mstar.model.omnivoice.config

Configuration for OmniVoice (k2-fsa/OmniVoice).

OmniVoice is a masked-diffusion TTS model, not an autoregressive one. A Qwen3-0.6B body runs bidirectionally over a fixed-length canvas laid out as

[ style tokens | text tokens | (ref audio tokens) | target MASK tokens ]

where every position carries num_audio_codebook rows. Text positions repeat the same token id across all rows; audio positions hold one codebook entry per row. Generation unmasks the target region over num_step iterations and the Higgs-Audio-v2 codec turns the finished token canvas into a waveform.

The architecture values are facts of the checkpoint’s config.json, hardcoded so constructing the model never touches the network. OmniVoiceModel._refresh_checkpoint_defaults re-reads the ones that would silently break parity if the checkpoint drifted.

Classes

OmniVoiceBackboneConfig([hidden_size, ...])	The Qwen3 body, read bidirectionally.
OmniVoiceConfig([backbone, generation, ...])	OmniVoice model configuration.
OmniVoiceGenerationDefaults([num_step, ...])	Per-request decoding knobs and their defaults.

class mstar.model.omnivoice.config.OmniVoiceBackboneConfig(hidden_size=1024, num_hidden_layers=28, num_attention_heads=16, num_key_value_heads=8, head_dim=128, intermediate_size=3072, vocab_size=151676, rms_norm_eps=1e-06, rope_theta=1000000.0, max_position_embeddings=40960, tie_word_embeddings=True)

Bases: object

The Qwen3 body, read bidirectionally.

Identical in shape to Qwen3ForCausalLM at 0.6B, with two differences that matter at serving time: attention is non-causal (every position sees the whole document, prefix included), and there is no KV cache — the target region changes every diffusion step, and because the prefix attends into that region its hidden states change too, so nothing is reusable across steps. use_cache in the checkpoint config is vestigial.

Parameters:

• hidden_size (int)

• num_hidden_layers (int)

• num_attention_heads (int)

• num_key_value_heads (int)

• head_dim (int)

• intermediate_size (int)

• vocab_size (int)

• rms_norm_eps (float)

• rope_theta (float)

• max_position_embeddings (int)

• tie_word_embeddings (bool)

head_dim: int = 128

hidden_size: int = 1024

intermediate_size: int = 3072

max_position_embeddings: int = 40960

num_attention_heads: int = 16

num_hidden_layers: int = 28

num_key_value_heads: int = 8

rms_norm_eps: float = 1e-06

rope_theta: float = 1000000.0

tie_word_embeddings: bool = True

vocab_size: int = 151676

class mstar.model.omnivoice.config.OmniVoiceConfig(backbone=<factory>, generation=<factory>, num_audio_codebook=8, audio_vocab_size=1025, audio_mask_id=1024, sample_rate=24000, frame_rate=25.0, max_target_seconds=30.0, max_batch_size=16, max_packed_tokens=24576, load_dtype='float16', checkpoint_dtype=None)

Bases: object

OmniVoice model configuration.

Parameters:

• backbone (OmniVoiceBackboneConfig)

• generation (OmniVoiceGenerationDefaults)

• num_audio_codebook (int)

• audio_vocab_size (int)

• audio_mask_id (int)

• sample_rate (int)

• frame_rate (float)

• max_target_seconds (float)

• max_batch_size (int)

• max_packed_tokens (int)

• load_dtype (str)

• checkpoint_dtype (str | None)

property audio_embedding_rows: int

Rows in audio_embeddings — one codebook’s vocab per row block.

audio_mask_id: int = 1024

audio_vocab_size: int = 1025

backbone: OmniVoiceBackboneConfig

checkpoint_dtype: str | None = None

frame_rate: float = 25.0

generation: OmniVoiceGenerationDefaults

load_dtype: str = 'float16'

max_batch_size: int = 16

max_packed_tokens: int = 24576

max_target_seconds: float = 30.0

num_audio_codebook: int = 8

sample_rate: int = 24000

seconds_for_target_tokens(tokens)

Parameters:

tokens (int)

Return type:

float

target_tokens_for_seconds(seconds)

Parameters:

seconds (float)

Return type:

int

class mstar.model.omnivoice.config.OmniVoiceGenerationDefaults(num_step=32, guidance_scale=2.0, t_shift=0.1, layer_penalty_factor=5.0, position_temperature=5.0, class_temperature=0.0, denoise=True, max_num_step=64)

Bases: object

Per-request decoding knobs and their defaults.

Every one of these is per-request: they ride step_metadata and are resolved once in OmniVoiceModel.get_initial_forward_pass_args. This is the reason to serve OmniVoice on a graph runtime at all — a fixed-pipeline port has to pick one value for the whole server.

Parameters:

• num_step (int)

• guidance_scale (float)

• t_shift (float)

• layer_penalty_factor (float)

• position_temperature (float)

• class_temperature (float)

• denoise (bool)

• max_num_step (int)

class_temperature: float = 0.0

denoise: bool = True

guidance_scale: float = 2.0

layer_penalty_factor: float = 5.0

max_num_step: int = 64

num_step: int = 32

position_temperature: float = 5.0

t_shift: float = 0.1