vllm_omni.diffusion.models.magi2.configuration_magi2 ¶

audio_guidance_scale: float = 7.0

audio_latent_channels: int = 64

audio_latent_fps: float = 25.0

audio_sample_rate: int = 44100

duration_seconds: float = 10.0

fps: float = 12.5

output_frames: int = 125

patch_size: tuple[int, int, int] = (1, 1, 1)

preview_steps: int = 100

shift: float = 7.0

video_guidance_scale: float = 5.0

video_latent_channels: int = 48

video_vae_stride: tuple[int, int, int] = (8, 16, 16)

alpha_init: float = 0.01

enabled: bool = True

num_streams: int = 4

sinkhorn_epsilon: float = 1e-12

sinkhorn_iterations: int = 20

expert_intermediate_size: int = 1280

layers: tuple[int, ...] = tuple(range(2, 38))

modality_shared_expert_intermediate_size: int = 1280

normalize_routing_weights: bool = True

num_experts: int = 256

num_heads: int = 12

routing_scale: float = 4.9

score_function: str = 'sigmoid'

shared_expert_intermediate_size: int = 1280

top_k: int = 6

attention_gating: bool = True

attention_sink_tokens: int = 1

attention_softcap: float = -1.0

audio_in_channels: int = 64

head_dim: int = 128

hidden_size: int = 3072

intermediate_factor: float = 4.0

mhc: Magi2MHCConfig = field(default_factory=Magi2MHCConfig)

moe: Magi2MoEConfig = field(default_factory=Magi2MoEConfig)

multimodal_layers: tuple[int, ...] = (0, 1, 38, 39)

num_attention_heads: int

num_heads_kv: int

num_heads_q: int

num_layers: int = 40

num_query_groups: int = 24

params_dtype: dtype = torch.bfloat16

text_in_channels: int = 5120

video_in_channels: int = 48

virtual_width: int