Class CogVideoXBlock

architecture/cogvideox_transformer_3d.py:42–161

r""" Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model. Parameters: dim (`int`): The number of channels in the input and output. num_attention_heads (`int`): The number of heads to use for multi-head attention. att

@maybe_allow_in_graph
class CogVideoXBlock(nn.Module):
    r"""
    Transformer block used in [CogVideoX](https://github.com/THUDM/CogVideo) model.

    Parameters:
        dim (`int`):
            The number of channels in the input and output.
        num_attention_heads (`int`):
            The number of heads to use for multi-head attention.
        attention_head_dim (`int`):
            The number of channels in each head.
        time_embed_dim (`int`):
            The number of channels in timestep embedding.
        dropout (`float`, defaults to `0.0`):
            The dropout probability to use.
        activation_fn (`str`, defaults to `"gelu-approximate"`):
            Activation function to be used in feed-forward.
        attention_bias (`bool`, defaults to `False`):
            Whether or not to use bias in attention projection layers.
        qk_norm (`bool`, defaults to `True`):
            Whether or not to use normalization after query and key projections in Attention.
        norm_elementwise_affine (`bool`, defaults to `True`):
            Whether to use learnable elementwise affine parameters for normalization.
        norm_eps (`float`, defaults to `1e-5`):
            Epsilon value for normalization layers.
        final_dropout (`bool` defaults to `False`):
            Whether to apply a final dropout after the last feed-forward layer.
        ff_inner_dim (`int`, *optional*, defaults to `None`):
            Custom hidden dimension of Feed-forward layer. If not provided, `4 * dim` is used.
        ff_bias (`bool`, defaults to `True`):
            Whether or not to use bias in Feed-forward layer.
        attention_out_bias (`bool`, defaults to `True`):
            Whether or not to use bias in Attention output projection layer.
    """

    def __init__(
        self,
        dim: int,
        num_attention_heads: int,
        attention_head_dim: int,
        time_embed_dim: int,
        dropout: float = 0.0,
        activation_fn: str = "gelu-approximate",
        attention_bias: bool = False,
        qk_norm: bool = True,
        norm_elementwise_affine: bool = True,
        norm_eps: float = 1e-5,
        final_dropout: bool = True,
        ff_inner_dim: Optional[int] = None,
        ff_bias: bool = True,
        attention_out_bias: bool = True,
    ):
        super().__init__()

        # 1. Self Attention
        self.norm1 = CogVideoXLayerNormZero(time_embed_dim, dim, norm_elementwise_affine, norm_eps, bias=True)

        self.attn1 = Attention(