Perform a robust GPU cleanup sequence. This includes synchronizing, emptying caches, collecting IPC handles and running the Python garbage co

Invert a torch or numpy matrix

Invert (N, 3, 4) SE3 matrices → (N, 3, 4). R' = R^T, t' = -R^T @ t.

Return whether the active online cache policy uses the dynamic memory bank.

Append one frame's K/V tensors to the two-stream cache. Token layout must be: [camera, reg0, ..., regN, scale, patch0, ..., patch

Accept (..., 3,4) or (..., 4,4) extrinsics, return (...,4,4) homogeneous matrix. Supports torch.Tensor or np.ndarray.

Apply basic activation function to pose parameters. Args: pose_enc: Tensor containing encoded pose parameters act_type: Acti

Compute cross-frame attention using FlashInfer BatchPrefillWithPagedKVCacheWrapper. When self.force_fp32 is True, gathers all vis

optical flow induced by camera motion

Helper function to slice, expand and flatten tokens. Args: token: Token tensor [1, 2, N, C] where first index is for first frames

Processes specialized tokens with shape (1, 2, X, C) for multi-frame processing: 1) Uses the first position (index=0) for the first frame onl

Transfer some variables to another device (i.e. GPU, CPU:torch, CPU:numpy). batch: list, tuple, dict of tensors or other things device: pyto

Validate path

Modified from https://github.com/MoonshotAI/Moonlight/blob/master/examples/toy_train.py#L49 Major change: G is (B, C, C) rather than (C, C).

modified from https://github.com/MoonshotAI/Moonlight/blob/master/examples/toy_train.py#L49 Major change: G is [b, d, d] rather than [d, d]

Apply the configured bounded-cache policy to a state object with updated KV tensors.

return_all_blocks: if True, return the features at the end of every block instead of just the features from the l

SE(3) closed-form inverse transform, numerically stable, supports single (4,4) or batched (N,4,4). For SE(3) matrix [R|t; 0 1], its inverse is [R

Lift 3x3 matrices to homogeneous 4x4 matrices.

Lift 3x3 matrices to homogeneous 4x4 matrices.

Load a scene through BenchmarkDataset's current unified data pipeline. Current SpatialBenchmark layout: <benchmark_root>/<density>/<datas

Max-pool over two spatial axes with padding.

Adds a new submap to the optimizer and creates constraints with the previous submap, and optionally updates the current submap's pose based o

Accept (..., 3,4) or (..., 4,4) extrinsics, return (...,4,4) homogeneous matrix. Supports torch.Tensor or np.ndarray.

Compute the inverse of each 4x4 (or 3x4) SE3 matrix in a batch. If `R` and `T` are provided, they must correspond to the rotation and transl

Args: - depthmap (BxHxW array): - pseudo_focal: [B,H,W] ; [B,2,H,W] or [B,1,H,W] Returns: pointmap of absolute coordi

Encode patch tokens from frames. Params: - frames: dict containing 'images' (B, nimgs, 3, H, W) in [-1, 1] range

Gather sequence dimension and scatter head dimension for Q, K, V tensors.

Return a snapshot of current GPU memory usage or None if CUDA not available. Keys in returned dict: total_gb, allocated_gb, reserved_gb, free_gb,

Get stored keyframe image.

Iteratively undistort the normalized tracks using the given distortion parameters. Args: params (torch.Tensor or numpy.ndarray): Dis

Convert rotations given as rotation matrices to quaternions. Args: matrix: Rotation matrices as tensor of shape (..., 3, 3). Re

Parse export_feat parameter

Convert pose encoding [t, q, fov] to homogeneous transform matrices.

Quaternion Order: XYZW or say ijkr, scalar-last Convert rotations given as quaternions to rotation matrices. Args: quaternions:

Quaternion Order: XYZW or say ijkr, scalar-last Convert rotations given as quaternions to rotation matrices. Args: quaternions: