mstar.model.orpheus.config

Classes

OrpheusModelConfig(num_hidden_layers, ...)

class mstar.model.orpheus.config.OrpheusModelConfig(num_hidden_layers: int = 28, num_attention_heads: int = 24, num_key_value_heads: int = 8, hidden_size: int = 3072, head_dim: int = 128, intermediate_size: int = 8192, max_position_embeddings: int = 131072, rms_norm_eps: float = 1e-05, rope_theta: float = 500000.0, vocab_size: int = 156940, tie_word_embeddings: bool = True, rope_scaling: dict = <factory>, start_token_id: int = 128259, end_token_ids: list[int] = <factory>, stop_token_id: int = 128258, pad_token_id: int = 128263, custom_token_base_id: int = 128256, snac_model_id: str = 'hubertsiuzdak/snac_24khz', tokens_per_frame: int = 7, sample_rate: int = 24000, snac_hop_samples: int = 2048, snac_chunk_schedule: tuple[int, ...]=(4, 2, 2, 2), snac_chunk_frames: int = 4, snac_left_context_frames: int = 3, snac_right_context_frames: int = 2, temperature: float = 0.6, top_p: float = 0.8, repetition_penalty: float = 1.3, ignore_eos: bool = False, max_new_tokens: int = 4096, available_voices: list[str] = <factory>, default_voice: str = 'tara')

Bases: object

Parameters:

• num_hidden_layers (int)

• num_attention_heads (int)

• num_key_value_heads (int)

• hidden_size (int)

• head_dim (int)

• intermediate_size (int)

• max_position_embeddings (int)

• rms_norm_eps (float)

• rope_theta (float)

• vocab_size (int)

• tie_word_embeddings (bool)

• rope_scaling (dict)

• start_token_id (int)

• end_token_ids (list[int])

• stop_token_id (int)

• pad_token_id (int)

• custom_token_base_id (int)

• snac_model_id (str)

• tokens_per_frame (int)

• sample_rate (int)

• snac_hop_samples (int)

• snac_chunk_schedule (tuple[int, ...])

• snac_chunk_frames (int)

• snac_left_context_frames (int)

• snac_right_context_frames (int)

• temperature (float)

• top_p (float)

• repetition_penalty (float)

• ignore_eos (bool)

• max_new_tokens (int)

• available_voices (list[str])

• default_voice (str)

available_voices: list[str]

custom_token_base_id: int = 128256

default_voice: str = 'tara'

end_token_ids: list[int]

head_dim: int = 128

hidden_size: int = 3072

ignore_eos: bool = False

intermediate_size: int = 8192

max_new_tokens: int = 4096

max_position_embeddings: int = 131072

num_attention_heads: int = 24

num_hidden_layers: int = 28

num_key_value_heads: int = 8

pad_token_id: int = 128263

repetition_penalty: float = 1.3

rms_norm_eps: float = 1e-05

rope_scaling: dict

rope_theta: float = 500000.0

sample_rate: int = 24000

snac_chunk_frames: int = 4

snac_chunk_policy()

The LLM -> SNAC edge’s chunking, in tokens.

A chunk’s leading context_items are its left context plus the previous chunk’s right context, which this chunk emits.

Return type:

ScheduledLeftContextChunkPolicy

snac_chunk_schedule: tuple[int, ...] = (4, 2, 2, 2)

snac_hop_samples: int = 2048

snac_left_context_frames: int = 3

snac_model_id: str = 'hubertsiuzdak/snac_24khz'

snac_right_context_frames: int = 2

snac_window_frames()

Every window size, in frames, a stream can hand the decoder.

The schedule’s windows and every final flush: the policy runs over each stream length until the windows repeat. The windows do not depend on when tokens arrive, only on how many there are. A window shorter than a frame (a stream shorter than a frame, or a final flush holding only the EOS token) decodes one silent frame, as in prepare_inputs.

Return type:

list[int]

start_token_id: int = 128259

stop_token_id: int = 128258

temperature: float = 0.6

tie_word_embeddings: bool = True

tokens_per_frame: int = 7

top_p: float = 0.8

vocab_size: int = 156940