Generates N random 4x4 Camera-to-World (c2w) homogeneous transformation matrices. The rotation (R) is generated using unit quaternions for u

Generate image/videos from noisy latents and compute the ODE regression loss. Input: - ode_latent: a tensor containing th

Get a pre-defined beta schedule for the given name. The beta schedule library consists of beta schedules which remain similar in the limi

Get current sequence length

Hash a string to an integer.

Download dummy data from a Hugging Face repository. Args: repo_id (str): The ID of the Hugging Face repository. filename (str): The

block expand strategy: each layer is expanded to continuous multiple layers Args: expand_ratio: expand ratio, default is 3, i.e.,

partial random init strategy: keep the first N layers, random init the remaining layers

cyclic copy strategy Args: zero_gate: whether to initialize the gate params of the non-first-appearing repeated layers to 0

interpolation init strategy

progressive init strategy (with noise)

partial random init strategy: keep the first N layers, random init the remaining layers

Return the intersection of the half-open integer interval r with the list of half-open integer intervals.

Decide whether a layer should run in chunk-causal (vs. fully bidirectional) mode. Chunk-causal mode applies when EITHER: 1. ``chunk_size``

Check if this is the first chunk.

This is a possible cleanup callback for cleanup_callback of MIndexedTar. It assumes that as long as there are some readers for a file, more r

This function is intended to align with the l2norm implementation in the FLA library.

Input: x: (..., C) weight: (2*D, C) bias: (2*D,) Output: y: (..., D)

Meta MovieGen style linear-quadratic scheduler. Output range matches time_uniform_flow: from ~0.999 to 0 Args: N: Number of samp

Kernel for computing the matmul C = A x B. A has shape (M, K), B has shape (K, N) and C has shape (M, N)

Load and process the input image

when the channel in FFN is not the same as the checkpoint, load the checkpoint