.arrow-backed dataset for tokenized audio and text samples. Assumes audio is saved as flat vlen int32 arrays (flattened [n_codebooks, seq_len

Custom implementation of ISTFT since torch.istft doesn't allow custom padding (other than `center=True`) with windowing. This is because the

ISTFT Head module for predicting STFT complex coefficients. Args: dim (int): Hidden dimension of the model. n_fft (int): Siz

Learning rate scheduler with a linear warmup and specificable decay.

Groups samples of similar lengths into bins to minimize padding.