Corrected device assignment error

modeling_chatglm.py

@@ -506,7 +506,7 @@ class NativeSparseAttention(CoreAttention):
             cmp_seqlens = torch.cat(
                 [
                     torch.zeros(1, dtype=torch.int32, device=query_states.device),
-                    torch.cumsum(y_seqlens, dim=0, device=query_states.device),
+                    torch.cumsum(y_seqlens, dim=0).to(query_states.device),
                 ],
                 dim=0,
             ).to(torch.int32)