Lazy-import conditioner classes to avoid circular imports.

Return the silence latent for the dataset that contains this file, or None.

Run `steps` iterations of fixed-magnitude Griffin–Lim in RAW STFT. Keeps the STFT frame grid identical to X0 on all intermediate proj

Return [1,1,F,1] weights: interior bins √2, DC/Nyquist 1 (for one-sided energy).

Pack complex U[B,C,F,M] → real z[B,2·C·F,M] as [Re, Im] along channels.

Sample the next token for a batch of codes, and translates from conditioning inputs to model inputs Handles CFG inference

Build a (seq_q, seq_k) additive mask for masked SDPA fallback. 0 inside the band [i - w_left, i + w_right], -inf outside.

Convert t to normalized logsnr in [0, 1] for timestep conditioning. Maps t through logsnr = log((1-t)/t), clamps to fixed range, then

x[B,2,T] → [B,2,T] with orthonormal mid/side.

Inverse of _pack_complex_to_channels. Z[B,2·C·F,M] → U[B,C,F,M] (complex).

Runs checks on the layout to ensure a valid pattern is defined. A pattern is considered invalid if: - Multiple timesteps for a sam

Build a congruent power-of-2 discrete timestep sampler. Samples from a nested grid of discrete t values where coarser step counts (1, 2, 4, .

Build dist_shift from type + 4 params (meaning depends on type).

Compute feature matching loss between real and fake hidden states. Matches features at the discriminator layers for GAN stability.

Compute Frechet Latent Distance (FLD) - FAD in the VAE latent space. Args: eval_embeddings: numpy array of latent embeddings (N, lat

Create a boolean padding mask from per-batch valid sequence lengths. Args: valid_lengths: Tensor of shape (batch_size,) with valid l

Declip the signal by scaling down if any samples are clipped

Design the optimal prototype filter for a multiband system given the desired specs. Parameters ---------- attenuation : float

Design prototype filter for PQMF. This method is based on `A Kaiser window approach for the design of prototype filters of cosine modulated fi

Detect DoRA variant from a state dict. Returns "dora-cols" (dim=0, per-input-feature) or "dora-rows" (dim=1, per-output-neuron) based on the

Evaluate the filter's objective value based on the criteria from https://ieeexplore.ieee.org/document/681427 Parameters ----------

Export generator LoRA weights as a safetensors file with embedded config.

Compute pairwise squared Euclidean distances for flat tensors. Args: x: Tensor of shape (B, D) y: Tensor of shape (B, D)

Torch version of float32_to_int16 for GPU tensors.

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

Generate a QMF bank of cosine modulated filters based on a given prototype filter. Parameters ---------- prototype_filter : tor

get urls of shards (tar files) for multiple datasets in one s3 bucket

Compute alpha via log-linear interpolation in seq_len.

recursively get a list of audio filenames

finds the loudest value in the entire clip and puts that into dB (full scale)

Get parameters that are inside FSDP-wrapped submodules. Args: module: The root module to search Returns: Set of paramet

recursively get a list of pre-encoded filenames

Compute adaptive logsnr_start: drops by `rate` per doubling of seq_len.

Combines context time features and features into mapping