Skip to content

lingbot_video

Architecture configuration for LingBot-Video Dense and MoE DiTs.

Classes

fastvideo.configs.models.dits.lingbot_video.LingBotVideoArchConfig dataclass 

LingBotVideoArchConfig(stacked_params_mapping: list[tuple[str, str, str]] = list(), _fsdp_shard_conditions: list = (lambda: [_is_lingbot_video_block])(), _compile_conditions: list = list(), param_names_mapping: dict = (lambda: {'^(.*)$': '\\1'})(), reverse_param_names_mapping: dict = dict(), lora_param_names_mapping: dict = dict(), cast_prompt_embeds_to_dit_dtype: bool = False, _supported_attention_backends: tuple[AttentionBackendEnum, ...] = (TORCH_SDPA, FLASH_ATTN), hidden_size: int = 2048, num_attention_heads: int = 16, num_channels_latents: int = 0, in_channels: int = 16, out_channels: int = 16, exclude_lora_layers: list[str] = list(), boundary_ratio: float | None = None, patch_size: tuple[int, int, int] = (1, 2, 2), depth: int = 24, intermediate_size: int = 6144, text_dim: int = 2560, freq_dim: int = 256, norm_eps: float = 1e-06, rope_theta: float = 256.0, axes_dims: tuple[int, int, int] = (32, 48, 48), axes_lens: tuple[int, int, int] = (8192, 1024, 1024), qkv_bias: bool = False, out_bias: bool = True, patch_embed_bias: bool = True, timestep_mlp_bias: bool = True, num_experts: int = 0, num_experts_per_tok: int = 8, moe_intermediate_size: int = 512, decoder_sparse_step: int = 1, mlp_only_layers: tuple[int, ...] = (), n_shared_experts: int | None = None, score_func: str = 'sigmoid', norm_topk_prob: bool = True, n_group: int | None = None, topk_group: int | None = None, routed_scaling_factor: float = 1.0)

Bases: DiTArchConfig

One-to-one representation of the released transformer config JSON.

Methods:

fastvideo.configs.models.dits.lingbot_video.LingBotVideoArchConfig.__post_init__ 
__post_init__() -> None

Populate FastVideo loader fields from the released architecture.

Source code in fastvideo/configs/models/dits/lingbot_video.py 
59
60
61
62
63
def __post_init__(self) -> None:
    """Populate FastVideo loader fields from the released architecture."""
    super().__post_init__()
    self.num_channels_latents = self.in_channels
    self.attention_head_dim = self.hidden_size // self.num_attention_heads

fastvideo.configs.models.dits.lingbot_video.LingBotVideoConfig dataclass 

LingBotVideoConfig(arch_config: DiTArchConfig = LingBotVideoArchConfig(), prefix: str = '', quant_config: QuantizationConfig | None = None)

Bases: DiTConfig

FastVideo component configuration for LingBot-Video transformers.