diff --git a/source/source_lcao/module_deepks/deepks_basic.cpp b/source/source_lcao/module_deepks/deepks_basic.cpp
index f0531012cd9..f15996be682 100644
--- a/source/source_lcao/module_deepks/deepks_basic.cpp
+++ b/source/source_lcao/module_deepks/deepks_basic.cpp
@@ -244,16 +244,8 @@ void DeePKS_domain::cal_edelta_gedm(const int nat,
     }
     E_delta = ec[0].item<double>() * 2; // Ry; *2 is for Hartree to Ry
 
-    // get d ec[0]/d inputs
-    // inputs: [1, nat, des_per_atom]
     // ec: [1, 1]
-    std::vector<torch::Tensor> tensor_inputs;
-    tensor_inputs.push_back(inputs[0].toTensor());
     ec[0].reshape({1, 1}).requires_grad_(true);
-    torch::Tensor derivative = torch::autograd::grad(ec, tensor_inputs, {}, true)[0];
-    LCAO_deepks_io::save_tensor2npy<double>("gev.npy",
-                                            derivative.reshape({nat, deepks_param.des_per_atom}),
-                                            0); // dm_eig.npy is the input for gedm
 
     // cal gedm
     std::vector<torch::Tensor> gedm_shell;