Set up environment variables and backends lazily. This is the main difference: we call our new lazy set_*_backend funcs.

Create sinusoidal position embeddings. Args: x: a 1-D Tensor of N indices Returns: an (N, D)

Connects four neighboring dual vertices to form a quadrilateral. The quadrilaterals are then split into triangles based on the gam

Return the attention configuration of the model.

Calculate serialization and partitioning for a set of coordinates. Args: tensor (SparseTensor): The input tensor. windo

Calculate serialization and partitioning for a set of coordinates. Args: tensor (SparseTensor): The input tensor. windo

construct a dense grid based on resolution

Convert primitive modules to float16.

Convert primitive modules to float32, undoing convert_module_to_f16().

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Convert the torso of the model to float16.

Args: cubes [Vx8] verts index for each cube value [Vx8xM] value to be scattered Operation: reduced[cubes[i][j]][

Encode the image. Args: image (Union[torch.Tensor, list[Image.Image]]): The image to encode Returns:

Context manager for inference that handles device placement and cleanup.

Load DINOv2 model with device, precision, and attention backend management

Load all required models with current configuration

Patchify a tensor. Args: x (torch.Tensor): (N, C, *spatial) tensor patch_size (int): Patch size

3D pixel shuffle.

Run the pipeline. Args: image (Image.Image): The image prompt. num_samples (int): The number of samples

Run the pipeline with multiple images as condition Args: images (List[Image.Image]): The multi-view images of the as

Sample x_{t-1} from the model using Euler method. Args: model: The model to sample from. x_t: T

Alias for set_attn for backwards compatibility

Set attention backend with validation

Set sparse backend with validation

Alias for set_backend for backwards compatibility Parameters: backend: The sparse backend to use algo: The algorithm to u

Set spconv algorithm with validation.

Apply serialized scaled dot product self attention to a sparse tensor. Args: qkv (SparseTensor): [N, *, 3, H, C] sparse tensor c

Unbind a sparse tensor along a dimension. Args: input (SparseTensor): Sparse tensor to unbind. dim (int): Dimension

Apply windowed scaled dot product self attention to a sparse tensor. Args: qkv (SparseTensor): [N, *, 3, H, C] sparse tensor con

Create sinusoidal timestep embeddings. Args: t: a 1-D Tensor of N indices, one per batch element. T

Convert a batch of network outputs to 3D representations. Args: x: The [N x * x C] sparse tensor output by the netwo

Unpatchify a tensor. Args: x (torch.Tensor): (N, C, *spatial) tensor patch_size (int): Patch size

Define input types with device-specific options.

Generates a mesh based on the specified sparse voxel structures. Args: cube_attrs [Nx21] : Sparse Tensor attrs about c

Function invoked when calling the pipeline. Args: image (`PIL.Image.Image`, `np.ndarray`, `torch.Tensor`, `List[PIL.Imag

Read current backend configuration

Read current backend configuration

Initialize the model manager with a specific model directory. Args: model_dir (str): Path to model directory