    converted_messages[-1]["content"].append({"type": "video", "num_frames": len(images), "time": timestamps})
+
+    messages = converted_messages
+
+    system_message = []
+
+    image_downsampling = kwargs.get('image_downsampling', model.config.spatial_merge_size)
+    # TODO: attention mask?
+    messages = system_message + messages
+    data_dict = vlprocessor(
+        images=images,
+        text=messages,
+        merge_size=image_downsampling,
+        return_labels=True,
+        return_tensors="pt",
+    )
+
+    torch_dtype = model.config.torch_dtype if hasattr(model.config, "torch_dtype") else torch.float16
+
+    # images = [x.to(torch_dtype).cuda(non_blocking=True) for x in data_dict["images"]]
+    # grid_thws = [x.cuda(non_blocking=True) for x in data_dict["grid_thws"]]
+
+    # 3. generate response according to visual signals and prompts. 
+    keywords = [tokenizer.eos_token]
+    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, data_dict["input_ids"].unsqueeze(0))
+
+    do_sample = kwargs.get('do_sample', False)
+    temperature = kwargs.get('temperature', 0.2 if do_sample else 1.0)
+    top_p = kwargs.get('top_p', 0.9 if do_sample else 1.0)
+    top_k = kwargs.get('top_k', 20 if do_sample else 50)
+    max_new_tokens = kwargs.get('max_new_tokens', 2048)
+
+    data_dict["modals"] = [modal]
+    data_dict = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data_dict.items()}
+    if "pixel_values" in data_dict:
+        data_dict["modals"] = data_dict["modals"] * len(data_dict["grid_sizes"])
+        data_dict["pixel_values"] = data_dict["pixel_values"].to(torch.bfloat16)
+
+    with torch.inference_mode():
+        output_ids = model.generate(
+            input_ids=data_dict["input_ids"].unsqueeze(0).cuda(),
+            pixel_values=data_dict["pixel_values"],
+            grid_sizes=data_dict["grid_sizes"],
+            merge_sizes=data_dict["merge_sizes"],
+            modals=data_dict["modals"],
+            do_sample=do_sample,
+            temperature=temperature,
+            max_new_tokens=max_new_tokens,
+            top_p=top_p,
+            top_k=top_k,
+            use_cache=True,
+            stopping_criteria=[stopping_criteria],
+            pad_token_id=tokenizer.eos_token_id,
+            masks=[masks],
+            mask_ids=mask_ids
+        )
+
+    outputs = tokenizer.decode(output_ids[0], skip_special_tokens=True)
+
+    return outputs
+
+def mm_infer_segmentation(images_or_videos, vlprocessor, instruct, model, tokenizer, modal='video', seg_start_idx=0, **kwargs):
+
+    image2maskids = kwargs.get('image2maskids', [])
+    img_size=1024
+    sam_transform = DirectResize(img_size)
+
+
+    if modal == 'image':
+        modal_token = DEFAULT_IMAGE_TOKEN
+        images = images_or_videos
+        timestamps = None
+    elif modal == 'video':
+        modal_token = DEFAULT_VIDEO_TOKEN
+        images, timestamps = images_or_videos
+    elif modal == 'text':
+        modal_token = ''
+    else:
+        raise ValueError(f"Unsupported modal: {modal}")
+
+
+    sam_images = []
+    sam_size = None
+    if len(images)>0:
+        for image in images:
+            sam_image = sam_transform.apply_image(np.array(image))
+            sam_images.append(sam_image)
+            if sam_size is None:
+                sam_size = sam_image.shape[:2] 
+        sam_images = np.array(sam_images)
+        sam_images = torch.from_numpy(sam_images).permute(0, 3, 1, 2).contiguous()
+        sam_images = sam_preprocess_batch(sam_images)
+
+
+    # 1. text preprocess (tag process & generate prompt).
+    if isinstance(instruct, str):
+        messages = [{'role': 'user', 'content': instruct}]
+    elif isinstance(instruct, list):
+        messages = copy.deepcopy(instruct)
+    else:
+        raise ValueError(f"Unsupported type of instruct: {type(instruct)}")
+
+    if all(not modal_token in message["content"] for message in messages):
+        warnings.warn(f"Image tag not found in the conversation, add it automatically at the beginning!")
+        messages[0]["content"] = modal_token + messages[0]["content"]
+
+    converted_messages = []
+    for message in messages:
+        chunks = message["content"].split(modal_token)
+        converted_messages.append({
+            "role": "user",
+            "content": []
+        })
+
+        for chunk_idx in range(1, 2 * len(chunks)):
+            if chunk_idx % 2 == 1:
+                chunk = chunks[chunk_idx // 2].strip()
+                converted_messages[-1]["content"].append({"type": "text",  "text": chunk}) if chunk else None
+            else:
+                if modal == 'image':
+                    converted_messages[-1]["content"].append({"type": "image"})
+                elif modal == 'video':
+                    converted_messages[-1]["content"].append({"type": "video", "num_frames": len(images), "time": timestamps})
+
+    messages = converted_messages
+
+    system_message = []
+
+    image_downsampling = kwargs.get('image_downsampling', model.config.spatial_merge_size)
+    # TODO: attention mask?
+    messages = system_message + messages
+    data_dict = vlprocessor(
+        images=images,
+        text=messages,
+        merge_size=image_downsampling,
+        return_labels=True,
+        return_tensors="pt",
+    )
+
+    torch_dtype = model.config.torch_dtype if hasattr(model.config, "torch_dtype") else torch.float16
+
+    keywords = [tokenizer.eos_token]
+    stopping_criteria = KeywordsStoppingCriteria(keywords, tokenizer, data_dict["input_ids"].unsqueeze(0))
+
+    do_sample = kwargs.get('do_sample', False)
+    temperature = kwargs.get('temperature', 0.2 if do_sample else 1.0)
+    top_p = kwargs.get('top_p', 0.9 if do_sample else 1.0)
+    top_k = kwargs.get('top_k', 20 if do_sample else 50)
+    max_new_tokens = kwargs.get('max_new_tokens', 2048)
+
+    torch_dtype = model.config.torch_dtype if hasattr(model.config, "torch_dtype") else torch.float16
+
+    data_dict["modals"] = [modal]
+    data_dict = {k: v.cuda() if isinstance(v, torch.Tensor) else v for k, v in data_dict.items()}
+    if "pixel_values" in data_dict:
+        data_dict["modals"] = data_dict["modals"] * len(data_dict["grid_sizes"])
+        data_dict["pixel_values"] = data_dict["pixel_values"].to(torch.bfloat16)
+
+    with torch.inference_mode():
+        output_ids, pred_masks = model.inference(
+            input_ids=data_dict["input_ids"].unsqueeze(0).cuda(),
+            pixel_values=data_dict["pixel_values"],
+            grid_sizes=data_dict["grid_sizes"],
+            merge_sizes=data_dict["merge_sizes"],
+            modals=data_dict["modals"],
+            sam_images=[sam_images],
+            sam_size=[sam_size],
+            image2maskids=[image2maskids],
+            do_sample=do_sample,
+            temperature=temperature,
+            max_new_tokens=max_new_tokens,
+            top_p=top_p,
+            top_k=top_k,
+            use_cache=True,
+            stopping_criteria=[stopping_criteria],
+            pad_token_id=tokenizer.eos_token_id,
+            seg_start_idx=seg_start_idx
+        )
+
+    outputs = tokenizer.batch_decode(output_ids, skip_special_tokens=True)[0].strip()
+    pred_masks_sigmoid = pred_masks.sigmoid()>0.5
+
+    return outputs, pred_masks_sigmoid
diff --git a/rynnec/constants.py b/rynnec/constants.py
new file mode 100644
index 0000000000000000000000000000000000000000..332d51dea3eb4f95e2197240b057503b1fa8b004
--- /dev/null
+++ b/rynnec/constants.py
@@ -0,0 +1,47 @@
+CONTROLLER_HEART_BEAT_EXPIRATION = 30
+WORKER_HEART_BEAT_INTERVAL = 15
+
+LOGDIR = "."
+
+# Model Constants
+IGNORE_INDEX = -100
+
+# Image arguments
+IMAGE_TOKEN_INDEX = -200
+DEFAULT_IMAGE_TOKEN = "<image>"
+DEFAULT_IMAGE_PATCH_TOKEN = "<im_patch>"
+DEFAULT_IM_START_TOKEN = "<im_start>"
+DEFAULT_IM_END_TOKEN = "<im_end>"
+IMAGE_PLACEHOLDER = "<image-placeholder>"
+
+# Video arguments
+VIDEO_TOKEN_INDEX = -201
+DEFAULT_VIDEO_TOKEN = "<video>"
+NUM_FRAMES = 128
+MAX_FRAMES = 768
+NUM_FRAMES_PER_SECOND = 1
+
+# Region arguments
+REGION_TOKEN = "<REGION>"
+REGION_TOKEN_REPLACE = "<region>"
+SEG_TOKEN = "[SEG]"
+
+# Audio arguments
+AUDIO_TOKEN_INDEX = -202
+DEFAULT_AUDIO_TOKEN = "<audio>"
+
+# Stream arguments
+STREAM_START_TOKEN = "<|stream_start|>"
+STREAM_END_TOKEN = "<|stream_end|>"
+STREAM_MAX_FRAMES = 400
+STREAM_FPS = 2
+STREAM_IMAGE_SIZE = 224
+STREAM_DOWNSAMPLING = 4
+
+MODAL_INDEX_MAP = {
+    "<image>": -200,
+    "<video>": -201,
+    "<audio>": -202,
+}
+
+subimage_token_num=196
\ No newline at end of file
diff --git a/rynnec/mm_utils.py b/rynnec/mm_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..3fc2f61a361bb2d07cde25795a0620c1cbafc269
--- /dev/null
+++ b/rynnec/mm_utils.py
@@ -0,0 +1,733 @@
+# Adopted from: https://github.com/DAMO-NLP-SG/VideoLLaMA3. 
+# Below is the original copyright:
+#    Copyright 2025 The VideoLLaMA3 team, Alibaba Group
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+import ast
+import os
+import re
+import math
+import base64
+import traceback
+from io import BytesIO
+from typing import Optional
+
+import torch
+import torchvision.transforms.functional as VF
+import torch.nn.functional as F
+import numpy as np
+from transformers import StoppingCriteria
+
+import cv2
+import imageio
+import ffmpeg
+from PIL import Image
+from decord import VideoReader, cpu
+
+from .constants import NUM_FRAMES, MAX_FRAMES, NUM_FRAMES_PER_SECOND, MODAL_INDEX_MAP, DEFAULT_IMAGE_TOKEN
+from pycocotools import mask as maskUtils
+
+from torchvision.transforms.functional import resize, to_pil_image  # type: ignore
+
+
+class DirectResize:
+    def __init__(self, target_length: int) -> None:
+        self.target_length = target_length
+
+    def apply_image(self, image: np.ndarray) -> np.ndarray:
+        """
+        Expects a numpy array with shape HxWxC in uint8 format.
+        """
+        img = to_pil_image(image, mode='RGB')
+        return np.array(img.resize((self.target_length, self.target_length)))
+
+def sam_preprocess_batch(x: torch.Tensor) -> torch.Tensor:
+    """
+    Normalize pixel values and pad to square input for a batch of images.
+    
+    Args:
+        images (torch.Tensor): A batch tensor of shape [N, C, H, W].
+    
+    Returns:
+        torch.Tensor: A batch tensor with normalized and padded images 
+                      (shape: [N, C, 1024, 1024]).
+    """
+    pixel_mean = torch.Tensor([123.675, 116.28, 103.53]).view(1, -1, 1, 1)
+    pixel_std = torch.Tensor([58.395, 57.12, 57.375]).view(1, -1, 1, 1)
+    img_size = 1024
+        
+    # Normalize colors
+    x = (x - pixel_mean) / pixel_std
+
+    # Pad
+    h, w = x.shape[-2:]
+    padh = img_size - h
+    padw = img_size - w
+    x = F.pad(x, (0, padw, 0, padh)) 
+    return x
+ 
+
+def sam_preprocess(x: torch.Tensor) -> torch.Tensor:
+    """Normalize pixel values and pad to a square input."""
+    pixel_mean = torch.Tensor([123.675, 116.28, 103.53]).view(-1, 1, 1)
+    pixel_std = torch.Tensor([58.395, 57.12, 57.375]).view(-1, 1, 1)
+    img_size = 1024
+        
+    # Normalize colors
+    x = (x - pixel_mean) / pixel_std
+
+    # Pad
+    h, w = x.shape[-2:]
+    padh = img_size - h
+    padw = img_size - w
+    x = F.pad(x, (0, padw, 0, padh))
+    return x
+ 
+
+def reshape_images_to_raw_grid(mm_features_raw, grid_thws):
+    start_idx=0
+    reshaped_features = []
+    # for thw_group in grid_thws:
+    for tensor_thw in grid_thws:
+        # for tensor_thw in thw_group:
+            t, H, W = tensor_thw.squeeze().tolist()
+            num_elements = H * W
+            for i in range(t):
+                split_tensor = mm_features_raw[start_idx:start_idx + num_elements].view(H, W, -1)
+                reshaped_features.append(split_tensor)
+
+                start_idx += num_elements
+
+    assert len(mm_features_raw)==start_idx
+    return reshaped_features
+
+def annToMask(mask_ann, h=None, w=None):
+    if isinstance(mask_ann, list):
+        rles = maskUtils.frPyObjects(mask_ann, h, w)
+        rle = maskUtils.merge(rles)
+    elif isinstance(mask_ann['counts'], list):
+        # uncompressed RLE
+        rle = maskUtils.frPyObjects(mask_ann, h, w)
+    else:
+        # rle
+        rle = mask_ann
+    mask = maskUtils.decode(rle)
+    return mask
+
+def chunk_list(input_list, chunk_size):
+    return [input_list[i:i + chunk_size] for i in range(0, len(input_list), chunk_size)]
+
+
+def load_image_from_base64(image):
+    return Image.open(BytesIO(base64.b64decode(image)))
+
+
+def expand2square(pil_img, background_color):
+    width, height = pil_img.size
+    if width == height:
+        return pil_img
+    elif width > height:
+        result = Image.new(pil_img.mode, (width, width), background_color)
+        result.paste(pil_img, (0, (width - height) // 2))
+        return result
+    else:
+        result = Image.new(pil_img.mode, (height, height), background_color)
+        result.paste(pil_img, ((height - width) // 2, 0))
+        return result
+
+
+def grid_divide(image, cell_size):
+    """
+    Divides an image into grid of a specified size.
+
+    Args:
+        image (PIL.Image.Image): The input image.
+        cell_size (int): The size of each cell.
+
+    Returns:
+        list: A list of PIL.Image.Image objects representing the patches.
+    """
+    grid = []
+    width, height = image.size
+    for i in range(0, height, cell_size):
+        row = []
+        for j in range(0, width, cell_size):
+            box = (j, i, j + cell_size, i + cell_size)
+            row.append(image.crop(box))
+        grid.append(row)
+
+    return grid
+
+
+def load_images(image_path):
+    if isinstance(image_path, str) and os.path.isfile(image_path):
+        # images = [cv2.cvtColor(cv2.imread(image_path), cv2.COLOR_BGR2RGB)]
+        images = [Image.open(image_path).convert('RGB')]
+    elif isinstance(image_path, str) and os.path.isdir(image_path):
+        # images = [cv2.cvtColor(cv2.imread(os.path.join(image_path, f)), cv2.COLOR_BGR2RGB) for f in sorted(os.listdir(image_path))]
+        images = [Image.open(os.path.join(image_path, f)).convert('RGB') for f in sorted(os.listdir(image_path))]
+    elif isinstance(image_path, list) and isinstance(image_path[0], str):
+        # images = [cv2.cvtColor(cv2.imread(f), cv2.COLOR_BGR2RGB) for f in image_path]
+        images = [Image.open(f).convert('RGB') for f in image_path]
+    elif isinstance(image_path, list) and isinstance(image_path[0], Image.Image):
+        images = image_path
+    elif isinstance(image_path, Image.Image):
+        images = [image_path]
+    else:
+        raise ValueError(f"Unsupported image path type: {image_path}")
+
+    return images
+
+
+def process_pad_image(image, padding_value=(0, 0, 0)):
+    image = expand2square(image, padding_value)
+
+    return [image]
+
+
+def find_closest_aspect_ratio(src_ratio, tgt_ratios, ori_size, tgt_size):
+    best_ratio_diff = float('inf')
+    best_ratio = (1, 1)
+    area = ori_size[0] * ori_size[1]
+    for ratio in tgt_ratios:
+        tgt_ratio = ratio[0] / ratio[1]
+        ratio_diff = abs(src_ratio - tgt_ratio)
+        if ratio_diff < best_ratio_diff:
+            best_ratio_diff = ratio_diff
+            best_ratio = ratio
+        elif ratio_diff == best_ratio_diff:
+            if area > 0.5 * tgt_size[0] * tgt_size[1] * ratio[0] * ratio[1]:
+                best_ratio = ratio
+
+    return best_ratio
+
+
+def process_dynamic_image(image, image_size=384, use_thumbnail=True):
+    # Grid Params:
+    min_num = 1
+    max_num = 12
+
+    if isinstance(image_size, int):
+        image_size = (image_size, image_size)
+
+    ori_size = image.size
+    aspect_ratio = ori_size[0] / ori_size[1]
+
+    # calculate the existing image aspect ratio
+    tgt_ratios = []
+    for n in range(min_num, max_num + 1):
+        tgt_ratios.extend([(i, j) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num])
+    tgt_ratios = set(tgt_ratios)
+    tgt_ratios = sorted(tgt_ratios, key=lambda x: x[0] * x[1])
+
+    # find the closest aspect ratio to the target
+    tgt_ratio = find_closest_aspect_ratio(aspect_ratio, tgt_ratios, ori_size, image_size)
+
+    # resize the image to the target size
+    tgt_width = image_size[0] * tgt_ratio[0]
+    tgt_height = image_size[1] * tgt_ratio[1]
+    resized_img = image.resize((tgt_width, tgt_height))
+
+    # NOTE: internvl2 style split the image into one column grids
+    # num_grids = tgt_ratio[0] * tgt_ratio[1]
+    # grid_images = []
+    # for i in range(num_grids):
+    #     box = (
+    #         (i %  tgt_ratio[0]) * image_size[0],
+    #         (i // tgt_ratio[0]) * image_size[1],
+    #         (i %  tgt_ratio[0] + 1) * image_size[0],
+    #         (i // tgt_ratio[0] + 1) * image_size[1],
+    #     )
+    #     # crop out the grid image
+    #     grid_images.append(resized_img.crop(box))
+    # assert len(grid_images) == num_grids
+    # grid_images = [grid_images]
+
+    # NOTE: eager implementation
+    # num_grids = tgt_ratio[0] * tgt_ratio[1]
+    # sub_grid_images = []
+    # tmp_grid_images = []
+    # for i in range(num_grids):
+    #     box = (
+    #         (i %  tgt_ratio[0]) * image_size[0],
+    #         (i // tgt_ratio[0]) * image_size[1],
+    #         (i %  tgt_ratio[0] + 1) * image_size[0],
+    #         (i // tgt_ratio[0] + 1) * image_size[1],
+    #     )
+    #     tmp_grid_images.append(resized_img.crop(box))
+
+    #     if (i + 1) % tgt_ratio[0] == 0:
+    #         sub_grid_images.append(tmp_grid_images)
+    #         tmp_grid_images = []
+
+    image_grid = grid_divide(resized_img, image_size[0])
+
+    if use_thumbnail:
+        thumbnail_img = image.resize((image_size[0], image_size[1]))
+        image_grid = [[thumbnail_img]] + image_grid
+
+    return image_grid
+
+
+def process_highres_image(image_path, image_size=384, use_thumbnail=True, padding_value=(0, 0, 0)):
+    # Grid Params:
+    grid_width = [1, 2, 3]
+    grid_width_real = [x * image_size for x in grid_width]
+
+    longest_side = max(image.size)
+    fit_grid_width_real = [x for x in grid_width_real if x >= longest_side]
+    if len(fit_grid_width_real) == 0:
+        select_size = max(grid_width_real)
+    else:
+        select_size = min(fit_grid_width_real)
+
+    image_padded = expand2square(image, padding_value)
+    image_padded = image_padded.resize((select_size, select_size))
+    image_grid = grid_divide(image_padded, image_size)
+
+    if use_thumbnail:
+        thumbnail_img = image.resize((image_size, image_size))
+        image_grid = [[thumbnail_img]] + image_grid
+
+    return image_grid
+
+
+def select_best_resolution(original_size, possible_resolutions):
+    """
+    Selects the best resolution from a list of possible resolutions based on the original size.
+
+    Args:
+        original_size (tuple): The original size of the image in the format (width, height).
+        possible_resolutions (list): A list of possible resolutions in the format [(width1, height1), (width2, height2), ...].
+
+    Returns:
+        tuple: The best fit resolution in the format (width, height).
+    """
+    original_width, original_height = original_size
+    best_fit = None
+    max_effective_resolution = 0
+    min_wasted_resolution = float('inf')
+
+    for width, height in possible_resolutions:
+        scale = min(width / original_width, height / original_height)
+        downscaled_width, downscaled_height = int(original_width * scale), int(original_height * scale)
+        effective_resolution = min(downscaled_width * downscaled_height, original_width * original_height)
+        wasted_resolution = (width * height) - effective_resolution
+
+        if effective_resolution > max_effective_resolution or (effective_resolution == max_effective_resolution and wasted_resolution < min_wasted_resolution):
+            max_effective_resolution = effective_resolution
+            min_wasted_resolution = wasted_resolution
+            best_fit = (width, height)
+
+    return best_fit
+
+
+def process_anyres_image(image, image_size=384, use_thumbnail=True, padding_value=(0, 0, 0)):
+    """
+    Process an image with variable resolutions.
+
+    Args:
+        image (PIL.Image.Image): The input image to be processed.
+        processor: The image processor object.
+
+    Returns:
+        torch.Tensor: A tensor containing the processed image patches.
+    """
+    # Grid Params:
+    possible_grids = [(1, 1), (1, 2), (1, 3), (2, 1), (2, 2), (2, 3)]
+    possible_resolutions = [(x * image_size, y * image_size) for x, y in possible_grids]
+
+    best_resolution = select_best_resolution(image.size, possible_resolutions)
+
+    # resize and padding image
+    nw, nh = best_resolution
+    ow, oh = image.size
+
+    scale_factor = min(nw / ow, nh / oh)
+    new_size = (int(ow * scale_factor), int(oh * scale_factor))
+
+    image_padded = Image.new("RGB", (nw, nh), padding_value)
+    image_padded.paste(image.resize(new_size), ((nw - new_size[0]) // 2, (nh - new_size[1]) // 2))
+
+    image_grid = grid_divide(image_padded, image_size)
+
+    if use_thumbnail:
+        thumbnail_img = image.resize((image_size, image_size))
+        image_grid = [[thumbnail_img]] + image_grid
+
+    return image_grid
+
+
+def process_adares_image(image_path, image_size=384, use_thumbnail=True):
+    # Grid Params:
+    min_num = 1
+    max_num = 12
+
+    if isinstance(image_size, int):
+        image_size = (image_size, image_size)
+
+    ori_size = image.size
+    aspect_ratio = ori_size[0] / ori_size[1]
+
+    # calculate the existing image aspect ratio
+    tgt_ratios = []
+    for n in range(min_num, max_num + 1):
+        tgt_ratios.extend([(i, j) for i in range(1, n + 1) for j in range(1, n + 1) if i * j <= max_num and i * j >= min_num])
+    tgt_ratios = set(tgt_ratios)
+    possible_resolutions = [(x * image_size[0], y * image_size[1]) for x, y in tgt_ratios]
+
+    # find the most possible resolution
+    best_resolution = select_best_resolution(ori_size, possible_resolutions)
+
+    # resize the image to the target size
+    resized_img = image.resize((best_resolution[0], best_resolution[1]))
+
+    image_grid = grid_divide(resized_img, image_size[0])
+
+    if use_thumbnail:
+        thumbnail_img = image.resize((image_size[0], image_size[1]))
+        image_grid = [[thumbnail_img]] + image_grid
+
+    return image_grid
+
+
+def process_images(image_path, processor, aspect_ratio='pad', image_size=384, use_thumbnail=True):
+    images = load_images(image_path)
+
+    padding_value = tuple(int(x*255) for x in processor.image_mean)
+
+    image_grids = []
+    for image in images:
+        if aspect_ratio == 'pad':
+            image_grid = process_pad_image(image, padding_value=padding_value)
+        elif aspect_ratio == 'dynamic':
+            image_grid = process_dynamic_image(image, image_size=image_size, use_thumbnail=use_thumbnail)
+        elif aspect_ratio == 'highres':
+            image_grid = process_highres_image(image, image_size=image_size, use_thumbnail=use_thumbnail, padding_value=padding_value)
+        elif aspect_ratio == 'anyres':
+            image_grid = process_anyres_image(image, image_size=image_size, use_thumbnail=use_thumbnail, padding_value=padding_value)
+        elif aspect_ratio == 'adares':
+            image_grid = process_adares_image(image, image_size=image_size, use_thumbnail=use_thumbnail)
+        else:
+            image_grid = [image]
+
+        image_grid = [processor.preprocess(image_row, return_tensors='pt', num_images=len(images)) for image_row in image_grid]
+        image_grids.append(image_grid)
+
+    return image_grids
+
+
+def frame_sample(duration, mode='uniform', num_frames=None, vid_fps=None, fps=None, must_sample_frames=None):
+    mask_ids = []
+    if mode == 'uniform':
+        assert num_frames is not None, "Number of frames must be provided for uniform sampling."
+        if duration <= num_frames:
+            video_ids = np.arange(duration).astype(int)
+            video_ids_list = video_ids.tolist()
+            for msf in must_sample_frames:
+                if msf not in video_ids_list:
+                    video_ids_list.append(msf)
+            video_ids_list.sort()
+            for msf in must_sample_frames:
+                mask_ids.append(video_ids_list.index(msf))
+            return np.array(video_ids_list), mask_ids
+        video_ids = np.linspace(0, duration-1, num_frames, dtype=int)
+        video_ids_list = video_ids.tolist()
+        if must_sample_frames is not None:
+            for msf in must_sample_frames:
+                if msf not in video_ids_list:
+                    video_ids_list.append(msf)
+            video_ids_list.sort()
+            for msf in must_sample_frames:
+                mask_ids.append(video_ids_list.index(msf))
+        return np.array(video_ids_list), mask_ids
+    elif mode == 'fps':
+        assert vid_fps is not None, "FPS must be provided for FPS sampling."
+        fps = fps if fps is not None else NUM_FRAMES_PER_SECOND
+        segment_len = min(vid_fps // fps, duration)
+        video_ids = np.arange(segment_len // 2, duration, segment_len, dtype=int)
+        video_ids_list = video_ids.tolist()
+        if must_sample_frames is not None:
+            for msf in must_sample_frames:
+                if msf not in video_ids_list:
+                    video_ids_list.append(msf)
+            video_ids_list.sort()
+            for msf in must_sample_frames:
+                mask_ids.append(video_ids_list.index(msf))
+        return np.array(video_ids_list), mask_ids
+
+    else:
+        raise ImportError(f'Unsupported frame sampling mode: {mode}')
+
+
+def load_video_from_ids(video_path, s=None, e=None, fps=None, max_frames=None, temporal_factor=1, must_sample_frames=None):
+    if s is not None and e is not None:
+        s = s if s >= 0. else 0.
+        e = e if e >= 0. else 0.
+        if s > e:
+            s, e = e, s
+        elif s == e:
+            e = s + 1
+
+    # 1. Loading Video
+    if os.path.isdir(video_path):
+        frame_files = sorted(os.listdir(video_path))
+
+        vid_fps = 3
+        num_frames_of_video = len(frame_files)
+    elif video_path.endswith('.gif'):
+        gif_reader = imageio.get_reader(video_path)
+
+        vid_fps = 25
+        num_frames_of_video = len(gif_reader)
+    else:
+        vreader = VideoReader(video_path, ctx=cpu(0), num_threads=2)
+        # vreader = VideoReader(video_path, ctx=cpu(0), num_threads=1)
+
+        vid_fps = vreader.get_avg_fps()
+        num_frames_of_video = len(vreader)
+
+    # 2. Determine frame range & Calculate frame indices
+    f_start = 0                       if s is None else max(int(s * vid_fps) - 1, 0)
+    f_end   = num_frames_of_video - 1 if e is None else min(int(e * vid_fps) - 1, num_frames_of_video - 1)
+    frame_indices = list(range(f_start, f_end + 1))
+
+    duration = len(frame_indices)
+    # 3. Sampling frame indices
+    max_frames = max_frames if max_frames is not None else MAX_FRAMES
+    if fps is not None and duration / vid_fps < max_frames:
+        sampled_ids, mask_ids = frame_sample(duration, mode='fps', vid_fps=vid_fps, fps=fps, must_sample_frames=must_sample_frames)
+        sampled_frame_indices = [frame_indices[i] for i in sampled_ids]
+    else:
+        sampled_ids, mask_ids = frame_sample(duration, mode='uniform', num_frames=max_frames, must_sample_frames=must_sample_frames)
+        sampled_frame_indices = [frame_indices[i] for i in sampled_ids]
+
+    # 4. Acquire frame data
+    if os.path.isdir(video_path):
+        frames = [cv2.cvtColor(cv2.imread(os.path.join(video_path, frame_files[frame_idx])), cv2.COLOR_BGR2RGB) for frame_idx in sampled_frame_indices]
+    elif video_path.endswith('.gif'):
+        frames = [cv2.cvtColor(frame, cv2.COLOR_RGBA2RGB) for idx, frame in enumerate(gif_reader) if idx in sampled_frame_indices]
+    else:
+        frames = vreader.get_batch(sampled_frame_indices).asnumpy()
+
+    # frames = frames.transpose(0, 3, 1, 2)
+    timestamps = [x / vid_fps for x in sampled_frame_indices]
+
+    if temporal_factor > 1:
+        pad_length = temporal_factor - len(frames) % temporal_factor
+        frames = np.concatenate([frames, frames[-1:].repeat(pad_length, axis=0)])
+        [timestamps.append(timestamps[-1] + 1 / fps) for _ in range(pad_length)]
+
+    # NOTE: pad the video with black frames
+    # while num_frames is not None and len(video_data) < num_frames:
+    #     video_data.append(Image.fromarray(np.zeros((*video_data[-1].size, 3), dtype=np.uint8)))
+
+    return frames, timestamps, mask_ids
+
+
+def load_video(
+    video_path: str,
+    start_time: Optional[float] = None,
+    end_time: Optional[float] = None,
+    fps: Optional[float] = None,
+    max_frames: Optional[float] = None,
+    size: Optional[int] = None,
+    size_divisible: int = 1,
+    precise_time: bool = False,
+    verbose: bool = False,
+    temporal_factor: int = 1
+):
+    """
+    Load and process a video file and return the frames and the timestamps of each frame.
+
+    Args:
+        video_path (str): Path to the video file.
+        start_time (float, optional): Start time in seconds. Defaults to None.
+        end_time (float, optional): End time in seconds. Defaults to None.
+        fps (float, optional): Frames per second. Defaults to None.
+        num_frames (float, optional): Number of frames to sample. Defaults to None.
+        size (int, optional): Size of the shortest side. Defaults to None.
+        size_divisible (int, optional): Size divisible by this number. Defaults to 1.
+        precise_time (bool, optional): Whether to use precise time. Defaults to False.
+        verbose (bool, optional): Print ffmpeg output. Defaults to False.
+
+    Returns:
+        frames (List[PIL.Image]): List of frames.
+        timestamps (List[float]): List of timestamps.
+    """
+    if start_time is not None and end_time is not None and end_time - start_time < 1:
+        return load_video_from_ids(video_path, start_time, end_time, fps=fps, max_frames=max_frames)
+    if os.path.isdir(video_path):
+        return load_video_from_ids(video_path, start_time, end_time, fps=fps, max_frames=max_frames)
+    if video_path.endswith('.gif'):
+        return load_video_from_ids(video_path, start_time, end_time, fps=fps, max_frames=max_frames)
+    probe = ffmpeg.probe(video_path)
+    duration = float(probe['format']['duration'])
+    video_stream = next((stream for stream in probe['streams'] if stream['codec_type'] == 'video'), None)
+    w, h = int(video_stream['width']), int(video_stream['height'])
+
+    kwargs, input_kwargs, output_kwargs = {}, {}, {}
+    do_trim = start_time is not None or end_time is not None
+    if start_time is not None:
+        new_start_time = max(float(video_stream['start_time']), start_time)
+        duration -= new_start_time - start_time
+        start_time = new_start_time
+    else:
+        start_time = float(video_stream['start_time'])
+    if end_time is not None:
+        duration = min(duration, end_time - start_time)
+    else:
+        duration = duration
+    if do_trim:
+        kwargs = {'ss': start_time, 't': duration}
+    if precise_time:
+        output_kwargs.update(kwargs)
+    else:
+        input_kwargs.update(kwargs)
+
+    if size is not None:
+        scale_factor = size / min(w, h)
+        new_w, new_h = round(w * scale_factor), round(h * scale_factor)
+    else:
+        new_w, new_h = w, h
+    new_w = new_w // size_divisible * size_divisible
+    new_h = new_h // size_divisible * size_divisible
+
+    # NOTE: It may result in unexpected number of frames in ffmpeg
+    # if calculate the fps directly according to max_frames
+    # NOTE: the below lines may hurt the performance
+    # if max_frames is not None and (fps is None or duration * fps > 2 * max_frames):
+    #     fps = max_frames / duration * 2
+
+    stream = ffmpeg.input(video_path, **input_kwargs)
+    if fps is not None:
+        stream = ffmpeg.filter(stream, "fps", fps=fps, round="down")
+    if new_w != w or new_h != h:
+        stream = ffmpeg.filter(stream, 'scale', new_w, new_h)
+    stream = ffmpeg.output(stream, "pipe:", format="rawvideo", pix_fmt="rgb24", **output_kwargs)
+    out, _ = ffmpeg.run(stream, capture_stdout=True, quiet=not verbose)
+
+    frames = np.frombuffer(out, np.uint8).reshape([-1, new_h, new_w, 3]).transpose([0, 3, 1, 2])
+
+    if fps is not None:
+        timestamps = np.arange(start_time, start_time + duration + 1 / fps, 1 / fps)[:len(frames)]
+    else:
+        timestamps = np.linspace(start_time, start_time + duration, len(frames))
+
+    max_frames = max_frames if max_frames is not None else MAX_FRAMES
+    if max_frames is not None and len(frames) > max_frames:
+        indices = np.linspace(0, len(frames) - 1, max_frames, dtype=int)
+        frames = frames[indices]
+        timestamps = [timestamps[i] for i in indices]
+
+    if temporal_factor > 1:
+        pad_length = temporal_factor - len(frames) % temporal_factor
+        frames = np.concatenate([frames, frames[-1:].repeat(pad_length, axis=0)])
+        [timestamps.append(timestamps[-1] + 1 / fps) for _ in range(pad_length)]
+
+    frames = [frame for frame in frames]
+
+    return frames, timestamps
+
+
+def process_video(video_path, processor, s=None, e=None, aspect_ratio='pad', num_frames=None):
+    fps = 1 if num_frames is None else None
+    # FFmpeg
+    frames, timestamps = load_video(video_path, s, e, fps=fps, max_frames=num_frames)
+    # Decord
+    # frames, timestamps = load_video_from_ids(video_path, s, e, fps=fps, max_frames=num_frames)
+
+    assert len(frames) == len(timestamps), "Number of frames and timestamps must match."
+
+    if aspect_ratio == 'pad':
+        frames = [expand2square(f, tuple(int(x*255) for x in processor.image_mean)) for f in frames]
+
+    if aspect_ratio == 'avt':
+        frames = [processor.preprocess(frame, return_tensors='pt', image_num=len(frames)) for frame in frames]
+        grid_frames = [frames]
+    else:
+        frames = processor.preprocess(frames, return_tensors='pt', image_num=len(frames))
+        grid_frames = [[frames]]
+
+    return grid_frames, timestamps
+
+
+def tokenizer_multimodal_token(prompt, tokenizer, multimodal_token=DEFAULT_IMAGE_TOKEN, return_tensors=None):
+    """Tokenize text and multimodal tag to input_ids.
+
+    Args:
+        prompt (str): Text prompt (w/ multimodal tag), e.g., '<video>\nDescribe the video.'
+        tokenizer (transformers.PreTrainedTokenizer): Tokenizer object.
+        multimodal_token (int): Token index corresponding to the multimodal tag.
+    """
+    multimodal_token_index = MODAL_INDEX_MAP.get(multimodal_token, None)
+    if multimodal_token_index is None:
+        input_ids = tokenizer(prompt, add_special_tokens=False).input_ids
+    else:
+        prompt_chunks = [tokenizer(chunk, add_special_tokens=False).input_ids for idx, chunk in enumerate(prompt.split(multimodal_token))]
+
+        input_ids = []
+        for i in range(1, 2 * len(prompt_chunks)):
+            if i % 2 == 1:
+                input_ids.extend(prompt_chunks[i // 2])
+            else:
+                input_ids.append(multimodal_token_index)
+
+    if return_tensors is not None:
+        if return_tensors == 'pt':
+            return torch.tensor(input_ids, dtype=torch.long)
+        raise ValueError(f'Unsupported tensor type: {return_tensors}')
+    return input_ids
+
+
+def get_model_name_from_path(model_path):
+    model_path = model_path.strip("/")
+    model_paths = model_path.split("/")
+    if model_paths[-1].startswith('checkpoint-'):
+        return model_paths[-2] + "_" + model_paths[-1]
+    else:
+        return model_paths[-1]
+
+
+class KeywordsStoppingCriteria(StoppingCriteria):
+    def __init__(self, keywords, tokenizer, input_ids):
+        self.keywords = keywords
+        self.keyword_ids = []
+        self.max_keyword_len = 0
+        for keyword in keywords:
+            cur_keyword_ids = tokenizer(keyword).input_ids
+            if len(cur_keyword_ids) > 1 and cur_keyword_ids[0] == tokenizer.bos_token_id:
+                cur_keyword_ids = cur_keyword_ids[1:]
+            if len(cur_keyword_ids) > self.max_keyword_len:
+                self.max_keyword_len = len(cur_keyword_ids)
+            self.keyword_ids.append(torch.tensor(cur_keyword_ids))
+        self.tokenizer = tokenizer
+        self.start_len = input_ids.shape[1]
+
+    def call_for_batch(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        offset = min(output_ids.shape[1] - self.start_len, self.max_keyword_len)
+        self.keyword_ids = [keyword_id.to(output_ids.device) for keyword_id in self.keyword_ids]
+        for keyword_id in self.keyword_ids:
+            if (output_ids[0, -keyword_id.shape[0]:] == keyword_id).all():
+                return True
+        outputs = self.tokenizer.batch_decode(output_ids[:, -offset:], skip_special_tokens=True)[0]
+        for keyword in self.keywords:
+            if keyword in outputs:
+                return True
+        return False
+
+    def __call__(self, output_ids: torch.LongTensor, scores: torch.FloatTensor, **kwargs) -> bool:
+        outputs = []
+        for i in range(output_ids.shape[0]):
+            outputs.append(self.call_for_batch(output_ids[i].unsqueeze(0), scores))
+        return all(outputs)
diff --git a/rynnec/model/__init__.py b/rynnec/model/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..6dc977849c165724eea8b3b06b43c20b6e6b9482
--- /dev/null
+++ b/rynnec/model/__init__.py
@@ -0,0 +1,196 @@
+# Adopted from https://github.com/haotian-liu/LLaVA. Below is the original copyright:
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+import os
+import warnings
+import shutil
+
+import torch
+from transformers import PretrainedConfig, AutoTokenizer, AutoModelForCausalLM, AutoConfig, BitsAndBytesConfig, AutoProcessor
+
+from .projector import load_mm_projector
+from .videollama3_encoder import Videollama3VisionEncoderModel, Videollama3VisionEncoderConfig
+from .rynnec_qwen2 import RynnecQwen2ForCausalLM, RynnecQwen2Config, Videollama3Qwen2Processor
+
+def apply_liger_kernel_to_rynnec():
+    from liger_kernel.transformers import (
+        apply_liger_kernel_to_mistral,
+        apply_liger_kernel_to_qwen2,
+        apply_liger_kernel_to_qwen3,
+        apply_liger_kernel_to_qwen3_moe,
+    )
+    from liger_kernel.transformers.rope import liger_rotary_pos_emb
+    from liger_kernel.transformers.layer_norm import LigerLayerNorm
+    from .videollama3_encoder import modeling_videollama3_encoder
+
+    apply_liger_kernel_to_mistral()
+    apply_liger_kernel_to_qwen2()
+
+    modeling_videollama3_encoder.apply_rotary_pos_emb_vision = liger_rotary_pos_emb
+    modeling_videollama3_encoder.LayerNorm = LigerLayerNorm
+
+
+def load_pretrained_model(model_path, model_base, model_name, load_8bit=False, load_4bit=False, device_map="auto", **kwargs):
+    if 'token' in kwargs:
+        token = kwargs['token']
+    else:
+        token = None
+
+    # NOTE: auto device_map by default
+    # if want to put model into a single device, you can set device_map={"": "cuda:0"}
+    kwargs = {"device_map": device_map, **kwargs}
+
+    config = AutoConfig.from_pretrained(model_path)
+    config._attn_implementation = kwargs.pop('attn_implementation', "flash_attention_2") # default to flash_attention_2
+
+    torch_dtype = config.torch_dtype if hasattr(config, "torch_dtype") else kwargs.pop('torch_dtype', torch.float16)
+
+    if load_8bit:
+        kwargs['load_in_8bit'] = True
+    elif load_4bit:
+        # NOTE: High-version Transformers will report: """ValueError: You can't pass `load_in_4bit`or `load_in_8bit` as a kwarg when passing `quantization_config` argument at the same time."""
+        # kwargs['load_in_4bit'] = True
+        kwargs['quantization_config'] = BitsAndBytesConfig(
+            load_in_4bit=True,
+            bnb_4bit_compute_dtype=torch_dtype,
+            bnb_4bit_use_double_quant=True,
+            bnb_4bit_quant_type='nf4'
+        )
+    else:
+        kwargs['torch_dtype'] = torch_dtype
+
+    # judge model type
+    model_type = config.model_type if hasattr(config, "model_type") else kwargs.pop('model_type', "rynnec_qwen2")
+
+    # judge pretrain/finetune
+    is_alignment = getattr(config, "tune_mm_mlp_adapter", False) or getattr(config, "is_alignment", False)
+
+    # NOTE: lora/qlora model loading
+    if 'lora' in model_name.lower() or 'qlora' in model_name.lower():
+    # if True:
+        cfg_pretrained = PretrainedConfig.from_pretrained(model_path, token=token)
+        # NOTE: AutoConfig will modify `_name_or_path` property to `model_path` if `model_path` is not None.
+        # cfg_pretrained = AutoConfig.from_pretrained(model_path, token=token)
+        model_base = model_base if model_base is not None else cfg_pretrained._name_or_path
+
+        # NOTE: remove qlora training quantization config 
+        if hasattr(cfg_pretrained, 'quantization_config'):
+            del cfg_pretrained.quantization_config
+        tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False, token=token)
+        print('Loading RynnEC from base model...')
+
+        config_raw = AutoConfig.from_pretrained(model_base)
+        new_vocab_size = config.vocab_size 
+        if config.vocab_size!=config_raw.vocab_size:
+            config.vocab_size = config_raw.vocab_size
+        config.training = False
+
+        if 'qwen2' in model_base.lower():
+            model = RynnecQwen2ForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=config, **kwargs)
+        else:
+            model = RynnecQwen2ForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=config, **kwargs)
+
+        model.config.mask_decoder_model = "./checkpoints/sam2_hiera_large.pt"
+       
+        token_num, tokem_dim = new_vocab_size, model.lm_head.in_features
+
+        if model.lm_head.weight.shape[0] != token_num:
+            model.lm_head.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
+            model.model.embed_tokens.weight = torch.nn.Parameter(torch.empty(token_num, tokem_dim, device=model.device, dtype=model.dtype))
+
+        print('Loading additional RynnEC weights...')
+        if os.path.exists(os.path.join(model_path, 'non_lora_trainables.bin')):
+            non_lora_trainables = torch.load(os.path.join(model_path, 'non_lora_trainables.bin'), map_location='cpu')
+        else:
+            # this is probably from HF Hub
+            from huggingface_hub import hf_hub_download
+            def load_from_hf(repo_id, filename, subfolder=None):
+                cache_file = hf_hub_download(
+                    repo_id=repo_id,
+                    filename=filename,
+                    subfolder=subfolder)
+                return torch.load(cache_file, map_location='cpu')
+            non_lora_trainables = load_from_hf(model_path, 'non_lora_trainables.bin')
+        
+        # add
+        sam2_model = torch.load(model.config.mask_decoder_model, map_location='cpu')['model']
+        prefix = "base_model.model.grounding_encoder.sam2_model."
+        for param_name in sam2_model.keys():
+            new_param_name = prefix + param_name
+            if new_param_name not in non_lora_trainables.keys():
+                non_lora_trainables[new_param_name] = sam2_model[param_name]
+
+        non_lora_trainables = {(k[11:] if k.startswith('base_model.') else k): v for k, v in non_lora_trainables.items()}
+        if any(k.startswith('model.model.') for k in non_lora_trainables):
+            non_lora_trainables = {(k[6:] if k.startswith('model.') else k): v for k, v in non_lora_trainables.items()}
+        model.load_state_dict(non_lora_trainables, strict=False)
+
+        from peft import PeftModel
+        print('Loading LoRA weights...')
+        model = PeftModel.from_pretrained(model, model_path)
+        print('Merging LoRA weights...')
+        model = model.merge_and_unload()
+        print('Model is loaded...')
+
+
+    elif model_base is not None or '-base' in model_name.lower() or is_alignment:
+        # NOTE: Base/Pretrain model loading
+        print('Loading RynnEC from base model...')
+        cfg_pretrained = PretrainedConfig.from_pretrained(model_path, token=token)
+        # NOTE: AutoConfig will modify `_name_or_path` property to `model_path` if `model_path` is not None.
+        # cfg_pretrained = AutoConfig.from_pretrained(model_path, token=token)
+        model_base = model_base if model_base is not None else cfg_pretrained._name_or_path
+
+        tokenizer = AutoTokenizer.from_pretrained(model_base, use_fast=False, token=token)
+
+        if model_type in ['rynnec', 'rynnec_qwen2']:
+            model = RynnecQwen2ForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=config, **kwargs)
+        else:
+            model = RynnecQwen2ForCausalLM.from_pretrained(model_base, low_cpu_mem_usage=True, config=config, **kwargs)
+
+        # NOTE; loading vision-language projector
+        # * old codes for loading local mm_projector.bin
+        # mm_projector_weights = torch.load(os.path.join(model_path, 'mm_projector.bin'), map_location='cpu')
+        # mm_projector_weights = {k: v.to(torch.float16) for k, v in mm_projector_weights.items()}
+        # model.load_state_dict(mm_projector_weights, strict=False)
+        # * new codes which supports loading mm_projector.bin both offline and online 
+        mm_projector_weights = load_mm_projector(model_path, token=token)
+        model.load_state_dict(mm_projector_weights, strict=False)
+    elif 'rynnec' in model_type:
+        # NOTE: SFT model loading
+        tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=False, token=token)
+
+        if model_type in ['rynnec_qwen2']:
+            model = RynnecQwen2ForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, config=config, **kwargs)
+        else:
+            model = AutoModelForCausalLM.from_pretrained(model_path, low_cpu_mem_usage=True, config=config, **kwargs)
+    else:
+        tokenizer = AutoTokenizer.from_pretrained(model_path, use_fast=True, token=token)
+        model = AutoModelForCausalLM.from_pretrained(model_path, config=config, **kwargs)
+
+    processor = None
+
+    # if "videollama" in model_type:
+    if True:
+        vision_encoder = model.get_vision_encoder()
+        processor = vision_encoder.image_processor
+
+    if hasattr(model.config, "max_sequence_length"):
+        context_len = model.config.max_sequence_length
+    else:
+        context_len = 2048
+
+    return tokenizer, model, processor, context_len
diff --git a/rynnec/model/encoder.py b/rynnec/model/encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..ecb347a5e3217b175b2f2269ad62842a4b008377
--- /dev/null
+++ b/rynnec/model/encoder.py
@@ -0,0 +1,282 @@
+import os
+
+import torch
+import torch.nn as nn
+from transformers import (CLIPImageProcessor, CLIPVisionConfig,
+                          CLIPVisionModel, SiglipImageProcessor,
+                          SiglipVisionConfig, SiglipVisionModel)
+
+from .videollama3_encoder import (Videollama3VisionEncoderConfig,
+    Videollama3VisionEncoderModel, Videollama3ImageProcessor)
+
+
+class CLIPVisionEncoder(nn.Module):
+
+    def __init__(self, vision_encoder, args, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+
+        self.vision_encoder_name = vision_encoder
+        self.select_layer = args.mm_vision_select_layer
+        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
+
+        if not delay_load:
+            self.attn_implementation = getattr(args, 'mm_attn_implementation', 'flash_attention_2')
+            self.load_model()
+        else:
+            # uncertain whether flash-attention-2 is supported during inference phase.
+            self.attn_implementation = 'sdpa' # 'eager'
+            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_encoder_name)
+
+    def load_model(self):
+        if self.is_loaded:
+            print('Vision tower is already loaded, `load model` call again, skipping.')
+            return
+
+        self.image_processor = CLIPImageProcessor.from_pretrained(self.vision_encoder_name)
+
+        self.vision_encoder = CLIPVisionModel.from_pretrained(self.vision_encoder_name,
+                                                            attn_implementation=self.attn_implementation)
+
+        self.is_loaded = True
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[self.select_layer]
+        if self.select_feature == 'patch':
+            image_features = image_features[:, 1:]
+        elif self.select_feature == 'cls_patch':
+            image_features = image_features
+        else:
+            raise ValueError(f'Unexpected select feature: {self.select_feature}')
+        return image_features
+
+    def forward(self, images, **kwargs):
+        images = torch.cat(images)
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_encoder(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_encoder(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_encoder.dtype
+
+    @property
+    def device(self):
+        return self.vision_encoder.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_encoder.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+    @property
+    def num_patches_per_side(self):
+        return self.config.image_size // self.config.patch_size
+
+    @property
+    def image_size(self):
+        return self.config.image_size
+
+
+class SiglipVisionEncoder(nn.Module):
+
+    def __init__(self, vision_encoder, args, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+
+        self.vision_encoder_name = vision_encoder
+        self.select_layer = args.mm_vision_select_layer
+        self.select_feature = getattr(args, 'mm_vision_select_feature', 'patch')
+
+        if not delay_load:
+            self.attn_implementation = getattr(args, 'mm_attn_implementation', 'flash_attention_2')
+            self.load_model()
+        else:
+            # uncertain whether flash-attention-2 is supported during inference phase.
+            self.attn_implementation = 'sdpa' # 'eager'
+            self.cfg_only = CLIPVisionConfig.from_pretrained(self.vision_encoder_name)
+
+    def load_model(self):
+        if self.is_loaded:
+            print('Vision tower is already loaded, `load model` call again, skipping.')
+            return
+
+        self.image_processor = SiglipImageProcessor.from_pretrained(self.vision_encoder_name)
+
+        self.vision_encoder = SiglipVisionModel.from_pretrained(self.vision_encoder_name,
+                                                              attn_implementation=self.attn_implementation)
+
+        self.is_loaded = True
+
+    def feature_select(self, image_forward_outs):
+        image_features = image_forward_outs.hidden_states[self.select_layer]
+        if self.select_feature == 'patch':
+            image_features = image_features
+        else:
+            raise ValueError(f'Unexpected select feature: {self.select_feature}')
+        return image_features
+
+    def forward(self, images, **kwargs):
+        images = torch.cat(images)
+        if type(images) is list:
+            image_features = []
+            for image in images:
+                image_forward_out = self.vision_encoder(image.to(device=self.device, dtype=self.dtype).unsqueeze(0), output_hidden_states=True)
+                image_feature = self.feature_select(image_forward_out).to(image.dtype)
+                image_features.append(image_feature)
+        else:
+            image_forward_outs = self.vision_encoder(images.to(device=self.device, dtype=self.dtype), output_hidden_states=True)
+            image_features = self.feature_select(image_forward_outs).to(images.dtype)
+
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_encoder.dtype
+
+    @property
+    def device(self):
+        return self.vision_encoder.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_encoder.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return (self.config.image_size // self.config.patch_size) ** 2
+
+    @property
+    def num_patches_per_side(self):
+        return self.config.image_size // self.config.patch_size
+
+    @property
+    def image_size(self):
+        return self.config.image_size
+
+
+class Videollama3VisionEncoder(nn.Module):
+
+    def __init__(self, vision_encoder, args, delay_load=False):
+        super().__init__()
+
+        self.is_loaded = False
+
+        self.vision_encoder_name = vision_encoder
+        self.args = args
+
+        if not delay_load:
+            self.attn_implementation = getattr(args, 'mm_attn_implementation', 'flash_attention_2')
+            self.load_model(self.args)
+        else:
+            # uncertain whether flash-attention-2 is supported during inference phase.
+            self.attn_implementation = 'sdpa' # 'eager'
+            self.cfg_only = Videollama3VisionEncoderConfig.from_pretrained(self.vision_encoder_name)
+
+    def load_model(self, args):
+        if self.is_loaded:
+            print('Vision tower is already loaded, `load model` call again, skipping.')
+            return
+
+        # merge_size is set to 1 by default, because STAGE1, STAGE1.5, STAGE2 are trained with merge_size=1
+        # for stage 3, the merge_size is set to 2 by argments. 
+        self.image_processor = Videollama3ImageProcessor.from_pretrained(self.vision_encoder_name)
+
+        # merge_size is fixed to 1 for STAGE1, STAGE1.5, STAGE2, STAGE3 in encoder and can be modified in connector.
+        self.cfg_only = Videollama3VisionEncoderConfig.from_pretrained(self.vision_encoder_name)
+
+        self.vision_encoder = Videollama3VisionEncoderModel.from_pretrained(
+            self.vision_encoder_name,
+            torch_dtype=args.torch_dtype,
+            attn_implementation=self.attn_implementation)
+
+        self.is_loaded = True
+
+    def forward(self, pixel_values, grid_sizes, merge_sizes, **kwargs):
+        image_features = self.vision_encoder(pixel_values, grid_sizes, merge_sizes)
+        return image_features
+
+    @property
+    def dummy_feature(self):
+        return torch.zeros(1, self.hidden_size, device=self.device, dtype=self.dtype)
+
+    @property
+    def dtype(self):
+        return self.vision_encoder.dtype
+
+    @property
+    def device(self):
+        return self.vision_encoder.device
+
+    @property
+    def config(self):
+        if self.is_loaded:
+            return self.vision_encoder.config
+        else:
+            return self.cfg_only
+
+    @property
+    def hidden_size(self):
+        return self.config.hidden_size
+
+    @property
+    def num_patches(self):
+        return -1
+
+    @property
+    def num_patches_per_side(self):
+        return -1
+
+    @property
+    def image_size(self):
+        return -1
+
+
+def build_vision_encoder(vision_encoder_cfg, **kwargs):
+    vision_encoder = getattr(vision_encoder_cfg, 'mm_vision_encoder', getattr(vision_encoder_cfg, 'vision_encoder', None))
+
+    if  'clip' in vision_encoder:
+        vision_encoder = CLIPVisionEncoder(vision_encoder, args=vision_encoder_cfg, **kwargs)
+    elif 'navit' in vision_encoder.lower() or 'damovl' in vision_encoder:
+        vision_encoder = Videollama3VisionEncoder(vision_encoder, args=vision_encoder_cfg, **kwargs)
+    elif 'siglip' in vision_encoder:
+        vision_encoder = SiglipVisionEncoder(vision_encoder, args=vision_encoder_cfg, **kwargs)
+    else:
+        raise ValueError(f'Unknown vision encoder: {vision_encoder}')
+
+    return vision_encoder
diff --git a/rynnec/model/extension/__init__.py b/rynnec/model/extension/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..a5a69d58ba75749633d98e11a6792654958c9e0c
--- /dev/null
+++ b/rynnec/model/extension/__init__.py
@@ -0,0 +1 @@
+from .sam2_base import SAM2Base
diff --git a/rynnec/model/extension/sam2_base.py b/rynnec/model/extension/sam2_base.py
new file mode 100644
index 0000000000000000000000000000000000000000..55325b67d019003d872b3e2e9ba4f7d890a26769
--- /dev/null
+++ b/rynnec/model/extension/sam2_base.py
@@ -0,0 +1,298 @@
+# Adopted from https://github.com/magic-research/Sa2VA/blob/main/projects/llava_sam2/models/extension/sam2_base.py.
+# Below is the original copyright:
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+import torch
+import torch.nn.functional as F
+
+from third_parts.sam2.modeling.sam2_base import SAM2Base as _SAM2Base
+from third_parts.sam2.modeling.sam2_base import NO_OBJ_SCORE
+
+
+class SAM2Base(_SAM2Base):
+
+    def track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
+        # to skip the memory encoder with `run_mem_encoder=False`. For example,
+        # in demo we might call `track_step` multiple times for each user click,
+        # and only encode the memory when the user finalizes their clicks. And in ablation
+        # settings like SAM training on static images, we don't need the memory encoder.
+        run_mem_encoder=True,
+        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
+        prev_sam_mask_logits=None,
+        ## Extension: LLM prompt
+        language_embd=None,
+    ):
+        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
+        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
+        if len(current_vision_feats) > 1:
+            high_res_features = [
+                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
+                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
+            ]
+        else:
+            high_res_features = None
+        if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
+            # When use_mask_input_as_output_without_sam=True, we directly output the mask input
+            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
+            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
+            sam_outputs = self._use_mask_as_output(
+                pix_feat, high_res_features, mask_inputs
+            )
+        else:
+            # fused the visual feature with previous memory features in the memory bank
+            pix_feat_with_mem = self._prepare_memory_conditioned_features(
+                frame_idx=frame_idx,
+                is_init_cond_frame=is_init_cond_frame,
+                current_vision_feats=current_vision_feats[-1:],
+                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
+                feat_sizes=feat_sizes[-1:],
+                output_dict=output_dict,
+                num_frames=num_frames,
+                track_in_reverse=track_in_reverse,
+            )
+            # apply SAM-style segmentation head
+            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
+            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
+            # (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
+            if prev_sam_mask_logits is not None:
+                assert point_inputs is not None and mask_inputs is None
+                mask_inputs = prev_sam_mask_logits
+            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
+            sam_outputs = self._forward_sam_heads(
+                backbone_features=pix_feat_with_mem,
+                point_inputs=point_inputs,
+                mask_inputs=mask_inputs,
+                high_res_features=high_res_features,
+                multimask_output=multimask_output,
+                # Inject language Embed if possible
+                language_embd=language_embd,
+            )
+        (
+            _,
+            _,
+            _,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            _,
+        ) = sam_outputs
+
+        current_out["pred_masks"] = low_res_masks
+        current_out["pred_masks_high_res"] = high_res_masks
+        current_out["obj_ptr"] = obj_ptr
+
+        # Finally run the memory encoder on the predicted mask to encode
+        # it into a new memory feature (that can be used in future frames)
+        if run_mem_encoder and self.num_maskmem > 0:
+            high_res_masks_for_mem_enc = high_res_masks
+            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+                current_vision_feats=current_vision_feats,
+                feat_sizes=feat_sizes,
+                pred_masks_high_res=high_res_masks_for_mem_enc,
+                is_mask_from_pts=(point_inputs is not None),
+            )
+            current_out["maskmem_features"] = maskmem_features
+            current_out["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            current_out["maskmem_features"] = None
+            current_out["maskmem_pos_enc"] = None
+
+        return current_out
+
+
+    def _forward_sam_heads(
+        self,
+        backbone_features,
+        point_inputs=None,
+        mask_inputs=None,
+        high_res_features=None,
+        multimask_output=False,
+        ## Extension: LLM prompt
+        language_embd=None,
+    ):
+        """
+        Forward SAM prompt encoders and mask heads.
+
+        Inputs:
+        - backbone_features: image features of [B, C, H, W] shape
+        - point_inputs: a dictionary with "point_coords" and "point_labels", where
+          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
+             absolute pixel-unit coordinate in (x, y) format of the P input points
+          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
+             positive clicks, 0 means negative clicks, and -1 means padding
+        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
+          same spatial size as the image.
+        - high_res_features: either 1) None or 2) or a list of length 2 containing
+          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
+          which will be used as high-resolution feature maps for SAM decoder.
+        - multimask_output: if it's True, we output 3 candidate masks and their 3
+          corresponding IoU estimates, and if it's False, we output only 1 mask and
+          its corresponding IoU estimate.
+
+        Outputs:
+        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
+          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
+          output mask logits (before sigmoid) for the low-resolution masks, with 4x
+          the resolution (1/4 stride) of the input backbone_features.
+        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
+          if `multimask_output=True` and M = 1 if `multimask_output=False`),
+          upsampled from the low-resolution masks, with shape size as the image
+          (stride is 1 pixel).
+        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
+          if `multimask_output=False`), the estimated IoU of each output mask.
+        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `low_res_multimasks`.
+        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `high_res_multimasks`.
+        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
+          based on the output token from the SAM mask decoder.
+        """
+        B = backbone_features.size(0)
+        device = backbone_features.device
+        assert backbone_features.size(1) == self.sam_prompt_embed_dim
+        assert backbone_features.size(2) == self.sam_image_embedding_size
+        assert backbone_features.size(3) == self.sam_image_embedding_size
+
+        # a) Handle point prompts
+        if point_inputs is not None:
+            sam_point_coords = point_inputs["point_coords"]
+            sam_point_labels = point_inputs["point_labels"]
+            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
+        else:
+            # If no points are provide, pad with an empty point (with label -1)
+            sam_point_coords = torch.zeros(B, 1, 2, device=device)
+            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)
+
+        # b) Handle mask prompts
+        if mask_inputs is not None:
+            # If mask_inputs is provided, downsize it into low-res mask input if needed
+            # and feed it as a dense mask prompt into the SAM mask encoder
+            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
+            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
+                sam_mask_prompt = F.interpolate(
+                    mask_inputs.float(),
+                    size=self.sam_prompt_encoder.mask_input_size,
+                    align_corners=False,
+                    mode="bilinear",
+                    antialias=True,  # use antialias for downsampling
+                )
+            else:
+                sam_mask_prompt = mask_inputs
+        else:
+            # Otherwise, simply feed None (and SAM's prompt encoder will add
+            # a learned `no_mask_embed` to indicate no mask input in this case).
+            sam_mask_prompt = None
+
+        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
+            points=(sam_point_coords, sam_point_labels),
+            boxes=None,
+            masks=sam_mask_prompt,
+        )
+
+        ## Extension: LLM prompt
+        if language_embd is not None:
+            # B N C
+            # print('sparse_embeddings ', sparse_embeddings.shape, 'language_embd ', language_embd.shape)
+            assert sparse_embeddings.size(0) == language_embd.size(0)
+            assert sparse_embeddings.size(2) == language_embd.size(2)
+            sparse_embeddings = torch.cat([sparse_embeddings, language_embd], dim=1)
+
+        (
+            low_res_multimasks,
+            ious,
+            sam_output_tokens,
+            object_score_logits,
+        ) = self.sam_mask_decoder(
+            image_embeddings=backbone_features,
+            image_pe=self.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=False,  # the image is already batched
+            high_res_features=high_res_features,
+        )
+        if self.pred_obj_scores:
+            is_obj_appearing = object_score_logits > 0
+
+            # Mask used for spatial memories is always a *hard* choice between obj and no obj,
+            # consistent with the actual mask prediction
+            # print('Do torch.where !!!')
+            # low_res_multimasks = torch.where(
+            #     is_obj_appearing[:, None, None],
+            #     low_res_multimasks,
+            #     NO_OBJ_SCORE,
+            # )
+
+        # convert masks from possibly bfloat16 (or float16) to float32
+        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
+        low_res_multimasks = low_res_multimasks.float()
+        high_res_multimasks = F.interpolate(
+            low_res_multimasks,
+            size=(self.image_size, self.image_size),
+            mode="bilinear",
+            align_corners=False,
+        )
+
+        sam_output_token = sam_output_tokens[:, 0]
+        if multimask_output:
+            # take the best mask prediction (with the highest IoU estimation)
+            best_iou_inds = torch.argmax(ious, dim=-1)
+            batch_inds = torch.arange(B, device=device)
+            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            if sam_output_tokens.size(1) > 1:
+                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
+        else:
+            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks
+
+        # Extract object pointer from the SAM output token (with occlusion handling)
+        obj_ptr = self.obj_ptr_proj(sam_output_token)
+        if self.pred_obj_scores:
+            # Allow *soft* no obj ptr, unlike for masks
+            if self.soft_no_obj_ptr:
+                # Only hard possible with gt
+                assert not self.teacher_force_obj_scores_for_mem
+                lambda_is_obj_appearing = object_score_logits.sigmoid()
+            else:
+                lambda_is_obj_appearing = is_obj_appearing.float()
+
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_multimasks,
+            high_res_multimasks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
diff --git a/rynnec/model/loss.py b/rynnec/model/loss.py
new file mode 100644
index 0000000000000000000000000000000000000000..983a5b6645cb7576c5878cfa52a8030fefb80beb
--- /dev/null
+++ b/rynnec/model/loss.py
@@ -0,0 +1,597 @@
+# Adopted from https://github.com/magic-research/Sa2VA.
+# Below is the original copyright:
+# coding=utf-8
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+linear_cross_entropy = None
+
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from rynnec.constants import IGNORE_INDEX
+from torch import Tensor
+import logging
+from huggingface_hub import hf_hub_download
+import functools
+from typing import Callable, Optional
+
+
+def reduce_loss(loss: Tensor, reduction: str) -> Tensor:
+    """Reduce loss as specified.
+
+    Args:
+        loss (Tensor): Elementwise loss tensor.
+        reduction (str): Options are "none", "mean" and "sum".
+
+    Return:
+        Tensor: Reduced loss tensor.
+    """
+    reduction_enum = F._Reduction.get_enum(reduction)
+    # none: 0, elementwise_mean:1, sum: 2
+    if reduction_enum == 0:
+        return loss
+    elif reduction_enum == 1:
+        return loss.mean()
+    elif reduction_enum == 2:
+        return loss.sum()
+
+
+def weight_reduce_loss(loss: Tensor,
+                       weight: Optional[Tensor] = None,
+                       reduction: str = 'mean',
+                       avg_factor: Optional[float] = None) -> Tensor:
+    """Apply element-wise weight and reduce loss.
+
+    Args:
+        loss (Tensor): Element-wise loss.
+        weight (Optional[Tensor], optional): Element-wise weights.
+            Defaults to None.
+        reduction (str, optional): Same as built-in losses of PyTorch.
+            Defaults to 'mean'.
+        avg_factor (Optional[float], optional): Average factor when
+            computing the mean of losses. Defaults to None.
+
+    Returns:
+        Tensor: Processed loss values.
+    """
+    # if weight is specified, apply element-wise weight
+    if weight is not None:
+        loss = loss * weight
+
+    # if avg_factor is not specified, just reduce the loss
+    if avg_factor is None:
+        loss = reduce_loss(loss, reduction)
+    else:
+        # if reduction is mean, then average the loss by avg_factor
+        if reduction == 'mean':
+            # Avoid causing ZeroDivisionError when avg_factor is 0.0,
+            # i.e., all labels of an image belong to ignore index.
+            eps = torch.finfo(torch.float32).eps
+            loss = loss.sum() / (avg_factor + eps)
+        # if reduction is 'none', then do nothing, otherwise raise an error
+        elif reduction != 'none':
+            raise ValueError('avg_factor can not be used with reduction="sum"')
+    return loss
+
+
+def dice_loss(pred,
+              target,
+              weight=None,
+              eps=1e-3,
+              reduction='mean',
+              naive_dice=False,
+              avg_factor=None):
+    """Calculate dice loss, there are two forms of dice loss is supported:
+
+        - the one proposed in `V-Net: Fully Convolutional Neural
+            Networks for Volumetric Medical Image Segmentation
+            <https://arxiv.org/abs/1606.04797>`_.
+        - the dice loss in which the power of the number in the
+            denominator is the first power instead of the second
+            power.
+
+    Args:
+        pred (torch.Tensor): The prediction, has a shape (n, *)
+        target (torch.Tensor): The learning label of the prediction,
+            shape (n, *), same shape of pred.
+        weight (torch.Tensor, optional): The weight of loss for each
+            prediction, has a shape (n,). Defaults to None.
+        eps (float): Avoid dividing by zero. Default: 1e-3.
+        reduction (str, optional): The method used to reduce the loss into
+            a scalar. Defaults to 'mean'.
+            Options are "none", "mean" and "sum".
+        naive_dice (bool, optional): If false, use the dice
+                loss defined in the V-Net paper, otherwise, use the
+                naive dice loss in which the power of the number in the
+                denominator is the first power instead of the second
+                power.Defaults to False.
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+    """
+
+    input = pred.flatten(1)
+    target = target.flatten(1).float()
+
+    a = torch.sum(input * target, 1)
+    if naive_dice:
+        b = torch.sum(input, 1)
+        c = torch.sum(target, 1)
+        d = (2 * a + eps) / (b + c + eps)
+    else:
+        b = torch.sum(input * input, 1) + eps
+        c = torch.sum(target * target, 1) + eps
+        d = (2 * a) / (b + c)
+
+    loss = 1 - d
+    if weight is not None:
+        assert weight.ndim == loss.ndim
+        assert len(weight) == len(pred)
+    loss = weight_reduce_loss(loss, weight, reduction, avg_factor)
+    return loss
+
+
+
+class DiceLoss(nn.Module):
+
+    def __init__(self,
+                 use_sigmoid=True,
+                 activate=True,
+                 reduction='mean',
+                 naive_dice=False,
+                 loss_weight=1.0,
+                 eps=1e-3):
+        """Compute dice loss.
+
+        Args:
+            use_sigmoid (bool, optional): Whether to the prediction is
+                used for sigmoid or softmax. Defaults to True.
+            activate (bool): Whether to activate the predictions inside,
+                this will disable the inside sigmoid operation.
+                Defaults to True.
+            reduction (str, optional): The method used
+                to reduce the loss. Options are "none",
+                "mean" and "sum". Defaults to 'mean'.
+            naive_dice (bool, optional): If false, use the dice
+                loss defined in the V-Net paper, otherwise, use the
+                naive dice loss in which the power of the number in the
+                denominator is the first power instead of the second
+                power. Defaults to False.
+            loss_weight (float, optional): Weight of loss. Defaults to 1.0.
+            eps (float): Avoid dividing by zero. Defaults to 1e-3.
+        """
+
+        super(DiceLoss, self).__init__()
+        self.use_sigmoid = use_sigmoid
+        self.reduction = reduction
+        self.naive_dice = naive_dice
+        self.loss_weight = loss_weight
+        self.eps = eps
+        self.activate = activate
+
+    def forward(self,
+                pred,
+                target,
+                weight=None,
+                reduction_override=None,
+                avg_factor=None):
+        """Forward function.
+
+        Args:
+            pred (torch.Tensor): The prediction, has a shape (n, *).
+            target (torch.Tensor): The label of the prediction,
+                shape (n, *), same shape of pred.
+            weight (torch.Tensor, optional): The weight of loss for each
+                prediction, has a shape (n,). Defaults to None.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The reduction method used to
+                override the original reduction method of the loss.
+                Options are "none", "mean" and "sum".
+
+        Returns:
+            torch.Tensor: The calculated loss
+        """
+
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+
+        if self.activate:
+            if self.use_sigmoid:
+                pred = pred.sigmoid()
+            else:
+                raise NotImplementedError
+
+        loss = self.loss_weight * dice_loss(
+            pred,
+            target,
+            weight,
+            eps=self.eps,
+            reduction=reduction,
+            naive_dice=self.naive_dice,
+            avg_factor=avg_factor)
+
+        return loss
+
+
+def cross_entropy_loss(
+    hidden_states,
+    lm_head,
+    position_ids,
+    labels,
+    reduction_scope="sequence",
+    **loss_kwargs
+):
+    batch_size = hidden_states.size(0)
+
+    shift_hidden_states = hidden_states[..., :-1, :]
+    shift_labels = labels[..., 1:]
+    mask = shift_labels != IGNORE_INDEX
+    shift_hidden_states = shift_hidden_states[mask].contiguous()
+    shift_labels = shift_labels[mask].contiguous()
+
+    if mask.sum() == 0:
+        print(f"Get labels={labels}. Found no sample to calculate loss!")
+        pseudo_logits = lm_head(hidden_states[:, 0:1])
+        loss = 0.0 * pseudo_logits.mean()
+        return loss
+
+    if "num_items_in_batch" not in loss_kwargs:
+        reduction = "mean"
+        denominator = None
+
+    elif reduction_scope == "batch":
+        reduction = "sum"
+        denominator = loss_kwargs["num_items_in_batch"]
+
+    elif reduction_scope == "sequence":
+        reduction = "none"
+
+        if batch_size == 1:
+            # NOTE: packed sequence
+            start_indices = torch.nonzero(position_ids[0] == 0)[:, 0]
+            end_indices = F.pad(start_indices[1:], (0, 1), value=position_ids.size(1))
+            batch_indices = torch.cat(
+                [
+                    torch.full((e - s,), fill_value=i, device=position_ids.device, dtype=torch.long)
+                    for i, (s, e) in enumerate(zip(start_indices, end_indices))
+                ],
+            ).unsqueeze(0)
+        else:
+            batch_indices = torch.arange(batch_size, device=position_ids.device)
+            batch_indices = batch_indices.unsqueeze(1).expand(-1, hidden_states.size(1))
+
+        shift_batch_indices = batch_indices[..., :-1]
+        shift_batch_indices = shift_batch_indices[mask].contiguous()
+        num_tokens = F.one_hot(shift_batch_indices).sum(dim=0)
+        denominator = num_tokens[shift_batch_indices] * loss_kwargs["num_items_in_batch"]
+
+    else:
+        raise ValueError(f"Unknown reduction scope: {reduction_scope}")
+
+    if linear_cross_entropy is None:
+        shift_logits = lm_head(shift_hidden_states)
+        loss = torch.nn.functional.cross_entropy(
+            shift_logits,
+            shift_labels,
+            reduction=reduction,
+        )
+    else:
+        loss = linear_cross_entropy(
+            shift_hidden_states,
+            lm_head.weight,
+            shift_labels,
+            bias=lm_head.bias,
+            reduction=reduction,
+            accum_e_fp32=True,
+            accum_c_fp32=True,
+        )
+
+    if denominator is not None:
+        loss = loss / denominator
+        if loss.ndim > 0:
+            loss = loss.sum()
+
+    return loss
+
+
+
+def cross_entropy(pred,
+                  label,
+                  weight=None,
+                  reduction='mean',
+                  avg_factor=None,
+                  class_weight=None,
+                  ignore_index=-100,
+                  avg_non_ignore=False):
+    """Calculate the CrossEntropy loss.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C), C is the number
+            of classes.
+        label (torch.Tensor): The learning label of the prediction.
+        weight (torch.Tensor, optional): Sample-wise loss weight.
+        reduction (str, optional): The method used to reduce the loss.
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+        class_weight (list[float], optional): The weight for each class.
+        ignore_index (int | None): The label index to be ignored.
+            If None, it will be set to default value. Default: -100.
+        avg_non_ignore (bool): The flag decides to whether the loss is
+            only averaged over non-ignored targets. Default: False.
+
+    Returns:
+        torch.Tensor: The calculated loss
+    """
+    # The default value of ignore_index is the same as F.cross_entropy
+    ignore_index = -100 if ignore_index is None else ignore_index
+    # element-wise losses
+    loss = F.cross_entropy(
+        pred,
+        label,
+        weight=class_weight,
+        reduction='none',
+        ignore_index=ignore_index)
+
+    # average loss over non-ignored elements
+    # pytorch's official cross_entropy average loss over non-ignored elements
+    # refer to https://github.com/pytorch/pytorch/blob/56b43f4fec1f76953f15a627694d4bba34588969/torch/nn/functional.py#L2660  # noqa
+    if (avg_factor is None) and avg_non_ignore and reduction == 'mean':
+        avg_factor = label.numel() - (label == ignore_index).sum().item()
+
+    # apply weights and do the reduction
+    if weight is not None:
+        weight = weight.float()
+    loss = weight_reduce_loss(
+        loss, weight=weight, reduction=reduction, avg_factor=avg_factor)
+
+    return loss
+
+
+def _expand_onehot_labels(labels, label_weights, label_channels, ignore_index):
+    """Expand onehot labels to match the size of prediction."""
+    bin_labels = labels.new_full((labels.size(0), label_channels), 0)
+    valid_mask = (labels >= 0) & (labels != ignore_index)
+    inds = torch.nonzero(
+        valid_mask & (labels < label_channels), as_tuple=False)
+
+    if inds.numel() > 0:
+        bin_labels[inds, labels[inds]] = 1
+
+    valid_mask = valid_mask.view(-1, 1).expand(labels.size(0),
+                                               label_channels).float()
+    if label_weights is None:
+        bin_label_weights = valid_mask
+    else:
+        bin_label_weights = label_weights.view(-1, 1).repeat(1, label_channels)
+        bin_label_weights *= valid_mask
+
+    return bin_labels, bin_label_weights, valid_mask
+
+
+def binary_cross_entropy(pred,
+                         label,
+                         weight=None,
+                         reduction='mean',
+                         avg_factor=None,
+                         class_weight=None,
+                         ignore_index=-100,
+                         avg_non_ignore=False):
+    """Calculate the binary CrossEntropy loss.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, 1) or (N, ).
+            When the shape of pred is (N, 1), label will be expanded to
+            one-hot format, and when the shape of pred is (N, ), label
+            will not be expanded to one-hot format.
+        label (torch.Tensor): The learning label of the prediction,
+            with shape (N, ).
+        weight (torch.Tensor, optional): Sample-wise loss weight.
+        reduction (str, optional): The method used to reduce the loss.
+            Options are "none", "mean" and "sum".
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+        class_weight (list[float], optional): The weight for each class.
+        ignore_index (int | None): The label index to be ignored.
+            If None, it will be set to default value. Default: -100.
+        avg_non_ignore (bool): The flag decides to whether the loss is
+            only averaged over non-ignored targets. Default: False.
+
+    Returns:
+        torch.Tensor: The calculated loss.
+    """
+    # The default value of ignore_index is the same as F.cross_entropy
+    ignore_index = -100 if ignore_index is None else ignore_index
+
+    if pred.dim() != label.dim():
+        label, weight, valid_mask = _expand_onehot_labels(
+            label, weight, pred.size(-1), ignore_index)
+    else:
+        # should mask out the ignored elements
+        valid_mask = ((label >= 0) & (label != ignore_index)).float()
+        if weight is not None:
+            # The inplace writing method will have a mismatched broadcast
+            # shape error if the weight and valid_mask dimensions
+            # are inconsistent such as (B,N,1) and (B,N,C).
+            weight = weight * valid_mask
+        else:
+            weight = valid_mask
+
+    # average loss over non-ignored elements
+    if (avg_factor is None) and avg_non_ignore and reduction == 'mean':
+        avg_factor = valid_mask.sum().item()
+
+    # weighted element-wise losses
+    weight = weight.float()
+    loss = F.binary_cross_entropy_with_logits(
+        pred, label.float(), pos_weight=class_weight, reduction='none')
+    # do the reduction for the weighted loss
+    loss = weight_reduce_loss(
+        loss, weight, reduction=reduction, avg_factor=avg_factor)
+
+    return loss
+
+
+def mask_cross_entropy(pred,
+                       target,
+                       label,
+                       reduction='mean',
+                       avg_factor=None,
+                       class_weight=None,
+                       ignore_index=None,
+                       **kwargs):
+    """Calculate the CrossEntropy loss for masks.
+
+    Args:
+        pred (torch.Tensor): The prediction with shape (N, C, *), C is the
+            number of classes. The trailing * indicates arbitrary shape.
+        target (torch.Tensor): The learning label of the prediction.
+        label (torch.Tensor): ``label`` indicates the class label of the mask
+            corresponding object. This will be used to select the mask in the
+            of the class which the object belongs to when the mask prediction
+            if not class-agnostic.
+        reduction (str, optional): The method used to reduce the loss.
+            Options are "none", "mean" and "sum".
+        avg_factor (int, optional): Average factor that is used to average
+            the loss. Defaults to None.
+        class_weight (list[float], optional): The weight for each class.
+        ignore_index (None): Placeholder, to be consistent with other loss.
+            Default: None.
+
+    Returns:
+        torch.Tensor: The calculated loss
+
+    Example:
+        >>> N, C = 3, 11
+        >>> H, W = 2, 2
+        >>> pred = torch.randn(N, C, H, W) * 1000
+        >>> target = torch.rand(N, H, W)
+        >>> label = torch.randint(0, C, size=(N,))
+        >>> reduction = 'mean'
+        >>> avg_factor = None
+        >>> class_weights = None
+        >>> loss = mask_cross_entropy(pred, target, label, reduction,
+        >>>                           avg_factor, class_weights)
+        >>> assert loss.shape == (1,)
+    """
+    assert ignore_index is None, 'BCE loss does not support ignore_index'
+    # TODO: handle these two reserved arguments
+    assert reduction == 'mean' and avg_factor is None
+    num_rois = pred.size()[0]
+    inds = torch.arange(0, num_rois, dtype=torch.long, device=pred.device)
+    pred_slice = pred[inds, label].squeeze(1)
+    return F.binary_cross_entropy_with_logits(
+        pred_slice, target, weight=class_weight, reduction='mean')[None]
+
+
+class CrossEntropyLoss(nn.Module):
+
+    def __init__(self,
+                 use_sigmoid=False,
+                 use_mask=False,
+                 reduction='mean',
+                 class_weight=None,
+                 ignore_index=None,
+                 loss_weight=1.0,
+                 avg_non_ignore=False):
+        """CrossEntropyLoss.
+
+        Args:
+            use_sigmoid (bool, optional): Whether the prediction uses sigmoid
+                of softmax. Defaults to False.
+            use_mask (bool, optional): Whether to use mask cross entropy loss.
+                Defaults to False.
+            reduction (str, optional): . Defaults to 'mean'.
+                Options are "none", "mean" and "sum".
+            class_weight (list[float], optional): Weight of each class.
+                Defaults to None.
+            ignore_index (int | None): The label index to be ignored.
+                Defaults to None.
+            loss_weight (float, optional): Weight of the loss. Defaults to 1.0.
+            avg_non_ignore (bool): The flag decides to whether the loss is
+                only averaged over non-ignored targets. Default: False.
+        """
+        super(CrossEntropyLoss, self).__init__()
+        assert (use_sigmoid is False) or (use_mask is False)
+        self.use_sigmoid = use_sigmoid
+        self.use_mask = use_mask
+        self.reduction = reduction
+        self.loss_weight = loss_weight
+        self.class_weight = class_weight
+        self.ignore_index = ignore_index
+        self.avg_non_ignore = avg_non_ignore
+        if ((ignore_index is not None) and not self.avg_non_ignore
+                and self.reduction == 'mean'):
+            warnings.warn(
+                'Default ``avg_non_ignore`` is False, if you would like to '
+                'ignore the certain label and average loss over non-ignore '
+                'labels, which is the same with PyTorch official '
+                'cross_entropy, set ``avg_non_ignore=True``.')
+
+        if self.use_sigmoid:
+            self.cls_criterion = binary_cross_entropy
+        elif self.use_mask:
+            self.cls_criterion = mask_cross_entropy
+        else:
+            self.cls_criterion = cross_entropy
+
+    def extra_repr(self):
+        """Extra repr."""
+        s = f'avg_non_ignore={self.avg_non_ignore}'
+        return s
+
+    def forward(self,
+                cls_score,
+                label,
+                weight=None,
+                avg_factor=None,
+                reduction_override=None,
+                ignore_index=None,
+                **kwargs):
+        """Forward function.
+
+        Args:
+            cls_score (torch.Tensor): The prediction.
+            label (torch.Tensor): The learning label of the prediction.
+            weight (torch.Tensor, optional): Sample-wise loss weight.
+            avg_factor (int, optional): Average factor that is used to average
+                the loss. Defaults to None.
+            reduction_override (str, optional): The method used to reduce the
+                loss. Options are "none", "mean" and "sum".
+            ignore_index (int | None): The label index to be ignored.
+                If not None, it will override the default value. Default: None.
+        Returns:
+            torch.Tensor: The calculated loss.
+        """
+        assert reduction_override in (None, 'none', 'mean', 'sum')
+        reduction = (
+            reduction_override if reduction_override else self.reduction)
+        if ignore_index is None:
+            ignore_index = self.ignore_index
+
+        if self.class_weight is not None:
+            class_weight = cls_score.new_tensor(
+                self.class_weight, device=cls_score.device)
+        else:
+            class_weight = None
+        loss_cls = self.loss_weight * self.cls_criterion(
+            cls_score,
+            label,
+            weight,
+            class_weight=class_weight,
+            reduction=reduction,
+            avg_factor=avg_factor,
+            ignore_index=ignore_index,
+            avg_non_ignore=self.avg_non_ignore,
+            **kwargs)
+        return loss_cls
diff --git a/rynnec/model/predictor/__init__.py b/rynnec/model/predictor/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..21f3bb13f0b8da1b1166f6ff826f7e605bee981e
--- /dev/null
+++ b/rynnec/model/predictor/__init__.py
@@ -0,0 +1 @@
+from .sam2_predictor import SAM2VideoPredictor
diff --git a/rynnec/model/predictor/sam2_predictor.py b/rynnec/model/predictor/sam2_predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..ff971ea3bf87de0ede3e3c8e10dec3215ba682ee
--- /dev/null
+++ b/rynnec/model/predictor/sam2_predictor.py
@@ -0,0 +1,724 @@
+# Adopted from https://github.com/magic-research/Sa2VA/blob/main/projects/llava_sam2/models/predictor/sam2_predictor.py.
+# Below is the original copyright:
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+from collections import OrderedDict
+
+import torch
+from tqdm import tqdm
+
+from ..extension import SAM2Base
+from third_parts.sam2.modeling.sam2_base import NO_OBJ_SCORE
+from third_parts.sam2.utils.misc import fill_holes_in_mask_scores
+
+
+def _obj_id_to_idx(inference_state, obj_id):
+    """Map client-side object id to model-side object index."""
+    obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
+    if obj_idx is not None:
+        return obj_idx
+
+    # This is a new object id not sent to the server before. We only allow adding
+    # new objects *before* the tracking starts.
+    allow_new_object = not inference_state["tracking_has_started"]
+    if allow_new_object:
+        # get the next object slot
+        obj_idx = len(inference_state["obj_id_to_idx"])
+        inference_state["obj_id_to_idx"][obj_id] = obj_idx
+        inference_state["obj_idx_to_id"][obj_idx] = obj_id
+        inference_state["obj_ids"] = list(inference_state["obj_id_to_idx"])
+        # set up input and output structures for this object
+        inference_state["point_inputs_per_obj"][obj_idx] = {}
+        inference_state["mask_inputs_per_obj"][obj_idx] = {}
+        inference_state["output_dict_per_obj"][obj_idx] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        inference_state["temp_output_dict_per_obj"][obj_idx] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        return obj_idx
+    else:
+        raise RuntimeError(
+            f"Cannot add new object id {obj_id} after tracking starts. "
+            f"All existing object ids: {inference_state['obj_ids']}. "
+            f"Please call 'reset_state' to restart from scratch."
+        )
+
+
+def _get_maskmem_pos_enc(inference_state, current_out):
+    """
+    `maskmem_pos_enc` is the same across frames and objects, so we cache it as
+    a constant in the inference session to reduce session storage size.
+    """
+    model_constants = inference_state["constants"]
+    # "out_maskmem_pos_enc" should be either a list of tensors or None
+    out_maskmem_pos_enc = current_out["maskmem_pos_enc"]
+    if out_maskmem_pos_enc is not None:
+        if "maskmem_pos_enc" not in model_constants:
+            assert isinstance(out_maskmem_pos_enc, list)
+            # only take the slice for one object, since it's same across objects
+            maskmem_pos_enc = [x[0:1].clone() for x in out_maskmem_pos_enc]
+            model_constants["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            maskmem_pos_enc = model_constants["maskmem_pos_enc"]
+        # expand the cached maskmem_pos_enc to the actual batch size
+        batch_size = out_maskmem_pos_enc[0].size(0)
+        expanded_maskmem_pos_enc = [
+            x.expand(batch_size, -1, -1, -1) for x in maskmem_pos_enc
+        ]
+    else:
+        expanded_maskmem_pos_enc = None
+    return expanded_maskmem_pos_enc
+
+
+def _obj_idx_to_id(inference_state, obj_idx):
+    """Map model-side object index to client-side object id."""
+    return inference_state["obj_idx_to_id"][obj_idx]
+
+
+def _get_obj_num(inference_state):
+    """Get the total number of unique object ids received so far in this session."""
+    return len(inference_state["obj_idx_to_id"])
+
+
+class SAM2VideoPredictor(SAM2Base):
+    """The predictor class to handle user interactions and manage inference states."""
+
+    def __init__(
+        self,
+        fill_hole_area=0,
+        # whether to apply non-overlapping constraints on the output object masks
+        non_overlap_masks=False,
+        # whether to clear non-conditioning memory of the surrounding frames (which may contain outdated information) after adding correction clicks;
+        # note that this would only apply to *single-object tracking* unless `clear_non_cond_mem_for_multi_obj` is also set to True)
+        clear_non_cond_mem_around_input=False,
+        # whether to also clear non-conditioning memory of the surrounding frames (only effective when `clear_non_cond_mem_around_input` is True).
+        clear_non_cond_mem_for_multi_obj=False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.fill_hole_area = fill_hole_area
+        self.non_overlap_masks = non_overlap_masks
+        self.clear_non_cond_mem_around_input = clear_non_cond_mem_around_input
+        self.clear_non_cond_mem_for_multi_obj = clear_non_cond_mem_for_multi_obj
+
+    def _get_image_feature(self, inference_state, frame_idx, batch_size):
+        """Compute the image features on a given frame."""
+        # Look up in the cache first
+        image, backbone_out = inference_state["cached_features"].get(
+            frame_idx, (None, None)
+        )
+        if backbone_out is None:
+            # Cache miss -- we will run inference on a single image
+            image = inference_state["images"][frame_idx].cuda().float().unsqueeze(0)
+            backbone_out = self.forward_image(image)
+            # Cache the most recent frame's feature (for repeated interactions with
+            # a frame; we can use an LRU cache for more frames in the future).
+            inference_state["cached_features"] = {frame_idx: (image, backbone_out)}
+
+        # expand the features to have the same dimension as the number of objects
+        expanded_image = image.expand(batch_size, -1, -1, -1)
+        expanded_backbone_out = {
+            "backbone_fpn": backbone_out["backbone_fpn"].copy(),
+            "vision_pos_enc": backbone_out["vision_pos_enc"].copy(),
+        }
+        for i, feat in enumerate(expanded_backbone_out["backbone_fpn"]):
+            expanded_backbone_out["backbone_fpn"][i] = feat.expand(
+                batch_size, -1, -1, -1
+            )
+        for i, pos in enumerate(expanded_backbone_out["vision_pos_enc"]):
+            pos = pos.expand(batch_size, -1, -1, -1)
+            expanded_backbone_out["vision_pos_enc"][i] = pos
+
+        features = self._prepare_backbone_features(expanded_backbone_out)
+        features = (expanded_image,) + features
+        return features
+
+
+    def _run_single_frame_inference(
+        self,
+        inference_state,
+        output_dict,
+        frame_idx,
+        batch_size,
+        is_init_cond_frame,
+        point_inputs,
+        mask_inputs,
+        reverse,
+        run_mem_encoder,
+        prev_sam_mask_logits=None,
+        ## Extension: LLM prompt
+        language_embd=None,
+    ):
+        """Run tracking on a single frame based on current inputs and previous memory."""
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # point and mask should not appear as input simultaneously on the same frame
+        assert point_inputs is None or mask_inputs is None
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=is_init_cond_frame,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            output_dict=output_dict,
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=reverse,
+            run_mem_encoder=run_mem_encoder,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+            language_embd=language_embd,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = current_out["maskmem_features"]
+        if maskmem_features is not None:
+            maskmem_features = maskmem_features.to(torch.bfloat16)
+            maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        pred_masks_gpu = current_out["pred_masks"]
+        # potentially fill holes in the predicted masks
+        if self.fill_hole_area > 0:
+            pred_masks_gpu = fill_holes_in_mask_scores(
+                pred_masks_gpu, self.fill_hole_area
+            )
+        pred_masks = pred_masks_gpu.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = _get_maskmem_pos_enc(inference_state, current_out)
+        # object pointer is a small tensor, so we always keep it on GPU memory for fast access
+        obj_ptr = current_out["obj_ptr"]
+        # make a compact version of this frame's output to reduce the state size
+        compact_current_out = {
+            "maskmem_features": maskmem_features,
+            "maskmem_pos_enc": maskmem_pos_enc,
+            "pred_masks": pred_masks,
+            "obj_ptr": obj_ptr,
+        }
+        return compact_current_out, pred_masks_gpu
+
+
+    def _consolidate_temp_output_across_obj(
+        self,
+        inference_state,
+        frame_idx,
+        is_cond,
+        run_mem_encoder,
+        consolidate_at_video_res=False,
+    ):
+        """
+        Consolidate the per-object temporary outputs in `temp_output_dict_per_obj` on
+        a frame into a single output for all objects, including
+        1) fill any missing objects either from `output_dict_per_obj` (if they exist in
+           `output_dict_per_obj` for this frame) or leave them as placeholder values
+           (if they don't exist in `output_dict_per_obj` for this frame);
+        2) if specified, rerun memory encoder after apply non-overlapping constraints
+           on the object scores.
+        """
+        batch_size = _get_obj_num(inference_state)
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+        # Optionally, we allow consolidating the temporary outputs at the original
+        # video resolution (to provide a better editing experience for mask prompts).
+        if consolidate_at_video_res:
+            assert not run_mem_encoder, "memory encoder cannot run at video resolution"
+            consolidated_H = inference_state["video_height"]
+            consolidated_W = inference_state["video_width"]
+            consolidated_mask_key = "pred_masks_video_res"
+        else:
+            consolidated_H = consolidated_W = self.image_size // 4
+            consolidated_mask_key = "pred_masks"
+
+        # Initialize `consolidated_out`. Its "maskmem_features" and "maskmem_pos_enc"
+        # will be added when rerunning the memory encoder after applying non-overlapping
+        # constraints to object scores. Its "pred_masks" are prefilled with a large
+        # negative value (NO_OBJ_SCORE) to represent missing objects.
+        consolidated_out = {
+            "maskmem_features": None,
+            "maskmem_pos_enc": None,
+            consolidated_mask_key: torch.full(
+                size=(batch_size, 1, consolidated_H, consolidated_W),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["storage_device"],
+            ),
+            "obj_ptr": torch.full(
+                size=(batch_size, self.hidden_dim),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+        }
+        empty_mask_ptr = None
+        for obj_idx in range(batch_size):
+            obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+            out = obj_temp_output_dict[storage_key].get(frame_idx, None)
+            # If the object doesn't appear in "temp_output_dict_per_obj" on this frame,
+            # we fall back and look up its previous output in "output_dict_per_obj".
+            # We look up both "cond_frame_outputs" and "non_cond_frame_outputs" in
+            # "output_dict_per_obj" to find a previous output for this object.
+            if out is None:
+                out = obj_output_dict["cond_frame_outputs"].get(frame_idx, None)
+            if out is None:
+                out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx, None)
+            # If the object doesn't appear in "output_dict_per_obj" either, we skip it
+            # and leave its mask scores to the default scores (i.e. the NO_OBJ_SCORE
+            # placeholder above) and set its object pointer to be a dummy pointer.
+            if out is None:
+                # Fill in dummy object pointers for those objects without any inputs or
+                # tracking outcomes on this frame (only do it under `run_mem_encoder=True`,
+                # i.e. when we need to build the memory for tracking).
+                if run_mem_encoder:
+                    if empty_mask_ptr is None:
+                        empty_mask_ptr = self._get_empty_mask_ptr(
+                            inference_state, frame_idx
+                        )
+                    # fill object pointer with a dummy pointer (based on an empty mask)
+                    consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = empty_mask_ptr
+                continue
+            # Add the temporary object output mask to consolidated output mask
+            obj_mask = out["pred_masks"]
+            consolidated_pred_masks = consolidated_out[consolidated_mask_key]
+            if obj_mask.shape[-2:] == consolidated_pred_masks.shape[-2:]:
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = obj_mask
+            else:
+                # Resize first if temporary object mask has a different resolution
+                resized_obj_mask = torch.nn.functional.interpolate(
+                    obj_mask,
+                    size=consolidated_pred_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                )
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = resized_obj_mask
+            consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = out["obj_ptr"]
+
+        # Optionally, apply non-overlapping constraints on the consolidated scores
+        # and rerun the memory encoder
+        if run_mem_encoder:
+            device = inference_state["device"]
+            high_res_masks = torch.nn.functional.interpolate(
+                consolidated_out["pred_masks"].to(device, non_blocking=True),
+                size=(self.image_size, self.image_size),
+                mode="bilinear",
+                align_corners=False,
+            )
+            if self.non_overlap_masks_for_mem_enc:
+                high_res_masks = self._apply_non_overlapping_constraints(high_res_masks)
+            maskmem_features, maskmem_pos_enc = self._run_memory_encoder(
+                inference_state=inference_state,
+                frame_idx=frame_idx,
+                batch_size=batch_size,
+                high_res_masks=high_res_masks,
+                is_mask_from_pts=True,  # these frames are what the user interacted with
+            )
+            consolidated_out["maskmem_features"] = maskmem_features
+            consolidated_out["maskmem_pos_enc"] = maskmem_pos_enc
+
+        return consolidated_out
+
+
+    def _get_orig_video_res_output(self, inference_state, any_res_masks):
+        """
+        Resize the object scores to the original video resolution (video_res_masks)
+        and apply non-overlapping constraints for final output.
+        """
+        device = inference_state["device"]
+        video_H = inference_state["video_height"]
+        video_W = inference_state["video_width"]
+        any_res_masks = any_res_masks.to(device, non_blocking=True)
+        if any_res_masks.shape[-2:] == (video_H, video_W):
+            video_res_masks = any_res_masks
+        else:
+            video_res_masks = torch.nn.functional.interpolate(
+                any_res_masks,
+                size=(video_H, video_W),
+                mode="bilinear",
+                align_corners=False,
+            )
+        if self.non_overlap_masks:
+            video_res_masks = self._apply_non_overlapping_constraints(video_res_masks)
+        return any_res_masks, video_res_masks
+
+    def init_state(
+        self,
+        images
+    ):
+        """Initialize a inference state."""
+        inference_state = {}
+        inference_state["images"] = images
+        inference_state["num_frames"] = len(images)
+        # whether to offload the video frames to CPU memory
+        # turning on this option saves the GPU memory with only a very small overhead
+        inference_state["offload_video_to_cpu"] = False
+        # whether to offload the inference state to CPU memory
+        # turning on this option saves the GPU memory at the cost of a lower tracking fps
+        # (e.g. in a test case of 768x768 model, fps dropped from 27 to 24 when tracking one object
+        # and from 24 to 21 when tracking two objects)
+        inference_state["offload_state_to_cpu"] = False
+        # the original video height and width, used for resizing final output scores
+        inference_state["video_height"] = self.image_size
+        inference_state["video_width"] = self.image_size
+        inference_state["device"] = torch.device("cuda")
+        inference_state["storage_device"] = torch.device("cuda")
+        # inputs on each frame
+        inference_state["point_inputs_per_obj"] = {}
+        inference_state["mask_inputs_per_obj"] = {}
+        # visual features on a small number of recently visited frames for quick interactions
+        inference_state["cached_features"] = {}
+        # values that don't change across frames (so we only need to hold one copy of them)
+        inference_state["constants"] = {}
+        # mapping between client-side object id and model-side object index
+        inference_state["obj_id_to_idx"] = OrderedDict()
+        inference_state["obj_idx_to_id"] = OrderedDict()
+        inference_state["obj_ids"] = []
+        # A storage to hold the model's tracking results and states on each frame
+        inference_state["output_dict"] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        # Slice (view) of each object tracking results, sharing the same memory with "output_dict"
+        inference_state["output_dict_per_obj"] = {}
+        # A temporary storage to hold new outputs when user interact with a frame
+        # to add clicks or mask (it's merged into "output_dict" before propagation starts)
+        inference_state["temp_output_dict_per_obj"] = {}
+        # Frames that already holds consolidated outputs from click or mask inputs
+        # (we directly use their consolidated outputs during tracking)
+        inference_state["consolidated_frame_inds"] = {
+            "cond_frame_outputs": set(),  # set containing frame indices
+            "non_cond_frame_outputs": set(),  # set containing frame indices
+        }
+        # metadata for each tracking frame (e.g. which direction it's tracked)
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"] = {}
+        return inference_state
+
+    def add_language_embd(
+            self,
+            inference_state,
+            frame_idx,
+            obj_id,
+            language_embd,
+            inference=False,
+    ):
+        obj_idx = _obj_id_to_idx(inference_state, obj_id)
+
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        # Get any previously predicted mask logits on this object and feed it along with
+        # the new clicks into the SAM mask decoder.
+        prev_sam_mask_logits = None
+        # lookup temporary output dict first, which contains the most recent output
+        # (if not found, then lookup conditioning and non-conditioning frame output)
+        prev_out = obj_temp_output_dict[storage_key].get(frame_idx)
+        if prev_out is None:
+            prev_out = obj_output_dict["cond_frame_outputs"].get(frame_idx)
+            if prev_out is None:
+                prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)
+
+        if prev_out is not None and prev_out["pred_masks"] is not None:
+            prev_sam_mask_logits = prev_out["pred_masks"].cuda(non_blocking=True)
+            # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
+            prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)
+
+        current_out, pred_mask_gpu = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=None,
+            mask_inputs=None,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+            ## Extension: LLM prompt
+            language_embd=language_embd,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        if inference:
+            _consolidated_out = self._consolidate_temp_output_across_obj(
+                inference_state,
+                frame_idx,
+                is_cond=is_cond,
+                run_mem_encoder=False,
+                consolidate_at_video_res=False,
+            )
+            # _, video_res_masks = self._get_orig_video_res_output(
+            #     inference_state, consolidated_out["pred_masks_video_res"]
+            # )
+        return frame_idx, obj_ids, pred_mask_gpu
+
+
+    def _clear_non_cond_mem_around_input(self, inference_state, frame_idx):
+        """
+        Remove the non-conditioning memory around the input frame. When users provide
+        correction clicks, the surrounding frames' non-conditioning memories can still
+        contain outdated object appearance information and could confuse the model.
+
+        This method clears those non-conditioning memories surrounding the interacted
+        frame to avoid giving the model both old and new information about the object.
+        """
+        r = self.memory_temporal_stride_for_eval
+        frame_idx_begin = frame_idx - r * self.num_maskmem
+        frame_idx_end = frame_idx + r * self.num_maskmem
+        output_dict = inference_state["output_dict"]
+        non_cond_frame_outputs = output_dict["non_cond_frame_outputs"]
+        for t in range(frame_idx_begin, frame_idx_end + 1):
+            non_cond_frame_outputs.pop(t, None)
+            for obj_output_dict in inference_state["output_dict_per_obj"].values():
+                obj_output_dict["non_cond_frame_outputs"].pop(t, None)
+
+    def _run_memory_encoder(
+        self, inference_state, frame_idx, batch_size, high_res_masks, is_mask_from_pts
+    ):
+        """
+        Run the memory encoder on `high_res_masks`. This is usually after applying
+        non-overlapping constraints to object scores. Since their scores changed, their
+        memory also need to be computed again with the memory encoder.
+        """
+        # Retrieve correct image features
+        _, _, current_vision_feats, _, feat_sizes = self._get_image_feature(
+            inference_state, frame_idx, batch_size
+        )
+        maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+            current_vision_feats=current_vision_feats,
+            feat_sizes=feat_sizes,
+            pred_masks_high_res=high_res_masks,
+            is_mask_from_pts=is_mask_from_pts,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = maskmem_features.to(torch.bfloat16)
+        maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = _get_maskmem_pos_enc(
+            inference_state, {"maskmem_pos_enc": maskmem_pos_enc}
+        )
+        return maskmem_features, maskmem_pos_enc
+
+    def _add_output_per_object(
+        self, inference_state, frame_idx, current_out, storage_key
+    ):
+        """
+        Split a multi-object output into per-object output slices and add them into
+        `output_dict_per_obj`. The resulting slices share the same tensor storage.
+        """
+        maskmem_features = current_out["maskmem_features"]
+        assert maskmem_features is None or isinstance(maskmem_features, torch.Tensor)
+
+        maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        assert maskmem_pos_enc is None or isinstance(maskmem_pos_enc, list)
+
+        output_dict_per_obj = inference_state["output_dict_per_obj"]
+        for obj_idx, obj_output_dict in output_dict_per_obj.items():
+            obj_slice = slice(obj_idx, obj_idx + 1)
+            obj_out = {
+                "maskmem_features": None,
+                "maskmem_pos_enc": None,
+                "pred_masks": current_out["pred_masks"][obj_slice],
+                "obj_ptr": current_out["obj_ptr"][obj_slice],
+            }
+            if maskmem_features is not None:
+                obj_out["maskmem_features"] = maskmem_features[obj_slice]
+            if maskmem_pos_enc is not None:
+                obj_out["maskmem_pos_enc"] = [x[obj_slice] for x in maskmem_pos_enc]
+            obj_output_dict[storage_key][frame_idx] = obj_out
+
+    @torch.inference_mode()
+    def propagate_in_video_preflight(self, inference_state):
+        """Prepare inference_state and consolidate temporary outputs before tracking."""
+        # Tracking has started and we don't allow adding new objects until session is reset.
+        inference_state["tracking_has_started"] = True
+        batch_size = _get_obj_num(inference_state)
+
+        # Consolidate per-object temporary outputs in "temp_output_dict_per_obj" and
+        # add them into "output_dict".
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        output_dict = inference_state["output_dict"]
+        # "consolidated_frame_inds" contains indices of those frames where consolidated
+        # temporary outputs have been added (either in this call or any previous calls
+        # to `propagate_in_video_preflight`).
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        for is_cond in [False, True]:
+            # Separately consolidate conditioning and non-conditioning temp outptus
+            storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+            # Find all the frames that contain temporary outputs for any objects
+            # (these should be the frames that have just received clicks for mask inputs
+            # via `add_new_points` or `add_new_mask`)
+            temp_frame_inds = set()
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
+            consolidated_frame_inds[storage_key].update(temp_frame_inds)
+            # consolidate the temprary output across all objects on this frame
+            for frame_idx in temp_frame_inds:
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state, frame_idx, is_cond=is_cond, run_mem_encoder=True
+                )
+                # merge them into "output_dict" and also create per-object slices
+                output_dict[storage_key][frame_idx] = consolidated_out
+                self._add_output_per_object(
+                    inference_state, frame_idx, consolidated_out, storage_key
+                )
+                clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+                    self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+                )
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+
+            # clear temporary outputs in `temp_output_dict_per_obj`
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                obj_temp_output_dict[storage_key].clear()
+
+        # edge case: if an output is added to "cond_frame_outputs", we remove any prior
+        # output on the same frame in "non_cond_frame_outputs"
+        for frame_idx in output_dict["cond_frame_outputs"]:
+            output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for obj_output_dict in inference_state["output_dict_per_obj"].values():
+            for frame_idx in obj_output_dict["cond_frame_outputs"]:
+                obj_output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+            assert frame_idx in output_dict["cond_frame_outputs"]
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+
+        # Make sure that the frame indices in "consolidated_frame_inds" are exactly those frames
+        # with either points or mask inputs (which should be true under a correct workflow).
+        all_consolidated_frame_inds = (
+            consolidated_frame_inds["cond_frame_outputs"]
+            | consolidated_frame_inds["non_cond_frame_outputs"]
+        )
+        input_frames_inds = set()
+        for point_inputs_per_frame in inference_state["point_inputs_per_obj"].values():
+            input_frames_inds.update(point_inputs_per_frame.keys())
+        for mask_inputs_per_frame in inference_state["mask_inputs_per_obj"].values():
+            input_frames_inds.update(mask_inputs_per_frame.keys())
+
+        # with language embd as input, there may not be point or box
+        # assert all_consolidated_frame_inds == input_frames_inds
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx=None,
+        max_frame_num_to_track=None,
+        reverse=False,
+    ):
+        """Propagate the input points across frames to track in the entire video."""
+        self.propagate_in_video_preflight(inference_state)
+
+        output_dict = inference_state["output_dict"]
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        obj_ids = inference_state["obj_ids"]
+        num_frames = inference_state["num_frames"]
+        batch_size = _get_obj_num(inference_state)
+        if len(output_dict["cond_frame_outputs"]) == 0:
+            raise RuntimeError("No points are provided; please add points first")
+        clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+            self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+        )
+
+        # set start index, end index, and processing order
+        if start_frame_idx is None:
+            # default: start from the earliest frame with input points
+            start_frame_idx = min(output_dict["cond_frame_outputs"])
+        if max_frame_num_to_track is None:
+            # default: track all the frames in the video
+            max_frame_num_to_track = num_frames
+        if reverse:
+            end_frame_idx = max(start_frame_idx - max_frame_num_to_track, 0)
+            if start_frame_idx > 0:
+                processing_order = range(start_frame_idx, end_frame_idx - 1, -1)
+            else:
+                processing_order = []  # skip reverse tracking if starting from frame 0
+        else:
+            end_frame_idx = min(
+                start_frame_idx + max_frame_num_to_track, num_frames - 1
+            )
+            processing_order = range(start_frame_idx, end_frame_idx + 1)
+
+        for frame_idx in tqdm(processing_order, desc="propagate in video"):
+            # We skip those frames already in consolidated outputs (these are frames
+            # that received input clicks or mask). Note that we cannot directly run
+            # batched forward on them via `_run_single_frame_inference` because the
+            # number of clicks on each object might be different.
+            if frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+                storage_key = "cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+            elif frame_idx in consolidated_frame_inds["non_cond_frame_outputs"]:
+                storage_key = "non_cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+            else:
+                storage_key = "non_cond_frame_outputs"
+                current_out, pred_masks = self._run_single_frame_inference(
+                    inference_state=inference_state,
+                    output_dict=output_dict,
+                    frame_idx=frame_idx,
+                    batch_size=batch_size,
+                    is_init_cond_frame=False,
+                    point_inputs=None,
+                    mask_inputs=None,
+                    reverse=reverse,
+                    run_mem_encoder=True,
+                )
+                output_dict[storage_key][frame_idx] = current_out
+            # Create slices of per-object outputs for subsequent interaction with each
+            # individual object after tracking.
+            self._add_output_per_object(
+                inference_state, frame_idx, current_out, storage_key
+            )
+            inference_state["frames_already_tracked"][frame_idx] = {"reverse": reverse}
+
+            # Resize the output mask to the original video resolution (we directly use
+            # the mask scores on GPU for output to avoid any CPU conversion in between)
+            _, video_res_masks = self._get_orig_video_res_output(
+                inference_state, pred_masks
+            )
+            yield frame_idx, obj_ids, video_res_masks
diff --git a/rynnec/model/processor.py b/rynnec/model/processor.py
new file mode 100644
index 0000000000000000000000000000000000000000..3eea643039b1a6ebba2b4b0a3b7f47530be1ecc5
--- /dev/null
+++ b/rynnec/model/processor.py
@@ -0,0 +1,401 @@
+# coding=utf-8
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""
+Processor class for VideoLLaMA3.
+"""
+from abc import ABCMeta, abstractmethod
+import copy
+import warnings
+from collections import defaultdict
+from typing import List, Union, Dict, Optional, Any
+
+import json
+import torch
+from transformers.feature_extraction_utils import BatchFeature
+from transformers.image_utils import ImageInput, VideoInput
+from transformers.processing_utils import ProcessingKwargs, ProcessorMixin, Unpack
+from transformers.tokenization_utils_base import PreTokenizedInput, TextInput
+
+from rynnec.constants import DEFAULT_IMAGE_TOKEN, IGNORE_INDEX
+from rynnec.mm_utils import load_video, load_images
+from rynnec.model.videollama3_encoder.image_processing_videollama3 import is_valid_image, is_valid_video
+
+
+class Videollama3ProcessorKwargs(ProcessingKwargs, total=False):
+    _defaults = {
+        "text_kwargs": {
+            "padding": False,
+        },
+    }
+
+
+class Videollama3BaseProcessor(ProcessorMixin, metaclass=ABCMeta):
+    r"""
+    Modified from Qwen2VLProcessor
+    Args:
+        image_processor ([`Qwen2VLImageProcessor`], *optional*):
+            The image processor is a required input.
+        tokenizer ([`Qwen2TokenizerFast`], *optional*):
+            The tokenizer is a required input.
+        chat_template (`str`, *optional*): A Jinja template which will be used to convert lists of messages
+            in a chat into a tokenizable string.
+    """
+
+    attributes = ["image_processor", "tokenizer"]
+    valid_kwargs = ["chat_template", "image_merge_size", "video_merge_size", "fps", "max_frames"]
+    image_processor_class = "AutoImageProcessor"
+    tokenizer_class = None
+    chat_template = None
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        image_merge_size: int = 1,
+        video_merge_size: int = 2,
+        fps=1,
+        max_frames=180,
+        **kwargs
+    ):
+        if chat_template is not None:
+            self.chat_template = chat_template
+
+        self.image_processor = image_processor
+        self.tokenizer = tokenizer
+        self.image_merge_size = image_merge_size
+        self.video_merge_size = video_merge_size
+        self.fps = fps
+        self.max_frames = max_frames
+
+        if self.chat_template is not None:
+            self.tokenizer.chat_template = self.chat_template
+
+        self.image_token = DEFAULT_IMAGE_TOKEN
+        self.think_start_token = "<think>"
+        self.think_end_token = "</think>"
+        self.tokenizer.add_tokens([self.image_token], special_tokens=True)
+        self.tokenizer.add_tokens([self.think_start_token, self.think_end_token], special_tokens=False)
+        self.image_token_id = self.tokenizer.convert_tokens_to_ids(self.image_token)
+        self.think_start_token_id = self.tokenizer.convert_tokens_to_ids(self.think_start_token)
+        self.think_end_token_id = self.tokenizer.convert_tokens_to_ids(self.think_end_token)
+        self.newline_token_id = self.tokenizer.encode("\n")[0]
+
+    def load_video(self, *args, **kwargs):
+        return load_video(*args, **kwargs)
+
+    def load_images(self, *args, **kwargs):
+        return load_images(*args, **kwargs)
+    
+    def _get_downsampled_grid_sizes(self, image_inputs: Dict[str, Any]):
+        grid_sizes = []
+        for grid_size, merge_size in zip(image_inputs.get("grid_sizes", []), image_inputs.get("merge_sizes", [])):
+            if not torch.all(grid_size[1:] % merge_size == 0):
+                warnings.warn(f"Grid size {grid_size} is not divisible by merge size. Some undesired errors may occur.")
+            if grid_size[0] == 1:
+                grid_sizes.append(grid_size[1:] / merge_size)
+            elif grid_size[0] > 1:
+                grid_sizes.extend([grid_size[1:] / merge_size] * grid_size[0])
+        return grid_sizes
+
+    def _get_visual_seq_len(self, grid_size: torch.Tensor):
+        num_tokens = int(grid_size.prod().item())
+        return num_tokens
+
+    @abstractmethod
+    def _process_text_with_label(
+        self,
+        text: List[Dict],
+        grid_sizes: torch.Tensor = None,
+        **kwargs,
+    ):
+        return {}
+
+    def _process_text_without_label(
+        self,
+        text: Union[List[str], List[Dict]],
+        grid_sizes: torch.Tensor = None,
+        **kwargs,
+    ):
+        if isinstance(text, (list, tuple)) and isinstance(text[0], dict):
+            warnings.warn("Input text is a list of messages. Automatically convert it to a string with 'apply_chat_template' with generation prompt.")
+            text = self.apply_chat_template(text, tokenize=False, add_generation_prompt=True)
+
+        if len(grid_sizes) > 0:
+            image_idx = 0
+            while self.image_token in text:
+                thw = grid_sizes[image_idx]
+                text = text.replace(self.image_token, "<placeholder>" * thw.prod().long(), 1)
+                image_idx += 1
+            text = text.replace("<placeholder>", self.image_token)
+            assert len(grid_sizes) == image_idx, "Number of images does not match the number of image tokens in the text."
+
+        text_inputs = self.tokenizer(text, **kwargs)
+        return text_inputs
+
+    def process_text(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput], List[Dict]],
+        image_inputs: Dict[str, torch.Tensor] = {},
+        return_labels: bool = False,
+        **kwargs,
+    ):
+        kwargs.pop("padding", None)
+        kwargs.pop("padding_side", None)
+
+        grid_sizes = []
+        for grid_size, merge_size in zip(image_inputs.get("grid_sizes", []), image_inputs.get("merge_sizes", [])):
+            if not torch.all(grid_size[1:] % merge_size == 0):
+                warnings.warn(f"Grid size {grid_size} is not divisible by merge size. Some undesired errors may occur.")
+            if grid_size[0] == 1:
+                grid_sizes.append(grid_size[1:] / merge_size)
+            elif grid_size[0] > 1:
+                grid_sizes.extend([grid_size[1:] / merge_size] * grid_size[0])
+
+        if return_labels:
+            return self._process_text_with_label(text, grid_sizes, **kwargs)
+        return self._process_text_without_label(text, grid_sizes, **kwargs)
+
+    def process_images(
+        self,
+        images: ImageInput = None,
+        merge_size: Optional[int] = 1,
+        **kwargs,
+    ):
+        if images is None:
+            return {}
+        image_inputs = self.image_processor(images=images, merge_size=merge_size, **kwargs)
+        return image_inputs
+
+    def __call__(
+        self,
+        text: Union[TextInput, PreTokenizedInput, List[TextInput], List[PreTokenizedInput], List[Dict]] = None,
+        images: ImageInput = None,
+        merge_size: Optional[int] = 1,
+        return_labels: bool = False,
+        **kwargs: Unpack[Videollama3ProcessorKwargs],
+    ) -> BatchFeature:
+        """
+        Main method to prepare for the model one or several sequences(s) and image(s). This method forwards the `text`
+        and `kwargs` arguments to Qwen2TokenizerFast's [`~Qwen2TokenizerFast.__call__`] if `text` is not `None` to encode
+        the text. To prepare the vision inputs, this method forwards the `vision_infos` and `kwrags` arguments to
+        Qwen2VLImageProcessor's [`~Qwen2VLImageProcessor.__call__`] if `vision_infos` is not `None`.
+
+        Args:
+            images (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Image.Image]`, `List[np.ndarray]`, `List[torch.Tensor]`):
+                The image or batch of images to be prepared. Each image can be a PIL image, NumPy array or PyTorch
+                tensor. Both channels-first and channels-last formats are supported.
+            text (`str`, `List[str]`, `List[List[str]]`):
+                The sequence or batch of sequences to be encoded. Each sequence can be a string or a list of strings
+                (pretokenized string). If the sequences are provided as list of strings (pretokenized), you must set
+                `is_split_into_words=True` (to lift the ambiguity with a batch of sequences).
+            return_tensors (`str` or [`~utils.TensorType`], *optional*):
+                If set, will return tensors of a particular framework. Acceptable values are:
+                - `'tf'`: Return TensorFlow `tf.constant` objects.
+                - `'pt'`: Return PyTorch `torch.Tensor` objects.
+                - `'np'`: Return NumPy `np.ndarray` objects.
+                - `'jax'`: Return JAX `jnp.ndarray` objects.
+
+        Returns:
+            [`BatchFeature`]: A [`BatchFeature`] with the following fields:
+
+            - **input_ids** -- List of token ids to be fed to a model. Returned when `text` is not `None`.
+            - **attention_mask** -- List of indices specifying which tokens should be attended to by the model (when
+              `return_attention_mask=True` or if *"attention_mask"* is in `self.model_input_names` and if `text` is not
+              `None`).
+            - **pixel_values** -- Pixel values to be fed to a model. Returned when `images` is not `None`.
+            - **grid_sizes** -- List of image 3D grid in LLM. Returned when `images` is not `None`.
+        """
+        output_kwargs = self._merge_kwargs(
+            Videollama3ProcessorKwargs,
+            tokenizer_init_kwargs=self.tokenizer.init_kwargs,
+            **kwargs,
+        )
+        output_kwargs["text_kwargs"].pop("padding", None)
+        output_kwargs["text_kwargs"].pop("padding_side", None)
+
+        image_inputs = self.process_images(images, merge_size, **output_kwargs["images_kwargs"])
+        text_inputs = self.process_text(text, image_inputs, return_labels, **output_kwargs["text_kwargs"])
+
+        return BatchFeature(data={**text_inputs, **image_inputs})
+
+    def batch_decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.batch_decode`]. Please
+        refer to the docstring of this method for more information.
+        """
+        return self.tokenizer.batch_decode(*args, **kwargs)
+
+    def decode(self, *args, **kwargs):
+        """
+        This method forwards all its arguments to Qwen2TokenizerFast's [`~PreTrainedTokenizer.decode`]. Please refer to
+        the docstring of this method for more information.
+        """
+        return self.tokenizer.decode(*args, **kwargs)
+
+    def _load_multimodal_data(self, conversation: List[Dict[str, Any]]):
+        multimodal_info = defaultdict(list)
+        new_conversation = []
+        for message in conversation:
+            new_message = {"role": message["role"]}
+            if not isinstance(message["content"], (list, tuple)):
+                new_message["content"] = message["content"]
+                new_conversation.append(new_message)
+                continue
+
+            new_contents = []
+            for content in message["content"]:
+                if not isinstance(content, dict):
+                    new_contents.append(content)
+                    continue
+                assert "type" in content, "Content must have 'type' field."
+                if content["type"] in ["image", "video"] and content["type"] in content and isinstance(content[content["type"]], dict):
+                    # TODO: support other types which are not compatible with json
+                    load_args = content[content["type"]]
+                    data_id = json.dumps({k: v for k, v in load_args.items() if not k in ["start_time", "end_time"]})
+                    new_content = copy.deepcopy(content)
+                    multimodal_info[data_id].append(new_content)
+                    new_contents.append(new_content)
+                else:
+                    new_contents.append(content)
+
+            new_message["content"] = new_contents
+            new_conversation.append(new_message)
+
+        for data_id, contents in multimodal_info.items():
+            data_type = contents[0]["type"]
+            if data_type == "image":
+                image = self.load_images(contents[0][data_type]["image_path"])[0]
+                for content in contents:
+                    content["image"] = image.copy()
+
+            elif data_type == "video":
+                # TODO: start_time is None?
+                start_times = [content["video"].get("start_time", 0.) for content in contents]
+                end_times = [content["video"].get("end_time", float("inf")) for content in contents]
+
+                load_args = contents[0][data_type]
+                start_time, end_time = min(start_times), max(end_times)
+                if start_time > 0:
+                    load_args["start_time"] = start_time
+                if end_time < float("inf"):
+                    load_args["end_time"] = end_time
+                images, timestamps = self.load_video(**load_args)
+
+                for content, start_time, end_time in zip(contents, start_times, end_times):
+                    cur_images, cur_timestamps = [], []
+                    for image, timestamp in zip(images, timestamps):
+                        if start_time <= timestamp <= end_time:
+                            cur_images.append(image.copy())
+                            cur_timestamps.append(timestamp)
+
+                    content[data_type] = cur_images
+                    content["num_frames"] = len(cur_images)
+                    content["timestamps"] = cur_timestamps
+
+        return new_conversation
+
+    def _gather_multimodal_data(self, conversation: List[Dict[str, Any]]):
+        images = []
+        for message in conversation:
+            if not isinstance(message["content"], (list, tuple)):
+                continue
+            for content in message["content"]:
+                if not isinstance(content, dict):
+                    continue
+                if content["type"] == "video":
+                    video = content["video"]
+                    assert is_valid_video(video), f"Invalid video data: {video}."
+                    images.append(video)
+                if content["type"] == "image":
+                    image = content["image"]
+                    assert is_valid_image(image), f"Invalid image data: {image}."
+                    images.append(image)
+        images = images if len(images) > 0 else None
+        return images
+
+    def apply_chat_template(
+        self,
+        conversation: List[Dict[str, Any]],
+        chat_template: Optional[str] = None,
+        tokenize: bool = False,
+        add_system_prompt: bool = False,
+        add_generation_prompt: bool = False,
+        add_think_prompt: bool = False,
+        return_dict: bool = False,
+        **kwargs,
+    ) -> str:
+        """
+        Similar to the `apply_chat_template` method on tokenizers, this method applies a Jinja template to input
+        conversations to turn them into a single tokenizable string.
+        Args:
+            conversation (`List[Dict, str, str]`):
+                The conversation to format.
+            chat_template (`Optional[str]`, *optional*):
+                The Jinja template to use for formatting the conversation. If not provided, the tokenizer's
+                chat template is used.
+            tokenize (`bool`, *optional*, defaults to `False`):
+                Whether to tokenize the output or not.
+            add_system_prompt (`bool`, *optional*, defaults to `False`):
+                Whether to add the system prompt to the output or not.
+            add_generation_prompt (`bool`, *optional*, defaults to `False`):
+                Whether to add the generation prompt to the output or not.
+            image_token (`Optional[str]`, *optional*, defaults to `<image>`):
+                The token to use for indicating images in the conversation.
+            **kwargs:
+                Additional keyword arguments
+        """
+
+        if chat_template is None:
+            if self.chat_template is not None:
+                chat_template = self.chat_template
+            else:
+                raise ValueError(
+                    "No chat template is set for this processor. Please either set the `chat_template` attribute, "
+                    "or provide a chat template as an argument. See "
+                    "https://huggingface.co/docs/transformers/main/en/chat_templating for more information."
+                )
+
+        images = None
+        if return_dict:
+            conversation = self._load_multimodal_data(conversation)
+            images = self._gather_multimodal_data(conversation)
+
+        prompt = self.tokenizer.apply_chat_template(
+            conversation,
+            chat_template=chat_template,
+            tokenize=tokenize,
+            add_system_prompt=add_system_prompt,
+            add_generation_prompt=add_generation_prompt,
+            add_think_prompt=add_think_prompt,
+            image_token=self.image_token,
+            **kwargs
+        )
+
+        out = {"text": prompt, "images": images}
+        if return_dict:
+            return out
+        return out["text"]
+
+    @property
+    def model_input_names(self):
+        tokenizer_input_names = self.tokenizer.model_input_names
+        image_processor_input_names = self.image_processor.model_input_names
+        return list(dict.fromkeys(tokenizer_input_names + image_processor_input_names))
diff --git a/rynnec/model/projector.py b/rynnec/model/projector.py
new file mode 100644
index 0000000000000000000000000000000000000000..df0319c83f349d9c8426e87b6e8783dcafc1539b
--- /dev/null
+++ b/rynnec/model/projector.py
@@ -0,0 +1,161 @@
+#    Copyright 2024 Alibaba DAMO Academy
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+import math
+import os
+import re
+
+import einops
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from timm.models.layers import LayerNorm, LayerNorm2d
+from timm.models.regnet import RegStage
+from transformers import TRANSFORMERS_CACHE
+
+
+def parse_snapshot_folder(repo_id, cache_dir=None, repo_type="model"):
+    revision = "main"
+    # 1. parse the downloaded cache folder
+    if cache_dir is None:
+        cache_dir = TRANSFORMERS_CACHE
+    else:
+        cache_dir = cache_dir
+    object_id = repo_id.replace("/", "--")
+    repo_cache = os.path.join(cache_dir, f"{repo_type}s--{object_id}")
+    # 2. resolve refs (for instance to convert main to the associated commit sha)
+    refs_dir = os.path.join(repo_cache, "refs")
+    if os.path.isdir(refs_dir):
+        revision_file = os.path.join(refs_dir, revision)
+        if os.path.isfile(revision_file):
+            with open(revision_file) as f:
+                revision = f.read()
+    # 3. acquire the snapshot folder
+    folder = os.path.join(repo_cache, "snapshots", revision)
+
+    return folder
+
+
+def load_mm_projector(model_path, cache_dir=None, token=None):
+    if os.path.exists(os.path.join(model_path, 'mm_projector.bin')):
+        is_local = True
+        folder = model_path
+    else:
+        is_local = False
+        folder = parse_snapshot_folder(model_path, cache_dir=cache_dir, repo_type="model")
+        if not os.path.exists(os.path.join(folder, 'mm_projector.bin')):
+            # downloading from remote repo
+            from huggingface_hub import snapshot_download
+            snapshot_download(repo_id=model_path, cache_dir=cache_dir, token=token)
+
+    mm_projector_weights = torch.load(os.path.join(folder, 'mm_projector.bin'), map_location='cpu')
+    mm_projector_weights = {k: v.to(torch.float16) for k, v in mm_projector_weights.items()}
+    return mm_projector_weights
+
+
+class IdentityMap(nn.Module):
+
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, *args, **kwargs):
+        return x
+
+    @property
+    def config(self):
+        return {"mm_projector_type": 'identity'}
+
+
+def build_mlp(depth, hidden_size, output_hidden_size):
+    modules = [nn.Linear(hidden_size, output_hidden_size)]
+    for _ in range(1, depth):
+        modules.append(nn.GELU())
+        modules.append(nn.Linear(output_hidden_size, output_hidden_size))
+    return nn.Sequential(*modules)
+
+
+class SimSpatialConv(nn.Module):
+
+    def __init__(self, mm_hidden_size, hidden_size, downsample=(2, 2), padding=1, depth=1, mlp_depth=2):
+        super().__init__()
+        self.encoder_hidden_size = encoder_hidden_size = mm_hidden_size
+        self.output_hidden_size = output_hidden_size = hidden_size
+        self.downsample = downsample
+        self.padding = padding
+        self.sampler = nn.Sequential(
+            nn.Conv2d(
+                in_channels=self.encoder_hidden_size,
+                out_channels=4 * self.encoder_hidden_size,
+                kernel_size=self.downsample,
+                stride=self.downsample,
+                padding=self.padding,
+                bias=True
+            ),
+            nn.SiLU(),
+        )
+        self.readout = build_mlp(mlp_depth, 4 * self.encoder_hidden_size, self.output_hidden_size)
+
+    def forward(self, x):
+        hw = int(x.size(1) ** 0.5)
+        x = einops.rearrange(x, "b (h w) d -> b d h w", h=hw, w=hw)
+        x = self.sampler(x)
+        x = einops.rearrange(x, "b d h w -> b (h w) d")
+        x = self.readout(x)
+        return x
+
+    def cal_proj_size(self, input_size):
+        if isinstance(input_size, int):
+            input_size = (input_size, input_size)
+        height = math.ceil((input_size[0] + self.padding) / self.downsample[0])
+        width  = math.ceil((input_size[1] + self.padding) / self.downsample[1])
+        return height * width
+
+
+class MlpGeluProjector(nn.Module):
+    def __init__(self, mm_hidden_size, hidden_size, projector_type):
+        super().__init__()
+
+        mlp_gelu_match = re.match(r"^mlp(\d+)x_gelu$", projector_type)
+        mlp_depth = int(mlp_gelu_match.group(1))
+
+        self.readout = build_mlp(mlp_depth, mm_hidden_size, hidden_size)
+
+    def forward(self, x):
+        x = self.readout(x)
+        return x
+
+    def cal_proj_size(self, input_size):
+        if isinstance(input_size, int):
+            input_size = (input_size, input_size)
+        height = input_size[0]
+        width  = input_size[1]
+        return height * width
+
+
+def build_vision_projector(config, mm_hidden_size, delay_load=False, **kwargs):
+    # rynnec projector only support image-wise operation now, i.e., prohibit the temporal aggregation
+    projector_type = getattr(config, 'mm_projector_type', 'linear')
+    hidden_size = config.hidden_size
+
+    if projector_type == "linear":
+        # NOTE: for both linear and mlp2x_gelu projector type, mean pooling is adopted to aggreate video features
+        return nn.Linear(mm_hidden_size, hidden_size)
+    elif  projector_type == "simp_spatial_conv":
+        return SimSpatialConv(mm_hidden_size, hidden_size)
+    elif projector_type.startswith("mlp"):
+        return MlpGeluProjector(mm_hidden_size, hidden_size, projector_type)
+    if projector_type == 'identity':
+        return IdentityMap()
+
+    raise ValueError(f'Unknown projector type: {projector_type}')
diff --git a/rynnec/model/region_encoder.py b/rynnec/model/region_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..8edb6f15c5aaf76750c832121fc43232f16dbd0e
--- /dev/null
+++ b/rynnec/model/region_encoder.py
@@ -0,0 +1,77 @@
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from functools import partial
+
+class MaskExtractor(nn.Module):
+    def __init__(self, config, mm_hidden_size, depth=2):
+        super(MaskExtractor, self).__init__()
+        self.mask_pooling = MaskPooling()
+        modules = [nn.Linear(mm_hidden_size, config.hidden_size)]
+        for _ in range(1, depth):
+            modules.append(nn.GELU())
+            modules.append(nn.Linear(config.hidden_size, config.hidden_size))
+        self.feat_linear =  nn.Sequential(*modules)
+
+    def forward(self, feats, masks):
+        query_feats = []
+        
+        if masks is None: #infer
+            return None
+
+        num_imgs = len(masks)
+        region_token_nums = []
+        image_idx = 0
+        for idx in range(num_imgs):
+            if masks[idx]==None:
+                continue
+            for mask_idx in range(len(masks[idx])):
+                mask = masks[idx][mask_idx].unsqueeze(0).unsqueeze(0).float()
+                if len(mask[0])==0:
+                    mask = torch.zeros((1, 1, 336, 336)).to(feats.device).float()
+
+                feat = feats[image_idx].unsqueeze(0)
+                image_idx+=1
+                
+                # h, w = feat.shape[1:3]
+                feat = feat.permute(0,3,1,2)
+
+                raw_dtype = feat.dtype
+                feat = feat.to(mask.dtype)
+                
+                mask_feat_raw = self.mask_pooling(feat, mask) # [n, 1024]
+
+                query_feats.append(mask_feat_raw)
+        if len(query_feats)==0:
+            return None
+        mask_feats = torch.cat(query_feats, dim=0)
+        mask_feats = mask_feats.to(feats[0].dtype)
+        mask_feats_linear = self.feat_linear(mask_feats)
+        return mask_feats_linear
+
+    
+class MaskPooling(nn.Module):
+    def __init__(self):
+        super().__init__()
+
+    def forward(self, x, mask):
+
+        if not x.shape[-2:] == mask.shape[-2:]:
+            # reshape mask to x
+            mask = F.interpolate(mask, size=x.shape[-2:], mode='bilinear', align_corners=False)
+
+        # b, c, h ,w = x.shape
+        # b, q, h, w = mask.shape
+        mask = (mask > 0).to(mask.dtype)
+        mask = mask.permute(1,0,2,3)
+        denorm = mask.sum(dim=(-1, -2), keepdim=True) + 1e-8
+
+        mask_pooled_x = (x * mask/denorm).sum(-1).sum(-1)
+
+        return mask_pooled_x
+
+
+def build_region_encoder(config, mm_hidden_size):
+
+    return MaskExtractor(config, mm_hidden_size)
+   
\ No newline at end of file
diff --git a/rynnec/model/rynnec_arch.py b/rynnec/model/rynnec_arch.py
new file mode 100644
index 0000000000000000000000000000000000000000..a0c7e50f8d2f6813ce62ceacbafe3c44e6ed4ae8
--- /dev/null
+++ b/rynnec/model/rynnec_arch.py
@@ -0,0 +1,271 @@
+# Adopted from: https://github.com/DAMO-NLP-SG/VideoLLaMA3. 
+# Adopted from https://github.com/haotian-liu/LLaVA. Below is the original copyright:
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+import os
+import math
+from abc import ABC, abstractmethod
+from typing import List, Optional, Tuple, Union
+
+import einops
+import torch
+import torch.distributed as dist
+import torch.nn as nn
+import numpy as np
+
+from ..constants import IGNORE_INDEX, MODAL_INDEX_MAP, NUM_FRAMES
+from .encoder import build_vision_encoder
+from .projector import build_vision_projector, load_mm_projector
+from .region_encoder import build_region_encoder
+from ..mm_utils import reshape_images_to_raw_grid
+
+
+def spatial_downsampling(features, grid_thws, stride=2):
+    n, c = features.shape
+
+    flatten_grid_thws = torch.cat([grid_thw for batch_grid_thws in grid_thws for grid_thw in batch_grid_thws])
+    split_sizes = [grid_thw.prod() for grid_thw in flatten_grid_thws]
+    features = torch.split(features, split_sizes)
+
+    new_features = []
+    for feature, grid_thw in zip(features, flatten_grid_thws):
+        # NOTE: adapted for reshape in image processor 
+        feature = feature.view(grid_thw[0], grid_thw[1] // stride, grid_thw[2] // stride, stride, stride,  c).permute(0, 1, 3, 2, 4, 5)
+        feature = feature.reshape(grid_thw[0], grid_thw[1], grid_thw[2], c).permute(0, 3, 1, 2)
+        # NOTE: previous version model is align_corners=True
+        new_feature = torch.nn.functional.interpolate(feature, (math.ceil(grid_thw[1] / stride), math.ceil(grid_thw[2] / stride)), mode='bilinear')
+        # new_feature = nn.functional.avg_pool2d(feature, stride)
+        # new_feature = nn.functional.max_pool2d(feature, stride)
+        new_features.append(new_feature.permute(0, 2, 3, 1).view(-1, c))
+    new_features = torch.cat(new_features)
+
+    return new_features
+
+
+class RynnecMetaModel:
+
+    def __init__(self, config):
+        super(RynnecMetaModel, self).__init__(config)
+
+        if hasattr(config, "vision_encoder") or hasattr(config, "mm_vision_encoder"):
+            self.vision_encoder = build_vision_encoder(config, delay_load=False)
+            self.mm_projector = build_vision_projector(config, config.mm_hidden_size)
+        self.region_encoder = build_region_encoder(config, config.mm_hidden_size) 
+
+    def get_vision_encoder(self):
+        vision_encoder = getattr(self, 'vision_encoder', None)
+        if type(vision_encoder) is list:
+            vision_encoder = vision_encoder[0]
+        return vision_encoder
+
+    def get_mm_projector(self):
+        return self.mm_projector
+
+    def initialize_vision_modules(self, model_args, fsdp=None):
+        vision_encoder = model_args.vision_encoder
+        mm_vision_select_layer = model_args.mm_vision_select_layer
+        mm_vision_select_feature = model_args.mm_vision_select_feature
+        pretrain_mm_projector = model_args.pretrain_mm_projector
+
+        self.config.mm_vision_encoder = vision_encoder
+
+        if self.get_vision_encoder() is None:
+            vision_encoder = build_vision_encoder(model_args)
+
+            if fsdp is not None and len(fsdp) > 0:
+                self.vision_encoder = [vision_encoder]
+            else:
+                self.vision_encoder = vision_encoder
+        else:
+            if fsdp is not None and len(fsdp) > 0:
+                vision_encoder = self.vision_encoder[0]
+            else:
+                vision_encoder = self.vision_encoder
+            # NOTE: only compatible with delay_load encoder
+            # vision_encoder.load_model(vision_encoder.cfg_only)
+
+        self.config.use_mm_proj = True
+        self.config.mm_projector_type = getattr(model_args, 'mm_projector_type', 'linear')
+        self.config.mm_hidden_size = vision_encoder.hidden_size
+        self.config.mm_vision_select_layer = mm_vision_select_layer
+        self.config.mm_vision_select_feature = mm_vision_select_feature
+
+        if getattr(self, 'mm_projector', None) is None:
+            self.mm_projector = build_vision_projector(self.config)
+        else:
+            # In case it is frozen by LoRA
+            for p in self.mm_projector.parameters():
+                p.requires_grad = True
+
+        if pretrain_mm_projector is not None:
+            if os.path.exists(pretrain_mm_projector):
+                is_local = True
+                if os.path.isdir(pretrain_mm_projector):
+                    mm_projector_weights = load_mm_projector(pretrain_mm_projector)
+                else:
+                    mm_projector_weights = torch.load(pretrain_mm_projector, map_location='cpu')
+            else:
+                # Support loading projector weights from remote HuggingFace model hub
+                is_local = False
+                pretrain_mm_projector = pretrain_mm_projector.replace('mm_projector.bin', '')
+                pretrain_mm_projector = pretrain_mm_projector.strip('/').strip('\\').strip()
+                mm_projector_weights = load_mm_projector(pretrain_mm_projector)
+
+            def get_w(weights, keyword):
+                return {k.split(keyword + '.')[1]: v for k, v in weights.items() if keyword in k}
+
+            self.mm_projector.load_state_dict(get_w(mm_projector_weights, 'mm_projector'), strict=False)
+
+
+class RynnecMetaForCausalLM(ABC):
+
+    @abstractmethod
+    def get_model(self):
+        pass
+
+    def num_frames(self):
+        if hasattr(self.config, 'num_frames'):
+            return self.config.num_frames
+        else:
+            return NUM_FRAMES
+
+    def spatial_merge_size(self):
+        if hasattr(self.config, 'spatial_merge_size'):
+            return self.config.spatial_merge_size
+        else:
+            return 1
+    
+    def get_vision_encoder(self):
+        return self.get_model().get_vision_encoder()
+
+    def get_mm_projector(self):
+        return self.get_model().get_mm_projector()
+
+    def encode_images(
+        self,
+        pixel_values: torch.FloatTensor,
+        grid_sizes: torch.LongTensor,
+        merge_sizes: torch.LongTensor,
+    ):
+        mm_features, mm_features_raw = self.get_model().get_vision_encoder()(
+            pixel_values=pixel_values,
+            grid_sizes=grid_sizes,
+            merge_sizes=merge_sizes,
+        )
+        mm_features = self.get_model().mm_projector(mm_features)
+        return mm_features, mm_features_raw
+
+    def _get_valid_visual_tokens(
+        self,
+        mm_features: torch.FloatTensor,
+        batched_num_patches: torch.LongTensor,
+        modals: List[str],
+    ):
+        valid_masks = []
+        for num_patches, modal in zip(batched_num_patches, modals):
+            valid_mask = torch.full((num_patches, ), modal != "text", dtype=torch.bool, device=mm_features.device)
+            valid_masks.append(valid_mask)
+        mm_features = mm_features[torch.cat(valid_masks)]
+        return mm_features
+
+    def prepare_inputs_labels_for_multimodal(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        labels: Optional[torch.LongTensor] = None,
+        pixel_values: Optional[torch.FloatTensor] = None,
+        grid_sizes: Optional[torch.LongTensor] = None,
+        merge_sizes: Optional[torch.LongTensor] = None,
+        modals: Optional[List[str]] = None,
+        masks=None, 
+        mask_ids = None
+    ):
+        vision_encoder = self.get_vision_encoder()
+        # NOTE: text-only situation
+        if vision_encoder is None or pixel_values is None or input_ids.shape[1] == 1:
+            return input_ids, attention_mask, position_ids, past_key_values, None, labels
+
+        # 1. flatten text inputs
+        B, N = input_ids.shape
+        input_ids = input_ids.view(B * N)
+        if attention_mask is not None:
+            attention_mask = attention_mask.view(B * N)
+        if position_ids is not None:
+            position_ids = position_ids.view(B * N)
+        if labels is not None:
+            labels = labels.view(B * N)
+
+        # 2. embed visual tokens
+        batched_num_patches = grid_sizes.prod(dim=1).div(merge_sizes ** 2).long()
+
+        mm_features, mm_features_raw = self.encode_images(pixel_values, grid_sizes, merge_sizes)
+        mm_features = mm_features.to(input_ids.device)
+        mm_features_raw = mm_features_raw.to(input_ids.device)
+        mm_features = self._get_valid_visual_tokens(mm_features, batched_num_patches, modals)
+
+        # 3. embed text tokens
+        image_selected = (input_ids == self.config.image_token_index)
+        # input_ids[image_selected] = 0
+        inputs_embeds = self.get_model().embed_tokens(input_ids).clone()
+
+        num_vision_tokens = image_selected.sum()
+        if mm_features.size(0) > num_vision_tokens:
+            print(f"Number of mm_features ({mm_features.size(0)}) exceeds the number of image tokens ({num_vision_tokens}). Automative truncated.")
+            mm_features = mm_features[:num_vision_tokens]
+
+        # 4. replace multimodal tokens with features
+        inputs_embeds[image_selected] = inputs_embeds[image_selected] * 0.0 + mm_features
+        
+        # 5. embed region tokens
+        try:
+
+            mask_selected = (input_ids == self.config.region_token_index)
+
+            if mask_selected.sum() > 0:
+                reshaped_features = reshape_images_to_raw_grid(mm_features_raw, grid_sizes)
+                mask_additional_image_features = []
+                idx = 0
+                new_masks = []
+                for bs in range(len(masks)):
+                    flag=True
+                    for ml in range(len(masks[bs])):
+                        if mask_ids[idx]>=0:
+                            mask_additional_image_features.append(reshaped_features[mask_ids[idx]])
+                        else:
+                            flag=False
+                        idx+=1
+                    if flag:
+                        new_masks.append(masks[bs])
+
+                mask_feats = self.get_model().region_encoder(mask_additional_image_features, new_masks)
+                inputs_embeds[mask_selected] = inputs_embeds[mask_selected]*0.0 + mask_feats
+            
+        except Exception as e:
+            print(e)
+
+
+        # 6. reshape back to batched format
+        C = inputs_embeds.shape[-1]
+        inputs_embeds = inputs_embeds.reshape(B, -1, C)
+        if attention_mask is not None:
+            attention_mask = attention_mask.view(B, -1)
+        if labels is not None:
+            labels = labels.view(B, -1)
+        if position_ids is not None:
+            position_ids = position_ids.view(B, -1)
+
+        return None, attention_mask, position_ids, past_key_values, inputs_embeds, labels
diff --git a/rynnec/model/rynnec_qwen2.py b/rynnec/model/rynnec_qwen2.py
new file mode 100644
index 0000000000000000000000000000000000000000..c189e56ceb4bdce46a0f6eb749c7e81c91d7593b
--- /dev/null
+++ b/rynnec/model/rynnec_qwen2.py
@@ -0,0 +1,638 @@
+# Adopted from: https://github.com/DAMO-NLP-SG/VideoLLaMA3. 
+# Adopted from: https://github.com/haotian-liu/LLaVA. 
+# Below is the original copyright:
+#    Copyright 2023 Haotian Liu
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+
+from typing import List, Optional, Tuple, Union, Dict
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from transformers import (AutoConfig, AutoModelForCausalLM, AutoProcessor, AutoImageProcessor,
+                          Qwen2Config, Qwen2ForCausalLM, Qwen2Model)
+from transformers.generation.utils import GenerateOutput
+# from transformers.modeling_outputs import CausalLMOutputWithPast
+from dataclasses import dataclass
+from transformers.utils import ModelOutput
+
+from .loss import cross_entropy_loss, CrossEntropyLoss, DiceLoss
+from .processor import Videollama3BaseProcessor
+from .rynnec_arch import RynnecMetaForCausalLM, RynnecMetaModel
+from .videollama3_encoder import Videollama3ImageProcessor
+from rynnec.constants import IGNORE_INDEX, DEFAULT_IMAGE_TOKEN
+from .sam2_train import SAM2TrainRunner
+from .sam2 import SAM2
+from .utils import genetate_video_pred_embeddings, process_video_gt_masks
+
+CHAT_TEMPLATE = """
+{%- set identifier = 'im' %}
+{% for message in messages %}
+    {% if message['role'] == 'stream' %}
+        {% set identifier = 'stream' %}
+    {% else %}
+        {% set identifier = 'im' %}
+    {% endif %}
+    {% if message['role'] is not none %}
+        {{- '<|' + identifier + '_start|>' + message['role'] + '\n' -}}
+    {% endif %}
+    {% if message['content'] is string %}
+        {{- message['content'] + '<|' + identifier + '_end|>\n' -}}
+    {% else %}
+        {% for content in message['content'] %}
+            {% if content['type'] == 'image' or 'image' in content or 'image_url' in content %}
+                {% if 'time' in content %}
+                    {{- 'Time ' + content['time'] | round(1) | string + 's: ' -}}
+                {% endif %}
+                {{- image_token + '\n' -}}
+            {% elif content['type'] == 'video' or 'video' in content or 'video_url' in content %}
+                {% for i in range(content['num_frames']) %}
+                    {% if 'timestamps' in content %}
+                        {{- 'Time ' + content['timestamps'][i] | round(1) | string + 's:' -}}
+                    {% endif %}
+                    {% if i < content['num_frames'] - 1 %}
+                        {{- image_token + ',' -}}
+                    {% else %}
+                        {{- image_token + '\n' -}}
+                    {% endif %}
+                {% endfor %}
+            {% elif content['type'] == 'text' or 'text' in content %}
+                {{- content['text'] -}}
+            {% endif %}
+        {% endfor %}
+        {% if message['role'] is not none %}
+            {{- '<|' + identifier + '_end|>\n' -}}
+        {% endif %}
+    {% endif %}
+{% endfor %}
+{% if add_generation_prompt %}
+    {{- '<|im_start|>assistant\n' -}}
+    {% if add_think_prompt %}
+        {{- '<think>\n' -}}
+    {% endif %}
+{% endif %}
+"""
+@dataclass
+class CausalLMOutputWithPast(ModelOutput):
+    loss: Optional[torch.FloatTensor] = None
+    logits: torch.FloatTensor = None
+    past_key_values: Optional[List[torch.FloatTensor]] = None
+    hidden_states: Optional[Tuple[torch.FloatTensor]] = None
+    attentions: Optional[Tuple[torch.FloatTensor]] = None
+    rope_deltas: Optional[torch.LongTensor] = None
+    ce_loss: Optional[torch.FloatTensor] = None
+    mask_bce_loss: Optional[torch.FloatTensor] = None
+    mask_dice_loss: Optional[torch.FloatTensor] = None
+    mask_loss: Optional[torch.FloatTensor] = None
+
+
+class Videollama3Qwen2Processor(Videollama3BaseProcessor):
+
+    tokenizer_class = ("Qwen2Tokenizer", "Qwen2TokenizerFast")
+    chat_template = CHAT_TEMPLATE
+
+    def __init__(
+        self,
+        image_processor=None,
+        tokenizer=None,
+        chat_template=None,
+        image_merge_size: int = 1,
+        video_merge_size: int = 2,
+        fps=1,
+        max_frames=180,
+        **kwargs
+    ):
+        super().__init__(image_processor, tokenizer, chat_template, **kwargs)
+        self.generation_prompt = self._infer_generation_prompt()
+        self.generation_prompt_ids = self.tokenizer.encode(self.generation_prompt, return_tensors="pt")
+        self.generation_prompt_length = len(self.generation_prompt_ids[0])
+
+    def _infer_generation_prompt(self):
+        pseudo_message = [{"role": "user", "content": ""}]
+        instruction = self.apply_chat_template(pseudo_message, tokenize=False, add_generation_prompt=True)
+        conversation = self.apply_chat_template(pseudo_message, tokenize=False, add_generation_prompt=False)
+        return instruction.replace(conversation, "")
+
+    def _process_text_with_label(
+        self,
+        text: List[Dict],
+        grid_sizes: torch.Tensor = None,
+        **kwargs,
+    ):
+        assert kwargs.pop("return_tensors", "pt") == "pt", "Only PyTorch tensors are supported when return_labels=True."
+        assert isinstance(text[0], dict), "When return_labels=True, text must be a list of messages."
+
+        input_ids_list = []
+        targets_list = []
+        image_idx = 0
+
+        for message_idx, message in enumerate(text):
+            # 1. set chat template and append image tokens
+            prompt = self.apply_chat_template([message], tokenize=False, add_generation_prompt=False)
+            prompt_chunks = prompt.split(DEFAULT_IMAGE_TOKEN)
+            prompt = []
+            for chunk_idx in range(len(prompt_chunks) - 1):
+                prompt.append(prompt_chunks[chunk_idx])
+                thw = grid_sizes[image_idx]
+                prompt.append(DEFAULT_IMAGE_TOKEN * thw.prod().long())
+                image_idx += 1
+            prompt.append(prompt_chunks[-1])
+            prompt = "".join(prompt)
+
+            input_ids = self.tokenizer.encode(prompt, return_tensors="pt")[0]
+            input_ids_list.append(input_ids)
+
+            targets = torch.full_like(input_ids, IGNORE_INDEX)
+            if message["role"] == "assistant" or message["role"] is None:
+                targets[self.generation_prompt_length:-1] = input_ids[self.generation_prompt_length:-1].clone()
+
+                # NOTE: mask out image tokens
+                vision_mask = input_ids == self.image_token_id
+                targets[vision_mask] = IGNORE_INDEX
+                vision_indices = torch.nonzero(vision_mask, as_tuple=True)[0]
+                targets[vision_indices + 1] = IGNORE_INDEX
+
+                # NOTE: mask out <think> or <think>\n
+                think_mask = targets == self.think_start_token_id
+                targets[think_mask] = IGNORE_INDEX
+                think_indices = torch.nonzero(think_mask, as_tuple=True)[0]
+                newline_mask = torch.zeros_like(think_mask)
+                newline_mask[think_indices + 1] = targets[think_indices + 1] == self.newline_token_id
+                targets[newline_mask] = IGNORE_INDEX
+
+            targets_list.append(targets)
+
+        assert len(grid_sizes) == image_idx, "Number of images does not match the number of image tokens in the text."
+
+        text_inputs = {
+            "input_ids": torch.cat(input_ids_list),
+            "labels": torch.cat(targets_list),
+        }
+
+        return text_inputs
+
+
+class RynnecQwen2Config(Qwen2Config):
+    model_type = "rynnec_qwen2"
+
+    def __init__(self, **kwargs):
+        super().__init__(**kwargs)
+        self.model_type = "rynnec_qwen2"
+
+
+class RynnecQwen2Model(RynnecMetaModel, Qwen2Model):
+    config_class = RynnecQwen2Config
+
+    def __init__(self, config: RynnecQwen2Config):
+        super(RynnecQwen2Model, self).__init__(config)
+
+        if hasattr(config, "mm_mask_decoder"): # inference
+            self.build_mask_decoder(config)
+        else: # training
+            if 'out_dim' not in config:
+                config.out_dim = 256        
+
+    def build_mask_decoder(self, config):
+            
+        # Projection layer for lisa
+        in_dim = config.hidden_size
+        out_dim = config.out_dim
+        text_fc = [
+            nn.Linear(in_dim, in_dim),
+            nn.ReLU(inplace=True),
+            nn.Linear(in_dim, out_dim),
+            nn.Dropout(0.0),
+        ]
+        self.text_hidden_fcs = nn.ModuleList([nn.Sequential(*text_fc)])
+        self.text_hidden_fcs.train()
+        for param in self.text_hidden_fcs.parameters():
+            param.requires_grad = True    
+
+
+class RynnecQwen2ForCausalLM(Qwen2ForCausalLM, RynnecMetaForCausalLM):
+    config_class = RynnecQwen2Config
+
+    def __init__(self, config, **kwargs):
+        super(Qwen2ForCausalLM, self).__init__(config)
+        self.model = RynnecQwen2Model(config)
+        self.vocab_size = config.vocab_size
+        self.lm_head = nn.Linear(config.hidden_size, config.vocab_size, bias=False)
+
+        # Initialize weights and apply final processing
+        self.post_init()
+        
+        if hasattr(config, "training") and config.training is True:
+            self.grounding_encoder = SAM2TrainRunner(ckpt_path=config.mask_decoder_model)
+            config.mm_mask_decoder = True
+        else:
+            self.grounding_encoder = SAM2(ckpt_path=config.mask_decoder_model)
+        
+        self.loss_mask = CrossEntropyLoss(
+            use_sigmoid=True,
+            reduction='mean',
+            loss_weight=2.0
+        )
+        self.loss_dice = DiceLoss(
+            use_sigmoid=True,
+            activate=True,
+            reduction='mean',
+            naive_dice=True,
+            eps=1.0,
+            loss_weight=0.5
+        )
+
+    def load_sam2_weights(self, model_path):
+        sam2_model = torch.load(model_path, map_location='cpu')['model']
+        prefix = "sam2_model."
+        new_state_dict = {}
+        for param_name in sam2_model.keys():
+            new_param_name = prefix + param_name
+            new_state_dict[new_param_name] = sam2_model[param_name]
+
+        self.grounding_encoder.load_state_dict(new_state_dict, strict=False)
+
+    def get_model(self):
+        return self.model
+    # NOTE: arguments are copied from transformers==4.46.3
+    def forward(
+        self,
+        input_ids: torch.LongTensor = None,
+        attention_mask: Optional[torch.Tensor] = None,
+        position_ids: Optional[torch.LongTensor] = None,
+        past_key_values: Optional[List[torch.FloatTensor]] = None,
+        inputs_embeds: Optional[torch.FloatTensor] = None,
+        labels: Optional[torch.LongTensor] = None,
+        use_cache: Optional[bool] = None,
+        output_attentions: Optional[bool] = None,
+        output_hidden_states: Optional[bool] = None,
+        return_dict: Optional[bool] = None,
+        cache_position: Optional[torch.LongTensor] = None,
+        num_logits_to_keep: int = 0,
+        # multimodal inputs
+        pixel_values: Optional[torch.FloatTensor] = None,
+        grid_sizes: Optional[torch.LongTensor] = None,
+        merge_sizes: Optional[torch.LongTensor] = None,
+        modals: Optional[List[str]] = None,
+        masks: Optional[List[torch.LongTensor]] = None,
+        mask_ids = None,
+        sam_images = None,
+        sam_size = None,
+        image2maskids = None,
+        **loss_kwargs,
+    ) -> Union[Tuple, CausalLMOutputWithPast]:
+        torch.cuda.empty_cache()
+        if inputs_embeds is None:
+            input_ids_raw = input_ids.clone()
+            (
+                input_ids,
+                attention_mask,
+                position_ids,
+                past_key_values,
+                inputs_embeds,
+                labels,
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                labels=labels,
+                pixel_values=pixel_values,
+                grid_sizes=grid_sizes,
+                merge_sizes=merge_sizes,
+                modals=modals,
+                masks=masks,
+                mask_ids=mask_ids
+            )
+
+        output_attentions = output_attentions if output_attentions is not None else self.config.output_attentions
+        output_hidden_states = (
+            output_hidden_states if output_hidden_states is not None else self.config.output_hidden_states
+        )
+        return_dict = return_dict if return_dict is not None else self.config.use_return_dict
+
+        # decoder outputs consists of (dec_features, layer_state, dec_hidden, dec_attn)
+        outputs = self.model(
+            input_ids=input_ids,
+            attention_mask=attention_mask,
+            position_ids=position_ids,
+            past_key_values=past_key_values,
+            inputs_embeds=inputs_embeds,
+            use_cache=use_cache,
+            output_attentions=output_attentions,
+            output_hidden_states=output_hidden_states,
+            return_dict=return_dict,
+            cache_position=cache_position,
+        )
+
+        hidden_states = outputs[0]
+        loss, logits = None, None
+        _valid = True
+        seg_valid = True
+
+        if labels is not None: #training
+
+            ce_loss = cross_entropy_loss(
+                hidden_states=hidden_states,
+                lm_head=self.lm_head,
+                position_ids=position_ids,
+                labels=labels,
+                reduction_scope=self.config.loss_reduction_scope,
+                **loss_kwargs,
+            )
+
+            if self.config.has_mask:
+
+                hidden_states_sam = []
+                hidden_states_sam.append(self.model.text_hidden_fcs[0](hidden_states))
+                hidden_states_sam = torch.stack(hidden_states_sam, dim=-1).sum(dim=-1)
+
+                bs = input_ids_raw.shape[0]
+                gt_masks_list = []
+                pred_masks_list = []
+                mask_bce_loss = 0
+                mask_dice_loss = 0
+                num_masks = 0
+                for i in range(bs):
+                    pred_masks = []
+                    pred_embeddings = []
+                    input_id = input_ids_raw[i]
+                    seg_token_mask = input_id[1:]==self.config.seg_token_index
+                    seg_token_mask = torch.cat(
+                        [
+                            seg_token_mask,
+                            torch.zeros((1)).bool().cuda(),
+                        ],
+                        dim=0,
+                    )
+
+                    pred_embedding = hidden_states_sam[i][seg_token_mask]
+                    if len(pred_embedding)>0:
+                        pred_embeddings.append(pred_embedding)
+                    else:
+                        pred_embeddings.append(hidden_states_sam[i, :1])
+        
+                
+                    gt_masks_video = []  # FIXME: Only support one segmentation now
+                    gt_mask = masks[i]
+                    mask_valid = True
+                    
+                    if len(image2maskids[i])==0:
+                        sam_images[i] = sam_images[i][:1]
+                        gt_masks_video.append(torch.zeros((len(sam_images[i]), 224, 224)).to(sam_images[0].device))
+                        mask_valid = False
+                        
+                    else:
+                        for mids in image2maskids[i]:
+                            for mid in mids:
+                                if mid is None:
+                                    gt_masks_video.append(torch.zeros((224, 224)).unsqueeze(0).to(gt_mask[0].device))
+                                else:
+                                    gt_masks_video.append(gt_mask[mid].unsqueeze(0))
+                    frames_per_batch = [len(sam_images[i])]
+                    try:
+                        pred_embeddings_list_video = genetate_video_pred_embeddings(pred_embeddings, frames_per_batch)
+
+                        # pred_embeddings_list_video, gt_masks_video = check_obj_number(pred_embeddings_list_video, gt_masks_video)
+
+                        g_pixel_values = sam_images[i]
+                        num_objs = len(pred_embeddings_list_video[0]) 
+                
+                        # with torch.no_grad():
+            
+                        sam_states = self.grounding_encoder.get_sam2_embeddings(g_pixel_values, expand_size=num_objs)
+                        language_embeddings = torch.cat(pred_embeddings_list_video, dim=0)[:, None]#.contiguous()
+
+                        num_frames = len(pred_embeddings_list_video)
+                        gt_masks_video = process_video_gt_masks(gt_masks_video, num_frames, num_objs)
+                        pred_masks = self.grounding_encoder.inject_language_embd(sam_states, language_embeddings, nf_nobj=(num_frames, num_objs))
+                        
+                        gt_masks = [F.interpolate(gt_mask.unsqueeze(0), size=pred_masks[0].shape[-2:], mode='nearest').squeeze(0) for gt_mask in gt_masks_video]
+                        gt_masks = torch.cat(gt_masks, dim=0)
+                        pred_masks = pred_masks.flatten(0, 1)
+
+                        if not mask_valid:
+                            pred_masks = pred_masks*0.0
+                    
+                        if len(pred_masks) != len(gt_masks):
+                            # drop this data
+                            print(f"Pred mask shape {pred_masks.shape} is not equal to gt_mask shape {gt_masks.shape} !!!")
+                            min_num = min(len(pred_masks), len(gt_masks))
+                            pred_masks = pred_masks[:min_num]
+                            gt_masks = gt_masks[:min_num]
+                            seg_valid = False
+
+                        if not seg_valid or not mask_valid:
+                            _scale = 0.0
+                        else:
+                            _scale = 1.0
+
+                        mask_bce_loss_ = self.loss_mask(pred_masks, gt_masks) * len(pred_masks) * _scale
+                        mask_dice_loss_ = self.loss_dice(pred_masks, gt_masks) * len(gt_masks) * _scale
+                        mask_bce_loss += mask_bce_loss_
+                        mask_dice_loss += mask_dice_loss_
+                        num_masks += len(pred_masks)
+                    except Exception as exp:
+                        print(exp) 
+                        _valid = False
+            
+
+                if num_masks>0:
+                    mask_bce_loss = mask_bce_loss / num_masks
+                    mask_dice_loss = mask_dice_loss / num_masks
+
+                mask_bce_loss = self.config.bce_loss_weight * mask_bce_loss 
+                mask_dice_loss = self.config.dice_loss_weight * mask_dice_loss 
+                if _valid==False:
+                    mask_bce_loss = mask_bce_loss * 0.0
+                    mask_dice_loss = mask_dice_loss* 0.0
+                
+                mask_loss = mask_bce_loss + mask_dice_loss
+                loss = mask_loss + ce_loss
+            else:
+                loss = ce_loss
+
+        else:
+            # Only compute necessary logits, and do not upcast them to float if we are not computing the loss
+            logits = self.lm_head(hidden_states[:, -num_logits_to_keep:, :])
+
+        if not return_dict:
+            output = (logits,) + outputs[1:]
+            return (loss,) + output if loss is not None else output
+
+        if loss is not None:
+            if self.config.has_mask:
+                return CausalLMOutputWithPast(
+                    loss=loss,
+                    ce_loss=ce_loss.detach(),
+                    mask_bce_loss=mask_bce_loss.detach(),
+                    mask_dice_loss=mask_dice_loss.detach(),
+                    mask_loss=mask_loss.detach(),
+                    logits=logits,
+                    past_key_values=outputs.past_key_values,
+                    hidden_states=outputs.hidden_states,
+                    attentions=outputs.attentions,
+                )
+            else:
+                return CausalLMOutputWithPast(
+                    loss=loss,
+                    logits=logits,
+                    past_key_values=outputs.past_key_values,
+                    hidden_states=outputs.hidden_states,
+                    attentions=outputs.attentions,
+                )
+        else: #infer
+            return CausalLMOutputWithPast(
+                loss=loss,
+                logits=logits,
+                past_key_values=outputs.past_key_values,
+                hidden_states=outputs.hidden_states,
+                attentions=outputs.attentions,
+            )
+
+    @torch.no_grad()
+    def inference(
+        self,
+        # multimodal inputs
+        pixel_values: Optional[torch.FloatTensor] = None,
+        grid_sizes: Optional[torch.LongTensor] = None,
+        merge_sizes: Optional[torch.LongTensor] = None,
+        modals: Optional[List[str]] = None,
+        masks: Optional[List[torch.LongTensor]] = None,
+        mask_ids = None,
+        sam_images = None,
+        sam_size = None,
+        image2maskids = None,
+        seg_start_idx = 0,
+        **kwargs,
+    ):
+        outputs = self.generate(
+            pixel_values=pixel_values,
+            grid_sizes=grid_sizes,
+            merge_sizes=merge_sizes,
+            modals=modals,
+            masks=masks,
+            mask_ids=mask_ids,
+            output_hidden_states=True,
+            return_dict_in_generate=True,
+            **kwargs
+        )
+
+        input_ids = kwargs.pop('input_ids')
+        last_hidden_state = []
+        for hs in outputs.hidden_states: # round
+            last_hidden_state.append(hs[-1])
+        last_hidden_state = torch.cat(last_hidden_state, dim=1)
+
+        output_ids = outputs.sequences
+
+        concat_ids = torch.cat((input_ids, output_ids), dim=1)
+        seg_token_mask = concat_ids[:, 1:] == self.config.seg_token_index
+
+        last_hidden_state_sam = self.model.text_hidden_fcs[0](last_hidden_state)
+
+        pred_embeddings = last_hidden_state_sam[seg_token_mask]
+        seg_token_counts = seg_token_mask.int().sum() 
+
+        if seg_token_counts>0:
+
+            g_pixel_values = torch.cat(sam_images, dim=0).contiguous()
+            num_objs = 1 #FIXME: Only support one segmentation now
+            if seg_start_idx>0:
+            # before start idx
+                g_pixel_values_beg = g_pixel_values[:seg_start_idx+1].flip(0)
+                num_frames = len(g_pixel_values_beg)
+                sam_states_beg = self.grounding_encoder.get_sam2_embeddings(g_pixel_values_beg)
+                pred_masks_beg = self.grounding_encoder.language_embd_inference(sam_states_beg, [pred_embeddings]*num_frames)
+            else:
+                pred_masks_beg = torch.zeros((1, 1, 1024, 1024)).to(pixel_values.device)
+            
+            if seg_start_idx<=len(g_pixel_values)-1:
+                g_pixel_values_end = g_pixel_values[seg_start_idx:]
+                num_frames = len(g_pixel_values_end)
+                sam_states_end = self.grounding_encoder.get_sam2_embeddings(g_pixel_values_end)
+                pred_masks_end = self.grounding_encoder.language_embd_inference(sam_states_end, [pred_embeddings]*num_frames)
+            else:
+                pred_masks_end = torch.zeros((0, 1, 1024, 1024)).to(pixel_values.device)
+            
+            pred_masks = torch.cat([pred_masks_beg[1:].flip(0), pred_masks_end], dim=0)
+
+        return output_ids, pred_masks
+
+        
+    @torch.no_grad()
+    def generate(
+        self,
+        # multimodal inputs
+        pixel_values: Optional[torch.FloatTensor] = None,
+        grid_sizes: Optional[torch.LongTensor] = None,
+        merge_sizes: Optional[torch.LongTensor] = None,
+        modals: Optional[List[str]] = None,
+        masks: Optional[List[torch.LongTensor]] = None,
+        mask_ids = None,
+        **kwargs,
+    ) -> Union[GenerateOutput, torch.LongTensor]:
+        input_ids = kwargs.pop("input_ids", None)
+        attention_mask = kwargs.pop("attention_mask", None)
+        position_ids = kwargs.pop("position_ids", None)
+        past_key_values = kwargs.pop("past_key_values", None)
+
+        if "inputs_embeds" in kwargs:
+            raise NotImplementedError("`inputs_embeds` is not supported")
+
+        if pixel_values is not None:
+            (
+                input_ids,
+                attention_mask,
+                position_ids,
+                past_key_values,
+                inputs_embeds,
+                labels,
+            ) = self.prepare_inputs_labels_for_multimodal(
+                input_ids=input_ids,
+                attention_mask=attention_mask,
+                position_ids=position_ids,
+                past_key_values=past_key_values,
+                labels=None,
+                pixel_values=pixel_values,
+                grid_sizes=grid_sizes,
+                merge_sizes=merge_sizes,
+                modals=modals,
+                masks=masks,
+                mask_ids=mask_ids
+            )
+        else:
+            inputs_embeds = self.get_model().embed_tokens(input_ids)
+
+        return super().generate(
+            position_ids=position_ids,
+            attention_mask=attention_mask,
+            inputs_embeds=inputs_embeds,
+            **kwargs
+        )
+
+    def prepare_inputs_for_generation(self, input_ids, past_key_values=None, inputs_embeds=None, **kwargs):
+        images = kwargs.pop("images", None)
+        _inputs = super().prepare_inputs_for_generation(
+            input_ids, past_key_values=past_key_values, inputs_embeds=inputs_embeds, **kwargs
+        )
+        if images is not None:
+            _inputs['images'] = images
+        return _inputs
+
+
+AutoConfig.register("rynnec_qwen2", RynnecQwen2Config)
+AutoModelForCausalLM.register(RynnecQwen2Config, RynnecQwen2ForCausalLM)
+AutoProcessor.register(RynnecQwen2Config, Videollama3Qwen2Processor)
+AutoImageProcessor.register(RynnecQwen2Config, Videollama3ImageProcessor)
diff --git a/rynnec/model/sam2.py b/rynnec/model/sam2.py
new file mode 100644
index 0000000000000000000000000000000000000000..8894f8770c1ae1ef18d1f723b8007dd4c38d5489
--- /dev/null
+++ b/rynnec/model/sam2.py
@@ -0,0 +1,133 @@
+# Adopted from https://github.com/magic-research/Sa2VA/blob/main/projects/llava_sam2/models/sam2.py.
+# Below is the original copyright:
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os.path
+
+import torch
+import torch.nn as nn
+from hydra import compose
+from hydra.utils import instantiate
+from omegaconf import OmegaConf
+
+from .utils import load_checkpoint_with_prefix, load_state_dict_to_model
+
+class SAM2(nn.Module):
+    def __init__(
+            self,
+            cfg_path: str = "sam2_hiera_l.yaml",
+            ckpt_path: str = "sam2_hiera_large.pt",
+            hydra_overrides_extra=None,
+            apply_postprocessing=True,
+    ):
+        super().__init__()
+
+        import third_parts.sam2 # noqa: F401
+
+        if hydra_overrides_extra is None:
+            hydra_overrides_extra = []
+        hydra_overrides = [
+            ## Extension: LLM prompt
+            "++model._target_=rynnec.model.predictor.SAM2VideoPredictor",
+        ]
+
+        if apply_postprocessing:
+            hydra_overrides_extra = hydra_overrides_extra.copy()
+            hydra_overrides_extra += [
+                # dynamically fall back to multi-mask if the single mask is not stable
+                "++model.sam_mask_decoder_extra_args.dynamic_multimask_via_stability=true",
+                "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_delta=0.05",
+                "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_thresh=0.98",
+                # the sigmoid mask logits on interacted frames with clicks in the memory encoder so that the encoded masks are exactly as what users see from clicking
+                # "++model.binarize_mask_from_pts_for_mem_enc=true",
+                # fill small holes in the low-res masks up to `fill_hole_area` (before resizing them to the original video resolution)
+                # "++model.fill_hole_area=8",
+            ]
+        hydra_overrides.extend(hydra_overrides_extra)
+
+        # Read config and init model
+        cfg = compose(config_name=cfg_path, overrides=hydra_overrides)
+        OmegaConf.resolve(cfg)
+        sam2_model = instantiate(cfg.model, _recursive_=True)
+        state_dict = load_checkpoint_with_prefix(ckpt_path)
+        load_state_dict_to_model(sam2_model, state_dict)
+
+        self.sam2_model = sam2_model
+
+        self.hidden_dim = self.sam2_model.hidden_dim
+
+        self.img_mean = (0.485, 0.456, 0.406)
+        self.img_std = (0.229, 0.224, 0.225)
+
+    def inject_language_embd(self, inference_state, language_embd):
+        num_frame = len(language_embd)
+        num_obj = len(language_embd[0])
+        mask_out = []
+        for frame_idx in range(num_frame):
+            frame_mask_out = []
+            for obj_idx in range(num_obj):
+                _language_embd = language_embd[frame_idx][obj_idx][None][None]
+                _, _, out_mask_logits = self.sam2_model.add_language_embd(inference_state, frame_idx, obj_idx + 100, _language_embd)
+                frame_mask_out.append(out_mask_logits)
+            frame_mask_out = torch.cat(frame_mask_out, dim=1)
+            mask_out.append(frame_mask_out)
+        mask_out = torch.cat(mask_out, dim=0)
+        return mask_out
+
+
+    def language_embd_inference(self, inference_state, language_embd):
+        num_frame = len(language_embd)
+        num_obj = len(language_embd[0])
+        mask_out = []
+        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+            for frame_idx in range(num_frame):
+                frame_mask_out = []
+
+                for obj_idx in range(num_obj):
+                    _language_embd = language_embd[frame_idx][obj_idx][None][None]
+                    _, _, out_mask_logits = self.sam2_model.add_language_embd(
+                        inference_state,
+                        frame_idx,
+                        obj_idx + 100,
+                        _language_embd,
+                        inference=True,
+                    )
+                    frame_mask_out.append(out_mask_logits)
+                frame_mask_out = torch.cat(frame_mask_out, dim=1)
+                mask_out.append(frame_mask_out)
+
+
+            mask_out = []
+            for out_frame_idx, out_obj_ids, out_mask_logits in self.sam2_model.propagate_in_video(inference_state):
+                mask_out.append(out_mask_logits)
+            mask_out = torch.cat(mask_out, dim=0)
+        return mask_out
+
+    def get_sam2_embeddings(self, images):
+        return self.sam2_model.init_state(images)
+
+    def forward(self, batch):
+        raise NotImplementedError
+
+    def preprocess_image(self, image: torch.Tensor, dtype=torch.float32) -> torch.Tensor:
+        image = image / 255.
+
+        img_mean = torch.tensor(self.img_mean, dtype=dtype, device=image.device)[:, None, None]
+        img_std = torch.tensor(self.img_std, dtype=dtype, device=image.device)[:, None, None]
+        image -= img_mean
+        image /= img_std
+
+        return image
diff --git a/rynnec/model/sam2_train.py b/rynnec/model/sam2_train.py
new file mode 100644
index 0000000000000000000000000000000000000000..d718f0a5732d7394b3964d969982a959dc759bf3
--- /dev/null
+++ b/rynnec/model/sam2_train.py
@@ -0,0 +1,134 @@
+# Adopted from https://github.com/magic-research/Sa2VA/blob/main/projects/llava_sam2/models/sam2_train.py.
+# Below is the original copyright:
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+import os.path
+
+import torch
+import torch.nn as nn
+
+from hydra import compose
+from hydra.utils import instantiate
+from omegaconf import OmegaConf
+
+from .utils import load_checkpoint_with_prefix, load_state_dict_to_model
+
+BASE_DIR = 'pretrained/'
+
+
+class SAM2TrainRunner(nn.Module):
+    def __init__(
+            self,
+            cfg_path: str = "sam2_hiera_l.yaml",
+            ckpt_path: str = "sam2_hiera_large.pt",
+            hydra_overrides_extra=None,
+            apply_postprocessing=True,
+    ):
+        super().__init__()
+
+        import third_parts.sam2 # noqa: F401
+
+        if hydra_overrides_extra is None:
+            hydra_overrides_extra = []
+        hydra_overrides = [
+            ## Extension: LLM prompt
+            "++model._target_=rynnec.model.extension.SAM2Base",
+        ]
+
+        if apply_postprocessing:
+            hydra_overrides_extra = hydra_overrides_extra.copy()
+            
+        hydra_overrides.extend(hydra_overrides_extra)
+
+        # Read config and init model
+        cfg = compose(config_name=cfg_path, overrides=hydra_overrides)
+        OmegaConf.resolve(cfg)
+        sam2_model = instantiate(cfg.model, _recursive_=True)
+        state_dict = load_checkpoint_with_prefix(ckpt_path)
+        load_state_dict_to_model(sam2_model, state_dict)
+
+        self.sam2_model = sam2_model
+
+        self.hidden_dim = self.sam2_model.hidden_dim
+        self.img_mean = (0.485, 0.456, 0.406)
+        self.img_std = (0.229, 0.224, 0.225)
+
+    def preprocess_image(self, image: torch.Tensor) -> torch.Tensor:
+        image = image / 255.
+        img_mean = torch.tensor(self.img_mean, dtype=image.dtype, device=image.device)[:, None, None]
+        img_std = torch.tensor(self.img_std, dtype=image.dtype, device=image.device)[:, None, None]
+        image -= img_mean
+        image /= img_std
+        return image
+
+    def inject_language_embd(self, sam_states, language_embd, nf_nobj=None):
+        high_res_features = [
+            x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
+            for x, s in zip(sam_states['current_vision_feats'][:-1], sam_states['feat_sizes'][:-1])
+        ]
+
+        B = sam_states['current_vision_feats'][-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = sam_states['feat_sizes'][-1]
+
+        if self.sam2_model.directly_add_no_mem_embed:
+            # directly add no-mem embedding (instead of using the transformer encoder)
+            pix_feat_with_mem = sam_states['current_vision_feats'][-1] + self.sam2_model.no_mem_embed
+            pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+        else:
+            raise NotImplementedError("directly add no memory embedding is not implemented")
+        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+            _, _, _, low_res_masks, high_res_masks, obj_ptr, _, = self.sam2_model._forward_sam_heads(
+                backbone_features=pix_feat_with_mem,
+                point_inputs=None,
+                mask_inputs=None,
+                high_res_features=high_res_features,
+                multimask_output=self.sam2_model._use_multimask(is_init_cond_frame=True, point_inputs=None),
+                # Inject language Embed if possible
+                language_embd=language_embd,
+            )
+
+        if nf_nobj is not None:
+            pred_masks = low_res_masks.squeeze(1)
+            pred_masks = pred_masks.unflatten(0, nf_nobj)
+        else:
+            pred_masks = low_res_masks
+        return pred_masks
+
+    def get_sam2_embeddings(self, images, expand_size=1):
+        # Step 1: inference the backbone with the images
+        with torch.autocast(device_type="cuda", dtype=torch.bfloat16):
+            feats = self.sam2_model.forward_image(images)
+
+        if expand_size > 1:
+            # feats['vision_features'] = feats['vision_features'][:, None].expand(-1, expand_size, -1, -1, -1).flatten(0, 1)
+            for i, feat in enumerate(feats["backbone_fpn"]):
+                feats["backbone_fpn"][i] = feat[:, None].expand(-1, expand_size, -1, -1, -1).flatten(0, 1)
+            for i, pos in enumerate(feats["vision_pos_enc"]):
+                pos = pos[:, None].expand(-1, expand_size, -1, -1, -1).flatten(0, 1)
+                feats["vision_pos_enc"][i] = pos
+
+        # Step 2: Process the features to output
+        _, current_vision_feats, current_vision_pos_embeds, feat_sizes = self.sam2_model._prepare_backbone_features(feats)
+
+        return {
+            "current_vision_feats": current_vision_feats,
+            "current_vision_pos_embeds": current_vision_pos_embeds,
+            "feat_sizes": feat_sizes,
+        }
+
+    def forward(self, batch):
+        raise NotImplementedError
diff --git a/rynnec/model/utils.py b/rynnec/model/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..2460069653c7ad49459a049b598c6e7b048ff7dc
--- /dev/null
+++ b/rynnec/model/utils.py
@@ -0,0 +1,61 @@
+import torch
+import torch.nn.functional as F
+import torch.nn as nn
+from torch import Tensor
+import logging
+from huggingface_hub import hf_hub_download
+import functools
+from typing import Callable, Optional
+
+def process_video_gt_masks(gt_masks, num_frames, num_objs):
+    gt_masks_processed = []
+    for i in range(num_frames):
+        for j in range(num_objs):
+            gt_masks_processed.append(gt_masks[j*num_frames+i])
+    return gt_masks_processed
+
+def load_checkpoint_with_prefix(filename, prefix=None, map_location='cpu', logger='current'):
+    HF_HUB_PREFIX = 'hf-hub:'
+    if filename.startswith(HF_HUB_PREFIX):
+        model_id = filename[len(HF_HUB_PREFIX):]
+        filename = hf_hub_download(model_id, 'pytorch_model.bin')
+
+    checkpoint = torch.load(filename, map_location=map_location)
+
+    if 'state_dict' in checkpoint:
+        state_dict = checkpoint['state_dict']
+    elif 'model' in checkpoint:
+        state_dict = checkpoint['model']
+    else:
+        state_dict = checkpoint
+    if not prefix:
+        return state_dict
+    if not prefix.endswith('.'):
+        prefix += '.'
+    prefix_len = len(prefix)
+
+    state_dict = {
+        k[prefix_len:]: v
+        for k, v in state_dict.items() if k.startswith(prefix)
+    }
+
+    assert state_dict, f'{prefix} is not in the pretrained model'
+    return state_dict
+
+
+def load_state_dict_to_model(model, state_dict,  logger='current'):
+    missing_keys, unexpected_keys = model.load_state_dict(state_dict)
+    if missing_keys:
+        raise RuntimeError()
+    if unexpected_keys:
+        raise RuntimeError()
+
+def genetate_video_pred_embeddings(pred_embeddings_list, frames_per_batch):
+    assert len(pred_embeddings_list) == len(frames_per_batch), \
+    f"Lengths do not match: len(pred_embeddings_list)={len(pred_embeddings_list)}, len(frames_per_batch)={len(frames_per_batch)}"
+    
+    pred_embeddings_list_video = []
+    for pred_embedding_batch, frame_nums in zip(pred_embeddings_list, frames_per_batch):
+        pred_embeddings_list_video += [pred_embedding_batch] * frame_nums
+    return pred_embeddings_list_video
+
diff --git a/rynnec/model/videollama3_encoder/__init__.py b/rynnec/model/videollama3_encoder/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..b5c8f93577a46c3eb647ae4c0e07c4650d17a239
--- /dev/null
+++ b/rynnec/model/videollama3_encoder/__init__.py
@@ -0,0 +1,3 @@
+from .configuration_videollama3_encoder import Videollama3VisionEncoderConfig
+from .image_processing_videollama3 import Videollama3ImageProcessor
+from .modeling_videollama3_encoder import Videollama3VisionEncoderModel
diff --git a/rynnec/model/videollama3_encoder/configuration_videollama3_encoder.py b/rynnec/model/videollama3_encoder/configuration_videollama3_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c30709762b254752af9f6b2f56982e7a64106aa
--- /dev/null
+++ b/rynnec/model/videollama3_encoder/configuration_videollama3_encoder.py
@@ -0,0 +1,71 @@
+# Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/siglip/configuration_siglip.py.
+# Below is the original copyright:
+# coding=utf-8
+# Copyright 2024 The HuggingFace Inc. team. All rights reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""VideoLLaMA3 vision encoder model configuration."""
+import os
+from typing import Union
+
+from transformers import PretrainedConfig
+from transformers.utils import logging
+
+
+logger = logging.get_logger(__name__)
+
+class Videollama3VisionEncoderConfig(PretrainedConfig):
+
+    model_type = "videollama3_vision_encoder"
+
+    def __init__(
+        self,
+        hidden_size=768,
+        intermediate_size=3072,
+        num_hidden_layers=12,
+        num_attention_heads=12,
+        num_channels=3,
+        patch_size=16,
+        hidden_act="gelu_pytorch_tanh",
+        layer_norm_eps=1e-6,
+        attention_dropout=0.0,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+
+        self.hidden_size = hidden_size
+        self.intermediate_size = intermediate_size
+        self.num_hidden_layers = num_hidden_layers
+        self.num_attention_heads = num_attention_heads
+        self.num_channels = num_channels
+        self.patch_size = patch_size
+        self.attention_dropout = attention_dropout
+        self.layer_norm_eps = layer_norm_eps
+        self.hidden_act = hidden_act
+
+    # @classmethod
+    # def from_pretrained(cls, pretrained_model_name_or_path: Union[str, os.PathLike], **kwargs) -> "PretrainedConfig":
+    #     cls._set_token_in_kwargs(kwargs)
+
+    #     config_dict, kwargs = cls.get_config_dict(pretrained_model_name_or_path, **kwargs)
+
+    #     p
+    #     config_dict = config_dict["vision_config"]
+
+    #     if "model_type" in config_dict and hasattr(cls, "model_type") and config_dict["model_type"] != cls.model_type:
+    #         logger.warning(
+    #             f"You are using a model of type {config_dict['model_type']} to instantiate a model of type "
+    #             f"{cls.model_type}. This is not supported for all configurations of models and can yield errors."
+    #         )
+
+    #     return cls.from_dict(config_dict, **kwargs)
diff --git a/rynnec/model/videollama3_encoder/image_processing_videollama3.py b/rynnec/model/videollama3_encoder/image_processing_videollama3.py
new file mode 100644
index 0000000000000000000000000000000000000000..eead08b5d4e67b7d579b633f416bee94aa5e68ff
--- /dev/null
+++ b/rynnec/model/videollama3_encoder/image_processing_videollama3.py
@@ -0,0 +1,473 @@
+# Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/image_processing_qwen2_vl.py.
+# Below is the original copyright:
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""Image processor class for VideoLLaMA3."""
+
+import math
+from typing import Dict, List, Optional, Union
+
+import numpy as np
+
+import torch
+from transformers.image_processing_utils import BaseImageProcessor, BatchFeature
+from transformers.image_utils import ImageInput
+from transformers.image_transforms import (
+    convert_to_rgb,
+    resize,
+    to_channel_dimension_format,
+)
+from transformers.image_utils import (
+    OPENAI_CLIP_MEAN,
+    OPENAI_CLIP_STD,
+    ChannelDimension,
+    ImageInput,
+    PILImageResampling,
+    VideoInput,
+    get_image_size,
+    infer_channel_dimension_format,
+    is_scaled_image,
+    is_valid_image,
+    make_list_of_images,
+    to_numpy_array,
+)
+from transformers.utils import TensorType, is_vision_available, logging
+
+
+logger = logging.get_logger(__name__)
+
+
+if is_vision_available():
+    from PIL import Image
+
+
+def is_valid_video(video) -> bool:
+    if isinstance(video, (list, tuple)):
+        return all(is_valid_image(frame) for frame in video)
+    elif isinstance(video, np.ndarray):
+        return video.ndim == 4
+    elif isinstance(video, torch.Tensor):
+        return video.ndim == 4
+    return False
+
+
+def make_batched_images(images) -> List[List[ImageInput]]:
+    """
+    Accepts images in list or nested list format, and makes a list of images for preprocessing.
+
+    Args:
+        images (`Union[List[List[ImageInput]], List[ImageInput], ImageInput]`):
+            The input image.
+
+    Returns:
+        list: A list of images.
+    """
+    if isinstance(images, (list, tuple)):
+        # list of images/videos
+        if not all(is_valid_video(image) or is_valid_image(image) for image in images):
+            raise ValueError(f"Could not make batched images from {images}")
+        return images
+    elif is_valid_video(images) or is_valid_image(images):
+        # single image/video
+        return [images]
+
+    raise ValueError(f"Could not make batched images from {images}")
+
+
+def simple_batched_resize(
+    images, factor: int = 28, min_tokens: int = 4 * 4, max_tokens: int = 16384, input_data_format: str = None
+):
+    min_pixels = min_tokens * factor * factor
+    max_pixels = max_tokens * factor * factor
+
+    num_images = 0
+    for image in images:
+        if is_valid_video(image):
+            num_images += len(image)
+        else:
+            num_images += 1
+
+    image_sizes = []
+    for image in images:
+        if is_valid_video(image):
+            image = image[0]
+        if isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            height, width = get_image_size(image, channel_dim=input_data_format)
+        image_sizes.append([height, width])
+
+    tmp_image_sizes = []
+    for height, width in image_sizes:
+        h_bar = round(height / factor) * factor
+        w_bar = round(width / factor) * factor
+        if h_bar * w_bar > (max_pixels // num_images):
+            beta = math.sqrt((height * width) / (max_pixels // num_images))
+            h_bar = math.floor(height / beta / factor) * factor
+            w_bar = math.floor(width / beta / factor) * factor
+        # per image min_pixels
+        if h_bar * w_bar < min_pixels:
+            beta = math.sqrt(min_pixels / (height * width))
+            h_bar = math.ceil(height * beta / factor) * factor
+            w_bar = math.ceil(width * beta / factor) * factor
+        tmp_image_sizes.append((h_bar, w_bar))
+    image_sizes = tmp_image_sizes
+    return image_sizes
+
+
+def batched_resize(
+    images, factors: List[int], min_tokens: int = 4 * 4, max_tokens: int = 16384, input_data_format: str = None
+):
+    image_sizes = []
+    for image in images:
+        if is_valid_video(image):
+            num_frame = len(image)
+            image = image[0]
+        else:
+            num_frame = 1
+        if isinstance(image, Image.Image):
+            width, height = image.size
+        else:
+            height, width = get_image_size(image, channel_dim=input_data_format)
+        image_sizes.append([num_frame, height, width])
+
+    # global max_pixels
+    smart_scale_factors = 1.0
+    total_tokens = 0
+    for (num_frame, height, width), factor in zip(image_sizes, factors):
+        total_tokens += num_frame * math.ceil(height / factor) * math.ceil(width / factor)
+
+    # TODO: add min_pixels
+    if total_tokens > max_tokens:
+        beta = math.sqrt(total_tokens / max_tokens)
+        tmp_image_sizes = []
+        for (_, height, width), factor in zip(image_sizes, factors):
+            h_bar = math.floor(height / beta / factor) * factor
+            w_bar = math.floor(width / beta / factor) * factor
+            tmp_image_sizes.append((h_bar, w_bar))
+        image_sizes = tmp_image_sizes
+    else:
+        tmp_image_sizes = []
+        for (_, height, width), factor in zip(image_sizes, factors):
+            height = round(height / factor) * factor
+            width = round(width / factor) * factor
+            tmp_image_sizes.append((height, width))
+        image_sizes = tmp_image_sizes
+
+    return image_sizes
+
+
+class Videollama3ImageProcessor(BaseImageProcessor):
+    r"""
+    Constructs a DAMOVL image processor that dynamically resizes images based on the original images.
+
+    Args:
+        do_resize (`bool`, *optional*, defaults to `True`):
+            Whether to resize the image's (height, width) dimensions.
+        resample (`PILImageResampling`, *optional*, defaults to `Resampling.BICUBIC`):
+            Resampling filter to use when resizing the image.
+        do_rescale (`bool`, *optional*, defaults to `True`):
+            Whether to rescale the image by the specified scale `rescale_factor`.
+        rescale_factor (`int` or `float`, *optional*, defaults to `1/255`):
+            Scale factor to use if rescaling the image.
+        do_normalize (`bool`, *optional*, defaults to `True`):
+            Whether to normalize the image.
+        image_mean (`float` or `List[float]`, *optional*, defaults to `[0.48145466, 0.4578275, 0.40821073]`):
+            Mean to use if normalizing the image. This is a float or list of floats for each channel in the image.
+        image_std (`float` or `List[float]`, *optional*, defaults to `[0.26862954, 0.26130258, 0.27577711]`):
+            Standard deviation to use if normalizing the image. This is a float or list of floats for each channel in the image.
+        do_convert_rgb (`bool`, *optional*, defaults to `True`):
+            Whether to convert the image to RGB.
+        min_pixels (`int`, *optional*, defaults to `56 * 56`):
+            The min pixels of the image to resize the image.
+        max_pixels (`int`, *optional*, defaults to `28 * 28 * 1280`):
+            The max pixels of the image to resize the image.
+        patch_size (`int`, *optional*, defaults to 14):
+            The spacial patch size of the vision encoder.
+    """
+
+    model_input_names = ["pixel_values", "grid_sizes", "merge_sizes"]
+
+    def __init__(
+        self,
+        do_resize: bool = True,
+        resample: PILImageResampling = PILImageResampling.BICUBIC,
+        do_rescale: bool = True,
+        rescale_factor: Union[int, float] = 1 / 255,
+        do_normalize: bool = True,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = True,
+        min_tokens: int = 4 * 4,
+        max_tokens: int = 16384,
+        patch_size: int = 14,
+        **kwargs,
+    ) -> None:
+        super().__init__(**kwargs)
+        self.do_resize = do_resize
+        self.resample = resample
+        self.do_rescale = do_rescale
+        self.rescale_factor = rescale_factor
+        self.do_normalize = do_normalize
+        self.image_mean = image_mean if image_mean is not None else OPENAI_CLIP_MEAN
+        self.image_std = image_std if image_std is not None else OPENAI_CLIP_STD
+        self.min_tokens = min_tokens
+        self.max_tokens = max_tokens
+        self.patch_size = patch_size
+        self.do_convert_rgb = do_convert_rgb
+
+    def _preprocess(
+        self,
+        images: Union[ImageInput, VideoInput],
+        target_size: List[int],
+        merge_size: int = 1,
+        do_resize: bool = None,
+        resample: PILImageResampling = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Preprocess an image or batch of images. Copy of the `preprocess` method from `CLIPImageProcessor`.
+
+        Args:
+            images (`ImageInput`):
+                Image or batch of images to preprocess. Expects pixel values ranging from 0 to 255. If pixel values range from 0 to 1, set `do_rescale=False`.
+            target_size (`List[int]`):
+                The target size to resize the image to. Should be a list of two integers: [target_height, target_width].
+            merge_size (`int`, *optional*, defaults to `1`):
+                The merge size after the vision encoder.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            resample (`PILImageResampling`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the `PILImageResampling` enums.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Scale factor to use if rescaling the image.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Mean to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Standard deviation to use if normalizing the image. Can be a float or a list of floats corresponding to the number of channels in the image.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            data_format (`ChannelDimension`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.   - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+        """
+        images = make_list_of_images(images)
+
+        if do_convert_rgb:
+            images = [convert_to_rgb(image) for image in images]
+
+        # All transformations expect numpy arrays.
+        images = [to_numpy_array(image) for image in images]
+
+        if is_scaled_image(images[0]) and do_rescale:
+            logger.warning_once(
+                "It looks like you are trying to rescale already rescaled images. If the input"
+                " images have pixel values between 0 and 1, set `do_rescale=False` to avoid rescaling them again."
+            )
+        if input_data_format is None:
+            # We assume that all images have the same channel dimension format.
+            input_data_format = infer_channel_dimension_format(images[0])
+
+        height, width = get_image_size(images[0], channel_dim=input_data_format)
+        resized_height, resized_width = height, width
+        processed_images = []
+        for image in images:
+            if do_resize:
+                resized_height, resized_width = target_size
+                image = resize(
+                    image, size=(resized_height, resized_width), resample=resample, input_data_format=input_data_format
+                )
+
+            if do_rescale:
+                image = self.rescale(image, scale=rescale_factor, input_data_format=input_data_format)
+
+            if do_normalize:
+                image = self.normalize(
+                    image=image, mean=image_mean, std=image_std, input_data_format=input_data_format
+                )
+
+            image = to_channel_dimension_format(image, data_format, input_channel_dim=input_data_format)
+            processed_images.append(image)
+
+        patches = np.array(processed_images)
+        if data_format == ChannelDimension.LAST:
+            patches = patches.transpose(0, 3, 1, 2)
+        t = patches.shape[0]
+        channel = patches.shape[1]
+        grid_h, grid_w = resized_height // self.patch_size, resized_width // self.patch_size
+        patches = patches.reshape(
+            t,
+            channel,
+            grid_h // merge_size,
+            merge_size,
+            self.patch_size,
+            grid_w // merge_size,
+            merge_size,
+            self.patch_size,
+        )
+        patches = patches.transpose(0, 2, 5, 3, 6, 1, 4, 7)
+        flatten_patches = patches.reshape(
+            t * grid_h * grid_w, channel * self.patch_size * self.patch_size
+        )
+
+        return flatten_patches, (t, grid_h, grid_w)
+
+    def preprocess(
+        self,
+        images: ImageInput,
+        do_resize: bool = None,
+        resample: PILImageResampling = None,
+        do_rescale: bool = None,
+        rescale_factor: float = None,
+        do_normalize: bool = None,
+        image_mean: Optional[Union[float, List[float]]] = None,
+        image_std: Optional[Union[float, List[float]]] = None,
+        do_convert_rgb: bool = None,
+        merge_size: Optional[Union[int, List[int]]] = None,
+        return_tensors: Optional[Union[str, TensorType]] = None,
+        data_format: Optional[ChannelDimension] = ChannelDimension.FIRST,
+        input_data_format: Optional[Union[str, ChannelDimension]] = None,
+    ):
+        """
+        Args:
+            images (`ImageInput`):
+                Image to preprocess. Expects a single or batch of images with pixel values ranging from 0 to 255. If
+                passing in images with pixel values between 0 and 1, set `do_rescale=False`.
+            do_resize (`bool`, *optional*, defaults to `self.do_resize`):
+                Whether to resize the image.
+            resample (`int`, *optional*, defaults to `self.resample`):
+                Resampling filter to use if resizing the image. This can be one of the enum `PILImageResampling`. Only
+                has an effect if `do_resize` is set to `True`.
+            do_rescale (`bool`, *optional*, defaults to `self.do_rescale`):
+                Whether to rescale the image.
+            rescale_factor (`float`, *optional*, defaults to `self.rescale_factor`):
+                Rescale factor to rescale the image by if `do_rescale` is set to `True`.
+            do_normalize (`bool`, *optional*, defaults to `self.do_normalize`):
+                Whether to normalize the image.
+            image_mean (`float` or `List[float]`, *optional*, defaults to `self.image_mean`):
+                Image mean to use for normalization. Only has an effect if `do_normalize` is set to `True`.
+            image_std (`float` or `List[float]`, *optional*, defaults to `self.image_std`):
+                Image standard deviation to use for normalization. Only has an effect if `do_normalize` is set to
+                `True`.
+            do_convert_rgb (`bool`, *optional*, defaults to `self.do_convert_rgb`):
+                Whether to convert the image to RGB.
+            return_tensors (`str` or `TensorType`, *optional*):
+                The type of tensors to return. Can be one of:
+                - Unset: Return a list of `np.ndarray`.
+                - `TensorType.TENSORFLOW` or `'tf'`: Return a batch of type `tf.Tensor`.
+                - `TensorType.PYTORCH` or `'pt'`: Return a batch of type `torch.Tensor`.
+                - `TensorType.NUMPY` or `'np'`: Return a batch of type `np.ndarray`.
+                - `TensorType.JAX` or `'jax'`: Return a batch of type `jax.numpy.ndarray`.
+            data_format (`ChannelDimension` or `str`, *optional*, defaults to `ChannelDimension.FIRST`):
+                The channel dimension format for the output image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - Unset: Use the channel dimension format of the input image.
+            input_data_format (`ChannelDimension` or `str`, *optional*):
+                The channel dimension format for the input image. If unset, the channel dimension format is inferred
+                from the input image. Can be one of:
+                - `"channels_first"` or `ChannelDimension.FIRST`: image in (num_channels, height, width) format.
+                - `"channels_last"` or `ChannelDimension.LAST`: image in (height, width, num_channels) format.
+                - `"none"` or `ChannelDimension.NONE`: image in (height, width) format.
+
+        """
+        do_resize = do_resize if do_resize is not None else self.do_resize
+        resample = resample if resample is not None else self.resample
+        do_rescale = do_rescale if do_rescale is not None else self.do_rescale
+        rescale_factor = rescale_factor if rescale_factor is not None else self.rescale_factor
+        do_normalize = do_normalize if do_normalize is not None else self.do_normalize
+        image_mean = image_mean if image_mean is not None else self.image_mean
+        image_std = image_std if image_std is not None else self.image_std
+        merge_size = merge_size if merge_size is not None else self.merge_size
+        do_convert_rgb = do_convert_rgb if do_convert_rgb is not None else self.do_convert_rgb
+
+        images = make_batched_images(images)
+
+        if isinstance(merge_size, (list, tuple)):
+            assert len(merge_size) == len(images), "Merge size must be the same length as images."
+            merge_sizes = merge_size
+        else:
+            merge_sizes = [merge_size for _ in images]
+
+        if all(merge_size == merge_sizes[0] for merge_size in merge_sizes):
+            target_sizes = simple_batched_resize(
+                images,
+                factor=self.patch_size * merge_sizes[0],
+                min_tokens=self.min_tokens,
+                max_tokens=self.max_tokens,
+                input_data_format=input_data_format,
+            )
+        else:
+            target_sizes = batched_resize(
+                images,
+                factors=[self.patch_size * merge_size for merge_size in merge_sizes],
+                min_tokens=self.min_tokens,
+                max_tokens=self.max_tokens,
+                input_data_format=input_data_format,
+            )
+
+        pixel_values, grid_sizes = [], []
+        for image, merge_size, target_size in zip(images, merge_sizes, target_sizes):
+            patches, grid_size = self._preprocess(
+                image,
+                target_size=target_size,
+                merge_size=merge_size,
+                do_resize=do_resize,
+                resample=resample,
+                do_rescale=do_rescale,
+                rescale_factor=rescale_factor,
+                do_normalize=do_normalize,
+                image_mean=image_mean,
+                image_std=image_std,
+                data_format=data_format,
+                do_convert_rgb=do_convert_rgb,
+                input_data_format=input_data_format,
+            )
+            pixel_values.append(patches)
+            grid_sizes.append(grid_size)
+
+        pixel_values = np.concatenate(pixel_values, axis=0)
+        grid_sizes = np.array(grid_sizes)
+        merge_sizes = np.array(merge_sizes)
+
+        data = {
+            "pixel_values": pixel_values,
+            "grid_sizes": grid_sizes,
+            "merge_sizes": merge_sizes,
+        }
+
+        return BatchFeature(data=data, tensor_type=return_tensors)
diff --git a/rynnec/model/videollama3_encoder/modeling_videollama3_encoder.py b/rynnec/model/videollama3_encoder/modeling_videollama3_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..cd368a46961f62fef4d26066825c9f7e95bfa58f
--- /dev/null
+++ b/rynnec/model/videollama3_encoder/modeling_videollama3_encoder.py
@@ -0,0 +1,555 @@
+# Adopted from https://github.com/huggingface/transformers/blob/main/src/transformers/models/qwen2_vl/modeling_qwen2_vl.py.
+# Below is the original copyright:
+# Copyright 2024 The Qwen team, Alibaba Group and the HuggingFace Inc. team. All rights reserved.
+#
+# This code is based on EleutherAI's GPT-NeoX library and the GPT-NeoX
+# and OPT implementations in this library. It has been modified from its
+# original forms to accommodate minor architectural differences compared
+# to GPT-NeoX and OPT used by the Meta AI team that trained the model.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+"""PyTorch VideoLLaMA3 vision encoder model."""
+
+import importlib.util
+import os.path as osp
+import math
+import warnings
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import torch.utils.checkpoint
+from torch.nn.init import _calculate_fan_in_and_fan_out
+
+from transformers.activations import ACT2FN
+from transformers.modeling_utils import PreTrainedModel
+from transformers.utils import is_flash_attn_2_available
+
+if is_flash_attn_2_available():
+    from flash_attn import flash_attn_varlen_func
+else:
+    flash_attn_varlen_func = None
+
+try:
+    from .configuration_videollama3_encoder import Videollama3VisionEncoderConfig
+except ImportError:
+    spec = importlib.util.spec_from_file_location(
+        "configuration_videollama3_encoder",
+        osp.join(osp.dirname(__file__), "configuration_videollama3_encoder.py"),
+    )
+    configuration_videollama3_encoder = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(configuration_videollama3_encoder)
+    Videollama3VisionEncoderConfig = getattr(
+        configuration_videollama3_encoder,
+        "Videollama3VisionEncoderConfig",
+    )
+
+LayerNorm = nn.LayerNorm
+
+
+def _trunc_normal_(tensor, mean, std, a, b):
+    # Cut & paste from PyTorch official master until it's in a few official releases - RW
+    # Method based on https://people.sc.fsu.edu/~jburkardt/presentations/truncated_normal.pdf
+    def norm_cdf(x):
+        # Computes standard normal cumulative distribution function
+        return (1.0 + math.erf(x / math.sqrt(2.0))) / 2.0
+
+    if (mean < a - 2 * std) or (mean > b + 2 * std):
+        warnings.warn(
+            "mean is more than 2 std from [a, b] in nn.init.trunc_normal_. "
+            "The distribution of values may be incorrect.",
+            stacklevel=2,
+        )
+
+    # Values are generated by using a truncated uniform distribution and
+    # then using the inverse CDF for the normal distribution.
+    # Get upper and lower cdf values
+    l = norm_cdf((a - mean) / std)
+    u = norm_cdf((b - mean) / std)
+
+    # Uniformly fill tensor with values from [l, u], then translate to
+    # [2l-1, 2u-1].
+    tensor.uniform_(2 * l - 1, 2 * u - 1)
+
+    # Use inverse cdf transform for normal distribution to get truncated
+    # standard normal
+    tensor.erfinv_()
+
+    # Transform to proper mean, std
+    tensor.mul_(std * math.sqrt(2.0))
+    tensor.add_(mean)
+
+    # Clamp to ensure it's in the proper range
+    tensor.clamp_(min=a, max=b)
+
+
+def trunc_normal_tf_(
+    tensor: torch.Tensor, mean: float = 0.0, std: float = 1.0, a: float = -2.0, b: float = 2.0
+) -> torch.Tensor:
+    """Fills the input Tensor with values drawn from a truncated
+    normal distribution. The values are effectively drawn from the
+    normal distribution :math:`\\mathcal{N}(\text{mean}, \text{std}^2)`
+    with values outside :math:`[a, b]` redrawn until they are within
+    the bounds. The method used for generating the random values works
+    best when :math:`a \\leq \text{mean} \\leq b`.
+
+    NOTE: this 'tf' variant behaves closer to Tensorflow / JAX impl where the
+    bounds [a, b] are applied when sampling the normal distribution with mean=0, std=1.0
+    and the result is subsequently scaled and shifted by the mean and std args.
+
+    Args:
+        tensor: an n-dimensional `torch.Tensor`
+        mean: the mean of the normal distribution
+        std: the standard deviation of the normal distribution
+        a: the minimum cutoff value
+        b: the maximum cutoff value
+    """
+    with torch.no_grad():
+        _trunc_normal_(tensor, 0, 1.0, a, b)
+        tensor.mul_(std).add_(mean)
+
+
+def variance_scaling_(tensor, scale=1.0, mode="fan_in", distribution="normal"):
+    fan_in, fan_out = _calculate_fan_in_and_fan_out(tensor)
+    if mode == "fan_in":
+        denom = fan_in
+    elif mode == "fan_out":
+        denom = fan_out
+    elif mode == "fan_avg":
+        denom = (fan_in + fan_out) / 2
+
+    variance = scale / denom
+
+    if distribution == "truncated_normal":
+        # constant is stddev of standard normal truncated to (-2, 2)
+        trunc_normal_tf_(tensor, std=math.sqrt(variance) / 0.87962566103423978)
+    elif distribution == "normal":
+        with torch.no_grad():
+            tensor.normal_(std=math.sqrt(variance))
+    elif distribution == "uniform":
+        bound = math.sqrt(3 * variance)
+        with torch.no_grad():
+            tensor.uniform_(-bound, bound)
+    else:
+        raise ValueError(f"invalid distribution {distribution}")
+
+
+def lecun_normal_(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="truncated_normal")
+
+
+def default_flax_embed_init(tensor):
+    variance_scaling_(tensor, mode="fan_in", distribution="normal")
+
+
+# Copied from transformers.models.llama.modeling_llama.rotate_half
+def rotate_half(x):
+    """Rotates half the hidden dims of the input."""
+    x1 = x[..., : x.shape[-1] // 2]
+    x2 = x[..., x.shape[-1] // 2 :]
+    return torch.cat((-x2, x1), dim=-1)
+
+
+# def apply_rotary_pos_emb_vision(tensor: torch.Tensor, freqs: torch.Tensor) -> torch.Tensor:
+#     orig_dtype = tensor.dtype
+#     tensor = tensor.float()
+#     cos = freqs.cos()
+#     sin = freqs.sin()
+#     cos = cos.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+#     sin = sin.unsqueeze(1).repeat(1, 1, 2).unsqueeze(0).float()
+#     output = (tensor * cos) + (rotate_half(tensor) * sin)
+#     output = output.to(orig_dtype)
+#     return output
+
+
+def apply_rotary_pos_emb_vision(q, k, cos, sin) -> torch.Tensor:
+    orig_dtype = q.dtype
+    q, k = q.float(), k.float()
+    cos = cos.unsqueeze(1).float()
+    sin = sin.unsqueeze(1).float()
+    q = (q * cos) + (rotate_half(q) * sin)
+    k = (k * cos) + (rotate_half(k) * sin)
+    return q.to(orig_dtype), k.to(orig_dtype)
+
+
+class VisionRotaryEmbedding(nn.Module):
+
+    def __init__(self, dim: int, theta: float = 10000.0) -> None:
+        super().__init__()
+        inv_freq = 1.0 / (theta ** (torch.arange(0, dim, 2, dtype=torch.float) / dim))
+        self.register_buffer("inv_freq", inv_freq, persistent=False)
+
+    def forward(self, seqlen: int) -> torch.Tensor:
+        seq = torch.arange(seqlen, device=self.inv_freq.device, dtype=self.inv_freq.dtype)
+        freqs = torch.outer(seq, self.inv_freq)
+        return freqs
+    
+
+class Videollama3VisionEmbeddings(nn.Module):
+
+    def __init__(self, config: Videollama3VisionEncoderConfig):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.patch_size = config.patch_size
+
+        self.patch_embedding = nn.Conv2d(
+            in_channels=config.num_channels,
+            out_channels=self.embed_dim,
+            kernel_size=self.patch_size,
+            stride=self.patch_size,
+            padding="valid",
+        )
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = hidden_states.view(
+            -1, self.config.num_channels, self.patch_size, self.patch_size
+        )
+        patch_embeds = self.patch_embedding(hidden_states)  # shape = [*, width, grid, grid]
+        # embeddings = patch_embeds.flatten(2).transpose(1, 2)
+        embeddings = patch_embeds.view(-1, self.embed_dim)
+
+        return embeddings
+
+
+class VisionAttention(nn.Module):
+    """Multi-headed attention from 'Attention Is All You Need' paper"""
+
+    # Copied from transformers.models.clip.modeling_clip.CLIPAttention.__init__
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.embed_dim = config.hidden_size
+        self.num_heads = config.num_attention_heads
+        self.head_dim = self.embed_dim // self.num_heads
+        if self.head_dim * self.num_heads != self.embed_dim:
+            raise ValueError(
+                f"embed_dim must be divisible by num_heads (got `embed_dim`: {self.embed_dim} and `num_heads`:"
+                f" {self.num_heads})."
+            )
+        self.scale = self.head_dim**-0.5
+        self.dropout = config.attention_dropout
+
+        self.k_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.v_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.q_proj = nn.Linear(self.embed_dim, self.embed_dim)
+        self.out_proj = nn.Linear(self.embed_dim, self.embed_dim)
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        """Input shape: Time x Channel"""
+
+        q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(q_len, self.num_heads, self.head_dim)
+        key_states = key_states.view(q_len, self.num_heads, self.head_dim)
+        value_states = value_states.view(q_len, self.num_heads, self.head_dim)
+
+        query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+        attention_mask = torch.zeros([1, q_len, q_len], device=query_states.device, dtype=torch.bool)
+        for i in range(1, len(cu_seqlens)):
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+
+        query_states = query_states.transpose(0, 1)
+        key_states = key_states.transpose(0, 1)
+        value_states = value_states.transpose(0, 1)
+
+        attn_weights = torch.matmul(query_states, key_states.transpose(1, 2)) / math.sqrt(self.head_dim)
+        attn_weights = attn_weights + attention_mask
+
+        # upcast attention to fp32
+        attn_weights = nn.functional.softmax(attn_weights, dim=-1, dtype=torch.float32).to(query_states.dtype)
+        attn_weights = nn.functional.dropout(attn_weights, p=self.dropout, training=self.training)
+        attn_output = torch.matmul(attn_weights, value_states)
+
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.reshape(q_len, -1)
+        attn_output = self.out_proj(attn_output)
+
+        return attn_output
+
+
+class VisionFlashAttention2(VisionAttention):
+
+    def __init__(self, *args, **kwargs):
+        super().__init__(*args, **kwargs)
+
+    # Adapted from transformers.models.llama.modeling_llama.LlamaFlashAttention2.forward
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        q_len, _ = hidden_states.size()
+
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        # Flash attention requires the input to have the shape
+        # batch_size x seq_length x head_dim x hidden_dim
+        # therefore we just need to keep the original shape
+        query_states = query_states.view(1, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        key_states = key_states.view(1, q_len, self.num_heads, self.head_dim).transpose(1, 2)
+        value_states = value_states.view(q_len, self.num_heads, self.head_dim)
+        # query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        # key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        query_states, key_states = apply_rotary_pos_emb_vision(
+            query_states,
+            key_states,
+            rotary_pos_emb.cos().unsqueeze(0).repeat(1, 1, 2),
+            rotary_pos_emb.sin().unsqueeze(0).repeat(1, 1, 2),
+        )
+        query_states = query_states.transpose(1, 2).squeeze(0)
+        key_states = key_states.transpose(1, 2).squeeze(0)
+
+        max_seqlen = (cu_seqlens[1:] - cu_seqlens[:-1]).max().item()
+        attn_output = flash_attn_varlen_func(query_states, key_states, value_states, cu_seqlens, cu_seqlens, max_seqlen, max_seqlen).reshape(
+            q_len, -1
+        )
+        attn_output = self.out_proj(attn_output)
+        
+        return attn_output
+
+
+class VisionSdpaAttention(VisionAttention):
+
+    def forward(
+        self,
+        hidden_states: torch.Tensor,
+        cu_seqlens: torch.Tensor,
+        rotary_pos_emb: torch.Tensor = None,
+    ) -> torch.Tensor:
+        seq_length = hidden_states.shape[0]
+        query_states = self.q_proj(hidden_states)
+        key_states = self.k_proj(hidden_states)
+        value_states = self.v_proj(hidden_states)
+
+        query_states = query_states.view(seq_length, self.num_heads, self.head_dim)
+        key_states = key_states.view(seq_length, self.num_heads, self.head_dim)
+        value_states = value_states.view(seq_length, self.num_heads, self.head_dim)
+
+        query_states = apply_rotary_pos_emb_vision(query_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+        key_states = apply_rotary_pos_emb_vision(key_states.unsqueeze(0), rotary_pos_emb).squeeze(0)
+
+        attention_mask = torch.zeros([1, seq_length, seq_length], device=query_states.device, dtype=torch.bool)
+        for i in range(1, len(cu_seqlens)):
+            attention_mask[..., cu_seqlens[i - 1] : cu_seqlens[i], cu_seqlens[i - 1] : cu_seqlens[i]] = True
+
+        query_states = query_states.transpose(0, 1)
+        key_states = key_states.transpose(0, 1)
+        value_states = value_states.transpose(0, 1)
+        attn_output = F.scaled_dot_product_attention(query_states, key_states, value_states, attention_mask, dropout_p=0.0)
+        attn_output = attn_output.transpose(0, 1)
+        attn_output = attn_output.reshape(seq_length, -1)
+        attn_output = self.out_proj(attn_output)
+        return attn_output
+
+
+VISION_ATTENTION_CLASSES = {
+    "eager": VisionAttention,
+    "flash_attention_2": VisionFlashAttention2,
+    "sdpa": VisionSdpaAttention,
+}
+
+
+# Copied from transformers.models.clip.modeling_clip.CLIPMLP with CLIP->Videollama3
+class Videollama3VisionMLP(nn.Module):
+
+    def __init__(self, config):
+        super().__init__()
+        self.config = config
+        self.activation_fn = ACT2FN[config.hidden_act]
+        self.fc1 = nn.Linear(config.hidden_size, config.intermediate_size)
+        self.fc2 = nn.Linear(config.intermediate_size, config.hidden_size)
+
+    def forward(self, hidden_states: torch.Tensor) -> torch.Tensor:
+        hidden_states = self.fc1(hidden_states)
+        hidden_states = self.activation_fn(hidden_states)
+        hidden_states = self.fc2(hidden_states)
+        return hidden_states
+
+
+class Videollama3VisionEncoderLayer(nn.Module):
+
+    def __init__(self, config: Videollama3VisionEncoderConfig):
+        super().__init__()
+        self.embed_dim = config.hidden_size
+        self.self_attn = VISION_ATTENTION_CLASSES[config._attn_implementation](config=config)
+        self.layer_norm1 = LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+        self.mlp = Videollama3VisionMLP(config)
+        self.layer_norm2 = LayerNorm(self.embed_dim, eps=config.layer_norm_eps)
+
+    # Ignore copy
+    def forward(self, hidden_states, cu_seqlens, rotary_pos_emb) -> torch.Tensor:
+        hidden_states = hidden_states + self.self_attn(
+            self.layer_norm1(hidden_states), cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb
+        )
+        hidden_states = hidden_states + self.mlp(self.layer_norm2(hidden_states))
+        return hidden_states
+
+
+class Videollama3VisionTransformerEncoder(nn.Module):
+
+    def __init__(self, config: Videollama3VisionEncoderConfig):
+        super().__init__()
+        self.config = config
+        head_dim = config.hidden_size // config.num_attention_heads
+        self.rotary_pos_emb = VisionRotaryEmbedding(head_dim // 2)
+        self.layers = nn.ModuleList([Videollama3VisionEncoderLayer(config) for _ in range(config.num_hidden_layers)])
+        self.gradient_checkpointing = False
+
+    def rot_pos_emb(self, grid_sizes, merge_sizes):
+        pos_ids = []
+        for (t, h, w), merge_size in zip(grid_sizes, merge_sizes):
+            hpos_ids = torch.arange(h).unsqueeze(1).expand(-1, w)
+            hpos_ids = hpos_ids.reshape(
+                h // merge_size,
+                merge_size,
+                w // merge_size,
+                merge_size,
+            )
+            hpos_ids = hpos_ids.permute(0, 2, 1, 3)
+            hpos_ids = hpos_ids.flatten()
+
+            wpos_ids = torch.arange(w).unsqueeze(0).expand(h, -1)
+            wpos_ids = wpos_ids.reshape(
+                h // merge_size,
+                merge_size,
+                w // merge_size,
+                merge_size,
+            )
+            wpos_ids = wpos_ids.permute(0, 2, 1, 3)
+            wpos_ids = wpos_ids.flatten()
+            pos_ids.append(torch.stack([hpos_ids, wpos_ids], dim=-1).repeat(t, 1))
+
+        pos_ids = torch.cat(pos_ids, dim=0)
+        max_grid_size = grid_sizes[:, 1:].max()
+        rotary_pos_emb_full = self.rotary_pos_emb(max_grid_size)
+        rotary_pos_emb = rotary_pos_emb_full[pos_ids].flatten(1)
+
+        return rotary_pos_emb
+
+    def forward(self, hidden_states, grid_sizes, merge_sizes) -> torch.Tensor:
+        rotary_pos_emb = self.rot_pos_emb(grid_sizes, merge_sizes)
+
+        cu_seqlens = torch.repeat_interleave(grid_sizes[:, 1] * grid_sizes[:, 2], grid_sizes[:, 0]).cumsum(dim=0, dtype=torch.int32)
+        cu_seqlens = F.pad(cu_seqlens, (1, 0), value=0)
+
+        for blk in self.layers:
+            if self.gradient_checkpointing and self.training:
+                hidden_states = self._gradient_checkpointing_func(
+                    blk.__call__,
+                    hidden_states,
+                    cu_seqlens,
+                    rotary_pos_emb
+                )
+            else:
+                hidden_states = blk(hidden_states, cu_seqlens=cu_seqlens, rotary_pos_emb=rotary_pos_emb)
+
+        return hidden_states
+
+
+class Videollama3VisionEncoderModel(PreTrainedModel):
+
+    config_class = Videollama3VisionEncoderConfig
+    base_model_prefix = "videollama3"
+    main_input_name = "pixel_values"
+    supports_gradient_checkpointing = True
+    _no_split_modules = [
+        "Videollama3VisionEncoderLayer",
+        "Videollama3VisionEmbeddings",
+    ]
+    _supports_flash_attn_2 = True
+    _supports_sdpa = True
+
+    def __init__(self, config: Videollama3VisionEncoderConfig):
+        super().__init__(config=config)
+        embed_dim = config.hidden_size
+
+        self.embeddings = Videollama3VisionEmbeddings(config)
+        self.encoder = Videollama3VisionTransformerEncoder(config)
+        self.post_layernorm = LayerNorm(embed_dim, eps=config.layer_norm_eps)
+
+        self.post_init()
+
+    def forward(self, pixel_values, grid_sizes, merge_sizes=None):
+        hidden_states = self.embeddings(pixel_values)
+        hidden_states = self.encoder(hidden_states, grid_sizes, merge_sizes)
+        hidden_states = self.post_layernorm(hidden_states)
+        hidden_states_raw = hidden_states.clone()
+
+        hidden_states_chunks = hidden_states.split(grid_sizes.prod(dim=1).tolist(), dim=0)
+        outputs = []
+
+        for hidden_states, grid_size, merge_size in zip(hidden_states_chunks, grid_sizes, merge_sizes):
+            # NOTE: previous implementation, which supports downsampling with any factor
+            c = hidden_states.shape[-1]
+            hidden_states = hidden_states.view(
+                grid_size[0], grid_size[1] // merge_size, grid_size[2] // merge_size, merge_size, merge_size,  c
+            ).permute(0, 1, 3, 2, 4, 5)
+            hidden_states = hidden_states.reshape(
+                grid_size[0], grid_size[1], grid_size[2], c
+            ).permute(0, 3, 1, 2)
+            hidden_states = torch.nn.functional.interpolate(
+                hidden_states,
+                size=(grid_size[1] // merge_size, grid_size[2] // merge_size),
+                mode='bilinear'
+            )
+            hidden_states = hidden_states.permute(0, 2, 3, 1).view(-1, c)
+
+            # NOTE: simplified implementation, which only supports downsampling with integer factor
+            # NOTE: this implementation is mathematically equivalent to the previous one when merge_size is 1 or 2 but may cause slightly different results
+            # hidden_states = hidden_states.view(-1, merge_size * merge_size, hidden_states.size(-1))
+            # hidden_states = hidden_states.mean(dim=1)
+
+            outputs.append(hidden_states)
+
+        return torch.cat(outputs, dim=0), hidden_states_raw
+
+    def _init_weights(self, module):
+        """Initialize the weights"""
+        if isinstance(module, nn.Embedding):
+            default_flax_embed_init(module.weight)
+        elif isinstance(module, VisionAttention):
+            nn.init.xavier_uniform_(module.q_proj.weight)
+            nn.init.xavier_uniform_(module.k_proj.weight)
+            nn.init.xavier_uniform_(module.v_proj.weight)
+            nn.init.xavier_uniform_(module.out_proj.weight)
+            nn.init.zeros_(module.q_proj.bias)
+            nn.init.zeros_(module.k_proj.bias)
+            nn.init.zeros_(module.v_proj.bias)
+            nn.init.zeros_(module.out_proj.bias)
+        elif isinstance(module, Videollama3VisionMLP):
+            nn.init.xavier_uniform_(module.fc1.weight)
+            nn.init.xavier_uniform_(module.fc2.weight)
+            nn.init.normal_(module.fc1.bias, std=1e-6)
+            nn.init.normal_(module.fc2.bias, std=1e-6)
+        elif isinstance(module, (nn.Linear, nn.Conv2d)):
+            lecun_normal_(module.weight)
+            if module.bias is not None:
+                nn.init.zeros_(module.bias)
+        elif isinstance(module, LayerNorm):
+            module.bias.data.zero_()
+            module.weight.data.fill_(1.0)
diff --git a/rynnec/rynnec_trainer.py b/rynnec/rynnec_trainer.py
new file mode 100644
index 0000000000000000000000000000000000000000..d6dfd10c4b037299d56e8db6f954d8b07bcc3820
--- /dev/null
+++ b/rynnec/rynnec_trainer.py
@@ -0,0 +1,496 @@
+# Adopted from: https://github.com/haotian-liu/LLaVA/blob/main/llava/train/llava_trainer.py
+import os
+import logging
+from typing import TYPE_CHECKING, Any, Callable, Dict, List, Optional, Tuple, Union
+
+import torch
+import torch.nn as nn
+from torch.utils.data import Sampler
+
+from transformers import Trainer
+from transformers.trainer import (
+    is_sagemaker_mp_enabled,
+    get_parameter_names,
+    has_length,
+    ALL_LAYERNORM_LAYERS,
+    logger,
+    TRAINER_STATE_NAME,
+)
+from transformers.utils import (
+    is_sagemaker_mp_enabled,
+    logging,
+)
+
+def maybe_zero_3(param, ignore_status=False, name=None):
+    from deepspeed import zero
+    from deepspeed.runtime.zero.partition_parameters import ZeroParamStatus
+    if hasattr(param, "ds_id"):
+        if param.ds_status == ZeroParamStatus.NOT_AVAILABLE:
+            if not ignore_status:
+                logging.warning(f"{name}: param.ds_status != ZeroParamStatus.NOT_AVAILABLE: {param.ds_status}")
+        with zero.GatheredParameters([param]):
+            param = param.data.detach().cpu().clone()
+    else:
+        param = param.detach().cpu().clone()
+    return param
+
+
+def get_mm_adapter_state_maybe_zero_3(named_params, keys_to_match):
+    to_return = {k: t for k, t in named_params if any(key_match in k for key_match in keys_to_match)}
+    to_return = {k: maybe_zero_3(v, ignore_status=True, name=k).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+# Borrowed from peft.utils.get_peft_model_state_dict
+def get_peft_state_maybe_zero_3(named_params, bias):
+    if bias == "none":
+        to_return = {k: t for k, t in named_params if "lora_" in k}
+    elif bias == "all":
+        to_return = {k: t for k, t in named_params if "lora_" in k or "bias" in k}
+    elif bias == "lora_only":
+        to_return = {}
+        maybe_lora_bias = {}
+        lora_bias_names = set()
+        for k, t in named_params:
+            if "lora_" in k:
+                to_return[k] = t
+                bias_name = k.split("lora_")[0] + "bias"
+                lora_bias_names.add(bias_name)
+            elif "bias" in k:
+                maybe_lora_bias[k] = t
+        for k, t in maybe_lora_bias:
+            if bias_name in lora_bias_names:
+                to_return[bias_name] = t
+    else:
+        raise NotImplementedError
+    to_return = {k: maybe_zero_3(v, ignore_status=True) for k, v in to_return.items()}
+    return to_return
+
+
+def get_peft_state_non_lora_maybe_zero_3(named_params, require_grad_only=True):
+    to_return = {k: t for k, t in named_params if "lora_" not in k}
+    if require_grad_only:
+        to_return = {k: t for k, t in to_return.items() if t.requires_grad}
+    to_return = {k: maybe_zero_3(v, ignore_status=True).cpu() for k, v in to_return.items()}
+    return to_return
+
+
+def find_all_linear_names(model):
+    cls = torch.nn.Linear
+    lora_module_names = set()
+    multimodal_keywords = ['mm_projector', 'vision_encoder', 'vision_resampler', 'text_hidden_fcs', 'region_encoder', 'grounding_encoder']
+    for name, module in model.named_modules():
+        if any(mm_keyword in name for mm_keyword in multimodal_keywords):
+            continue
+        if isinstance(module, cls):
+            if 'lm_head' in name:
+                continue
+            lora_module_names.add(name)
+    
+    return list(lora_module_names)
+
+def safe_save_model_for_hf_trainer(trainer: Trainer,
+                                   output_dir: str):
+    """Collects the state dict and dump to disk."""
+
+    if getattr(trainer.args, "is_alignment", False):
+        # Only save Adapter
+        keys_to_match = ['mm_projector']
+
+        weight_to_save = get_mm_adapter_state_maybe_zero_3(trainer.model.named_parameters(), keys_to_match)
+        trainer.model.config.save_pretrained(output_dir)
+
+        current_folder = output_dir.split('/')[-1]
+        parent_folder = os.path.dirname(output_dir)
+        # if trainer.args.local_rank == 0 or trainer.args.local_rank == -1:
+        if torch.distributed.get_rank() == 0:
+            if current_folder.startswith('checkpoint-'):
+                mm_projector_folder = os.path.join(parent_folder, "mm_projector")
+                os.makedirs(mm_projector_folder, exist_ok=True)
+                torch.save(weight_to_save, os.path.join(mm_projector_folder, f'{current_folder}.bin'))
+            else:
+                torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
+        return
+
+    if trainer.deepspeed:
+        torch.cuda.synchronize()
+        trainer.save_model(output_dir)
+        return
+
+    state_dict = trainer.model.state_dict()
+    if trainer.args.should_save:
+        cpu_state_dict = {
+            key: value.cpu()
+            for key, value in state_dict.items()
+        }
+        del state_dict
+        trainer._save(output_dir, state_dict=cpu_state_dict)  # noqa
+
+
+def split_to_even_chunks(indices, lengths, num_chunks):
+    """
+    Split a list of indices into `chunks` chunks of roughly equal lengths.
+    """
+
+    if len(indices) % num_chunks != 0:
+        return [indices[i::num_chunks] for i in range(num_chunks)]
+
+    num_indices_per_chunk = len(indices) // num_chunks
+
+    chunks = [[] for _ in range(num_chunks)]
+    chunks_lengths = [0 for _ in range(num_chunks)]
+    for index in indices:
+        shortest_chunk = chunks_lengths.index(min(chunks_lengths))
+        chunks[shortest_chunk].append(index)
+        chunks_lengths[shortest_chunk] += lengths[index]
+        if len(chunks[shortest_chunk]) == num_indices_per_chunk:
+            chunks_lengths[shortest_chunk] = float("inf")
+
+    return chunks
+
+
+def get_modality_length_grouped_indices(lengths, batch_size, world_size, generator=None):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    assert all(l != 0 for l in lengths), "Should not have zero length."
+    if all(l > 0 for l in lengths) or all(l < 0 for l in lengths):
+        # all samples are in the same modality
+        return get_length_grouped_indices(lengths, batch_size, world_size, generator=generator)
+    mm_indices, mm_lengths = zip(*[(i, l) for i, l in enumerate(lengths) if l > 0])
+    lang_indices, lang_lengths = zip(*[(i, -l) for i, l in enumerate(lengths) if l < 0])
+
+    mm_shuffle = [mm_indices[i] for i in get_length_grouped_indices(mm_lengths, batch_size, world_size, generator=None)]
+    lang_shuffle = [lang_indices[i] for i in get_length_grouped_indices(lang_lengths, batch_size, world_size, generator=None)]
+    megabatch_size = world_size * batch_size
+    mm_megabatches = [mm_shuffle[i : i + megabatch_size] for i in range(0, len(mm_shuffle), megabatch_size)]
+    lang_megabatches = [lang_shuffle[i : i + megabatch_size] for i in range(0, len(lang_shuffle), megabatch_size)]
+
+    # last_mm = mm_megabatches[-1]
+    # last_lang = lang_megabatches[-1]
+    # additional_batch = last_mm + last_lang
+    megabatches = mm_megabatches[:-1] + lang_megabatches[:-1]
+    megabatch_indices = torch.randperm(len(megabatches), generator=generator)
+    megabatches = [megabatches[i] for i in megabatch_indices]
+
+    # if len(additional_batch) > 0:
+    #     megabatches.append(sorted(additional_batch))
+
+    return [i for megabatch in megabatches for i in megabatch]
+
+
+def get_length_grouped_indices(lengths, batch_size, world_size, generator=None, merge=True):
+    # We need to use torch for the random part as a distributed sampler will set the random seed for torch.
+    indices = torch.randperm(len(lengths), generator=generator)
+    megabatch_size = world_size * batch_size
+    megabatches = [indices[i : i + megabatch_size].tolist() for i in range(0, len(lengths), megabatch_size)]
+    megabatches = [sorted(megabatch, key=lambda i: lengths[i], reverse=True) for megabatch in megabatches]
+    megabatches = [split_to_even_chunks(megabatch, lengths, world_size) for megabatch in megabatches]
+
+    return [i for megabatch in megabatches for batch in megabatch for i in batch]
+
+
+class LengthGroupedSampler(Sampler):
+    r"""
+    Sampler that samples indices in a way that groups together features of the dataset of roughly the same length while
+    keeping a bit of randomness.
+    """
+
+    def __init__(
+        self,
+        batch_size: int,
+        world_size: int,
+        lengths: Optional[List[int]] = None,
+        generator=None,
+        group_by_modality: bool = False,
+    ):
+        if lengths is None:
+            raise ValueError("Lengths must be provided.")
+
+        self.batch_size = batch_size
+        self.world_size = world_size
+        self.lengths = lengths
+        self.generator = generator
+        self.group_by_modality = group_by_modality
+
+    def __len__(self):
+        return len(self.lengths)
+
+    def __iter__(self):
+        if self.group_by_modality:
+            indices = get_modality_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        else:
+            indices = get_length_grouped_indices(self.lengths, self.batch_size, self.world_size, generator=self.generator)
+        return iter(indices)
+
+
+class RynnECTrainer(Trainer):
+
+    def _get_train_sampler(self) -> Optional[torch.utils.data.Sampler]:
+        print('_get_train_sampler')
+        print('world size: ', self.args.world_size * self.args.gradient_accumulation_steps)
+        if self.train_dataset is None or not has_length(self.train_dataset):
+            return None
+        print('group_by_modality_length...')
+        if self.args.group_by_modality_length:
+            lengths = self.train_dataset.modality_lengths
+            return LengthGroupedSampler(
+                self.args.train_batch_size,
+                world_size=self.args.world_size * self.args.gradient_accumulation_steps,
+                lengths=lengths,
+                group_by_modality=True,
+            )
+        else:
+            return super()._get_train_sampler()
+
+    def update_history_loss_dict(self,outputs):
+        if not hasattr(self,'history_loss_dict'):
+            self.history_loss_dict = {}
+        for name, value in outputs.items():
+            if 'loss' in name and name != 'loss':
+                if name not in self.history_loss_dict:
+                    self.history_loss_dict[name] = value.item()
+                else:
+                    if value != 0:
+                        self.history_loss_dict[name] = value.item()
+    def compute_loss(self, model, inputs, return_outputs=False, num_items_in_batch=None):
+        """
+        How the loss is computed by Trainer. By default, all models return the loss in the first element.
+
+        Subclass and override for custom behavior.
+        """
+        if (self.label_smoother is not None or self.compute_loss_func is not None) and "labels" in inputs:
+            labels = inputs.pop("labels")
+        else:
+            labels = None
+        if self.model_accepts_loss_kwargs:
+            loss_kwargs = {}
+            if num_items_in_batch is not None:
+                loss_kwargs["num_items_in_batch"] = num_items_in_batch
+            inputs = {**inputs, **loss_kwargs}
+        outputs = model(**inputs)
+        # Save past state if it exists
+        # TODO: this needs to be fixed and made cleaner later.
+        if self.args.past_index >= 0:
+            self._past = outputs[self.args.past_index]
+
+        if labels is not None:
+            unwrapped_model = self.accelerator.unwrap_model(model)
+            if _is_peft_model(unwrapped_model):
+                model_name = unwrapped_model.base_model.model._get_name()
+            else:
+                model_name = unwrapped_model._get_name()
+            # User-defined compute_loss function
+            if self.compute_loss_func is not None:
+                loss = self.compute_loss_func(outputs, labels, num_items_in_batch=num_items_in_batch)
+            elif model_name in MODEL_FOR_CAUSAL_LM_MAPPING_NAMES.values():
+                loss = self.label_smoother(outputs, labels, shift_labels=True)
+            else:
+                loss = self.label_smoother(outputs, labels)
+        else:
+            if isinstance(outputs, dict) and "loss" not in outputs:
+                raise ValueError(
+                    "The model did not return a loss from the inputs, only the following keys: "
+                    f"{','.join(outputs.keys())}. For reference, the inputs it received are {','.join(inputs.keys())}."
+                )
+            # We don't use .loss here since the model may return tuples instead of ModelOutput.
+            loss = outputs["loss"] if isinstance(outputs, dict) else outputs[0]
+            if isinstance(outputs, dict) and 'mask_bce_loss' in outputs:
+                loss_dict = {}
+                for name,value in outputs.items():
+                    if 'loss' in name and name != 'loss':
+                        loss_value = value.item()
+                        if loss_value == 0 and hasattr(self,'history_loss_dict'):
+                            loss_value = self.history_loss_dict[name]
+                        loss_dict[name] = loss_value
+                self.update_history_loss_dict(outputs)
+                self.log(loss_dict)
+
+        if self.args.average_tokens_across_devices and self.model_accepts_loss_kwargs:
+            loss *= self.accelerator.num_processes
+
+        return (loss, outputs) if return_outputs else loss
+
+
+    def create_optimizer(self):
+        """
+        Setup the optimizer.
+
+        We provide a reasonable default that works well. If you want to use something else, you can pass a tuple in the
+        Trainer's init through `optimizers`, or subclass and override this method in a subclass.
+        """
+        if is_sagemaker_mp_enabled():
+            return super().create_optimizer()
+
+        opt_model = self.model
+
+        if self.optimizer is None:
+            optimized_parameters = [(n, p) for n, p in opt_model.named_parameters() if p.requires_grad]
+            optimizer_grouped_parameters = []
+
+            decay_parameters = get_parameter_names(opt_model, ALL_LAYERNORM_LAYERS)
+            decay_parameters = [name for name in decay_parameters if "bias" not in name]
+
+            if self.args.llm_lr is not None:
+                lm_parameters = [
+                    name for name, _ in optimized_parameters if "vision_encoder" not in name and "mm_projector" not in name and "region_encoder" not in name and "grounding_encoder" not in name
+                ]
+                decay_lm_parameters = [name for name in lm_parameters if name in decay_parameters]
+                nodecay_lm_parameters = [name for name in lm_parameters if name not in decay_parameters]
+                optimizer_grouped_parameters.extend([
+                    {
+                        "params": [p for n, p in optimized_parameters if n in decay_lm_parameters],
+                        "weight_decay": self.args.weight_decay,
+                        "lr": self.args.llm_lr,
+                    },
+                    {
+                        "params": [p for n, p in optimized_parameters if n in nodecay_lm_parameters],
+                        "weight_decay": 0.0,
+                        "lr": self.args.llm_lr,
+                    }
+                ])
+
+            if self.args.mm_projector_lr is not None:
+                projector_parameters = [name for name, _ in optimized_parameters if "mm_projector" in name]
+                decay_projector_parameters = [name for name in projector_parameters if name in decay_parameters]
+                nodecay_projector_parameters = [name for name in projector_parameters if name not in decay_parameters]
+                optimizer_grouped_parameters.extend([
+                    {
+                        "params": [p for n, p in optimized_parameters if n in decay_projector_parameters], 
+                        "weight_decay": self.args.weight_decay,
+                        "lr": self.args.mm_projector_lr,
+                    },
+                    {
+                        "params": [p for n, p in optimized_parameters if n in nodecay_projector_parameters],
+                        "weight_decay": 0.0,
+                        "lr": self.args.mm_projector_lr,
+                    }
+                ])
+
+            if self.args.vision_encoder_lr is not None:
+                vision_encoder_parameters = [name for name, _ in optimized_parameters if "vision_encoder" in name]
+                decay_vision_encoder_parameters = [name for name in vision_encoder_parameters if name in decay_parameters]
+                nodecay_vision_encoder_parameters = [name for name in vision_encoder_parameters if name not in decay_parameters]
+                optimizer_grouped_parameters.extend([
+                    {
+                        "params": [p for n, p in optimized_parameters if n in decay_vision_encoder_parameters], 
+                        "weight_decay": self.args.weight_decay,
+                        "lr": self.args.vision_encoder_lr,
+                    },
+                    {
+                        "params": [p for n, p in optimized_parameters if n in nodecay_vision_encoder_parameters],
+                        "weight_decay": 0.0,
+                        "lr": self.args.vision_encoder_lr,
+                    }
+                ])
+
+            if self.args.region_encoder_lr is not None:
+                projector_parameters = [name for name, _ in optimized_parameters if "region_encoder" in name]
+                decay_projector_parameters = [name for name in projector_parameters if name in decay_parameters]
+                nodecay_projector_parameters = [name for name in projector_parameters if name not in decay_parameters]
+                optimizer_grouped_parameters.extend([
+                    {
+                        "params": [p for n, p in optimized_parameters if n in decay_projector_parameters], 
+                        "weight_decay": self.args.weight_decay,
+                        "lr": self.args.region_encoder_lr,
+                    },
+                    {
+                        "params": [p for n, p in optimized_parameters if n in nodecay_projector_parameters],
+                        "weight_decay": 0.0,
+                        "lr": self.args.region_encoder_lr,
+                    }
+                ])
+            if self.args.sam_decoder_lr is not None:
+                projector_parameters = [name for name, _ in optimized_parameters if "grounding_encoder" in name and "image_encoder" not in name]
+                decay_projector_parameters = [name for name in projector_parameters if name in decay_parameters]
+                nodecay_projector_parameters = [name for name in projector_parameters if name not in decay_parameters]
+                optimizer_grouped_parameters.extend([
+                    {
+                        "params": [p for n, p in optimized_parameters if n in decay_projector_parameters], 
+                        "weight_decay": self.args.weight_decay,
+                        "lr": self.args.sam_decoder_lr,
+                    },
+                    {
+                        "params": [p for n, p in optimized_parameters if n in nodecay_projector_parameters],
+                        "weight_decay": 0.0,
+                        "lr": self.args.sam_decoder_lr,
+                    }
+                ])
+
+            optimizer_cls, optimizer_kwargs = Trainer.get_optimizer_cls_and_kwargs(self.args)
+
+            self.optimizer = optimizer_cls(optimizer_grouped_parameters, **optimizer_kwargs)
+            if optimizer_cls.__name__ == "Adam8bit":
+                import bitsandbytes
+
+                manager = bitsandbytes.optim.GlobalOptimManager.get_instance()
+
+                skipped = 0
+                for module in opt_model.modules():
+                    if isinstance(module, nn.Embedding):
+                        skipped += sum({p.data_ptr(): p.numel() for p in module.parameters()}.values())
+                        logger.info(f"skipped {module}: {skipped/2**20}M params")
+                        manager.register_module_override(module, "weight", {"optim_bits": 32})
+                        logger.debug(f"bitsandbytes: will optimize {module} in fp32")
+                logger.info(f"skipped: {skipped/2**20}M params")
+
+        return self.optimizer
+
+    def _save_checkpoint(self, model, trial, metrics=None):
+        if getattr(self.args, 'is_alignment', False):
+            from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+            checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
+
+            run_dir = self._get_output_dir(trial=trial)
+            output_dir = os.path.join(run_dir, checkpoint_folder)
+
+            # Only save Adapter
+            keys_to_match = ['mm_projector', 'vision_resampler']
+
+            weight_to_save = get_mm_adapter_state_maybe_zero_3(self.model.named_parameters(), keys_to_match)
+
+            if self.args.local_rank == 0 or self.args.local_rank == -1:
+                self.model.config.save_pretrained(output_dir)
+                torch.save(weight_to_save, os.path.join(output_dir, f'mm_projector.bin'))
+            # Save optimizer and scheduler
+            self._save_optimizer_and_scheduler(output_dir)
+            # Save RNG state
+            self._save_rng_state(output_dir)
+            self.state.save_to_json(os.path.join(output_dir, TRAINER_STATE_NAME))
+            self.args.distributed_state.wait_for_everyone()
+        else:
+            # NOTE: Supporting save complete lora checkpoint during training.
+            if self.args.lora_enable:
+                from transformers.trainer_utils import PREFIX_CHECKPOINT_DIR
+                checkpoint_folder = f"{PREFIX_CHECKPOINT_DIR}-{self.state.global_step}"
+
+                run_dir = self._get_output_dir(trial=trial)
+                output_dir = os.path.join(run_dir, checkpoint_folder)
+
+                state_dict = get_peft_state_maybe_zero_3(self.model.named_parameters(), self.args.lora_bias)
+                non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(self.model.named_parameters())
+           
+                # add for qwen2
+                if hasattr(self.model, 'base_model') and hasattr(self.model.base_model, 'lm_head'):
+                    lm_head_weight = self.model.base_model.lm_head.weight.cpu() 
+                    non_lora_state_dict['base_model.lm_head.weight'] = lm_head_weight
+                    print("add base_model.lm_head.weight")
+                else:
+                    print("The model does not have 'base_model.lm_head.weight' attribute.")
+
+
+                if self.args.local_rank == 0 or self.args.local_rank == -1:
+                    # save for acquring `config.json`
+                    self.model.config.save_pretrained(output_dir)
+                    # save for acquring `adapter_config.json`, `adapter_model.bin`
+                    # self.model.save_pretrained(output_dir, state_dict=state_dict)
+                    torch.save(non_lora_state_dict, os.path.join(output_dir, 'non_lora_trainables.bin'))
+
+                # save for acquring lora adapter parameters & trainer states: `adapter_config.json`, `adapter_model.safetensors`
+                super(RynnECTrainer, self)._save_checkpoint(model, trial, metrics)
+            else:
+                super(RynnECTrainer, self)._save_checkpoint(model, trial, metrics)
+
+    def _save(self, output_dir: Optional[str] = None, state_dict=None):
+        if getattr(self.args, 'is_alignment', False):
+            pass
+        else:
+            super(RynnECTrainer, self)._save(output_dir, state_dict)
diff --git a/rynnec/train.py b/rynnec/train.py
new file mode 100644
index 0000000000000000000000000000000000000000..697320cca59db48131fd0c7c59720e8a58e031fc
--- /dev/null
+++ b/rynnec/train.py
@@ -0,0 +1,832 @@
+# Adopted from https://github.com/DAMO-NLP-SG/VideoLLaMA3. Below is the original copyright:
+# Adopted from https://github.com/haotian-liu/LLaVA. Below is the original copyright:
+# Adopted from https://github.com/lm-sys/FastChat. Below is the original copyright:
+# Adopted from tatsu-lab@stanford_alpaca. Below is the original copyright:
+#    Copyright 2023 Rohan Taori, Ishaan Gulrajani, Tianyi Zhang, Yann Dubois, Xuechen Li
+#
+#    Licensed under the Apache License, Version 2.0 (the "License");
+#    you may not use this file except in compliance with the License.
+#    You may obtain a copy of the License at
+#
+#        http://www.apache.org/licenses/LICENSE-2.0
+#
+#    Unless required by applicable law or agreed to in writing, software
+#    distributed under the License is distributed on an "AS IS" BASIS,
+#    WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+#    See the License for the specific language governing permissions and
+#    limitations under the License.
+
+import math
+import copy
+import json
+import os
+import pathlib
+import random
+import re
+import sys
+import warnings
+import traceback
+from packaging import version
+from dataclasses import dataclass, field
+from typing import Dict, List, Optional, Sequence
+import numpy as np
+import pyarrow as pa
+
+# torch-related packages
+# NOTE: torch must be imported before transformers. Otherwise, `Segmentation fault (core dumped)` will occur.
+import torch
+import transformers
+from packaging import version
+import datasets
+from datasets import load_dataset, concatenate_datasets
+from torch.utils.data import Dataset
+from transformers.models.mixtral.modeling_mixtral import MixtralSparseMoeBlock
+from transformers import logging
+# logging.set_verbosity_error()
+
+sys.path.append('./')
+
+from rynnec.constants import (IGNORE_INDEX, MODAL_INDEX_MAP,
+    NUM_FRAMES, DEFAULT_IMAGE_TOKEN, STREAM_MAX_FRAMES,
+    STREAM_DOWNSAMPLING, STREAM_FPS, STREAM_IMAGE_SIZE,
+    STREAM_START_TOKEN, STREAM_END_TOKEN, REGION_TOKEN, SEG_TOKEN, REGION_TOKEN_REPLACE)
+from rynnec.mm_utils import (load_images, load_video, DirectResize, load_video_from_ids,
+                                  tokenizer_multimodal_token, annToMask, sam_preprocess_batch)
+from rynnec.model import *
+from rynnec.rynnec_trainer import (
+    RynnECTrainer, find_all_linear_names, get_peft_state_maybe_zero_3,
+    get_peft_state_non_lora_maybe_zero_3, safe_save_model_for_hf_trainer)
+
+# NOTE: fast tokenizer warning issue: https://github.com/huggingface/transformers/issues/5486
+os.environ["TOKENIZERS_PARALLELISM"] = "true"
+
+local_rank = None
+
+
+def rank0_print(*args):
+    if local_rank == 0:
+        print(*args)
+
+
+def set_seed(seed=42):
+    """
+    Set the random seed for reproducible results.
+
+    :param seed: An integer value to be used as the random seed.
+    """
+    torch.manual_seed(seed)
+    torch.cuda.manual_seed(seed)
+    torch.cuda.manual_seed_all(seed)  # for multi-GPU setups
+    torch.backends.cudnn.deterministic = True
+    torch.backends.cudnn.benchmark = False
+
+
+def int_with_none(value):
+    if value == 'None':
+        return None
+    return int(value)
+
+
+@dataclass
+class ModelArguments:
+    # LLM Arguments
+    model_type: Optional[str] = field(default="rynnec", metadata={"help": "Model type selected in the list: " + ", ".join('rynnec_qwen2')})
+    model_path: Optional[str] = field(default="lmsys/vicuna-7b-v1.5")
+    version: Optional[str] = field(default="v1", metadata={"help": "Version of the conversation template."})
+    freeze_backbone: bool = field(default=False, metadata={"help": "Whether to freeze the LLM backbone."})
+    # Connector Arguments
+    mm_projector_type: Optional[str] = field(default='linear')
+    pretrain_mm_projector: Optional[str] = field(default=None)
+    # Vision tower Arguments
+    vision_encoder: Optional[str] = field(default=None)
+    mm_vision_select_layer: Optional[int] = field(default=-1)
+    mm_vision_select_feature: Optional[str] = field(default="patch")
+    mm_attn_implementation: Optional[str] = field(default="flash_attention_2")
+    # Token downsampling Arguments
+    spatial_merge_size: Optional[int] = field(default=1)
+    mm_max_length: Optional[int] = field(default=10240)
+    use_token_compression: Optional[bool] = field(default=False)
+    mask_decoder_model: Optional[str] = field(default="./checkpoints/sam2_hiera_large.pt")
+    load_sam2_weight: Optional[bool] = field(default=False)
+    training: Optional[bool] = field(default=True)
+    has_mask: Optional[bool] = field(default=True)
+
+@dataclass
+class DataArguments:
+    # Path Arguments
+    data_path: List[str] = field(default=None, metadata={"help": "Path to the training data."})
+    # image_folder: Optional[str] = field(default=None)
+    # video_folder: Optional[str] = field(default=None)
+    data_folder: Optional[str] = field(default=None)
+    # Loading Arguments
+    is_multimodal: bool = False
+    fps: Optional[int] = field(default=None)
+    max_frames: Optional[int_with_none] = field(default=None)
+    # Preprocess Arguments
+    image_aspect_ratio: str = 'square'
+    use_batch_flattening: bool = field(default=False, metadata={"help": "Whether to flatten the in-batch sequences of variable lengths."})
+    dataset_cache_dir: Optional[str] = field(default=None)
+
+
+@dataclass
+class TrainingArguments(transformers.TrainingArguments):
+    # shut auto processing (_remove_unused_columns) of transformers Trainer
+    remove_unused_columns: bool = field(default=False)
+
+    optim: str = field(default="adamw_torch")
+    # Training learning rate Arguments
+    vision_encoder_lr: Optional[float] = None
+    mm_projector_lr: Optional[float] = None
+    llm_lr: Optional[float] = None
+    region_encoder_lr: Optional[float] = None
+    sam_encoder_lr: Optional[float] = None
+    sam_decoder_lr: Optional[float] = None
+    # Training Data Arguments
+    group_by_modality_length: bool = field(default=False)
+    model_max_length: int = field(
+        default=512,
+        metadata={
+            "help":
+            "Maximum sequence length. Sequences will be right padded (and possibly truncated)."
+        },
+    )
+    # Lora or Quant Arguments
+    double_quant: bool = field(
+        default=True,
+        metadata={"help": "Compress the quantization statistics through double quantization."}
+    )
+    quant_type: str = field(
+        default="nf4",
+        metadata={"help": "Quantization data type to use. Should be one of `fp4` or `nf4`."}
+    )
+    bits: int = field(
+        default=16,
+        metadata={"help": "How many bits to use."}
+    )
+    lora_enable: bool = False
+    lora_r: int = 64
+    lora_alpha: int = 16
+    lora_dropout: float = 0.05
+    lora_weight_path: str = ""
+    lora_bias: str = "none"
+
+    use_workload_balancing: bool = field(default=False, metadata={"help": "Whether to use data balancing."})
+    loss_reduction_scope: str = field(default="batch", metadata={"help": "Loss reduction scope."})
+    context_parallel_size: int = field(default=1, metadata={"help": "Context parallel size."})
+    use_liger_kernel: bool = field(default=False, metadata={"help": "Whether to use Liger Kernel."})
+
+
+
+class LazySupervisedDataset(Dataset):
+    """Dataset for supervised fine-tuning."""
+
+    def __init__(self, data_path: str, vlprocessor, data_args: DataArguments):
+        super(LazySupervisedDataset, self).__init__()
+        data_objs = []
+
+        try:
+            for data in data_path:
+                # NOTE: load_dataset can process both json or jsonl files
+                if data.endswith(".json") or data.endswith(".jsonl"):
+                    data_objs.append(load_dataset("json", data_files=data, cache_dir=data_args.dataset_cache_dir)["train"])
+                else:
+                    raise Exception(f"Unsupported file format (<{data}>)!")
+            list_data_dict = concatenate_datasets(data_objs)
+        except:
+            traceback.print_exc()
+            # NOTE: compatible with the old version
+            list_data_dict = []
+            for data in data_path:
+                if data.endswith(".json"):
+                    data = json.load(open(data, "r"))
+                    for i in data:
+                        i['id'] = len(list_data_dict)
+                        list_data_dict.append(i)
+                elif data.endswith(".jsonl"):
+                    with open(data, "r", encoding="utf-8") as fp:
+                        for line in fp:
+                            line = line.strip()
+                            obj = json.loads(line)
+                            obj["id"] = len(list_data_dict)
+                            list_data_dict.append(obj)
+                else:
+                    raise Exception(f"Unsupported file format (<{data}>)!!!")
+
+
+        rank0_print("Formatting inputs...Skip in lazy mode")
+        self.vlprocessor = vlprocessor
+        self.list_data_dict = list_data_dict
+        self.data_args = data_args
+
+        img_size=1024
+        self.img_size = img_size
+        self.sam_transform = DirectResize(img_size)
+
+    def __len__(self):
+        return len(self.list_data_dict)
+
+    @property
+    def lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            img_tokens = 576 if 'image' in sample else 0
+            length_list.append(sum(len(conv['value'].split()) for conv in sample['conversations']) + img_tokens)
+        return length_list
+
+    @property
+    def modality_lengths(self):
+        length_list = []
+        for sample in self.list_data_dict:
+            cur_len = sum(len(conv['value'].split()) for conv in sample['conversations'])
+            if cur_len==0:
+                cur_len = 1
+            cur_len = cur_len if 'masks' in sample and sample['masks'] is not None and ('seg' not in sample or sample['seg'] is None) else -cur_len
+            length_list.append(cur_len)
+        return length_list
+
+    def _convert_normal(self, data_dict):
+        data_folder = self.data_args.data_folder
+        conversation = copy.deepcopy(data_dict["conversations"])
+
+        # data sanity check and repair
+        start_idx = 0
+        for sentence in conversation:
+            if sentence["from"] == "human" or sentence["from"] == "system":
+                break
+            start_idx += 1
+        if start_idx > 0:
+            warnings.warn(f"Find {start_idx} non-user sentences at the beginning of the conversation, remove them automatically!")
+            conversation = conversation[start_idx:]
+        assert len(conversation) > 1, f"Invalid conversation"
+
+        mask_ids = []
+
+        if 'image' in data_dict and data_dict['image'] is not None:
+            modal = 'image'
+            if all(not "<image>" in sentence["value"] for sentence in conversation):
+                warnings.warn(f"Image tag not found in the conversation, add it automatically at the beginning!")
+                conversation[0]["value"] = "<image>" + conversation[0]["value"]
+            image_file = data_dict['image']
+            if isinstance(image_file, list):
+                image_file = [os.path.join(data_folder, f) for f in image_file]
+            else:
+                image_file = os.path.join(data_folder, image_file)
+            images = load_images(image_file)
+            
+            masks = []
+            if 'masks' in data_dict and data_dict['masks'] is not None:
+                if 'height' in data_dict:
+                    h = data_dict['height']
+                    w = data_dict['width']
+                else:
+                    h = None
+                    w = None
+                
+                if isinstance(data_dict['masks'], str):
+                    masks_ = json.load(open(data_dict['masks']))
+                else:
+                    masks_= data_dict['masks']
+                image2maskids = []
+                mask_idx = 0
+                for ann in masks_:
+                    image2maskids_ = []
+                    mask = annToMask(ann, h, w)
+                    masks.append(mask)
+                    mask_ids.append(0)
+                    image2maskids_.append(mask_idx)
+                    mask_idx += 1
+                    image2maskids.append(image2maskids_)
+                masks = np.stack(masks, axis=0)
+                masks = torch.from_numpy(masks)
+                
+                seg_flag = False
+                for conv in conversation:
+                    conv['value'] = conv['value'].replace(REGION_TOKEN_REPLACE, f'[{REGION_TOKEN}]')
+                    if SEG_TOKEN in conv['value']:
+                        seg_flag = True
+                if seg_flag is False:
+                    image2maskids = []
+                else:
+                    mask_ids = [-10000 for i in range(len(mask_ids))]
+            else:
+                image2maskids = []
+                masks = torch.zeros((1, 336, 336))
+                mask_ids.append(-10000)
+                
+        elif 'video' in data_dict and data_dict['video'] is not None:
+            modal = 'video'
+            if all(not "<video>" in sentence["value"] for sentence in conversation):
+                warnings.warn(f"Video tag not found in the conversation, add it automatically at the beginning!")
+                conversation[0]["value"] = "<video>" + conversation[0]["value"]
+            if 'video_root' in data_dict and data_dict['video_root'] is not None:
+                video_root = data_dict['video_root']
+                video_file = [os.path.join(video_root,d) for d in data_dict['video']]
+            else:
+                video_file = data_dict['video']
+
+            if not isinstance(video_file, list):
+                video_file = [video_file]
+            if isinstance(video_file, list) and len(video_file) == 1 and ('timestamps' not in data_dict or data_dict['timestamps'] is None):
+                video_file = os.path.join(data_folder, video_file[0])
+                must_sample_frames = []
+                if 'masks' in data_dict and data_dict['masks'] is not None:
+                    if isinstance(data_dict['masks'], str):
+                        masks_ = json.load(open(data_dict['masks']))
+                    else:
+                        masks_= data_dict['masks']
+                    for ann in masks_:
+                        for k in ann.keys():
+                            must_sample_frames.append(int(k))
+                images, timestamps, mask_ids = load_video_from_ids(video_file, fps=self.data_args.fps, max_frames=self.data_args.max_frames, must_sample_frames=must_sample_frames)
+            elif isinstance(video_file, list): #images
+                images = []
+                for vf in video_file:
+                    images+=load_images(os.path.join(data_folder, vf))
+                timestamps = data_dict['timestamps']
+
+            else:
+                raise ValueError(f"Unsupported video format: {video_file}")
+            images = [images]
+            masks = []
+            mask_nums = []
+            image2maskids = []
+            maskid = 0
+            
+            if 'masks' in data_dict and data_dict['masks'] is not None:
+                if 'mask_ids' in data_dict and data_dict['mask_ids'] is not None:
+                    mask_ids = data_dict["mask_ids"]
+                if 'height' in data_dict:
+                    h = data_dict['height']
+                    w = data_dict['width']
+                else:
+                    h = None
+                    w = None
+
+                if isinstance(data_dict['masks'], str):
+                    masks_ = json.load(open(data_dict['masks']))
+                else:
+                    masks_= data_dict['masks']
+                for ann in masks_:
+                    image2maskids_ = [None]*len(video_file)
+                    for k in ann.keys():
+                        mask = annToMask(ann[k], h, w)
+                        masks.append(mask)
+                        image2maskids_[mask_ids[maskid]] = maskid
+                        maskid+=1
+                    image2maskids.append(image2maskids_)
+
+                    mask_nums.append(len(ann.keys()))
+                masks = np.stack(masks, axis=0)
+                masks = torch.from_numpy(masks)
+                
+                conv_i = 0
+                region_num = 0
+                seg_flag = False
+                for idx in range(len(mask_nums)):
+                    while '<region>' not in conversation[conv_i]['value'] and conv_i<len(conversation)-1:
+                        conv_i+=1
+                    conversation[conv_i]['value'] = conversation[conv_i]['value'].replace('<region>', "["+REGION_TOKEN*mask_nums[idx]+"]", 1)
+                    region_num += mask_nums[idx]
+                    if '[SEG]' in conversation[conv_i]['value']:
+                        seg_flag = True
+
+                if seg_flag is False:
+                    image2maskids = []
+                else:
+                    mask_ids = [-10000 for i in range(len(mask_ids))]
+                # assert region_num == len(masks), f"error in {conversation}"
+
+            else:
+                image2maskids = []
+                masks = torch.zeros((1, 336, 336))
+                mask_ids.append(-10000)
+            
+        else:
+            modal = 'text'
+            image2maskids = []
+            images = None
+            masks = torch.zeros((1, 336, 336))
+            sam_size = (336, 336)
+            sam_images = torch.zeros(1, 3, self.img_size, self.img_size)
+            mask_ids = [-10000]
+
+        if images is not None and len(images)>0:
+            sam_images = []
+            sam_size = None
+            if modal=='video':
+                for image in images[0]:
+                    sam_image = self.sam_transform.apply_image(np.array(image))
+                    sam_images.append(sam_image)
+                    if sam_size is None:
+                        sam_size = sam_image.shape[:2] 
+            else:
+                for image in images:
+                    sam_image = self.sam_transform.apply_image(np.array(image))
+                    sam_images.append(sam_image)
+                    if sam_size is None:
+                        sam_size = sam_image.shape[:2] 
+            sam_images = np.array(sam_images)
+            sam_images = torch.from_numpy(sam_images).permute(0, 3, 1, 2).contiguous()
+            sam_images = sam_preprocess_batch(sam_images)
+
+        messages = []
+        for conv in conversation:
+            if conv["from"] == "human":
+                # replace video tag to image tag for unified processing
+                # conv["value"] = conv["value"].replace("<video>", "<image>" * len(images))
+                chunks = conv["value"].split("<image>" if modal == 'image' else "<video>")
+                messages.append({
+                    "role": "user",
+                    "content": []
+                })
+
+                for chunk_idx in range(1, 2 * len(chunks)):
+                    if chunk_idx % 2 == 1:
+                        chunk = chunks[chunk_idx // 2].strip()
+                        messages[-1]["content"].append({"type": "text",  "text": chunk}) if chunk else None
+                    else:
+                        if modal == 'image':
+                            messages[-1]["content"].append({"type": "image"})
+                        elif modal == 'video':
+                            messages[-1]["content"].append({"type": "video", "num_frames": len(images[0]), "time": timestamps})
+            else:
+                messages.append({
+                    "role": "assistant",
+                    "content": conv['value']
+                })
+
+        # TODO: dynamic downsampling
+        # image_downsampling = self.data_args.spatial_merge_size
+        image_downsampling = 2 if modal == "video" else 1
+
+        return modal, images, messages, image_downsampling, masks, mask_ids, sam_images, sam_size, image2maskids
+
+    def __getitem__(self, i) -> Dict[str, torch.Tensor]:
+        data_dict = self.list_data_dict[i]
+
+        try:
+            modal, images, messages, image_downsampling, masks, mask_ids, sam_images, sam_size, image2maskids = self._convert_normal(data_dict)
+
+            data_dict = self.vlprocessor(
+                images=images,
+                text=messages,
+                merge_size=image_downsampling,
+                return_labels=True,
+                return_tensors="pt",
+            )
+
+            if modal == 'text':
+                unit_size = self.vlprocessor.image_processor.patch_size**2 * 3
+                data_dict['pixel_values'] = torch.zeros(self.vlprocessor.image_merge_size**2, unit_size)
+                data_dict['grid_sizes'] = torch.as_tensor([[1, self.vlprocessor.image_merge_size, self.vlprocessor.image_merge_size]])
+                data_dict['merge_sizes'] = torch.as_tensor([self.vlprocessor.image_merge_size])
+            elif modal == 'image' or modal == 'video':
+                assert len(data_dict['pixel_values']) > 0 and len(data_dict['grid_sizes']) > 0, f"Invalid image data: {data_dict['pixel_values']}, {data_dict['grid_sizes']}"
+
+            data_dict['modals'] = [modal] if isinstance(modal, str) else modal
+            data_dict['masks'] = masks
+            data_dict['mask_ids'] = mask_ids
+            data_dict['idx'] = i
+            data_dict['sam_images'] = sam_images
+            data_dict['sam_size'] = sam_size
+            data_dict['image2maskids'] = image2maskids
+
+        except Exception as e:
+            traceback.print_exc()
+            backup_idx = random.randint(0, len(self.list_data_dict) - 1)
+            print(f"Encounted error when process {i}-th example: {data_dict}, use {backup_idx}-th example instead!!!")
+            return self.__getitem__(backup_idx)
+
+        return data_dict
+
+
+@dataclass
+class DataCollatorForSupervisedDataset(object):
+    """Collate examples for supervised fine-tuning."""
+
+    vlprocessor: transformers.ProcessorMixin
+
+    def __call__(self, instances: Sequence[Dict]) -> Dict[str, torch.Tensor]:
+        input_ids, labels = tuple([instance[key] for instance in instances]
+                                  for key in ("input_ids", "labels"))
+        input_ids = torch.nn.utils.rnn.pad_sequence(
+            input_ids,
+            batch_first=True,
+            padding_value=self.vlprocessor.tokenizer.pad_token_id)
+        labels = torch.nn.utils.rnn.pad_sequence(labels,
+                                                 batch_first=True,
+                                                 padding_value=IGNORE_INDEX)
+        input_ids = input_ids[:, :self.vlprocessor.tokenizer.model_max_length]
+        labels = labels[:, :self.vlprocessor.tokenizer.model_max_length]
+        attention_mask = input_ids.ne(self.vlprocessor.tokenizer.pad_token_id)
+        position_ids = attention_mask.cumsum(dim=-1) - 1
+
+        batch = dict(
+            input_ids=input_ids,
+            labels=labels,
+            attention_mask=input_ids.ne(self.vlprocessor.tokenizer.pad_token_id),
+            position_ids=position_ids
+        )
+        
+        # work for 'images' argument in `prepare_inputs_labels_for_multimodal`
+        batch["pixel_values"] = torch.cat([x["pixel_values"] for x in instances])
+        batch["grid_sizes"] = torch.cat([x["grid_sizes"] for x in instances])
+        batch["merge_sizes"] = torch.cat([x["merge_sizes"] for x in instances])
+        batch["modals"] = sum([x["modals"] for x in instances], [])
+
+        batch['mask_ids'] = []
+        mask_idx_start = 0
+        for instance in instances:
+            if len(instance['mask_ids'])>0:
+                batch['mask_ids'].extend([idx+mask_idx_start for idx in instance['mask_ids']])      
+            # print(int(instance['grid_sizes'][0][0]))      
+            
+            mask_idx_start += int(instance['grid_sizes'][0][0])         
+        batch["masks"] = [x["masks"] for x in instances]
+        batch["sam_images"] = [x["sam_images"] for x in instances]
+        batch["sam_size"] = [x["sam_size"] for x in instances]
+        batch["image2maskids"] = [x["image2maskids"] for x in instances]
+        batch["idxes"] = [x["idx"] for x in instances]
+        return batch
+
+
+def make_supervised_data_module(vlprocessor, data_args) -> Dict:
+    """Make dataset and collator for supervised fine-tuning."""
+    train_dataset = LazySupervisedDataset(
+        vlprocessor=vlprocessor,
+        # data_folder=data_args.data_folder,
+        data_path=data_args.data_path,
+        data_args=data_args
+    )
+    data_collator = DataCollatorForSupervisedDataset(vlprocessor=vlprocessor)
+    return dict(train_dataset=train_dataset,
+                eval_dataset=None,
+                data_collator=data_collator)
+
+
+def train(attn_implementation=None):
+    global local_rank
+    set_seed(42)
+
+    parser = transformers.HfArgumentParser((ModelArguments, DataArguments, TrainingArguments))
+    model_args, data_args, training_args = parser.parse_args_into_dataclasses()
+
+    local_rank = training_args.local_rank
+
+    if local_rank == 0:
+        print('------model args------')
+        print(model_args)
+        print('------data args------')
+        print(data_args)
+        print('------training args------')
+        print(training_args)
+
+    compute_dtype = (torch.float16 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
+
+    bnb_model_from_pretrained_args = {}
+    if training_args.bits in [4, 8]:
+        from transformers import BitsAndBytesConfig
+        bnb_model_from_pretrained_args.update(dict(
+            # device_map={"": training_args.device},
+            # BUG: High version transformers report error:
+            # ValueError: You can't pass `load_in_4bit`or `load_in_8bit` as a kwarg when passing `quantization_config` argument at the same time
+            # load_in_4bit=training_args.bits == 4,
+            # load_in_8bit=training_args.bits == 8,
+            quantization_config=BitsAndBytesConfig(
+                load_in_4bit=training_args.bits == 4,
+                load_in_8bit=training_args.bits == 8,
+                llm_int8_skip_modules=["mm_projector"],
+                llm_int8_threshold=6.0,
+                llm_int8_has_fp16_weight=False,
+                bnb_4bit_compute_dtype=compute_dtype,
+                bnb_4bit_use_double_quant=training_args.double_quant,
+                bnb_4bit_quant_type=training_args.quant_type, # {'fp4', 'nf4'}
+                bnb_4bit_quant_storage=compute_dtype,
+            )
+        ))
+
+    config = RynnecQwen2Config.from_pretrained(model_args.model_path)
+
+    config._attn_implementation = attn_implementation
+    # NOTE: active spatial_merge_size arguments
+    config.spatial_merge_size = model_args.spatial_merge_size
+    config.mm_max_length = model_args.mm_max_length
+    config.use_token_compression = model_args.use_token_compression
+    config.loss_reduction_scope = training_args.loss_reduction_scope
+    config.mask_decoder_model = model_args.mask_decoder_model
+    config.training = model_args.training
+    config.has_mask = model_args.has_mask
+
+    if model_args.vision_encoder is not None:
+        model = RynnecQwen2ForCausalLM.from_pretrained(
+            model_args.model_path,
+            config=config,
+            torch_dtype=compute_dtype,
+            do_sample=True,
+            **bnb_model_from_pretrained_args
+        )
+        if 'mixtral' in model_args.model_type:
+            import deepspeed
+            deepspeed.utils.set_z3_leaf_modules(model, [MixtralSparseMoeBlock])
+    else:
+        model = transformers.LlamaForCausalLM.from_pretrained(
+            model_args.model_path,
+            config=config,
+            torch_dtype=compute_dtype,
+            do_sample=True,
+            **bnb_model_from_pretrained_args
+        )
+    model.config.use_cache = False
+    if model_args.freeze_backbone:
+        model.model.requires_grad_(False)
+
+    if training_args.bits in [4, 8]:
+        from peft import prepare_model_for_kbit_training
+        model.config.torch_dtype=(torch.float32 if training_args.fp16 else (torch.bfloat16 if training_args.bf16 else torch.float32))
+        model = prepare_model_for_kbit_training(model, use_gradient_checkpointing=training_args.gradient_checkpointing)
+
+    if training_args.gradient_checkpointing:
+        if hasattr(model, "enable_input_require_grads"):
+            model.enable_input_require_grads()
+        else:
+            def make_inputs_require_grad(module, input, output):
+                output.requires_grad_(True)
+            model.get_input_embeddings().register_forward_hook(make_inputs_require_grad)
+
+    if training_args.lora_enable:
+        from peft import LoraConfig, get_peft_model
+        lora_config = LoraConfig(
+            r=training_args.lora_r,
+            lora_alpha=training_args.lora_alpha,
+            target_modules=find_all_linear_names(model),
+            lora_dropout=training_args.lora_dropout,
+            bias=training_args.lora_bias,
+            task_type="CAUSAL_LM",
+        )
+        if training_args.bits == 16:
+            if training_args.bf16:
+                model.to(torch.bfloat16)
+            if training_args.fp16:
+                model.to(torch.float16)
+        rank0_print("Adding LoRA adapters...")
+        model = get_peft_model(model, lora_config)
+
+    tokenizer = transformers.AutoTokenizer.from_pretrained(
+        model_args.model_path,
+        model_max_length=training_args.model_max_length,
+        padding_side="right",
+        use_fast=True,
+    )
+
+    if tokenizer.pad_token is None:
+        tokenizer.pad_token = tokenizer.unk_token
+
+    if model_args.vision_encoder is not None:
+        # initialize vision encoder + multi-modal projector
+        model.get_model().initialize_vision_modules(model_args=model_args, fsdp=training_args.fsdp)
+
+        if model_args.load_sam2_weight is True:
+            model.get_model().build_mask_decoder(model.get_model().config)
+            model.load_sam2_weights(model_args.mask_decoder_model)
+
+        vision_encoder = model.get_vision_encoder()
+        vision_encoder.to(dtype=compute_dtype, device=training_args.device)
+
+        vision_encoder.image_processor.max_tokens = model_args.mm_max_length
+        mm_projector = model.get_mm_projector()
+        mm_projector.to(dtype=compute_dtype if training_args.bf16 else torch.float16, device=training_args.device)
+
+        data_args.is_multimodal = True
+
+        model.config.tokenizer_padding_side = tokenizer.padding_side
+        model.config.tokenizer_model_max_length = tokenizer.model_max_length
+
+        if training_args.bits in [4, 8]:
+            model.get_model().mm_projector.to(dtype=compute_dtype, device=training_args.device)
+
+        # decoupled learning rate
+        model.config.llm_lr = training_args.llm_lr
+        model.config.vision_encoder_lr = training_args.vision_encoder_lr
+        model.config.mm_projector_lr = training_args.mm_projector_lr
+        model.config.region_encoder_lr = training_args.region_encoder_lr
+        model.config.sam_decoder_lr = training_args.sam_decoder_lr
+        model.config.sam_encoder_lr = training_args.sam_encoder_lr
+        model.config.dice_loss_weight = 0.5
+        model.config.bce_loss_weight = 2.0
+
+        if model.config.llm_lr is None:
+            for p in model.get_model().parameters():
+                p.requires_grad = False
+            for p in model.get_model().vision_encoder.parameters():
+                p.requires_grad = True
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = True
+            for p in model.get_model().region_encoder.parameters():
+                p.requires_grad = True
+
+
+        if model.config.vision_encoder_lr is None:
+            for p in model.get_model().vision_encoder.parameters():
+                p.requires_grad = False
+
+        if model.config.mm_projector_lr is None:
+            for p in model.get_model().mm_projector.parameters():
+                p.requires_grad = False
+                
+        if model.config.region_encoder_lr is None:
+            for p in model.get_model().region_encoder.parameters():
+                p.requires_grad = False
+
+        if model.config.sam_decoder_lr is None:
+            for p in model.grounding_encoder.sam2_model.sam_mask_decoder.parameters():
+                p.requires_grad = False
+        else:
+            for p in model.grounding_encoder.sam2_model.sam_mask_decoder.parameters():
+                p.requires_grad = True
+
+        if model.config.sam_encoder_lr is None:
+            for p in model.grounding_encoder.sam2_model.image_encoder.parameters():
+                p.requires_grad = False
+
+        if training_args.lora_enable:
+            for n, p in model.named_parameters():
+                if any(
+                    [
+                        x in n
+                        for x in ["lm_head", "embed_tokens", "text_hidden_fcs"]
+                    ]
+                ):
+                    # print(n)
+                    p.requires_grad = True
+    
+        model.config.max_frames = getattr(data_args, 'max_frames', NUM_FRAMES)
+        model.config.image_aspect_ratio = data_args.image_aspect_ratio if 'qwen2vl' not in model_args.vision_encoder else 'qwen2vl'
+
+        # NOTE: complement data_args via model hyperparameters
+        # 1. acquire image size
+        model.config.image_size = data_args.image_size = vision_encoder.image_size
+        # 2. calculate the number of tokens in the image
+        model.config.image_token_length = data_args.image_token_length = mm_projector.cal_proj_size(vision_encoder.num_patches_per_side)
+        # 3. check if alignment
+        model.config.is_alignment = training_args.is_alignment = data_args.is_alignment = (
+            model.config.mm_projector_lr is not None and
+            model.config.llm_lr is None and
+            model.config.vision_encoder_lr is None
+        )
+        # 4. set spatial merge size as default
+        model.config.spatial_merge_size = data_args.spatial_merge_size = model_args.spatial_merge_size
+        tokenizer.add_tokens([DEFAULT_IMAGE_TOKEN, STREAM_START_TOKEN, STREAM_END_TOKEN], special_tokens=True)
+        tokenizer.add_tokens([REGION_TOKEN], special_tokens=True)
+        num_new_tokens = tokenizer.add_tokens([SEG_TOKEN], special_tokens=True)
+        model.resize_token_embeddings(len(tokenizer))
+
+        model.config.image_token_index = tokenizer.convert_tokens_to_ids(DEFAULT_IMAGE_TOKEN)
+        model.config.region_token_index = tokenizer.convert_tokens_to_ids(REGION_TOKEN)
+        model.config.seg_token_index = tokenizer.convert_tokens_to_ids(SEG_TOKEN)
+
+        vlprocessor = Videollama3Qwen2Processor(vision_encoder.image_processor, tokenizer)
+
+    if training_args.bits in [4, 8]:
+        from peft.tuners.lora import LoraLayer
+        for name, module in model.named_modules():
+            if isinstance(module, LoraLayer):
+                if training_args.bf16:
+                    module = module.to(torch.bfloat16)
+            if 'norm' in name:
+                module = module.to(torch.float32)
+            if 'lm_head' in name or 'embed_tokens' in name:
+                if hasattr(module, 'weight'):
+                    if training_args.bf16 and module.weight.dtype == torch.float32:
+                        module = module.to(torch.bfloat16)
+
+    if local_rank == 0:
+        print("Current model:", model)
+        print("Model config:", model.config)
+
+    
+    data_module = make_supervised_data_module(vlprocessor=vlprocessor, data_args=data_args)
+
+    # select a Trainer
+    trainer = RynnECTrainer(model=model, tokenizer=tokenizer, args=training_args, **data_module)
+
+    if list(pathlib.Path(training_args.output_dir).glob("checkpoint-*")):
+        trainer.train(resume_from_checkpoint=True)
+    else:
+        trainer.train()
+    trainer.save_state()
+
+    model.config.use_cache = True
+
+    if training_args.lora_enable:
+        state_dict = get_peft_state_maybe_zero_3(model.named_parameters(), training_args.lora_bias)
+        non_lora_state_dict = get_peft_state_non_lora_maybe_zero_3(model.named_parameters())
+        if training_args.local_rank == 0 or training_args.local_rank == -1:
+            model.config.save_pretrained(training_args.output_dir)
+            model.save_pretrained(training_args.output_dir, state_dict=state_dict)
+            torch.save(non_lora_state_dict, os.path.join(training_args.output_dir, 'non_lora_trainables.bin'))
+    else:
+        safe_save_model_for_hf_trainer(trainer=trainer, output_dir=training_args.output_dir)
+
+
+if __name__ == "__main__":
+    train(attn_implementation="flash_attention_2")
diff --git a/third_parts/sam2/__init__.py b/third_parts/sam2/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..e35d3f5bc43d604ccc7574c212f421dc4b76cde0
--- /dev/null
+++ b/third_parts/sam2/__init__.py
@@ -0,0 +1,9 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from hydra import initialize_config_module
+
+initialize_config_module("third_parts.sam2.sam2_configs", version_base="1.2")
diff --git a/third_parts/sam2/automatic_mask_generator.py b/third_parts/sam2/automatic_mask_generator.py
new file mode 100644
index 0000000000000000000000000000000000000000..9c4c46d814347140ea2f7a01e8109bcf161103c0
--- /dev/null
+++ b/third_parts/sam2/automatic_mask_generator.py
@@ -0,0 +1,434 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+# Adapted from https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/automatic_mask_generator.py
+from typing import Any, Dict, List, Optional, Tuple
+
+import numpy as np
+import torch
+from torchvision.ops.boxes import batched_nms, box_area  # type: ignore
+
+from third_parts.sam2.modeling.sam2_base import SAM2Base
+from third_parts.sam2.sam2_image_predictor import SAM2ImagePredictor
+from third_parts.sam2.utils.amg import (
+    area_from_rle,
+    batch_iterator,
+    batched_mask_to_box,
+    box_xyxy_to_xywh,
+    build_all_layer_point_grids,
+    calculate_stability_score,
+    coco_encode_rle,
+    generate_crop_boxes,
+    is_box_near_crop_edge,
+    mask_to_rle_pytorch,
+    MaskData,
+    remove_small_regions,
+    rle_to_mask,
+    uncrop_boxes_xyxy,
+    uncrop_masks,
+    uncrop_points,
+)
+
+
+class SAM2AutomaticMaskGenerator:
+    def __init__(
+        self,
+        model: SAM2Base,
+        points_per_side: Optional[int] = 32,
+        points_per_batch: int = 64,
+        pred_iou_thresh: float = 0.8,
+        stability_score_thresh: float = 0.95,
+        stability_score_offset: float = 1.0,
+        mask_threshold: float = 0.0,
+        box_nms_thresh: float = 0.7,
+        crop_n_layers: int = 0,
+        crop_nms_thresh: float = 0.7,
+        crop_overlap_ratio: float = 512 / 1500,
+        crop_n_points_downscale_factor: int = 1,
+        point_grids: Optional[List[np.ndarray]] = None,
+        min_mask_region_area: int = 0,
+        output_mode: str = "binary_mask",
+        use_m2m: bool = False,
+        multimask_output: bool = True,
+    ) -> None:
+        """
+        Using a SAM 2 model, generates masks for the entire image.
+        Generates a grid of point prompts over the image, then filters
+        low quality and duplicate masks. The default settings are chosen
+        for SAM 2 with a HieraL backbone.
+
+        Arguments:
+          model (Sam): The SAM 2 model to use for mask prediction.
+          points_per_side (int or None): The number of points to be sampled
+            along one side of the image. The total number of points is
+            points_per_side**2. If None, 'point_grids' must provide explicit
+            point sampling.
+          points_per_batch (int): Sets the number of points run simultaneously
+            by the model. Higher numbers may be faster but use more GPU memory.
+          pred_iou_thresh (float): A filtering threshold in [0,1], using the
+            model's predicted mask quality.
+          stability_score_thresh (float): A filtering threshold in [0,1], using
+            the stability of the mask under changes to the cutoff used to binarize
+            the model's mask predictions.
+          stability_score_offset (float): The amount to shift the cutoff when
+            calculated the stability score.
+          mask_threshold (float): Threshold for binarizing the mask logits
+          box_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks.
+          crop_n_layers (int): If >0, mask prediction will be run again on
+            crops of the image. Sets the number of layers to run, where each
+            layer has 2**i_layer number of image crops.
+          crop_nms_thresh (float): The box IoU cutoff used by non-maximal
+            suppression to filter duplicate masks between different crops.
+          crop_overlap_ratio (float): Sets the degree to which crops overlap.
+            In the first crop layer, crops will overlap by this fraction of
+            the image length. Later layers with more crops scale down this overlap.
+          crop_n_points_downscale_factor (int): The number of points-per-side
+            sampled in layer n is scaled down by crop_n_points_downscale_factor**n.
+          point_grids (list(np.ndarray) or None): A list over explicit grids
+            of points used for sampling, normalized to [0,1]. The nth grid in the
+            list is used in the nth crop layer. Exclusive with points_per_side.
+          min_mask_region_area (int): If >0, postprocessing will be applied
+            to remove disconnected regions and holes in masks with area smaller
+            than min_mask_region_area. Requires opencv.
+          output_mode (str): The form masks are returned in. Can be 'binary_mask',
+            'uncompressed_rle', or 'coco_rle'. 'coco_rle' requires pycocotools.
+            For large resolutions, 'binary_mask' may consume large amounts of
+            memory.
+          use_m2m (bool): Whether to add a one step refinement using previous mask predictions.
+          multimask_output (bool): Whether to output multimask at each point of the grid.
+        """
+
+        assert (points_per_side is None) != (
+            point_grids is None
+        ), "Exactly one of points_per_side or point_grid must be provided."
+        if points_per_side is not None:
+            self.point_grids = build_all_layer_point_grids(
+                points_per_side,
+                crop_n_layers,
+                crop_n_points_downscale_factor,
+            )
+        elif point_grids is not None:
+            self.point_grids = point_grids
+        else:
+            raise ValueError("Can't have both points_per_side and point_grid be None.")
+
+        assert output_mode in [
+            "binary_mask",
+            "uncompressed_rle",
+            "coco_rle",
+        ], f"Unknown output_mode {output_mode}."
+        if output_mode == "coco_rle":
+            try:
+                from pycocotools import mask as mask_utils  # type: ignore  # noqa: F401
+            except ImportError as e:
+                print("Please install pycocotools")
+                raise e
+
+        self.predictor = SAM2ImagePredictor(
+            model,
+            max_hole_area=min_mask_region_area,
+            max_sprinkle_area=min_mask_region_area,
+        )
+        self.points_per_batch = points_per_batch
+        self.pred_iou_thresh = pred_iou_thresh
+        self.stability_score_thresh = stability_score_thresh
+        self.stability_score_offset = stability_score_offset
+        self.mask_threshold = mask_threshold
+        self.box_nms_thresh = box_nms_thresh
+        self.crop_n_layers = crop_n_layers
+        self.crop_nms_thresh = crop_nms_thresh
+        self.crop_overlap_ratio = crop_overlap_ratio
+        self.crop_n_points_downscale_factor = crop_n_points_downscale_factor
+        self.min_mask_region_area = min_mask_region_area
+        self.output_mode = output_mode
+        self.use_m2m = use_m2m
+        self.multimask_output = multimask_output
+
+    @torch.no_grad()
+    def generate(self, image: np.ndarray) -> List[Dict[str, Any]]:
+        """
+        Generates masks for the given image.
+
+        Arguments:
+          image (np.ndarray): The image to generate masks for, in HWC uint8 format.
+
+        Returns:
+           list(dict(str, any)): A list over records for masks. Each record is
+             a dict containing the following keys:
+               segmentation (dict(str, any) or np.ndarray): The mask. If
+                 output_mode='binary_mask', is an array of shape HW. Otherwise,
+                 is a dictionary containing the RLE.
+               bbox (list(float)): The box around the mask, in XYWH format.
+               area (int): The area in pixels of the mask.
+               predicted_iou (float): The model's own prediction of the mask's
+                 quality. This is filtered by the pred_iou_thresh parameter.
+               point_coords (list(list(float))): The point coordinates input
+                 to the model to generate this mask.
+               stability_score (float): A measure of the mask's quality. This
+                 is filtered on using the stability_score_thresh parameter.
+               crop_box (list(float)): The crop of the image used to generate
+                 the mask, given in XYWH format.
+        """
+
+        # Generate masks
+        mask_data = self._generate_masks(image)
+
+        # Encode masks
+        if self.output_mode == "coco_rle":
+            mask_data["segmentations"] = [
+                coco_encode_rle(rle) for rle in mask_data["rles"]
+            ]
+        elif self.output_mode == "binary_mask":
+            mask_data["segmentations"] = [rle_to_mask(rle) for rle in mask_data["rles"]]
+        else:
+            mask_data["segmentations"] = mask_data["rles"]
+
+        # Write mask records
+        curr_anns = []
+        for idx in range(len(mask_data["segmentations"])):
+            ann = {
+                "segmentation": mask_data["segmentations"][idx],
+                "area": area_from_rle(mask_data["rles"][idx]),
+                "bbox": box_xyxy_to_xywh(mask_data["boxes"][idx]).tolist(),
+                "predicted_iou": mask_data["iou_preds"][idx].item(),
+                "point_coords": [mask_data["points"][idx].tolist()],
+                "stability_score": mask_data["stability_score"][idx].item(),
+                "crop_box": box_xyxy_to_xywh(mask_data["crop_boxes"][idx]).tolist(),
+            }
+            curr_anns.append(ann)
+
+        return curr_anns
+
+    def _generate_masks(self, image: np.ndarray) -> MaskData:
+        orig_size = image.shape[:2]
+        crop_boxes, layer_idxs = generate_crop_boxes(
+            orig_size, self.crop_n_layers, self.crop_overlap_ratio
+        )
+
+        # Iterate over image crops
+        data = MaskData()
+        for crop_box, layer_idx in zip(crop_boxes, layer_idxs):
+            crop_data = self._process_crop(image, crop_box, layer_idx, orig_size)
+            data.cat(crop_data)
+
+        # Remove duplicate masks between crops
+        if len(crop_boxes) > 1:
+            # Prefer masks from smaller crops
+            scores = 1 / box_area(data["crop_boxes"])
+            scores = scores.to(data["boxes"].device)
+            keep_by_nms = batched_nms(
+                data["boxes"].float(),
+                scores,
+                torch.zeros_like(data["boxes"][:, 0]),  # categories
+                iou_threshold=self.crop_nms_thresh,
+            )
+            data.filter(keep_by_nms)
+        data.to_numpy()
+        return data
+
+    def _process_crop(
+        self,
+        image: np.ndarray,
+        crop_box: List[int],
+        crop_layer_idx: int,
+        orig_size: Tuple[int, ...],
+    ) -> MaskData:
+        # Crop the image and calculate embeddings
+        x0, y0, x1, y1 = crop_box
+        cropped_im = image[y0:y1, x0:x1, :]
+        cropped_im_size = cropped_im.shape[:2]
+        self.predictor.set_image(cropped_im)
+
+        # Get points for this crop
+        points_scale = np.array(cropped_im_size)[None, ::-1]
+        points_for_image = self.point_grids[crop_layer_idx] * points_scale
+
+        # Generate masks for this crop in batches
+        data = MaskData()
+        for (points,) in batch_iterator(self.points_per_batch, points_for_image):
+            batch_data = self._process_batch(
+                points, cropped_im_size, crop_box, orig_size, normalize=True
+            )
+            data.cat(batch_data)
+            del batch_data
+        self.predictor.reset_predictor()
+
+        # Remove duplicates within this crop.
+        keep_by_nms = batched_nms(
+            data["boxes"].float(),
+            data["iou_preds"],
+            torch.zeros_like(data["boxes"][:, 0]),  # categories
+            iou_threshold=self.box_nms_thresh,
+        )
+        data.filter(keep_by_nms)
+
+        # Return to the original image frame
+        data["boxes"] = uncrop_boxes_xyxy(data["boxes"], crop_box)
+        data["points"] = uncrop_points(data["points"], crop_box)
+        data["crop_boxes"] = torch.tensor([crop_box for _ in range(len(data["rles"]))])
+
+        return data
+
+    def _process_batch(
+        self,
+        points: np.ndarray,
+        im_size: Tuple[int, ...],
+        crop_box: List[int],
+        orig_size: Tuple[int, ...],
+        normalize=False,
+    ) -> MaskData:
+        orig_h, orig_w = orig_size
+
+        # Run model on this batch
+        points = torch.as_tensor(points, device=self.predictor.device)
+        in_points = self.predictor._transforms.transform_coords(
+            points, normalize=normalize, orig_hw=im_size
+        )
+        in_labels = torch.ones(
+            in_points.shape[0], dtype=torch.int, device=in_points.device
+        )
+        masks, iou_preds, low_res_masks = self.predictor._predict(
+            in_points[:, None, :],
+            in_labels[:, None],
+            multimask_output=self.multimask_output,
+            return_logits=True,
+        )
+
+        # Serialize predictions and store in MaskData
+        data = MaskData(
+            masks=masks.flatten(0, 1),
+            iou_preds=iou_preds.flatten(0, 1),
+            points=points.repeat_interleave(masks.shape[1], dim=0),
+            low_res_masks=low_res_masks.flatten(0, 1),
+        )
+        del masks
+
+        if not self.use_m2m:
+            # Filter by predicted IoU
+            if self.pred_iou_thresh > 0.0:
+                keep_mask = data["iou_preds"] > self.pred_iou_thresh
+                data.filter(keep_mask)
+
+            # Calculate and filter by stability score
+            data["stability_score"] = calculate_stability_score(
+                data["masks"], self.mask_threshold, self.stability_score_offset
+            )
+            if self.stability_score_thresh > 0.0:
+                keep_mask = data["stability_score"] >= self.stability_score_thresh
+                data.filter(keep_mask)
+        else:
+            # One step refinement using previous mask predictions
+            in_points = self.predictor._transforms.transform_coords(
+                data["points"], normalize=normalize, orig_hw=im_size
+            )
+            labels = torch.ones(
+                in_points.shape[0], dtype=torch.int, device=in_points.device
+            )
+            masks, ious = self.refine_with_m2m(
+                in_points, labels, data["low_res_masks"], self.points_per_batch
+            )
+            data["masks"] = masks.squeeze(1)
+            data["iou_preds"] = ious.squeeze(1)
+
+            if self.pred_iou_thresh > 0.0:
+                keep_mask = data["iou_preds"] > self.pred_iou_thresh
+                data.filter(keep_mask)
+
+            data["stability_score"] = calculate_stability_score(
+                data["masks"], self.mask_threshold, self.stability_score_offset
+            )
+            if self.stability_score_thresh > 0.0:
+                keep_mask = data["stability_score"] >= self.stability_score_thresh
+                data.filter(keep_mask)
+
+        # Threshold masks and calculate boxes
+        data["masks"] = data["masks"] > self.mask_threshold
+        data["boxes"] = batched_mask_to_box(data["masks"])
+
+        # Filter boxes that touch crop boundaries
+        keep_mask = ~is_box_near_crop_edge(
+            data["boxes"], crop_box, [0, 0, orig_w, orig_h]
+        )
+        if not torch.all(keep_mask):
+            data.filter(keep_mask)
+
+        # Compress to RLE
+        data["masks"] = uncrop_masks(data["masks"], crop_box, orig_h, orig_w)
+        data["rles"] = mask_to_rle_pytorch(data["masks"])
+        del data["masks"]
+
+        return data
+
+    @staticmethod
+    def postprocess_small_regions(
+        mask_data: MaskData, min_area: int, nms_thresh: float
+    ) -> MaskData:
+        """
+        Removes small disconnected regions and holes in masks, then reruns
+        box NMS to remove any new duplicates.
+
+        Edits mask_data in place.
+
+        Requires open-cv as a dependency.
+        """
+        if len(mask_data["rles"]) == 0:
+            return mask_data
+
+        # Filter small disconnected regions and holes
+        new_masks = []
+        scores = []
+        for rle in mask_data["rles"]:
+            mask = rle_to_mask(rle)
+
+            mask, changed = remove_small_regions(mask, min_area, mode="holes")
+            unchanged = not changed
+            mask, changed = remove_small_regions(mask, min_area, mode="islands")
+            unchanged = unchanged and not changed
+
+            new_masks.append(torch.as_tensor(mask).unsqueeze(0))
+            # Give score=0 to changed masks and score=1 to unchanged masks
+            # so NMS will prefer ones that didn't need postprocessing
+            scores.append(float(unchanged))
+
+        # Recalculate boxes and remove any new duplicates
+        masks = torch.cat(new_masks, dim=0)
+        boxes = batched_mask_to_box(masks)
+        keep_by_nms = batched_nms(
+            boxes.float(),
+            torch.as_tensor(scores),
+            torch.zeros_like(boxes[:, 0]),  # categories
+            iou_threshold=nms_thresh,
+        )
+
+        # Only recalculate RLEs for masks that have changed
+        for i_mask in keep_by_nms:
+            if scores[i_mask] == 0.0:
+                mask_torch = masks[i_mask].unsqueeze(0)
+                mask_data["rles"][i_mask] = mask_to_rle_pytorch(mask_torch)[0]
+                mask_data["boxes"][i_mask] = boxes[i_mask]  # update res directly
+        mask_data.filter(keep_by_nms)
+
+        return mask_data
+
+    def refine_with_m2m(self, points, point_labels, low_res_masks, points_per_batch):
+        new_masks = []
+        new_iou_preds = []
+
+        for cur_points, cur_point_labels, low_res_mask in batch_iterator(
+            points_per_batch, points, point_labels, low_res_masks
+        ):
+            best_masks, best_iou_preds, _ = self.predictor._predict(
+                cur_points[:, None, :],
+                cur_point_labels[:, None],
+                mask_input=low_res_mask[:, None, :],
+                multimask_output=False,
+                return_logits=True,
+            )
+            new_masks.append(best_masks)
+            new_iou_preds.append(best_iou_preds)
+        masks = torch.cat(new_masks, dim=0)
+        return masks, torch.cat(new_iou_preds, dim=0)
diff --git a/third_parts/sam2/build_sam.py b/third_parts/sam2/build_sam.py
new file mode 100644
index 0000000000000000000000000000000000000000..8a4c873cda7895096593754a35ef83f8494b6d60
--- /dev/null
+++ b/third_parts/sam2/build_sam.py
@@ -0,0 +1,89 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+import torch
+from hydra import compose
+from hydra.utils import instantiate
+from omegaconf import OmegaConf
+
+
+def build_sam2(
+    config_file,
+    ckpt_path=None,
+    device="cuda",
+    mode="eval",
+    hydra_overrides_extra=[],
+    apply_postprocessing=True,
+):
+
+    if apply_postprocessing:
+        hydra_overrides_extra = hydra_overrides_extra.copy()
+        hydra_overrides_extra += [
+            # dynamically fall back to multi-mask if the single mask is not stable
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_via_stability=true",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_delta=0.05",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_thresh=0.98",
+        ]
+    # Read config and init model
+    cfg = compose(config_name=config_file, overrides=hydra_overrides_extra)
+    OmegaConf.resolve(cfg)
+    model = instantiate(cfg.model, _recursive_=True)
+    _load_checkpoint(model, ckpt_path)
+    model = model.to(device)
+    if mode == "eval":
+        model.eval()
+    return model
+
+
+def build_sam2_video_predictor(
+    config_file,
+    ckpt_path=None,
+    device="cuda",
+    mode="eval",
+    hydra_overrides_extra=[],
+    apply_postprocessing=True,
+):
+    hydra_overrides = [
+        "++model._target_=third_parts.sam2.sam2_video_predictor.SAM2VideoPredictor",
+    ]
+    if apply_postprocessing:
+        hydra_overrides_extra = hydra_overrides_extra.copy()
+        hydra_overrides_extra += [
+            # dynamically fall back to multi-mask if the single mask is not stable
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_via_stability=true",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_delta=0.05",
+            "++model.sam_mask_decoder_extra_args.dynamic_multimask_stability_thresh=0.98",
+            # the sigmoid mask logits on interacted frames with clicks in the memory encoder so that the encoded masks are exactly as what users see from clicking
+            "++model.binarize_mask_from_pts_for_mem_enc=true",
+            # fill small holes in the low-res masks up to `fill_hole_area` (before resizing them to the original video resolution)
+            "++model.fill_hole_area=8",
+        ]
+    hydra_overrides.extend(hydra_overrides_extra)
+
+    # Read config and init model
+    cfg = compose(config_name=config_file, overrides=hydra_overrides)
+    OmegaConf.resolve(cfg)
+    model = instantiate(cfg.model, _recursive_=True)
+    _load_checkpoint(model, ckpt_path)
+    model = model.to(device)
+    if mode == "eval":
+        model.eval()
+    return model
+
+
+def _load_checkpoint(model, ckpt_path):
+    if ckpt_path is not None:
+        sd = torch.load(ckpt_path, map_location="cpu")["model"]
+        missing_keys, unexpected_keys = model.load_state_dict(sd)
+        if missing_keys:
+            logging.error(missing_keys)
+            raise RuntimeError()
+        if unexpected_keys:
+            logging.error(unexpected_keys)
+            raise RuntimeError()
+        logging.info("Loaded checkpoint sucessfully")
diff --git a/third_parts/sam2/csrc/connected_components.cu b/third_parts/sam2/csrc/connected_components.cu
new file mode 100644
index 0000000000000000000000000000000000000000..ced21eb32eaaadb818d441c1322b99d1bf068f45
--- /dev/null
+++ b/third_parts/sam2/csrc/connected_components.cu
@@ -0,0 +1,289 @@
+// Copyright (c) Meta Platforms, Inc. and affiliates.
+// All rights reserved.
+
+// This source code is licensed under the license found in the
+// LICENSE file in the root directory of this source tree.
+
+// adapted from https://github.com/zsef123/Connected_components_PyTorch
+// with license found in the LICENSE_cctorch file in the root directory.
+#include <ATen/cuda/CUDAContext.h>
+#include <cuda.h>
+#include <cuda_runtime.h>
+#include <torch/extension.h>
+#include <torch/script.h>
+#include <vector>
+
+// 2d
+#define BLOCK_ROWS 16
+#define BLOCK_COLS 16
+
+namespace cc2d {
+
+template <typename T>
+__device__ __forceinline__ unsigned char hasBit(T bitmap, unsigned char pos) {
+  return (bitmap >> pos) & 1;
+}
+
+__device__ int32_t find(const int32_t* s_buf, int32_t n) {
+  while (s_buf[n] != n)
+    n = s_buf[n];
+  return n;
+}
+
+__device__ int32_t find_n_compress(int32_t* s_buf, int32_t n) {
+  const int32_t id = n;
+  while (s_buf[n] != n) {
+    n = s_buf[n];
+    s_buf[id] = n;
+  }
+  return n;
+}
+
+__device__ void union_(int32_t* s_buf, int32_t a, int32_t b) {
+  bool done;
+  do {
+    a = find(s_buf, a);
+    b = find(s_buf, b);
+
+    if (a < b) {
+      int32_t old = atomicMin(s_buf + b, a);
+      done = (old == b);
+      b = old;
+    } else if (b < a) {
+      int32_t old = atomicMin(s_buf + a, b);
+      done = (old == a);
+      a = old;
+    } else
+      done = true;
+
+  } while (!done);
+}
+
+__global__ void
+init_labeling(int32_t* label, const uint32_t W, const uint32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row < H && col < W)
+    label[idx] = idx;
+}
+
+__global__ void
+merge(uint8_t* img, int32_t* label, const uint32_t W, const uint32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  uint32_t P = 0;
+
+  if (img[idx])
+    P |= 0x777;
+  if (row + 1 < H && img[idx + W])
+    P |= 0x777 << 4;
+  if (col + 1 < W && img[idx + 1])
+    P |= 0x777 << 1;
+
+  if (col == 0)
+    P &= 0xEEEE;
+  if (col + 1 >= W)
+    P &= 0x3333;
+  else if (col + 2 >= W)
+    P &= 0x7777;
+
+  if (row == 0)
+    P &= 0xFFF0;
+  if (row + 1 >= H)
+    P &= 0xFF;
+
+  if (P > 0) {
+    // If need check about top-left pixel(if flag the first bit) and hit the
+    // top-left pixel
+    if (hasBit(P, 0) && img[idx - W - 1]) {
+      union_(label, idx, idx - 2 * W - 2); // top left block
+    }
+
+    if ((hasBit(P, 1) && img[idx - W]) || (hasBit(P, 2) && img[idx - W + 1]))
+      union_(label, idx, idx - 2 * W); // top bottom block
+
+    if (hasBit(P, 3) && img[idx + 2 - W])
+      union_(label, idx, idx - 2 * W + 2); // top right block
+
+    if ((hasBit(P, 4) && img[idx - 1]) || (hasBit(P, 8) && img[idx + W - 1]))
+      union_(label, idx, idx - 2); // just left block
+  }
+}
+
+__global__ void compression(int32_t* label, const int32_t W, const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row < H && col < W)
+    find_n_compress(label, idx);
+}
+
+__global__ void final_labeling(
+    const uint8_t* img,
+    int32_t* label,
+    const int32_t W,
+    const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y) * 2;
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x) * 2;
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  int32_t y = label[idx] + 1;
+
+  if (img[idx])
+    label[idx] = y;
+  else
+    label[idx] = 0;
+
+  if (col + 1 < W) {
+    if (img[idx + 1])
+      label[idx + 1] = y;
+    else
+      label[idx + 1] = 0;
+
+    if (row + 1 < H) {
+      if (img[idx + W + 1])
+        label[idx + W + 1] = y;
+      else
+        label[idx + W + 1] = 0;
+    }
+  }
+
+  if (row + 1 < H) {
+    if (img[idx + W])
+      label[idx + W] = y;
+    else
+      label[idx + W] = 0;
+  }
+}
+
+__global__ void init_counting(
+    const int32_t* label,
+    int32_t* count_init,
+    const int32_t W,
+    const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y);
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x);
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  int32_t y = label[idx];
+  if (y > 0) {
+    int32_t count_idx = y - 1;
+    atomicAdd(count_init + count_idx, 1);
+  }
+}
+
+__global__ void final_counting(
+    const int32_t* label,
+    const int32_t* count_init,
+    int32_t* count_final,
+    const int32_t W,
+    const int32_t H) {
+  const uint32_t row = (blockIdx.y * blockDim.y + threadIdx.y);
+  const uint32_t col = (blockIdx.x * blockDim.x + threadIdx.x);
+  const uint32_t idx = row * W + col;
+
+  if (row >= H || col >= W)
+    return;
+
+  int32_t y = label[idx];
+  if (y > 0) {
+    int32_t count_idx = y - 1;
+    count_final[idx] = count_init[count_idx];
+  } else {
+    count_final[idx] = 0;
+  }
+}
+
+} // namespace cc2d
+
+std::vector<torch::Tensor> get_connected_componnets(
+    const torch::Tensor& inputs) {
+  AT_ASSERTM(inputs.is_cuda(), "inputs must be a CUDA tensor");
+  AT_ASSERTM(inputs.ndimension() == 4, "inputs must be [N, 1, H, W] shape");
+  AT_ASSERTM(
+      inputs.scalar_type() == torch::kUInt8, "inputs must be a uint8 type");
+
+  const uint32_t N = inputs.size(0);
+  const uint32_t C = inputs.size(1);
+  const uint32_t H = inputs.size(2);
+  const uint32_t W = inputs.size(3);
+
+  AT_ASSERTM(C == 1, "inputs must be [N, 1, H, W] shape");
+  AT_ASSERTM((H % 2) == 0, "height must be an even number");
+  AT_ASSERTM((W % 2) == 0, "width must be an even number");
+
+  // label must be uint32_t
+  auto label_options =
+      torch::TensorOptions().dtype(torch::kInt32).device(inputs.device());
+  torch::Tensor labels = torch::zeros({N, C, H, W}, label_options);
+  torch::Tensor counts_init = torch::zeros({N, C, H, W}, label_options);
+  torch::Tensor counts_final = torch::zeros({N, C, H, W}, label_options);
+
+  dim3 grid = dim3(
+      ((W + 1) / 2 + BLOCK_COLS - 1) / BLOCK_COLS,
+      ((H + 1) / 2 + BLOCK_ROWS - 1) / BLOCK_ROWS);
+  dim3 block = dim3(BLOCK_COLS, BLOCK_ROWS);
+  dim3 grid_count =
+      dim3((W + BLOCK_COLS) / BLOCK_COLS, (H + BLOCK_ROWS) / BLOCK_ROWS);
+  dim3 block_count = dim3(BLOCK_COLS, BLOCK_ROWS);
+  cudaStream_t stream = at::cuda::getCurrentCUDAStream();
+
+  for (int n = 0; n < N; n++) {
+    uint32_t offset = n * H * W;
+
+    cc2d::init_labeling<<<grid, block, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset, W, H);
+    cc2d::merge<<<grid, block, 0, stream>>>(
+        inputs.data_ptr<uint8_t>() + offset,
+        labels.data_ptr<int32_t>() + offset,
+        W,
+        H);
+    cc2d::compression<<<grid, block, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset, W, H);
+    cc2d::final_labeling<<<grid, block, 0, stream>>>(
+        inputs.data_ptr<uint8_t>() + offset,
+        labels.data_ptr<int32_t>() + offset,
+        W,
+        H);
+
+    // get the counting of each pixel
+    cc2d::init_counting<<<grid_count, block_count, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset,
+        counts_init.data_ptr<int32_t>() + offset,
+        W,
+        H);
+    cc2d::final_counting<<<grid_count, block_count, 0, stream>>>(
+        labels.data_ptr<int32_t>() + offset,
+        counts_init.data_ptr<int32_t>() + offset,
+        counts_final.data_ptr<int32_t>() + offset,
+        W,
+        H);
+  }
+
+  // returned values are [labels, counts]
+  std::vector<torch::Tensor> outputs;
+  outputs.push_back(labels);
+  outputs.push_back(counts_final);
+  return outputs;
+}
+
+PYBIND11_MODULE(TORCH_EXTENSION_NAME, m) {
+  m.def(
+      "get_connected_componnets",
+      &get_connected_componnets,
+      "get_connected_componnets");
+}
diff --git a/third_parts/sam2/modeling/__init__.py b/third_parts/sam2/modeling/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5277f46157403e47fd830fc519144b97ef69d4ae
--- /dev/null
+++ b/third_parts/sam2/modeling/__init__.py
@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
diff --git a/third_parts/sam2/modeling/backbones/__init__.py b/third_parts/sam2/modeling/backbones/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5277f46157403e47fd830fc519144b97ef69d4ae
--- /dev/null
+++ b/third_parts/sam2/modeling/backbones/__init__.py
@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
diff --git a/third_parts/sam2/modeling/backbones/hieradet.py b/third_parts/sam2/modeling/backbones/hieradet.py
new file mode 100644
index 0000000000000000000000000000000000000000..a7163dfb60404bb5e277c752f70b120511921612
--- /dev/null
+++ b/third_parts/sam2/modeling/backbones/hieradet.py
@@ -0,0 +1,295 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from functools import partial
+from typing import List, Tuple, Union
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from third_parts.sam2.modeling.backbones.utils import (
+    PatchEmbed,
+    window_partition,
+    window_unpartition,
+)
+
+from third_parts.sam2.modeling.sam2_utils import DropPath, MLP
+
+
+def do_pool(x: torch.Tensor, pool: nn.Module, norm: nn.Module = None) -> torch.Tensor:
+    if pool is None:
+        return x
+    # (B, H, W, C) -> (B, C, H, W)
+    x = x.permute(0, 3, 1, 2)
+    x = pool(x)
+    # (B, C, H', W') -> (B, H', W', C)
+    x = x.permute(0, 2, 3, 1)
+    if norm:
+        x = norm(x)
+
+    return x
+
+
+class MultiScaleAttention(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: int,
+        num_heads: int,
+        q_pool: nn.Module = None,
+    ):
+        super().__init__()
+
+        self.dim = dim
+        self.dim_out = dim_out
+
+        self.num_heads = num_heads
+        head_dim = dim_out // num_heads
+        self.scale = head_dim**-0.5
+
+        self.q_pool = q_pool
+        self.qkv = nn.Linear(dim, dim_out * 3)
+        self.proj = nn.Linear(dim_out, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        B, H, W, _ = x.shape
+        # qkv with shape (B, H * W, 3, nHead, C)
+        qkv = self.qkv(x).reshape(B, H * W, 3, self.num_heads, -1)
+        # q, k, v with shape (B, H * W, nheads, C)
+        q, k, v = torch.unbind(qkv, 2)
+
+        # Q pooling (for downsample at stage changes)
+        if self.q_pool:
+            q = do_pool(q.reshape(B, H, W, -1), self.q_pool)
+            H, W = q.shape[1:3]  # downsampled shape
+            q = q.reshape(B, H * W, self.num_heads, -1)
+
+        # Torch's SDPA expects [B, nheads, H*W, C] so we transpose
+        x = F.scaled_dot_product_attention(
+            q.transpose(1, 2),
+            k.transpose(1, 2),
+            v.transpose(1, 2),
+        )
+        # Transpose back
+        x = x.transpose(1, 2)
+        x = x.reshape(B, H, W, -1)
+
+        x = self.proj(x)
+
+        return x
+
+
+class MultiScaleBlock(nn.Module):
+    def __init__(
+        self,
+        dim: int,
+        dim_out: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        drop_path: float = 0.0,
+        norm_layer: Union[nn.Module, str] = "LayerNorm",
+        q_stride: Tuple[int, int] = None,
+        act_layer: nn.Module = nn.GELU,
+        window_size: int = 0,
+    ):
+        super().__init__()
+
+        if isinstance(norm_layer, str):
+            norm_layer = partial(getattr(nn, norm_layer), eps=1e-6)
+
+        self.dim = dim
+        self.dim_out = dim_out
+        self.norm1 = norm_layer(dim)
+
+        self.window_size = window_size
+
+        self.pool, self.q_stride = None, q_stride
+        if self.q_stride:
+            self.pool = nn.MaxPool2d(
+                kernel_size=q_stride, stride=q_stride, ceil_mode=False
+            )
+
+        self.attn = MultiScaleAttention(
+            dim,
+            dim_out,
+            num_heads=num_heads,
+            q_pool=self.pool,
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+        self.norm2 = norm_layer(dim_out)
+        self.mlp = MLP(
+            dim_out,
+            int(dim_out * mlp_ratio),
+            dim_out,
+            num_layers=2,
+            activation=act_layer,
+        )
+
+        if dim != dim_out:
+            self.proj = nn.Linear(dim, dim_out)
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        shortcut = x  # B, H, W, C
+        x = self.norm1(x)
+
+        # Skip connection
+        if self.dim != self.dim_out:
+            shortcut = do_pool(self.proj(x), self.pool)
+
+        # Window partition
+        window_size = self.window_size
+        if window_size > 0:
+            H, W = x.shape[1], x.shape[2]
+            x, pad_hw = window_partition(x, window_size)
+
+        # Window Attention + Q Pooling (if stage change)
+        x = self.attn(x)
+        if self.q_stride:
+            # Shapes have changed due to Q pooling
+            window_size = self.window_size // self.q_stride[0]
+            H, W = shortcut.shape[1:3]
+
+            pad_h = (window_size - H % window_size) % window_size
+            pad_w = (window_size - W % window_size) % window_size
+            pad_hw = (H + pad_h, W + pad_w)
+
+        # Reverse window partition
+        if self.window_size > 0:
+            x = window_unpartition(x, window_size, pad_hw, (H, W))
+
+        x = shortcut + self.drop_path(x)
+        # MLP
+        x = x + self.drop_path(self.mlp(self.norm2(x)))
+        return x
+
+
+class Hiera(nn.Module):
+    """
+    Reference: https://arxiv.org/abs/2306.00989
+    """
+
+    def __init__(
+        self,
+        embed_dim: int = 96,  # initial embed dim
+        num_heads: int = 1,  # initial number of heads
+        drop_path_rate: float = 0.0,  # stochastic depth
+        q_pool: int = 3,  # number of q_pool stages
+        q_stride: Tuple[int, int] = (2, 2),  # downsample stride bet. stages
+        stages: Tuple[int, ...] = (2, 3, 16, 3),  # blocks per stage
+        dim_mul: float = 2.0,  # dim_mul factor at stage shift
+        head_mul: float = 2.0,  # head_mul factor at stage shift
+        window_pos_embed_bkg_spatial_size: Tuple[int, int] = (14, 14),
+        # window size per stage, when not using global att.
+        window_spec: Tuple[int, ...] = (
+            8,
+            4,
+            14,
+            7,
+        ),
+        # global attn in these blocks
+        global_att_blocks: Tuple[int, ...] = (
+            12,
+            16,
+            20,
+        ),
+        return_interm_layers=True,  # return feats from every stage
+    ):
+        super().__init__()
+
+        assert len(stages) == len(window_spec)
+        self.window_spec = window_spec
+
+        depth = sum(stages)
+        self.q_stride = q_stride
+        self.stage_ends = [sum(stages[:i]) - 1 for i in range(1, len(stages) + 1)]
+        assert 0 <= q_pool <= len(self.stage_ends[:-1])
+        self.q_pool_blocks = [x + 1 for x in self.stage_ends[:-1]][:q_pool]
+        self.return_interm_layers = return_interm_layers
+
+        self.patch_embed = PatchEmbed(
+            embed_dim=embed_dim,
+        )
+        # Which blocks have global att?
+        self.global_att_blocks = global_att_blocks
+
+        # Windowed positional embedding (https://arxiv.org/abs/2311.05613)
+        self.window_pos_embed_bkg_spatial_size = window_pos_embed_bkg_spatial_size
+        self.pos_embed = nn.Parameter(
+            torch.zeros(1, embed_dim, *self.window_pos_embed_bkg_spatial_size)
+        )
+        self.pos_embed_window = nn.Parameter(
+            torch.zeros(1, embed_dim, self.window_spec[0], self.window_spec[0])
+        )
+
+        dpr = [
+            x.item() for x in torch.linspace(0, drop_path_rate, depth)
+        ]  # stochastic depth decay rule
+
+        cur_stage = 1
+        self.blocks = nn.ModuleList()
+
+        for i in range(depth):
+            dim_out = embed_dim
+            # lags by a block, so first block of
+            # next stage uses an initial window size
+            # of previous stage and final window size of current stage
+            window_size = self.window_spec[cur_stage - 1]
+
+            if self.global_att_blocks is not None:
+                window_size = 0 if i in self.global_att_blocks else window_size
+
+            if i - 1 in self.stage_ends:
+                dim_out = int(embed_dim * dim_mul)
+                num_heads = int(num_heads * head_mul)
+                cur_stage += 1
+
+            block = MultiScaleBlock(
+                dim=embed_dim,
+                dim_out=dim_out,
+                num_heads=num_heads,
+                drop_path=dpr[i],
+                q_stride=self.q_stride if i in self.q_pool_blocks else None,
+                window_size=window_size,
+            )
+
+            embed_dim = dim_out
+            self.blocks.append(block)
+
+        self.channel_list = (
+            [self.blocks[i].dim_out for i in self.stage_ends[::-1]]
+            if return_interm_layers
+            else [self.blocks[-1].dim_out]
+        )
+
+    def _get_pos_embed(self, hw: Tuple[int, int]) -> torch.Tensor:
+        h, w = hw
+        window_embed = self.pos_embed_window
+        pos_embed = F.interpolate(self.pos_embed, size=(h, w), mode="bicubic")
+        pos_embed = pos_embed + window_embed.tile(
+            [x // y for x, y in zip(pos_embed.shape, window_embed.shape)]
+        )
+        pos_embed = pos_embed.permute(0, 2, 3, 1)
+        return pos_embed
+
+    def forward(self, x: torch.Tensor) -> List[torch.Tensor]:
+        x = self.patch_embed(x)
+        # x: (B, H, W, C)
+
+        # Add pos embed
+        x = x + self._get_pos_embed(x.shape[1:3])
+
+        outputs = []
+        for i, blk in enumerate(self.blocks):
+            x = blk(x)
+            if (i == self.stage_ends[-1]) or (
+                i in self.stage_ends and self.return_interm_layers
+            ):
+                feats = x.permute(0, 3, 1, 2)
+                outputs.append(feats)
+
+        return outputs
diff --git a/third_parts/sam2/modeling/backbones/image_encoder.py b/third_parts/sam2/modeling/backbones/image_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..5f92baf47dcab96385ff99899fd3e3a642c1cf9c
--- /dev/null
+++ b/third_parts/sam2/modeling/backbones/image_encoder.py
@@ -0,0 +1,133 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+class ImageEncoder(nn.Module):
+    def __init__(
+        self,
+        trunk: nn.Module,
+        neck: nn.Module,
+        scalp: int = 0,
+    ):
+        super().__init__()
+        self.trunk = trunk
+        self.neck = neck
+        self.scalp = scalp
+        assert (
+            self.trunk.channel_list == self.neck.backbone_channel_list
+        ), f"Channel dims of trunk and neck do not match. Trunk: {self.trunk.channel_list}, neck: {self.neck.backbone_channel_list}"
+
+    def forward(self, sample: torch.Tensor):
+        # Forward through backbone
+        features, pos = self.neck(self.trunk(sample))
+        if self.scalp > 0:
+            # Discard the lowest resolution features
+            features, pos = features[: -self.scalp], pos[: -self.scalp]
+
+        src = features[-1]
+        output = {
+            "vision_features": src,
+            "vision_pos_enc": pos,
+            "backbone_fpn": features,
+        }
+        return output
+
+
+class FpnNeck(nn.Module):
+    """
+    A modified variant of Feature Pyramid Network (FPN) neck
+    (we remove output conv and also do bicubic interpolation similar to ViT
+    pos embed interpolation)
+    """
+
+    def __init__(
+        self,
+        position_encoding: nn.Module,
+        d_model: int,
+        backbone_channel_list: List[int],
+        kernel_size: int = 1,
+        stride: int = 1,
+        padding: int = 0,
+        fpn_interp_model: str = "bilinear",
+        fuse_type: str = "sum",
+        fpn_top_down_levels: Optional[List[int]] = None,
+    ):
+        """Initialize the neck
+        :param trunk: the backbone
+        :param position_encoding: the positional encoding to use
+        :param d_model: the dimension of the model
+        :param neck_norm: the normalization to use
+        """
+        super().__init__()
+        self.position_encoding = position_encoding
+        self.convs = nn.ModuleList()
+        self.backbone_channel_list = backbone_channel_list
+        for dim in backbone_channel_list:
+            current = nn.Sequential()
+            current.add_module(
+                "conv",
+                nn.Conv2d(
+                    in_channels=dim,
+                    out_channels=d_model,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                ),
+            )
+
+            self.convs.append(current)
+        self.fpn_interp_model = fpn_interp_model
+        assert fuse_type in ["sum", "avg"]
+        self.fuse_type = fuse_type
+
+        # levels to have top-down features in its outputs
+        # e.g. if fpn_top_down_levels is [2, 3], then only outputs of level 2 and 3
+        # have top-down propagation, while outputs of level 0 and level 1 have only
+        # lateral features from the same backbone level.
+        if fpn_top_down_levels is None:
+            # default is to have top-down features on all levels
+            fpn_top_down_levels = range(len(self.convs))
+        self.fpn_top_down_levels = list(fpn_top_down_levels)
+
+    def forward(self, xs: List[torch.Tensor]):
+
+        out = [None] * len(self.convs)
+        pos = [None] * len(self.convs)
+        assert len(xs) == len(self.convs)
+        # fpn forward pass
+        # see https://github.com/facebookresearch/detectron2/blob/main/detectron2/modeling/backbone/fpn.py
+        prev_features = None
+        # forward in top-down order (from low to high resolution)
+        n = len(self.convs) - 1
+        for i in range(n, -1, -1):
+            x = xs[i]
+            lateral_features = self.convs[n - i](x)
+            if i in self.fpn_top_down_levels and prev_features is not None:
+                top_down_features = F.interpolate(
+                    prev_features.to(dtype=torch.float32),
+                    scale_factor=2.0,
+                    mode=self.fpn_interp_model,
+                    align_corners=(
+                        None if self.fpn_interp_model == "nearest" else False
+                    ),
+                    antialias=False,
+                )
+                prev_features = lateral_features + top_down_features
+                if self.fuse_type == "avg":
+                    prev_features /= 2
+            else:
+                prev_features = lateral_features
+            x_out = prev_features
+            out[i] = x_out
+            pos[i] = self.position_encoding(x_out).to(x_out.dtype)
+
+        return out, pos
diff --git a/third_parts/sam2/modeling/backbones/utils.py b/third_parts/sam2/modeling/backbones/utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..32d55c7545f064de133a5ff0200ba1ece9b504b7
--- /dev/null
+++ b/third_parts/sam2/modeling/backbones/utils.py
@@ -0,0 +1,95 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+"""Some utilities for backbones, in particular for windowing"""
+
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def window_partition(x, window_size):
+    """
+    Partition into non-overlapping windows with padding if needed.
+    Args:
+        x (tensor): input tokens with [B, H, W, C].
+        window_size (int): window size.
+    Returns:
+        windows: windows after partition with [B * num_windows, window_size, window_size, C].
+        (Hp, Wp): padded height and width before partition
+    """
+    B, H, W, C = x.shape
+
+    pad_h = (window_size - H % window_size) % window_size
+    pad_w = (window_size - W % window_size) % window_size
+    if pad_h > 0 or pad_w > 0:
+        x = F.pad(x, (0, 0, 0, pad_w, 0, pad_h))
+    Hp, Wp = H + pad_h, W + pad_w
+
+    x = x.view(B, Hp // window_size, window_size, Wp // window_size, window_size, C)
+    windows = (
+        x.permute(0, 1, 3, 2, 4, 5).contiguous().view(-1, window_size, window_size, C)
+    )
+    return windows, (Hp, Wp)
+
+
+def window_unpartition(windows, window_size, pad_hw, hw):
+    """
+    Window unpartition into original sequences and removing padding.
+    Args:
+        x (tensor): input tokens with [B * num_windows, window_size, window_size, C].
+        window_size (int): window size.
+        pad_hw (Tuple): padded height and width (Hp, Wp).
+        hw (Tuple): original height and width (H, W) before padding.
+    Returns:
+        x: unpartitioned sequences with [B, H, W, C].
+    """
+    Hp, Wp = pad_hw
+    H, W = hw
+    B = windows.shape[0] // (Hp * Wp // window_size // window_size)
+    x = windows.view(
+        B, Hp // window_size, Wp // window_size, window_size, window_size, -1
+    )
+    x = x.permute(0, 1, 3, 2, 4, 5).contiguous().view(B, Hp, Wp, -1)
+
+    if Hp > H or Wp > W:
+        x = x[:, :H, :W, :].contiguous()
+    return x
+
+
+class PatchEmbed(nn.Module):
+    """
+    Image to Patch Embedding.
+    """
+
+    def __init__(
+        self,
+        kernel_size: Tuple[int, ...] = (7, 7),
+        stride: Tuple[int, ...] = (4, 4),
+        padding: Tuple[int, ...] = (3, 3),
+        in_chans: int = 3,
+        embed_dim: int = 768,
+    ):
+        """
+        Args:
+            kernel_size (Tuple): kernel size of the projection layer.
+            stride (Tuple): stride of the projection layer.
+            padding (Tuple): padding size of the projection layer.
+            in_chans (int): Number of input image channels.
+            embed_dim (int):  embed_dim (int): Patch embedding dimension.
+        """
+        super().__init__()
+        self.proj = nn.Conv2d(
+            in_chans, embed_dim, kernel_size=kernel_size, stride=stride, padding=padding
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        x = self.proj(x)
+        # B C H W -> B H W C
+        x = x.permute(0, 2, 3, 1)
+        return x
diff --git a/third_parts/sam2/modeling/memory_attention.py b/third_parts/sam2/modeling/memory_attention.py
new file mode 100644
index 0000000000000000000000000000000000000000..64097aed192b180cc37345cd3b3819e68257168e
--- /dev/null
+++ b/third_parts/sam2/modeling/memory_attention.py
@@ -0,0 +1,169 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional
+
+import torch
+from torch import nn, Tensor
+
+from third_parts.sam2.modeling.sam.transformer import RoPEAttention
+
+from third_parts.sam2.modeling.sam2_utils import get_activation_fn, get_clones
+
+
+class MemoryAttentionLayer(nn.Module):
+
+    def __init__(
+        self,
+        activation: str,
+        cross_attention: nn.Module,
+        d_model: int,
+        dim_feedforward: int,
+        dropout: float,
+        pos_enc_at_attn: bool,
+        pos_enc_at_cross_attn_keys: bool,
+        pos_enc_at_cross_attn_queries: bool,
+        self_attention: nn.Module,
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.dim_feedforward = dim_feedforward
+        self.dropout_value = dropout
+        self.self_attn = self_attention
+        self.cross_attn_image = cross_attention
+
+        # Implementation of Feedforward model
+        self.linear1 = nn.Linear(d_model, dim_feedforward)
+        self.dropout = nn.Dropout(dropout)
+        self.linear2 = nn.Linear(dim_feedforward, d_model)
+
+        self.norm1 = nn.LayerNorm(d_model)
+        self.norm2 = nn.LayerNorm(d_model)
+        self.norm3 = nn.LayerNorm(d_model)
+        self.dropout1 = nn.Dropout(dropout)
+        self.dropout2 = nn.Dropout(dropout)
+        self.dropout3 = nn.Dropout(dropout)
+
+        self.activation_str = activation
+        self.activation = get_activation_fn(activation)
+
+        # Where to add pos enc
+        self.pos_enc_at_attn = pos_enc_at_attn
+        self.pos_enc_at_cross_attn_queries = pos_enc_at_cross_attn_queries
+        self.pos_enc_at_cross_attn_keys = pos_enc_at_cross_attn_keys
+
+    def _forward_sa(self, tgt, query_pos):
+        # Self-Attention
+        tgt2 = self.norm1(tgt)
+        q = k = tgt2 + query_pos if self.pos_enc_at_attn else tgt2
+        tgt2 = self.self_attn(q, k, v=tgt2)
+        tgt = tgt + self.dropout1(tgt2)
+        return tgt
+
+    def _forward_ca(self, tgt, memory, query_pos, pos, num_k_exclude_rope=0):
+        kwds = {}
+        if num_k_exclude_rope > 0:
+            assert isinstance(self.cross_attn_image, RoPEAttention)
+            kwds = {"num_k_exclude_rope": num_k_exclude_rope}
+
+        # Cross-Attention
+        tgt2 = self.norm2(tgt)
+        tgt2 = self.cross_attn_image(
+            q=tgt2 + query_pos if self.pos_enc_at_cross_attn_queries else tgt2,
+            k=memory + pos if self.pos_enc_at_cross_attn_keys else memory,
+            v=memory,
+            **kwds,
+        )
+        tgt = tgt + self.dropout2(tgt2)
+        return tgt
+
+    def forward(
+        self,
+        tgt,
+        memory,
+        pos: Optional[Tensor] = None,
+        query_pos: Optional[Tensor] = None,
+        num_k_exclude_rope: int = 0,
+    ) -> torch.Tensor:
+
+        # Self-Attn, Cross-Attn
+        tgt = self._forward_sa(tgt, query_pos)
+        tgt = self._forward_ca(tgt, memory, query_pos, pos, num_k_exclude_rope)
+        # MLP
+        tgt2 = self.norm3(tgt)
+        tgt2 = self.linear2(self.dropout(self.activation(self.linear1(tgt2))))
+        tgt = tgt + self.dropout3(tgt2)
+        return tgt
+
+
+class MemoryAttention(nn.Module):
+    def __init__(
+        self,
+        d_model: int,
+        pos_enc_at_input: bool,
+        layer: nn.Module,
+        num_layers: int,
+        batch_first: bool = True,  # Do layers expect batch first input?
+    ):
+        super().__init__()
+        self.d_model = d_model
+        self.layers = get_clones(layer, num_layers)
+        self.num_layers = num_layers
+        self.norm = nn.LayerNorm(d_model)
+        self.pos_enc_at_input = pos_enc_at_input
+        self.batch_first = batch_first
+
+    def forward(
+        self,
+        curr: torch.Tensor,  # self-attention inputs
+        memory: torch.Tensor,  # cross-attention inputs
+        curr_pos: Optional[Tensor] = None,  # pos_enc for self-attention inputs
+        memory_pos: Optional[Tensor] = None,  # pos_enc for cross-attention inputs
+        num_obj_ptr_tokens: int = 0,  # number of object pointer *tokens*
+    ):
+        if isinstance(curr, list):
+            assert isinstance(curr_pos, list)
+            assert len(curr) == len(curr_pos) == 1
+            curr, curr_pos = (
+                curr[0],
+                curr_pos[0],
+            )
+
+        assert (
+            curr.shape[1] == memory.shape[1]
+        ), "Batch size must be the same for curr and memory"
+
+        output = curr
+        if self.pos_enc_at_input and curr_pos is not None:
+            output = output + 0.1 * curr_pos
+
+        if self.batch_first:
+            # Convert to batch first
+            output = output.transpose(0, 1)
+            curr_pos = curr_pos.transpose(0, 1)
+            memory = memory.transpose(0, 1)
+            memory_pos = memory_pos.transpose(0, 1)
+
+        for layer in self.layers:
+            kwds = {}
+            if isinstance(layer.cross_attn_image, RoPEAttention):
+                kwds = {"num_k_exclude_rope": num_obj_ptr_tokens}
+
+            output = layer(
+                tgt=output,
+                memory=memory,
+                pos=memory_pos,
+                query_pos=curr_pos,
+                **kwds,
+            )
+        normed_output = self.norm(output)
+
+        if self.batch_first:
+            # Convert back to seq first
+            normed_output = normed_output.transpose(0, 1)
+            curr_pos = curr_pos.transpose(0, 1)
+
+        return normed_output
diff --git a/third_parts/sam2/modeling/memory_encoder.py b/third_parts/sam2/modeling/memory_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..4c61d87fcc60c6005f2c98e24892389611e30deb
--- /dev/null
+++ b/third_parts/sam2/modeling/memory_encoder.py
@@ -0,0 +1,181 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Tuple
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from third_parts.sam2.modeling.sam2_utils import DropPath, get_clones, LayerNorm2d
+
+
+class MaskDownSampler(nn.Module):
+    """
+    Progressively downsample a mask by total_stride, each time by stride.
+    Note that LayerNorm is applied per *token*, like in ViT.
+
+    With each downsample (by a factor stride**2), channel capacity increases by the same factor.
+    In the end, we linearly project to embed_dim channels.
+    """
+
+    def __init__(
+        self,
+        embed_dim=256,
+        kernel_size=4,
+        stride=4,
+        padding=0,
+        total_stride=16,
+        activation=nn.GELU,
+    ):
+        super().__init__()
+        num_layers = int(math.log2(total_stride) // math.log2(stride))
+        assert stride**num_layers == total_stride
+        self.encoder = nn.Sequential()
+        mask_in_chans, mask_out_chans = 1, 1
+        for _ in range(num_layers):
+            mask_out_chans = mask_in_chans * (stride**2)
+            self.encoder.append(
+                nn.Conv2d(
+                    mask_in_chans,
+                    mask_out_chans,
+                    kernel_size=kernel_size,
+                    stride=stride,
+                    padding=padding,
+                )
+            )
+            self.encoder.append(LayerNorm2d(mask_out_chans))
+            self.encoder.append(activation())
+            mask_in_chans = mask_out_chans
+
+        self.encoder.append(nn.Conv2d(mask_out_chans, embed_dim, kernel_size=1))
+
+    def forward(self, x):
+        return self.encoder(x)
+
+
+# Lightly adapted from ConvNext (https://github.com/facebookresearch/ConvNeXt)
+class CXBlock(nn.Module):
+    r"""ConvNeXt Block. There are two equivalent implementations:
+    (1) DwConv -> LayerNorm (channels_first) -> 1x1 Conv -> GELU -> 1x1 Conv; all in (N, C, H, W)
+    (2) DwConv -> Permute to (N, H, W, C); LayerNorm (channels_last) -> Linear -> GELU -> Linear; Permute back
+    We use (2) as we find it slightly faster in PyTorch
+
+    Args:
+        dim (int): Number of input channels.
+        drop_path (float): Stochastic depth rate. Default: 0.0
+        layer_scale_init_value (float): Init value for Layer Scale. Default: 1e-6.
+    """
+
+    def __init__(
+        self,
+        dim,
+        kernel_size=7,
+        padding=3,
+        drop_path=0.0,
+        layer_scale_init_value=1e-6,
+        use_dwconv=True,
+    ):
+        super().__init__()
+        self.dwconv = nn.Conv2d(
+            dim,
+            dim,
+            kernel_size=kernel_size,
+            padding=padding,
+            groups=dim if use_dwconv else 1,
+        )  # depthwise conv
+        self.norm = LayerNorm2d(dim, eps=1e-6)
+        self.pwconv1 = nn.Linear(
+            dim, 4 * dim
+        )  # pointwise/1x1 convs, implemented with linear layers
+        self.act = nn.GELU()
+        self.pwconv2 = nn.Linear(4 * dim, dim)
+        self.gamma = (
+            nn.Parameter(layer_scale_init_value * torch.ones((dim)), requires_grad=True)
+            if layer_scale_init_value > 0
+            else None
+        )
+        self.drop_path = DropPath(drop_path) if drop_path > 0.0 else nn.Identity()
+
+    def forward(self, x):
+        input = x
+        x = self.dwconv(x)
+        x = self.norm(x)
+        x = x.permute(0, 2, 3, 1)  # (N, C, H, W) -> (N, H, W, C)
+        x = self.pwconv1(x)
+        x = self.act(x)
+        x = self.pwconv2(x)
+        if self.gamma is not None:
+            x = self.gamma * x
+        x = x.permute(0, 3, 1, 2)  # (N, H, W, C) -> (N, C, H, W)
+
+        x = input + self.drop_path(x)
+        return x
+
+
+class Fuser(nn.Module):
+    def __init__(self, layer, num_layers, dim=None, input_projection=False):
+        super().__init__()
+        self.proj = nn.Identity()
+        self.layers = get_clones(layer, num_layers)
+
+        if input_projection:
+            assert dim is not None
+            self.proj = nn.Conv2d(dim, dim, kernel_size=1)
+
+    def forward(self, x):
+        # normally x: (N, C, H, W)
+        x = self.proj(x)
+        for layer in self.layers:
+            x = layer(x)
+        return x
+
+
+class MemoryEncoder(nn.Module):
+    def __init__(
+        self,
+        out_dim,
+        mask_downsampler,
+        fuser,
+        position_encoding,
+        in_dim=256,  # in_dim of pix_feats
+    ):
+        super().__init__()
+
+        self.mask_downsampler = mask_downsampler
+
+        self.pix_feat_proj = nn.Conv2d(in_dim, in_dim, kernel_size=1)
+        self.fuser = fuser
+        self.position_encoding = position_encoding
+        self.out_proj = nn.Identity()
+        if out_dim != in_dim:
+            self.out_proj = nn.Conv2d(in_dim, out_dim, kernel_size=1)
+
+    def forward(
+        self,
+        pix_feat: torch.Tensor,
+        masks: torch.Tensor,
+        skip_mask_sigmoid: bool = False,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        ## Process masks
+        # sigmoid, so that less domain shift from gt masks which are bool
+        if not skip_mask_sigmoid:
+            masks = F.sigmoid(masks)
+        masks = self.mask_downsampler(masks)
+
+        ## Fuse pix_feats and downsampled masks
+        # in case the visual features are on CPU, cast them to CUDA
+        pix_feat = pix_feat.to(masks.device)
+
+        x = self.pix_feat_proj(pix_feat)
+        x = x + masks
+        x = self.fuser(x)
+        x = self.out_proj(x)
+
+        pos = self.position_encoding(x).to(x.dtype)
+
+        return {"vision_features": x, "vision_pos_enc": [pos]}
diff --git a/third_parts/sam2/modeling/position_encoding.py b/third_parts/sam2/modeling/position_encoding.py
new file mode 100644
index 0000000000000000000000000000000000000000..85dc1e375279f8bcacb8652f205ab41af0bb21c3
--- /dev/null
+++ b/third_parts/sam2/modeling/position_encoding.py
@@ -0,0 +1,221 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from typing import Any, Optional, Tuple
+
+import numpy as np
+
+import torch
+from torch import nn
+
+
+class PositionEmbeddingSine(nn.Module):
+    """
+    This is a more standard version of the position embedding, very similar to the one
+    used by the Attention is all you need paper, generalized to work on images.
+    """
+
+    def __init__(
+        self,
+        num_pos_feats,
+        temperature: int = 10000,
+        normalize: bool = True,
+        scale: Optional[float] = None,
+    ):
+        super().__init__()
+        assert num_pos_feats % 2 == 0, "Expecting even model width"
+        self.num_pos_feats = num_pos_feats // 2
+        self.temperature = temperature
+        self.normalize = normalize
+        if scale is not None and normalize is False:
+            raise ValueError("normalize should be True if scale is passed")
+        if scale is None:
+            scale = 2 * math.pi
+        self.scale = scale
+
+        self.cache = {}
+
+    def _encode_xy(self, x, y):
+        # The positions are expected to be normalized
+        assert len(x) == len(y) and x.ndim == y.ndim == 1
+        x_embed = x * self.scale
+        y_embed = y * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, None] / dim_t
+        pos_y = y_embed[:, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, 0::2].sin(), pos_x[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        pos_y = torch.stack(
+            (pos_y[:, 0::2].sin(), pos_y[:, 1::2].cos()), dim=2
+        ).flatten(1)
+        return pos_x, pos_y
+
+    @torch.no_grad()
+    def encode_boxes(self, x, y, w, h):
+        pos_x, pos_y = self._encode_xy(x, y)
+        pos = torch.cat((pos_y, pos_x, h[:, None], w[:, None]), dim=1)
+        return pos
+
+    encode = encode_boxes  # Backwards compatibility
+
+    @torch.no_grad()
+    def encode_points(self, x, y, labels):
+        (bx, nx), (by, ny), (bl, nl) = x.shape, y.shape, labels.shape
+        assert bx == by and nx == ny and bx == bl and nx == nl
+        pos_x, pos_y = self._encode_xy(x.flatten(), y.flatten())
+        pos_x, pos_y = pos_x.reshape(bx, nx, -1), pos_y.reshape(by, ny, -1)
+        pos = torch.cat((pos_y, pos_x, labels[:, :, None]), dim=2)
+        return pos
+
+    @torch.no_grad()
+    def forward(self, x: torch.Tensor):
+        cache_key = (x.shape[-2], x.shape[-1])
+        if cache_key in self.cache:
+            return self.cache[cache_key][None].repeat(x.shape[0], 1, 1, 1)
+        y_embed = (
+            torch.arange(1, x.shape[-2] + 1, dtype=torch.float32, device=x.device)
+            .view(1, -1, 1)
+            .repeat(x.shape[0], 1, x.shape[-1])
+        )
+        x_embed = (
+            torch.arange(1, x.shape[-1] + 1, dtype=torch.float32, device=x.device)
+            .view(1, 1, -1)
+            .repeat(x.shape[0], x.shape[-2], 1)
+        )
+
+        if self.normalize:
+            eps = 1e-6
+            y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self.scale
+            x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self.scale
+
+        dim_t = torch.arange(self.num_pos_feats, dtype=torch.float32, device=x.device)
+        dim_t = self.temperature ** (2 * (dim_t // 2) / self.num_pos_feats)
+
+        pos_x = x_embed[:, :, :, None] / dim_t
+        pos_y = y_embed[:, :, :, None] / dim_t
+        pos_x = torch.stack(
+            (pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos_y = torch.stack(
+            (pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4
+        ).flatten(3)
+        pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2)
+        self.cache[cache_key] = pos[0]
+        return pos
+
+
+class PositionEmbeddingRandom(nn.Module):
+    """
+    Positional encoding using random spatial frequencies.
+    """
+
+    def __init__(self, num_pos_feats: int = 64, scale: Optional[float] = None) -> None:
+        super().__init__()
+        if scale is None or scale <= 0.0:
+            scale = 1.0
+        self.register_buffer(
+            "positional_encoding_gaussian_matrix",
+            scale * torch.randn((2, num_pos_feats)),
+        )
+        self.first = True
+
+    def _pe_encoding(self, coords: torch.Tensor) -> torch.Tensor:
+        """Positionally encode points that are normalized to [0,1]."""
+        # assuming coords are in [0, 1]^2 square and have d_1 x ... x d_n x 2 shape
+        coords = 2 * coords - 1
+        coords = coords.to(self.positional_encoding_gaussian_matrix.dtype)
+        if self.first:
+            self.positional_encoding_gaussian_matrix = self.positional_encoding_gaussian_matrix.to(coords.device)
+            self.first = False
+        coords = coords @ self.positional_encoding_gaussian_matrix
+        coords = 2 * np.pi * coords
+        # outputs d_1 x ... x d_n x C shape
+        return torch.cat([torch.sin(coords), torch.cos(coords)], dim=-1)
+
+    def forward(self, size: Tuple[int, int]) -> torch.Tensor:
+        """Generate positional encoding for a grid of the specified size."""
+        h, w = size
+        device: Any = self.positional_encoding_gaussian_matrix.device
+        grid = torch.ones((h, w), device=device, dtype=torch.float32)
+        y_embed = grid.cumsum(dim=0) - 0.5
+        x_embed = grid.cumsum(dim=1) - 0.5
+        y_embed = y_embed / h
+        x_embed = x_embed / w
+
+        pe = self._pe_encoding(torch.stack([x_embed, y_embed], dim=-1))
+        return pe.permute(2, 0, 1)  # C x H x W
+
+    def forward_with_coords(
+        self, coords_input: torch.Tensor, image_size: Tuple[int, int]
+    ) -> torch.Tensor:
+        """Positionally encode points that are not normalized to [0,1]."""
+        coords = coords_input.clone()
+        coords[:, :, 0] = coords[:, :, 0] / image_size[1]
+        coords[:, :, 1] = coords[:, :, 1] / image_size[0]
+        return self._pe_encoding(coords.to(torch.float))  # B x N x C
+
+
+# Rotary Positional Encoding, adapted from:
+# 1. https://github.com/meta-llama/codellama/blob/main/llama/model.py
+# 2. https://github.com/naver-ai/rope-vit
+# 3. https://github.com/lucidrains/rotary-embedding-torch
+
+
+def init_t_xy(end_x: int, end_y: int):
+    t = torch.arange(end_x * end_y, dtype=torch.float32)
+    t_x = (t % end_x).float()
+    t_y = torch.div(t, end_x, rounding_mode="floor").float()
+    return t_x, t_y
+
+
+def compute_axial_cis(dim: int, end_x: int, end_y: int, theta: float = 10000.0):
+    freqs_x = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+    freqs_y = 1.0 / (theta ** (torch.arange(0, dim, 4)[: (dim // 4)].float() / dim))
+
+    t_x, t_y = init_t_xy(end_x, end_y)
+    freqs_x = torch.outer(t_x, freqs_x)
+    freqs_y = torch.outer(t_y, freqs_y)
+    freqs_cis_x = torch.polar(torch.ones_like(freqs_x), freqs_x)
+    freqs_cis_y = torch.polar(torch.ones_like(freqs_y), freqs_y)
+    return torch.cat([freqs_cis_x, freqs_cis_y], dim=-1)
+
+
+def reshape_for_broadcast(freqs_cis: torch.Tensor, x: torch.Tensor):
+    ndim = x.ndim
+    assert 0 <= 1 < ndim
+    assert freqs_cis.shape == (x.shape[-2], x.shape[-1])
+    shape = [d if i >= ndim - 2 else 1 for i, d in enumerate(x.shape)]
+    return freqs_cis.view(*shape)
+
+
+def apply_rotary_enc(
+    xq: torch.Tensor,
+    xk: torch.Tensor,
+    freqs_cis: torch.Tensor,
+    repeat_freqs_k: bool = False,
+):
+    xq_ = torch.view_as_complex(xq.float().reshape(*xq.shape[:-1], -1, 2))
+    xk_ = (
+        torch.view_as_complex(xk.float().reshape(*xk.shape[:-1], -1, 2))
+        if xk.shape[-2] != 0
+        else None
+    )
+    freqs_cis = reshape_for_broadcast(freqs_cis, xq_)
+    xq_out = torch.view_as_real(xq_ * freqs_cis).flatten(3)
+    if xk_ is None:
+        # no keys to rotate, due to dropout
+        return xq_out.type_as(xq).to(xq.device), xk
+    # repeat freqs along seq_len dim to match k seq_len
+    if repeat_freqs_k:
+        r = xk_.shape[-2] // xq_.shape[-2]
+        freqs_cis = freqs_cis.repeat(*([1] * (freqs_cis.ndim - 2)), r, 1)
+    xk_out = torch.view_as_real(xk_ * freqs_cis).flatten(3)
+    return xq_out.type_as(xq).to(xq.device), xk_out.type_as(xk).to(xk.device)
diff --git a/third_parts/sam2/modeling/sam/__init__.py b/third_parts/sam2/modeling/sam/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5277f46157403e47fd830fc519144b97ef69d4ae
--- /dev/null
+++ b/third_parts/sam2/modeling/sam/__init__.py
@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
diff --git a/third_parts/sam2/modeling/sam/mask_decoder.py b/third_parts/sam2/modeling/sam/mask_decoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..55825cd6ede175ff327da6fa8e7627cfb8979f58
--- /dev/null
+++ b/third_parts/sam2/modeling/sam/mask_decoder.py
@@ -0,0 +1,299 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import List, Optional, Tuple, Type
+
+import torch
+from torch import nn
+
+from third_parts.sam2.modeling.sam2_utils import LayerNorm2d, MLP
+
+
+class MaskDecoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        transformer_dim: int,
+        transformer: nn.Module,
+        num_multimask_outputs: int = 3,
+        activation: Type[nn.Module] = nn.GELU,
+        iou_head_depth: int = 3,
+        iou_head_hidden_dim: int = 256,
+        use_high_res_features: bool = False,
+        iou_prediction_use_sigmoid=False,
+        dynamic_multimask_via_stability=False,
+        dynamic_multimask_stability_delta=0.05,
+        dynamic_multimask_stability_thresh=0.98,
+        pred_obj_scores: bool = False,
+        pred_obj_scores_mlp: bool = False,
+        use_multimask_token_for_obj_ptr: bool = False,
+    ) -> None:
+        """
+        Predicts masks given an image and prompt embeddings, using a
+        transformer architecture.
+
+        Arguments:
+          transformer_dim (int): the channel dimension of the transformer
+          transformer (nn.Module): the transformer used to predict masks
+          num_multimask_outputs (int): the number of masks to predict
+            when disambiguating masks
+          activation (nn.Module): the type of activation to use when
+            upscaling masks
+          iou_head_depth (int): the depth of the MLP used to predict
+            mask quality
+          iou_head_hidden_dim (int): the hidden dimension of the MLP
+            used to predict mask quality
+        """
+        super().__init__()
+        self.transformer_dim = transformer_dim
+        self.transformer = transformer
+
+        self.num_multimask_outputs = num_multimask_outputs
+
+        self.iou_token = nn.Embedding(1, transformer_dim)
+        self.num_mask_tokens = num_multimask_outputs + 1
+        self.mask_tokens = nn.Embedding(self.num_mask_tokens, transformer_dim)
+
+        self.pred_obj_scores = pred_obj_scores
+        if self.pred_obj_scores:
+            self.obj_score_token = nn.Embedding(1, transformer_dim)
+        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
+
+        self.output_upscaling = nn.Sequential(
+            nn.ConvTranspose2d(
+                transformer_dim, transformer_dim // 4, kernel_size=2, stride=2
+            ),
+            LayerNorm2d(transformer_dim // 4),
+            activation(),
+            nn.ConvTranspose2d(
+                transformer_dim // 4, transformer_dim // 8, kernel_size=2, stride=2
+            ),
+            activation(),
+        )
+        self.use_high_res_features = use_high_res_features
+        if use_high_res_features:
+            self.conv_s0 = nn.Conv2d(
+                transformer_dim, transformer_dim // 8, kernel_size=1, stride=1
+            )
+            self.conv_s1 = nn.Conv2d(
+                transformer_dim, transformer_dim // 4, kernel_size=1, stride=1
+            )
+
+        self.output_hypernetworks_mlps = nn.ModuleList(
+            [
+                MLP(transformer_dim, transformer_dim, transformer_dim // 8, 3)
+                for i in range(self.num_mask_tokens)
+            ]
+        )
+
+        self.iou_prediction_head = MLP(
+            transformer_dim,
+            iou_head_hidden_dim,
+            self.num_mask_tokens,
+            iou_head_depth,
+            sigmoid_output=iou_prediction_use_sigmoid,
+        )
+        if self.pred_obj_scores:
+            self.pred_obj_score_head = nn.Linear(transformer_dim, 1)
+            if pred_obj_scores_mlp:
+                self.pred_obj_score_head = MLP(transformer_dim, transformer_dim, 1, 3)
+
+        # When outputting a single mask, optionally we can dynamically fall back to the best
+        # multimask output token if the single mask output token gives low stability scores.
+        self.dynamic_multimask_via_stability = dynamic_multimask_via_stability
+        self.dynamic_multimask_stability_delta = dynamic_multimask_stability_delta
+        self.dynamic_multimask_stability_thresh = dynamic_multimask_stability_thresh
+
+    def forward(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        multimask_output: bool,
+        repeat_image: bool,
+        high_res_features: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Predict masks given image and prompt embeddings.
+
+        Arguments:
+          image_embeddings (torch.Tensor): the embeddings from the image encoder
+          image_pe (torch.Tensor): positional encoding with the shape of image_embeddings
+          sparse_prompt_embeddings (torch.Tensor): the embeddings of the points and boxes
+          dense_prompt_embeddings (torch.Tensor): the embeddings of the mask inputs
+          multimask_output (bool): Whether to return multiple masks or a single
+            mask.
+
+        Returns:
+          torch.Tensor: batched predicted masks
+          torch.Tensor: batched predictions of mask quality
+          torch.Tensor: batched SAM token for mask output
+        """
+        masks, iou_pred, mask_tokens_out, object_score_logits = self.predict_masks(
+            image_embeddings=image_embeddings,
+            image_pe=image_pe,
+            sparse_prompt_embeddings=sparse_prompt_embeddings,
+            dense_prompt_embeddings=dense_prompt_embeddings,
+            repeat_image=repeat_image,
+            high_res_features=high_res_features,
+        )
+
+        # Select the correct mask or masks for output
+        if multimask_output:
+            masks = masks[:, 1:, :, :]
+            iou_pred = iou_pred[:, 1:]
+        elif self.dynamic_multimask_via_stability and not self.training:
+            masks, iou_pred = self._dynamic_multimask_via_stability(masks, iou_pred)
+        else:
+            masks = masks[:, 0:1, :, :]
+            iou_pred = iou_pred[:, 0:1]
+
+        if multimask_output and self.use_multimask_token_for_obj_ptr:
+            sam_tokens_out = mask_tokens_out[:, 1:]  # [b, 3, c] shape
+        else:
+            # Take the mask output token. Here we *always* use the token for single mask output.
+            # At test time, even if we track after 1-click (and using multimask_output=True),
+            # we still take the single mask token here. The rationale is that we always track
+            # after multiple clicks during training, so the past tokens seen during training
+            # are always the single mask token (and we'll let it be the object-memory token).
+            sam_tokens_out = mask_tokens_out[:, 0:1]  # [b, 1, c] shape
+
+        # Prepare output
+        return masks, iou_pred, sam_tokens_out, object_score_logits
+
+    def predict_masks(
+        self,
+        image_embeddings: torch.Tensor,
+        image_pe: torch.Tensor,
+        sparse_prompt_embeddings: torch.Tensor,
+        dense_prompt_embeddings: torch.Tensor,
+        repeat_image: bool,
+        high_res_features: Optional[List[torch.Tensor]] = None,
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """Predicts masks. See 'forward' for more details."""
+        # Concatenate output tokens
+        s = 0
+        if self.pred_obj_scores:
+            output_tokens = torch.cat(
+                [
+                    self.obj_score_token.weight,
+                    self.iou_token.weight,
+                    self.mask_tokens.weight,
+                ],
+                dim=0,
+            )
+            s = 1
+        else:
+            output_tokens = torch.cat(
+                [self.iou_token.weight, self.mask_tokens.weight], dim=0
+            )
+        output_tokens = output_tokens.unsqueeze(0).expand(
+            sparse_prompt_embeddings.size(0), -1, -1
+        )
+        tokens = torch.cat((output_tokens, sparse_prompt_embeddings), dim=1)
+
+        # Expand per-image data in batch direction to be per-mask
+        if repeat_image:
+            src = torch.repeat_interleave(image_embeddings, tokens.shape[0], dim=0)
+        else:
+            assert image_embeddings.shape[0] == tokens.shape[0]
+            src = image_embeddings
+        src = src + dense_prompt_embeddings
+        assert (
+            image_pe.size(0) == 1
+        ), "image_pe should have size 1 in batch dim (from `get_dense_pe()`)"
+        pos_src = torch.repeat_interleave(image_pe, tokens.shape[0], dim=0)
+        b, c, h, w = src.shape
+
+        # Run the transformer
+        # print('src: ', src.dtype, 'pos_src:', pos_src.dtype, 'tokens:', tokens.dtype)
+        _dtype = pos_src.dtype
+        src = src.to(_dtype)
+        tokens = tokens.to(_dtype)
+        hs, src = self.transformer(src, pos_src, tokens)
+        iou_token_out = hs[:, s, :]
+        mask_tokens_out = hs[:, s + 1 : (s + 1 + self.num_mask_tokens), :]
+
+        # Upscale mask embeddings and predict masks using the mask tokens
+        src = src.transpose(1, 2).view(b, c, h, w)
+        if not self.use_high_res_features:
+            upscaled_embedding = self.output_upscaling(src)
+        else:
+            dc1, ln1, act1, dc2, act2 = self.output_upscaling
+            feat_s0, feat_s1 = high_res_features
+            upscaled_embedding = act1(ln1(dc1(src) + feat_s1))
+            upscaled_embedding = act2(dc2(upscaled_embedding) + feat_s0)
+
+        hyper_in_list: List[torch.Tensor] = []
+        for i in range(self.num_mask_tokens):
+            hyper_in_list.append(
+                self.output_hypernetworks_mlps[i](mask_tokens_out[:, i, :])
+            )
+        hyper_in = torch.stack(hyper_in_list, dim=1)
+        b, c, h, w = upscaled_embedding.shape
+        masks = (hyper_in @ upscaled_embedding.view(b, c, h * w)).view(b, -1, h, w)
+
+        # Generate mask quality predictions
+        iou_pred = self.iou_prediction_head(iou_token_out)
+        if self.pred_obj_scores:
+            assert s == 1
+            object_score_logits = self.pred_obj_score_head(hs[:, 0, :])
+        else:
+            # Obj scores logits - default to 10.0, i.e. assuming the object is present, sigmoid(10)=1
+            object_score_logits = 10.0 * iou_pred.new_ones(iou_pred.shape[0], 1)
+
+        return masks, iou_pred, mask_tokens_out, object_score_logits
+
+    def _get_stability_scores(self, mask_logits):
+        """
+        Compute stability scores of the mask logits based on the IoU between upper and
+        lower thresholds, similar to https://github.com/fairinternal/onevision/pull/568.
+        """
+        mask_logits = mask_logits.flatten(-2)
+        stability_delta = self.dynamic_multimask_stability_delta
+        area_i = torch.sum(mask_logits > stability_delta, dim=-1).float()
+        area_u = torch.sum(mask_logits > -stability_delta, dim=-1).float()
+        stability_scores = torch.where(area_u > 0, area_i / area_u, 1.0)
+        return stability_scores
+
+    def _dynamic_multimask_via_stability(self, all_mask_logits, all_iou_scores):
+        """
+        When outputting a single mask, if the stability score from the current single-mask
+        output (based on output token 0) falls below a threshold, we instead select from
+        multi-mask outputs (based on output token 1~3) the mask with the highest predicted
+        IoU score. This is intended to ensure a valid mask for both clicking and tracking.
+        """
+        # The best mask from multimask output tokens (1~3)
+        multimask_logits = all_mask_logits[:, 1:, :, :]
+        multimask_iou_scores = all_iou_scores[:, 1:]
+        best_scores_inds = torch.argmax(multimask_iou_scores, dim=-1)
+        batch_inds = torch.arange(
+            multimask_iou_scores.size(0), device=all_iou_scores.device
+        )
+        best_multimask_logits = multimask_logits[batch_inds, best_scores_inds]
+        best_multimask_logits = best_multimask_logits.unsqueeze(1)
+        best_multimask_iou_scores = multimask_iou_scores[batch_inds, best_scores_inds]
+        best_multimask_iou_scores = best_multimask_iou_scores.unsqueeze(1)
+
+        # The mask from singlemask output token 0 and its stability score
+        singlemask_logits = all_mask_logits[:, 0:1, :, :]
+        singlemask_iou_scores = all_iou_scores[:, 0:1]
+        stability_scores = self._get_stability_scores(singlemask_logits)
+        is_stable = stability_scores >= self.dynamic_multimask_stability_thresh
+
+        # Dynamically fall back to best multimask output upon low stability scores.
+        mask_logits_out = torch.where(
+            is_stable[..., None, None].expand_as(singlemask_logits),
+            singlemask_logits,
+            best_multimask_logits,
+        )
+        iou_scores_out = torch.where(
+            is_stable.expand_as(singlemask_iou_scores),
+            singlemask_iou_scores,
+            best_multimask_iou_scores,
+        )
+        return mask_logits_out, iou_scores_out
diff --git a/third_parts/sam2/modeling/sam/prompt_encoder.py b/third_parts/sam2/modeling/sam/prompt_encoder.py
new file mode 100644
index 0000000000000000000000000000000000000000..5c7e7c97875f2d4bfd11ebc2cb604c56bb901236
--- /dev/null
+++ b/third_parts/sam2/modeling/sam/prompt_encoder.py
@@ -0,0 +1,182 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from typing import Optional, Tuple, Type
+
+import torch
+from torch import nn
+
+from third_parts.sam2.modeling.position_encoding import PositionEmbeddingRandom
+
+from third_parts.sam2.modeling.sam2_utils import LayerNorm2d
+
+
+class PromptEncoder(nn.Module):
+    def __init__(
+        self,
+        embed_dim: int,
+        image_embedding_size: Tuple[int, int],
+        input_image_size: Tuple[int, int],
+        mask_in_chans: int,
+        activation: Type[nn.Module] = nn.GELU,
+    ) -> None:
+        """
+        Encodes prompts for input to SAM's mask decoder.
+
+        Arguments:
+          embed_dim (int): The prompts' embedding dimension
+          image_embedding_size (tuple(int, int)): The spatial size of the
+            image embedding, as (H, W).
+          input_image_size (int): The padded size of the image as input
+            to the image encoder, as (H, W).
+          mask_in_chans (int): The number of hidden channels used for
+            encoding input masks.
+          activation (nn.Module): The activation to use when encoding
+            input masks.
+        """
+        super().__init__()
+        self.embed_dim = embed_dim
+        self.input_image_size = input_image_size
+        self.image_embedding_size = image_embedding_size
+        self.pe_layer = PositionEmbeddingRandom(embed_dim // 2)
+
+        self.num_point_embeddings: int = 4  # pos/neg point + 2 box corners
+        point_embeddings = [
+            nn.Embedding(1, embed_dim) for i in range(self.num_point_embeddings)
+        ]
+        self.point_embeddings = nn.ModuleList(point_embeddings)
+        self.not_a_point_embed = nn.Embedding(1, embed_dim)
+
+        self.mask_input_size = (
+            4 * image_embedding_size[0],
+            4 * image_embedding_size[1],
+        )
+        self.mask_downscaling = nn.Sequential(
+            nn.Conv2d(1, mask_in_chans // 4, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans // 4),
+            activation(),
+            nn.Conv2d(mask_in_chans // 4, mask_in_chans, kernel_size=2, stride=2),
+            LayerNorm2d(mask_in_chans),
+            activation(),
+            nn.Conv2d(mask_in_chans, embed_dim, kernel_size=1),
+        )
+        self.no_mask_embed = nn.Embedding(1, embed_dim)
+
+    def get_dense_pe(self) -> torch.Tensor:
+        """
+        Returns the positional encoding used to encode point prompts,
+        applied to a dense set of points the shape of the image encoding.
+
+        Returns:
+          torch.Tensor: Positional encoding with shape
+            1x(embed_dim)x(embedding_h)x(embedding_w)
+        """
+        return self.pe_layer(self.image_embedding_size).unsqueeze(0)
+
+    def _embed_points(
+        self,
+        points: torch.Tensor,
+        labels: torch.Tensor,
+        pad: bool,
+    ) -> torch.Tensor:
+        """Embeds point prompts."""
+        points = points + 0.5  # Shift to center of pixel
+        if pad:
+            padding_point = torch.zeros((points.shape[0], 1, 2), device=points.device)
+            padding_label = -torch.ones((labels.shape[0], 1), device=labels.device)
+            points = torch.cat([points, padding_point], dim=1)
+            labels = torch.cat([labels, padding_label], dim=1)
+        point_embedding = self.pe_layer.forward_with_coords(
+            points, self.input_image_size
+        )
+        point_embedding[labels == -1] = 0.0
+        point_embedding[labels == -1] += self.not_a_point_embed.weight
+        point_embedding[labels == 0] += self.point_embeddings[0].weight
+        point_embedding[labels == 1] += self.point_embeddings[1].weight
+        point_embedding[labels == 2] += self.point_embeddings[2].weight
+        point_embedding[labels == 3] += self.point_embeddings[3].weight
+        return point_embedding
+
+    def _embed_boxes(self, boxes: torch.Tensor) -> torch.Tensor:
+        """Embeds box prompts."""
+        boxes = boxes + 0.5  # Shift to center of pixel
+        coords = boxes.reshape(-1, 2, 2)
+        corner_embedding = self.pe_layer.forward_with_coords(
+            coords, self.input_image_size
+        )
+        corner_embedding[:, 0, :] += self.point_embeddings[2].weight
+        corner_embedding[:, 1, :] += self.point_embeddings[3].weight
+        return corner_embedding
+
+    def _embed_masks(self, masks: torch.Tensor) -> torch.Tensor:
+        """Embeds mask inputs."""
+        mask_embedding = self.mask_downscaling(masks)
+        return mask_embedding
+
+    def _get_batch_size(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> int:
+        """
+        Gets the batch size of the output given the batch size of the input prompts.
+        """
+        if points is not None:
+            return points[0].shape[0]
+        elif boxes is not None:
+            return boxes.shape[0]
+        elif masks is not None:
+            return masks.shape[0]
+        else:
+            return 1
+
+    def _get_device(self) -> torch.device:
+        return self.point_embeddings[0].weight.device
+
+    def forward(
+        self,
+        points: Optional[Tuple[torch.Tensor, torch.Tensor]],
+        boxes: Optional[torch.Tensor],
+        masks: Optional[torch.Tensor],
+    ) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Embeds different types of prompts, returning both sparse and dense
+        embeddings.
+
+        Arguments:
+          points (tuple(torch.Tensor, torch.Tensor) or none): point coordinates
+            and labels to embed.
+          boxes (torch.Tensor or none): boxes to embed
+          masks (torch.Tensor or none): masks to embed
+
+        Returns:
+          torch.Tensor: sparse embeddings for the points and boxes, with shape
+            BxNx(embed_dim), where N is determined by the number of input points
+            and boxes.
+          torch.Tensor: dense embeddings for the masks, in the shape
+            Bx(embed_dim)x(embed_H)x(embed_W)
+        """
+        bs = self._get_batch_size(points, boxes, masks)
+        sparse_embeddings = torch.empty(
+            (bs, 0, self.embed_dim), device=self._get_device()
+        )
+        if points is not None:
+            coords, labels = points
+            point_embeddings = self._embed_points(coords, labels, pad=(boxes is None))
+            sparse_embeddings = torch.cat([sparse_embeddings, point_embeddings], dim=1)
+        if boxes is not None:
+            box_embeddings = self._embed_boxes(boxes)
+            sparse_embeddings = torch.cat([sparse_embeddings, box_embeddings], dim=1)
+
+        if masks is not None:
+            dense_embeddings = self._embed_masks(masks)
+        else:
+            dense_embeddings = self.no_mask_embed.weight.reshape(1, -1, 1, 1).expand(
+                bs, -1, self.image_embedding_size[0], self.image_embedding_size[1]
+            )
+
+        return sparse_embeddings, dense_embeddings
diff --git a/third_parts/sam2/modeling/sam/transformer.py b/third_parts/sam2/modeling/sam/transformer.py
new file mode 100644
index 0000000000000000000000000000000000000000..58d6f8bb43fbe4f6ba4f2bbbab13650c380aecf7
--- /dev/null
+++ b/third_parts/sam2/modeling/sam/transformer.py
@@ -0,0 +1,328 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+import warnings
+from functools import partial
+from typing import Tuple, Type
+
+import torch
+import torch.nn.functional as F
+from torch import nn, Tensor
+
+from third_parts.sam2.modeling.position_encoding import apply_rotary_enc, compute_axial_cis
+
+from third_parts.sam2.modeling.sam2_utils import MLP
+from third_parts.sam2.utils.misc import get_sdpa_settings
+
+warnings.simplefilter(action="ignore", category=FutureWarning)
+# OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = get_sdpa_settings()
+OLD_GPU, USE_FLASH_ATTN, MATH_KERNEL_ON = True, True, True
+
+
+class TwoWayTransformer(nn.Module):
+    def __init__(
+        self,
+        depth: int,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+    ) -> None:
+        """
+        A transformer decoder that attends to an input image using
+        queries whose positional embedding is supplied.
+
+        Args:
+          depth (int): number of layers in the transformer
+          embedding_dim (int): the channel dimension for the input embeddings
+          num_heads (int): the number of heads for multihead attention. Must
+            divide embedding_dim
+          mlp_dim (int): the channel dimension internal to the MLP block
+          activation (nn.Module): the activation to use in the MLP block
+        """
+        super().__init__()
+        self.depth = depth
+        self.embedding_dim = embedding_dim
+        self.num_heads = num_heads
+        self.mlp_dim = mlp_dim
+        self.layers = nn.ModuleList()
+
+        for i in range(depth):
+            self.layers.append(
+                TwoWayAttentionBlock(
+                    embedding_dim=embedding_dim,
+                    num_heads=num_heads,
+                    mlp_dim=mlp_dim,
+                    activation=activation,
+                    attention_downsample_rate=attention_downsample_rate,
+                    skip_first_layer_pe=(i == 0),
+                )
+            )
+
+        self.final_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm_final_attn = nn.LayerNorm(embedding_dim)
+
+    def forward(
+        self,
+        image_embedding: Tensor,
+        image_pe: Tensor,
+        point_embedding: Tensor,
+    ) -> Tuple[Tensor, Tensor]:
+        """
+        Args:
+          image_embedding (torch.Tensor): image to attend to. Should be shape
+            B x embedding_dim x h x w for any h and w.
+          image_pe (torch.Tensor): the positional encoding to add to the image. Must
+            have the same shape as image_embedding.
+          point_embedding (torch.Tensor): the embedding to add to the query points.
+            Must have shape B x N_points x embedding_dim for any N_points.
+
+        Returns:
+          torch.Tensor: the processed point_embedding
+          torch.Tensor: the processed image_embedding
+        """
+        # BxCxHxW -> BxHWxC == B x N_image_tokens x C
+        bs, c, h, w = image_embedding.shape
+        image_embedding = image_embedding.flatten(2).permute(0, 2, 1)
+        image_pe = image_pe.flatten(2).permute(0, 2, 1)
+
+        # Prepare queries
+        queries = point_embedding
+        keys = image_embedding
+
+        # Apply transformer blocks and final layernorm
+        for layer in self.layers:
+            queries, keys = layer(
+                queries=queries,
+                keys=keys,
+                query_pe=point_embedding,
+                key_pe=image_pe,
+            )
+
+        # Apply the final attention layer from the points to the image
+        q = queries + point_embedding
+        k = keys + image_pe
+        attn_out = self.final_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm_final_attn(queries)
+
+        return queries, keys
+
+
+class TwoWayAttentionBlock(nn.Module):
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        mlp_dim: int = 2048,
+        activation: Type[nn.Module] = nn.ReLU,
+        attention_downsample_rate: int = 2,
+        skip_first_layer_pe: bool = False,
+    ) -> None:
+        """
+        A transformer block with four layers: (1) self-attention of sparse
+        inputs, (2) cross attention of sparse inputs to dense inputs, (3) mlp
+        block on sparse inputs, and (4) cross attention of dense inputs to sparse
+        inputs.
+
+        Arguments:
+          embedding_dim (int): the channel dimension of the embeddings
+          num_heads (int): the number of heads in the attention layers
+          mlp_dim (int): the hidden dimension of the mlp block
+          activation (nn.Module): the activation of the mlp block
+          skip_first_layer_pe (bool): skip the PE on the first layer
+        """
+        super().__init__()
+        self.self_attn = Attention(embedding_dim, num_heads)
+        self.norm1 = nn.LayerNorm(embedding_dim)
+
+        self.cross_attn_token_to_image = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+        self.norm2 = nn.LayerNorm(embedding_dim)
+
+        self.mlp = MLP(
+            embedding_dim, mlp_dim, embedding_dim, num_layers=2, activation=activation
+        )
+        self.norm3 = nn.LayerNorm(embedding_dim)
+
+        self.norm4 = nn.LayerNorm(embedding_dim)
+        self.cross_attn_image_to_token = Attention(
+            embedding_dim, num_heads, downsample_rate=attention_downsample_rate
+        )
+
+        self.skip_first_layer_pe = skip_first_layer_pe
+
+    def forward(
+        self, queries: Tensor, keys: Tensor, query_pe: Tensor, key_pe: Tensor
+    ) -> Tuple[Tensor, Tensor]:
+        # Self attention block
+        if self.skip_first_layer_pe:
+            queries = self.self_attn(q=queries, k=queries, v=queries)
+        else:
+            q = queries + query_pe
+            attn_out = self.self_attn(q=q, k=q, v=queries)
+            queries = queries + attn_out
+        queries = self.norm1(queries)
+
+        # Cross attention block, tokens attending to image embedding
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_token_to_image(q=q, k=k, v=keys)
+        queries = queries + attn_out
+        queries = self.norm2(queries)
+
+        # MLP block
+        mlp_out = self.mlp(queries)
+        queries = queries + mlp_out
+        queries = self.norm3(queries)
+
+        # Cross attention block, image embedding attending to tokens
+        q = queries + query_pe
+        k = keys + key_pe
+        attn_out = self.cross_attn_image_to_token(q=k, k=q, v=queries)
+        keys = keys + attn_out
+        keys = self.norm4(keys)
+
+        return queries, keys
+
+
+class Attention(nn.Module):
+    """
+    An attention layer that allows for downscaling the size of the embedding
+    after projection to queries, keys, and values.
+    """
+
+    def __init__(
+        self,
+        embedding_dim: int,
+        num_heads: int,
+        downsample_rate: int = 1,
+        dropout: float = 0.0,
+        kv_in_dim: int = None,
+    ) -> None:
+        super().__init__()
+        self.embedding_dim = embedding_dim
+        self.kv_in_dim = kv_in_dim if kv_in_dim is not None else embedding_dim
+        self.internal_dim = embedding_dim // downsample_rate
+        self.num_heads = num_heads
+        assert (
+            self.internal_dim % num_heads == 0
+        ), "num_heads must divide embedding_dim."
+
+        self.q_proj = nn.Linear(embedding_dim, self.internal_dim)
+        self.k_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.v_proj = nn.Linear(self.kv_in_dim, self.internal_dim)
+        self.out_proj = nn.Linear(self.internal_dim, embedding_dim)
+
+        self.dropout_p = dropout
+
+    def _separate_heads(self, x: Tensor, num_heads: int) -> Tensor:
+        b, n, c = x.shape
+        x = x.reshape(b, n, num_heads, c // num_heads)
+        return x.transpose(1, 2)  # B x N_heads x N_tokens x C_per_head
+
+    def _recombine_heads(self, x: Tensor) -> Tensor:
+        b, n_heads, n_tokens, c_per_head = x.shape
+        x = x.transpose(1, 2)
+        return x.reshape(b, n_tokens, n_heads * c_per_head)  # B x N_tokens x C
+
+    def forward(self, q: Tensor, k: Tensor, v: Tensor) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        with torch.backends.cuda.sdp_kernel(
+            enable_flash=USE_FLASH_ATTN,
+            # if Flash attention kernel is off, then math kernel needs to be enabled
+            enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
+            enable_mem_efficient=OLD_GPU,
+        ):
+            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out
+
+
+class RoPEAttention(Attention):
+    """Attention with rotary position encoding."""
+
+    def __init__(
+        self,
+        *args,
+        rope_theta=10000.0,
+        # whether to repeat q rope to match k length
+        # this is needed for cross-attention to memories
+        rope_k_repeat=False,
+        feat_sizes=(32, 32),  # [w, h] for stride 16 feats at 512 resolution
+        **kwargs,
+    ):
+        super().__init__(*args, **kwargs)
+
+        self.compute_cis = partial(
+            compute_axial_cis, dim=self.internal_dim // self.num_heads, theta=rope_theta
+        )
+        freqs_cis = self.compute_cis(end_x=feat_sizes[0], end_y=feat_sizes[1])
+        self.freqs_cis = freqs_cis
+        self.rope_k_repeat = rope_k_repeat
+
+    def forward(
+        self, q: Tensor, k: Tensor, v: Tensor, num_k_exclude_rope: int = 0
+    ) -> Tensor:
+        # Input projections
+        q = self.q_proj(q)
+        k = self.k_proj(k)
+        v = self.v_proj(v)
+
+        # Separate into heads
+        q = self._separate_heads(q, self.num_heads)
+        k = self._separate_heads(k, self.num_heads)
+        v = self._separate_heads(v, self.num_heads)
+
+        # Apply rotary position encoding
+        w = h = math.sqrt(q.shape[-2])
+        self.freqs_cis = self.freqs_cis.to(q.device)
+        if self.freqs_cis.shape[0] != q.shape[-2]:
+            self.freqs_cis = self.compute_cis(end_x=w, end_y=h).to(q.device)
+        if q.shape[-2] != k.shape[-2]:
+            assert self.rope_k_repeat
+
+        num_k_rope = k.size(-2) - num_k_exclude_rope
+        q, k[:, :, :num_k_rope] = apply_rotary_enc(
+            q,
+            k[:, :, :num_k_rope],
+            freqs_cis=self.freqs_cis,
+            repeat_freqs_k=self.rope_k_repeat,
+        )
+
+        dropout_p = self.dropout_p if self.training else 0.0
+        # Attention
+        with torch.backends.cuda.sdp_kernel(
+            enable_flash=USE_FLASH_ATTN,
+            # if Flash attention kernel is off, then math kernel needs to be enabled
+            enable_math=(OLD_GPU and dropout_p > 0.0) or MATH_KERNEL_ON,
+            enable_mem_efficient=OLD_GPU,
+        ):
+            out = F.scaled_dot_product_attention(q, k, v, dropout_p=dropout_p)
+
+        out = self._recombine_heads(out)
+        out = self.out_proj(out)
+
+        return out
diff --git a/third_parts/sam2/modeling/sam2_base.py b/third_parts/sam2/modeling/sam2_base.py
new file mode 100644
index 0000000000000000000000000000000000000000..335268257a2806c3839b4b6c8730bb1318488602
--- /dev/null
+++ b/third_parts/sam2/modeling/sam2_base.py
@@ -0,0 +1,830 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.distributed
+import torch.nn.functional as F
+
+from torch.nn.init import trunc_normal_
+
+from third_parts.sam2.modeling.sam.mask_decoder import MaskDecoder
+from third_parts.sam2.modeling.sam.prompt_encoder import PromptEncoder
+from third_parts.sam2.modeling.sam.transformer import TwoWayTransformer
+from third_parts.sam2.modeling.sam2_utils import get_1d_sine_pe, MLP, select_closest_cond_frames
+
+# a large negative value as a placeholder score for missing objects
+NO_OBJ_SCORE = -1024.0
+
+
+class SAM2Base(torch.nn.Module):
+    def __init__(
+        self,
+        image_encoder,
+        memory_attention,
+        memory_encoder,
+        num_maskmem=7,  # default 1 input frame + 6 previous frames
+        image_size=512,
+        backbone_stride=16,  # stride of the image backbone output
+        sigmoid_scale_for_mem_enc=1.0,  # scale factor for mask sigmoid prob
+        sigmoid_bias_for_mem_enc=0.0,  # bias factor for mask sigmoid prob
+        # During evaluation, whether to binarize the sigmoid mask logits on interacted frames with clicks
+        binarize_mask_from_pts_for_mem_enc=False,
+        use_mask_input_as_output_without_sam=False,  # on frames with mask input, whether to directly output the input mask without using a SAM prompt encoder + mask decoder
+        # The maximum number of conditioning frames to participate in the memory attention (-1 means no limit; if there are more conditioning frames than this limit,
+        # we only cross-attend to the temporally closest `max_cond_frames_in_attn` conditioning frames in the encoder when tracking each frame). This gives the model
+        # a temporal locality when handling a large number of annotated frames (since closer frames should be more important) and also avoids GPU OOM.
+        max_cond_frames_in_attn=-1,
+        # on the first frame, whether to directly add the no-memory embedding to the image feature
+        # (instead of using the transformer encoder)
+        directly_add_no_mem_embed=False,
+        # whether to use high-resolution feature maps in the SAM mask decoder
+        use_high_res_features_in_sam=False,
+        # whether to output multiple (3) masks for the first click on initial conditioning frames
+        multimask_output_in_sam=False,
+        # the minimum and maximum number of clicks to use multimask_output_in_sam (only relevant when `multimask_output_in_sam=True`;
+        # default is 1 for both, meaning that only the first click gives multimask output; also note that a box counts as two points)
+        multimask_min_pt_num=1,
+        multimask_max_pt_num=1,
+        # whether to also use multimask output for tracking (not just for the first click on initial conditioning frames; only relevant when `multimask_output_in_sam=True`)
+        multimask_output_for_tracking=False,
+        # Whether to use multimask tokens for obj ptr; Only relevant when both
+        # use_obj_ptrs_in_encoder=True and multimask_output_for_tracking=True
+        use_multimask_token_for_obj_ptr: bool = False,
+        # whether to use sigmoid to restrict ious prediction to [0-1]
+        iou_prediction_use_sigmoid=False,
+        # The memory bank's temporal stride during evaluation (i.e. the `r` parameter in XMem and Cutie; XMem and Cutie use r=5).
+        # For r>1, the (self.num_maskmem - 1) non-conditioning memory frames consist of
+        # (self.num_maskmem - 2) nearest frames from every r-th frames, plus the last frame.
+        memory_temporal_stride_for_eval=1,
+        # if `add_all_frames_to_correct_as_cond` is True, we also append to the conditioning frame list any frame that receives a later correction click
+        # if `add_all_frames_to_correct_as_cond` is False, we conditioning frame list to only use those initial conditioning frames
+        add_all_frames_to_correct_as_cond=False,
+        # whether to apply non-overlapping constraints on the object masks in the memory encoder during evaluation (to avoid/alleviate superposing masks)
+        non_overlap_masks_for_mem_enc=False,
+        # whether to cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+        use_obj_ptrs_in_encoder=False,
+        # the maximum number of object pointers from other frames in encoder cross attention (only relevant when `use_obj_ptrs_in_encoder=True`)
+        max_obj_ptrs_in_encoder=16,
+        # whether to add temporal positional encoding to the object pointers in the encoder (only relevant when `use_obj_ptrs_in_encoder=True`)
+        add_tpos_enc_to_obj_ptrs=True,
+        # whether to add an extra linear projection layer for the temporal positional encoding in the object pointers to avoid potential interference
+        # with spatial positional encoding (only relevant when both `use_obj_ptrs_in_encoder=True` and `add_tpos_enc_to_obj_ptrs=True`)
+        proj_tpos_enc_in_obj_ptrs=False,
+        # whether to only attend to object pointers in the past (before the current frame) in the encoder during evaluation
+        # (only relevant when `use_obj_ptrs_in_encoder=True`; this might avoid pointer information too far in the future to distract the initial tracking)
+        only_obj_ptrs_in_the_past_for_eval=False,
+        # Whether to predict if there is an object in the frame
+        pred_obj_scores: bool = False,
+        # Whether to use an MLP to predict object scores
+        pred_obj_scores_mlp: bool = False,
+        # Only relevant if pred_obj_scores=True and use_obj_ptrs_in_encoder=True;
+        # Whether to have a fixed no obj pointer when there is no object present
+        # or to use it as an additive embedding with obj_ptr produced by decoder
+        fixed_no_obj_ptr: bool = False,
+        # Soft no object, i.e. mix in no_obj_ptr softly,
+        # hope to make recovery easier if there is a mistake and mitigate accumulation of errors
+        soft_no_obj_ptr: bool = False,
+        use_mlp_for_obj_ptr_proj: bool = False,
+        # extra arguments used to construct the SAM mask decoder; if not None, it should be a dict of kwargs to be passed into `MaskDecoder` class.
+        sam_mask_decoder_extra_args=None,
+        compile_image_encoder: bool = False,
+    ):
+        super().__init__()
+
+        # Part 1: the image backbone
+        self.image_encoder = image_encoder
+        # Use level 0, 1, 2 for high-res setting, or just level 2 for the default setting
+        self.use_high_res_features_in_sam = use_high_res_features_in_sam
+        self.num_feature_levels = 3 if use_high_res_features_in_sam else 1
+        self.use_obj_ptrs_in_encoder = use_obj_ptrs_in_encoder
+        self.max_obj_ptrs_in_encoder = max_obj_ptrs_in_encoder
+        if use_obj_ptrs_in_encoder:
+            # A conv layer to downsample the mask prompt to stride 4 (the same stride as
+            # low-res SAM mask logits) and to change its scales from 0~1 to SAM logit scale,
+            # so that it can be fed into the SAM mask decoder to generate a pointer.
+            self.mask_downsample = torch.nn.Conv2d(1, 1, kernel_size=4, stride=4)
+        self.add_tpos_enc_to_obj_ptrs = add_tpos_enc_to_obj_ptrs
+        if proj_tpos_enc_in_obj_ptrs:
+            assert add_tpos_enc_to_obj_ptrs  # these options need to be used together
+        self.proj_tpos_enc_in_obj_ptrs = proj_tpos_enc_in_obj_ptrs
+        self.only_obj_ptrs_in_the_past_for_eval = only_obj_ptrs_in_the_past_for_eval
+
+        # Part 2: memory attention to condition current frame's visual features
+        # with memories (and obj ptrs) from past frames
+        self.memory_attention = memory_attention
+        self.hidden_dim = memory_attention.d_model
+
+        # Part 3: memory encoder for the previous frame's outputs
+        self.memory_encoder = memory_encoder
+        self.mem_dim = self.hidden_dim
+        if hasattr(self.memory_encoder, "out_proj") and hasattr(
+            self.memory_encoder.out_proj, "weight"
+        ):
+            # if there is compression of memories along channel dim
+            self.mem_dim = self.memory_encoder.out_proj.weight.shape[0]
+        self.num_maskmem = num_maskmem  # Number of memories accessible
+        # Temporal encoding of the memories
+        self.maskmem_tpos_enc = torch.nn.Parameter(
+            torch.zeros(num_maskmem, 1, 1, self.mem_dim)
+        )
+        trunc_normal_(self.maskmem_tpos_enc, std=0.02)
+        # a single token to indicate no memory embedding from previous frames
+        self.no_mem_embed = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        self.no_mem_pos_enc = torch.nn.Parameter(torch.zeros(1, 1, self.hidden_dim))
+        trunc_normal_(self.no_mem_embed, std=0.02)
+        trunc_normal_(self.no_mem_pos_enc, std=0.02)
+        self.directly_add_no_mem_embed = directly_add_no_mem_embed
+        # Apply sigmoid to the output raw mask logits (to turn them from
+        # range (-inf, +inf) to range (0, 1)) before feeding them into the memory encoder
+        self.sigmoid_scale_for_mem_enc = sigmoid_scale_for_mem_enc
+        self.sigmoid_bias_for_mem_enc = sigmoid_bias_for_mem_enc
+        self.binarize_mask_from_pts_for_mem_enc = binarize_mask_from_pts_for_mem_enc
+        self.non_overlap_masks_for_mem_enc = non_overlap_masks_for_mem_enc
+        self.memory_temporal_stride_for_eval = memory_temporal_stride_for_eval
+        # On frames with mask input, whether to directly output the input mask without
+        # using a SAM prompt encoder + mask decoder
+        self.use_mask_input_as_output_without_sam = use_mask_input_as_output_without_sam
+        self.multimask_output_in_sam = multimask_output_in_sam
+        self.multimask_min_pt_num = multimask_min_pt_num
+        self.multimask_max_pt_num = multimask_max_pt_num
+        self.multimask_output_for_tracking = multimask_output_for_tracking
+        self.use_multimask_token_for_obj_ptr = use_multimask_token_for_obj_ptr
+        self.iou_prediction_use_sigmoid = iou_prediction_use_sigmoid
+
+        # Part 4: SAM-style prompt encoder (for both mask and point inputs)
+        # and SAM-style mask decoder for the final mask output
+        self.image_size = image_size
+        self.backbone_stride = backbone_stride
+        self.sam_mask_decoder_extra_args = sam_mask_decoder_extra_args
+        self.pred_obj_scores = pred_obj_scores
+        self.pred_obj_scores_mlp = pred_obj_scores_mlp
+        self.fixed_no_obj_ptr = fixed_no_obj_ptr
+        self.soft_no_obj_ptr = soft_no_obj_ptr
+        if self.fixed_no_obj_ptr:
+            assert self.pred_obj_scores
+            assert self.use_obj_ptrs_in_encoder
+        if self.pred_obj_scores and self.use_obj_ptrs_in_encoder:
+            self.no_obj_ptr = torch.nn.Parameter(torch.zeros(1, self.hidden_dim))
+            trunc_normal_(self.no_obj_ptr, std=0.02)
+        self.use_mlp_for_obj_ptr_proj = use_mlp_for_obj_ptr_proj
+
+        self._build_sam_heads()
+        self.add_all_frames_to_correct_as_cond = add_all_frames_to_correct_as_cond
+        self.max_cond_frames_in_attn = max_cond_frames_in_attn
+
+        # Model compilation
+        if compile_image_encoder:
+            # Compile the forward function (not the full module) to allow loading checkpoints.
+            print(
+                "Image encoder compilation is enabled. First forward pass will be slow."
+            )
+            self.image_encoder.forward = torch.compile(
+                self.image_encoder.forward,
+                mode="max-autotune",
+                fullgraph=True,
+                dynamic=False,
+            )
+
+    @property
+    def device(self):
+        return next(self.parameters()).device
+
+    def forward(self, *args, **kwargs):
+        raise NotImplementedError(
+            "Please use the corresponding methods in SAM2VideoPredictor for inference."
+            "See notebooks/video_predictor_example.ipynb for an example."
+        )
+
+    def _build_sam_heads(self):
+        """Build SAM-style prompt encoder and mask decoder."""
+        self.sam_prompt_embed_dim = self.hidden_dim
+        self.sam_image_embedding_size = self.image_size // self.backbone_stride
+
+        # build PromptEncoder and MaskDecoder from SAM
+        # (their hyperparameters like `mask_in_chans=16` are from SAM code)
+        self.sam_prompt_encoder = PromptEncoder(
+            embed_dim=self.sam_prompt_embed_dim,
+            image_embedding_size=(
+                self.sam_image_embedding_size,
+                self.sam_image_embedding_size,
+            ),
+            input_image_size=(self.image_size, self.image_size),
+            mask_in_chans=16,
+        )
+        self.sam_mask_decoder = MaskDecoder(
+            num_multimask_outputs=3,
+            transformer=TwoWayTransformer(
+                depth=2,
+                embedding_dim=self.sam_prompt_embed_dim,
+                mlp_dim=2048,
+                num_heads=8,
+            ),
+            transformer_dim=self.sam_prompt_embed_dim,
+            iou_head_depth=3,
+            iou_head_hidden_dim=256,
+            use_high_res_features=self.use_high_res_features_in_sam,
+            iou_prediction_use_sigmoid=self.iou_prediction_use_sigmoid,
+            pred_obj_scores=self.pred_obj_scores,
+            pred_obj_scores_mlp=self.pred_obj_scores_mlp,
+            use_multimask_token_for_obj_ptr=self.use_multimask_token_for_obj_ptr,
+            **(self.sam_mask_decoder_extra_args or {}),
+        )
+        if self.use_obj_ptrs_in_encoder:
+            # a linear projection on SAM output tokens to turn them into object pointers
+            self.obj_ptr_proj = torch.nn.Linear(self.hidden_dim, self.hidden_dim)
+            if self.use_mlp_for_obj_ptr_proj:
+                self.obj_ptr_proj = MLP(
+                    self.hidden_dim, self.hidden_dim, self.hidden_dim, 3
+                )
+        else:
+            self.obj_ptr_proj = torch.nn.Identity()
+        if self.proj_tpos_enc_in_obj_ptrs:
+            # a linear projection on temporal positional encoding in object pointers to
+            # avoid potential interference with spatial positional encoding
+            self.obj_ptr_tpos_proj = torch.nn.Linear(self.hidden_dim, self.mem_dim)
+        else:
+            self.obj_ptr_tpos_proj = torch.nn.Identity()
+
+    def _forward_sam_heads(
+        self,
+        backbone_features,
+        point_inputs=None,
+        mask_inputs=None,
+        high_res_features=None,
+        multimask_output=False,
+    ):
+        """
+        Forward SAM prompt encoders and mask heads.
+
+        Inputs:
+        - backbone_features: image features of [B, C, H, W] shape
+        - point_inputs: a dictionary with "point_coords" and "point_labels", where
+          1) "point_coords" has [B, P, 2] shape and float32 dtype and contains the
+             absolute pixel-unit coordinate in (x, y) format of the P input points
+          2) "point_labels" has shape [B, P] and int32 dtype, where 1 means
+             positive clicks, 0 means negative clicks, and -1 means padding
+        - mask_inputs: a mask of [B, 1, H*16, W*16] shape, float or bool, with the
+          same spatial size as the image.
+        - high_res_features: either 1) None or 2) or a list of length 2 containing
+          two feature maps of [B, C, 4*H, 4*W] and [B, C, 2*H, 2*W] shapes respectively,
+          which will be used as high-resolution feature maps for SAM decoder.
+        - multimask_output: if it's True, we output 3 candidate masks and their 3
+          corresponding IoU estimates, and if it's False, we output only 1 mask and
+          its corresponding IoU estimate.
+
+        Outputs:
+        - low_res_multimasks: [B, M, H*4, W*4] shape (where M = 3 if
+          `multimask_output=True` and M = 1 if `multimask_output=False`), the SAM
+          output mask logits (before sigmoid) for the low-resolution masks, with 4x
+          the resolution (1/4 stride) of the input backbone_features.
+        - high_res_multimasks: [B, M, H*16, W*16] shape (where M = 3
+          if `multimask_output=True` and M = 1 if `multimask_output=False`),
+          upsampled from the low-resolution masks, with shape size as the image
+          (stride is 1 pixel).
+        - ious, [B, M] shape, where (where M = 3 if `multimask_output=True` and M = 1
+          if `multimask_output=False`), the estimated IoU of each output mask.
+        - low_res_masks: [B, 1, H*4, W*4] shape, the best mask in `low_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `low_res_multimasks`.
+        - high_res_masks: [B, 1, H*16, W*16] shape, the best mask in `high_res_multimasks`.
+          If `multimask_output=True`, it's the mask with the highest IoU estimate.
+          If `multimask_output=False`, it's the same as `high_res_multimasks`.
+        - obj_ptr: [B, C] shape, the object pointer vector for the output mask, extracted
+          based on the output token from the SAM mask decoder.
+        """
+        B = backbone_features.size(0)
+        device = backbone_features.device
+        assert backbone_features.size(1) == self.sam_prompt_embed_dim
+        assert backbone_features.size(2) == self.sam_image_embedding_size
+        assert backbone_features.size(3) == self.sam_image_embedding_size
+
+        # a) Handle point prompts
+        if point_inputs is not None:
+            sam_point_coords = point_inputs["point_coords"]
+            sam_point_labels = point_inputs["point_labels"]
+            assert sam_point_coords.size(0) == B and sam_point_labels.size(0) == B
+        else:
+            # If no points are provide, pad with an empty point (with label -1)
+            sam_point_coords = torch.zeros(B, 1, 2, device=device)
+            sam_point_labels = -torch.ones(B, 1, dtype=torch.int32, device=device)
+
+        # b) Handle mask prompts
+        if mask_inputs is not None:
+            # If mask_inputs is provided, downsize it into low-res mask input if needed
+            # and feed it as a dense mask prompt into the SAM mask encoder
+            assert len(mask_inputs.shape) == 4 and mask_inputs.shape[:2] == (B, 1)
+            if mask_inputs.shape[-2:] != self.sam_prompt_encoder.mask_input_size:
+                sam_mask_prompt = F.interpolate(
+                    mask_inputs.float(),
+                    size=self.sam_prompt_encoder.mask_input_size,
+                    align_corners=False,
+                    mode="bilinear",
+                    antialias=True,  # use antialias for downsampling
+                )
+            else:
+                sam_mask_prompt = mask_inputs
+        else:
+            # Otherwise, simply feed None (and SAM's prompt encoder will add
+            # a learned `no_mask_embed` to indicate no mask input in this case).
+            sam_mask_prompt = None
+
+        sparse_embeddings, dense_embeddings = self.sam_prompt_encoder(
+            points=(sam_point_coords, sam_point_labels),
+            boxes=None,
+            masks=sam_mask_prompt,
+        )
+        (
+            low_res_multimasks,
+            ious,
+            sam_output_tokens,
+            object_score_logits,
+        ) = self.sam_mask_decoder(
+            image_embeddings=backbone_features,
+            image_pe=self.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=False,  # the image is already batched
+            high_res_features=high_res_features,
+        )
+        if self.pred_obj_scores:
+            is_obj_appearing = object_score_logits > 0
+
+            # Mask used for spatial memories is always a *hard* choice between obj and no obj,
+            # consistent with the actual mask prediction
+            low_res_multimasks = torch.where(
+                is_obj_appearing[:, None, None],
+                low_res_multimasks,
+                NO_OBJ_SCORE,
+            )
+
+        # convert masks from possibly bfloat16 (or float16) to float32
+        # (older PyTorch versions before 2.1 don't support `interpolate` on bf16)
+        _dtype = low_res_multimasks.dtype
+        # low_res_multimasks = low_res_multimasks.float()
+        high_res_multimasks = F.interpolate(
+            low_res_multimasks.float(),
+            size=(self.image_size, self.image_size),
+            mode="bilinear",
+            align_corners=False,
+        ).to(_dtype)
+
+        sam_output_token = sam_output_tokens[:, 0]
+        if multimask_output:
+            # take the best mask prediction (with the highest IoU estimation)
+            best_iou_inds = torch.argmax(ious, dim=-1)
+            batch_inds = torch.arange(B, device=device)
+            low_res_masks = low_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            high_res_masks = high_res_multimasks[batch_inds, best_iou_inds].unsqueeze(1)
+            if sam_output_tokens.size(1) > 1:
+                sam_output_token = sam_output_tokens[batch_inds, best_iou_inds]
+        else:
+            low_res_masks, high_res_masks = low_res_multimasks, high_res_multimasks
+
+        # Extract object pointer from the SAM output token (with occlusion handling)
+        obj_ptr = self.obj_ptr_proj(sam_output_token)
+        if self.pred_obj_scores:
+            # Allow *soft* no obj ptr, unlike for masks
+            if self.soft_no_obj_ptr:
+                # Only hard possible with gt
+                assert not self.teacher_force_obj_scores_for_mem
+                lambda_is_obj_appearing = object_score_logits.sigmoid()
+            else:
+                lambda_is_obj_appearing = is_obj_appearing.float()
+
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_multimasks,
+            high_res_multimasks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def _use_mask_as_output(self, backbone_features, high_res_features, mask_inputs):
+        """
+        Directly turn binary `mask_inputs` into a output mask logits without using SAM.
+        (same input and output shapes as in _forward_sam_heads above).
+        """
+        # Use -10/+10 as logits for neg/pos pixels (very close to 0/1 in prob after sigmoid).
+        out_scale, out_bias = 20.0, -10.0  # sigmoid(-10.0)=4.5398e-05
+        mask_inputs_float = mask_inputs.float()
+        high_res_masks = mask_inputs_float * out_scale + out_bias
+        low_res_masks = F.interpolate(
+            high_res_masks,
+            size=(high_res_masks.size(-2) // 4, high_res_masks.size(-1) // 4),
+            align_corners=False,
+            mode="bilinear",
+            antialias=True,  # use antialias for downsampling
+        )
+        # a dummy IoU prediction of all 1's under mask input
+        ious = mask_inputs.new_ones(mask_inputs.size(0), 1).float()
+        if not self.use_obj_ptrs_in_encoder:
+            # all zeros as a dummy object pointer (of shape [B, C])
+            obj_ptr = torch.zeros(
+                mask_inputs.size(0), self.hidden_dim, device=mask_inputs.device
+            )
+        else:
+            # produce an object pointer using the SAM decoder from the mask input
+            _, _, _, _, _, obj_ptr, _ = self._forward_sam_heads(
+                backbone_features=backbone_features,
+                mask_inputs=self.mask_downsample(mask_inputs_float),
+                high_res_features=high_res_features,
+            )
+        # In this method, we are treating mask_input as output, e.g. using it directly to create spatial mem;
+        # Below, we follow the same design axiom to use mask_input to decide if obj appears or not instead of relying
+        # on the object_scores from the SAM decoder.
+        is_obj_appearing = torch.any(mask_inputs.flatten(1).float() > 0.0, dim=1)
+        is_obj_appearing = is_obj_appearing[..., None]
+        lambda_is_obj_appearing = is_obj_appearing.float()
+        object_score_logits = out_scale * lambda_is_obj_appearing + out_bias
+        if self.pred_obj_scores:
+            if self.fixed_no_obj_ptr:
+                obj_ptr = lambda_is_obj_appearing * obj_ptr
+            obj_ptr = obj_ptr + (1 - lambda_is_obj_appearing) * self.no_obj_ptr
+
+        return (
+            low_res_masks,
+            high_res_masks,
+            ious,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            object_score_logits,
+        )
+
+    def forward_image(self, img_batch: torch.Tensor):
+        """Get the image feature on the input batch."""
+        backbone_out = self.image_encoder(img_batch)
+        if self.use_high_res_features_in_sam:
+            # precompute projected level 0 and level 1 features in SAM decoder
+            # to avoid running it again on every SAM click
+            backbone_out["backbone_fpn"][0] = self.sam_mask_decoder.conv_s0(
+                backbone_out["backbone_fpn"][0]
+            )
+            backbone_out["backbone_fpn"][1] = self.sam_mask_decoder.conv_s1(
+                backbone_out["backbone_fpn"][1]
+            )
+        return backbone_out
+
+    def _prepare_backbone_features(self, backbone_out):
+        """Prepare and flatten visual features."""
+        backbone_out = backbone_out.copy()
+        assert len(backbone_out["backbone_fpn"]) == len(backbone_out["vision_pos_enc"])
+        assert len(backbone_out["backbone_fpn"]) >= self.num_feature_levels
+
+        feature_maps = backbone_out["backbone_fpn"][-self.num_feature_levels :]
+        vision_pos_embeds = backbone_out["vision_pos_enc"][-self.num_feature_levels :]
+
+        feat_sizes = [(x.shape[-2], x.shape[-1]) for x in vision_pos_embeds]
+        # flatten NxCxHxW to HWxNxC
+        vision_feats = [x.flatten(2).permute(2, 0, 1) for x in feature_maps]
+        vision_pos_embeds = [x.flatten(2).permute(2, 0, 1) for x in vision_pos_embeds]
+
+        return backbone_out, vision_feats, vision_pos_embeds, feat_sizes
+
+    def _prepare_memory_conditioned_features(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+    ):
+        """Fuse the current frame's visual feature map with previous memory."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        device = current_vision_feats[-1].device
+        # The case of `self.num_maskmem == 0` below is primarily used for reproducing SAM on images.
+        # In this case, we skip the fusion with any memory.
+        if self.num_maskmem == 0:  # Disable memory and skip fusion
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+            return pix_feat
+
+        num_obj_ptr_tokens = 0
+        # Step 1: condition the visual features of the current frame on previous memories
+        if not is_init_cond_frame:
+            # Retrieve the memories encoded with the maskmem backbone
+            to_cat_memory, to_cat_memory_pos_embed = [], []
+            # Add conditioning frames's output first (all cond frames have t_pos=0 for
+            # when getting temporal positional embedding below)
+            assert len(output_dict["cond_frame_outputs"]) > 0
+            # Select a maximum number of temporally closest cond frames for cross attention
+            cond_outputs = output_dict["cond_frame_outputs"]
+            selected_cond_outputs, unselected_cond_outputs = select_closest_cond_frames(
+                frame_idx, cond_outputs, self.max_cond_frames_in_attn
+            )
+            t_pos_and_prevs = [(0, out) for out in selected_cond_outputs.values()]
+            # Add last (self.num_maskmem - 1) frames before current frame for non-conditioning memory
+            # the earliest one has t_pos=1 and the latest one has t_pos=self.num_maskmem-1
+            # We also allow taking the memory frame non-consecutively (with r>1), in which case
+            # we take (self.num_maskmem - 2) frames among every r-th frames plus the last frame.
+            r = self.memory_temporal_stride_for_eval
+            for t_pos in range(1, self.num_maskmem):
+                t_rel = self.num_maskmem - t_pos  # how many frames before current frame
+                if t_rel == 1:
+                    # for t_rel == 1, we take the last frame (regardless of r)
+                    if not track_in_reverse:
+                        # the frame immediately before this frame (i.e. frame_idx - 1)
+                        prev_frame_idx = frame_idx - t_rel
+                    else:
+                        # the frame immediately after this frame (i.e. frame_idx + 1)
+                        prev_frame_idx = frame_idx + t_rel
+                else:
+                    # for t_rel >= 2, we take the memory frame from every r-th frames
+                    if not track_in_reverse:
+                        # first find the nearest frame among every r-th frames before this frame
+                        # for r=1, this would be (frame_idx - 2)
+                        prev_frame_idx = ((frame_idx - 2) // r) * r
+                        # then seek further among every r-th frames
+                        prev_frame_idx = prev_frame_idx - (t_rel - 2) * r
+                    else:
+                        # first find the nearest frame among every r-th frames after this frame
+                        # for r=1, this would be (frame_idx + 2)
+                        prev_frame_idx = -(-(frame_idx + 2) // r) * r
+                        # then seek further among every r-th frames
+                        prev_frame_idx = prev_frame_idx + (t_rel - 2) * r
+                out = output_dict["non_cond_frame_outputs"].get(prev_frame_idx, None)
+                if out is None:
+                    # If an unselected conditioning frame is among the last (self.num_maskmem - 1)
+                    # frames, we still attend to it as if it's a non-conditioning frame.
+                    out = unselected_cond_outputs.get(prev_frame_idx, None)
+                t_pos_and_prevs.append((t_pos, out))
+
+            for t_pos, prev in t_pos_and_prevs:
+                if prev is None:
+                    continue  # skip padding frames
+                # "maskmem_features" might have been offloaded to CPU in demo use cases,
+                # so we load it back to GPU (it's a no-op if it's already on GPU).
+                feats = prev["maskmem_features"].cuda(non_blocking=True)
+                to_cat_memory.append(feats.flatten(2).permute(2, 0, 1))
+                # Spatial positional encoding (it might have been offloaded to CPU in eval)
+                maskmem_enc = prev["maskmem_pos_enc"][-1].cuda()
+                maskmem_enc = maskmem_enc.flatten(2).permute(2, 0, 1)
+                # Temporal positional encoding
+                maskmem_enc = (
+                    maskmem_enc + self.maskmem_tpos_enc[self.num_maskmem - t_pos - 1]
+                )
+                to_cat_memory_pos_embed.append(maskmem_enc)
+
+            # Construct the list of past object pointers
+            if self.use_obj_ptrs_in_encoder:
+                max_obj_ptrs_in_encoder = min(num_frames, self.max_obj_ptrs_in_encoder)
+                # First add those object pointers from selected conditioning frames
+                # (optionally, only include object pointers in the past during evaluation)
+                if not self.training and self.only_obj_ptrs_in_the_past_for_eval:
+                    ptr_cond_outputs = {
+                        t: out
+                        for t, out in selected_cond_outputs.items()
+                        if (t >= frame_idx if track_in_reverse else t <= frame_idx)
+                    }
+                else:
+                    ptr_cond_outputs = selected_cond_outputs
+                pos_and_ptrs = [
+                    # Temporal pos encoding contains how far away each pointer is from current frame
+                    (abs(frame_idx - t), out["obj_ptr"])
+                    for t, out in ptr_cond_outputs.items()
+                ]
+                # Add up to (max_obj_ptrs_in_encoder - 1) non-conditioning frames before current frame
+                for t_diff in range(1, max_obj_ptrs_in_encoder):
+                    t = frame_idx + t_diff if track_in_reverse else frame_idx - t_diff
+                    if t < 0 or (num_frames is not None and t >= num_frames):
+                        break
+                    out = output_dict["non_cond_frame_outputs"].get(
+                        t, unselected_cond_outputs.get(t, None)
+                    )
+                    if out is not None:
+                        pos_and_ptrs.append((t_diff, out["obj_ptr"]))
+                # If we have at least one object pointer, add them to the across attention
+                if len(pos_and_ptrs) > 0:
+                    pos_list, ptrs_list = zip(*pos_and_ptrs)
+                    # stack object pointers along dim=0 into [ptr_seq_len, B, C] shape
+                    obj_ptrs = torch.stack(ptrs_list, dim=0)
+                    # a temporal positional embedding based on how far each object pointer is from
+                    # the current frame (sine embedding normalized by the max pointer num).
+                    if self.add_tpos_enc_to_obj_ptrs:
+                        t_diff_max = max_obj_ptrs_in_encoder - 1
+                        tpos_dim = C if self.proj_tpos_enc_in_obj_ptrs else self.mem_dim
+                        obj_pos = torch.tensor(pos_list, device=device)
+                        obj_pos = get_1d_sine_pe(obj_pos / t_diff_max, dim=tpos_dim)
+                        obj_pos = self.obj_ptr_tpos_proj(obj_pos)
+                        obj_pos = obj_pos.unsqueeze(1).expand(-1, B, self.mem_dim)
+                    else:
+                        obj_pos = obj_ptrs.new_zeros(len(pos_list), B, self.mem_dim)
+                    if self.mem_dim < C:
+                        # split a pointer into (C // self.mem_dim) tokens for self.mem_dim < C
+                        obj_ptrs = obj_ptrs.reshape(
+                            -1, B, C // self.mem_dim, self.mem_dim
+                        )
+                        obj_ptrs = obj_ptrs.permute(0, 2, 1, 3).flatten(0, 1)
+                        obj_pos = obj_pos.repeat_interleave(C // self.mem_dim, dim=0)
+                    to_cat_memory.append(obj_ptrs)
+                    to_cat_memory_pos_embed.append(obj_pos)
+                    num_obj_ptr_tokens = obj_ptrs.shape[0]
+                else:
+                    num_obj_ptr_tokens = 0
+        else:
+            # for initial conditioning frames, encode them without using any previous memory
+            if self.directly_add_no_mem_embed:
+                # directly add no-mem embedding (instead of using the transformer encoder)
+                pix_feat_with_mem = current_vision_feats[-1] + self.no_mem_embed
+                pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+                return pix_feat_with_mem
+
+            # Use a dummy token on the first frame (to avoid emtpy memory input to tranformer encoder)
+            to_cat_memory = [self.no_mem_embed.expand(1, B, self.mem_dim)]
+            to_cat_memory_pos_embed = [self.no_mem_pos_enc.expand(1, B, self.mem_dim)]
+
+        # Step 2: Concatenate the memories and forward through the transformer encoder
+        memory = torch.cat(to_cat_memory, dim=0)
+        memory_pos_embed = torch.cat(to_cat_memory_pos_embed, dim=0)
+
+        pix_feat_with_mem = self.memory_attention(
+            curr=current_vision_feats,
+            curr_pos=current_vision_pos_embeds,
+            memory=memory,
+            memory_pos=memory_pos_embed,
+            num_obj_ptr_tokens=num_obj_ptr_tokens,
+        )
+        # reshape the output (HW)BC => BCHW
+        pix_feat_with_mem = pix_feat_with_mem.permute(1, 2, 0).view(B, C, H, W)
+        return pix_feat_with_mem
+
+    def _encode_new_memory(
+        self,
+        current_vision_feats,
+        feat_sizes,
+        pred_masks_high_res,
+        is_mask_from_pts,
+    ):
+        """Encode the current image and its prediction into a memory feature."""
+        B = current_vision_feats[-1].size(1)  # batch size on this frame
+        C = self.hidden_dim
+        H, W = feat_sizes[-1]  # top-level (lowest-resolution) feature size
+        # top-level feature, (HW)BC => BCHW
+        pix_feat = current_vision_feats[-1].permute(1, 2, 0).view(B, C, H, W)
+        if self.non_overlap_masks_for_mem_enc and not self.training:
+            # optionally, apply non-overlapping constraints to the masks (it's applied
+            # in the batch dimension and should only be used during eval, where all
+            # the objects come from the same video under batch size 1).
+            pred_masks_high_res = self._apply_non_overlapping_constraints(
+                pred_masks_high_res
+            )
+        # scale the raw mask logits with a temperature before applying sigmoid
+        binarize = self.binarize_mask_from_pts_for_mem_enc and is_mask_from_pts
+        if binarize and not self.training:
+            mask_for_mem = (pred_masks_high_res > 0).float()
+        else:
+            # apply sigmoid on the raw mask logits to turn them into range (0, 1)
+            mask_for_mem = torch.sigmoid(pred_masks_high_res)
+        # apply scale and bias terms to the sigmoid probabilities
+        if self.sigmoid_scale_for_mem_enc != 1.0:
+            mask_for_mem = mask_for_mem * self.sigmoid_scale_for_mem_enc
+        if self.sigmoid_bias_for_mem_enc != 0.0:
+            mask_for_mem = mask_for_mem + self.sigmoid_bias_for_mem_enc
+        maskmem_out = self.memory_encoder(
+            pix_feat, mask_for_mem, skip_mask_sigmoid=True  # sigmoid already applied
+        )
+        maskmem_features = maskmem_out["vision_features"]
+        maskmem_pos_enc = maskmem_out["vision_pos_enc"]
+
+        return maskmem_features, maskmem_pos_enc
+
+    def track_step(
+        self,
+        frame_idx,
+        is_init_cond_frame,
+        current_vision_feats,
+        current_vision_pos_embeds,
+        feat_sizes,
+        point_inputs,
+        mask_inputs,
+        output_dict,
+        num_frames,
+        track_in_reverse=False,  # tracking in reverse time order (for demo usage)
+        # Whether to run the memory encoder on the predicted masks. Sometimes we might want
+        # to skip the memory encoder with `run_mem_encoder=False`. For example,
+        # in demo we might call `track_step` multiple times for each user click,
+        # and only encode the memory when the user finalizes their clicks. And in ablation
+        # settings like SAM training on static images, we don't need the memory encoder.
+        run_mem_encoder=True,
+        # The previously predicted SAM mask logits (which can be fed together with new clicks in demo).
+        prev_sam_mask_logits=None,
+    ):
+        current_out = {"point_inputs": point_inputs, "mask_inputs": mask_inputs}
+        # High-resolution feature maps for the SAM head, reshape (HW)BC => BCHW
+        if len(current_vision_feats) > 1:
+            high_res_features = [
+                x.permute(1, 2, 0).view(x.size(1), x.size(2), *s)
+                for x, s in zip(current_vision_feats[:-1], feat_sizes[:-1])
+            ]
+        else:
+            high_res_features = None
+        if mask_inputs is not None and self.use_mask_input_as_output_without_sam:
+            # When use_mask_input_as_output_without_sam=True, we directly output the mask input
+            # (see it as a GT mask) without using a SAM prompt encoder + mask decoder.
+            pix_feat = current_vision_feats[-1].permute(1, 2, 0)
+            pix_feat = pix_feat.view(-1, self.hidden_dim, *feat_sizes[-1])
+            sam_outputs = self._use_mask_as_output(
+                pix_feat, high_res_features, mask_inputs
+            )
+        else:
+            # fused the visual feature with previous memory features in the memory bank
+            pix_feat_with_mem = self._prepare_memory_conditioned_features(
+                frame_idx=frame_idx,
+                is_init_cond_frame=is_init_cond_frame,
+                current_vision_feats=current_vision_feats[-1:],
+                current_vision_pos_embeds=current_vision_pos_embeds[-1:],
+                feat_sizes=feat_sizes[-1:],
+                output_dict=output_dict,
+                num_frames=num_frames,
+                track_in_reverse=track_in_reverse,
+            )
+            # apply SAM-style segmentation head
+            # here we might feed previously predicted low-res SAM mask logits into the SAM mask decoder,
+            # e.g. in demo where such logits come from earlier interaction instead of correction sampling
+            # (in this case, any `mask_inputs` shouldn't reach here as they are sent to _use_mask_as_output instead)
+            if prev_sam_mask_logits is not None:
+                assert point_inputs is not None and mask_inputs is None
+                mask_inputs = prev_sam_mask_logits
+            multimask_output = self._use_multimask(is_init_cond_frame, point_inputs)
+            sam_outputs = self._forward_sam_heads(
+                backbone_features=pix_feat_with_mem,
+                point_inputs=point_inputs,
+                mask_inputs=mask_inputs,
+                high_res_features=high_res_features,
+                multimask_output=multimask_output,
+            )
+        (
+            _,
+            _,
+            _,
+            low_res_masks,
+            high_res_masks,
+            obj_ptr,
+            _,
+        ) = sam_outputs
+
+        current_out["pred_masks"] = low_res_masks
+        current_out["pred_masks_high_res"] = high_res_masks
+        current_out["obj_ptr"] = obj_ptr
+
+        # Finally run the memory encoder on the predicted mask to encode
+        # it into a new memory feature (that can be used in future frames)
+        if run_mem_encoder and self.num_maskmem > 0:
+            high_res_masks_for_mem_enc = high_res_masks
+            maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+                current_vision_feats=current_vision_feats,
+                feat_sizes=feat_sizes,
+                pred_masks_high_res=high_res_masks_for_mem_enc,
+                is_mask_from_pts=(point_inputs is not None),
+            )
+            current_out["maskmem_features"] = maskmem_features
+            current_out["maskmem_pos_enc"] = maskmem_pos_enc
+        else:
+            current_out["maskmem_features"] = None
+            current_out["maskmem_pos_enc"] = None
+
+        return current_out
+
+    def _use_multimask(self, is_init_cond_frame, point_inputs):
+        """Whether to use multimask output in the SAM head."""
+        num_pts = 0 if point_inputs is None else point_inputs["point_labels"].size(1)
+        multimask_output = (
+            self.multimask_output_in_sam
+            and (is_init_cond_frame or self.multimask_output_for_tracking)
+            and (self.multimask_min_pt_num <= num_pts <= self.multimask_max_pt_num)
+        )
+        return multimask_output
+
+    def _apply_non_overlapping_constraints(self, pred_masks):
+        """
+        Apply non-overlapping constraints to the object scores in pred_masks. Here we
+        keep only the highest scoring object at each spatial location in pred_masks.
+        """
+        batch_size = pred_masks.size(0)
+        if batch_size == 1:
+            return pred_masks
+
+        device = pred_masks.device
+        # "max_obj_inds": object index of the object with the highest score at each location
+        max_obj_inds = torch.argmax(pred_masks, dim=0, keepdim=True)
+        # "batch_obj_inds": object index of each object slice (along dim 0) in `pred_masks`
+        batch_obj_inds = torch.arange(batch_size, device=device)[:, None, None, None]
+        keep = max_obj_inds == batch_obj_inds
+        # suppress overlapping regions' scores below -10.0 so that the foreground regions
+        # don't overlap (here sigmoid(-10.0)=4.5398e-05)
+        pred_masks = torch.where(keep, pred_masks, torch.clamp(pred_masks, max=-10.0))
+        return pred_masks
diff --git a/third_parts/sam2/modeling/sam2_utils.py b/third_parts/sam2/modeling/sam2_utils.py
new file mode 100644
index 0000000000000000000000000000000000000000..6d9705963efc57d74b7d1bff31692d7d293a46ad
--- /dev/null
+++ b/third_parts/sam2/modeling/sam2_utils.py
@@ -0,0 +1,149 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+
+import copy
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+
+def select_closest_cond_frames(frame_idx, cond_frame_outputs, max_cond_frame_num):
+    """
+    Select up to `max_cond_frame_num` conditioning frames from `cond_frame_outputs`
+    that are temporally closest to the current frame at `frame_idx`. Here, we take
+    - a) the closest conditioning frame before `frame_idx` (if any);
+    - b) the closest conditioning frame after `frame_idx` (if any);
+    - c) any other temporally closest conditioning frames until reaching a total
+         of `max_cond_frame_num` conditioning frames.
+
+    Outputs:
+    - selected_outputs: selected items (keys & values) from `cond_frame_outputs`.
+    - unselected_outputs: items (keys & values) not selected in `cond_frame_outputs`.
+    """
+    if max_cond_frame_num == -1 or len(cond_frame_outputs) <= max_cond_frame_num:
+        selected_outputs = cond_frame_outputs
+        unselected_outputs = {}
+    else:
+        assert max_cond_frame_num >= 2, "we should allow using 2+ conditioning frames"
+        selected_outputs = {}
+
+        # the closest conditioning frame before `frame_idx` (if any)
+        idx_before = max((t for t in cond_frame_outputs if t < frame_idx), default=None)
+        if idx_before is not None:
+            selected_outputs[idx_before] = cond_frame_outputs[idx_before]
+
+        # the closest conditioning frame after `frame_idx` (if any)
+        idx_after = min((t for t in cond_frame_outputs if t >= frame_idx), default=None)
+        if idx_after is not None:
+            selected_outputs[idx_after] = cond_frame_outputs[idx_after]
+
+        # add other temporally closest conditioning frames until reaching a total
+        # of `max_cond_frame_num` conditioning frames.
+        num_remain = max_cond_frame_num - len(selected_outputs)
+        inds_remain = sorted(
+            (t for t in cond_frame_outputs if t not in selected_outputs),
+            key=lambda x: abs(x - frame_idx),
+        )[:num_remain]
+        selected_outputs.update((t, cond_frame_outputs[t]) for t in inds_remain)
+        unselected_outputs = {
+            t: v for t, v in cond_frame_outputs.items() if t not in selected_outputs
+        }
+
+    return selected_outputs, unselected_outputs
+
+
+def get_1d_sine_pe(pos_inds, dim, temperature=10000):
+    """
+    Get 1D sine positional embedding as in the original Transformer paper.
+    """
+    pe_dim = dim // 2
+    dim_t = torch.arange(pe_dim, dtype=torch.float32, device=pos_inds.device)
+    dim_t = temperature ** (2 * (dim_t // 2) / pe_dim)
+
+    pos_embed = pos_inds.unsqueeze(-1) / dim_t
+    pos_embed = torch.cat([pos_embed.sin(), pos_embed.cos()], dim=-1)
+    return pos_embed
+
+
+def get_activation_fn(activation):
+    """Return an activation function given a string"""
+    if activation == "relu":
+        return F.relu
+    if activation == "gelu":
+        return F.gelu
+    if activation == "glu":
+        return F.glu
+    raise RuntimeError(f"activation should be relu/gelu, not {activation}.")
+
+
+def get_clones(module, N):
+    return nn.ModuleList([copy.deepcopy(module) for i in range(N)])
+
+
+class DropPath(nn.Module):
+    # adapted from https://github.com/huggingface/pytorch-image-models/blob/main/timm/layers/drop.py
+    def __init__(self, drop_prob=0.0, scale_by_keep=True):
+        super(DropPath, self).__init__()
+        self.drop_prob = drop_prob
+        self.scale_by_keep = scale_by_keep
+
+    def forward(self, x):
+        if self.drop_prob == 0.0 or not self.training:
+            return x
+        keep_prob = 1 - self.drop_prob
+        shape = (x.shape[0],) + (1,) * (x.ndim - 1)
+        random_tensor = x.new_empty(shape).bernoulli_(keep_prob)
+        if keep_prob > 0.0 and self.scale_by_keep:
+            random_tensor.div_(keep_prob)
+        return x * random_tensor
+
+
+# Lightly adapted from
+# https://github.com/facebookresearch/MaskFormer/blob/main/mask_former/modeling/transformer/transformer_predictor.py # noqa
+class MLP(nn.Module):
+    def __init__(
+        self,
+        input_dim: int,
+        hidden_dim: int,
+        output_dim: int,
+        num_layers: int,
+        activation: nn.Module = nn.ReLU,
+        sigmoid_output: bool = False,
+    ) -> None:
+        super().__init__()
+        self.num_layers = num_layers
+        h = [hidden_dim] * (num_layers - 1)
+        self.layers = nn.ModuleList(
+            nn.Linear(n, k) for n, k in zip([input_dim] + h, h + [output_dim])
+        )
+        self.sigmoid_output = sigmoid_output
+        self.act = activation()
+
+    def forward(self, x):
+        for i, layer in enumerate(self.layers):
+            x = self.act(layer(x)) if i < self.num_layers - 1 else layer(x)
+        if self.sigmoid_output:
+            x = F.sigmoid(x)
+        return x
+
+
+# From https://github.com/facebookresearch/detectron2/blob/main/detectron2/layers/batch_norm.py # noqa
+# Itself from https://github.com/facebookresearch/ConvNeXt/blob/d1fa8f6fef0a165b27399986cc2bdacc92777e40/models/convnext.py#L119  # noqa
+class LayerNorm2d(nn.Module):
+    def __init__(self, num_channels: int, eps: float = 1e-6) -> None:
+        super().__init__()
+        self.weight = nn.Parameter(torch.ones(num_channels))
+        self.bias = nn.Parameter(torch.zeros(num_channels))
+        self.eps = eps
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        u = x.mean(1, keepdim=True)
+        s = (x - u).pow(2).mean(1, keepdim=True)
+        x = (x - u) / torch.sqrt(s + self.eps)
+        x = self.weight[:, None, None] * x + self.bias[:, None, None]
+        return x
diff --git a/third_parts/sam2/sam2_configs/__init__.py b/third_parts/sam2/sam2_configs/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5277f46157403e47fd830fc519144b97ef69d4ae
--- /dev/null
+++ b/third_parts/sam2/sam2_configs/__init__.py
@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
diff --git a/third_parts/sam2/sam2_configs/sam2_hiera_b+.yaml b/third_parts/sam2/sam2_configs/sam2_hiera_b+.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..509ce3d030a56ea08026c894223ccc3bc1de9b90
--- /dev/null
+++ b/third_parts/sam2/sam2_configs/sam2_hiera_b+.yaml
@@ -0,0 +1,113 @@
+# @package _global_
+
+# Model
+model:
+  _target_: third_parts.sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: third_parts.sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: third_parts.sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 112
+      num_heads: 2
+    neck:
+      _target_: third_parts.sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [896, 448, 224, 112]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: third_parts.sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: third_parts.sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: third_parts.sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: third_parts.sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: third_parts.sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: third_parts.sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False
diff --git a/third_parts/sam2/sam2_configs/sam2_hiera_l.yaml b/third_parts/sam2/sam2_configs/sam2_hiera_l.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..f00a7db45789148463c39be186f6a7b53e1f21ad
--- /dev/null
+++ b/third_parts/sam2/sam2_configs/sam2_hiera_l.yaml
@@ -0,0 +1,117 @@
+# @package _global_
+
+# Model
+model:
+  _target_: third_parts.sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: third_parts.sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: third_parts.sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 144
+      num_heads: 2
+      stages: [2, 6, 36, 4]
+      global_att_blocks: [23, 33, 43]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+      window_spec: [8, 4, 16, 8]
+    neck:
+      _target_: third_parts.sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [1152, 576, 288, 144]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: third_parts.sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: third_parts.sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: third_parts.sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: third_parts.sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: third_parts.sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: third_parts.sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False
diff --git a/third_parts/sam2/sam2_configs/sam2_hiera_s.yaml b/third_parts/sam2/sam2_configs/sam2_hiera_s.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..08a767bd833e8d73b66960f7ba6998eeaaf673c8
--- /dev/null
+++ b/third_parts/sam2/sam2_configs/sam2_hiera_s.yaml
@@ -0,0 +1,116 @@
+# @package _global_
+
+# Model
+model:
+  _target_: third_parts.sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: third_parts.sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: third_parts.sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 11, 2]
+      global_att_blocks: [7, 10, 13]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: third_parts.sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: third_parts.sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: third_parts.sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: third_parts.sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: third_parts.sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: third_parts.sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: third_parts.sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  compile_image_encoder: False
diff --git a/third_parts/sam2/sam2_configs/sam2_hiera_t.yaml b/third_parts/sam2/sam2_configs/sam2_hiera_t.yaml
new file mode 100644
index 0000000000000000000000000000000000000000..bb681b73fc95527566b28ffeeb71d6408dae7b3c
--- /dev/null
+++ b/third_parts/sam2/sam2_configs/sam2_hiera_t.yaml
@@ -0,0 +1,118 @@
+# @package _global_
+
+# Model
+model:
+  _target_: third_parts.sam2.modeling.sam2_base.SAM2Base
+  image_encoder:
+    _target_: third_parts.sam2.modeling.backbones.image_encoder.ImageEncoder
+    scalp: 1
+    trunk:
+      _target_: third_parts.sam2.modeling.backbones.hieradet.Hiera
+      embed_dim: 96
+      num_heads: 1
+      stages: [1, 2, 7, 2]
+      global_att_blocks: [5, 7, 9]
+      window_pos_embed_bkg_spatial_size: [7, 7]
+    neck:
+      _target_: third_parts.sam2.modeling.backbones.image_encoder.FpnNeck
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 256
+        normalize: true
+        scale: null
+        temperature: 10000
+      d_model: 256
+      backbone_channel_list: [768, 384, 192, 96]
+      fpn_top_down_levels: [2, 3]  # output level 0 and 1 directly use the backbone features
+      fpn_interp_model: nearest
+
+  memory_attention:
+    _target_: third_parts.sam2.modeling.memory_attention.MemoryAttention
+    d_model: 256
+    pos_enc_at_input: true
+    layer:
+      _target_: third_parts.sam2.modeling.memory_attention.MemoryAttentionLayer
+      activation: relu
+      dim_feedforward: 2048
+      dropout: 0.1
+      pos_enc_at_attn: false
+      self_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+      d_model: 256
+      pos_enc_at_cross_attn_keys: true
+      pos_enc_at_cross_attn_queries: false
+      cross_attention:
+        _target_: third_parts.sam2.modeling.sam.transformer.RoPEAttention
+        rope_theta: 10000.0
+        feat_sizes: [32, 32]
+        rope_k_repeat: True
+        embedding_dim: 256
+        num_heads: 1
+        downsample_rate: 1
+        dropout: 0.1
+        kv_in_dim: 64
+    num_layers: 4
+
+  memory_encoder:
+      _target_: third_parts.sam2.modeling.memory_encoder.MemoryEncoder
+      out_dim: 64
+      position_encoding:
+        _target_: third_parts.sam2.modeling.position_encoding.PositionEmbeddingSine
+        num_pos_feats: 64
+        normalize: true
+        scale: null
+        temperature: 10000
+      mask_downsampler:
+        _target_: third_parts.sam2.modeling.memory_encoder.MaskDownSampler
+        kernel_size: 3
+        stride: 2
+        padding: 1
+      fuser:
+        _target_: third_parts.sam2.modeling.memory_encoder.Fuser
+        layer:
+          _target_: third_parts.sam2.modeling.memory_encoder.CXBlock
+          dim: 256
+          kernel_size: 7
+          padding: 3
+          layer_scale_init_value: 1e-6
+          use_dwconv: True  # depth-wise convs
+        num_layers: 2
+
+  num_maskmem: 7
+  image_size: 1024
+  # apply scaled sigmoid on mask logits for memory encoder, and directly feed input mask as output mask
+  # SAM decoder
+  sigmoid_scale_for_mem_enc: 20.0
+  sigmoid_bias_for_mem_enc: -10.0
+  use_mask_input_as_output_without_sam: true
+  # Memory
+  directly_add_no_mem_embed: true
+  # use high-resolution feature map in the SAM mask decoder
+  use_high_res_features_in_sam: true
+  # output 3 masks on the first click on initial conditioning frames
+  multimask_output_in_sam: true
+  # SAM heads
+  iou_prediction_use_sigmoid: True
+  # cross-attend to object pointers from other frames (based on SAM output tokens) in the encoder
+  use_obj_ptrs_in_encoder: true
+  add_tpos_enc_to_obj_ptrs: false
+  only_obj_ptrs_in_the_past_for_eval: true
+  # object occlusion prediction
+  pred_obj_scores: true
+  pred_obj_scores_mlp: true
+  fixed_no_obj_ptr: true
+  # multimask tracking settings
+  multimask_output_for_tracking: true
+  use_multimask_token_for_obj_ptr: true
+  multimask_min_pt_num: 0
+  multimask_max_pt_num: 1
+  use_mlp_for_obj_ptr_proj: true
+  # Compilation flag
+  # HieraT does not currently support compilation, should always be set to False
+  compile_image_encoder: False
diff --git a/third_parts/sam2/sam2_image_predictor.py b/third_parts/sam2/sam2_image_predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..d8d1d6ac5178e05c75ccb0683d27b4e83e7014b9
--- /dev/null
+++ b/third_parts/sam2/sam2_image_predictor.py
@@ -0,0 +1,446 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import logging
+
+from typing import List, Optional, Tuple, Union
+
+import numpy as np
+import torch
+from PIL.Image import Image
+
+from third_parts.sam2.modeling.sam2_base import SAM2Base
+
+from third_parts.sam2.utils.transforms import SAM2Transforms
+
+
+class SAM2ImagePredictor:
+    def __init__(
+        self,
+        sam_model: SAM2Base,
+        mask_threshold=0.0,
+        max_hole_area=0.0,
+        max_sprinkle_area=0.0,
+    ) -> None:
+        """
+        Uses SAM-2 to calculate the image embedding for an image, and then
+        allow repeated, efficient mask prediction given prompts.
+
+        Arguments:
+          sam_model (Sam-2): The model to use for mask prediction.
+          mask_threshold (float): The threshold to use when converting mask logits
+            to binary masks. Masks are thresholded at 0 by default.
+          fill_hole_area (int): If fill_hole_area > 0, we fill small holes in up to
+            the maximum area of fill_hole_area in low_res_masks.
+        """
+        super().__init__()
+        self.model = sam_model
+        self._transforms = SAM2Transforms(
+            resolution=self.model.image_size,
+            mask_threshold=mask_threshold,
+            max_hole_area=max_hole_area,
+            max_sprinkle_area=max_sprinkle_area,
+        )
+
+        # Predictor state
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        # Whether the predictor is set for single image or a batch of images
+        self._is_batch = False
+
+        # Predictor config
+        self.mask_threshold = mask_threshold
+
+        # Spatial dim for backbone feature maps
+        self._bb_feat_sizes = [
+            (256, 256),
+            (128, 128),
+            (64, 64),
+        ]
+
+    @torch.no_grad()
+    def set_image(
+        self,
+        image: Union[np.ndarray, Image],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image, allowing
+        masks to be predicted with the 'predict' method.
+
+        Arguments:
+          image (np.ndarray or PIL Image): The input image to embed in RGB format. The image should be in HWC format if np.ndarray, or WHC format if PIL Image
+          with pixel values in [0, 255].
+          image_format (str): The color format of the image, in ['RGB', 'BGR'].
+        """
+        self.reset_predictor()
+        # Transform the image to the form expected by the model
+        if isinstance(image, np.ndarray):
+            logging.info("For numpy array image, we assume (HxWxC) format")
+            self._orig_hw = [image.shape[:2]]
+        elif isinstance(image, Image):
+            w, h = image.size
+            self._orig_hw = [(h, w)]
+        else:
+            raise NotImplementedError("Image format not supported")
+
+        input_image = self._transforms(image)
+        input_image = input_image[None, ...].to(self.device)
+
+        assert (
+            len(input_image.shape) == 4 and input_image.shape[1] == 3
+        ), f"input_image must be of size 1x3xHxW, got {input_image.shape}"
+        logging.info("Computing image embeddings for the provided image...")
+        backbone_out = self.model.forward_image(input_image)
+        _, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        if self.model.directly_add_no_mem_embed:
+            vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(1, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        logging.info("Image embeddings computed.")
+
+    @torch.no_grad()
+    def set_image_batch(
+        self,
+        image_list: List[Union[np.ndarray]],
+    ) -> None:
+        """
+        Calculates the image embeddings for the provided image batch, allowing
+        masks to be predicted with the 'predict_batch' method.
+
+        Arguments:
+          image_list (List[np.ndarray]): The input images to embed in RGB format. The image should be in HWC format if np.ndarray
+          with pixel values in [0, 255].
+        """
+        self.reset_predictor()
+        assert isinstance(image_list, list)
+        self._orig_hw = []
+        for image in image_list:
+            assert isinstance(
+                image, np.ndarray
+            ), "Images are expected to be an np.ndarray in RGB format, and of shape  HWC"
+            self._orig_hw.append(image.shape[:2])
+        # Transform the image to the form expected by the model
+        img_batch = self._transforms.forward_batch(image_list)
+        img_batch = img_batch.to(self.device)
+        batch_size = img_batch.shape[0]
+        assert (
+            len(img_batch.shape) == 4 and img_batch.shape[1] == 3
+        ), f"img_batch must be of size Bx3xHxW, got {img_batch.shape}"
+        logging.info("Computing image embeddings for the provided images...")
+        backbone_out = self.model.forward_image(img_batch)
+        _, vision_feats, _, _ = self.model._prepare_backbone_features(backbone_out)
+        # Add no_mem_embed, which is added to the lowest rest feat. map during training on videos
+        if self.model.directly_add_no_mem_embed:
+            vision_feats[-1] = vision_feats[-1] + self.model.no_mem_embed
+
+        feats = [
+            feat.permute(1, 2, 0).view(batch_size, -1, *feat_size)
+            for feat, feat_size in zip(vision_feats[::-1], self._bb_feat_sizes[::-1])
+        ][::-1]
+        self._features = {"image_embed": feats[-1], "high_res_feats": feats[:-1]}
+        self._is_image_set = True
+        self._is_batch = True
+        logging.info("Image embeddings computed.")
+
+    def predict_batch(
+        self,
+        point_coords_batch: List[np.ndarray] = None,
+        point_labels_batch: List[np.ndarray] = None,
+        box_batch: List[np.ndarray] = None,
+        mask_input_batch: List[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[List[np.ndarray], List[np.ndarray], List[np.ndarray]]:
+        """This function is very similar to predict(...), however it is used for batched mode, when the model is expected to generate predictions on multiple images.
+        It returns a tupele of lists of masks, ious, and low_res_masks_logits.
+        """
+        assert self._is_batch, "This function should only be used when in batched mode"
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image_batch(...) before mask prediction."
+            )
+        num_images = len(self._features["image_embed"])
+        all_masks = []
+        all_ious = []
+        all_low_res_masks = []
+        for img_idx in range(num_images):
+            # Transform input prompts
+            point_coords = (
+                point_coords_batch[img_idx] if point_coords_batch is not None else None
+            )
+            point_labels = (
+                point_labels_batch[img_idx] if point_labels_batch is not None else None
+            )
+            box = box_batch[img_idx] if box_batch is not None else None
+            mask_input = (
+                mask_input_batch[img_idx] if mask_input_batch is not None else None
+            )
+            mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+                point_coords,
+                point_labels,
+                box,
+                mask_input,
+                normalize_coords,
+                img_idx=img_idx,
+            )
+            masks, iou_predictions, low_res_masks = self._predict(
+                unnorm_coords,
+                labels,
+                unnorm_box,
+                mask_input,
+                multimask_output,
+                return_logits=return_logits,
+                img_idx=img_idx,
+            )
+            masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+            iou_predictions_np = (
+                iou_predictions.squeeze(0).float().detach().cpu().numpy()
+            )
+            low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+            all_masks.append(masks_np)
+            all_ious.append(iou_predictions_np)
+            all_low_res_masks.append(low_res_masks_np)
+
+        return all_masks, all_ious, all_low_res_masks
+
+    def predict(
+        self,
+        point_coords: Optional[np.ndarray] = None,
+        point_labels: Optional[np.ndarray] = None,
+        box: Optional[np.ndarray] = None,
+        mask_input: Optional[np.ndarray] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        normalize_coords=True,
+    ) -> Tuple[np.ndarray, np.ndarray, np.ndarray]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+
+        Arguments:
+          point_coords (np.ndarray or None): A Nx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (np.ndarray or None): A length N array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          box (np.ndarray or None): A length 4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form 1xHxW, where
+            for SAM, H=W=256.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+          normalize_coords (bool): If true, the point coordinates will be normalized to the range [0,1] and point_coords is expected to be wrt. image dimensions.
+
+        Returns:
+          (np.ndarray): The output masks in CxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (np.ndarray): An array of length C containing the model's
+            predictions for the quality of each mask.
+          (np.ndarray): An array of shape CxHxW, where C is the number
+            of masks and H=W=256. These low resolution logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        # Transform input prompts
+
+        mask_input, unnorm_coords, labels, unnorm_box = self._prep_prompts(
+            point_coords, point_labels, box, mask_input, normalize_coords
+        )
+
+        masks, iou_predictions, low_res_masks = self._predict(
+            unnorm_coords,
+            labels,
+            unnorm_box,
+            mask_input,
+            multimask_output,
+            return_logits=return_logits,
+        )
+
+        masks_np = masks.squeeze(0).float().detach().cpu().numpy()
+        iou_predictions_np = iou_predictions.squeeze(0).float().detach().cpu().numpy()
+        low_res_masks_np = low_res_masks.squeeze(0).float().detach().cpu().numpy()
+        return masks_np, iou_predictions_np, low_res_masks_np
+
+    def _prep_prompts(
+        self, point_coords, point_labels, box, mask_logits, normalize_coords, img_idx=-1
+    ):
+
+        unnorm_coords, labels, unnorm_box, mask_input = None, None, None, None
+        if point_coords is not None:
+            assert (
+                point_labels is not None
+            ), "point_labels must be supplied if point_coords is supplied."
+            point_coords = torch.as_tensor(
+                point_coords, dtype=torch.float, device=self.device
+            )
+            unnorm_coords = self._transforms.transform_coords(
+                point_coords, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )
+            labels = torch.as_tensor(point_labels, dtype=torch.int, device=self.device)
+            if len(unnorm_coords.shape) == 2:
+                unnorm_coords, labels = unnorm_coords[None, ...], labels[None, ...]
+        if box is not None:
+            box = torch.as_tensor(box, dtype=torch.float, device=self.device)
+            unnorm_box = self._transforms.transform_boxes(
+                box, normalize=normalize_coords, orig_hw=self._orig_hw[img_idx]
+            )  # Bx2x2
+        if mask_logits is not None:
+            mask_input = torch.as_tensor(
+                mask_logits, dtype=torch.float, device=self.device
+            )
+            if len(mask_input.shape) == 3:
+                mask_input = mask_input[None, :, :, :]
+        return mask_input, unnorm_coords, labels, unnorm_box
+
+    @torch.no_grad()
+    def _predict(
+        self,
+        point_coords: Optional[torch.Tensor],
+        point_labels: Optional[torch.Tensor],
+        boxes: Optional[torch.Tensor] = None,
+        mask_input: Optional[torch.Tensor] = None,
+        multimask_output: bool = True,
+        return_logits: bool = False,
+        img_idx: int = -1,
+    ) -> Tuple[torch.Tensor, torch.Tensor, torch.Tensor]:
+        """
+        Predict masks for the given input prompts, using the currently set image.
+        Input prompts are batched torch tensors and are expected to already be
+        transformed to the input frame using SAM2Transforms.
+
+        Arguments:
+          point_coords (torch.Tensor or None): A BxNx2 array of point prompts to the
+            model. Each point is in (X,Y) in pixels.
+          point_labels (torch.Tensor or None): A BxN array of labels for the
+            point prompts. 1 indicates a foreground point and 0 indicates a
+            background point.
+          boxes (np.ndarray or None): A Bx4 array given a box prompt to the
+            model, in XYXY format.
+          mask_input (np.ndarray): A low resolution mask input to the model, typically
+            coming from a previous prediction iteration. Has form Bx1xHxW, where
+            for SAM, H=W=256. Masks returned by a previous iteration of the
+            predict method do not need further transformation.
+          multimask_output (bool): If true, the model will return three masks.
+            For ambiguous input prompts (such as a single click), this will often
+            produce better masks than a single prediction. If only a single
+            mask is needed, the model's predicted quality score can be used
+            to select the best mask. For non-ambiguous prompts, such as multiple
+            input prompts, multimask_output=False can give better results.
+          return_logits (bool): If true, returns un-thresholded masks logits
+            instead of a binary mask.
+
+        Returns:
+          (torch.Tensor): The output masks in BxCxHxW format, where C is the
+            number of masks, and (H, W) is the original image size.
+          (torch.Tensor): An array of shape BxC containing the model's
+            predictions for the quality of each mask.
+          (torch.Tensor): An array of shape BxCxHxW, where C is the number
+            of masks and H=W=256. These low res logits can be passed to
+            a subsequent iteration as mask input.
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) before mask prediction."
+            )
+
+        if point_coords is not None:
+            concat_points = (point_coords, point_labels)
+        else:
+            concat_points = None
+
+        # Embed prompts
+        if boxes is not None:
+            box_coords = boxes.reshape(-1, 2, 2)
+            box_labels = torch.tensor([[2, 3]], dtype=torch.int, device=boxes.device)
+            box_labels = box_labels.repeat(boxes.size(0), 1)
+            # we merge "boxes" and "points" into a single "concat_points" input (where
+            # boxes are added at the beginning) to sam_prompt_encoder
+            if concat_points is not None:
+                concat_coords = torch.cat([box_coords, concat_points[0]], dim=1)
+                concat_labels = torch.cat([box_labels, concat_points[1]], dim=1)
+                concat_points = (concat_coords, concat_labels)
+            else:
+                concat_points = (box_coords, box_labels)
+
+        sparse_embeddings, dense_embeddings = self.model.sam_prompt_encoder(
+            points=concat_points,
+            boxes=None,
+            masks=mask_input,
+        )
+
+        # Predict masks
+        batched_mode = (
+            concat_points is not None and concat_points[0].shape[0] > 1
+        )  # multi object prediction
+        high_res_features = [
+            feat_level[img_idx].unsqueeze(0)
+            for feat_level in self._features["high_res_feats"]
+        ]
+        low_res_masks, iou_predictions, _, _ = self.model.sam_mask_decoder(
+            image_embeddings=self._features["image_embed"][img_idx].unsqueeze(0),
+            image_pe=self.model.sam_prompt_encoder.get_dense_pe(),
+            sparse_prompt_embeddings=sparse_embeddings,
+            dense_prompt_embeddings=dense_embeddings,
+            multimask_output=multimask_output,
+            repeat_image=batched_mode,
+            high_res_features=high_res_features,
+        )
+
+        # Upscale the masks to the original image resolution
+        masks = self._transforms.postprocess_masks(
+            low_res_masks, self._orig_hw[img_idx]
+        )
+        low_res_masks = torch.clamp(low_res_masks, -32.0, 32.0)
+        if not return_logits:
+            masks = masks > self.mask_threshold
+
+        return masks, iou_predictions, low_res_masks
+
+    def get_image_embedding(self) -> torch.Tensor:
+        """
+        Returns the image embeddings for the currently set image, with
+        shape 1xCxHxW, where C is the embedding dimension and (H,W) are
+        the embedding spatial dimension of SAM (typically C=256, H=W=64).
+        """
+        if not self._is_image_set:
+            raise RuntimeError(
+                "An image must be set with .set_image(...) to generate an embedding."
+            )
+        assert (
+            self._features is not None
+        ), "Features must exist if an image has been set."
+        return self._features["image_embed"]
+
+    @property
+    def device(self) -> torch.device:
+        return self.model.device
+
+    def reset_predictor(self) -> None:
+        """
+        Resets the image embeddings and other state variables.
+        """
+        self._is_image_set = False
+        self._features = None
+        self._orig_hw = None
+        self._is_batch = False
diff --git a/third_parts/sam2/sam2_video_predictor.py b/third_parts/sam2/sam2_video_predictor.py
new file mode 100644
index 0000000000000000000000000000000000000000..5bad0d4e69f5d8bc2160d61acb248df501b07b92
--- /dev/null
+++ b/third_parts/sam2/sam2_video_predictor.py
@@ -0,0 +1,898 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+from collections import OrderedDict
+
+import torch
+
+from tqdm import tqdm
+
+from third_parts.sam2.modeling.sam2_base import NO_OBJ_SCORE, SAM2Base
+from third_parts.sam2.utils.misc import concat_points, fill_holes_in_mask_scores, load_video_frames
+
+
+class SAM2VideoPredictor(SAM2Base):
+    """The predictor class to handle user interactions and manage inference states."""
+
+    def __init__(
+        self,
+        fill_hole_area=0,
+        # whether to apply non-overlapping constraints on the output object masks
+        non_overlap_masks=False,
+        # whether to clear non-conditioning memory of the surrounding frames (which may contain outdated information) after adding correction clicks;
+        # note that this would only apply to *single-object tracking* unless `clear_non_cond_mem_for_multi_obj` is also set to True)
+        clear_non_cond_mem_around_input=False,
+        # whether to also clear non-conditioning memory of the surrounding frames (only effective when `clear_non_cond_mem_around_input` is True).
+        clear_non_cond_mem_for_multi_obj=False,
+        **kwargs,
+    ):
+        super().__init__(**kwargs)
+        self.fill_hole_area = fill_hole_area
+        self.non_overlap_masks = non_overlap_masks
+        self.clear_non_cond_mem_around_input = clear_non_cond_mem_around_input
+        self.clear_non_cond_mem_for_multi_obj = clear_non_cond_mem_for_multi_obj
+
+    @torch.inference_mode()
+    def init_state(
+        self,
+        video_path,
+        offload_video_to_cpu=False,
+        offload_state_to_cpu=False,
+        async_loading_frames=False,
+    ):
+        """Initialize a inference state."""
+        images, video_height, video_width = load_video_frames(
+            video_path=video_path,
+            image_size=self.image_size,
+            offload_video_to_cpu=offload_video_to_cpu,
+            async_loading_frames=async_loading_frames,
+        )
+        inference_state = {}
+        inference_state["images"] = images
+        inference_state["num_frames"] = len(images)
+        # whether to offload the video frames to CPU memory
+        # turning on this option saves the GPU memory with only a very small overhead
+        inference_state["offload_video_to_cpu"] = offload_video_to_cpu
+        # whether to offload the inference state to CPU memory
+        # turning on this option saves the GPU memory at the cost of a lower tracking fps
+        # (e.g. in a test case of 768x768 model, fps dropped from 27 to 24 when tracking one object
+        # and from 24 to 21 when tracking two objects)
+        inference_state["offload_state_to_cpu"] = offload_state_to_cpu
+        # the original video height and width, used for resizing final output scores
+        inference_state["video_height"] = video_height
+        inference_state["video_width"] = video_width
+        inference_state["device"] = torch.device("cuda")
+        if offload_state_to_cpu:
+            inference_state["storage_device"] = torch.device("cpu")
+        else:
+            inference_state["storage_device"] = torch.device("cuda")
+        # inputs on each frame
+        inference_state["point_inputs_per_obj"] = {}
+        inference_state["mask_inputs_per_obj"] = {}
+        # visual features on a small number of recently visited frames for quick interactions
+        inference_state["cached_features"] = {}
+        # values that don't change across frames (so we only need to hold one copy of them)
+        inference_state["constants"] = {}
+        # mapping between client-side object id and model-side object index
+        inference_state["obj_id_to_idx"] = OrderedDict()
+        inference_state["obj_idx_to_id"] = OrderedDict()
+        inference_state["obj_ids"] = []
+        # A storage to hold the model's tracking results and states on each frame
+        inference_state["output_dict"] = {
+            "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+        }
+        # Slice (view) of each object tracking results, sharing the same memory with "output_dict"
+        inference_state["output_dict_per_obj"] = {}
+        # A temporary storage to hold new outputs when user interact with a frame
+        # to add clicks or mask (it's merged into "output_dict" before propagation starts)
+        inference_state["temp_output_dict_per_obj"] = {}
+        # Frames that already holds consolidated outputs from click or mask inputs
+        # (we directly use their consolidated outputs during tracking)
+        inference_state["consolidated_frame_inds"] = {
+            "cond_frame_outputs": set(),  # set containing frame indices
+            "non_cond_frame_outputs": set(),  # set containing frame indices
+        }
+        # metadata for each tracking frame (e.g. which direction it's tracked)
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"] = {}
+        # Warm up the visual backbone and cache the image feature on frame 0
+        self._get_image_feature(inference_state, frame_idx=0, batch_size=1)
+        return inference_state
+
+    def _obj_id_to_idx(self, inference_state, obj_id):
+        """Map client-side object id to model-side object index."""
+        obj_idx = inference_state["obj_id_to_idx"].get(obj_id, None)
+        if obj_idx is not None:
+            return obj_idx
+
+        # This is a new object id not sent to the server before. We only allow adding
+        # new objects *before* the tracking starts.
+        allow_new_object = not inference_state["tracking_has_started"]
+        if allow_new_object:
+            # get the next object slot
+            obj_idx = len(inference_state["obj_id_to_idx"])
+            inference_state["obj_id_to_idx"][obj_id] = obj_idx
+            inference_state["obj_idx_to_id"][obj_idx] = obj_id
+            inference_state["obj_ids"] = list(inference_state["obj_id_to_idx"])
+            # set up input and output structures for this object
+            inference_state["point_inputs_per_obj"][obj_idx] = {}
+            inference_state["mask_inputs_per_obj"][obj_idx] = {}
+            inference_state["output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            inference_state["temp_output_dict_per_obj"][obj_idx] = {
+                "cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+                "non_cond_frame_outputs": {},  # dict containing {frame_idx: <out>}
+            }
+            return obj_idx
+        else:
+            raise RuntimeError(
+                f"Cannot add new object id {obj_id} after tracking starts. "
+                f"All existing object ids: {inference_state['obj_ids']}. "
+                f"Please call 'reset_state' to restart from scratch."
+            )
+
+    def _obj_idx_to_id(self, inference_state, obj_idx):
+        """Map model-side object index to client-side object id."""
+        return inference_state["obj_idx_to_id"][obj_idx]
+
+    def _get_obj_num(self, inference_state):
+        """Get the total number of unique object ids received so far in this session."""
+        return len(inference_state["obj_idx_to_id"])
+
+    @torch.inference_mode()
+    def add_new_points(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        points,
+        labels,
+        clear_old_points=True,
+        normalize_coords=True,
+    ):
+        """Add new points to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if not isinstance(points, torch.Tensor):
+            points = torch.tensor(points, dtype=torch.float32)
+        if not isinstance(labels, torch.Tensor):
+            labels = torch.tensor(labels, dtype=torch.int32)
+        if points.dim() == 2:
+            points = points.unsqueeze(0)  # add batch dimension
+        if labels.dim() == 1:
+            labels = labels.unsqueeze(0)  # add batch dimension
+        if normalize_coords:
+            video_H = inference_state["video_height"]
+            video_W = inference_state["video_width"]
+            points = points / torch.tensor([video_W, video_H]).to(points.device)
+        # scale the (normalized) coordinates by the model's internal image size
+        points = points * self.image_size
+        points = points.to(inference_state["device"])
+        labels = labels.to(inference_state["device"])
+
+        if not clear_old_points:
+            point_inputs = point_inputs_per_frame.get(frame_idx, None)
+        else:
+            point_inputs = None
+        point_inputs = concat_points(point_inputs, points, labels)
+
+        point_inputs_per_frame[frame_idx] = point_inputs
+        mask_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        # Get any previously predicted mask logits on this object and feed it along with
+        # the new clicks into the SAM mask decoder.
+        prev_sam_mask_logits = None
+        # lookup temporary output dict first, which contains the most recent output
+        # (if not found, then lookup conditioning and non-conditioning frame output)
+        prev_out = obj_temp_output_dict[storage_key].get(frame_idx)
+        if prev_out is None:
+            prev_out = obj_output_dict["cond_frame_outputs"].get(frame_idx)
+            if prev_out is None:
+                prev_out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx)
+
+        if prev_out is not None and prev_out["pred_masks"] is not None:
+            prev_sam_mask_logits = prev_out["pred_masks"].cuda(non_blocking=True)
+            # Clamp the scale of prev_sam_mask_logits to avoid rare numerical issues.
+            prev_sam_mask_logits = torch.clamp(prev_sam_mask_logits, -32.0, 32.0)
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=point_inputs,
+            mask_inputs=None,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    @torch.inference_mode()
+    def add_new_mask(
+        self,
+        inference_state,
+        frame_idx,
+        obj_id,
+        mask,
+    ):
+        """Add new mask to a frame."""
+        obj_idx = self._obj_id_to_idx(inference_state, obj_id)
+        point_inputs_per_frame = inference_state["point_inputs_per_obj"][obj_idx]
+        mask_inputs_per_frame = inference_state["mask_inputs_per_obj"][obj_idx]
+
+        if not isinstance(mask, torch.Tensor):
+            mask = torch.tensor(mask, dtype=torch.bool)
+        assert mask.dim() == 2
+        mask_H, mask_W = mask.shape
+        mask_inputs_orig = mask[None, None]  # add batch and channel dimension
+        mask_inputs_orig = mask_inputs_orig.float().to(inference_state["device"])
+
+        # resize the mask if it doesn't match the model's image size
+        if mask_H != self.image_size or mask_W != self.image_size:
+            mask_inputs = torch.nn.functional.interpolate(
+                mask_inputs_orig,
+                size=(self.image_size, self.image_size),
+                align_corners=False,
+                mode="bilinear",
+                antialias=True,  # use antialias for downsampling
+            )
+            mask_inputs = (mask_inputs >= 0.5).float()
+        else:
+            mask_inputs = mask_inputs_orig
+
+        mask_inputs_per_frame[frame_idx] = mask_inputs
+        point_inputs_per_frame.pop(frame_idx, None)
+        # If this frame hasn't been tracked before, we treat it as an initial conditioning
+        # frame, meaning that the inputs points are to generate segments on this frame without
+        # using any memory from other frames, like in SAM. Otherwise (if it has been tracked),
+        # the input points will be used to correct the already tracked masks.
+        is_init_cond_frame = frame_idx not in inference_state["frames_already_tracked"]
+        # whether to track in reverse time order
+        if is_init_cond_frame:
+            reverse = False
+        else:
+            reverse = inference_state["frames_already_tracked"][frame_idx]["reverse"]
+        obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+        obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+        # Add a frame to conditioning output if it's an initial conditioning frame or
+        # if the model sees all frames receiving clicks/mask as conditioning frames.
+        is_cond = is_init_cond_frame or self.add_all_frames_to_correct_as_cond
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+
+        current_out, _ = self._run_single_frame_inference(
+            inference_state=inference_state,
+            output_dict=obj_output_dict,  # run on the slice of a single object
+            frame_idx=frame_idx,
+            batch_size=1,  # run on the slice of a single object
+            is_init_cond_frame=is_init_cond_frame,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            reverse=reverse,
+            # Skip the memory encoder when adding clicks or mask. We execute the memory encoder
+            # at the beginning of `propagate_in_video` (after user finalize their clicks). This
+            # allows us to enforce non-overlapping constraints on all objects before encoding
+            # them into memory.
+            run_mem_encoder=False,
+        )
+        # Add the output to the output dict (to be used as future memory)
+        obj_temp_output_dict[storage_key][frame_idx] = current_out
+
+        # Resize the output mask to the original video resolution
+        obj_ids = inference_state["obj_ids"]
+        consolidated_out = self._consolidate_temp_output_across_obj(
+            inference_state,
+            frame_idx,
+            is_cond=is_cond,
+            run_mem_encoder=False,
+            consolidate_at_video_res=True,
+        )
+        _, video_res_masks = self._get_orig_video_res_output(
+            inference_state, consolidated_out["pred_masks_video_res"]
+        )
+        return frame_idx, obj_ids, video_res_masks
+
+    def _get_orig_video_res_output(self, inference_state, any_res_masks):
+        """
+        Resize the object scores to the original video resolution (video_res_masks)
+        and apply non-overlapping constraints for final output.
+        """
+        device = inference_state["device"]
+        video_H = inference_state["video_height"]
+        video_W = inference_state["video_width"]
+        any_res_masks = any_res_masks.to(device, non_blocking=True)
+        if any_res_masks.shape[-2:] == (video_H, video_W):
+            video_res_masks = any_res_masks
+        else:
+            video_res_masks = torch.nn.functional.interpolate(
+                any_res_masks,
+                size=(video_H, video_W),
+                mode="bilinear",
+                align_corners=False,
+            )
+        if self.non_overlap_masks:
+            video_res_masks = self._apply_non_overlapping_constraints(video_res_masks)
+        return any_res_masks, video_res_masks
+
+    def _consolidate_temp_output_across_obj(
+        self,
+        inference_state,
+        frame_idx,
+        is_cond,
+        run_mem_encoder,
+        consolidate_at_video_res=False,
+    ):
+        """
+        Consolidate the per-object temporary outputs in `temp_output_dict_per_obj` on
+        a frame into a single output for all objects, including
+        1) fill any missing objects either from `output_dict_per_obj` (if they exist in
+           `output_dict_per_obj` for this frame) or leave them as placeholder values
+           (if they don't exist in `output_dict_per_obj` for this frame);
+        2) if specified, rerun memory encoder after apply non-overlapping constraints
+           on the object scores.
+        """
+        batch_size = self._get_obj_num(inference_state)
+        storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+        # Optionally, we allow consolidating the temporary outputs at the original
+        # video resolution (to provide a better editing experience for mask prompts).
+        if consolidate_at_video_res:
+            assert not run_mem_encoder, "memory encoder cannot run at video resolution"
+            consolidated_H = inference_state["video_height"]
+            consolidated_W = inference_state["video_width"]
+            consolidated_mask_key = "pred_masks_video_res"
+        else:
+            consolidated_H = consolidated_W = self.image_size // 4
+            consolidated_mask_key = "pred_masks"
+
+        # Initialize `consolidated_out`. Its "maskmem_features" and "maskmem_pos_enc"
+        # will be added when rerunning the memory encoder after applying non-overlapping
+        # constraints to object scores. Its "pred_masks" are prefilled with a large
+        # negative value (NO_OBJ_SCORE) to represent missing objects.
+        consolidated_out = {
+            "maskmem_features": None,
+            "maskmem_pos_enc": None,
+            consolidated_mask_key: torch.full(
+                size=(batch_size, 1, consolidated_H, consolidated_W),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["storage_device"],
+            ),
+            "obj_ptr": torch.full(
+                size=(batch_size, self.hidden_dim),
+                fill_value=NO_OBJ_SCORE,
+                dtype=torch.float32,
+                device=inference_state["device"],
+            ),
+        }
+        empty_mask_ptr = None
+        for obj_idx in range(batch_size):
+            obj_temp_output_dict = inference_state["temp_output_dict_per_obj"][obj_idx]
+            obj_output_dict = inference_state["output_dict_per_obj"][obj_idx]
+            out = obj_temp_output_dict[storage_key].get(frame_idx, None)
+            # If the object doesn't appear in "temp_output_dict_per_obj" on this frame,
+            # we fall back and look up its previous output in "output_dict_per_obj".
+            # We look up both "cond_frame_outputs" and "non_cond_frame_outputs" in
+            # "output_dict_per_obj" to find a previous output for this object.
+            if out is None:
+                out = obj_output_dict["cond_frame_outputs"].get(frame_idx, None)
+            if out is None:
+                out = obj_output_dict["non_cond_frame_outputs"].get(frame_idx, None)
+            # If the object doesn't appear in "output_dict_per_obj" either, we skip it
+            # and leave its mask scores to the default scores (i.e. the NO_OBJ_SCORE
+            # placeholder above) and set its object pointer to be a dummy pointer.
+            if out is None:
+                # Fill in dummy object pointers for those objects without any inputs or
+                # tracking outcomes on this frame (only do it under `run_mem_encoder=True`,
+                # i.e. when we need to build the memory for tracking).
+                if run_mem_encoder:
+                    if empty_mask_ptr is None:
+                        empty_mask_ptr = self._get_empty_mask_ptr(
+                            inference_state, frame_idx
+                        )
+                    # fill object pointer with a dummy pointer (based on an empty mask)
+                    consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = empty_mask_ptr
+                continue
+            # Add the temporary object output mask to consolidated output mask
+            obj_mask = out["pred_masks"]
+            consolidated_pred_masks = consolidated_out[consolidated_mask_key]
+            if obj_mask.shape[-2:] == consolidated_pred_masks.shape[-2:]:
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = obj_mask
+            else:
+                # Resize first if temporary object mask has a different resolution
+                resized_obj_mask = torch.nn.functional.interpolate(
+                    obj_mask,
+                    size=consolidated_pred_masks.shape[-2:],
+                    mode="bilinear",
+                    align_corners=False,
+                )
+                consolidated_pred_masks[obj_idx : obj_idx + 1] = resized_obj_mask
+            consolidated_out["obj_ptr"][obj_idx : obj_idx + 1] = out["obj_ptr"]
+
+        # Optionally, apply non-overlapping constraints on the consolidated scores
+        # and rerun the memory encoder
+        if run_mem_encoder:
+            device = inference_state["device"]
+            high_res_masks = torch.nn.functional.interpolate(
+                consolidated_out["pred_masks"].to(device, non_blocking=True),
+                size=(self.image_size, self.image_size),
+                mode="bilinear",
+                align_corners=False,
+            )
+            if self.non_overlap_masks_for_mem_enc:
+                high_res_masks = self._apply_non_overlapping_constraints(high_res_masks)
+            maskmem_features, maskmem_pos_enc = self._run_memory_encoder(
+                inference_state=inference_state,
+                frame_idx=frame_idx,
+                batch_size=batch_size,
+                high_res_masks=high_res_masks,
+                is_mask_from_pts=True,  # these frames are what the user interacted with
+            )
+            consolidated_out["maskmem_features"] = maskmem_features
+            consolidated_out["maskmem_pos_enc"] = maskmem_pos_enc
+
+        return consolidated_out
+
+    def _get_empty_mask_ptr(self, inference_state, frame_idx):
+        """Get a dummy object pointer based on an empty mask on the current frame."""
+        # A dummy (empty) mask with a single object
+        batch_size = 1
+        mask_inputs = torch.zeros(
+            (batch_size, 1, self.image_size, self.image_size),
+            dtype=torch.float32,
+            device=inference_state["device"],
+        )
+
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # Feed the empty mask and image feature above to get a dummy object pointer
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=True,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=None,
+            mask_inputs=mask_inputs,
+            output_dict={},
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=False,
+            run_mem_encoder=False,
+            prev_sam_mask_logits=None,
+        )
+        return current_out["obj_ptr"]
+
+    @torch.inference_mode()
+    def propagate_in_video_preflight(self, inference_state):
+        """Prepare inference_state and consolidate temporary outputs before tracking."""
+        # Tracking has started and we don't allow adding new objects until session is reset.
+        inference_state["tracking_has_started"] = True
+        batch_size = self._get_obj_num(inference_state)
+
+        # Consolidate per-object temporary outputs in "temp_output_dict_per_obj" and
+        # add them into "output_dict".
+        temp_output_dict_per_obj = inference_state["temp_output_dict_per_obj"]
+        output_dict = inference_state["output_dict"]
+        # "consolidated_frame_inds" contains indices of those frames where consolidated
+        # temporary outputs have been added (either in this call or any previous calls
+        # to `propagate_in_video_preflight`).
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        for is_cond in [False, True]:
+            # Separately consolidate conditioning and non-conditioning temp outptus
+            storage_key = "cond_frame_outputs" if is_cond else "non_cond_frame_outputs"
+            # Find all the frames that contain temporary outputs for any objects
+            # (these should be the frames that have just received clicks for mask inputs
+            # via `add_new_points` or `add_new_mask`)
+            temp_frame_inds = set()
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                temp_frame_inds.update(obj_temp_output_dict[storage_key].keys())
+            consolidated_frame_inds[storage_key].update(temp_frame_inds)
+            # consolidate the temprary output across all objects on this frame
+            for frame_idx in temp_frame_inds:
+                consolidated_out = self._consolidate_temp_output_across_obj(
+                    inference_state, frame_idx, is_cond=is_cond, run_mem_encoder=True
+                )
+                # merge them into "output_dict" and also create per-object slices
+                output_dict[storage_key][frame_idx] = consolidated_out
+                self._add_output_per_object(
+                    inference_state, frame_idx, consolidated_out, storage_key
+                )
+                clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+                    self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+                )
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+
+            # clear temporary outputs in `temp_output_dict_per_obj`
+            for obj_temp_output_dict in temp_output_dict_per_obj.values():
+                obj_temp_output_dict[storage_key].clear()
+
+        # edge case: if an output is added to "cond_frame_outputs", we remove any prior
+        # output on the same frame in "non_cond_frame_outputs"
+        for frame_idx in output_dict["cond_frame_outputs"]:
+            output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for obj_output_dict in inference_state["output_dict_per_obj"].values():
+            for frame_idx in obj_output_dict["cond_frame_outputs"]:
+                obj_output_dict["non_cond_frame_outputs"].pop(frame_idx, None)
+        for frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+            assert frame_idx in output_dict["cond_frame_outputs"]
+            consolidated_frame_inds["non_cond_frame_outputs"].discard(frame_idx)
+
+        # Make sure that the frame indices in "consolidated_frame_inds" are exactly those frames
+        # with either points or mask inputs (which should be true under a correct workflow).
+        all_consolidated_frame_inds = (
+            consolidated_frame_inds["cond_frame_outputs"]
+            | consolidated_frame_inds["non_cond_frame_outputs"]
+        )
+        input_frames_inds = set()
+        for point_inputs_per_frame in inference_state["point_inputs_per_obj"].values():
+            input_frames_inds.update(point_inputs_per_frame.keys())
+        for mask_inputs_per_frame in inference_state["mask_inputs_per_obj"].values():
+            input_frames_inds.update(mask_inputs_per_frame.keys())
+        assert all_consolidated_frame_inds == input_frames_inds
+
+    @torch.inference_mode()
+    def propagate_in_video(
+        self,
+        inference_state,
+        start_frame_idx=None,
+        max_frame_num_to_track=None,
+        reverse=False,
+    ):
+        """Propagate the input points across frames to track in the entire video."""
+        self.propagate_in_video_preflight(inference_state)
+
+        output_dict = inference_state["output_dict"]
+        consolidated_frame_inds = inference_state["consolidated_frame_inds"]
+        obj_ids = inference_state["obj_ids"]
+        num_frames = inference_state["num_frames"]
+        batch_size = self._get_obj_num(inference_state)
+        if len(output_dict["cond_frame_outputs"]) == 0:
+            raise RuntimeError("No points are provided; please add points first")
+        clear_non_cond_mem = self.clear_non_cond_mem_around_input and (
+            self.clear_non_cond_mem_for_multi_obj or batch_size <= 1
+        )
+
+        # set start index, end index, and processing order
+        if start_frame_idx is None:
+            # default: start from the earliest frame with input points
+            start_frame_idx = min(output_dict["cond_frame_outputs"])
+        if max_frame_num_to_track is None:
+            # default: track all the frames in the video
+            max_frame_num_to_track = num_frames
+        if reverse:
+            end_frame_idx = max(start_frame_idx - max_frame_num_to_track, 0)
+            if start_frame_idx > 0:
+                processing_order = range(start_frame_idx, end_frame_idx - 1, -1)
+            else:
+                processing_order = []  # skip reverse tracking if starting from frame 0
+        else:
+            end_frame_idx = min(
+                start_frame_idx + max_frame_num_to_track, num_frames - 1
+            )
+            processing_order = range(start_frame_idx, end_frame_idx + 1)
+
+        for frame_idx in tqdm(processing_order, desc="propagate in video"):
+            # We skip those frames already in consolidated outputs (these are frames
+            # that received input clicks or mask). Note that we cannot directly run
+            # batched forward on them via `_run_single_frame_inference` because the
+            # number of clicks on each object might be different.
+            if frame_idx in consolidated_frame_inds["cond_frame_outputs"]:
+                storage_key = "cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+                if clear_non_cond_mem:
+                    # clear non-conditioning memory of the surrounding frames
+                    self._clear_non_cond_mem_around_input(inference_state, frame_idx)
+            elif frame_idx in consolidated_frame_inds["non_cond_frame_outputs"]:
+                storage_key = "non_cond_frame_outputs"
+                current_out = output_dict[storage_key][frame_idx]
+                pred_masks = current_out["pred_masks"]
+            else:
+                storage_key = "non_cond_frame_outputs"
+                current_out, pred_masks = self._run_single_frame_inference(
+                    inference_state=inference_state,
+                    output_dict=output_dict,
+                    frame_idx=frame_idx,
+                    batch_size=batch_size,
+                    is_init_cond_frame=False,
+                    point_inputs=None,
+                    mask_inputs=None,
+                    reverse=reverse,
+                    run_mem_encoder=True,
+                )
+                output_dict[storage_key][frame_idx] = current_out
+            # Create slices of per-object outputs for subsequent interaction with each
+            # individual object after tracking.
+            self._add_output_per_object(
+                inference_state, frame_idx, current_out, storage_key
+            )
+            inference_state["frames_already_tracked"][frame_idx] = {"reverse": reverse}
+
+            # Resize the output mask to the original video resolution (we directly use
+            # the mask scores on GPU for output to avoid any CPU conversion in between)
+            _, video_res_masks = self._get_orig_video_res_output(
+                inference_state, pred_masks
+            )
+            yield frame_idx, obj_ids, video_res_masks
+
+    def _add_output_per_object(
+        self, inference_state, frame_idx, current_out, storage_key
+    ):
+        """
+        Split a multi-object output into per-object output slices and add them into
+        `output_dict_per_obj`. The resulting slices share the same tensor storage.
+        """
+        maskmem_features = current_out["maskmem_features"]
+        assert maskmem_features is None or isinstance(maskmem_features, torch.Tensor)
+
+        maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        assert maskmem_pos_enc is None or isinstance(maskmem_pos_enc, list)
+
+        output_dict_per_obj = inference_state["output_dict_per_obj"]
+        for obj_idx, obj_output_dict in output_dict_per_obj.items():
+            obj_slice = slice(obj_idx, obj_idx + 1)
+            obj_out = {
+                "maskmem_features": None,
+                "maskmem_pos_enc": None,
+                "pred_masks": current_out["pred_masks"][obj_slice],
+                "obj_ptr": current_out["obj_ptr"][obj_slice],
+            }
+            if maskmem_features is not None:
+                obj_out["maskmem_features"] = maskmem_features[obj_slice]
+            if maskmem_pos_enc is not None:
+                obj_out["maskmem_pos_enc"] = [x[obj_slice] for x in maskmem_pos_enc]
+            obj_output_dict[storage_key][frame_idx] = obj_out
+
+    @torch.inference_mode()
+    def reset_state(self, inference_state):
+        """Remove all input points or mask in all frames throughout the video."""
+        self._reset_tracking_results(inference_state)
+        # Remove all object ids
+        inference_state["obj_id_to_idx"].clear()
+        inference_state["obj_idx_to_id"].clear()
+        inference_state["obj_ids"].clear()
+        inference_state["point_inputs_per_obj"].clear()
+        inference_state["mask_inputs_per_obj"].clear()
+        inference_state["output_dict_per_obj"].clear()
+        inference_state["temp_output_dict_per_obj"].clear()
+
+    def _reset_tracking_results(self, inference_state):
+        """Reset all tracking inputs and results across the videos."""
+        for v in inference_state["point_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["mask_inputs_per_obj"].values():
+            v.clear()
+        for v in inference_state["output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        for v in inference_state["temp_output_dict_per_obj"].values():
+            v["cond_frame_outputs"].clear()
+            v["non_cond_frame_outputs"].clear()
+        inference_state["output_dict"]["cond_frame_outputs"].clear()
+        inference_state["output_dict"]["non_cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["cond_frame_outputs"].clear()
+        inference_state["consolidated_frame_inds"]["non_cond_frame_outputs"].clear()
+        inference_state["tracking_has_started"] = False
+        inference_state["frames_already_tracked"].clear()
+
+    def _get_image_feature(self, inference_state, frame_idx, batch_size):
+        """Compute the image features on a given frame."""
+        # Look up in the cache first
+        image, backbone_out = inference_state["cached_features"].get(
+            frame_idx, (None, None)
+        )
+        if backbone_out is None:
+            # Cache miss -- we will run inference on a single image
+            image = inference_state["images"][frame_idx].cuda().float().unsqueeze(0)
+            backbone_out = self.forward_image(image)
+            # Cache the most recent frame's feature (for repeated interactions with
+            # a frame; we can use an LRU cache for more frames in the future).
+            inference_state["cached_features"] = {frame_idx: (image, backbone_out)}
+
+        # expand the features to have the same dimension as the number of objects
+        expanded_image = image.expand(batch_size, -1, -1, -1)
+        expanded_backbone_out = {
+            "backbone_fpn": backbone_out["backbone_fpn"].copy(),
+            "vision_pos_enc": backbone_out["vision_pos_enc"].copy(),
+        }
+        for i, feat in enumerate(expanded_backbone_out["backbone_fpn"]):
+            expanded_backbone_out["backbone_fpn"][i] = feat.expand(
+                batch_size, -1, -1, -1
+            )
+        for i, pos in enumerate(expanded_backbone_out["vision_pos_enc"]):
+            pos = pos.expand(batch_size, -1, -1, -1)
+            expanded_backbone_out["vision_pos_enc"][i] = pos
+
+        features = self._prepare_backbone_features(expanded_backbone_out)
+        features = (expanded_image,) + features
+        return features
+
+    def _run_single_frame_inference(
+        self,
+        inference_state,
+        output_dict,
+        frame_idx,
+        batch_size,
+        is_init_cond_frame,
+        point_inputs,
+        mask_inputs,
+        reverse,
+        run_mem_encoder,
+        prev_sam_mask_logits=None,
+    ):
+        """Run tracking on a single frame based on current inputs and previous memory."""
+        # Retrieve correct image features
+        (
+            _,
+            _,
+            current_vision_feats,
+            current_vision_pos_embeds,
+            feat_sizes,
+        ) = self._get_image_feature(inference_state, frame_idx, batch_size)
+
+        # point and mask should not appear as input simultaneously on the same frame
+        assert point_inputs is None or mask_inputs is None
+        current_out = self.track_step(
+            frame_idx=frame_idx,
+            is_init_cond_frame=is_init_cond_frame,
+            current_vision_feats=current_vision_feats,
+            current_vision_pos_embeds=current_vision_pos_embeds,
+            feat_sizes=feat_sizes,
+            point_inputs=point_inputs,
+            mask_inputs=mask_inputs,
+            output_dict=output_dict,
+            num_frames=inference_state["num_frames"],
+            track_in_reverse=reverse,
+            run_mem_encoder=run_mem_encoder,
+            prev_sam_mask_logits=prev_sam_mask_logits,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = current_out["maskmem_features"]
+        if maskmem_features is not None:
+            maskmem_features = maskmem_features.to(torch.bfloat16)
+            maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        pred_masks_gpu = current_out["pred_masks"]
+        # potentially fill holes in the predicted masks
+        if self.fill_hole_area > 0:
+            pred_masks_gpu = fill_holes_in_mask_scores(
+                pred_masks_gpu, self.fill_hole_area
+            )
+        pred_masks = pred_masks_gpu.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(inference_state, current_out)
+        # object pointer is a small tensor, so we always keep it on GPU memory for fast access
+        obj_ptr = current_out["obj_ptr"]
+        # make a compact version of this frame's output to reduce the state size
+        compact_current_out = {
+            "maskmem_features": maskmem_features,
+            "maskmem_pos_enc": maskmem_pos_enc,
+            "pred_masks": pred_masks,
+            "obj_ptr": obj_ptr,
+        }
+        return compact_current_out, pred_masks_gpu
+
+    def _run_memory_encoder(
+        self, inference_state, frame_idx, batch_size, high_res_masks, is_mask_from_pts
+    ):
+        """
+        Run the memory encoder on `high_res_masks`. This is usually after applying
+        non-overlapping constraints to object scores. Since their scores changed, their
+        memory also need to be computed again with the memory encoder.
+        """
+        # Retrieve correct image features
+        _, _, current_vision_feats, _, feat_sizes = self._get_image_feature(
+            inference_state, frame_idx, batch_size
+        )
+        maskmem_features, maskmem_pos_enc = self._encode_new_memory(
+            current_vision_feats=current_vision_feats,
+            feat_sizes=feat_sizes,
+            pred_masks_high_res=high_res_masks,
+            is_mask_from_pts=is_mask_from_pts,
+        )
+
+        # optionally offload the output to CPU memory to save GPU space
+        storage_device = inference_state["storage_device"]
+        maskmem_features = maskmem_features.to(torch.bfloat16)
+        maskmem_features = maskmem_features.to(storage_device, non_blocking=True)
+        # "maskmem_pos_enc" is the same across frames, so we only need to store one copy of it
+        maskmem_pos_enc = self._get_maskmem_pos_enc(
+            inference_state, {"maskmem_pos_enc": maskmem_pos_enc}
+        )
+        return maskmem_features, maskmem_pos_enc
+
+    def _get_maskmem_pos_enc(self, inference_state, current_out):
+        """
+        `maskmem_pos_enc` is the same across frames and objects, so we cache it as
+        a constant in the inference session to reduce session storage size.
+        """
+        model_constants = inference_state["constants"]
+        # "out_maskmem_pos_enc" should be either a list of tensors or None
+        out_maskmem_pos_enc = current_out["maskmem_pos_enc"]
+        if out_maskmem_pos_enc is not None:
+            if "maskmem_pos_enc" not in model_constants:
+                assert isinstance(out_maskmem_pos_enc, list)
+                # only take the slice for one object, since it's same across objects
+                maskmem_pos_enc = [x[0:1].clone() for x in out_maskmem_pos_enc]
+                model_constants["maskmem_pos_enc"] = maskmem_pos_enc
+            else:
+                maskmem_pos_enc = model_constants["maskmem_pos_enc"]
+            # expand the cached maskmem_pos_enc to the actual batch size
+            batch_size = out_maskmem_pos_enc[0].size(0)
+            expanded_maskmem_pos_enc = [
+                x.expand(batch_size, -1, -1, -1) for x in maskmem_pos_enc
+            ]
+        else:
+            expanded_maskmem_pos_enc = None
+        return expanded_maskmem_pos_enc
+
+    def _clear_non_cond_mem_around_input(self, inference_state, frame_idx):
+        """
+        Remove the non-conditioning memory around the input frame. When users provide
+        correction clicks, the surrounding frames' non-conditioning memories can still
+        contain outdated object appearance information and could confuse the model.
+
+        This method clears those non-conditioning memories surrounding the interacted
+        frame to avoid giving the model both old and new information about the object.
+        """
+        r = self.memory_temporal_stride_for_eval
+        frame_idx_begin = frame_idx - r * self.num_maskmem
+        frame_idx_end = frame_idx + r * self.num_maskmem
+        output_dict = inference_state["output_dict"]
+        non_cond_frame_outputs = output_dict["non_cond_frame_outputs"]
+        for t in range(frame_idx_begin, frame_idx_end + 1):
+            non_cond_frame_outputs.pop(t, None)
+            for obj_output_dict in inference_state["output_dict_per_obj"].values():
+                obj_output_dict["non_cond_frame_outputs"].pop(t, None)
diff --git a/third_parts/sam2/utils/__init__.py b/third_parts/sam2/utils/__init__.py
new file mode 100644
index 0000000000000000000000000000000000000000..5277f46157403e47fd830fc519144b97ef69d4ae
--- /dev/null
+++ b/third_parts/sam2/utils/__init__.py
@@ -0,0 +1,5 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
diff --git a/third_parts/sam2/utils/amg.py b/third_parts/sam2/utils/amg.py
new file mode 100644
index 0000000000000000000000000000000000000000..986842960cf5deca00614b7b1cde1ab77dad7e6e
--- /dev/null
+++ b/third_parts/sam2/utils/amg.py
@@ -0,0 +1,348 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import math
+from copy import deepcopy
+from itertools import product
+from typing import Any, Dict, Generator, ItemsView, List, Tuple
+
+import numpy as np
+import torch
+
+# Very lightly adapted from https://github.com/facebookresearch/segment-anything/blob/main/segment_anything/utils/amg.py
+
+
+class MaskData:
+    """
+    A structure for storing masks and their related data in batched format.
+    Implements basic filtering and concatenation.
+    """
+
+    def __init__(self, **kwargs) -> None:
+        for v in kwargs.values():
+            assert isinstance(
+                v, (list, np.ndarray, torch.Tensor)
+            ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats = dict(**kwargs)
+
+    def __setitem__(self, key: str, item: Any) -> None:
+        assert isinstance(
+            item, (list, np.ndarray, torch.Tensor)
+        ), "MaskData only supports list, numpy arrays, and torch tensors."
+        self._stats[key] = item
+
+    def __delitem__(self, key: str) -> None:
+        del self._stats[key]
+
+    def __getitem__(self, key: str) -> Any:
+        return self._stats[key]
+
+    def items(self) -> ItemsView[str, Any]:
+        return self._stats.items()
+
+    def filter(self, keep: torch.Tensor) -> None:
+        for k, v in self._stats.items():
+            if v is None:
+                self._stats[k] = None
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = v[torch.as_tensor(keep, device=v.device)]
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = v[keep.detach().cpu().numpy()]
+            elif isinstance(v, list) and keep.dtype == torch.bool:
+                self._stats[k] = [a for i, a in enumerate(v) if keep[i]]
+            elif isinstance(v, list):
+                self._stats[k] = [v[i] for i in keep]
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def cat(self, new_stats: "MaskData") -> None:
+        for k, v in new_stats.items():
+            if k not in self._stats or self._stats[k] is None:
+                self._stats[k] = deepcopy(v)
+            elif isinstance(v, torch.Tensor):
+                self._stats[k] = torch.cat([self._stats[k], v], dim=0)
+            elif isinstance(v, np.ndarray):
+                self._stats[k] = np.concatenate([self._stats[k], v], axis=0)
+            elif isinstance(v, list):
+                self._stats[k] = self._stats[k] + deepcopy(v)
+            else:
+                raise TypeError(f"MaskData key {k} has an unsupported type {type(v)}.")
+
+    def to_numpy(self) -> None:
+        for k, v in self._stats.items():
+            if isinstance(v, torch.Tensor):
+                self._stats[k] = v.float().detach().cpu().numpy()
+
+
+def is_box_near_crop_edge(
+    boxes: torch.Tensor, crop_box: List[int], orig_box: List[int], atol: float = 20.0
+) -> torch.Tensor:
+    """Filter masks at the edge of a crop, but not at the edge of the original image."""
+    crop_box_torch = torch.as_tensor(crop_box, dtype=torch.float, device=boxes.device)
+    orig_box_torch = torch.as_tensor(orig_box, dtype=torch.float, device=boxes.device)
+    boxes = uncrop_boxes_xyxy(boxes, crop_box).float()
+    near_crop_edge = torch.isclose(boxes, crop_box_torch[None, :], atol=atol, rtol=0)
+    near_image_edge = torch.isclose(boxes, orig_box_torch[None, :], atol=atol, rtol=0)
+    near_crop_edge = torch.logical_and(near_crop_edge, ~near_image_edge)
+    return torch.any(near_crop_edge, dim=1)
+
+
+def box_xyxy_to_xywh(box_xyxy: torch.Tensor) -> torch.Tensor:
+    box_xywh = deepcopy(box_xyxy)
+    box_xywh[2] = box_xywh[2] - box_xywh[0]
+    box_xywh[3] = box_xywh[3] - box_xywh[1]
+    return box_xywh
+
+
+def batch_iterator(batch_size: int, *args) -> Generator[List[Any], None, None]:
+    assert len(args) > 0 and all(
+        len(a) == len(args[0]) for a in args
+    ), "Batched iteration must have inputs of all the same size."
+    n_batches = len(args[0]) // batch_size + int(len(args[0]) % batch_size != 0)
+    for b in range(n_batches):
+        yield [arg[b * batch_size : (b + 1) * batch_size] for arg in args]
+
+
+def mask_to_rle_pytorch(tensor: torch.Tensor) -> List[Dict[str, Any]]:
+    """
+    Encodes masks to an uncompressed RLE, in the format expected by
+    pycoco tools.
+    """
+    # Put in fortran order and flatten h,w
+    b, h, w = tensor.shape
+    tensor = tensor.permute(0, 2, 1).flatten(1)
+
+    # Compute change indices
+    diff = tensor[:, 1:] ^ tensor[:, :-1]
+    change_indices = diff.nonzero()
+
+    # Encode run length
+    out = []
+    for i in range(b):
+        cur_idxs = change_indices[change_indices[:, 0] == i, 1]
+        cur_idxs = torch.cat(
+            [
+                torch.tensor([0], dtype=cur_idxs.dtype, device=cur_idxs.device),
+                cur_idxs + 1,
+                torch.tensor([h * w], dtype=cur_idxs.dtype, device=cur_idxs.device),
+            ]
+        )
+        btw_idxs = cur_idxs[1:] - cur_idxs[:-1]
+        counts = [] if tensor[i, 0] == 0 else [0]
+        counts.extend(btw_idxs.detach().cpu().tolist())
+        out.append({"size": [h, w], "counts": counts})
+    return out
+
+
+def rle_to_mask(rle: Dict[str, Any]) -> np.ndarray:
+    """Compute a binary mask from an uncompressed RLE."""
+    h, w = rle["size"]
+    mask = np.empty(h * w, dtype=bool)
+    idx = 0
+    parity = False
+    for count in rle["counts"]:
+        mask[idx : idx + count] = parity
+        idx += count
+        parity ^= True
+    mask = mask.reshape(w, h)
+    return mask.transpose()  # Put in C order
+
+
+def area_from_rle(rle: Dict[str, Any]) -> int:
+    return sum(rle["counts"][1::2])
+
+
+def calculate_stability_score(
+    masks: torch.Tensor, mask_threshold: float, threshold_offset: float
+) -> torch.Tensor:
+    """
+    Computes the stability score for a batch of masks. The stability
+    score is the IoU between the binary masks obtained by thresholding
+    the predicted mask logits at high and low values.
+    """
+    # One mask is always contained inside the other.
+    # Save memory by preventing unnecessary cast to torch.int64
+    intersections = (
+        (masks > (mask_threshold + threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    unions = (
+        (masks > (mask_threshold - threshold_offset))
+        .sum(-1, dtype=torch.int16)
+        .sum(-1, dtype=torch.int32)
+    )
+    return intersections / unions
+
+
+def build_point_grid(n_per_side: int) -> np.ndarray:
+    """Generates a 2D grid of points evenly spaced in [0,1]x[0,1]."""
+    offset = 1 / (2 * n_per_side)
+    points_one_side = np.linspace(offset, 1 - offset, n_per_side)
+    points_x = np.tile(points_one_side[None, :], (n_per_side, 1))
+    points_y = np.tile(points_one_side[:, None], (1, n_per_side))
+    points = np.stack([points_x, points_y], axis=-1).reshape(-1, 2)
+    return points
+
+
+def build_all_layer_point_grids(
+    n_per_side: int, n_layers: int, scale_per_layer: int
+) -> List[np.ndarray]:
+    """Generates point grids for all crop layers."""
+    points_by_layer = []
+    for i in range(n_layers + 1):
+        n_points = int(n_per_side / (scale_per_layer**i))
+        points_by_layer.append(build_point_grid(n_points))
+    return points_by_layer
+
+
+def generate_crop_boxes(
+    im_size: Tuple[int, ...], n_layers: int, overlap_ratio: float
+) -> Tuple[List[List[int]], List[int]]:
+    """
+    Generates a list of crop boxes of different sizes. Each layer
+    has (2**i)**2 boxes for the ith layer.
+    """
+    crop_boxes, layer_idxs = [], []
+    im_h, im_w = im_size
+    short_side = min(im_h, im_w)
+
+    # Original image
+    crop_boxes.append([0, 0, im_w, im_h])
+    layer_idxs.append(0)
+
+    def crop_len(orig_len, n_crops, overlap):
+        return int(math.ceil((overlap * (n_crops - 1) + orig_len) / n_crops))
+
+    for i_layer in range(n_layers):
+        n_crops_per_side = 2 ** (i_layer + 1)
+        overlap = int(overlap_ratio * short_side * (2 / n_crops_per_side))
+
+        crop_w = crop_len(im_w, n_crops_per_side, overlap)
+        crop_h = crop_len(im_h, n_crops_per_side, overlap)
+
+        crop_box_x0 = [int((crop_w - overlap) * i) for i in range(n_crops_per_side)]
+        crop_box_y0 = [int((crop_h - overlap) * i) for i in range(n_crops_per_side)]
+
+        # Crops in XYWH format
+        for x0, y0 in product(crop_box_x0, crop_box_y0):
+            box = [x0, y0, min(x0 + crop_w, im_w), min(y0 + crop_h, im_h)]
+            crop_boxes.append(box)
+            layer_idxs.append(i_layer + 1)
+
+    return crop_boxes, layer_idxs
+
+
+def uncrop_boxes_xyxy(boxes: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0, x0, y0]], device=boxes.device)
+    # Check if boxes has a channel dimension
+    if len(boxes.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return boxes + offset
+
+
+def uncrop_points(points: torch.Tensor, crop_box: List[int]) -> torch.Tensor:
+    x0, y0, _, _ = crop_box
+    offset = torch.tensor([[x0, y0]], device=points.device)
+    # Check if points has a channel dimension
+    if len(points.shape) == 3:
+        offset = offset.unsqueeze(1)
+    return points + offset
+
+
+def uncrop_masks(
+    masks: torch.Tensor, crop_box: List[int], orig_h: int, orig_w: int
+) -> torch.Tensor:
+    x0, y0, x1, y1 = crop_box
+    if x0 == 0 and y0 == 0 and x1 == orig_w and y1 == orig_h:
+        return masks
+    # Coordinate transform masks
+    pad_x, pad_y = orig_w - (x1 - x0), orig_h - (y1 - y0)
+    pad = (x0, pad_x - x0, y0, pad_y - y0)
+    return torch.nn.functional.pad(masks, pad, value=0)
+
+
+def remove_small_regions(
+    mask: np.ndarray, area_thresh: float, mode: str
+) -> Tuple[np.ndarray, bool]:
+    """
+    Removes small disconnected regions and holes in a mask. Returns the
+    mask and an indicator of if the mask has been modified.
+    """
+    import cv2  # type: ignore
+
+    assert mode in ["holes", "islands"]
+    correct_holes = mode == "holes"
+    working_mask = (correct_holes ^ mask).astype(np.uint8)
+    n_labels, regions, stats, _ = cv2.connectedComponentsWithStats(working_mask, 8)
+    sizes = stats[:, -1][1:]  # Row 0 is background label
+    small_regions = [i + 1 for i, s in enumerate(sizes) if s < area_thresh]
+    if len(small_regions) == 0:
+        return mask, False
+    fill_labels = [0] + small_regions
+    if not correct_holes:
+        fill_labels = [i for i in range(n_labels) if i not in fill_labels]
+        # If every region is below threshold, keep largest
+        if len(fill_labels) == 0:
+            fill_labels = [int(np.argmax(sizes)) + 1]
+    mask = np.isin(regions, fill_labels)
+    return mask, True
+
+
+def coco_encode_rle(uncompressed_rle: Dict[str, Any]) -> Dict[str, Any]:
+    from pycocotools import mask as mask_utils  # type: ignore
+
+    h, w = uncompressed_rle["size"]
+    rle = mask_utils.frPyObjects(uncompressed_rle, h, w)
+    rle["counts"] = rle["counts"].decode("utf-8")  # Necessary to serialize with json
+    return rle
+
+
+def batched_mask_to_box(masks: torch.Tensor) -> torch.Tensor:
+    """
+    Calculates boxes in XYXY format around masks. Return [0,0,0,0] for
+    an empty mask. For input shape C1xC2x...xHxW, the output shape is C1xC2x...x4.
+    """
+    # torch.max below raises an error on empty inputs, just skip in this case
+    if torch.numel(masks) == 0:
+        return torch.zeros(*masks.shape[:-2], 4, device=masks.device)
+
+    # Normalize shape to CxHxW
+    shape = masks.shape
+    h, w = shape[-2:]
+    if len(shape) > 2:
+        masks = masks.flatten(0, -3)
+    else:
+        masks = masks.unsqueeze(0)
+
+    # Get top and bottom edges
+    in_height, _ = torch.max(masks, dim=-1)
+    in_height_coords = in_height * torch.arange(h, device=in_height.device)[None, :]
+    bottom_edges, _ = torch.max(in_height_coords, dim=-1)
+    in_height_coords = in_height_coords + h * (~in_height)
+    top_edges, _ = torch.min(in_height_coords, dim=-1)
+
+    # Get left and right edges
+    in_width, _ = torch.max(masks, dim=-2)
+    in_width_coords = in_width * torch.arange(w, device=in_width.device)[None, :]
+    right_edges, _ = torch.max(in_width_coords, dim=-1)
+    in_width_coords = in_width_coords + w * (~in_width)
+    left_edges, _ = torch.min(in_width_coords, dim=-1)
+
+    # If the mask is empty the right edge will be to the left of the left edge.
+    # Replace these boxes with [0, 0, 0, 0]
+    empty_filter = (right_edges < left_edges) | (bottom_edges < top_edges)
+    out = torch.stack([left_edges, top_edges, right_edges, bottom_edges], dim=-1)
+    out = out * (~empty_filter).unsqueeze(-1)
+
+    # Return to original shape
+    if len(shape) > 2:
+        out = out.reshape(*shape[:-2], 4)
+    else:
+        out = out[0]
+
+    return out
diff --git a/third_parts/sam2/utils/misc.py b/third_parts/sam2/utils/misc.py
new file mode 100644
index 0000000000000000000000000000000000000000..b354b91cb534a34f8c51ac23dc83d70c8c28ed6a
--- /dev/null
+++ b/third_parts/sam2/utils/misc.py
@@ -0,0 +1,249 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import os
+import warnings
+from threading import Thread
+
+import numpy as np
+import torch
+from PIL import Image
+from tqdm import tqdm
+
+
+def get_sdpa_settings():
+    if torch.cuda.is_available():
+        old_gpu = torch.cuda.get_device_properties(0).major < 7
+        # only use Flash Attention on Ampere (8.0) or newer GPUs
+        use_flash_attn = torch.cuda.get_device_properties(0).major >= 8
+        if not use_flash_attn:
+            warnings.warn(
+                "Flash Attention is disabled as it requires a GPU with Ampere (8.0) CUDA capability.",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        # keep math kernel for PyTorch versions before 2.2 (Flash Attention v2 is only
+        # available on PyTorch 2.2+, while Flash Attention v1 cannot handle all cases)
+        pytorch_version = tuple(int(v) for v in torch.__version__.split(".")[:2])
+        if pytorch_version < (2, 2):
+            warnings.warn(
+                f"You are using PyTorch {torch.__version__} without Flash Attention v2 support. "
+                "Consider upgrading to PyTorch 2.2+ for Flash Attention v2 (which could be faster).",
+                category=UserWarning,
+                stacklevel=2,
+            )
+        math_kernel_on = pytorch_version < (2, 2) or not use_flash_attn
+    else:
+        old_gpu = True
+        use_flash_attn = False
+        math_kernel_on = True
+
+    return old_gpu, use_flash_attn, math_kernel_on
+
+
+def get_connected_components(mask):
+    """
+    Get the connected components (8-connectivity) of binary masks of shape (N, 1, H, W).
+
+    Inputs:
+    - mask: A binary mask tensor of shape (N, 1, H, W), where 1 is foreground and 0 is
+            background.
+
+    Outputs:
+    - labels: A tensor of shape (N, 1, H, W) containing the connected component labels
+              for foreground pixels and 0 for background pixels.
+    - counts: A tensor of shape (N, 1, H, W) containing the area of the connected
+              components for foreground pixels and 0 for background pixels.
+    """
+    from torch.utils.cpp_extension import load
+    get_connected_componnets = load(
+        name="get_connected_componnets",
+        sources=["third_parts/sam2/csrc/connected_components.cu"],
+        verbose=True,
+        extra_cuda_cflags=[
+            "-DCUDA_HAS_FP16=1",
+            "-D__CUDA_NO_HALF_OPERATORS__",
+            "-D__CUDA_NO_HALF_CONVERSIONS__",
+            "-D__CUDA_NO_HALF2_OPERATORS__",
+        ]
+    )
+
+    return get_connected_componnets.get_connected_componnets(mask.to(torch.uint8).contiguous())
+
+
+def mask_to_box(masks: torch.Tensor):
+    """
+    compute bounding box given an input mask
+
+    Inputs:
+    - masks: [B, 1, H, W] boxes, dtype=torch.Tensor
+
+    Returns:
+    - box_coords: [B, 1, 4], contains (x, y) coordinates of top left and bottom right box corners, dtype=torch.Tensor
+    """
+    B, _, h, w = masks.shape
+    device = masks.device
+    xs = torch.arange(w, device=device, dtype=torch.int32)
+    ys = torch.arange(h, device=device, dtype=torch.int32)
+    grid_xs, grid_ys = torch.meshgrid(xs, ys, indexing="xy")
+    grid_xs = grid_xs[None, None, ...].expand(B, 1, h, w)
+    grid_ys = grid_ys[None, None, ...].expand(B, 1, h, w)
+    min_xs, _ = torch.min(torch.where(masks, grid_xs, w).flatten(-2), dim=-1)
+    max_xs, _ = torch.max(torch.where(masks, grid_xs, -1).flatten(-2), dim=-1)
+    min_ys, _ = torch.min(torch.where(masks, grid_ys, h).flatten(-2), dim=-1)
+    max_ys, _ = torch.max(torch.where(masks, grid_ys, -1).flatten(-2), dim=-1)
+    bbox_coords = torch.stack((min_xs, min_ys, max_xs, max_ys), dim=-1)
+
+    return bbox_coords
+
+
+def _load_img_as_tensor(img_path, image_size):
+    img_pil = Image.open(img_path)
+    img_np = np.array(img_pil.convert("RGB").resize((image_size, image_size)))
+    if img_np.dtype == np.uint8:  # np.uint8 is expected for JPEG images
+        img_np = img_np / 255.0
+    else:
+        raise RuntimeError(f"Unknown image dtype: {img_np.dtype} on {img_path}")
+    img = torch.from_numpy(img_np).permute(2, 0, 1)
+    video_width, video_height = img_pil.size  # the original video size
+    return img, video_height, video_width
+
+
+class AsyncVideoFrameLoader:
+    """
+    A list of video frames to be load asynchronously without blocking session start.
+    """
+
+    def __init__(self, img_paths, image_size, offload_video_to_cpu, img_mean, img_std):
+        self.img_paths = img_paths
+        self.image_size = image_size
+        self.offload_video_to_cpu = offload_video_to_cpu
+        self.img_mean = img_mean
+        self.img_std = img_std
+        # items in `self._images` will be loaded asynchronously
+        self.images = [None] * len(img_paths)
+        # catch and raise any exceptions in the async loading thread
+        self.exception = None
+        # video_height and video_width be filled when loading the first image
+        self.video_height = None
+        self.video_width = None
+
+        # load the first frame to fill video_height and video_width and also
+        # to cache it (since it's most likely where the user will click)
+        self.__getitem__(0)
+
+        # load the rest of frames asynchronously without blocking the session start
+        def _load_frames():
+            try:
+                for n in tqdm(range(len(self.images)), desc="frame loading (JPEG)"):
+                    self.__getitem__(n)
+            except Exception as e:
+                self.exception = e
+
+        self.thread = Thread(target=_load_frames, daemon=True)
+        self.thread.start()
+
+    def __getitem__(self, index):
+        if self.exception is not None:
+            raise RuntimeError("Failure in frame loading thread") from self.exception
+
+        img = self.images[index]
+        if img is not None:
+            return img
+
+        img, video_height, video_width = _load_img_as_tensor(
+            self.img_paths[index], self.image_size
+        )
+        self.video_height = video_height
+        self.video_width = video_width
+        # normalize by mean and std
+        img -= self.img_mean
+        img /= self.img_std
+        if not self.offload_video_to_cpu:
+            img = img.cuda(non_blocking=True)
+        self.images[index] = img
+        return img
+
+    def __len__(self):
+        return len(self.images)
+
+
+def load_video_frames(
+    video_path,
+    image_size,
+    offload_video_to_cpu,
+    img_mean=(0.485, 0.456, 0.406),
+    img_std=(0.229, 0.224, 0.225),
+    async_loading_frames=False,
+):
+    """
+    Load the video frames from a directory of JPEG files ("<frame_index>.jpg" format).
+
+    The frames are resized to image_size x image_size and are loaded to GPU if
+    `offload_video_to_cpu` is `False` and to CPU if `offload_video_to_cpu` is `True`.
+
+    You can load a frame asynchronously by setting `async_loading_frames` to `True`.
+    """
+    if isinstance(video_path, str) and os.path.isdir(video_path):
+        jpg_folder = video_path
+    else:
+        raise NotImplementedError("Only JPEG frames are supported at this moment")
+
+    frame_names = [
+        p
+        for p in os.listdir(jpg_folder)
+        if os.path.splitext(p)[-1] in [".jpg", ".jpeg", ".JPG", ".JPEG"]
+    ]
+    frame_names.sort(key=lambda p: int(os.path.splitext(p)[0]))
+    num_frames = len(frame_names)
+    if num_frames == 0:
+        raise RuntimeError(f"no images found in {jpg_folder}")
+    img_paths = [os.path.join(jpg_folder, frame_name) for frame_name in frame_names]
+    img_mean = torch.tensor(img_mean, dtype=torch.float32)[:, None, None]
+    img_std = torch.tensor(img_std, dtype=torch.float32)[:, None, None]
+
+    if async_loading_frames:
+        lazy_images = AsyncVideoFrameLoader(
+            img_paths, image_size, offload_video_to_cpu, img_mean, img_std
+        )
+        return lazy_images, lazy_images.video_height, lazy_images.video_width
+
+    images = torch.zeros(num_frames, 3, image_size, image_size, dtype=torch.float32)
+    for n, img_path in enumerate(tqdm(img_paths, desc="frame loading (JPEG)")):
+        images[n], video_height, video_width = _load_img_as_tensor(img_path, image_size)
+    if not offload_video_to_cpu:
+        images = images.cuda()
+        img_mean = img_mean.cuda()
+        img_std = img_std.cuda()
+    # normalize by mean and std
+    images -= img_mean
+    images /= img_std
+    return images, video_height, video_width
+
+
+def fill_holes_in_mask_scores(mask, max_area):
+    """
+    A post processor to fill small holes in mask scores with area under `max_area`.
+    """
+    # Holes are those connected components in background with area <= self.max_area
+    # (background regions are those with mask scores <= 0)
+    assert max_area > 0, "max_area must be positive"
+    labels, areas = get_connected_components(mask <= 0)
+    is_hole = (labels > 0) & (areas <= max_area)
+    # We fill holes with a small positive mask score (0.1) to change them to foreground.
+    mask = torch.where(is_hole, 0.1, mask)
+    return mask
+
+
+def concat_points(old_point_inputs, new_points, new_labels):
+    """Add new points and labels to previous point inputs (add at the end)."""
+    if old_point_inputs is None:
+        points, labels = new_points, new_labels
+    else:
+        points = torch.cat([old_point_inputs["point_coords"], new_points], dim=1)
+        labels = torch.cat([old_point_inputs["point_labels"], new_labels], dim=1)
+
+    return {"point_coords": points, "point_labels": labels}
diff --git a/third_parts/sam2/utils/transforms.py b/third_parts/sam2/utils/transforms.py
new file mode 100644
index 0000000000000000000000000000000000000000..d877a460bc7b115f4a58e34b31d927b223097b71
--- /dev/null
+++ b/third_parts/sam2/utils/transforms.py
@@ -0,0 +1,99 @@
+# Copyright (c) Meta Platforms, Inc. and affiliates.
+# All rights reserved.
+
+# This source code is licensed under the license found in the
+# LICENSE file in the root directory of this source tree.
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from torchvision.transforms import Normalize, Resize, ToTensor
+
+
+class SAM2Transforms(nn.Module):
+    def __init__(
+        self, resolution, mask_threshold, max_hole_area=0.0, max_sprinkle_area=0.0
+    ):
+        """
+        Transforms for SAM2.
+        """
+        super().__init__()
+        self.resolution = resolution
+        self.mask_threshold = mask_threshold
+        self.max_hole_area = max_hole_area
+        self.max_sprinkle_area = max_sprinkle_area
+        self.mean = [0.485, 0.456, 0.406]
+        self.std = [0.229, 0.224, 0.225]
+        self.to_tensor = ToTensor()
+        self.transforms = torch.jit.script(
+            nn.Sequential(
+                Resize((self.resolution, self.resolution)),
+                Normalize(self.mean, self.std),
+            )
+        )
+
+    def __call__(self, x):
+        x = self.to_tensor(x)
+        return self.transforms(x)
+
+    def forward_batch(self, img_list):
+        img_batch = [self.transforms(self.to_tensor(img)) for img in img_list]
+        img_batch = torch.stack(img_batch, dim=0)
+        return img_batch
+
+    def transform_coords(
+        self, coords: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a torch tensor with length 2 in the last dimension. The coordinates can be in absolute image or normalized coordinates,
+        If the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+
+        Returns
+            Un-normalized coordinates in the range of [0, 1] which is expected by the SAM2 model.
+        """
+        if normalize:
+            assert orig_hw is not None
+            h, w = orig_hw
+            coords = coords.clone()
+            coords[..., 0] = coords[..., 0] / w
+            coords[..., 1] = coords[..., 1] / h
+
+        coords = coords * self.resolution  # unnormalize coords
+        return coords
+
+    def transform_boxes(
+        self, boxes: torch.Tensor, normalize=False, orig_hw=None
+    ) -> torch.Tensor:
+        """
+        Expects a tensor of shape Bx4. The coordinates can be in absolute image or normalized coordinates,
+        if the coords are in absolute image coordinates, normalize should be set to True and original image size is required.
+        """
+        boxes = self.transform_coords(boxes.reshape(-1, 2, 2), normalize, orig_hw)
+        return boxes
+
+    def postprocess_masks(self, masks: torch.Tensor, orig_hw) -> torch.Tensor:
+        """
+        Perform PostProcessing on output masks.
+        """
+        from third_parts.sam2.utils.misc import get_connected_components
+
+        masks = masks.float()
+        if self.max_hole_area > 0:
+            # Holes are those connected components in background with area <= self.fill_hole_area
+            # (background regions are those with mask scores <= self.mask_threshold)
+            mask_flat = masks.flatten(0, 1).unsqueeze(1)  # flatten as 1-channel image
+            labels, areas = get_connected_components(mask_flat <= self.mask_threshold)
+            is_hole = (labels > 0) & (areas <= self.max_hole_area)
+            is_hole = is_hole.reshape_as(masks)
+            # We fill holes with a small positive mask score (10.0) to change them to foreground.
+            masks = torch.where(is_hole, self.mask_threshold + 10.0, masks)
+
+        if self.max_sprinkle_area > 0:
+            labels, areas = get_connected_components(mask_flat > self.mask_threshold)
+            is_hole = (labels > 0) & (areas <= self.max_sprinkle_area)
+            is_hole = is_hole.reshape_as(masks)
+            # We fill holes with negative mask score (-10.0) to change them to background.
+            masks = torch.where(is_hole, self.mask_threshold - 10.0, masks)
+
+        masks = F.interpolate(masks, orig_hw, mode="bilinear", align_corners=False)
+        return masks