Pass `hidden_states` into `backbone` and `multi_head`, applying classifier-free guidance if `cfg_scale != 1.0`.

A helper to build the 'cond_dict' that the model expects. By default, it will generate a random speaker embedding

Generate a speaker embedding from an audio clip.

Sample next token from logits using either top_k/p/min_p OR using NovelAI's Unified Sampler. Args: logits (torch.Tensor): Input logit

"Prefill" mode: we already have `prefix_hidden_states`, and we want to append new embeddings, then compute the logits.

k/v: (batch_size, seqlen, nheads, head_dim) or (batch_size, 1, nheads, head_dim)

Sample next token using min-p sampling. Args: scores (torch.FloatTensor): Input logits with token candidates on the last dimension.

Sample next token from top K values along the last dimension of the input probs tensor. Args: probs (torch.Tensor): Input probabilities w

Sample next token from top P probabilities along the last dimension of the input probs tensor. Args: probs (torch.Tensor): Input probabil

Sample next token using unified sampling approach that combines linear scaling, confidence, and quadratic terms. Args: probs (torch.T

See https://arxiv.org/abs/1909.05858 Apply repetition penalty over a sliding window of the last `repetition_penalty_window` tokens. logits: (b

torch.multinomial with arbitrary number of dimensions, and number of candidates on the last dimension. Args: input (torch.Tensor): The in

Pad the weight of an embedding or linear layer to a multiple of `multiple`.

Single-step decode. Prepares the hidden states, possibly replicates them for CFG, and then delegates to `_compute_logits`. B

Args: texts: list of texts to convert to phonemes languages: ISO 639-1 -or otherwise eSpeak compatible- language code