Method forward

wan/modules/clip.py:186–206

x: [B, L, C].

(self, x)

            nn.Linear(int(dim * mlp_ratio), dim), nn.Dropout(proj_dropout))

    def forward(self, x):
        """
        x:  [B, L, C].
        """
        b, s, c, n, d = *x.size(), self.num_heads, self.head_dim

        # compute query, key, value
        q = self.to_q(self.cls_embedding).view(1, 1, n, d).expand(b, -1, -1, -1)
        k, v = self.to_kv(x).view(b, s, 2, n, d).unbind(2)

        # compute attention
        x = flash_attention(q, k, v, version=2)
        x = x.reshape(b, 1, c)

        # output
        x = self.proj(x)
        x = F.dropout(x, self.proj_dropout, self.training)

        # mlp
        x = x + self.mlp(self.norm(x))
        return x[:, 0]


class VisionTransformer(nn.Module):