feat(mamba): extract last time step from Mamba2Encoder output

layers/MambaSeries.py

@@ -23,7 +23,7 @@ class Mamba2Encoder(nn.Module):
         expand=2,
         headdim=64,
         # 堆叠层数
-        n_layers=2,
+        n_layers=1,
     ):
         super().__init__()
         self.patch_num = patch_num
@@ -57,3 +57,13 @@ class Mamba2Encoder(nn.Module):
         u = x.reshape(bs * n_vars, patch_num, self.d_model)  # u: [bs*nvars, patch_num, d_model]
 
         # 3) 通过 n_layers 层 Mamba2 进行建模（在 patch_num 维度上）
+        for m in self.mambas:
+            u = m(u)  # 形状保持 [bs*nvars, patch_num, d_model]
+
+        # 4) 仅取最后一个时间步
+        y_last = u[:, -1, :]  # y_last: [bs*nvars, d_model]
+
+        # 5) 还原回 (bs, nvars, d_model)
+        y_last = y_last.view(bs, n_vars, self.d_model)  # y_last: [bs, nvars, d_model]
+
+        return y_last  # [bs, nvars, d_model]