Set up Gradio app for Trellis

.gitattributes

@@ -33,3 +33,7 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 *.zip filter=lfs diff=lfs merge=lfs -text
 *.zst filter=lfs diff=lfs merge=lfs -text
 *tfevents* filter=lfs diff=lfs merge=lfs -text
+*.webp filter=lfs diff=lfs merge=lfs -text
+*.png filter=lfs diff=lfs merge=lfs -text
+*.jpg filter=lfs diff=lfs merge=lfs -text
+*.jpeg filter=lfs diff=lfs merge=lfs -text

.gitignore

@@ -0,0 +1,49 @@
+# Python
+__pycache__/
+*.py[cod]
+*$py.class
+*.so
+.Python
+build/
+develop-eggs/
+dist/
+downloads/
+eggs/
+.eggs/
+lib/
+lib64/
+parts/
+sdist/
+var/
+wheels/
+*.egg-info/
+.installed.cfg
+*.egg
+
+# Virtual environments
+venv/
+env/
+ENV/
+.venv
+
+# IDE
+.vscode/
+.idea/
+*.swp
+*.swo
+*~
+.DS_Store
+
+# Temporary files
+tmp/
+*.tmp
+*.log
+
+# Model cache (if any)
+.cache/
+*.pth
+*.pt
+*.ckpt
+
+# Jupyter
+.ipynb_checkpoints/

README.md

@@ -8,7 +8,65 @@ sdk_version: 6.1.0
 app_file: app.py
 pinned: false
 license: mit
-short_description: Space to work with microsoft/TRELLIS image to 3D model.
+short_description: Convert image to high-quality 3D model via microsoft/TRELLIS
+hardware: zero-gpu-t4
 ---
 
-Check out the configuration reference at https://huggingface.co/docs/hub/spaces-config-reference
+# TRELLIS Image to 3D
+
+Convert 2D images into high-quality 3D models using [TRELLIS](https://trellis3d.github.io/), Microsoft's scalable and versatile 3D generation model.
+
+## 🚀 Features
+
+- **Single Image to 3D**: Generate 3D models from a single input image
+- **Multi-Image Support**: Use multiple views of an object for better reconstruction (experimental)
+- **Multiple Output Formats**:
+  - GLB files for use in 3D applications and game engines
+  - Gaussian Splatting (.ply) files for advanced rendering
+- **Interactive 3D Viewer**: Preview your generated models directly in the browser
+- **Automatic Background Removal**: Uses alpha channel or automatic background removal
+- **Configurable Generation**: Adjust sampling steps and guidance strength for fine-tuned results
+
+## 📖 How to Use
+
+1. **Upload an Image**: Click on the image input area and select an image, or choose from the example images below
+2. **Configure Settings** (optional): Expand "Generation Settings" to adjust:
+   - Seed for reproducibility
+   - Sparse Structure Generation parameters (Stage 1)
+   - Structured Latent Generation parameters (Stage 2)
+3. **Generate**: Click "Generate & Extract GLB" to create your 3D model
+4. **Download**: Once generation is complete, download the GLB file or extract Gaussian splatting data
+
+## 💡 Tips for Best Results
+
+- **Image Quality**: Use clear, well-lit images with good contrast
+- **Alpha Channel**: Images with transparent backgrounds (alpha channel) work best
+- **Object Focus**: Ensure the main object is clearly visible and centered
+- **Background**: The tool automatically removes backgrounds if no alpha channel is present
+
+## 🔧 Technical Details
+
+- **Model**: [microsoft/TRELLIS](https://huggingface.co/microsoft/TRELLIS)
+- **Hardware**: ZeroGPU (T4) - GPU resources are allocated on-demand
+- **Processing Time**: Typically 2-5 minutes depending on image complexity and GPU availability
+
+## 📚 Resources
+
+- [TRELLIS Project Page](https://trellis3d.github.io/)
+- [Paper](https://huggingface.co/papers/2412.01506)
+- [Model Card](https://huggingface.co/microsoft/TRELLIS)
+
+## 📝 Output Formats
+
+- **GLB**: Universal 3D format compatible with most 3D software, game engines, and web viewers
+- **Gaussian Splatting (.ply)**: Advanced point-based representation for high-quality rendering
+
+## ⚠️ Notes
+
+- Multi-image mode is experimental and may not produce optimal results for all image sets
+- Gaussian splatting files can be large (~50MB) and may take time to download
+- Processing requires GPU resources - you may need to wait if all GPUs are in use
+
+---
+
+Built with [Gradio](https://gradio.app/) and powered by [Hugging Face Spaces](https://huggingface.co/spaces)

app.py

@@ -1,7 +1,425 @@
 import gradio as gr
+import spaces
+from gradio_litmodel3d import LitModel3D
 
-def greet(name):
-    return "Hello " + name + "!!"
+import os
+import shutil
+os.environ['SPCONV_ALGO'] = 'native'
+from typing import *
+import torch
+import numpy as np
+import imageio
+from easydict import EasyDict as edict
+from PIL import Image
+from trellis.pipelines import TrellisImageTo3DPipeline
+from trellis.representations import Gaussian, MeshExtractResult
+from trellis.utils import render_utils, postprocessing_utils
 
-demo = gr.Interface(fn=greet, inputs="text", outputs="text")
-demo.launch()
+
+MAX_SEED = np.iinfo(np.int32).max
+TMP_DIR = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'tmp')
+os.makedirs(TMP_DIR, exist_ok=True)
+
+
+def start_session(req: gr.Request):
+    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
+    os.makedirs(user_dir, exist_ok=True)
+    
+    
+def end_session(req: gr.Request):
+    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
+    if os.path.exists(user_dir):
+        shutil.rmtree(user_dir)
+
+
+def preprocess_image(image: Image.Image) -> Image.Image:
+    """
+    Preprocess the input image for 3D generation.
+    
+    This function is called when a user uploads an image or selects an example.
+    It applies background removal and other preprocessing steps necessary for
+    optimal 3D model generation.
+
+    Args:
+        image (Image.Image): The input image from the user
+
+    Returns:
+        Image.Image: The preprocessed image ready for 3D generation
+    """
+    processed_image = pipeline.preprocess_image(image)
+    return processed_image
+
+
+def preprocess_images(images: List[Tuple[Image.Image, str]]) -> List[Image.Image]:
+    """
+    Preprocess a list of input images for multi-image 3D generation.
+    
+    This function is called when users upload multiple images in the gallery.
+    It processes each image to prepare them for the multi-image 3D generation pipeline.
+    
+    Args:
+        images (List[Tuple[Image.Image, str]]): The input images from the gallery
+        
+    Returns:
+        List[Image.Image]: The preprocessed images ready for 3D generation
+    """
+    images = [image[0] for image in images]
+    processed_images = [pipeline.preprocess_image(image) for image in images]
+    return processed_images
+
+
+def pack_state(gs: Gaussian, mesh: MeshExtractResult) -> dict:
+    return {
+        'gaussian': {
+            **gs.init_params,
+            '_xyz': gs._xyz.cpu().numpy(),
+            '_features_dc': gs._features_dc.cpu().numpy(),
+            '_scaling': gs._scaling.cpu().numpy(),
+            '_rotation': gs._rotation.cpu().numpy(),
+            '_opacity': gs._opacity.cpu().numpy(),
+        },
+        'mesh': {
+            'vertices': mesh.vertices.cpu().numpy(),
+            'faces': mesh.faces.cpu().numpy(),
+        },
+    }
+    
+    
+def unpack_state(state: dict) -> Tuple[Gaussian, edict, str]:
+    gs = Gaussian(
+        aabb=state['gaussian']['aabb'],
+        sh_degree=state['gaussian']['sh_degree'],
+        mininum_kernel_size=state['gaussian']['mininum_kernel_size'],
+        scaling_bias=state['gaussian']['scaling_bias'],
+        opacity_bias=state['gaussian']['opacity_bias'],
+        scaling_activation=state['gaussian']['scaling_activation'],
+    )
+    gs._xyz = torch.tensor(state['gaussian']['_xyz'], device='cuda')
+    gs._features_dc = torch.tensor(state['gaussian']['_features_dc'], device='cuda')
+    gs._scaling = torch.tensor(state['gaussian']['_scaling'], device='cuda')
+    gs._rotation = torch.tensor(state['gaussian']['_rotation'], device='cuda')
+    gs._opacity = torch.tensor(state['gaussian']['_opacity'], device='cuda')
+    
+    mesh = edict(
+        vertices=torch.tensor(state['mesh']['vertices'], device='cuda'),
+        faces=torch.tensor(state['mesh']['faces'], device='cuda'),
+    )
+    
+    return gs, mesh
+
+
+def get_seed(randomize_seed: bool, seed: int) -> int:
+    """
+    Get the random seed for generation.
+    
+    This function is called by the generate button to determine whether to use
+    a random seed or the user-specified seed value.
+    
+    Args:
+        randomize_seed (bool): Whether to generate a random seed
+        seed (int): The user-specified seed value
+        
+    Returns:
+        int: The seed to use for generation
+    """
+    return np.random.randint(0, MAX_SEED) if randomize_seed else seed
+
+
+@spaces.GPU(duration=120)
+def generate_and_extract_glb(
+    image: Image.Image,
+    multiimages: List[Tuple[Image.Image, str]],
+    is_multiimage: bool,
+    seed: int,
+    ss_guidance_strength: float,
+    ss_sampling_steps: int,
+    slat_guidance_strength: float,
+    slat_sampling_steps: int,
+    multiimage_algo: Literal["multidiffusion", "stochastic"],
+    mesh_simplify: float,
+    texture_size: int,
+    req: gr.Request,
+) -> Tuple[dict, str, str, str]:
+    """
+    Convert an image to a 3D model and extract GLB file.
+
+    Args:
+        image (Image.Image): The input image.
+        multiimages (List[Tuple[Image.Image, str]]): The input images in multi-image mode.
+        is_multiimage (bool): Whether is in multi-image mode.
+        seed (int): The random seed.
+        ss_guidance_strength (float): The guidance strength for sparse structure generation.
+        ss_sampling_steps (int): The number of sampling steps for sparse structure generation.
+        slat_guidance_strength (float): The guidance strength for structured latent generation.
+        slat_sampling_steps (int): The number of sampling steps for structured latent generation.
+        multiimage_algo (Literal["multidiffusion", "stochastic"]): The algorithm for multi-image generation.
+        mesh_simplify (float): The mesh simplification factor.
+        texture_size (int): The texture resolution.
+
+    Returns:
+        dict: The information of the generated 3D model.
+        str: The path to the video of the 3D model.
+        str: The path to the extracted GLB file.
+        str: The path to the extracted GLB file (for download).
+    """
+    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
+    
+    # Generate 3D model
+    if not is_multiimage:
+        outputs = pipeline.run(
+            image,
+            seed=seed,
+            formats=["gaussian", "mesh"],
+            preprocess_image=False,
+            sparse_structure_sampler_params={
+                "steps": ss_sampling_steps,
+                "cfg_strength": ss_guidance_strength,
+            },
+            slat_sampler_params={
+                "steps": slat_sampling_steps,
+                "cfg_strength": slat_guidance_strength,
+            },
+        )
+    else:
+        outputs = pipeline.run_multi_image(
+            [image[0] for image in multiimages],
+            seed=seed,
+            formats=["gaussian", "mesh"],
+            preprocess_image=False,
+            sparse_structure_sampler_params={
+                "steps": ss_sampling_steps,
+                "cfg_strength": ss_guidance_strength,
+            },
+            slat_sampler_params={
+                "steps": slat_sampling_steps,
+                "cfg_strength": slat_guidance_strength,
+            },
+            mode=multiimage_algo,
+        )
+    
+    # Render video
+    video = render_utils.render_video(outputs['gaussian'][0], num_frames=120)['color']
+    video_geo = render_utils.render_video(outputs['mesh'][0], num_frames=120)['normal']
+    video = [np.concatenate([video[i], video_geo[i]], axis=1) for i in range(len(video))]
+    video_path = os.path.join(user_dir, 'sample.mp4')
+    imageio.mimsave(video_path, video, fps=15)
+    
+    # Extract GLB
+    gs = outputs['gaussian'][0]
+    mesh = outputs['mesh'][0]
+    glb = postprocessing_utils.to_glb(gs, mesh, simplify=mesh_simplify, texture_size=texture_size, verbose=False)
+    glb_path = os.path.join(user_dir, 'sample.glb')
+    glb.export(glb_path)
+    
+    # Pack state for optional Gaussian extraction
+    state = pack_state(gs, mesh)
+    
+    torch.cuda.empty_cache()
+    return state, video_path, glb_path, glb_path
+
+
+@spaces.GPU
+def extract_gaussian(state: dict, req: gr.Request) -> Tuple[str, str]:
+    """
+    Extract a Gaussian splatting file from the generated 3D model.
+    
+    This function is called when the user clicks "Extract Gaussian" button.
+    It converts the 3D model state into a .ply file format containing
+    Gaussian splatting data for advanced 3D applications.
+
+    Args:
+        state (dict): The state of the generated 3D model containing Gaussian data
+        req (gr.Request): Gradio request object for session management
+
+    Returns:
+        Tuple[str, str]: Paths to the extracted Gaussian file (for display and download)
+    """
+    user_dir = os.path.join(TMP_DIR, str(req.session_hash))
+    gs, _ = unpack_state(state)
+    gaussian_path = os.path.join(user_dir, 'sample.ply')
+    gs.save_ply(gaussian_path)
+    torch.cuda.empty_cache()
+    return gaussian_path, gaussian_path
+
+
+def prepare_multi_example() -> List[Image.Image]:
+    # Multi-image examples removed - using only assets/images
+    return []
+
+
+def split_image(image: Image.Image) -> List[Image.Image]:
+    """
+    Split a multi-view image into separate view images.
+    
+    This function is called when users select multi-image examples that contain
+    multiple views in a single concatenated image. It automatically splits them
+    based on alpha channel boundaries and preprocesses each view.
+    
+    Args:
+        image (Image.Image): A concatenated image containing multiple views
+        
+    Returns:
+        List[Image.Image]: List of individual preprocessed view images
+    """
+    image = np.array(image)
+    alpha = image[..., 3]
+    alpha = np.any(alpha>0, axis=0)
+    start_pos = np.where(~alpha[:-1] & alpha[1:])[0].tolist()
+    end_pos = np.where(alpha[:-1] & ~alpha[1:])[0].tolist()
+    images = []
+    for s, e in zip(start_pos, end_pos):
+        images.append(Image.fromarray(image[:, s:e+1]))
+    return [preprocess_image(image) for image in images]
+
+
+with gr.Blocks(delete_cache=(600, 600)) as demo:
+    gr.Markdown("""
+    ## Image to 3D Asset with [TRELLIS](https://trellis3d.github.io/)
+    * Upload an image and click "Generate & Extract GLB" to create a 3D asset and automatically extract the GLB file.
+    * If you want the Gaussian file as well, click "Extract Gaussian" after generation.
+    * If the image has alpha channel, it will be used as the mask. Otherwise, we use `rembg` to remove the background.
+    
+    ✨New: 1) Experimental multi-image support. 2) Gaussian file extraction.
+    """)
+    
+    with gr.Row():
+        with gr.Column():
+            with gr.Tabs() as input_tabs:
+                with gr.Tab(label="Single Image", id=0) as single_image_input_tab:
+                    image_prompt = gr.Image(label="Image Prompt", format="png", image_mode="RGBA", type="pil", height=300)
+                with gr.Tab(label="Multiple Images", id=1) as multiimage_input_tab:
+                    multiimage_prompt = gr.Gallery(label="Image Prompt", format="png", type="pil", height=300, columns=3)
+                    gr.Markdown("""
+                        Input different views of the object in separate images. 
+                        
+                        *NOTE: this is an experimental algorithm without training a specialized model. It may not produce the best results for all images, especially those having different poses or inconsistent details.*
+                    """)
+            
+            with gr.Accordion(label="Generation Settings", open=False):
+                seed = gr.Slider(0, MAX_SEED, label="Seed", value=0, step=1)
+                randomize_seed = gr.Checkbox(label="Randomize Seed", value=True)
+                gr.Markdown("Stage 1: Sparse Structure Generation")
+                with gr.Row():
+                    ss_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=7.5, step=0.1)
+                    ss_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
+                gr.Markdown("Stage 2: Structured Latent Generation")
+                with gr.Row():
+                    slat_guidance_strength = gr.Slider(0.0, 10.0, label="Guidance Strength", value=3.0, step=0.1)
+                    slat_sampling_steps = gr.Slider(1, 50, label="Sampling Steps", value=12, step=1)
+                multiimage_algo = gr.Radio(["stochastic", "multidiffusion"], label="Multi-image Algorithm", value="stochastic")
+            
+            with gr.Accordion(label="GLB Extraction Settings", open=False):
+                mesh_simplify = gr.Slider(0.9, 0.98, label="Simplify", value=0.95, step=0.01)
+                texture_size = gr.Slider(512, 2048, label="Texture Size", value=1024, step=512)
+
+            generate_btn = gr.Button("Generate & Extract GLB", variant="primary")
+            extract_gs_btn = gr.Button("Extract Gaussian", interactive=False)
+            gr.Markdown("""
+                        *NOTE: Gaussian file can be very large (~50MB), it will take a while to display and download.*
+                        """)
+
+        with gr.Column():
+            video_output = gr.Video(label="Generated 3D Asset", autoplay=True, loop=True, height=300)
+            model_output = LitModel3D(label="Extracted GLB/Gaussian", exposure=10.0, height=300)
+            
+            with gr.Row():
+                download_glb = gr.DownloadButton(label="Download GLB", interactive=False)
+                download_gs = gr.DownloadButton(label="Download Gaussian", interactive=False)  
+    
+    is_multiimage = gr.State(False)
+    output_buf = gr.State()
+
+    # Example images at the bottom of the page
+    with gr.Row() as single_image_example:
+        if os.path.exists("assets/images"):
+            examples = gr.Examples(
+                examples=[
+                    f'assets/images/{image}'
+                    for image in os.listdir("assets/images")
+                    if image.endswith(('.png', '.jpg', '.jpeg', '.webp'))
+                ],
+                inputs=[image_prompt],
+                fn=preprocess_image,
+                outputs=[image_prompt],
+                run_on_click=True,
+                examples_per_page=64,
+            )
+        else:
+            examples = gr.Examples(examples=[], inputs=[image_prompt])
+    
+    with gr.Row(visible=False) as multiimage_example:
+        examples_multi = gr.Examples(
+            examples=prepare_multi_example(),
+            inputs=[image_prompt],
+            fn=split_image,
+            outputs=[multiimage_prompt],
+            run_on_click=True,
+            examples_per_page=8,
+        )
+
+    # Handlers
+    demo.load(start_session)
+    demo.unload(end_session)
+    
+    single_image_input_tab.select(
+        lambda: tuple([False, gr.Row.update(visible=True), gr.Row.update(visible=False)]),
+        outputs=[is_multiimage, single_image_example, multiimage_example]
+    )
+    multiimage_input_tab.select(
+        lambda: tuple([True, gr.Row.update(visible=False), gr.Row.update(visible=True)]),
+        outputs=[is_multiimage, single_image_example, multiimage_example]
+    )
+    
+    image_prompt.upload(
+        preprocess_image,
+        inputs=[image_prompt],
+        outputs=[image_prompt],
+    )
+    multiimage_prompt.upload(
+        preprocess_images,
+        inputs=[multiimage_prompt],
+        outputs=[multiimage_prompt],
+    )
+
+    generate_btn.click(
+        get_seed,
+        inputs=[randomize_seed, seed],
+        outputs=[seed],
+    ).then(
+        generate_and_extract_glb,
+        inputs=[image_prompt, multiimage_prompt, is_multiimage, seed, ss_guidance_strength, ss_sampling_steps, slat_guidance_strength, slat_sampling_steps, multiimage_algo, mesh_simplify, texture_size],
+        outputs=[output_buf, video_output, model_output, download_glb],
+    ).then(
+        lambda: tuple([gr.Button(interactive=True), gr.Button(interactive=True)]),
+        outputs=[extract_gs_btn, download_glb],
+    )
+
+    video_output.clear(
+        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False), gr.Button(interactive=False)]),
+        outputs=[extract_gs_btn, download_glb, download_gs],
+    )
+    
+    extract_gs_btn.click(
+        extract_gaussian,
+        inputs=[output_buf],
+        outputs=[model_output, download_gs],
+    ).then(
+        lambda: gr.Button(interactive=True),
+        outputs=[download_gs],
+    )
+
+    model_output.clear(
+        lambda: tuple([gr.Button(interactive=False), gr.Button(interactive=False)]),
+        outputs=[download_glb, download_gs],
+    )
+    
+
+# Launch the Gradio app
+if __name__ == "__main__":
+    pipeline = TrellisImageTo3DPipeline.from_pretrained("microsoft/TRELLIS")
+    pipeline.cuda()
+    try:
+        pipeline.preprocess_image(Image.fromarray(np.zeros((512, 512, 3), dtype=np.uint8)))    # Preload rembg
+    except:
+        pass
+    demo.launch()

requirements.txt

@@ -0,0 +1,28 @@
+--extra-index-url https://download.pytorch.org/whl/cu121
+
+torch==2.4.0
+torchvision==0.19.0
+pillow==10.4.0
+imageio==2.36.1
+imageio-ffmpeg==0.5.1
+tqdm==4.67.1
+easydict==1.13
+opencv-python-headless==4.10.0.84
+scipy==1.14.1
+rembg==2.0.60
+onnxruntime==1.20.1
+trimesh==4.5.3
+xatlas==0.0.9
+pyvista==0.44.2
+pymeshfix==0.17.0
+igraph==0.11.8
+git+https://github.com/EasternJournalist/utils3d.git@9a4eb15e4021b67b12c460c7057d642626897ec8
+xformers==0.0.27.post2
+spconv-cu120==2.3.6
+transformers==4.46.3
+gradio_litmodel3d==0.0.1
+pydantic==2.10.6
+https://github.com/Dao-AILab/flash-attention/releases/download/v2.7.0.post2/flash_attn-2.7.0.post2+cu12torch2.4cxx11abiFALSE-cp310-cp310-linux_x86_64.whl
+https://huggingface.co/spaces/JeffreyXiang/TRELLIS/resolve/main/wheels/diff_gaussian_rasterization-0.0.0-cp310-cp310-linux_x86_64.whl?download=true
+https://huggingface.co/spaces/JeffreyXiang/TRELLIS/resolve/main/wheels/nvdiffrast-0.3.3-cp310-cp310-linux_x86_64.whl?download=true
+

trellis/__init__.py

@@ -0,0 +1,6 @@
+from . import models
+from . import modules
+from . import pipelines
+from . import renderers
+from . import representations
+from . import utils

trellis/models/__init__.py

@@ -0,0 +1,70 @@
+import importlib
+
+__attributes = {
+    'SparseStructureEncoder': 'sparse_structure_vae',
+    'SparseStructureDecoder': 'sparse_structure_vae',
+    'SparseStructureFlowModel': 'sparse_structure_flow',
+    'SLatEncoder': 'structured_latent_vae',
+    'SLatGaussianDecoder': 'structured_latent_vae',
+    'SLatRadianceFieldDecoder': 'structured_latent_vae',
+    'SLatMeshDecoder': 'structured_latent_vae',
+    'SLatFlowModel': 'structured_latent_flow',
+}
+
+__submodules = []
+
+__all__ = list(__attributes.keys()) + __submodules
+
+def __getattr__(name):
+    if name not in globals():
+        if name in __attributes:
+            module_name = __attributes[name]
+            module = importlib.import_module(f".{module_name}", __name__)
+            globals()[name] = getattr(module, name)
+        elif name in __submodules:
+            module = importlib.import_module(f".{name}", __name__)
+            globals()[name] = module
+        else:
+            raise AttributeError(f"module {__name__} has no attribute {name}")
+    return globals()[name]
+
+
+def from_pretrained(path: str, **kwargs):
+    """
+    Load a model from a pretrained checkpoint.
+
+    Args:
+        path: The path to the checkpoint. Can be either local path or a Hugging Face model name.
+              NOTE: config file and model file should take the name f'{path}.json' and f'{path}.safetensors' respectively.
+        **kwargs: Additional arguments for the model constructor.
+    """
+    import os
+    import json
+    from safetensors.torch import load_file
+    is_local = os.path.exists(f"{path}.json") and os.path.exists(f"{path}.safetensors")
+
+    if is_local:
+        config_file = f"{path}.json"
+        model_file = f"{path}.safetensors"
+    else:
+        from huggingface_hub import hf_hub_download
+        path_parts = path.split('/')
+        repo_id = f'{path_parts[0]}/{path_parts[1]}'
+        model_name = '/'.join(path_parts[2:])
+        config_file = hf_hub_download(repo_id, f"{model_name}.json")
+        model_file = hf_hub_download(repo_id, f"{model_name}.safetensors")
+
+    with open(config_file, 'r') as f:
+        config = json.load(f)
+    model = __getattr__(config['name'])(**config['args'], **kwargs)
+    model.load_state_dict(load_file(model_file))
+
+    return model
+
+
+# For Pylance
+if __name__ == '__main__':
+    from .sparse_structure_vae import SparseStructureEncoder, SparseStructureDecoder
+    from .sparse_structure_flow import SparseStructureFlowModel
+    from .structured_latent_vae import SLatEncoder, SLatGaussianDecoder, SLatRadianceFieldDecoder, SLatMeshDecoder
+    from .structured_latent_flow import SLatFlowModel

trellis/models/sparse_structure_flow.py

@@ -0,0 +1,200 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ..modules.utils import convert_module_to_f16, convert_module_to_f32
+from ..modules.transformer import AbsolutePositionEmbedder, ModulatedTransformerCrossBlock
+from ..modules.spatial import patchify, unpatchify
+
+
+class TimestepEmbedder(nn.Module):
+    """
+    Embeds scalar timesteps into vector representations.
+    """
+    def __init__(self, hidden_size, frequency_embedding_size=256):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(frequency_embedding_size, hidden_size, bias=True),
+            nn.SiLU(),
+            nn.Linear(hidden_size, hidden_size, bias=True),
+        )
+        self.frequency_embedding_size = frequency_embedding_size
+
+    @staticmethod
+    def timestep_embedding(t, dim, max_period=10000):
+        """
+        Create sinusoidal timestep embeddings.
+
+        Args:
+            t: a 1-D Tensor of N indices, one per batch element.
+                These may be fractional.
+            dim: the dimension of the output.
+            max_period: controls the minimum frequency of the embeddings.
+
+        Returns:
+            an (N, D) Tensor of positional embeddings.
+        """
+        # https://github.com/openai/glide-text2im/blob/main/glide_text2im/nn.py
+        half = dim // 2
+        freqs = torch.exp(
+            -np.log(max_period) * torch.arange(start=0, end=half, dtype=torch.float32) / half
+        ).to(device=t.device)
+        args = t[:, None].float() * freqs[None]
+        embedding = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
+        if dim % 2:
+            embedding = torch.cat([embedding, torch.zeros_like(embedding[:, :1])], dim=-1)
+        return embedding
+
+    def forward(self, t):
+        t_freq = self.timestep_embedding(t, self.frequency_embedding_size)
+        t_emb = self.mlp(t_freq)
+        return t_emb
+
+
+class SparseStructureFlowModel(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        cond_channels: int,
+        out_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        patch_size: int = 2,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        share_mod: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.cond_channels = cond_channels
+        self.out_channels = out_channels
+        self.num_blocks = num_blocks
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.patch_size = patch_size
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.qk_rms_norm = qk_rms_norm
+        self.qk_rms_norm_cross = qk_rms_norm_cross
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.t_embedder = TimestepEmbedder(model_channels)
+        if share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(model_channels, 6 * model_channels, bias=True)
+            )
+
+        if pe_mode == "ape":
+            pos_embedder = AbsolutePositionEmbedder(model_channels, 3)
+            coords = torch.meshgrid(*[torch.arange(res, device=self.device) for res in [resolution // patch_size] * 3], indexing='ij')
+            coords = torch.stack(coords, dim=-1).reshape(-1, 3)
+            pos_emb = pos_embedder(coords)
+            self.register_buffer("pos_emb", pos_emb)
+
+        self.input_layer = nn.Linear(in_channels * patch_size**3, model_channels)
+            
+        self.blocks = nn.ModuleList([
+            ModulatedTransformerCrossBlock(
+                model_channels,
+                cond_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode='full',
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                share_mod=share_mod,
+                qk_rms_norm=self.qk_rms_norm,
+                qk_rms_norm_cross=self.qk_rms_norm_cross,
+            )
+            for _ in range(num_blocks)
+        ])
+
+        self.out_layer = nn.Linear(model_channels, out_channels * patch_size**3)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+        # Initialize timestep embedding MLP:
+        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        # Zero-out adaLN modulation layers in DiT blocks:
+        if self.share_mod:
+            nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
+        else:
+            for block in self.blocks:
+                nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+                nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: torch.Tensor, t: torch.Tensor, cond: torch.Tensor) -> torch.Tensor:
+        assert [*x.shape] == [x.shape[0], self.in_channels, *[self.resolution] * 3], \
+                f"Input shape mismatch, got {x.shape}, expected {[x.shape[0], self.in_channels, *[self.resolution] * 3]}"
+
+        h = patchify(x, self.patch_size)
+        h = h.view(*h.shape[:2], -1).permute(0, 2, 1).contiguous()
+
+        h = self.input_layer(h)
+        h = h + self.pos_emb[None]
+        t_emb = self.t_embedder(t)
+        if self.share_mod:
+            t_emb = self.adaLN_modulation(t_emb)
+        t_emb = t_emb.type(self.dtype)
+        h = h.type(self.dtype)
+        cond = cond.type(self.dtype)
+        for block in self.blocks:
+            h = block(h, t_emb, cond)
+        h = h.type(x.dtype)
+        h = F.layer_norm(h, h.shape[-1:])
+        h = self.out_layer(h)
+
+        h = h.permute(0, 2, 1).view(h.shape[0], h.shape[2], *[self.resolution // self.patch_size] * 3)
+        h = unpatchify(h, self.patch_size).contiguous()
+
+        return h

trellis/models/sparse_structure_vae.py

@@ -0,0 +1,306 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ..modules.norm import GroupNorm32, ChannelLayerNorm32
+from ..modules.spatial import pixel_shuffle_3d
+from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+
+
+def norm_layer(norm_type: str, *args, **kwargs) -> nn.Module:
+    """
+    Return a normalization layer.
+    """
+    if norm_type == "group":
+        return GroupNorm32(32, *args, **kwargs)
+    elif norm_type == "layer":
+        return ChannelLayerNorm32(*args, **kwargs)
+    else:
+        raise ValueError(f"Invalid norm type {norm_type}")
+
+
+class ResBlock3d(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        out_channels: Optional[int] = None,
+        norm_type: Literal["group", "layer"] = "layer",
+    ):
+        super().__init__()
+        self.channels = channels
+        self.out_channels = out_channels or channels
+
+        self.norm1 = norm_layer(norm_type, channels)
+        self.norm2 = norm_layer(norm_type, self.out_channels)
+        self.conv1 = nn.Conv3d(channels, self.out_channels, 3, padding=1)
+        self.conv2 = zero_module(nn.Conv3d(self.out_channels, self.out_channels, 3, padding=1))
+        self.skip_connection = nn.Conv3d(channels, self.out_channels, 1) if channels != self.out_channels else nn.Identity()
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.norm1(x)
+        h = F.silu(h)
+        h = self.conv1(h)
+        h = self.norm2(h)
+        h = F.silu(h)
+        h = self.conv2(h)
+        h = h + self.skip_connection(x)
+        return h
+
+
+class DownsampleBlock3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        mode: Literal["conv", "avgpool"] = "conv",
+    ):
+        assert mode in ["conv", "avgpool"], f"Invalid mode {mode}"
+
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if mode == "conv":
+            self.conv = nn.Conv3d(in_channels, out_channels, 2, stride=2)
+        elif mode == "avgpool":
+            assert in_channels == out_channels, "Pooling mode requires in_channels to be equal to out_channels"
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if hasattr(self, "conv"):
+            return self.conv(x)
+        else:
+            return F.avg_pool3d(x, 2)
+
+
+class UpsampleBlock3d(nn.Module):
+    def __init__(
+        self,
+        in_channels: int,
+        out_channels: int,
+        mode: Literal["conv", "nearest"] = "conv",
+    ):
+        assert mode in ["conv", "nearest"], f"Invalid mode {mode}"
+
+        super().__init__()
+        self.in_channels = in_channels
+        self.out_channels = out_channels
+
+        if mode == "conv":
+            self.conv = nn.Conv3d(in_channels, out_channels*8, 3, padding=1)
+        elif mode == "nearest":
+            assert in_channels == out_channels, "Nearest mode requires in_channels to be equal to out_channels"
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if hasattr(self, "conv"):
+            x = self.conv(x)
+            return pixel_shuffle_3d(x, 2)
+        else:
+            return F.interpolate(x, scale_factor=2, mode="nearest")
+        
+
+class SparseStructureEncoder(nn.Module):
+    """
+    Encoder for Sparse Structure (\mathcal{E}_S in the paper Sec. 3.3).
+    
+    Args:
+        in_channels (int): Channels of the input.
+        latent_channels (int): Channels of the latent representation.
+        num_res_blocks (int): Number of residual blocks at each resolution.
+        channels (List[int]): Channels of the encoder blocks.
+        num_res_blocks_middle (int): Number of residual blocks in the middle.
+        norm_type (Literal["group", "layer"]): Type of normalization layer.
+        use_fp16 (bool): Whether to use FP16.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        latent_channels: int,
+        num_res_blocks: int,
+        channels: List[int],
+        num_res_blocks_middle: int = 2,
+        norm_type: Literal["group", "layer"] = "layer",
+        use_fp16: bool = False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.latent_channels = latent_channels
+        self.num_res_blocks = num_res_blocks
+        self.channels = channels
+        self.num_res_blocks_middle = num_res_blocks_middle
+        self.norm_type = norm_type
+        self.use_fp16 = use_fp16
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.input_layer = nn.Conv3d(in_channels, channels[0], 3, padding=1)
+
+        self.blocks = nn.ModuleList([])
+        for i, ch in enumerate(channels):
+            self.blocks.extend([
+                ResBlock3d(ch, ch)
+                for _ in range(num_res_blocks)
+            ])
+            if i < len(channels) - 1:
+                self.blocks.append(
+                    DownsampleBlock3d(ch, channels[i+1])
+                )
+        
+        self.middle_block = nn.Sequential(*[
+            ResBlock3d(channels[-1], channels[-1])
+            for _ in range(num_res_blocks_middle)
+        ])
+
+        self.out_layer = nn.Sequential(
+            norm_layer(norm_type, channels[-1]),
+            nn.SiLU(),
+            nn.Conv3d(channels[-1], latent_channels*2, 3, padding=1)
+        )
+
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.use_fp16 = True
+        self.dtype = torch.float16
+        self.blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.use_fp16 = False
+        self.dtype = torch.float32
+        self.blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+
+    def forward(self, x: torch.Tensor, sample_posterior: bool = False, return_raw: bool = False) -> torch.Tensor:
+        h = self.input_layer(x)
+        h = h.type(self.dtype)
+
+        for block in self.blocks:
+            h = block(h)
+        h = self.middle_block(h)
+
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+
+        mean, logvar = h.chunk(2, dim=1)
+
+        if sample_posterior:
+            std = torch.exp(0.5 * logvar)
+            z = mean + std * torch.randn_like(std)
+        else:
+            z = mean
+            
+        if return_raw:
+            return z, mean, logvar
+        return z
+        
+
+class SparseStructureDecoder(nn.Module):
+    """
+    Decoder for Sparse Structure (\mathcal{D}_S in the paper Sec. 3.3).
+    
+    Args:
+        out_channels (int): Channels of the output.
+        latent_channels (int): Channels of the latent representation.
+        num_res_blocks (int): Number of residual blocks at each resolution.
+        channels (List[int]): Channels of the decoder blocks.
+        num_res_blocks_middle (int): Number of residual blocks in the middle.
+        norm_type (Literal["group", "layer"]): Type of normalization layer.
+        use_fp16 (bool): Whether to use FP16.
+    """ 
+    def __init__(
+        self,
+        out_channels: int,
+        latent_channels: int,
+        num_res_blocks: int,
+        channels: List[int],
+        num_res_blocks_middle: int = 2,
+        norm_type: Literal["group", "layer"] = "layer",
+        use_fp16: bool = False,
+    ):
+        super().__init__()
+        self.out_channels = out_channels
+        self.latent_channels = latent_channels
+        self.num_res_blocks = num_res_blocks
+        self.channels = channels
+        self.num_res_blocks_middle = num_res_blocks_middle
+        self.norm_type = norm_type
+        self.use_fp16 = use_fp16
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        self.input_layer = nn.Conv3d(latent_channels, channels[0], 3, padding=1)
+
+        self.middle_block = nn.Sequential(*[
+            ResBlock3d(channels[0], channels[0])
+            for _ in range(num_res_blocks_middle)
+        ])
+
+        self.blocks = nn.ModuleList([])
+        for i, ch in enumerate(channels):
+            self.blocks.extend([
+                ResBlock3d(ch, ch)
+                for _ in range(num_res_blocks)
+            ])
+            if i < len(channels) - 1:
+                self.blocks.append(
+                    UpsampleBlock3d(ch, channels[i+1])
+                )
+
+        self.out_layer = nn.Sequential(
+            norm_layer(norm_type, channels[-1]),
+            nn.SiLU(),
+            nn.Conv3d(channels[-1], out_channels, 3, padding=1)
+        )
+
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+    
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.use_fp16 = True
+        self.dtype = torch.float16
+        self.blocks.apply(convert_module_to_f16)
+        self.middle_block.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.use_fp16 = False
+        self.dtype = torch.float32
+        self.blocks.apply(convert_module_to_f32)
+        self.middle_block.apply(convert_module_to_f32)
+    
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.input_layer(x)
+        
+        h = h.type(self.dtype)
+                
+        h = self.middle_block(h)
+        for block in self.blocks:
+            h = block(h)
+
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+        return h

trellis/models/structured_latent_flow.py

@@ -0,0 +1,262 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ..modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+from ..modules.transformer import AbsolutePositionEmbedder
+from ..modules.norm import LayerNorm32
+from ..modules import sparse as sp
+from ..modules.sparse.transformer import ModulatedSparseTransformerCrossBlock
+from .sparse_structure_flow import TimestepEmbedder
+
+
+class SparseResBlock3d(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        emb_channels: int,
+        out_channels: Optional[int] = None,
+        downsample: bool = False,
+        upsample: bool = False,
+    ):
+        super().__init__()
+        self.channels = channels
+        self.emb_channels = emb_channels
+        self.out_channels = out_channels or channels
+        self.downsample = downsample
+        self.upsample = upsample
+        
+        assert not (downsample and upsample), "Cannot downsample and upsample at the same time"
+
+        self.norm1 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
+        self.norm2 = LayerNorm32(self.out_channels, elementwise_affine=False, eps=1e-6)
+        self.conv1 = sp.SparseConv3d(channels, self.out_channels, 3)
+        self.conv2 = zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3))
+        self.emb_layers = nn.Sequential(
+            nn.SiLU(),
+            nn.Linear(emb_channels, 2 * self.out_channels, bias=True),
+        )
+        self.skip_connection = sp.SparseLinear(channels, self.out_channels) if channels != self.out_channels else nn.Identity()
+        self.updown = None
+        if self.downsample:
+            self.updown = sp.SparseDownsample(2)
+        elif self.upsample:
+            self.updown = sp.SparseUpsample(2)
+
+    def _updown(self, x: sp.SparseTensor) -> sp.SparseTensor:
+        if self.updown is not None:
+            x = self.updown(x)
+        return x
+
+    def forward(self, x: sp.SparseTensor, emb: torch.Tensor) -> sp.SparseTensor:
+        emb_out = self.emb_layers(emb).type(x.dtype)
+        scale, shift = torch.chunk(emb_out, 2, dim=1)
+
+        x = self._updown(x)
+        h = x.replace(self.norm1(x.feats))
+        h = h.replace(F.silu(h.feats))
+        h = self.conv1(h)
+        h = h.replace(self.norm2(h.feats)) * (1 + scale) + shift
+        h = h.replace(F.silu(h.feats))
+        h = self.conv2(h)
+        h = h + self.skip_connection(x)
+
+        return h
+    
+
+class SLatFlowModel(nn.Module):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        cond_channels: int,
+        out_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        patch_size: int = 2,
+        num_io_res_blocks: int = 2,
+        io_block_channels: List[int] = None,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        use_skip_connection: bool = True,
+        share_mod: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+    ):
+        super().__init__()
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.cond_channels = cond_channels
+        self.out_channels = out_channels
+        self.num_blocks = num_blocks
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.patch_size = patch_size
+        self.num_io_res_blocks = num_io_res_blocks
+        self.io_block_channels = io_block_channels
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.use_skip_connection = use_skip_connection
+        self.share_mod = share_mod
+        self.qk_rms_norm = qk_rms_norm
+        self.qk_rms_norm_cross = qk_rms_norm_cross
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        assert int(np.log2(patch_size)) == np.log2(patch_size), "Patch size must be a power of 2"
+        assert np.log2(patch_size) == len(io_block_channels), "Number of IO ResBlocks must match the number of stages"
+
+        self.t_embedder = TimestepEmbedder(model_channels)
+        if share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(model_channels, 6 * model_channels, bias=True)
+            )
+
+        if pe_mode == "ape":
+            self.pos_embedder = AbsolutePositionEmbedder(model_channels)
+
+        self.input_layer = sp.SparseLinear(in_channels, io_block_channels[0])
+        self.input_blocks = nn.ModuleList([])
+        for chs, next_chs in zip(io_block_channels, io_block_channels[1:] + [model_channels]):
+            self.input_blocks.extend([
+                SparseResBlock3d(
+                    chs,
+                    model_channels,
+                    out_channels=chs,
+                )
+                for _ in range(num_io_res_blocks-1)
+            ])
+            self.input_blocks.append(
+                SparseResBlock3d(
+                    chs,
+                    model_channels,
+                    out_channels=next_chs,
+                    downsample=True,
+                )
+            )
+            
+        self.blocks = nn.ModuleList([
+            ModulatedSparseTransformerCrossBlock(
+                model_channels,
+                cond_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode='full',
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                share_mod=self.share_mod,
+                qk_rms_norm=self.qk_rms_norm,
+                qk_rms_norm_cross=self.qk_rms_norm_cross,
+            )
+            for _ in range(num_blocks)
+        ])
+
+        self.out_blocks = nn.ModuleList([])
+        for chs, prev_chs in zip(reversed(io_block_channels), [model_channels] + list(reversed(io_block_channels[1:]))):
+            self.out_blocks.append(
+                SparseResBlock3d(
+                    prev_chs * 2 if self.use_skip_connection else prev_chs,
+                    model_channels,
+                    out_channels=chs,
+                    upsample=True,
+                )
+            )
+            self.out_blocks.extend([
+                SparseResBlock3d(
+                    chs * 2 if self.use_skip_connection else chs,
+                    model_channels,
+                    out_channels=chs,
+                )
+                for _ in range(num_io_res_blocks-1)
+            ])
+        self.out_layer = sp.SparseLinear(io_block_channels[0], out_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.input_blocks.apply(convert_module_to_f16)
+        self.blocks.apply(convert_module_to_f16)
+        self.out_blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.input_blocks.apply(convert_module_to_f32)
+        self.blocks.apply(convert_module_to_f32)
+        self.out_blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+        # Initialize timestep embedding MLP:
+        nn.init.normal_(self.t_embedder.mlp[0].weight, std=0.02)
+        nn.init.normal_(self.t_embedder.mlp[2].weight, std=0.02)
+
+        # Zero-out adaLN modulation layers in DiT blocks:
+        if self.share_mod:
+            nn.init.constant_(self.adaLN_modulation[-1].weight, 0)
+            nn.init.constant_(self.adaLN_modulation[-1].bias, 0)
+        else:
+            for block in self.blocks:
+                nn.init.constant_(block.adaLN_modulation[-1].weight, 0)
+                nn.init.constant_(block.adaLN_modulation[-1].bias, 0)
+
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: sp.SparseTensor, t: torch.Tensor, cond: torch.Tensor) -> sp.SparseTensor:
+        h = self.input_layer(x).type(self.dtype)
+        t_emb = self.t_embedder(t)
+        if self.share_mod:
+            t_emb = self.adaLN_modulation(t_emb)
+        t_emb = t_emb.type(self.dtype)
+        cond = cond.type(self.dtype)
+
+        skips = []
+        # pack with input blocks
+        for block in self.input_blocks:
+            h = block(h, t_emb)
+            skips.append(h.feats)
+        
+        if self.pe_mode == "ape":
+            h = h + self.pos_embedder(h.coords[:, 1:]).type(self.dtype)
+        for block in self.blocks:
+            h = block(h, t_emb, cond)
+
+        # unpack with output blocks
+        for block, skip in zip(self.out_blocks, reversed(skips)):
+            if self.use_skip_connection:
+                h = block(h.replace(torch.cat([h.feats, skip], dim=1)), t_emb)
+            else:
+                h = block(h, t_emb)
+
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h.type(x.dtype))
+        return h

trellis/models/structured_latent_vae/__init__.py

@@ -0,0 +1,4 @@
+from .encoder import SLatEncoder
+from .decoder_gs import SLatGaussianDecoder
+from .decoder_rf import SLatRadianceFieldDecoder
+from .decoder_mesh import SLatMeshDecoder

trellis/models/structured_latent_vae/base.py

@@ -0,0 +1,117 @@
+from typing import *
+import torch
+import torch.nn as nn
+from ...modules.utils import convert_module_to_f16, convert_module_to_f32
+from ...modules import sparse as sp
+from ...modules.transformer import AbsolutePositionEmbedder
+from ...modules.sparse.transformer import SparseTransformerBlock
+
+
+def block_attn_config(self):
+    """
+    Return the attention configuration of the model.
+    """
+    for i in range(self.num_blocks):
+        if self.attn_mode == "shift_window":
+            yield "serialized", self.window_size, 0, (16 * (i % 2),) * 3, sp.SerializeMode.Z_ORDER
+        elif self.attn_mode == "shift_sequence":
+            yield "serialized", self.window_size, self.window_size // 2 * (i % 2), (0, 0, 0), sp.SerializeMode.Z_ORDER
+        elif self.attn_mode == "shift_order":
+            yield "serialized", self.window_size, 0, (0, 0, 0), sp.SerializeModes[i % 4]
+        elif self.attn_mode == "full":
+            yield "full", None, None, None, None
+        elif self.attn_mode == "swin":
+            yield "windowed", self.window_size, None, self.window_size // 2 * (i % 2), None
+
+
+class SparseTransformerBase(nn.Module):
+    """
+    Sparse Transformer without output layers.
+    Serve as the base class for encoder and decoder.
+    """
+    def __init__(
+        self,
+        in_channels: int,
+        model_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        window_size: Optional[int] = None,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        self.model_channels = model_channels
+        self.num_blocks = num_blocks
+        self.window_size = window_size
+        self.num_heads = num_heads or model_channels // num_head_channels
+        self.mlp_ratio = mlp_ratio
+        self.attn_mode = attn_mode
+        self.pe_mode = pe_mode
+        self.use_fp16 = use_fp16
+        self.use_checkpoint = use_checkpoint
+        self.qk_rms_norm = qk_rms_norm
+        self.dtype = torch.float16 if use_fp16 else torch.float32
+
+        if pe_mode == "ape":
+            self.pos_embedder = AbsolutePositionEmbedder(model_channels)
+
+        self.input_layer = sp.SparseLinear(in_channels, model_channels)
+        self.blocks = nn.ModuleList([
+            SparseTransformerBlock(
+                model_channels,
+                num_heads=self.num_heads,
+                mlp_ratio=self.mlp_ratio,
+                attn_mode=attn_mode,
+                window_size=window_size,
+                shift_sequence=shift_sequence,
+                shift_window=shift_window,
+                serialize_mode=serialize_mode,
+                use_checkpoint=self.use_checkpoint,
+                use_rope=(pe_mode == "rope"),
+                qk_rms_norm=self.qk_rms_norm,
+            )
+            for attn_mode, window_size, shift_sequence, shift_window, serialize_mode in block_attn_config(self)
+        ])
+
+    @property
+    def device(self) -> torch.device:
+        """
+        Return the device of the model.
+        """
+        return next(self.parameters()).device
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        self.blocks.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        self.blocks.apply(convert_module_to_f32)
+
+    def initialize_weights(self) -> None:
+        # Initialize transformer layers:
+        def _basic_init(module):
+            if isinstance(module, nn.Linear):
+                torch.nn.init.xavier_uniform_(module.weight)
+                if module.bias is not None:
+                    nn.init.constant_(module.bias, 0)
+        self.apply(_basic_init)
+
+    def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
+        h = self.input_layer(x)
+        if self.pe_mode == "ape":
+            h = h + self.pos_embedder(x.coords[:, 1:])
+        h = h.type(self.dtype)
+        for block in self.blocks:
+            h = block(h)
+        return h

trellis/models/structured_latent_vae/decoder_gs.py

@@ -0,0 +1,122 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ...modules import sparse as sp
+from ...utils.random_utils import hammersley_sequence
+from .base import SparseTransformerBase
+from ...representations import Gaussian
+
+
+class SLatGaussianDecoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+        representation_config: dict = None,
+    ):
+        super().__init__(
+            in_channels=latent_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.rep_config = representation_config
+        self._calc_layout()
+        self.out_layer = sp.SparseLinear(model_channels, self.out_channels)
+        self._build_perturbation()
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def _build_perturbation(self) -> None:
+        perturbation = [hammersley_sequence(3, i, self.rep_config['num_gaussians']) for i in range(self.rep_config['num_gaussians'])]
+        perturbation = torch.tensor(perturbation).float() * 2 - 1
+        perturbation = perturbation / self.rep_config['voxel_size']
+        perturbation = torch.atanh(perturbation).to(self.device)
+        self.register_buffer('offset_perturbation', perturbation)
+
+    def _calc_layout(self) -> None:
+        self.layout = {
+            '_xyz' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
+            '_features_dc' : {'shape': (self.rep_config['num_gaussians'], 1, 3), 'size': self.rep_config['num_gaussians'] * 3},
+            '_scaling' : {'shape': (self.rep_config['num_gaussians'], 3), 'size': self.rep_config['num_gaussians'] * 3},
+            '_rotation' : {'shape': (self.rep_config['num_gaussians'], 4), 'size': self.rep_config['num_gaussians'] * 4},
+            '_opacity' : {'shape': (self.rep_config['num_gaussians'], 1), 'size': self.rep_config['num_gaussians']},
+        }
+        start = 0
+        for k, v in self.layout.items():
+            v['range'] = (start, start + v['size'])
+            start += v['size']
+        self.out_channels = start
+    
+    def to_representation(self, x: sp.SparseTensor) -> List[Gaussian]:
+        """
+        Convert a batch of network outputs to 3D representations.
+
+        Args:
+            x: The [N x * x C] sparse tensor output by the network.
+
+        Returns:
+            list of representations
+        """
+        ret = []
+        for i in range(x.shape[0]):
+            representation = Gaussian(
+                sh_degree=0,
+                aabb=[-0.5, -0.5, -0.5, 1.0, 1.0, 1.0],
+                mininum_kernel_size = self.rep_config['3d_filter_kernel_size'],
+                scaling_bias = self.rep_config['scaling_bias'],
+                opacity_bias = self.rep_config['opacity_bias'],
+                scaling_activation = self.rep_config['scaling_activation']
+            )
+            xyz = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
+            for k, v in self.layout.items():
+                if k == '_xyz':
+                    offset = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape'])
+                    offset = offset * self.rep_config['lr'][k]
+                    if self.rep_config['perturb_offset']:
+                        offset = offset + self.offset_perturbation
+                    offset = torch.tanh(offset) / self.resolution * 0.5 * self.rep_config['voxel_size']
+                    _xyz = xyz.unsqueeze(1) + offset
+                    setattr(representation, k, _xyz.flatten(0, 1))
+                else:
+                    feats = x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']).flatten(0, 1)
+                    feats = feats * self.rep_config['lr'][k]
+                    setattr(representation, k, feats)
+            ret.append(representation)
+        return ret
+
+    def forward(self, x: sp.SparseTensor) -> List[Gaussian]:
+        h = super().forward(x)
+        h = h.type(x.dtype)
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h)
+        return self.to_representation(h)

trellis/models/structured_latent_vae/decoder_mesh.py

@@ -0,0 +1,167 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ...modules.utils import zero_module, convert_module_to_f16, convert_module_to_f32
+from ...modules import sparse as sp
+from .base import SparseTransformerBase
+from ...representations import MeshExtractResult
+from ...representations.mesh import SparseFeatures2Mesh
+
+
+class SparseSubdivideBlock3d(nn.Module):
+    """
+    A 3D subdivide block that can subdivide the sparse tensor.
+
+    Args:
+        channels: channels in the inputs and outputs.
+        out_channels: if specified, the number of output channels.
+        num_groups: the number of groups for the group norm.
+    """
+    def __init__(
+        self,
+        channels: int,
+        resolution: int,
+        out_channels: Optional[int] = None,
+        num_groups: int = 32
+    ):
+        super().__init__()
+        self.channels = channels
+        self.resolution = resolution
+        self.out_resolution = resolution * 2
+        self.out_channels = out_channels or channels
+
+        self.act_layers = nn.Sequential(
+            sp.SparseGroupNorm32(num_groups, channels),
+            sp.SparseSiLU()
+        )
+        
+        self.sub = sp.SparseSubdivide()
+        
+        self.out_layers = nn.Sequential(
+            sp.SparseConv3d(channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}"),
+            sp.SparseGroupNorm32(num_groups, self.out_channels),
+            sp.SparseSiLU(),
+            zero_module(sp.SparseConv3d(self.out_channels, self.out_channels, 3, indice_key=f"res_{self.out_resolution}")),
+        )
+        
+        if self.out_channels == channels:
+            self.skip_connection = nn.Identity()
+        else:
+            self.skip_connection = sp.SparseConv3d(channels, self.out_channels, 1, indice_key=f"res_{self.out_resolution}")
+        
+    def forward(self, x: sp.SparseTensor) -> sp.SparseTensor:
+        """
+        Apply the block to a Tensor, conditioned on a timestep embedding.
+
+        Args:
+            x: an [N x C x ...] Tensor of features.
+        Returns:
+            an [N x C x ...] Tensor of outputs.
+        """
+        h = self.act_layers(x)
+        h = self.sub(h)
+        x = self.sub(x)
+        h = self.out_layers(h)
+        h = h + self.skip_connection(x)
+        return h
+
+
+class SLatMeshDecoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+        representation_config: dict = None,
+    ):
+        super().__init__(
+            in_channels=latent_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.rep_config = representation_config
+        self.mesh_extractor = SparseFeatures2Mesh(res=self.resolution*4, use_color=self.rep_config.get('use_color', False))
+        self.out_channels = self.mesh_extractor.feats_channels
+        self.upsample = nn.ModuleList([
+            SparseSubdivideBlock3d(
+                channels=model_channels,
+                resolution=resolution,
+                out_channels=model_channels // 4
+            ),
+            SparseSubdivideBlock3d(
+                channels=model_channels // 4,
+                resolution=resolution * 2,
+                out_channels=model_channels // 8
+            )
+        ])
+        self.out_layer = sp.SparseLinear(model_channels // 8, self.out_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def convert_to_fp16(self) -> None:
+        """
+        Convert the torso of the model to float16.
+        """
+        super().convert_to_fp16()
+        self.upsample.apply(convert_module_to_f16)
+
+    def convert_to_fp32(self) -> None:
+        """
+        Convert the torso of the model to float32.
+        """
+        super().convert_to_fp32()
+        self.upsample.apply(convert_module_to_f32)  
+    
+    def to_representation(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
+        """
+        Convert a batch of network outputs to 3D representations.
+
+        Args:
+            x: The [N x * x C] sparse tensor output by the network.
+
+        Returns:
+            list of representations
+        """
+        ret = []
+        for i in range(x.shape[0]):
+            mesh = self.mesh_extractor(x[i], training=self.training)
+            ret.append(mesh)
+        return ret
+
+    def forward(self, x: sp.SparseTensor) -> List[MeshExtractResult]:
+        h = super().forward(x)
+        for block in self.upsample:
+            h = block(h)
+        h = h.type(x.dtype)
+        h = self.out_layer(h)
+        return self.to_representation(h)

trellis/models/structured_latent_vae/decoder_rf.py

@@ -0,0 +1,104 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+import numpy as np
+from ...modules import sparse as sp
+from .base import SparseTransformerBase
+from ...representations import Strivec
+
+
+class SLatRadianceFieldDecoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+        representation_config: dict = None,
+    ):
+        super().__init__(
+            in_channels=latent_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.rep_config = representation_config
+        self._calc_layout()
+        self.out_layer = sp.SparseLinear(model_channels, self.out_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def _calc_layout(self) -> None:
+        self.layout = {
+            'trivec': {'shape': (self.rep_config['rank'], 3, self.rep_config['dim']), 'size': self.rep_config['rank'] * 3 * self.rep_config['dim']},
+            'density': {'shape': (self.rep_config['rank'],), 'size': self.rep_config['rank']},
+            'features_dc': {'shape': (self.rep_config['rank'], 1, 3), 'size': self.rep_config['rank'] * 3},
+        }
+        start = 0
+        for k, v in self.layout.items():
+            v['range'] = (start, start + v['size'])
+            start += v['size']
+        self.out_channels = start    
+    
+    def to_representation(self, x: sp.SparseTensor) -> List[Strivec]:
+        """
+        Convert a batch of network outputs to 3D representations.
+
+        Args:
+            x: The [N x * x C] sparse tensor output by the network.
+
+        Returns:
+            list of representations
+        """
+        ret = []
+        for i in range(x.shape[0]):
+            representation = Strivec(
+                sh_degree=0,
+                resolution=self.resolution,
+                aabb=[-0.5, -0.5, -0.5, 1, 1, 1],
+                rank=self.rep_config['rank'],
+                dim=self.rep_config['dim'],
+                device='cuda',
+            )
+            representation.density_shift = 0.0
+            representation.position = (x.coords[x.layout[i]][:, 1:].float() + 0.5) / self.resolution
+            representation.depth = torch.full((representation.position.shape[0], 1), int(np.log2(self.resolution)), dtype=torch.uint8, device='cuda')
+            for k, v in self.layout.items():
+                setattr(representation, k, x.feats[x.layout[i]][:, v['range'][0]:v['range'][1]].reshape(-1, *v['shape']))
+            representation.trivec = representation.trivec + 1
+            ret.append(representation)
+        return ret
+
+    def forward(self, x: sp.SparseTensor) -> List[Strivec]:
+        h = super().forward(x)
+        h = h.type(x.dtype)
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h)
+        return self.to_representation(h)

trellis/models/structured_latent_vae/encoder.py

@@ -0,0 +1,72 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from ...modules import sparse as sp
+from .base import SparseTransformerBase
+
+
+class SLatEncoder(SparseTransformerBase):
+    def __init__(
+        self,
+        resolution: int,
+        in_channels: int,
+        model_channels: int,
+        latent_channels: int,
+        num_blocks: int,
+        num_heads: Optional[int] = None,
+        num_head_channels: Optional[int] = 64,
+        mlp_ratio: float = 4,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "swin",
+        window_size: int = 8,
+        pe_mode: Literal["ape", "rope"] = "ape",
+        use_fp16: bool = False,
+        use_checkpoint: bool = False,
+        qk_rms_norm: bool = False,
+    ):
+        super().__init__(
+            in_channels=in_channels,
+            model_channels=model_channels,
+            num_blocks=num_blocks,
+            num_heads=num_heads,
+            num_head_channels=num_head_channels,
+            mlp_ratio=mlp_ratio,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            pe_mode=pe_mode,
+            use_fp16=use_fp16,
+            use_checkpoint=use_checkpoint,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.resolution = resolution
+        self.out_layer = sp.SparseLinear(model_channels, 2 * latent_channels)
+
+        self.initialize_weights()
+        if use_fp16:
+            self.convert_to_fp16()
+
+    def initialize_weights(self) -> None:
+        super().initialize_weights()
+        # Zero-out output layers:
+        nn.init.constant_(self.out_layer.weight, 0)
+        nn.init.constant_(self.out_layer.bias, 0)
+
+    def forward(self, x: sp.SparseTensor, sample_posterior=True, return_raw=False):
+        h = super().forward(x)
+        h = h.type(x.dtype)
+        h = h.replace(F.layer_norm(h.feats, h.feats.shape[-1:]))
+        h = self.out_layer(h)
+        
+        # Sample from the posterior distribution
+        mean, logvar = h.feats.chunk(2, dim=-1)
+        if sample_posterior:
+            std = torch.exp(0.5 * logvar)
+            z = mean + std * torch.randn_like(std)
+        else:
+            z = mean
+        z = h.replace(z)
+            
+        if return_raw:
+            return z, mean, logvar
+        else:
+            return z

trellis/modules/attention/__init__.py

@@ -0,0 +1,36 @@
+from typing import *
+
+BACKEND = 'flash_attn' 
+DEBUG = False
+
+def __from_env():
+    import os
+    
+    global BACKEND
+    global DEBUG
+    
+    env_attn_backend = os.environ.get('ATTN_BACKEND')
+    env_sttn_debug = os.environ.get('ATTN_DEBUG')
+    
+    if env_attn_backend is not None and env_attn_backend in ['xformers', 'flash_attn', 'sdpa', 'naive']:
+        BACKEND = env_attn_backend
+    if env_sttn_debug is not None:
+        DEBUG = env_sttn_debug == '1'
+
+    print(f"[ATTENTION] Using backend: {BACKEND}")
+        
+
+__from_env()
+    
+
+def set_backend(backend: Literal['xformers', 'flash_attn']):
+    global BACKEND
+    BACKEND = backend
+
+def set_debug(debug: bool):
+    global DEBUG
+    DEBUG = debug
+
+
+from .full_attn import *
+from .modules import *

trellis/modules/attention/full_attn.py

@@ -0,0 +1,140 @@
+from typing import *
+import torch
+import math
+from . import DEBUG, BACKEND
+
+if BACKEND == 'xformers':
+    import xformers.ops as xops
+elif BACKEND == 'flash_attn':
+    import flash_attn
+elif BACKEND == 'sdpa':
+    from torch.nn.functional import scaled_dot_product_attention as sdpa
+elif BACKEND == 'naive':
+    pass
+else:
+    raise ValueError(f"Unknown attention backend: {BACKEND}")
+
+
+__all__ = [
+    'scaled_dot_product_attention',
+]
+
+
+def _naive_sdpa(q, k, v):
+    """
+    Naive implementation of scaled dot product attention.
+    """
+    q = q.permute(0, 2, 1, 3)   # [N, H, L, C]
+    k = k.permute(0, 2, 1, 3)   # [N, H, L, C]
+    v = v.permute(0, 2, 1, 3)   # [N, H, L, C]
+    scale_factor = 1 / math.sqrt(q.size(-1))
+    attn_weight = q @ k.transpose(-2, -1) * scale_factor
+    attn_weight = torch.softmax(attn_weight, dim=-1)
+    out = attn_weight @ v
+    out = out.permute(0, 2, 1, 3)   # [N, L, H, C]
+    return out
+
+
+@overload
+def scaled_dot_product_attention(qkv: torch.Tensor) -> torch.Tensor:
+    """
+    Apply scaled dot product attention.
+
+    Args:
+        qkv (torch.Tensor): A [N, L, 3, H, C] tensor containing Qs, Ks, and Vs.
+    """
+    ...
+
+@overload
+def scaled_dot_product_attention(q: torch.Tensor, kv: torch.Tensor) -> torch.Tensor:
+    """
+    Apply scaled dot product attention.
+
+    Args:
+        q (torch.Tensor): A [N, L, H, C] tensor containing Qs.
+        kv (torch.Tensor): A [N, L, 2, H, C] tensor containing Ks and Vs.
+    """
+    ...
+
+@overload
+def scaled_dot_product_attention(q: torch.Tensor, k: torch.Tensor, v: torch.Tensor) -> torch.Tensor:
+    """
+    Apply scaled dot product attention.
+
+    Args:
+        q (torch.Tensor): A [N, L, H, Ci] tensor containing Qs.
+        k (torch.Tensor): A [N, L, H, Ci] tensor containing Ks.
+        v (torch.Tensor): A [N, L, H, Co] tensor containing Vs.
+
+    Note:
+        k and v are assumed to have the same coordinate map.
+    """
+    ...
+
+def scaled_dot_product_attention(*args, **kwargs):
+    arg_names_dict = {
+        1: ['qkv'],
+        2: ['q', 'kv'],
+        3: ['q', 'k', 'v']
+    }
+    num_all_args = len(args) + len(kwargs)
+    assert num_all_args in arg_names_dict, f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
+    for key in arg_names_dict[num_all_args][len(args):]:
+        assert key in kwargs, f"Missing argument {key}"
+
+    if num_all_args == 1:
+        qkv = args[0] if len(args) > 0 else kwargs['qkv']
+        assert len(qkv.shape) == 5 and qkv.shape[2] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, L, 3, H, C]"
+        device = qkv.device
+
+    elif num_all_args == 2:
+        q = args[0] if len(args) > 0 else kwargs['q']
+        kv = args[1] if len(args) > 1 else kwargs['kv']
+        assert q.shape[0] == kv.shape[0], f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
+        assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
+        assert len(kv.shape) == 5, f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
+        device = q.device
+
+    elif num_all_args == 3:
+        q = args[0] if len(args) > 0 else kwargs['q']
+        k = args[1] if len(args) > 1 else kwargs['k']
+        v = args[2] if len(args) > 2 else kwargs['v']
+        assert q.shape[0] == k.shape[0] == v.shape[0], f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
+        assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
+        assert len(k.shape) == 4, f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
+        assert len(v.shape) == 4, f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
+        device = q.device    
+
+    if BACKEND == 'xformers':
+        if num_all_args == 1:
+            q, k, v = qkv.unbind(dim=2)
+        elif num_all_args == 2:
+            k, v = kv.unbind(dim=2)
+        out = xops.memory_efficient_attention(q, k, v)
+    elif BACKEND == 'flash_attn':
+        if num_all_args == 1:
+            out = flash_attn.flash_attn_qkvpacked_func(qkv)
+        elif num_all_args == 2:
+            out = flash_attn.flash_attn_kvpacked_func(q, kv)
+        elif num_all_args == 3:
+            out = flash_attn.flash_attn_func(q, k, v)
+    elif BACKEND == 'sdpa':
+        if num_all_args == 1:
+            q, k, v = qkv.unbind(dim=2)
+        elif num_all_args == 2:
+            k, v = kv.unbind(dim=2)
+        q = q.permute(0, 2, 1, 3)   # [N, H, L, C]
+        k = k.permute(0, 2, 1, 3)   # [N, H, L, C]
+        v = v.permute(0, 2, 1, 3)   # [N, H, L, C]
+        out = sdpa(q, k, v)         # [N, H, L, C]
+        out = out.permute(0, 2, 1, 3)   # [N, L, H, C]
+    elif BACKEND == 'naive':
+        if num_all_args == 1:
+            q, k, v = qkv.unbind(dim=2)
+        elif num_all_args == 2:
+            k, v = kv.unbind(dim=2)
+        out = _naive_sdpa(q, k, v)
+    else:
+        raise ValueError(f"Unknown attention module: {BACKEND}")
+    
+    return out

trellis/modules/attention/modules.py

@@ -0,0 +1,146 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .full_attn import scaled_dot_product_attention
+
+
+class MultiHeadRMSNorm(nn.Module):
+    def __init__(self, dim: int, heads: int):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(heads, dim))
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return (F.normalize(x.float(), dim = -1) * self.gamma * self.scale).to(x.dtype)
+
+
+class RotaryPositionEmbedder(nn.Module):
+    def __init__(self, hidden_size: int, in_channels: int = 3):
+        super().__init__()
+        assert hidden_size % 2 == 0, "Hidden size must be divisible by 2"
+        self.hidden_size = hidden_size
+        self.in_channels = in_channels
+        self.freq_dim = hidden_size // in_channels // 2
+        self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
+        self.freqs = 1.0 / (10000 ** self.freqs)
+        
+    def _get_phases(self, indices: torch.Tensor) -> torch.Tensor:
+        self.freqs = self.freqs.to(indices.device)
+        phases = torch.outer(indices, self.freqs)
+        phases = torch.polar(torch.ones_like(phases), phases)
+        return phases
+        
+    def _rotary_embedding(self, x: torch.Tensor, phases: torch.Tensor) -> torch.Tensor:
+        x_complex = torch.view_as_complex(x.float().reshape(*x.shape[:-1], -1, 2))
+        x_rotated = x_complex * phases
+        x_embed = torch.view_as_real(x_rotated).reshape(*x_rotated.shape[:-1], -1).to(x.dtype)
+        return x_embed
+        
+    def forward(self, q: torch.Tensor, k: torch.Tensor, indices: Optional[torch.Tensor] = None) -> Tuple[torch.Tensor, torch.Tensor]:
+        """
+        Args:
+            q (sp.SparseTensor): [..., N, D] tensor of queries
+            k (sp.SparseTensor): [..., N, D] tensor of keys
+            indices (torch.Tensor): [..., N, C] tensor of spatial positions
+        """
+        if indices is None:
+            indices = torch.arange(q.shape[-2], device=q.device)
+            if len(q.shape) > 2:
+                indices = indices.unsqueeze(0).expand(q.shape[:-2] + (-1,))
+        
+        phases = self._get_phases(indices.reshape(-1)).reshape(*indices.shape[:-1], -1)
+        if phases.shape[1] < self.hidden_size // 2:
+            phases = torch.cat([phases, torch.polar(
+                torch.ones(*phases.shape[:-1], self.hidden_size // 2 - phases.shape[1], device=phases.device),
+                torch.zeros(*phases.shape[:-1], self.hidden_size // 2 - phases.shape[1], device=phases.device)
+            )], dim=-1)
+        q_embed = self._rotary_embedding(q, phases)
+        k_embed = self._rotary_embedding(k, phases)
+        return q_embed, k_embed
+    
+
+class MultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        num_heads: int,
+        ctx_channels: Optional[int]=None,
+        type: Literal["self", "cross"] = "self",
+        attn_mode: Literal["full", "windowed"] = "full",
+        window_size: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        qkv_bias: bool = True,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+    ):
+        super().__init__()
+        assert channels % num_heads == 0
+        assert type in ["self", "cross"], f"Invalid attention type: {type}"
+        assert attn_mode in ["full", "windowed"], f"Invalid attention mode: {attn_mode}"
+        assert type == "self" or attn_mode == "full", "Cross-attention only supports full attention"
+        
+        if attn_mode == "windowed":
+            raise NotImplementedError("Windowed attention is not yet implemented")
+        
+        self.channels = channels
+        self.head_dim = channels // num_heads
+        self.ctx_channels = ctx_channels if ctx_channels is not None else channels
+        self.num_heads = num_heads
+        self._type = type
+        self.attn_mode = attn_mode
+        self.window_size = window_size
+        self.shift_window = shift_window
+        self.use_rope = use_rope
+        self.qk_rms_norm = qk_rms_norm
+
+        if self._type == "self":
+            self.to_qkv = nn.Linear(channels, channels * 3, bias=qkv_bias)
+        else:
+            self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
+            self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
+            
+        if self.qk_rms_norm:
+            self.q_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
+            self.k_rms_norm = MultiHeadRMSNorm(self.head_dim, num_heads)
+            
+        self.to_out = nn.Linear(channels, channels)
+
+        if use_rope:
+            self.rope = RotaryPositionEmbedder(channels)
+    
+    def forward(self, x: torch.Tensor, context: Optional[torch.Tensor] = None, indices: Optional[torch.Tensor] = None) -> torch.Tensor:
+        B, L, C = x.shape
+        if self._type == "self":
+            qkv = self.to_qkv(x)
+            qkv = qkv.reshape(B, L, 3, self.num_heads, -1)
+            if self.use_rope:
+                q, k, v = qkv.unbind(dim=2)
+                q, k = self.rope(q, k, indices)
+                qkv = torch.stack([q, k, v], dim=2)
+            if self.attn_mode == "full":
+                if self.qk_rms_norm:
+                    q, k, v = qkv.unbind(dim=2)
+                    q = self.q_rms_norm(q)
+                    k = self.k_rms_norm(k)
+                    h = scaled_dot_product_attention(q, k, v)
+                else:
+                    h = scaled_dot_product_attention(qkv)
+            elif self.attn_mode == "windowed":
+                raise NotImplementedError("Windowed attention is not yet implemented")
+        else:
+            Lkv = context.shape[1]
+            q = self.to_q(x)
+            kv = self.to_kv(context)
+            q = q.reshape(B, L, self.num_heads, -1)
+            kv = kv.reshape(B, Lkv, 2, self.num_heads, -1)
+            if self.qk_rms_norm:
+                q = self.q_rms_norm(q)
+                k, v = kv.unbind(dim=2)
+                k = self.k_rms_norm(k)
+                h = scaled_dot_product_attention(q, k, v)
+            else:
+                h = scaled_dot_product_attention(q, kv)
+        h = h.reshape(B, L, -1)
+        h = self.to_out(h)
+        return h

trellis/modules/norm.py

@@ -0,0 +1,25 @@
+import torch
+import torch.nn as nn
+
+
+class LayerNorm32(nn.LayerNorm):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return super().forward(x.float()).type(x.dtype)
+    
+
+class GroupNorm32(nn.GroupNorm):
+    """
+    A GroupNorm layer that converts to float32 before the forward pass.
+    """
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return super().forward(x.float()).type(x.dtype)
+    
+    
+class ChannelLayerNorm32(LayerNorm32):
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        DIM = x.dim()
+        x = x.permute(0, *range(2, DIM), 1).contiguous()
+        x = super().forward(x)
+        x = x.permute(0, DIM-1, *range(1, DIM-1)).contiguous()
+        return x
+    

trellis/modules/sparse/__init__.py

@@ -0,0 +1,102 @@
+from typing import *
+
+BACKEND = 'spconv' 
+DEBUG = False
+ATTN = 'flash_attn'
+
+def __from_env():
+    import os
+    
+    global BACKEND
+    global DEBUG
+    global ATTN
+    
+    env_sparse_backend = os.environ.get('SPARSE_BACKEND')
+    env_sparse_debug = os.environ.get('SPARSE_DEBUG')
+    env_sparse_attn = os.environ.get('SPARSE_ATTN_BACKEND')
+    if env_sparse_attn is None:
+        env_sparse_attn = os.environ.get('ATTN_BACKEND')
+
+    if env_sparse_backend is not None and env_sparse_backend in ['spconv', 'torchsparse']:
+        BACKEND = env_sparse_backend
+    if env_sparse_debug is not None:
+        DEBUG = env_sparse_debug == '1'
+    if env_sparse_attn is not None and env_sparse_attn in ['xformers', 'flash_attn']:
+        ATTN = env_sparse_attn
+        
+    print(f"[SPARSE] Backend: {BACKEND}, Attention: {ATTN}")
+        
+
+__from_env()
+    
+
+def set_backend(backend: Literal['spconv', 'torchsparse']):
+    global BACKEND
+    BACKEND = backend
+
+def set_debug(debug: bool):
+    global DEBUG
+    DEBUG = debug
+
+def set_attn(attn: Literal['xformers', 'flash_attn']):
+    global ATTN
+    ATTN = attn
+    
+    
+import importlib
+
+__attributes = {
+    'SparseTensor': 'basic',
+    'sparse_batch_broadcast': 'basic',
+    'sparse_batch_op': 'basic',
+    'sparse_cat': 'basic',
+    'sparse_unbind': 'basic',
+    'SparseGroupNorm': 'norm',
+    'SparseLayerNorm': 'norm',
+    'SparseGroupNorm32': 'norm',
+    'SparseLayerNorm32': 'norm',
+    'SparseReLU': 'nonlinearity',
+    'SparseSiLU': 'nonlinearity',
+    'SparseGELU': 'nonlinearity',
+    'SparseActivation': 'nonlinearity',
+    'SparseLinear': 'linear',
+    'sparse_scaled_dot_product_attention': 'attention',
+    'SerializeMode': 'attention',
+    'sparse_serialized_scaled_dot_product_self_attention': 'attention',
+    'sparse_windowed_scaled_dot_product_self_attention': 'attention',
+    'SparseMultiHeadAttention': 'attention',
+    'SparseConv3d': 'conv',
+    'SparseInverseConv3d': 'conv',
+    'SparseDownsample': 'spatial',
+    'SparseUpsample': 'spatial',
+    'SparseSubdivide' : 'spatial'
+}
+
+__submodules = ['transformer']
+
+__all__ = list(__attributes.keys()) + __submodules
+
+def __getattr__(name):
+    if name not in globals():
+        if name in __attributes:
+            module_name = __attributes[name]
+            module = importlib.import_module(f".{module_name}", __name__)
+            globals()[name] = getattr(module, name)
+        elif name in __submodules:
+            module = importlib.import_module(f".{name}", __name__)
+            globals()[name] = module
+        else:
+            raise AttributeError(f"module {__name__} has no attribute {name}")
+    return globals()[name]
+
+
+# For Pylance
+if __name__ == '__main__':
+    from .basic import *
+    from .norm import *
+    from .nonlinearity import *
+    from .linear import *
+    from .attention import *
+    from .conv import *
+    from .spatial import *
+    import transformer

trellis/modules/sparse/attention/__init__.py

@@ -0,0 +1,4 @@
+from .full_attn import *
+from .serialized_attn import *
+from .windowed_attn import *
+from .modules import *

trellis/modules/sparse/attention/full_attn.py

@@ -0,0 +1,215 @@
+from typing import *
+import torch
+from .. import SparseTensor
+from .. import DEBUG, ATTN
+
+if ATTN == 'xformers':
+    import xformers.ops as xops
+elif ATTN == 'flash_attn':
+    import flash_attn
+else:
+    raise ValueError(f"Unknown attention module: {ATTN}")
+
+
+__all__ = [
+    'sparse_scaled_dot_product_attention',
+]
+
+
+@overload
+def sparse_scaled_dot_product_attention(qkv: SparseTensor) -> SparseTensor:
+    """
+    Apply scaled dot product attention to a sparse tensor.
+
+    Args:
+        qkv (SparseTensor): A [N, *, 3, H, C] sparse tensor containing Qs, Ks, and Vs.
+    """
+    ...
+
+@overload
+def sparse_scaled_dot_product_attention(q: SparseTensor, kv: Union[SparseTensor, torch.Tensor]) -> SparseTensor:
+    """
+    Apply scaled dot product attention to a sparse tensor.
+
+    Args:
+        q (SparseTensor): A [N, *, H, C] sparse tensor containing Qs.
+        kv (SparseTensor or torch.Tensor): A [N, *, 2, H, C] sparse tensor or a [N, L, 2, H, C] dense tensor containing Ks and Vs.
+    """
+    ...
+
+@overload
+def sparse_scaled_dot_product_attention(q: torch.Tensor, kv: SparseTensor) -> torch.Tensor:
+    """
+    Apply scaled dot product attention to a sparse tensor.
+
+    Args:
+        q (SparseTensor): A [N, L, H, C] dense tensor containing Qs.
+        kv (SparseTensor or torch.Tensor): A [N, *, 2, H, C] sparse tensor containing Ks and Vs.
+    """
+    ...
+
+@overload
+def sparse_scaled_dot_product_attention(q: SparseTensor, k: SparseTensor, v: SparseTensor) -> SparseTensor:
+    """
+    Apply scaled dot product attention to a sparse tensor.
+
+    Args:
+        q (SparseTensor): A [N, *, H, Ci] sparse tensor containing Qs.
+        k (SparseTensor): A [N, *, H, Ci] sparse tensor containing Ks.
+        v (SparseTensor): A [N, *, H, Co] sparse tensor containing Vs.
+
+    Note:
+        k and v are assumed to have the same coordinate map.
+    """
+    ...
+
+@overload
+def sparse_scaled_dot_product_attention(q: SparseTensor, k: torch.Tensor, v: torch.Tensor) -> SparseTensor:
+    """
+    Apply scaled dot product attention to a sparse tensor.
+
+    Args:
+        q (SparseTensor): A [N, *, H, Ci] sparse tensor containing Qs.
+        k (torch.Tensor): A [N, L, H, Ci] dense tensor containing Ks.
+        v (torch.Tensor): A [N, L, H, Co] dense tensor containing Vs.
+    """
+    ...
+
+@overload
+def sparse_scaled_dot_product_attention(q: torch.Tensor, k: SparseTensor, v: SparseTensor) -> torch.Tensor:
+    """
+    Apply scaled dot product attention to a sparse tensor.
+
+    Args:
+        q (torch.Tensor): A [N, L, H, Ci] dense tensor containing Qs.
+        k (SparseTensor): A [N, *, H, Ci] sparse tensor containing Ks.
+        v (SparseTensor): A [N, *, H, Co] sparse tensor containing Vs.
+    """
+    ...
+
+def sparse_scaled_dot_product_attention(*args, **kwargs):
+    arg_names_dict = {
+        1: ['qkv'],
+        2: ['q', 'kv'],
+        3: ['q', 'k', 'v']
+    }
+    num_all_args = len(args) + len(kwargs)
+    assert num_all_args in arg_names_dict, f"Invalid number of arguments, got {num_all_args}, expected 1, 2, or 3"
+    for key in arg_names_dict[num_all_args][len(args):]:
+        assert key in kwargs, f"Missing argument {key}"
+
+    if num_all_args == 1:
+        qkv = args[0] if len(args) > 0 else kwargs['qkv']
+        assert isinstance(qkv, SparseTensor), f"qkv must be a SparseTensor, got {type(qkv)}"
+        assert len(qkv.shape) == 4 and qkv.shape[1] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
+        device = qkv.device
+
+        s = qkv
+        q_seqlen = [qkv.layout[i].stop - qkv.layout[i].start for i in range(qkv.shape[0])]
+        kv_seqlen = q_seqlen
+        qkv = qkv.feats     # [T, 3, H, C]
+
+    elif num_all_args == 2:
+        q = args[0] if len(args) > 0 else kwargs['q']
+        kv = args[1] if len(args) > 1 else kwargs['kv']
+        assert isinstance(q, SparseTensor) and isinstance(kv, (SparseTensor, torch.Tensor)) or \
+               isinstance(q, torch.Tensor) and isinstance(kv, SparseTensor), \
+               f"Invalid types, got {type(q)} and {type(kv)}"
+        assert q.shape[0] == kv.shape[0], f"Batch size mismatch, got {q.shape[0]} and {kv.shape[0]}"
+        device = q.device
+
+        if isinstance(q, SparseTensor):
+            assert len(q.shape) == 3, f"Invalid shape for q, got {q.shape}, expected [N, *, H, C]"
+            s = q
+            q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
+            q = q.feats     # [T_Q, H, C]
+        else:
+            assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, C]"
+            s = None
+            N, L, H, C = q.shape
+            q_seqlen = [L] * N
+            q = q.reshape(N * L, H, C)   # [T_Q, H, C]
+
+        if isinstance(kv, SparseTensor):
+            assert len(kv.shape) == 4 and kv.shape[1] == 2, f"Invalid shape for kv, got {kv.shape}, expected [N, *, 2, H, C]"
+            kv_seqlen = [kv.layout[i].stop - kv.layout[i].start for i in range(kv.shape[0])]
+            kv = kv.feats     # [T_KV, 2, H, C]
+        else:
+            assert len(kv.shape) == 5, f"Invalid shape for kv, got {kv.shape}, expected [N, L, 2, H, C]"
+            N, L, _, H, C = kv.shape
+            kv_seqlen = [L] * N
+            kv = kv.reshape(N * L, 2, H, C)   # [T_KV, 2, H, C]
+
+    elif num_all_args == 3:
+        q = args[0] if len(args) > 0 else kwargs['q']
+        k = args[1] if len(args) > 1 else kwargs['k']
+        v = args[2] if len(args) > 2 else kwargs['v']
+        assert isinstance(q, SparseTensor) and isinstance(k, (SparseTensor, torch.Tensor)) and type(k) == type(v) or \
+               isinstance(q, torch.Tensor) and isinstance(k, SparseTensor) and isinstance(v, SparseTensor), \
+               f"Invalid types, got {type(q)}, {type(k)}, and {type(v)}"
+        assert q.shape[0] == k.shape[0] == v.shape[0], f"Batch size mismatch, got {q.shape[0]}, {k.shape[0]}, and {v.shape[0]}"
+        device = q.device
+
+        if isinstance(q, SparseTensor):
+            assert len(q.shape) == 3, f"Invalid shape for q, got {q.shape}, expected [N, *, H, Ci]"
+            s = q
+            q_seqlen = [q.layout[i].stop - q.layout[i].start for i in range(q.shape[0])]
+            q = q.feats     # [T_Q, H, Ci]
+        else:
+            assert len(q.shape) == 4, f"Invalid shape for q, got {q.shape}, expected [N, L, H, Ci]"
+            s = None
+            N, L, H, CI = q.shape
+            q_seqlen = [L] * N
+            q = q.reshape(N * L, H, CI)  # [T_Q, H, Ci]
+
+        if isinstance(k, SparseTensor):
+            assert len(k.shape) == 3, f"Invalid shape for k, got {k.shape}, expected [N, *, H, Ci]"
+            assert len(v.shape) == 3, f"Invalid shape for v, got {v.shape}, expected [N, *, H, Co]"
+            kv_seqlen = [k.layout[i].stop - k.layout[i].start for i in range(k.shape[0])]
+            k = k.feats     # [T_KV, H, Ci]
+            v = v.feats     # [T_KV, H, Co]
+        else:
+            assert len(k.shape) == 4, f"Invalid shape for k, got {k.shape}, expected [N, L, H, Ci]"
+            assert len(v.shape) == 4, f"Invalid shape for v, got {v.shape}, expected [N, L, H, Co]"
+            N, L, H, CI, CO = *k.shape, v.shape[-1]
+            kv_seqlen = [L] * N
+            k = k.reshape(N * L, H, CI)     # [T_KV, H, Ci]
+            v = v.reshape(N * L, H, CO)     # [T_KV, H, Co]
+
+    if DEBUG:
+        if s is not None:
+            for i in range(s.shape[0]):
+                assert (s.coords[s.layout[i]] == i).all(), f"SparseScaledDotProductSelfAttention: batch index mismatch"
+        if num_all_args in [2, 3]:
+            assert q.shape[:2] == [1, sum(q_seqlen)], f"SparseScaledDotProductSelfAttention: q shape mismatch"
+        if num_all_args == 3:
+            assert k.shape[:2] == [1, sum(kv_seqlen)], f"SparseScaledDotProductSelfAttention: k shape mismatch"
+            assert v.shape[:2] == [1, sum(kv_seqlen)], f"SparseScaledDotProductSelfAttention: v shape mismatch"
+
+    if ATTN == 'xformers':
+        if num_all_args == 1:
+            q, k, v = qkv.unbind(dim=1)
+        elif num_all_args == 2:
+            k, v = kv.unbind(dim=1)
+        q = q.unsqueeze(0)
+        k = k.unsqueeze(0)
+        v = v.unsqueeze(0)
+        mask = xops.fmha.BlockDiagonalMask.from_seqlens(q_seqlen, kv_seqlen)
+        out = xops.memory_efficient_attention(q, k, v, mask)[0]
+    elif ATTN == 'flash_attn':
+        cu_seqlens_q = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(q_seqlen), dim=0)]).int().to(device)
+        if num_all_args in [2, 3]:
+            cu_seqlens_kv = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(kv_seqlen), dim=0)]).int().to(device)
+        if num_all_args == 1:
+            out = flash_attn.flash_attn_varlen_qkvpacked_func(qkv, cu_seqlens_q, max(q_seqlen))
+        elif num_all_args == 2:
+            out = flash_attn.flash_attn_varlen_kvpacked_func(q, kv, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen))
+        elif num_all_args == 3:
+            out = flash_attn.flash_attn_varlen_func(q, k, v, cu_seqlens_q, cu_seqlens_kv, max(q_seqlen), max(kv_seqlen))
+    else:
+        raise ValueError(f"Unknown attention module: {ATTN}")
+    
+    if s is not None:
+        return s.replace(out)
+    else:
+        return out.reshape(N, L, H, -1)

trellis/modules/sparse/attention/modules.py

@@ -0,0 +1,139 @@
+from typing import *
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+from .. import SparseTensor
+from .full_attn import sparse_scaled_dot_product_attention
+from .serialized_attn import SerializeMode, sparse_serialized_scaled_dot_product_self_attention
+from .windowed_attn import sparse_windowed_scaled_dot_product_self_attention
+from ...attention import RotaryPositionEmbedder
+
+
+class SparseMultiHeadRMSNorm(nn.Module):
+    def __init__(self, dim: int, heads: int):
+        super().__init__()
+        self.scale = dim ** 0.5
+        self.gamma = nn.Parameter(torch.ones(heads, dim))
+
+    def forward(self, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
+        x_type = x.dtype
+        x = x.float()
+        if isinstance(x, SparseTensor):
+            x = x.replace(F.normalize(x.feats, dim=-1))
+        else:
+            x = F.normalize(x, dim=-1)            
+        return (x * self.gamma * self.scale).to(x_type)
+
+
+class SparseMultiHeadAttention(nn.Module):
+    def __init__(
+        self,
+        channels: int,
+        num_heads: int,
+        ctx_channels: Optional[int] = None,
+        type: Literal["self", "cross"] = "self",
+        attn_mode: Literal["full", "serialized", "windowed"] = "full",
+        window_size: Optional[int] = None,
+        shift_sequence: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        serialize_mode: Optional[SerializeMode] = None,
+        qkv_bias: bool = True,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+    ):
+        super().__init__()
+        assert channels % num_heads == 0
+        assert type in ["self", "cross"], f"Invalid attention type: {type}"
+        assert attn_mode in ["full", "serialized", "windowed"], f"Invalid attention mode: {attn_mode}"
+        assert type == "self" or attn_mode == "full", "Cross-attention only supports full attention"
+        assert type == "self" or use_rope is False, "Rotary position embeddings only supported for self-attention"
+        self.channels = channels
+        self.ctx_channels = ctx_channels if ctx_channels is not None else channels
+        self.num_heads = num_heads
+        self._type = type
+        self.attn_mode = attn_mode
+        self.window_size = window_size
+        self.shift_sequence = shift_sequence
+        self.shift_window = shift_window
+        self.serialize_mode = serialize_mode
+        self.use_rope = use_rope
+        self.qk_rms_norm = qk_rms_norm
+
+        if self._type == "self":
+            self.to_qkv = nn.Linear(channels, channels * 3, bias=qkv_bias)
+        else:
+            self.to_q = nn.Linear(channels, channels, bias=qkv_bias)
+            self.to_kv = nn.Linear(self.ctx_channels, channels * 2, bias=qkv_bias)
+        
+        if self.qk_rms_norm:
+            self.q_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
+            self.k_rms_norm = SparseMultiHeadRMSNorm(channels // num_heads, num_heads)
+            
+        self.to_out = nn.Linear(channels, channels)
+
+        if use_rope:
+            self.rope = RotaryPositionEmbedder(channels)
+
+    @staticmethod
+    def _linear(module: nn.Linear, x: Union[SparseTensor, torch.Tensor]) -> Union[SparseTensor, torch.Tensor]:
+        if isinstance(x, SparseTensor):
+            return x.replace(module(x.feats))
+        else:
+            return module(x)
+
+    @staticmethod
+    def _reshape_chs(x: Union[SparseTensor, torch.Tensor], shape: Tuple[int, ...]) -> Union[SparseTensor, torch.Tensor]:
+        if isinstance(x, SparseTensor):
+            return x.reshape(*shape)
+        else:
+            return x.reshape(*x.shape[:2], *shape)
+
+    def _fused_pre(self, x: Union[SparseTensor, torch.Tensor], num_fused: int) -> Union[SparseTensor, torch.Tensor]:
+        if isinstance(x, SparseTensor):
+            x_feats = x.feats.unsqueeze(0)
+        else:
+            x_feats = x
+        x_feats = x_feats.reshape(*x_feats.shape[:2], num_fused, self.num_heads, -1)
+        return x.replace(x_feats.squeeze(0)) if isinstance(x, SparseTensor) else x_feats
+
+    def _rope(self, qkv: SparseTensor) -> SparseTensor:
+        q, k, v = qkv.feats.unbind(dim=1)   # [T, H, C]
+        q, k = self.rope(q, k, qkv.coords[:, 1:])
+        qkv = qkv.replace(torch.stack([q, k, v], dim=1)) 
+        return qkv
+    
+    def forward(self, x: Union[SparseTensor, torch.Tensor], context: Optional[Union[SparseTensor, torch.Tensor]] = None) -> Union[SparseTensor, torch.Tensor]:
+        if self._type == "self":
+            qkv = self._linear(self.to_qkv, x)
+            qkv = self._fused_pre(qkv, num_fused=3)
+            if self.use_rope:
+                qkv = self._rope(qkv)
+            if self.qk_rms_norm:
+                q, k, v = qkv.unbind(dim=1)
+                q = self.q_rms_norm(q)
+                k = self.k_rms_norm(k)
+                qkv = qkv.replace(torch.stack([q.feats, k.feats, v.feats], dim=1))
+            if self.attn_mode == "full":
+                h = sparse_scaled_dot_product_attention(qkv)
+            elif self.attn_mode == "serialized":
+                h = sparse_serialized_scaled_dot_product_self_attention(
+                    qkv, self.window_size, serialize_mode=self.serialize_mode, shift_sequence=self.shift_sequence, shift_window=self.shift_window
+                )
+            elif self.attn_mode == "windowed":
+                h = sparse_windowed_scaled_dot_product_self_attention(
+                    qkv, self.window_size, shift_window=self.shift_window
+                )
+        else:
+            q = self._linear(self.to_q, x)
+            q = self._reshape_chs(q, (self.num_heads, -1))
+            kv = self._linear(self.to_kv, context)
+            kv = self._fused_pre(kv, num_fused=2)
+            if self.qk_rms_norm:
+                q = self.q_rms_norm(q)
+                k, v = kv.unbind(dim=1)
+                k = self.k_rms_norm(k)
+                kv = kv.replace(torch.stack([k.feats, v.feats], dim=1))
+            h = sparse_scaled_dot_product_attention(q, kv)
+        h = self._reshape_chs(h, (-1,))
+        h = self._linear(self.to_out, h)
+        return h

trellis/modules/sparse/attention/serialized_attn.py

@@ -0,0 +1,193 @@
+from typing import *
+from enum import Enum
+import torch
+import math
+from .. import SparseTensor
+from .. import DEBUG, ATTN
+
+if ATTN == 'xformers':
+    import xformers.ops as xops
+elif ATTN == 'flash_attn':
+    import flash_attn
+else:
+    raise ValueError(f"Unknown attention module: {ATTN}")
+
+
+__all__ = [
+    'sparse_serialized_scaled_dot_product_self_attention',
+]
+
+
+class SerializeMode(Enum):
+    Z_ORDER = 0
+    Z_ORDER_TRANSPOSED = 1
+    HILBERT = 2
+    HILBERT_TRANSPOSED = 3
+
+
+SerializeModes = [
+    SerializeMode.Z_ORDER,
+    SerializeMode.Z_ORDER_TRANSPOSED,
+    SerializeMode.HILBERT,
+    SerializeMode.HILBERT_TRANSPOSED
+]
+
+
+def calc_serialization(
+    tensor: SparseTensor,
+    window_size: int,
+    serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
+    shift_sequence: int = 0,
+    shift_window: Tuple[int, int, int] = (0, 0, 0)
+) -> Tuple[torch.Tensor, torch.Tensor, List[int]]:
+    """
+    Calculate serialization and partitioning for a set of coordinates.
+
+    Args:
+        tensor (SparseTensor): The input tensor.
+        window_size (int): The window size to use.
+        serialize_mode (SerializeMode): The serialization mode to use.
+        shift_sequence (int): The shift of serialized sequence.
+        shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
+
+    Returns:
+        (torch.Tensor, torch.Tensor): Forwards and backwards indices.
+    """
+    fwd_indices = []
+    bwd_indices = []
+    seq_lens = []
+    seq_batch_indices = []
+    offsets = [0]
+    
+    if 'vox2seq' not in globals():
+        import vox2seq
+
+    # Serialize the input
+    serialize_coords = tensor.coords[:, 1:].clone()
+    serialize_coords += torch.tensor(shift_window, dtype=torch.int32, device=tensor.device).reshape(1, 3)
+    if serialize_mode == SerializeMode.Z_ORDER:
+        code = vox2seq.encode(serialize_coords, mode='z_order', permute=[0, 1, 2])
+    elif serialize_mode == SerializeMode.Z_ORDER_TRANSPOSED:
+        code = vox2seq.encode(serialize_coords, mode='z_order', permute=[1, 0, 2])
+    elif serialize_mode == SerializeMode.HILBERT:
+        code = vox2seq.encode(serialize_coords, mode='hilbert', permute=[0, 1, 2])
+    elif serialize_mode == SerializeMode.HILBERT_TRANSPOSED:
+        code = vox2seq.encode(serialize_coords, mode='hilbert', permute=[1, 0, 2])
+    else:
+        raise ValueError(f"Unknown serialize mode: {serialize_mode}")
+    
+    for bi, s in enumerate(tensor.layout):
+        num_points = s.stop - s.start
+        num_windows = (num_points + window_size - 1) // window_size
+        valid_window_size = num_points / num_windows
+        to_ordered = torch.argsort(code[s.start:s.stop])
+        if num_windows == 1:
+            fwd_indices.append(to_ordered)
+            bwd_indices.append(torch.zeros_like(to_ordered).scatter_(0, to_ordered, torch.arange(num_points, device=tensor.device)))
+            fwd_indices[-1] += s.start
+            bwd_indices[-1] += offsets[-1]
+            seq_lens.append(num_points)
+            seq_batch_indices.append(bi)
+            offsets.append(offsets[-1] + seq_lens[-1])
+        else:
+            # Partition the input
+            offset = 0
+            mids = [(i + 0.5) * valid_window_size + shift_sequence for i in range(num_windows)]
+            split = [math.floor(i * valid_window_size + shift_sequence) for i in range(num_windows + 1)]
+            bwd_index = torch.zeros((num_points,), dtype=torch.int64, device=tensor.device)
+            for i in range(num_windows):
+                mid = mids[i]
+                valid_start = split[i]
+                valid_end = split[i + 1]
+                padded_start = math.floor(mid - 0.5 * window_size)
+                padded_end = padded_start + window_size
+                fwd_indices.append(to_ordered[torch.arange(padded_start, padded_end, device=tensor.device) % num_points])
+                offset += valid_start - padded_start
+                bwd_index.scatter_(0, fwd_indices[-1][valid_start-padded_start:valid_end-padded_start], torch.arange(offset, offset + valid_end - valid_start, device=tensor.device))
+                offset += padded_end - valid_start
+                fwd_indices[-1] += s.start
+            seq_lens.extend([window_size] * num_windows)
+            seq_batch_indices.extend([bi] * num_windows)
+            bwd_indices.append(bwd_index + offsets[-1])
+            offsets.append(offsets[-1] + num_windows * window_size)
+
+    fwd_indices = torch.cat(fwd_indices)
+    bwd_indices = torch.cat(bwd_indices)
+
+    return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
+    
+
+def sparse_serialized_scaled_dot_product_self_attention(
+    qkv: SparseTensor,
+    window_size: int,
+    serialize_mode: SerializeMode = SerializeMode.Z_ORDER,
+    shift_sequence: int = 0,
+    shift_window: Tuple[int, int, int] = (0, 0, 0)
+) -> SparseTensor:
+    """
+    Apply serialized scaled dot product self attention to a sparse tensor.
+
+    Args:
+        qkv (SparseTensor): [N, *, 3, H, C] sparse tensor containing Qs, Ks, and Vs.
+        window_size (int): The window size to use.
+        serialize_mode (SerializeMode): The serialization mode to use.
+        shift_sequence (int): The shift of serialized sequence.
+        shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
+        shift (int): The shift to use.
+    """
+    assert len(qkv.shape) == 4 and qkv.shape[1] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
+
+    serialization_spatial_cache_name = f'serialization_{serialize_mode}_{window_size}_{shift_sequence}_{shift_window}'
+    serialization_spatial_cache = qkv.get_spatial_cache(serialization_spatial_cache_name)
+    if serialization_spatial_cache is None:
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_serialization(qkv, window_size, serialize_mode, shift_sequence, shift_window)
+        qkv.register_spatial_cache(serialization_spatial_cache_name, (fwd_indices, bwd_indices, seq_lens, seq_batch_indices))
+    else:
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = serialization_spatial_cache
+
+    M = fwd_indices.shape[0]
+    T = qkv.feats.shape[0]
+    H = qkv.feats.shape[2]
+    C = qkv.feats.shape[3]
+    
+    qkv_feats = qkv.feats[fwd_indices]      # [M, 3, H, C]
+
+    if DEBUG:
+        start = 0
+        qkv_coords = qkv.coords[fwd_indices]
+        for i in range(len(seq_lens)):
+            assert (qkv_coords[start:start+seq_lens[i], 0] == seq_batch_indices[i]).all(), f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
+            start += seq_lens[i]
+
+    if all([seq_len == window_size for seq_len in seq_lens]):
+        B = len(seq_lens)
+        N = window_size
+        qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
+        if ATTN == 'xformers':
+            q, k, v = qkv_feats.unbind(dim=2)                       # [B, N, H, C]
+            out = xops.memory_efficient_attention(q, k, v)          # [B, N, H, C]
+        elif ATTN == 'flash_attn':
+            out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)   # [B, N, H, C]
+        else:
+            raise ValueError(f"Unknown attention module: {ATTN}")
+        out = out.reshape(B * N, H, C)                              # [M, H, C]
+    else:
+        if ATTN == 'xformers':
+            q, k, v = qkv_feats.unbind(dim=1)                       # [M, H, C]
+            q = q.unsqueeze(0)                                      # [1, M, H, C]
+            k = k.unsqueeze(0)                                      # [1, M, H, C]
+            v = v.unsqueeze(0)                                      # [1, M, H, C]
+            mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
+            out = xops.memory_efficient_attention(q, k, v, mask)[0] # [M, H, C]
+        elif ATTN == 'flash_attn':
+            cu_seqlens = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)], dim=0) \
+                        .to(qkv.device).int()
+            out = flash_attn.flash_attn_varlen_qkvpacked_func(qkv_feats, cu_seqlens, max(seq_lens)) # [M, H, C]
+
+    out = out[bwd_indices]      # [T, H, C]
+
+    if DEBUG:
+        qkv_coords = qkv_coords[bwd_indices]
+        assert torch.equal(qkv_coords, qkv.coords), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
+
+    return qkv.replace(out)

trellis/modules/sparse/attention/windowed_attn.py

@@ -0,0 +1,135 @@
+from typing import *
+import torch
+import math
+from .. import SparseTensor
+from .. import DEBUG, ATTN
+
+if ATTN == 'xformers':
+    import xformers.ops as xops
+elif ATTN == 'flash_attn':
+    import flash_attn
+else:
+    raise ValueError(f"Unknown attention module: {ATTN}")
+
+
+__all__ = [
+    'sparse_windowed_scaled_dot_product_self_attention',
+]
+
+
+def calc_window_partition(
+    tensor: SparseTensor,
+    window_size: Union[int, Tuple[int, ...]],
+    shift_window: Union[int, Tuple[int, ...]] = 0
+) -> Tuple[torch.Tensor, torch.Tensor, List[int], List[int]]:
+    """
+    Calculate serialization and partitioning for a set of coordinates.
+
+    Args:
+        tensor (SparseTensor): The input tensor.
+        window_size (int): The window size to use.
+        shift_window (Tuple[int, ...]): The shift of serialized coordinates.
+
+    Returns:
+        (torch.Tensor): Forwards indices.
+        (torch.Tensor): Backwards indices.
+        (List[int]): Sequence lengths.
+        (List[int]): Sequence batch indices.
+    """
+    DIM = tensor.coords.shape[1] - 1
+    shift_window = (shift_window,) * DIM if isinstance(shift_window, int) else shift_window
+    window_size = (window_size,) * DIM if isinstance(window_size, int) else window_size
+    shifted_coords = tensor.coords.clone().detach()
+    shifted_coords[:, 1:] += torch.tensor(shift_window, device=tensor.device, dtype=torch.int32).unsqueeze(0)
+
+    MAX_COORDS = shifted_coords[:, 1:].max(dim=0).values.tolist()
+    NUM_WINDOWS = [math.ceil((mc + 1) / ws) for mc, ws in zip(MAX_COORDS, window_size)]
+    OFFSET = torch.cumprod(torch.tensor([1] + NUM_WINDOWS[::-1]), dim=0).tolist()[::-1]
+
+    shifted_coords[:, 1:] //= torch.tensor(window_size, device=tensor.device, dtype=torch.int32).unsqueeze(0)
+    shifted_indices = (shifted_coords * torch.tensor(OFFSET, device=tensor.device, dtype=torch.int32).unsqueeze(0)).sum(dim=1)
+    fwd_indices = torch.argsort(shifted_indices)
+    bwd_indices = torch.empty_like(fwd_indices)
+    bwd_indices[fwd_indices] = torch.arange(fwd_indices.shape[0], device=tensor.device)
+    seq_lens = torch.bincount(shifted_indices)
+    seq_batch_indices = torch.arange(seq_lens.shape[0], device=tensor.device, dtype=torch.int32) // OFFSET[0]
+    mask = seq_lens != 0
+    seq_lens = seq_lens[mask].tolist()
+    seq_batch_indices = seq_batch_indices[mask].tolist()
+
+    return fwd_indices, bwd_indices, seq_lens, seq_batch_indices
+    
+
+def sparse_windowed_scaled_dot_product_self_attention(
+    qkv: SparseTensor,
+    window_size: int,
+    shift_window: Tuple[int, int, int] = (0, 0, 0)
+) -> SparseTensor:
+    """
+    Apply windowed scaled dot product self attention to a sparse tensor.
+
+    Args:
+        qkv (SparseTensor): [N, *, 3, H, C] sparse tensor containing Qs, Ks, and Vs.
+        window_size (int): The window size to use.
+        shift_window (Tuple[int, int, int]): The shift of serialized coordinates.
+        shift (int): The shift to use.
+    """
+    assert len(qkv.shape) == 4 and qkv.shape[1] == 3, f"Invalid shape for qkv, got {qkv.shape}, expected [N, *, 3, H, C]"
+
+    serialization_spatial_cache_name = f'window_partition_{window_size}_{shift_window}'
+    serialization_spatial_cache = qkv.get_spatial_cache(serialization_spatial_cache_name)
+    if serialization_spatial_cache is None:
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = calc_window_partition(qkv, window_size, shift_window)
+        qkv.register_spatial_cache(serialization_spatial_cache_name, (fwd_indices, bwd_indices, seq_lens, seq_batch_indices))
+    else:
+        fwd_indices, bwd_indices, seq_lens, seq_batch_indices = serialization_spatial_cache
+
+    M = fwd_indices.shape[0]
+    T = qkv.feats.shape[0]
+    H = qkv.feats.shape[2]
+    C = qkv.feats.shape[3]
+    
+    qkv_feats = qkv.feats[fwd_indices]      # [M, 3, H, C]
+
+    if DEBUG:
+        start = 0
+        qkv_coords = qkv.coords[fwd_indices]
+        for i in range(len(seq_lens)):
+            seq_coords = qkv_coords[start:start+seq_lens[i]]
+            assert (seq_coords[:, 0] == seq_batch_indices[i]).all(), f"SparseWindowedScaledDotProductSelfAttention: batch index mismatch"
+            assert (seq_coords[:, 1:].max(dim=0).values - seq_coords[:, 1:].min(dim=0).values < window_size).all(), \
+                    f"SparseWindowedScaledDotProductSelfAttention: window size exceeded"
+            start += seq_lens[i]
+
+    if all([seq_len == window_size for seq_len in seq_lens]):
+        B = len(seq_lens)
+        N = window_size
+        qkv_feats = qkv_feats.reshape(B, N, 3, H, C)
+        if ATTN == 'xformers':
+            q, k, v = qkv_feats.unbind(dim=2)                       # [B, N, H, C]
+            out = xops.memory_efficient_attention(q, k, v)          # [B, N, H, C]
+        elif ATTN == 'flash_attn':
+            out = flash_attn.flash_attn_qkvpacked_func(qkv_feats)   # [B, N, H, C]
+        else:
+            raise ValueError(f"Unknown attention module: {ATTN}")
+        out = out.reshape(B * N, H, C)                              # [M, H, C]
+    else:
+        if ATTN == 'xformers':
+            q, k, v = qkv_feats.unbind(dim=1)                       # [M, H, C]
+            q = q.unsqueeze(0)                                      # [1, M, H, C]
+            k = k.unsqueeze(0)                                      # [1, M, H, C]
+            v = v.unsqueeze(0)                                      # [1, M, H, C]
+            mask = xops.fmha.BlockDiagonalMask.from_seqlens(seq_lens)
+            out = xops.memory_efficient_attention(q, k, v, mask)[0] # [M, H, C]
+        elif ATTN == 'flash_attn':
+            cu_seqlens = torch.cat([torch.tensor([0]), torch.cumsum(torch.tensor(seq_lens), dim=0)], dim=0) \
+                        .to(qkv.device).int()
+            out = flash_attn.flash_attn_varlen_qkvpacked_func(qkv_feats, cu_seqlens, max(seq_lens)) # [M, H, C]
+
+    out = out[bwd_indices]      # [T, H, C]
+
+    if DEBUG:
+        qkv_coords = qkv_coords[bwd_indices]
+        assert torch.equal(qkv_coords, qkv.coords), "SparseWindowedScaledDotProductSelfAttention: coordinate mismatch"
+
+    return qkv.replace(out)

trellis/modules/sparse/basic.py

@@ -0,0 +1,459 @@
+from typing import *
+import torch
+import torch.nn as nn
+from . import BACKEND, DEBUG
+SparseTensorData = None # Lazy import
+
+
+__all__ = [
+    'SparseTensor',
+    'sparse_batch_broadcast',
+    'sparse_batch_op',
+    'sparse_cat',
+    'sparse_unbind',
+]
+
+
+class SparseTensor:
+    """
+    Sparse tensor with support for both torchsparse and spconv backends.
+    
+    Parameters:
+    - feats (torch.Tensor): Features of the sparse tensor.
+    - coords (torch.Tensor): Coordinates of the sparse tensor.
+    - shape (torch.Size): Shape of the sparse tensor.
+    - layout (List[slice]): Layout of the sparse tensor for each batch
+    - data (SparseTensorData): Sparse tensor data used for convolusion
+
+    NOTE:
+    - Data corresponding to a same batch should be contiguous.
+    - Coords should be in [0, 1023]
+    """
+    @overload
+    def __init__(self, feats: torch.Tensor, coords: torch.Tensor, shape: Optional[torch.Size] = None, layout: Optional[List[slice]] = None, **kwargs): ...
+
+    @overload
+    def __init__(self, data, shape: Optional[torch.Size] = None, layout: Optional[List[slice]] = None, **kwargs): ...
+
+    def __init__(self, *args, **kwargs):
+        # Lazy import of sparse tensor backend
+        global SparseTensorData
+        if SparseTensorData is None:
+            import importlib
+            if BACKEND == 'torchsparse':
+                SparseTensorData = importlib.import_module('torchsparse').SparseTensor
+            elif BACKEND == 'spconv':
+                SparseTensorData = importlib.import_module('spconv.pytorch').SparseConvTensor
+                
+        method_id = 0
+        if len(args) != 0:
+            method_id = 0 if isinstance(args[0], torch.Tensor) else 1
+        else:
+            method_id = 1 if 'data' in kwargs else 0
+
+        if method_id == 0:
+            feats, coords, shape, layout = args + (None,) * (4 - len(args))
+            if 'feats' in kwargs:
+                feats = kwargs['feats']
+                del kwargs['feats']
+            if 'coords' in kwargs:
+                coords = kwargs['coords']
+                del kwargs['coords']
+            if 'shape' in kwargs:
+                shape = kwargs['shape']
+                del kwargs['shape']
+            if 'layout' in kwargs:
+                layout = kwargs['layout']
+                del kwargs['layout']
+
+            if shape is None:
+                shape = self.__cal_shape(feats, coords)
+            if layout is None:
+                layout = self.__cal_layout(coords, shape[0])
+            if BACKEND == 'torchsparse':
+                self.data = SparseTensorData(feats, coords, **kwargs)
+            elif BACKEND == 'spconv':
+                spatial_shape = list(coords.max(0)[0] + 1)[1:]
+                self.data = SparseTensorData(feats.reshape(feats.shape[0], -1), coords, spatial_shape, shape[0], **kwargs)
+                self.data._features = feats
+        elif method_id == 1:
+            data, shape, layout = args + (None,) * (3 - len(args))
+            if 'data' in kwargs:
+                data = kwargs['data']
+                del kwargs['data']
+            if 'shape' in kwargs:
+                shape = kwargs['shape']
+                del kwargs['shape']
+            if 'layout' in kwargs:
+                layout = kwargs['layout']
+                del kwargs['layout']
+
+            self.data = data
+            if shape is None:
+                shape = self.__cal_shape(self.feats, self.coords)
+            if layout is None:
+                layout = self.__cal_layout(self.coords, shape[0])
+
+        self._shape = shape
+        self._layout = layout
+        self._scale = kwargs.get('scale', (1, 1, 1))
+        self._spatial_cache = kwargs.get('spatial_cache', {})
+
+        if DEBUG:
+            try:
+                assert self.feats.shape[0] == self.coords.shape[0], f"Invalid feats shape: {self.feats.shape}, coords shape: {self.coords.shape}"
+                assert self.shape == self.__cal_shape(self.feats, self.coords), f"Invalid shape: {self.shape}"
+                assert self.layout == self.__cal_layout(self.coords, self.shape[0]), f"Invalid layout: {self.layout}"
+                for i in range(self.shape[0]):
+                    assert torch.all(self.coords[self.layout[i], 0] == i), f"The data of batch {i} is not contiguous"
+            except Exception as e:
+                print('Debugging information:')
+                print(f"- Shape: {self.shape}")
+                print(f"- Layout: {self.layout}")
+                print(f"- Scale: {self._scale}")
+                print(f"- Coords: {self.coords}")
+                raise e
+        
+    def __cal_shape(self, feats, coords):
+        shape = []
+        shape.append(coords[:, 0].max().item() + 1)
+        shape.extend([*feats.shape[1:]])
+        return torch.Size(shape)
+    
+    def __cal_layout(self, coords, batch_size):
+        seq_len = torch.bincount(coords[:, 0], minlength=batch_size)
+        offset = torch.cumsum(seq_len, dim=0) 
+        layout = [slice((offset[i] - seq_len[i]).item(), offset[i].item()) for i in range(batch_size)]
+        return layout
+    
+    @property
+    def shape(self) -> torch.Size:
+        return self._shape
+    
+    def dim(self) -> int:
+        return len(self.shape)
+    
+    @property
+    def layout(self) -> List[slice]:
+        return self._layout
+
+    @property
+    def feats(self) -> torch.Tensor:
+        if BACKEND == 'torchsparse':
+            return self.data.F
+        elif BACKEND == 'spconv':
+            return self.data.features
+    
+    @feats.setter
+    def feats(self, value: torch.Tensor):
+        if BACKEND == 'torchsparse':
+            self.data.F = value
+        elif BACKEND == 'spconv':
+            self.data.features = value
+
+    @property
+    def coords(self) -> torch.Tensor:
+        if BACKEND == 'torchsparse':
+            return self.data.C
+        elif BACKEND == 'spconv':
+            return self.data.indices
+        
+    @coords.setter
+    def coords(self, value: torch.Tensor):
+        if BACKEND == 'torchsparse':
+            self.data.C = value
+        elif BACKEND == 'spconv':
+            self.data.indices = value
+
+    @property
+    def dtype(self):
+        return self.feats.dtype
+
+    @property
+    def device(self):
+        return self.feats.device
+
+    @overload
+    def to(self, dtype: torch.dtype) -> 'SparseTensor': ...
+
+    @overload
+    def to(self, device: Optional[Union[str, torch.device]] = None, dtype: Optional[torch.dtype] = None) -> 'SparseTensor': ...
+
+    def to(self, *args, **kwargs) -> 'SparseTensor':
+        device = None
+        dtype = None
+        if len(args) == 2:
+            device, dtype = args
+        elif len(args) == 1:
+            if isinstance(args[0], torch.dtype):
+                dtype = args[0]
+            else:
+                device = args[0]
+        if 'dtype' in kwargs:
+            assert dtype is None, "to() received multiple values for argument 'dtype'"
+            dtype = kwargs['dtype']
+        if 'device' in kwargs:
+            assert device is None, "to() received multiple values for argument 'device'"
+            device = kwargs['device']
+        
+        new_feats = self.feats.to(device=device, dtype=dtype)
+        new_coords = self.coords.to(device=device)
+        return self.replace(new_feats, new_coords)
+
+    def type(self, dtype):
+        new_feats = self.feats.type(dtype)
+        return self.replace(new_feats)
+
+    def cpu(self) -> 'SparseTensor':
+        new_feats = self.feats.cpu()
+        new_coords = self.coords.cpu()
+        return self.replace(new_feats, new_coords)
+    
+    def cuda(self) -> 'SparseTensor':
+        new_feats = self.feats.cuda()
+        new_coords = self.coords.cuda()
+        return self.replace(new_feats, new_coords)
+
+    def half(self) -> 'SparseTensor':
+        new_feats = self.feats.half()
+        return self.replace(new_feats)
+    
+    def float(self) -> 'SparseTensor':
+        new_feats = self.feats.float()
+        return self.replace(new_feats)
+    
+    def detach(self) -> 'SparseTensor':
+        new_coords = self.coords.detach()
+        new_feats = self.feats.detach()
+        return self.replace(new_feats, new_coords)
+
+    def dense(self) -> torch.Tensor:
+        if BACKEND == 'torchsparse':
+            return self.data.dense()
+        elif BACKEND == 'spconv':
+            return self.data.dense()
+
+    def reshape(self, *shape) -> 'SparseTensor':
+        new_feats = self.feats.reshape(self.feats.shape[0], *shape)
+        return self.replace(new_feats)
+    
+    def unbind(self, dim: int) -> List['SparseTensor']:
+        return sparse_unbind(self, dim)
+
+    def replace(self, feats: torch.Tensor, coords: Optional[torch.Tensor] = None) -> 'SparseTensor':
+        new_shape = [self.shape[0]]
+        new_shape.extend(feats.shape[1:])
+        if BACKEND == 'torchsparse':
+            new_data = SparseTensorData(
+                feats=feats,
+                coords=self.data.coords if coords is None else coords,
+                stride=self.data.stride,
+                spatial_range=self.data.spatial_range,
+            )
+            new_data._caches = self.data._caches
+        elif BACKEND == 'spconv':
+            new_data = SparseTensorData(
+                self.data.features.reshape(self.data.features.shape[0], -1),
+                self.data.indices,
+                self.data.spatial_shape,
+                self.data.batch_size,
+                self.data.grid,
+                self.data.voxel_num,
+                self.data.indice_dict
+            )
+            new_data._features = feats
+            new_data.benchmark = self.data.benchmark
+            new_data.benchmark_record = self.data.benchmark_record
+            new_data.thrust_allocator = self.data.thrust_allocator
+            new_data._timer = self.data._timer
+            new_data.force_algo = self.data.force_algo
+            new_data.int8_scale = self.data.int8_scale
+            if coords is not None:
+                new_data.indices = coords
+        new_tensor = SparseTensor(new_data, shape=torch.Size(new_shape), layout=self.layout, scale=self._scale, spatial_cache=self._spatial_cache)
+        return new_tensor
+
+    @staticmethod
+    def full(aabb, dim, value, dtype=torch.float32, device=None) -> 'SparseTensor':
+        N, C = dim
+        x = torch.arange(aabb[0], aabb[3] + 1)
+        y = torch.arange(aabb[1], aabb[4] + 1)
+        z = torch.arange(aabb[2], aabb[5] + 1)
+        coords = torch.stack(torch.meshgrid(x, y, z, indexing='ij'), dim=-1).reshape(-1, 3)
+        coords = torch.cat([
+            torch.arange(N).view(-1, 1).repeat(1, coords.shape[0]).view(-1, 1),
+            coords.repeat(N, 1),
+        ], dim=1).to(dtype=torch.int32, device=device)
+        feats = torch.full((coords.shape[0], C), value, dtype=dtype, device=device)
+        return SparseTensor(feats=feats, coords=coords)
+
+    def __merge_sparse_cache(self, other: 'SparseTensor') -> dict:
+        new_cache = {}
+        for k in set(list(self._spatial_cache.keys()) + list(other._spatial_cache.keys())):
+            if k in self._spatial_cache:
+                new_cache[k] = self._spatial_cache[k]
+            if k in other._spatial_cache:
+                if k not in new_cache:
+                    new_cache[k] = other._spatial_cache[k]
+                else:
+                    new_cache[k].update(other._spatial_cache[k])
+        return new_cache
+
+    def __neg__(self) -> 'SparseTensor':
+        return self.replace(-self.feats)
+    
+    def __elemwise__(self, other: Union[torch.Tensor, 'SparseTensor'], op: callable) -> 'SparseTensor':
+        if isinstance(other, torch.Tensor):
+            try:
+                other = torch.broadcast_to(other, self.shape)
+                other = sparse_batch_broadcast(self, other)
+            except:
+                pass
+        if isinstance(other, SparseTensor):
+            other = other.feats
+        new_feats = op(self.feats, other)
+        new_tensor = self.replace(new_feats)
+        if isinstance(other, SparseTensor):
+            new_tensor._spatial_cache = self.__merge_sparse_cache(other)
+        return new_tensor
+
+    def __add__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, torch.add)
+
+    def __radd__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, torch.add)
+    
+    def __sub__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, torch.sub)
+    
+    def __rsub__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, lambda x, y: torch.sub(y, x))
+
+    def __mul__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, torch.mul)
+
+    def __rmul__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, torch.mul)
+
+    def __truediv__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, torch.div)
+
+    def __rtruediv__(self, other: Union[torch.Tensor, 'SparseTensor', float]) -> 'SparseTensor':
+        return self.__elemwise__(other, lambda x, y: torch.div(y, x))
+
+    def __getitem__(self, idx):
+        if isinstance(idx, int):
+            idx = [idx]
+        elif isinstance(idx, slice):
+            idx = range(*idx.indices(self.shape[0]))
+        elif isinstance(idx, torch.Tensor):
+            if idx.dtype == torch.bool:
+                assert idx.shape == (self.shape[0],), f"Invalid index shape: {idx.shape}"
+                idx = idx.nonzero().squeeze(1)
+            elif idx.dtype in [torch.int32, torch.int64]:
+                assert len(idx.shape) == 1, f"Invalid index shape: {idx.shape}"
+            else:
+                raise ValueError(f"Unknown index type: {idx.dtype}")
+        else:
+            raise ValueError(f"Unknown index type: {type(idx)}")
+        
+        coords = []
+        feats = []
+        for new_idx, old_idx in enumerate(idx):
+            coords.append(self.coords[self.layout[old_idx]].clone())
+            coords[-1][:, 0] = new_idx
+            feats.append(self.feats[self.layout[old_idx]])
+        coords = torch.cat(coords, dim=0).contiguous()
+        feats = torch.cat(feats, dim=0).contiguous()
+        return SparseTensor(feats=feats, coords=coords)
+
+    def register_spatial_cache(self, key, value) -> None:
+        """
+        Register a spatial cache.
+        The spatial cache can be any thing you want to cache.
+        The registery and retrieval of the cache is based on current scale.
+        """
+        scale_key = str(self._scale)
+        if scale_key not in self._spatial_cache:
+            self._spatial_cache[scale_key] = {}
+        self._spatial_cache[scale_key][key] = value
+
+    def get_spatial_cache(self, key=None):
+        """
+        Get a spatial cache.
+        """
+        scale_key = str(self._scale)
+        cur_scale_cache = self._spatial_cache.get(scale_key, {})
+        if key is None:
+            return cur_scale_cache
+        return cur_scale_cache.get(key, None)
+
+
+def sparse_batch_broadcast(input: SparseTensor, other: torch.Tensor) -> torch.Tensor:
+    """
+    Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
+    
+    Args:
+        input (torch.Tensor): 1D tensor to broadcast.
+        target (SparseTensor): Sparse tensor to broadcast to.
+        op (callable): Operation to perform after broadcasting. Defaults to torch.add.
+    """
+    coords, feats = input.coords, input.feats
+    broadcasted = torch.zeros_like(feats)
+    for k in range(input.shape[0]):
+        broadcasted[input.layout[k]] = other[k]
+    return broadcasted
+
+
+def sparse_batch_op(input: SparseTensor, other: torch.Tensor, op: callable = torch.add) -> SparseTensor:
+    """
+    Broadcast a 1D tensor to a sparse tensor along the batch dimension then perform an operation.
+    
+    Args:
+        input (torch.Tensor): 1D tensor to broadcast.
+        target (SparseTensor): Sparse tensor to broadcast to.
+        op (callable): Operation to perform after broadcasting. Defaults to torch.add.
+    """
+    return input.replace(op(input.feats, sparse_batch_broadcast(input, other)))
+
+
+def sparse_cat(inputs: List[SparseTensor], dim: int = 0) -> SparseTensor:
+    """
+    Concatenate a list of sparse tensors.
+    
+    Args:
+        inputs (List[SparseTensor]): List of sparse tensors to concatenate.
+    """
+    if dim == 0:
+        start = 0
+        coords = []
+        for input in inputs:
+            coords.append(input.coords.clone())
+            coords[-1][:, 0] += start
+            start += input.shape[0]
+        coords = torch.cat(coords, dim=0)
+        feats = torch.cat([input.feats for input in inputs], dim=0)
+        output = SparseTensor(
+            coords=coords,
+            feats=feats,
+        )
+    else:
+        feats = torch.cat([input.feats for input in inputs], dim=dim)
+        output = inputs[0].replace(feats)
+
+    return output
+
+
+def sparse_unbind(input: SparseTensor, dim: int) -> List[SparseTensor]:
+    """
+    Unbind a sparse tensor along a dimension.
+    
+    Args:
+        input (SparseTensor): Sparse tensor to unbind.
+        dim (int): Dimension to unbind.
+    """
+    if dim == 0:
+        return [input[i] for i in range(input.shape[0])]
+    else:
+        feats = input.feats.unbind(dim)
+        return [input.replace(f) for f in feats]

trellis/modules/sparse/conv/__init__.py

@@ -0,0 +1,21 @@
+from .. import BACKEND
+
+
+SPCONV_ALGO = 'auto'    # 'auto', 'implicit_gemm', 'native'
+
+def __from_env():
+    import os
+        
+    global SPCONV_ALGO
+    env_spconv_algo = os.environ.get('SPCONV_ALGO')
+    if env_spconv_algo is not None and env_spconv_algo in ['auto', 'implicit_gemm', 'native']:
+        SPCONV_ALGO = env_spconv_algo
+    print(f"[SPARSE][CONV] spconv algo: {SPCONV_ALGO}")
+        
+
+__from_env()
+
+if BACKEND == 'torchsparse':
+    from .conv_torchsparse import *
+elif BACKEND == 'spconv':
+    from .conv_spconv import *

trellis/modules/sparse/conv/conv_spconv.py

@@ -0,0 +1,80 @@
+import torch
+import torch.nn as nn
+from .. import SparseTensor
+from .. import DEBUG
+from . import SPCONV_ALGO
+
+class SparseConv3d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, padding=None, bias=True, indice_key=None):
+        super(SparseConv3d, self).__init__()
+        if 'spconv' not in globals():
+            import spconv.pytorch as spconv
+        algo = None
+        if SPCONV_ALGO == 'native':
+            algo = spconv.ConvAlgo.Native
+        elif SPCONV_ALGO == 'implicit_gemm':
+            algo = spconv.ConvAlgo.MaskImplicitGemm
+        if stride == 1 and (padding is None):
+            self.conv = spconv.SubMConv3d(in_channels, out_channels, kernel_size, dilation=dilation, bias=bias, indice_key=indice_key, algo=algo)
+        else:
+            self.conv = spconv.SparseConv3d(in_channels, out_channels, kernel_size, stride=stride, dilation=dilation, padding=padding, bias=bias, indice_key=indice_key, algo=algo)
+        self.stride = tuple(stride) if isinstance(stride, (list, tuple)) else (stride, stride, stride)
+        self.padding = padding
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        spatial_changed = any(s != 1 for s in self.stride) or (self.padding is not None)
+        new_data = self.conv(x.data)
+        new_shape = [x.shape[0], self.conv.out_channels]
+        new_layout = None if spatial_changed else x.layout
+
+        if spatial_changed and (x.shape[0] != 1):
+            # spconv was non-1 stride will break the contiguous of the output tensor, sort by the coords
+            fwd = new_data.indices[:, 0].argsort()
+            bwd = torch.zeros_like(fwd).scatter_(0, fwd, torch.arange(fwd.shape[0], device=fwd.device))
+            sorted_feats = new_data.features[fwd]
+            sorted_coords = new_data.indices[fwd]
+            unsorted_data = new_data
+            new_data = spconv.SparseConvTensor(sorted_feats, sorted_coords, unsorted_data.spatial_shape, unsorted_data.batch_size)  # type: ignore
+
+        out = SparseTensor(
+            new_data, shape=torch.Size(new_shape), layout=new_layout,
+            scale=tuple([s * stride for s, stride in zip(x._scale, self.stride)]),
+            spatial_cache=x._spatial_cache,
+        )
+
+        if spatial_changed and (x.shape[0] != 1):
+            out.register_spatial_cache(f'conv_{self.stride}_unsorted_data', unsorted_data)
+            out.register_spatial_cache(f'conv_{self.stride}_sort_bwd', bwd)
+ 
+        return out
+
+
+class SparseInverseConv3d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):
+        super(SparseInverseConv3d, self).__init__()
+        if 'spconv' not in globals():
+            import spconv.pytorch as spconv
+        self.conv = spconv.SparseInverseConv3d(in_channels, out_channels, kernel_size, bias=bias, indice_key=indice_key)
+        self.stride = tuple(stride) if isinstance(stride, (list, tuple)) else (stride, stride, stride)
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        spatial_changed = any(s != 1 for s in self.stride)
+        if spatial_changed:
+            # recover the original spconv order
+            data = x.get_spatial_cache(f'conv_{self.stride}_unsorted_data')
+            bwd = x.get_spatial_cache(f'conv_{self.stride}_sort_bwd')
+            data = data.replace_feature(x.feats[bwd])
+            if DEBUG:
+                assert torch.equal(data.indices, x.coords[bwd]), 'Recover the original order failed'
+        else:
+            data = x.data
+
+        new_data = self.conv(data)
+        new_shape = [x.shape[0], self.conv.out_channels]
+        new_layout = None if spatial_changed else x.layout
+        out = SparseTensor(
+            new_data, shape=torch.Size(new_shape), layout=new_layout,
+            scale=tuple([s // stride for s, stride in zip(x._scale, self.stride)]),
+            spatial_cache=x._spatial_cache,
+        )
+        return out

trellis/modules/sparse/conv/conv_torchsparse.py

@@ -0,0 +1,38 @@
+import torch
+import torch.nn as nn
+from .. import SparseTensor
+
+
+class SparseConv3d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):
+        super(SparseConv3d, self).__init__()
+        if 'torchsparse' not in globals():
+            import torchsparse
+        self.conv = torchsparse.nn.Conv3d(in_channels, out_channels, kernel_size, stride, 0, dilation, bias)
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        out = self.conv(x.data)
+        new_shape = [x.shape[0], self.conv.out_channels]
+        out = SparseTensor(out, shape=torch.Size(new_shape), layout=x.layout if all(s == 1 for s in self.conv.stride) else None)
+        out._spatial_cache = x._spatial_cache
+        out._scale = tuple([s * stride for s, stride in zip(x._scale, self.conv.stride)])
+        return out
+
+
+class SparseInverseConv3d(nn.Module):
+    def __init__(self, in_channels, out_channels, kernel_size, stride=1, dilation=1, bias=True, indice_key=None):
+        super(SparseInverseConv3d, self).__init__()
+        if 'torchsparse' not in globals():
+            import torchsparse
+        self.conv = torchsparse.nn.Conv3d(in_channels, out_channels, kernel_size, stride, 0, dilation, bias, transposed=True)
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        out = self.conv(x.data)        
+        new_shape = [x.shape[0], self.conv.out_channels]
+        out = SparseTensor(out, shape=torch.Size(new_shape), layout=x.layout if all(s == 1 for s in self.conv.stride) else None)
+        out._spatial_cache = x._spatial_cache
+        out._scale = tuple([s // stride for s, stride in zip(x._scale, self.conv.stride)])
+        return out
+
+
+

trellis/modules/sparse/linear.py

@@ -0,0 +1,15 @@
+import torch
+import torch.nn as nn
+from . import SparseTensor
+
+__all__ = [
+    'SparseLinear'
+]
+
+
+class SparseLinear(nn.Linear):
+    def __init__(self, in_features, out_features, bias=True):
+        super(SparseLinear, self).__init__(in_features, out_features, bias)
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        return input.replace(super().forward(input.feats))

trellis/modules/sparse/nonlinearity.py

@@ -0,0 +1,35 @@
+import torch
+import torch.nn as nn
+from . import SparseTensor
+
+__all__ = [
+    'SparseReLU',
+    'SparseSiLU',
+    'SparseGELU',
+    'SparseActivation'
+]
+
+
+class SparseReLU(nn.ReLU):
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        return input.replace(super().forward(input.feats))
+    
+
+class SparseSiLU(nn.SiLU):
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        return input.replace(super().forward(input.feats))
+
+
+class SparseGELU(nn.GELU):
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        return input.replace(super().forward(input.feats))
+
+
+class SparseActivation(nn.Module):
+    def __init__(self, activation: nn.Module):
+        super().__init__()
+        self.activation = activation
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        return input.replace(self.activation(input.feats))
+    

trellis/modules/sparse/norm.py

@@ -0,0 +1,58 @@
+import torch
+import torch.nn as nn
+from . import SparseTensor
+from . import DEBUG
+
+__all__ = [
+    'SparseGroupNorm',
+    'SparseLayerNorm',
+    'SparseGroupNorm32',
+    'SparseLayerNorm32',
+]
+
+
+class SparseGroupNorm(nn.GroupNorm):
+    def __init__(self, num_groups, num_channels, eps=1e-5, affine=True):
+        super(SparseGroupNorm, self).__init__(num_groups, num_channels, eps, affine)
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        nfeats = torch.zeros_like(input.feats)
+        for k in range(input.shape[0]):
+            if DEBUG:
+                assert (input.coords[input.layout[k], 0] == k).all(), f"SparseGroupNorm: batch index mismatch"
+            bfeats = input.feats[input.layout[k]]
+            bfeats = bfeats.permute(1, 0).reshape(1, input.shape[1], -1)
+            bfeats = super().forward(bfeats)
+            bfeats = bfeats.reshape(input.shape[1], -1).permute(1, 0)
+            nfeats[input.layout[k]] = bfeats
+        return input.replace(nfeats)
+
+
+class SparseLayerNorm(nn.LayerNorm):
+    def __init__(self, normalized_shape, eps=1e-5, elementwise_affine=True):
+        super(SparseLayerNorm, self).__init__(normalized_shape, eps, elementwise_affine)
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        nfeats = torch.zeros_like(input.feats)
+        for k in range(input.shape[0]):
+            bfeats = input.feats[input.layout[k]]
+            bfeats = bfeats.permute(1, 0).reshape(1, input.shape[1], -1)
+            bfeats = super().forward(bfeats)
+            bfeats = bfeats.reshape(input.shape[1], -1).permute(1, 0)
+            nfeats[input.layout[k]] = bfeats
+        return input.replace(nfeats)
+
+
+class SparseGroupNorm32(SparseGroupNorm):
+    """
+    A GroupNorm layer that converts to float32 before the forward pass.
+    """
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        return super().forward(x.float()).type(x.dtype)
+
+class SparseLayerNorm32(SparseLayerNorm):
+    """
+    A LayerNorm layer that converts to float32 before the forward pass.
+    """
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        return super().forward(x.float()).type(x.dtype)

trellis/modules/sparse/spatial.py

@@ -0,0 +1,110 @@
+from typing import *
+import torch
+import torch.nn as nn
+from . import SparseTensor
+
+__all__ = [
+    'SparseDownsample',
+    'SparseUpsample',
+    'SparseSubdivide'
+]
+
+
+class SparseDownsample(nn.Module):
+    """
+    Downsample a sparse tensor by a factor of `factor`.
+    Implemented as average pooling.
+    """
+    def __init__(self, factor: Union[int, Tuple[int, ...], List[int]]):
+        super(SparseDownsample, self).__init__()
+        self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        DIM = input.coords.shape[-1] - 1
+        factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
+        assert DIM == len(factor), 'Input coordinates must have the same dimension as the downsample factor.'
+
+        coord = list(input.coords.unbind(dim=-1))
+        for i, f in enumerate(factor):
+            coord[i+1] = coord[i+1] // f
+
+        MAX = [coord[i+1].max().item() + 1 for i in range(DIM)]
+        OFFSET = torch.cumprod(torch.tensor(MAX[::-1]), 0).tolist()[::-1] + [1]
+        code = sum([c * o for c, o in zip(coord, OFFSET)])
+        code, idx = code.unique(return_inverse=True)
+
+        new_feats = torch.scatter_reduce(
+            torch.zeros(code.shape[0], input.feats.shape[1], device=input.feats.device, dtype=input.feats.dtype),
+            dim=0,
+            index=idx.unsqueeze(1).expand(-1, input.feats.shape[1]),
+            src=input.feats,
+            reduce='mean'
+        )
+        new_coords = torch.stack(
+            [code // OFFSET[0]] +
+            [(code // OFFSET[i+1]) % MAX[i] for i in range(DIM)],
+            dim=-1
+        )
+        out = SparseTensor(new_feats, new_coords, input.shape,)
+        out._scale = tuple([s // f for s, f in zip(input._scale, factor)])
+        out._spatial_cache = input._spatial_cache
+
+        out.register_spatial_cache(f'upsample_{factor}_coords', input.coords)
+        out.register_spatial_cache(f'upsample_{factor}_layout', input.layout)
+        out.register_spatial_cache(f'upsample_{factor}_idx', idx)
+
+        return out
+
+
+class SparseUpsample(nn.Module):
+    """
+    Upsample a sparse tensor by a factor of `factor`.
+    Implemented as nearest neighbor interpolation.
+    """
+    def __init__(self, factor: Union[int, Tuple[int, int, int], List[int]]):
+        super(SparseUpsample, self).__init__()
+        self.factor = tuple(factor) if isinstance(factor, (list, tuple)) else factor
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        DIM = input.coords.shape[-1] - 1
+        factor = self.factor if isinstance(self.factor, tuple) else (self.factor,) * DIM
+        assert DIM == len(factor), 'Input coordinates must have the same dimension as the upsample factor.'
+
+        new_coords = input.get_spatial_cache(f'upsample_{factor}_coords')
+        new_layout = input.get_spatial_cache(f'upsample_{factor}_layout')
+        idx = input.get_spatial_cache(f'upsample_{factor}_idx')
+        if any([x is None for x in [new_coords, new_layout, idx]]):
+            raise ValueError('Upsample cache not found. SparseUpsample must be paired with SparseDownsample.')
+        new_feats = input.feats[idx]
+        out = SparseTensor(new_feats, new_coords, input.shape, new_layout)
+        out._scale = tuple([s * f for s, f in zip(input._scale, factor)])
+        out._spatial_cache = input._spatial_cache
+        return out
+    
+class SparseSubdivide(nn.Module):
+    """
+    Upsample a sparse tensor by a factor of `factor`.
+    Implemented as nearest neighbor interpolation.
+    """
+    def __init__(self):
+        super(SparseSubdivide, self).__init__()
+
+    def forward(self, input: SparseTensor) -> SparseTensor:
+        DIM = input.coords.shape[-1] - 1
+        # upsample scale=2^DIM
+        n_cube = torch.ones([2] * DIM, device=input.device, dtype=torch.int)
+        n_coords = torch.nonzero(n_cube)
+        n_coords = torch.cat([torch.zeros_like(n_coords[:, :1]), n_coords], dim=-1)
+        factor = n_coords.shape[0]
+        assert factor == 2 ** DIM
+        # print(n_coords.shape)
+        new_coords = input.coords.clone()
+        new_coords[:, 1:] *= 2
+        new_coords = new_coords.unsqueeze(1) + n_coords.unsqueeze(0).to(new_coords.dtype)
+        
+        new_feats = input.feats.unsqueeze(1).expand(input.feats.shape[0], factor, *input.feats.shape[1:])
+        out = SparseTensor(new_feats.flatten(0, 1), new_coords.flatten(0, 1), input.shape)
+        out._scale = input._scale * 2
+        out._spatial_cache = input._spatial_cache
+        return out
+

trellis/modules/sparse/transformer/__init__.py

@@ -0,0 +1,2 @@
+from .blocks import *
+from .modulated import *

trellis/modules/sparse/transformer/blocks.py

@@ -0,0 +1,151 @@
+from typing import *
+import torch
+import torch.nn as nn
+from ..basic import SparseTensor
+from ..linear import SparseLinear
+from ..nonlinearity import SparseGELU
+from ..attention import SparseMultiHeadAttention, SerializeMode
+from ...norm import LayerNorm32
+
+
+class SparseFeedForwardNet(nn.Module):
+    def __init__(self, channels: int, mlp_ratio: float = 4.0):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            SparseLinear(channels, int(channels * mlp_ratio)),
+            SparseGELU(approximate="tanh"),
+            SparseLinear(int(channels * mlp_ratio), channels),
+        )
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        return self.mlp(x)
+
+
+class SparseTransformerBlock(nn.Module):
+    """
+    Sparse Transformer block (MSA + FFN).
+    """
+    def __init__(
+        self,
+        channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        window_size: Optional[int] = None,
+        shift_sequence: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        serialize_mode: Optional[SerializeMode] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qkv_bias: bool = True,
+        ln_affine: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.attn = SparseMultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_sequence=shift_sequence,
+            shift_window=shift_window,
+            serialize_mode=serialize_mode,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.mlp = SparseFeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+
+    def _forward(self, x: SparseTensor) -> SparseTensor:
+        h = x.replace(self.norm1(x.feats))
+        h = self.attn(h)
+        x = x + h
+        h = x.replace(self.norm2(x.feats))
+        h = self.mlp(h)
+        x = x + h
+        return x
+
+    def forward(self, x: SparseTensor) -> SparseTensor:
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, use_reentrant=False)
+        else:
+            return self._forward(x)
+
+
+class SparseTransformerCrossBlock(nn.Module):
+    """
+    Sparse Transformer cross-attention block (MSA + MCA + FFN).
+    """
+    def __init__(
+        self,
+        channels: int,
+        ctx_channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        window_size: Optional[int] = None,
+        shift_sequence: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        serialize_mode: Optional[SerializeMode] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+        qkv_bias: bool = True,
+        ln_affine: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.norm3 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.self_attn = SparseMultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            type="self",
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_sequence=shift_sequence,
+            shift_window=shift_window,
+            serialize_mode=serialize_mode,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.cross_attn = SparseMultiHeadAttention(
+            channels,
+            ctx_channels=ctx_channels,
+            num_heads=num_heads,
+            type="cross",
+            attn_mode="full",
+            qkv_bias=qkv_bias,
+            qk_rms_norm=qk_rms_norm_cross,
+        )
+        self.mlp = SparseFeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+
+    def _forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor):
+        h = x.replace(self.norm1(x.feats))
+        h = self.self_attn(h)
+        x = x + h
+        h = x.replace(self.norm2(x.feats))
+        h = self.cross_attn(h, context)
+        x = x + h
+        h = x.replace(self.norm3(x.feats))
+        h = self.mlp(h)
+        x = x + h
+        return x
+
+    def forward(self, x: SparseTensor, context: torch.Tensor):
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
+        else:
+            return self._forward(x, context)

trellis/modules/sparse/transformer/modulated.py

@@ -0,0 +1,166 @@
+from typing import *
+import torch
+import torch.nn as nn
+from ..basic import SparseTensor
+from ..attention import SparseMultiHeadAttention, SerializeMode
+from ...norm import LayerNorm32
+from .blocks import SparseFeedForwardNet
+
+
+class ModulatedSparseTransformerBlock(nn.Module):
+    """
+    Sparse Transformer block (MSA + FFN) with adaptive layer norm conditioning.
+    """
+    def __init__(
+        self,
+        channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        window_size: Optional[int] = None,
+        shift_sequence: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        serialize_mode: Optional[SerializeMode] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qkv_bias: bool = True,
+        share_mod: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.attn = SparseMultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_sequence=shift_sequence,
+            shift_window=shift_window,
+            serialize_mode=serialize_mode,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.mlp = SparseFeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+        if not share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(channels, 6 * channels, bias=True)
+            )
+
+    def _forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
+        if self.share_mod:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+        else:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+        h = x.replace(self.norm1(x.feats))
+        h = h * (1 + scale_msa) + shift_msa
+        h = self.attn(h)
+        h = h * gate_msa
+        x = x + h
+        h = x.replace(self.norm2(x.feats))
+        h = h * (1 + scale_mlp) + shift_mlp
+        h = self.mlp(h)
+        h = h * gate_mlp
+        x = x + h
+        return x
+
+    def forward(self, x: SparseTensor, mod: torch.Tensor) -> SparseTensor:
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, use_reentrant=False)
+        else:
+            return self._forward(x, mod)
+
+
+class ModulatedSparseTransformerCrossBlock(nn.Module):
+    """
+    Sparse Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
+    """
+    def __init__(
+        self,
+        channels: int,
+        ctx_channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "shift_window", "shift_sequence", "shift_order", "swin"] = "full",
+        window_size: Optional[int] = None,
+        shift_sequence: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        serialize_mode: Optional[SerializeMode] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+        qkv_bias: bool = True,
+        share_mod: bool = False,
+
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
+        self.norm3 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.self_attn = SparseMultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            type="self",
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_sequence=shift_sequence,
+            shift_window=shift_window,
+            serialize_mode=serialize_mode,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.cross_attn = SparseMultiHeadAttention(
+            channels,
+            ctx_channels=ctx_channels,
+            num_heads=num_heads,
+            type="cross",
+            attn_mode="full",
+            qkv_bias=qkv_bias,
+            qk_rms_norm=qk_rms_norm_cross,
+        )
+        self.mlp = SparseFeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+        if not share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(channels, 6 * channels, bias=True)
+            )
+
+    def _forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor) -> SparseTensor:
+        if self.share_mod:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+        else:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+        h = x.replace(self.norm1(x.feats))
+        h = h * (1 + scale_msa) + shift_msa
+        h = self.self_attn(h)
+        h = h * gate_msa
+        x = x + h
+        h = x.replace(self.norm2(x.feats))
+        h = self.cross_attn(h, context)
+        x = x + h
+        h = x.replace(self.norm3(x.feats))
+        h = h * (1 + scale_mlp) + shift_mlp
+        h = self.mlp(h)
+        h = h * gate_mlp
+        x = x + h
+        return x
+
+    def forward(self, x: SparseTensor, mod: torch.Tensor, context: torch.Tensor) -> SparseTensor:
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, context, use_reentrant=False)
+        else:
+            return self._forward(x, mod, context)

trellis/modules/spatial.py

@@ -0,0 +1,48 @@
+import torch
+
+
+def pixel_shuffle_3d(x: torch.Tensor, scale_factor: int) -> torch.Tensor:
+    """
+    3D pixel shuffle.
+    """
+    B, C, H, W, D = x.shape
+    C_ = C // scale_factor**3
+    x = x.reshape(B, C_, scale_factor, scale_factor, scale_factor, H, W, D)
+    x = x.permute(0, 1, 5, 2, 6, 3, 7, 4)
+    x = x.reshape(B, C_, H*scale_factor, W*scale_factor, D*scale_factor)
+    return x
+
+
+def patchify(x: torch.Tensor, patch_size: int):
+    """
+    Patchify a tensor.
+
+    Args:
+        x (torch.Tensor): (N, C, *spatial) tensor
+        patch_size (int): Patch size
+    """
+    DIM = x.dim() - 2
+    for d in range(2, DIM + 2):
+        assert x.shape[d] % patch_size == 0, f"Dimension {d} of input tensor must be divisible by patch size, got {x.shape[d]} and {patch_size}"
+
+    x = x.reshape(*x.shape[:2], *sum([[x.shape[d] // patch_size, patch_size] for d in range(2, DIM + 2)], []))
+    x = x.permute(0, 1, *([2 * i + 3 for i in range(DIM)] + [2 * i + 2 for i in range(DIM)]))
+    x = x.reshape(x.shape[0], x.shape[1] * (patch_size ** DIM), *(x.shape[-DIM:]))
+    return x
+
+
+def unpatchify(x: torch.Tensor, patch_size: int):
+    """
+    Unpatchify a tensor.
+
+    Args:
+        x (torch.Tensor): (N, C, *spatial) tensor
+        patch_size (int): Patch size
+    """
+    DIM = x.dim() - 2
+    assert x.shape[1] % (patch_size ** DIM) == 0, f"Second dimension of input tensor must be divisible by patch size to unpatchify, got {x.shape[1]} and {patch_size ** DIM}"
+
+    x = x.reshape(x.shape[0], x.shape[1] // (patch_size ** DIM), *([patch_size] * DIM), *(x.shape[-DIM:]))
+    x = x.permute(0, 1, *(sum([[2 + DIM + i, 2 + i] for i in range(DIM)], [])))
+    x = x.reshape(x.shape[0], x.shape[1], *[x.shape[2 + 2 * i] * patch_size for i in range(DIM)])
+    return x

trellis/modules/transformer/__init__.py

@@ -0,0 +1,2 @@
+from .blocks import *
+from .modulated import *

trellis/modules/transformer/blocks.py

@@ -0,0 +1,182 @@
+from typing import *
+import torch
+import torch.nn as nn
+from ..attention import MultiHeadAttention
+from ..norm import LayerNorm32
+
+
+class AbsolutePositionEmbedder(nn.Module):
+    """
+    Embeds spatial positions into vector representations.
+    """
+    def __init__(self, channels: int, in_channels: int = 3):
+        super().__init__()
+        self.channels = channels
+        self.in_channels = in_channels
+        self.freq_dim = channels // in_channels // 2
+        self.freqs = torch.arange(self.freq_dim, dtype=torch.float32) / self.freq_dim
+        self.freqs = 1.0 / (10000 ** self.freqs)
+        
+    def _sin_cos_embedding(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Create sinusoidal position embeddings.
+
+        Args:
+            x: a 1-D Tensor of N indices
+
+        Returns:
+            an (N, D) Tensor of positional embeddings.
+        """
+        self.freqs = self.freqs.to(x.device)
+        out = torch.outer(x, self.freqs)
+        out = torch.cat([torch.sin(out), torch.cos(out)], dim=-1)
+        return out
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        """
+        Args:
+            x (torch.Tensor): (N, D) tensor of spatial positions
+        """
+        N, D = x.shape
+        assert D == self.in_channels, "Input dimension must match number of input channels"
+        embed = self._sin_cos_embedding(x.reshape(-1))
+        embed = embed.reshape(N, -1)
+        if embed.shape[1] < self.channels:
+            embed = torch.cat([embed, torch.zeros(N, self.channels - embed.shape[1], device=embed.device)], dim=-1)
+        return embed
+
+
+class FeedForwardNet(nn.Module):
+    def __init__(self, channels: int, mlp_ratio: float = 4.0):
+        super().__init__()
+        self.mlp = nn.Sequential(
+            nn.Linear(channels, int(channels * mlp_ratio)),
+            nn.GELU(approximate="tanh"),
+            nn.Linear(int(channels * mlp_ratio), channels),
+        )
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        return self.mlp(x)
+
+
+class TransformerBlock(nn.Module):
+    """
+    Transformer block (MSA + FFN).
+    """
+    def __init__(
+        self,
+        channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "windowed"] = "full",
+        window_size: Optional[int] = None,
+        shift_window: Optional[int] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qkv_bias: bool = True,
+        ln_affine: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.attn = MultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_window=shift_window,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.mlp = FeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+
+    def _forward(self, x: torch.Tensor) -> torch.Tensor:
+        h = self.norm1(x)
+        h = self.attn(h)
+        x = x + h
+        h = self.norm2(x)
+        h = self.mlp(h)
+        x = x + h
+        return x
+
+    def forward(self, x: torch.Tensor) -> torch.Tensor:
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, use_reentrant=False)
+        else:
+            return self._forward(x)
+
+
+class TransformerCrossBlock(nn.Module):
+    """
+    Transformer cross-attention block (MSA + MCA + FFN).
+    """
+    def __init__(
+        self,
+        channels: int,
+        ctx_channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "windowed"] = "full",
+        window_size: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+        qkv_bias: bool = True,
+        ln_affine: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.norm1 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.norm3 = LayerNorm32(channels, elementwise_affine=ln_affine, eps=1e-6)
+        self.self_attn = MultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            type="self",
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_window=shift_window,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.cross_attn = MultiHeadAttention(
+            channels,
+            ctx_channels=ctx_channels,
+            num_heads=num_heads,
+            type="cross",
+            attn_mode="full",
+            qkv_bias=qkv_bias,
+            qk_rms_norm=qk_rms_norm_cross,
+        )
+        self.mlp = FeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+
+    def _forward(self, x: torch.Tensor, context: torch.Tensor):
+        h = self.norm1(x)
+        h = self.self_attn(h)
+        x = x + h
+        h = self.norm2(x)
+        h = self.cross_attn(h, context)
+        x = x + h
+        h = self.norm3(x)
+        h = self.mlp(h)
+        x = x + h
+        return x
+
+    def forward(self, x: torch.Tensor, context: torch.Tensor):
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, context, use_reentrant=False)
+        else:
+            return self._forward(x, context)
+        

trellis/modules/transformer/modulated.py

@@ -0,0 +1,157 @@
+from typing import *
+import torch
+import torch.nn as nn
+from ..attention import MultiHeadAttention
+from ..norm import LayerNorm32
+from .blocks import FeedForwardNet
+
+
+class ModulatedTransformerBlock(nn.Module):
+    """
+    Transformer block (MSA + FFN) with adaptive layer norm conditioning.
+    """
+    def __init__(
+        self,
+        channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "windowed"] = "full",
+        window_size: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qkv_bias: bool = True,
+        share_mod: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.attn = MultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_window=shift_window,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.mlp = FeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+        if not share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(channels, 6 * channels, bias=True)
+            )
+
+    def _forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:
+        if self.share_mod:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+        else:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+        h = self.norm1(x)
+        h = h * (1 + scale_msa.unsqueeze(1)) + shift_msa.unsqueeze(1)
+        h = self.attn(h)
+        h = h * gate_msa.unsqueeze(1)
+        x = x + h
+        h = self.norm2(x)
+        h = h * (1 + scale_mlp.unsqueeze(1)) + shift_mlp.unsqueeze(1)
+        h = self.mlp(h)
+        h = h * gate_mlp.unsqueeze(1)
+        x = x + h
+        return x
+
+    def forward(self, x: torch.Tensor, mod: torch.Tensor) -> torch.Tensor:
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, use_reentrant=False)
+        else:
+            return self._forward(x, mod)
+
+
+class ModulatedTransformerCrossBlock(nn.Module):
+    """
+    Transformer cross-attention block (MSA + MCA + FFN) with adaptive layer norm conditioning.
+    """
+    def __init__(
+        self,
+        channels: int,
+        ctx_channels: int,
+        num_heads: int,
+        mlp_ratio: float = 4.0,
+        attn_mode: Literal["full", "windowed"] = "full",
+        window_size: Optional[int] = None,
+        shift_window: Optional[Tuple[int, int, int]] = None,
+        use_checkpoint: bool = False,
+        use_rope: bool = False,
+        qk_rms_norm: bool = False,
+        qk_rms_norm_cross: bool = False,
+        qkv_bias: bool = True,
+        share_mod: bool = False,
+    ):
+        super().__init__()
+        self.use_checkpoint = use_checkpoint
+        self.share_mod = share_mod
+        self.norm1 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.norm2 = LayerNorm32(channels, elementwise_affine=True, eps=1e-6)
+        self.norm3 = LayerNorm32(channels, elementwise_affine=False, eps=1e-6)
+        self.self_attn = MultiHeadAttention(
+            channels,
+            num_heads=num_heads,
+            type="self",
+            attn_mode=attn_mode,
+            window_size=window_size,
+            shift_window=shift_window,
+            qkv_bias=qkv_bias,
+            use_rope=use_rope,
+            qk_rms_norm=qk_rms_norm,
+        )
+        self.cross_attn = MultiHeadAttention(
+            channels,
+            ctx_channels=ctx_channels,
+            num_heads=num_heads,
+            type="cross",
+            attn_mode="full",
+            qkv_bias=qkv_bias,
+            qk_rms_norm=qk_rms_norm_cross,
+        )
+        self.mlp = FeedForwardNet(
+            channels,
+            mlp_ratio=mlp_ratio,
+        )
+        if not share_mod:
+            self.adaLN_modulation = nn.Sequential(
+                nn.SiLU(),
+                nn.Linear(channels, 6 * channels, bias=True)
+            )
+
+    def _forward(self, x: torch.Tensor, mod: torch.Tensor, context: torch.Tensor):
+        if self.share_mod:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = mod.chunk(6, dim=1)
+        else:
+            shift_msa, scale_msa, gate_msa, shift_mlp, scale_mlp, gate_mlp = self.adaLN_modulation(mod).chunk(6, dim=1)
+        h = self.norm1(x)
+        h = h * (1 + scale_msa.unsqueeze(1)) + shift_msa.unsqueeze(1)
+        h = self.self_attn(h)
+        h = h * gate_msa.unsqueeze(1)
+        x = x + h
+        h = self.norm2(x)
+        h = self.cross_attn(h, context)
+        x = x + h
+        h = self.norm3(x)
+        h = h * (1 + scale_mlp.unsqueeze(1)) + shift_mlp.unsqueeze(1)
+        h = self.mlp(h)
+        h = h * gate_mlp.unsqueeze(1)
+        x = x + h
+        return x
+
+    def forward(self, x: torch.Tensor, mod: torch.Tensor, context: torch.Tensor):
+        if self.use_checkpoint:
+            return torch.utils.checkpoint.checkpoint(self._forward, x, mod, context, use_reentrant=False)
+        else:
+            return self._forward(x, mod, context)
+        

trellis/modules/utils.py

@@ -0,0 +1,54 @@
+import torch.nn as nn
+from ..modules import sparse as sp
+
+FP16_MODULES = (
+    nn.Conv1d,
+    nn.Conv2d,
+    nn.Conv3d,
+    nn.ConvTranspose1d,
+    nn.ConvTranspose2d,
+    nn.ConvTranspose3d,
+    nn.Linear,
+    sp.SparseConv3d,
+    sp.SparseInverseConv3d,
+    sp.SparseLinear,
+)
+
+def convert_module_to_f16(l):
+    """
+    Convert primitive modules to float16.
+    """
+    if isinstance(l, FP16_MODULES):
+        for p in l.parameters():
+            p.data = p.data.half()
+
+
+def convert_module_to_f32(l):
+    """
+    Convert primitive modules to float32, undoing convert_module_to_f16().
+    """
+    if isinstance(l, FP16_MODULES):
+        for p in l.parameters():
+            p.data = p.data.float()
+
+
+def zero_module(module):
+    """
+    Zero out the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().zero_()
+    return module
+
+
+def scale_module(module, scale):
+    """
+    Scale the parameters of a module and return it.
+    """
+    for p in module.parameters():
+        p.detach().mul_(scale)
+    return module
+
+
+def modulate(x, shift, scale):
+    return x * (1 + scale.unsqueeze(1)) + shift.unsqueeze(1)

trellis/pipelines/__init__.py

@@ -0,0 +1,24 @@
+from . import samplers
+from .trellis_image_to_3d import TrellisImageTo3DPipeline
+
+
+def from_pretrained(path: str):
+    """
+    Load a pipeline from a model folder or a Hugging Face model hub.
+
+    Args:
+        path: The path to the model. Can be either local path or a Hugging Face model name.
+    """
+    import os
+    import json
+    is_local = os.path.exists(f"{path}/pipeline.json")
+
+    if is_local:
+        config_file = f"{path}/pipeline.json"
+    else:
+        from huggingface_hub import hf_hub_download
+        config_file = hf_hub_download(path, "pipeline.json")
+
+    with open(config_file, 'r') as f:
+        config = json.load(f)
+    return globals()[config['name']].from_pretrained(path)

trellis/pipelines/base.py

@@ -0,0 +1,66 @@
+from typing import *
+import torch
+import torch.nn as nn
+from .. import models
+
+
+class Pipeline:
+    """
+    A base class for pipelines.
+    """
+    def __init__(
+        self,
+        models: dict[str, nn.Module] = None,
+    ):
+        if models is None:
+            return
+        self.models = models
+        for model in self.models.values():
+            model.eval()
+
+    @staticmethod
+    def from_pretrained(path: str) -> "Pipeline":
+        """
+        Load a pretrained model.
+        """
+        import os
+        import json
+        is_local = os.path.exists(f"{path}/pipeline.json")
+
+        if is_local:
+            config_file = f"{path}/pipeline.json"
+        else:
+            from huggingface_hub import hf_hub_download
+            config_file = hf_hub_download(path, "pipeline.json")
+
+        with open(config_file, 'r') as f:
+            args = json.load(f)['args']
+
+        _models = {
+            k: models.from_pretrained(f"{path}/{v}")
+            for k, v in args['models'].items()
+        }
+
+        new_pipeline = Pipeline(_models)
+        new_pipeline._pretrained_args = args
+        return new_pipeline
+
+    @property
+    def device(self) -> torch.device:
+        for model in self.models.values():
+            if hasattr(model, 'device'):
+                return model.device
+        for model in self.models.values():
+            if hasattr(model, 'parameters'):
+                return next(model.parameters()).device
+        raise RuntimeError("No device found.")
+
+    def to(self, device: torch.device) -> None:
+        for model in self.models.values():
+            model.to(device)
+
+    def cuda(self) -> None:
+        self.to(torch.device("cuda"))
+
+    def cpu(self) -> None:
+        self.to(torch.device("cpu"))

trellis/pipelines/samplers/__init__.py

@@ -0,0 +1,2 @@
+from .base import Sampler
+from .flow_euler import FlowEulerSampler, FlowEulerCfgSampler, FlowEulerGuidanceIntervalSampler

trellis/pipelines/samplers/base.py

@@ -0,0 +1,20 @@
+from typing import *
+from abc import ABC, abstractmethod
+
+
+class Sampler(ABC):
+    """
+    A base class for samplers.
+    """
+
+    @abstractmethod
+    def sample(
+        self,
+        model,
+        **kwargs
+    ):
+        """
+        Sample from a model.
+        """
+        pass
+    

trellis/pipelines/samplers/classifier_free_guidance_mixin.py

@@ -0,0 +1,12 @@
+from typing import *
+
+
+class ClassifierFreeGuidanceSamplerMixin:
+    """
+    A mixin class for samplers that apply classifier-free guidance.
+    """
+
+    def _inference_model(self, model, x_t, t, cond, neg_cond, cfg_strength, **kwargs):
+        pred = super()._inference_model(model, x_t, t, cond, **kwargs)
+        neg_pred = super()._inference_model(model, x_t, t, neg_cond, **kwargs)
+        return (1 + cfg_strength) * pred - cfg_strength * neg_pred

trellis/pipelines/samplers/flow_euler.py

@@ -0,0 +1,199 @@
+from typing import *
+import torch
+import numpy as np
+from tqdm import tqdm
+from easydict import EasyDict as edict
+from .base import Sampler
+from .classifier_free_guidance_mixin import ClassifierFreeGuidanceSamplerMixin
+from .guidance_interval_mixin import GuidanceIntervalSamplerMixin
+
+
+class FlowEulerSampler(Sampler):
+    """
+    Generate samples from a flow-matching model using Euler sampling.
+
+    Args:
+        sigma_min: The minimum scale of noise in flow.
+    """
+    def __init__(
+        self,
+        sigma_min: float,
+    ):
+        self.sigma_min = sigma_min
+
+    def _eps_to_xstart(self, x_t, t, eps):
+        assert x_t.shape == eps.shape
+        return (x_t - (self.sigma_min + (1 - self.sigma_min) * t) * eps) / (1 - t)
+
+    def _xstart_to_eps(self, x_t, t, x_0):
+        assert x_t.shape == x_0.shape
+        return (x_t - (1 - t) * x_0) / (self.sigma_min + (1 - self.sigma_min) * t)
+
+    def _v_to_xstart_eps(self, x_t, t, v):
+        assert x_t.shape == v.shape
+        eps = (1 - t) * v + x_t
+        x_0 = (1 - self.sigma_min) * x_t - (self.sigma_min + (1 - self.sigma_min) * t) * v
+        return x_0, eps
+
+    def _inference_model(self, model, x_t, t, cond=None, **kwargs):
+        t = torch.tensor([1000 * t] * x_t.shape[0], device=x_t.device, dtype=torch.float32)
+        return model(x_t, t, cond, **kwargs)
+
+    def _get_model_prediction(self, model, x_t, t, cond=None, **kwargs):
+        pred_v = self._inference_model(model, x_t, t, cond, **kwargs)
+        pred_x_0, pred_eps = self._v_to_xstart_eps(x_t=x_t, t=t, v=pred_v)
+        return pred_x_0, pred_eps, pred_v
+
+    @torch.no_grad()
+    def sample_once(
+        self,
+        model,
+        x_t,
+        t: float,
+        t_prev: float,
+        cond: Optional[Any] = None,
+        **kwargs
+    ):
+        """
+        Sample x_{t-1} from the model using Euler method.
+        
+        Args:
+            model: The model to sample from.
+            x_t: The [N x C x ...] tensor of noisy inputs at time t.
+            t: The current timestep.
+            t_prev: The previous timestep.
+            cond: conditional information.
+            **kwargs: Additional arguments for model inference.
+
+        Returns:
+            a dict containing the following
+            - 'pred_x_prev': x_{t-1}.
+            - 'pred_x_0': a prediction of x_0.
+        """
+        pred_x_0, pred_eps, pred_v = self._get_model_prediction(model, x_t, t, cond, **kwargs)
+        pred_x_prev = x_t - (t - t_prev) * pred_v
+        return edict({"pred_x_prev": pred_x_prev, "pred_x_0": pred_x_0})
+
+    @torch.no_grad()
+    def sample(
+        self,
+        model,
+        noise,
+        cond: Optional[Any] = None,
+        steps: int = 50,
+        rescale_t: float = 1.0,
+        verbose: bool = True,
+        **kwargs
+    ):
+        """
+        Generate samples from the model using Euler method.
+        
+        Args:
+            model: The model to sample from.
+            noise: The initial noise tensor.
+            cond: conditional information.
+            steps: The number of steps to sample.
+            rescale_t: The rescale factor for t.
+            verbose: If True, show a progress bar.
+            **kwargs: Additional arguments for model_inference.
+
+        Returns:
+            a dict containing the following
+            - 'samples': the model samples.
+            - 'pred_x_t': a list of prediction of x_t.
+            - 'pred_x_0': a list of prediction of x_0.
+        """
+        sample = noise
+        t_seq = np.linspace(1, 0, steps + 1)
+        t_seq = rescale_t * t_seq / (1 + (rescale_t - 1) * t_seq)
+        t_pairs = list((t_seq[i], t_seq[i + 1]) for i in range(steps))
+        ret = edict({"samples": None, "pred_x_t": [], "pred_x_0": []})
+        for t, t_prev in tqdm(t_pairs, desc="Sampling", disable=not verbose):
+            out = self.sample_once(model, sample, t, t_prev, cond, **kwargs)
+            sample = out.pred_x_prev
+            ret.pred_x_t.append(out.pred_x_prev)
+            ret.pred_x_0.append(out.pred_x_0)
+        ret.samples = sample
+        return ret
+
+
+class FlowEulerCfgSampler(ClassifierFreeGuidanceSamplerMixin, FlowEulerSampler):
+    """
+    Generate samples from a flow-matching model using Euler sampling with classifier-free guidance.
+    """
+    @torch.no_grad()
+    def sample(
+        self,
+        model,
+        noise,
+        cond,
+        neg_cond,
+        steps: int = 50,
+        rescale_t: float = 1.0,
+        cfg_strength: float = 3.0,
+        verbose: bool = True,
+        **kwargs
+    ):
+        """
+        Generate samples from the model using Euler method.
+        
+        Args:
+            model: The model to sample from.
+            noise: The initial noise tensor.
+            cond: conditional information.
+            neg_cond: negative conditional information.
+            steps: The number of steps to sample.
+            rescale_t: The rescale factor for t.
+            cfg_strength: The strength of classifier-free guidance.
+            verbose: If True, show a progress bar.
+            **kwargs: Additional arguments for model_inference.
+
+        Returns:
+            a dict containing the following
+            - 'samples': the model samples.
+            - 'pred_x_t': a list of prediction of x_t.
+            - 'pred_x_0': a list of prediction of x_0.
+        """
+        return super().sample(model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, **kwargs)
+
+
+class FlowEulerGuidanceIntervalSampler(GuidanceIntervalSamplerMixin, FlowEulerSampler):
+    """
+    Generate samples from a flow-matching model using Euler sampling with classifier-free guidance and interval.
+    """
+    @torch.no_grad()
+    def sample(
+        self,
+        model,
+        noise,
+        cond,
+        neg_cond,
+        steps: int = 50,
+        rescale_t: float = 1.0,
+        cfg_strength: float = 3.0,
+        cfg_interval: Tuple[float, float] = (0.0, 1.0),
+        verbose: bool = True,
+        **kwargs
+    ):
+        """
+        Generate samples from the model using Euler method.
+        
+        Args:
+            model: The model to sample from.
+            noise: The initial noise tensor.
+            cond: conditional information.
+            neg_cond: negative conditional information.
+            steps: The number of steps to sample.
+            rescale_t: The rescale factor for t.
+            cfg_strength: The strength of classifier-free guidance.
+            cfg_interval: The interval for classifier-free guidance.
+            verbose: If True, show a progress bar.
+            **kwargs: Additional arguments for model_inference.
+
+        Returns:
+            a dict containing the following
+            - 'samples': the model samples.
+            - 'pred_x_t': a list of prediction of x_t.
+            - 'pred_x_0': a list of prediction of x_0.
+        """
+        return super().sample(model, noise, cond, steps, rescale_t, verbose, neg_cond=neg_cond, cfg_strength=cfg_strength, cfg_interval=cfg_interval, **kwargs)