def forward()

in janus/models/siglip_vit.py [0:0]

• 26 lines of code
• 3 McCabe index (conditional complexity)

    def forward(self, x: torch.Tensor) -> torch.Tensor:
        B, N, C = x.shape
        qkv = (
            self.qkv(x)
            .reshape(B, N, 3, self.num_heads, self.head_dim)
            .permute(2, 0, 3, 1, 4)
        )
        q, k, v = qkv.unbind(0)
        q, k = self.q_norm(q), self.k_norm(k)

        if self.fused_attn:
            x = F.scaled_dot_product_attention(
                q,
                k,
                v,
                dropout_p=self.attn_drop.p if self.training else 0.0,
            )
        else:
            q = q * self.scale
            attn = q @ k.transpose(-2, -1)
            attn = attn.softmax(dim=-1)
            attn = self.attn_drop(attn)
            x = attn @ v

        x = x.transpose(1, 2).reshape(B, N, C)
        x = self.proj(x)
        x = self.proj_drop(x)
        return x