Corrected device assignment error

modeling_chatglm.py

@@ -505,7 +505,7 @@ class NativeSparseAttention(CoreAttention):
             y_seqlens[seqlens < self.kernel_size] = 0
             cmp_seqlens = torch.cat(
                 [
-                    torch.zeros(1, dtype=torch.int32, device="cuda"),
+                    torch.zeros(1, dtype=torch.int32, device=y_seqlens.device),
                     torch.cumsum(y_seqlens, dim=0),
                 ],
                 dim=0,