Compute FAD (and optionally CLAP score) from pre-extracted embeddings. Args: eval_embeddings: numpy array of embeddings (M, embeddin

Augment padding mask by randomly extending the valid region to include silence. This helps the model learn to handle silence at the end of a

Create a MultiConditioner from a conditioning config dictionary Args: config: the conditioning config dictionary device: the

Discrete Cosine Transform, Type II (a.k.a. the DCT) For the meaning of the parameter `norm`, see: https://docs.scipy.org/doc/scipy-0.14.

Design a lowpass filter using the Kaiser window. Parameters ---------- angular_cutoff : float The angular frequency cuto

Disable LoRA layers. If lora_index is None, disables all. If specified, disables only that index.

Enable LoRA layers. If lora_index is None, enables all. If specified, enables only that index.

Enable/disable LoRA layers by name filter. If lora_index is specified, only affects that index.

Generate audio from a batch of codes

Generate audio from a prompt using a diffusion model. Args: model: The diffusion model to use for generation. steps: The num

Gets context channels at `layer` and checks that shape is correct

See `pad_for_conv1d`.

Get rank of current process.

Return True if the model has at least one LoRAParametrization on any weight.

The inverse to DCT-II, which is a scaled Discrete Cosine Transform, Type III Our definition of idct is that idct(dct(x)) == x For the m

checks if entire clip is 'silence' below some dB threshold

Load and return a CLAP model. Select one of the following clap models from https://github.com/LAION-AI/CLAP: - music_speech_audioset

Load a LoRA checkpoint from either .safetensors or .ckpt format. Returns: Tuple of (state_dict, lora_config) where lora_config may be an

Compute the MMD loss along a chosen feature axis by collapsing all other dimensions. Args: x: Tensor of arbitrary shape.

Tiny wrapper around F.pad, just to allow for reflect padding on small input. If this is the case, we insert extra 0 padding to the right before th

Squeeze 2D DoRA magnitude tensors to 1D for loading. Modifies in-place.

this projects to lower dimenions, grabbing the first _`proj_dims`_ dimensions

Resolve legacy "dora" adapter_type to the correct variant. If adapter_type is "dora" and a checkpoint state_dict is provided, attempts to det

Round with straight through gradients.

Set lora strength. If lora_index is None, sets all LoRAs. If specified, sets only that index.

Performs spherical interpolation with magnitude interpolation and calculates local velocity. Args: tensor1: First tensor of shap

Strip silence from the end of an audio tensor. Args: audio: tensor [channels, samples] sample_rate: audio sample rate thr

All-reduce gradients for non-FSDP trainable parameters. Call this after backward() but before optimizer.step() to ensure gradient synchr

Returns the scaling factors for the clean image and for the noise, given a timestep.

Remove padding from x, handling properly zero padding. Only for 1d!

Newton-Schulz iteration to compute the polar factor of G. Produces approximately U @ V^T where G = U @ S @ V^T is the SVD. Uses a quinti

Absolute Threshold of Hearing in dB SPL (ISO 226:2003 approximation)

Build a flattened codebooks representation as a dictionary of inner step and the actual codebook indices corresponding to the flattened codebo

Build scatter indexes corresponding to the pattern, up to the provided sequence_steps. Args: timesteps (int): Maximum number of t

Helper method to call transformer and handle early exit logic.

Compute cross entropy between multi-codebook targets and model's logits. The cross entropy is computed per codebook to provide codebook-level

Vectorized exponential moving average

Generate inpainting demos from batch data with forced mask types.

Generate full t2m demos from specified prompts (FULL_MASK).

Generate a mask with 1-N non-overlapping contiguous spans covering a target ratio. Samples a target mask ratio from U[mask_ratio_range], the

Return the list of trainable parameter names for this adapter type.