Set the active number of codebooks used by the quantizer.

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

_flatten_streaming_state(state_dict, state_dict_metadata, state, prefix) Helper function for recursively flattening the streaming state into

Load Moshi LM with CPU offloading using accelerate. This function distributes model layers across GPU and CPU based on available GPU memory.

If voice_prompt_dir is None: - download voices.tgz from HF - extract it once - return extracted directory If voice_prompt_d

If voice_prompt_dir is None: - download voices.tgz from HF - extract it once - return extracted directory If voice_prompt_d

Args: q (torch.Tensor): queries, shape `[B, T, H, D]`. k (torch.Tensor): keys, shape `[B, T, H, D]`. offset (int): curren

Create cert.pem and key.pem using mkcert if they don't already exist. Returns (cert_file, key_file) or (None, None) if mkcert unavailable.

Create sinusoidal positional embedding, with shape `[B, T, C]`. Args: positions (torch.Tensor): LongTensor of positions. dim (int

Return a 440 Hz 'silent' sinewave of the given duration.

Main entry: create SSL context if certificates can be created. Returns (ssl_context, protocol_str) where protocol_str is "http" or "https".

Decode a single step of model tokens to PCM using Mimi. tokens is shaped [B, dep_q+1, 1]; channels 1..dep_q are the agent audio codebooks. Re

Set whether ema should be applied to the codebooks.

Takes an iterator of samples, batches them, encodes them; and yields the encoded samples one sample at a time in the same order.

Ensure mkcert exists; if not, download the correct binary. Returns ------- str | None Path to mkcert binary, or None if

Install mkcert CA if not installed yet. Parameters ---------- mkcert_bin : str Path to mkcert binary.

Return the complete streaming state, including that of sub-modules.

Indicate whether we are in a function that is CUDA Graphed (or will be soon).

Yields audio samples in intervals of sample_interval_size

Parse CLI args and run offline inference.

Normalize **mono** audio to a target LUFS level.

Run offline inference using an input WAV as the user-side stream. - Loads/initializes models and tokenizer - Warms up execution - Loads s