Skip to content

vllm_omni.model_executor.models.minimax_music3.constants

Inference-contract constants for MiniMax Music 3.

These values are fixed by the checkpoint, not tuning knobs. Changing any of them changes the audio, so they are declared once here and imported everywhere rather than re-derived per call site.

AR_CFG_SCALE module-attribute 

AR_CFG_SCALE = 1.5

AR_CFG_TOP_K module-attribute 

AR_CFG_TOP_K = 50

AR_CHUNK_FRAMES module-attribute 

AR_CHUNK_FRAMES = 200

AR_CHUNK_HOP_FRAMES module-attribute 

AR_CHUNK_HOP_FRAMES = 100

AR_HIDDEN_SIZE module-attribute 

AR_HIDDEN_SIZE = BACKBONE_HIDDEN_SIZE * NUM_CODEBOOKS

AR_TOP_K module-attribute 

AR_TOP_K = 50

AUDIO_FRAME_RATE module-attribute 

AUDIO_FRAME_RATE = 25

AUDIO_VOCAB_SIZE module-attribute 

AUDIO_VOCAB_SIZE = 1024

BACKBONE_HIDDEN_SIZE module-attribute 

BACKBONE_HIDDEN_SIZE = 4096

BLEND_MEL_FRAMES module-attribute 

BLEND_MEL_FRAMES = HOP_MEL_FRAMES // 4

C0_VOCAB_SIZE module-attribute 

C0_VOCAB_SIZE = 16384

CONDITION_INPUT_HOP module-attribute 

CONDITION_INPUT_HOP = 960

CONDITION_INPUT_SAMPLE_RATE module-attribute 

CONDITION_INPUT_SAMPLE_RATE = 24000

CONDITION_OUTPUT_HOP module-attribute 

CONDITION_OUTPUT_HOP = 512

CONDITION_OUTPUT_SAMPLE_RATE module-attribute 

CONDITION_OUTPUT_SAMPLE_RATE = 44100

DAV_HOP_SAMPLES module-attribute 

DAV_HOP_SAMPLES = 512

DAV_SAMPLE_RATE module-attribute 

DAV_SAMPLE_RATE = 44100

DEFAULT_DIT_CFG_SCALE module-attribute 

DEFAULT_DIT_CFG_SCALE = 1.7

DEFAULT_DIT_STEPS module-attribute 

DEFAULT_DIT_STEPS = 30

DEFAULT_MAX_AUDIO_FRAMES module-attribute 

DEFAULT_MAX_AUDIO_FRAMES = 9000

DIT_LATENT_CHANNELS module-attribute 

DIT_LATENT_CHANNELS = 128

HOP_MEL_FRAMES module-attribute 

HOP_MEL_FRAMES = 344

MAX_AUDIO_FRAMES module-attribute 

MAX_AUDIO_FRAMES = 9000

MAX_PROMPT_TOKENS module-attribute 

MAX_PROMPT_TOKENS = 5000

NUM_CODEBOOKS module-attribute 

NUM_CODEBOOKS = 8

OUTPUT_CHANNELS module-attribute 

OUTPUT_CHANNELS = 2

OUTPUT_SAMPLE_RATE module-attribute 

OUTPUT_SAMPLE_RATE = 32000

SUPPORTED_ACOUSTIC_DTYPES module-attribute 

SUPPORTED_ACOUSTIC_DTYPES = frozenset(
    {"float32", "bfloat16"}
)