diff --git a/source/source_base/module_container/ATen/core/tensor_accessor.h b/source/source_base/module_container/ATen/core/tensor_accessor.h
index 9fbea6f0b81..80e85881953 100644
--- a/source/source_base/module_container/ATen/core/tensor_accessor.h
+++ b/source/source_base/module_container/ATen/core/tensor_accessor.h
@@ -1,10 +1,9 @@
 #ifndef ATEN_CORE_TENSOR_ACCESSOR_H_
 #define ATEN_CORE_TENSOR_ACCESSOR_H_
 
-#include <cstddef> // Include the <cstddef> header file to define size_t
+#include <cstddef>
 #include <cstdint>
 #include <base/macros/macros.h>
-#include <base/utils/array_ref.h>
 
 namespace container {
 
@@ -33,12 +32,17 @@ class TensorAccessorBase {
             const index_t* strides)
             : data_(data), sizes_(sizes), strides_(strides) {}
     
-    AT_HOST int_array_ref sizes() const {
-        return {sizes_, N};
+    // NOTE: Previously returned int_array_ref (from <base/utils/array_ref.h>),
+    // which supported .size(), iteration, etc. Now returns raw const pointer.
+    // If you need array_ref functionality, include <base/utils/array_ref.h> directly
+    // and wrap the result: int_array_ref(accessor.sizes(), N).
+    // mohan add 20260605
+    AT_HOST_DEVICE const index_t* sizes() const {
+        return sizes_;
     }
 
-    AT_HOST int_array_ref strides() const {
-        return {strides_, N};
+    AT_HOST_DEVICE const index_t* strides() const {
+        return strides_;
     }
 
     AT_HOST_DEVICE index_t stride(index_t idx) const {
diff --git a/source/source_base/module_container/ATen/core/tensor_enums.h b/source/source_base/module_container/ATen/core/tensor_enums.h
new file mode 100644
index 00000000000..4e3df3cb154
--- /dev/null
+++ b/source/source_base/module_container/ATen/core/tensor_enums.h
@@ -0,0 +1,96 @@
+// Extracted from tensor_types.h to break the circular dependency:
+// macros.h -> tensor_types.h -> cuda.h/rocm.h -> macros.h
+// This file provides only enums and template specializations needed by macros.h,
+// without any GPU library dependencies.
+// mohan add 20260605
+#ifndef ATEN_CORE_TENSOR_ENUMS_H_
+#define ATEN_CORE_TENSOR_ENUMS_H_
+
+#include <complex>
+
+#include "source_base/module_device/types.h"
+
+namespace container {
+
+enum class DataType {
+    DT_INVALID = 0,
+    DT_FLOAT = 1,
+    DT_DOUBLE = 2,
+    DT_INT = 3,
+    DT_INT64 = 4,
+    DT_COMPLEX = 5,
+    DT_COMPLEX_DOUBLE = 6,
+};
+
+struct DEVICE_CPU {};
+struct DEVICE_GPU {};
+
+enum class DeviceType {
+    UnKnown = 0,
+    CpuDevice = 1,
+    GpuDevice = 2,
+};
+
+template <typename T>
+struct DeviceTypeToEnum {
+    static constexpr DeviceType value = {};
+};
+
+template <>
+struct DeviceTypeToEnum<DEVICE_CPU> {
+    static constexpr DeviceType value = DeviceType::CpuDevice;
+};
+
+template <>
+struct DeviceTypeToEnum<DEVICE_GPU> {
+    static constexpr DeviceType value = DeviceType::GpuDevice;
+};
+
+template <>
+struct DeviceTypeToEnum<base_device::DEVICE_CPU> {
+    static constexpr DeviceType value = DeviceType::CpuDevice;
+};
+
+template <>
+struct DeviceTypeToEnum<base_device::DEVICE_GPU> {
+    static constexpr DeviceType value = DeviceType::GpuDevice;
+};
+
+template <typename T>
+struct DataTypeToEnum {
+    static constexpr DataType value = {};
+};
+
+template <>
+struct DataTypeToEnum<int> {
+    static constexpr DataType value = DataType::DT_INT;
+};
+
+template <>
+struct DataTypeToEnum<float> {
+    static constexpr DataType value = DataType::DT_FLOAT;
+};
+
+template <>
+struct DataTypeToEnum<double> {
+    static constexpr DataType value = DataType::DT_DOUBLE;
+};
+
+template <>
+struct DataTypeToEnum<int64_t> {
+    static constexpr DataType value = DataType::DT_INT64;
+};
+
+template <>
+struct DataTypeToEnum<std::complex<float>> {
+    static constexpr DataType value = DataType::DT_COMPLEX;
+};
+
+template <>
+struct DataTypeToEnum<std::complex<double>> {
+    static constexpr DataType value = DataType::DT_COMPLEX_DOUBLE;
+};
+
+} // namespace container
+
+#endif // ATEN_CORE_TENSOR_ENUMS_H_
diff --git a/source/source_base/module_container/ATen/core/tensor_types.h b/source/source_base/module_container/ATen/core/tensor_types.h
index b87907825ac..5dbdcc1e3e3 100644
--- a/source/source_base/module_container/ATen/core/tensor_types.h
+++ b/source/source_base/module_container/ATen/core/tensor_types.h
@@ -5,26 +5,23 @@
 #ifndef ATEN_CORE_TENSOR_TYPES_H_
 #define ATEN_CORE_TENSOR_TYPES_H_
 
-#include <regex>
 #include <string>
-#include <vector>
-#include <numeric>
 #include <complex>
-#include <utility>
-#include <iomanip>
 #include <iostream>
 #include <stdexcept>
-#include <algorithm>
-#include <unordered_map>
-#include <initializer_list>
 
 #include "source_base/module_device/types.h"
-
-#if defined(__CUDACC__)
-#include <base/macros/cuda.h>
-#elif defined(__HIPCC__)
-#include <base/macros/rocm.h>
-#endif // defined(__CUDACC__) || defined(__HIPCC__)
+#include "ATen/core/tensor_enums.h"
+
+// NOTE: Previously this file included <base/macros/cuda.h> or <base/macros/rocm.h>,
+// which transitively provided GetTypeThrust, GetTypeCuda, GetTypeRocm, etc.
+// Now replaced with <thrust/complex.h> which only provides thrust::complex type.
+// If you need GetTypeThrust/GetTypeCuda/GetTypeRocm, include <base/macros/cuda.h>
+// or <base/macros/rocm.h> directly in your .cpp file.
+// mohan add 20260605
+#if defined(__CUDACC__) || defined(__HIPCC__)
+#include <thrust/complex.h>
+#endif
 
 namespace container {
 
@@ -41,41 +38,6 @@ static inline bool element_compare(T& a, T& b) {
     }
 }
 
-/**
-@brief Enumeration of data types for tensors.
-The DataType enum lists the supported data types for tensors. Each data type
-is identified by a unique value. The DT_INVALID value is reserved for invalid
-data types.
-*/
-enum class DataType {
-    DT_INVALID = 0, ///< Invalid data type */
-    DT_FLOAT = 1, ///< Single-precision floating point */
-    DT_DOUBLE = 2, ///< Double-precision floating point */
-    DT_INT = 3, ///< 32-bit integer */
-    DT_INT64 = 4, ///< 64-bit integer */
-    DT_COMPLEX = 5, ///< 32-bit complex */
-    DT_COMPLEX_DOUBLE = 6, /**< 64-bit complex */
-// ... other data types
-};
-
-/**
- *@struct DEVICE_CPU, DEVICE_GPU
- *@brief A tag type for identifying CPU and GPU devices.
-*/
-struct DEVICE_CPU;
-struct DEVICE_GPU;
-
-struct DEVICE_CPU {};
-struct DEVICE_GPU {};
-/**
- * @brief The type of memory used by an allocator.
- */
-enum class DeviceType {
-    UnKnown = 0,  ///< Memory type is unknown.
-    CpuDevice = 1,     ///< Memory type is CPU.
-    GpuDevice = 2,     ///< Memory type is GPU(CUDA or ROCm).
-};
-
 /**
  * @brief Template struct to determine the return type based on the input type.
  *
@@ -141,85 +103,6 @@ struct ContainerToPsi<container::DEVICE_GPU> {
     using type = base_device::DEVICE_GPU; /**< The return type specialization for std::complex<double>. */
 };
 
-
-/**
- * @brief Template struct for mapping a Device Type to its corresponding enum value.
- *
- * @param T The DataType to map to its enum value.
- *
- * @return The enumeration value corresponding to the data type.
- *  This method uses template specialization to map each supported data type to its
- *  corresponding enumeration value. If the template argument T is not a supported
- *  data type, this method will cause a compile-time error.
- *  Example usage:
- *      DataTypeToEnum<float>::value; // Returns DataType::DT_FLOAT
- */
-template <typename T>
-struct DeviceTypeToEnum {
-    static constexpr DeviceType value = {};
-};
-// Specializations of DeviceTypeToEnum for supported devices.
-template <>
-struct DeviceTypeToEnum<DEVICE_CPU> {
-    static constexpr DeviceType value = DeviceType::CpuDevice;
-};
-template <>
-struct DeviceTypeToEnum<DEVICE_GPU> {
-    static constexpr DeviceType value = DeviceType::GpuDevice;
-};
-template <>
-struct DeviceTypeToEnum<base_device::DEVICE_CPU>
-{
-    static constexpr DeviceType value = DeviceType::CpuDevice;
-};
-template <>
-struct DeviceTypeToEnum<base_device::DEVICE_GPU>
-{
-    static constexpr DeviceType value = DeviceType::GpuDevice;
-};
-
-/**
- * @brief Template struct for mapping a DataType to its corresponding enum value.
- *
- * @param T The DataType to map to its enum value.
- *
- * @return The enumeration value corresponding to the data type.
- *  This method uses template specialization to map each supported data type to its
- *  corresponding enumeration value. If the template argument T is not a supported
- *  data type, this method will cause a compile-time error.
- *  Example usage:
- *      DataTypeToEnum<float>::value; // Returns DataType::DT_FLOAT
- */
-template <typename T>
-struct DataTypeToEnum {
-    static constexpr DataType value = {};
-};
-// Specializations of DataTypeToEnum for supported types.
-template <>
-struct DataTypeToEnum<int> {
-    static constexpr DataType value = DataType::DT_INT;
-};
-template <>
-struct DataTypeToEnum<float> {
-    static constexpr DataType value = DataType::DT_FLOAT;
-};
-template <>
-struct DataTypeToEnum<double> {
-    static constexpr DataType value = DataType::DT_DOUBLE;
-};
-template <>
-struct DataTypeToEnum<int64_t> {
-    static constexpr DataType value = DataType::DT_INT64;
-};
-template <>
-struct DataTypeToEnum<std::complex<float>> {
-    static constexpr DataType value = DataType::DT_COMPLEX;
-};
-template <>
-struct DataTypeToEnum<std::complex<double>> {
-    static constexpr DataType value = DataType::DT_COMPLEX_DOUBLE;
-};
-
 #if defined(__CUDACC__) || defined(__HIPCC__)
 template <>
 struct DataTypeToEnum<thrust::complex<float>> {
diff --git a/source/source_base/module_container/ATen/core/tensor_utils.h b/source/source_base/module_container/ATen/core/tensor_utils.h
index 115ddfedb73..20f567d2843 100644
--- a/source/source_base/module_container/ATen/core/tensor_utils.h
+++ b/source/source_base/module_container/ATen/core/tensor_utils.h
@@ -1,6 +1,8 @@
 #ifndef ATEN_CORE_TENSOR_UTILS_H_
 #define ATEN_CORE_TENSOR_UTILS_H_
 
+#include <iomanip>
+
 #include <ATen/core/tensor.h>
 #include <ATen/core/tensor_shape.h>
 
diff --git a/source/source_base/module_container/ATen/ops/einsum_op.cpp b/source/source_base/module_container/ATen/ops/einsum_op.cpp
index 40440db9407..eeb60aea2c2 100644
--- a/source/source_base/module_container/ATen/ops/einsum_op.cpp
+++ b/source/source_base/module_container/ATen/ops/einsum_op.cpp
@@ -2,6 +2,9 @@
 
 #include <array>
 #include <algorithm>
+#include <functional>
+#include <numeric>
+#include <regex>
 
 #include <ATen/ops/linalg_op.h>
 #include <ATen/core/tensor_types.h>
diff --git a/source/source_base/module_container/ATen/ops/einsum_op.h b/source/source_base/module_container/ATen/ops/einsum_op.h
index 6850f0445cb..feabf08d58b 100644
--- a/source/source_base/module_container/ATen/ops/einsum_op.h
+++ b/source/source_base/module_container/ATen/ops/einsum_op.h
@@ -1,6 +1,8 @@
 #ifndef ATEN_KERNELS_EINSUM_OP_H_
 #define ATEN_KERNELS_EINSUM_OP_H_
 
+#include <unordered_map>
+
 #include <ATen/core/tensor.h>
 #include <ATen/core/tensor_map.h>
 
diff --git a/source/source_base/module_container/base/macros/macros.h b/source/source_base/module_container/base/macros/macros.h
index 645c6aee896..fbcae6fe8a2 100644
--- a/source/source_base/module_container/base/macros/macros.h
+++ b/source/source_base/module_container/base/macros/macros.h
@@ -10,7 +10,14 @@
 #include <base/macros/rocm.h>
 #endif
 
-#include <ATen/core/tensor_types.h>
+// NOTE: Previously included <ATen/core/tensor_types.h> which pulled in
+// <base/macros/cuda.h> or <base/macros/rocm.h> transitively.
+// Now only includes <ATen/core/tensor_enums.h> which provides DataType,
+// DeviceType, DataTypeToEnum, DeviceTypeToEnum without GPU dependencies.
+// If you need GetTypeReal, PsiToContainer, ContainerToPsi, etc.,
+// include <ATen/core/tensor_types.h> directly.
+// mohan add 20260605
+#include <ATen/core/tensor_enums.h>
 
 #define DISALLOW_COPY_AND_ASSIGN(TypeName) \
   TypeName(const TypeName&) = delete;      \
diff --git a/source/source_base/module_device/output_device.cpp b/source/source_base/module_device/output_device.cpp
index 3d4587d4a99..db41786b969 100644
--- a/source/source_base/module_device/output_device.cpp
+++ b/source/source_base/module_device/output_device.cpp
@@ -7,6 +7,7 @@
 #include <base/macros/macros.h>
 #include <cstring>
 #include <iostream>
+#include <numeric>
 #include <set>
 #ifdef __MPI
 #include "mpi.h"
diff --git a/source/source_esolver/esolver_ks_lcao.cpp b/source/source_esolver/esolver_ks_lcao.cpp
index 08499df2bdf..5b864866041 100644
--- a/source/source_esolver/esolver_ks_lcao.cpp
+++ b/source/source_esolver/esolver_ks_lcao.cpp
@@ -9,6 +9,7 @@
 #include "source_estate/module_charge/symmetry_rho.h"
 #include "source_lcao/LCAO_domain.h" // need DeePKS_init
 #include "source_lcao/FORCE_STRESS.h"
+#include "source_lcao/module_gint/gint.h"
 #include "source_estate/elecstate_lcao.h"
 #include "source_lcao/hamilt_lcao.h"
 #include "source_hsolver/hsolver_lcao.h"
diff --git a/source/source_esolver/esolver_ks_lcao.h b/source/source_esolver/esolver_ks_lcao.h
index bfd80e50fe3..03ef22c5658 100644
--- a/source/source_esolver/esolver_ks_lcao.h
+++ b/source/source_esolver/esolver_ks_lcao.h
@@ -4,7 +4,6 @@
 #include "esolver_ks.h"
 #include "source_lcao/record_adj.h" // adjacent atoms
 #include "source_basis/module_nao/two_center_bundle.h" // nao basis
-#include "source_lcao/module_gint/gint.h" // gint
 #include "source_lcao/module_gint/gint_info.h"
 #include "source_estate/module_charge/gint_precision_controller.h"
 #include "source_lcao/setup_deepks.h" // for deepks, mohan add 20251008
diff --git a/source/source_esolver/lcao_others.cpp b/source/source_esolver/lcao_others.cpp
index 57174c47ceb..1c08e815f7a 100644
--- a/source/source_esolver/lcao_others.cpp
+++ b/source/source_esolver/lcao_others.cpp
@@ -3,6 +3,7 @@
 #include "source_estate/module_charge/symmetry_rho.h"
 #include "source_lcao/hamilt_lcao.h"
 #include "source_lcao/module_dftu/dftu.h"
+#include "source_lcao/module_gint/gint.h"
 #include "source_base/formatter.h"
 #include "source_base/timer.h"
 #include "source_cell/module_neighbor/sltk_atom_arrange.h"
diff --git a/source/source_estate/module_pot/H_TDDFT_pw.cpp b/source/source_estate/module_pot/H_TDDFT_pw.cpp
index babece4d784..422045eab1c 100644
--- a/source/source_estate/module_pot/H_TDDFT_pw.cpp
+++ b/source/source_estate/module_pot/H_TDDFT_pw.cpp
@@ -3,9 +3,7 @@
 #include "source_base/constants.h"
 #include "source_base/math_integral.h"
 #include "source_base/timer.h"
-#include "source_io/module_parameter/input_conv.h"
 #include "source_io/module_parameter/parameter.h"
-#include "source_lcao/module_rt/evolve_elec.h"
 
 namespace elecstate
 {
diff --git a/source/source_hsolver/diago_bpcg.h b/source/source_hsolver/diago_bpcg.h
index 27f528024ba..9eb64585df3 100644
--- a/source/source_hsolver/diago_bpcg.h
+++ b/source/source_hsolver/diago_bpcg.h
@@ -1,6 +1,8 @@
 #ifndef DIAGO_BPCG_H_
 #define DIAGO_BPCG_H_
 
+#include <functional>
+
 #include "source_base/kernels/math_kernel_op.h"
 #include "source_base/module_device/memory_op.h"
 #include "source_base/module_device/types.h"
diff --git a/source/source_io/module_ctrl/ctrl_scf_lcao.cpp b/source/source_io/module_ctrl/ctrl_scf_lcao.cpp
index 7318a3f2b7b..30a220ef37d 100644
--- a/source/source_io/module_ctrl/ctrl_scf_lcao.cpp
+++ b/source/source_io/module_ctrl/ctrl_scf_lcao.cpp
@@ -12,7 +12,7 @@
 #include "../module_hs/cal_pLpR.h"                            // use AngularMomentumCalculator()
 #include "source_io/module_hs/output_mat_sparse.h"                   // use ModuleIO::output_mat_sparse()
 #include "../module_hs/write_HS_R.h"                          // use ModuleIO::write_hsr()
-#include "../module_mulliken/output_mulliken.h"                     // use cal_mag()
+#include "../module_mulliken/cal_mag.h"                          // use cal_mag()
 #include "../module_wannier/to_wannier90_lcao.h"                   // use toWannier90_LCAO
 #include "../module_wannier/to_wannier90_lcao_in_pw.h"             // use toWannier90_LCAO_IN_PW
 #include "../module_hs/write_HS.h"                            // use ModuleIO::write_hsk()
diff --git a/source/source_io/module_current/td_current_io_comm.cpp b/source/source_io/module_current/td_current_io_comm.cpp
index 66392a80394..a6e28ebaea9 100644
--- a/source/source_io/module_current/td_current_io_comm.cpp
+++ b/source/source_io/module_current/td_current_io_comm.cpp
@@ -7,10 +7,8 @@
 #include "source_base/timer.h"
 #include "source_base/tool_threading.h"
 #include "source_base/vector3.h"
-#include "source_estate/module_dm/cal_dm_psi.h"
 #include "source_estate/module_pot/H_TDDFT_pw.h"
 #include "source_io/module_parameter/parameter.h"
-#include "source_lcao/LCAO_domain.h"
 #include "source_lcao/module_hcontainer/hcontainer_funcs.h"
 #include "source_lcao/module_rt/td_folding.h"
 #include "source_lcao/module_rt/td_info.h"
diff --git a/source/source_io/module_mulliken/cal_mag.h b/source/source_io/module_mulliken/cal_mag.h
new file mode 100644
index 00000000000..b0ec41974a6
--- /dev/null
+++ b/source/source_io/module_mulliken/cal_mag.h
@@ -0,0 +1,115 @@
+#ifndef CAL_MAG_H
+#define CAL_MAG_H
+
+#include "source_basis/module_ao/parallel_orbitals.h"
+#include "source_basis/module_ao/ORB_read.h"
+#include "source_basis/module_nao/two_center_bundle.h"
+#include "source_cell/cell_index.h"
+#include "source_cell/klist.h"
+#include "source_cell/module_neighbor/sltk_grid_driver.h"
+#include "source_cell/unitcell.h"
+#include "source_estate/module_dm/density_matrix.h"
+#include "source_hamilt/hamilt.h"
+#include "source_io/module_mulliken/output_dmk.h"
+#include "source_io/module_mulliken/output_mulliken.h"
+#include "source_io/module_mulliken/output_sk.h"
+#include "source_io/module_parameter/parameter.h"
+#include "source_lcao/module_operator_lcao/dspin_lcao.h"
+
+#include <vector>
+
+namespace ModuleIO
+{
+
+template <typename TK>
+void cal_mag(Parallel_Orbitals* pv,
+             hamilt::Hamilt<TK>* p_ham,
+             K_Vectors& kv,
+             elecstate::DensityMatrix<TK,double>* dm,
+             const TwoCenterBundle& two_center_bundle,
+             const LCAO_Orbitals& orb,
+             UnitCell& ucell,
+             const Grid_Driver& gd,
+             const int istep,
+             const bool print)
+{
+    if (PARAM.inp.out_mul)
+    {
+        auto cell_index
+            = CellIndex(ucell.get_atomLabels(),
+			ucell.get_atomCounts(), ucell.get_lnchiCounts(), PARAM.inp.nspin);
+        auto out_s_k = ModuleIO::Output_Sk<TK>(p_ham, pv, PARAM.inp.nspin, kv.get_nks());
+        auto out_dm_k = ModuleIO::Output_DMK<TK>(dm, pv, PARAM.inp.nspin, kv.get_nks());
+
+        auto mulp = ModuleIO::Output_Mulliken<TK>(&(out_s_k),
+			&(out_dm_k), pv, &cell_index, kv.isk, PARAM.inp.nspin);
+        auto atom_chg = mulp.get_atom_chg();
+        ucell.atom_mulliken = mulp.get_atom_mulliken(atom_chg);
+        if (print && GlobalV::MY_RANK == 0)
+        {
+            cell_index.write_orb_info(PARAM.globalv.global_out_dir);
+            mulp.write(istep, PARAM.globalv.global_out_dir);
+            mulp.print_atom_mag(atom_chg, GlobalV::ofs_running);
+        }
+    }
+    if (PARAM.inp.onsite_radius > 0)
+    {
+        std::vector<std::vector<double>> atom_mag(ucell.nat, std::vector<double>(PARAM.inp.nspin, 0.0));
+        std::vector<ModuleBase::Vector3<int>> constrain(ucell.nat, ModuleBase::Vector3<int>(1, 1, 1));
+        const hamilt::HContainer<double>* dmr = dm->get_DMR_pointer(1);
+        std::vector<double> moments;
+        std::vector<double> mag_x(ucell.nat, 0.0);
+        std::vector<double> mag_y(ucell.nat, 0.0);
+        std::vector<double> mag_z(ucell.nat, 0.0);
+        auto atomLabels = ucell.get_atomLabels();
+
+        if(PARAM.inp.nspin == 2)
+        {
+            auto sc_lambda = new hamilt::DeltaSpin<hamilt::OperatorLCAO<TK, double>>(nullptr,
+		kv.kvec_d,
+		dynamic_cast<hamilt::HamiltLCAO<TK, double>*>(p_ham)->getHR(),
+		ucell,
+		&gd,
+		two_center_bundle.overlap_orb_onsite.get(),
+		orb.cutoffs());
+
+	    dm->switch_dmr(2);
+	    moments = sc_lambda->cal_moment(dmr, constrain);
+	    dm->switch_dmr(0);
+
+	    delete sc_lambda;
+
+            for(int iat=0;iat<ucell.nat;iat++)
+            {
+                atom_mag[iat][0] = 0.0;
+                atom_mag[iat][1] = moments[iat];
+            }
+        }
+        else if(PARAM.inp.nspin == 4)
+        {
+            auto sc_lambda = new hamilt::DeltaSpin<hamilt::OperatorLCAO<std::complex<double>, std::complex<double>>>(
+                nullptr,
+                kv.kvec_d,
+                dynamic_cast<hamilt::HamiltLCAO<std::complex<double>, std::complex<double>>*>(p_ham)->getHR(),
+                ucell,
+                &gd,
+                two_center_bundle.overlap_orb_onsite.get(),
+                orb.cutoffs());
+            moments = sc_lambda->cal_moment(dmr, constrain);
+            delete sc_lambda;
+
+            for(int iat=0;iat<ucell.nat;iat++)
+            {
+                atom_mag[iat][0] = 0.0;
+                atom_mag[iat][1] = moments[iat*3];
+                atom_mag[iat][2] = moments[iat*3+1];
+                atom_mag[iat][3] = moments[iat*3+2];
+            }
+        }
+        ucell.atom_mulliken = atom_mag;
+    }
+}
+
+} // namespace ModuleIO
+
+#endif // CAL_MAG_H
diff --git a/source/source_io/module_mulliken/output_mulliken.cpp b/source/source_io/module_mulliken/output_mulliken.cpp
index 624bbcd3dca..3594bf1d04f 100644
--- a/source/source_io/module_mulliken/output_mulliken.cpp
+++ b/source/source_io/module_mulliken/output_mulliken.cpp
@@ -1,4 +1,6 @@
 #include "source_io/module_mulliken/output_mulliken.h"
+#include "source_io/module_mulliken/output_dmk.h"
+#include "source_io/module_mulliken/output_sk.h"
 #include "source_io/module_parameter/parameter.h"
 #include "source_base/formatter.h"
 #include "source_base/name_angular.h"
diff --git a/source/source_io/module_mulliken/output_mulliken.h b/source/source_io/module_mulliken/output_mulliken.h
index 2fd1bebd98b..933d721aa9e 100644
--- a/source/source_io/module_mulliken/output_mulliken.h
+++ b/source/source_io/module_mulliken/output_mulliken.h
@@ -4,11 +4,6 @@
 #include "source_base/matrix.h"
 #include "source_basis/module_ao/parallel_orbitals.h"
 #include "source_cell/cell_index.h"
-#include "source_estate/elecstate_lcao.h"
-#include "source_io/module_mulliken/output_dmk.h"
-#include "source_io/module_mulliken/output_sk.h"
-#include "source_lcao/module_operator_lcao/dspin_lcao.h"
-#include "source_estate/module_dm/density_matrix.h" // mohan add 2025-11-04
 
 #include <map>
 #include <vector>
@@ -16,6 +11,12 @@
 namespace ModuleIO
 {
 
+template <typename TK>
+class Output_Sk;
+
+template <typename TK>
+class Output_DMK;
+
 /// @brief the output interface to write the Mulliken population charges
 template <typename TK>
 class Output_Mulliken
@@ -85,101 +86,6 @@ class Output_Mulliken
     ModuleBase::matrix orbMulP_;
 };
 
-template <typename TK>
-void cal_mag(Parallel_Orbitals* pv,
-             hamilt::Hamilt<TK>* p_ham,
-             K_Vectors& kv,
-             elecstate::DensityMatrix<TK,double>* dm, // mohan add 2025-11-04
-             const TwoCenterBundle& two_center_bundle,
-             const LCAO_Orbitals& orb,
-             UnitCell& ucell,
-             const Grid_Driver& gd,
-             const int istep,
-             const bool print)
-{
-    // 1) calculate and output Mulliken population charges and magnetic moments
-    if (PARAM.inp.out_mul)
-    {
-        auto cell_index
-            = CellIndex(ucell.get_atomLabels(), 
-			ucell.get_atomCounts(), ucell.get_lnchiCounts(), PARAM.inp.nspin);
-        auto out_s_k = ModuleIO::Output_Sk<TK>(p_ham, pv, PARAM.inp.nspin, kv.get_nks());
-        auto out_dm_k = ModuleIO::Output_DMK<TK>(dm, pv, PARAM.inp.nspin, kv.get_nks());
-
-        auto mulp = ModuleIO::Output_Mulliken<TK>(&(out_s_k), 
-			&(out_dm_k), pv, &cell_index, kv.isk, PARAM.inp.nspin);
-        auto atom_chg = mulp.get_atom_chg();
-        /// used in updating mag info in STRU file
-        ucell.atom_mulliken = mulp.get_atom_mulliken(atom_chg);
-        if (print && GlobalV::MY_RANK == 0)
-        {
-            /// write the Orbital file
-            cell_index.write_orb_info(PARAM.globalv.global_out_dir);
-            /// write mulliken.txt
-            mulp.write(istep, PARAM.globalv.global_out_dir);
-            /// write atomic mag info in running log file
-            mulp.print_atom_mag(atom_chg, GlobalV::ofs_running);
-        }
-    }
-    // 2) calculate and output the magnetizations of each atom with projection method
-    if (PARAM.inp.onsite_radius > 0)
-    {
-        std::vector<std::vector<double>> atom_mag(ucell.nat, std::vector<double>(PARAM.inp.nspin, 0.0));
-        std::vector<ModuleBase::Vector3<int>> constrain(ucell.nat, ModuleBase::Vector3<int>(1, 1, 1));
-        const hamilt::HContainer<double>* dmr = dm->get_DMR_pointer(1);
-        std::vector<double> moments;
-        std::vector<double> mag_x(ucell.nat, 0.0);
-        std::vector<double> mag_y(ucell.nat, 0.0);
-        std::vector<double> mag_z(ucell.nat, 0.0);
-        auto atomLabels = ucell.get_atomLabels();
-
-        if(PARAM.inp.nspin == 2)
-        {
-            auto sc_lambda = new hamilt::DeltaSpin<hamilt::OperatorLCAO<TK, double>>(nullptr,
-		kv.kvec_d,
-		dynamic_cast<hamilt::HamiltLCAO<TK, double>*>(p_ham)->getHR(),
-		ucell,
-		&gd,
-		two_center_bundle.overlap_orb_onsite.get(),
-		orb.cutoffs());
-
-	    dm->switch_dmr(2);
-	    moments = sc_lambda->cal_moment(dmr, constrain);
-	    dm->switch_dmr(0);
-
-	    delete sc_lambda;
-
-            for(int iat=0;iat<ucell.nat;iat++)
-            {
-                atom_mag[iat][0] = 0.0;
-                atom_mag[iat][1] = moments[iat];
-            }
-        }
-        else if(PARAM.inp.nspin == 4)
-        {
-            auto sc_lambda = new hamilt::DeltaSpin<hamilt::OperatorLCAO<std::complex<double>, std::complex<double>>>(
-                nullptr,
-                kv.kvec_d,
-                dynamic_cast<hamilt::HamiltLCAO<std::complex<double>, std::complex<double>>*>(p_ham)->getHR(),
-                ucell,
-                &gd,
-                two_center_bundle.overlap_orb_onsite.get(),
-                orb.cutoffs());
-            moments = sc_lambda->cal_moment(dmr, constrain);
-            delete sc_lambda;
-
-            for(int iat=0;iat<ucell.nat;iat++)
-            {
-                atom_mag[iat][0] = 0.0;
-                atom_mag[iat][1] = moments[iat*3];
-                atom_mag[iat][2] = moments[iat*3+1];
-                atom_mag[iat][3] = moments[iat*3+2];
-            }
-        }
-        ucell.atom_mulliken = atom_mag;
-    }
-}
-
 } // namespace ModuleIO
 
 #endif // OUTPUT_MULLIKEN_H
diff --git a/source/source_io/module_mulliken/output_sk.cpp b/source/source_io/module_mulliken/output_sk.cpp
index 4d6dc0eab1c..b3f3772f107 100644
--- a/source/source_io/module_mulliken/output_sk.cpp
+++ b/source/source_io/module_mulliken/output_sk.cpp
@@ -1,5 +1,6 @@
 #include "output_sk.h"
 
+#include "source_lcao/hamilt_lcao.h"
 #include "source_base/tool_quit.h"
 
 namespace ModuleIO
diff --git a/source/source_io/module_mulliken/output_sk.h b/source/source_io/module_mulliken/output_sk.h
index ca56dae4e61..1fd1f294abb 100644
--- a/source/source_io/module_mulliken/output_sk.h
+++ b/source/source_io/module_mulliken/output_sk.h
@@ -2,7 +2,7 @@
 #define OUTPUT_SK_H
 
 #include "source_basis/module_ao/parallel_orbitals.h"
-#include "source_lcao/hamilt_lcao.h"
+#include "source_hamilt/hamilt.h"
 
 namespace ModuleIO
 {
diff --git a/source/source_io/module_parameter/input_conv.cpp b/source/source_io/module_parameter/input_conv.cpp
index 9d88945d88e..952460c8ded 100644
--- a/source/source_io/module_parameter/input_conv.cpp
+++ b/source/source_io/module_parameter/input_conv.cpp
@@ -25,7 +25,6 @@
 #include "source_basis/module_ao/ORB_read.h"
 #include "source_estate/module_pot/H_TDDFT_pw.h"
 #include "source_lcao/FORCE_STRESS.h"
-#include "source_lcao/module_rt/evolve_elec.h"
 #include "source_lcao/module_rt/td_info.h"
 #endif
 #ifdef __PEXSI
diff --git a/source/source_io/test/for_testing_input_conv.h b/source/source_io/test/for_testing_input_conv.h
index 9efdfc2c705..b3d24448771 100644
--- a/source/source_io/test/for_testing_input_conv.h
+++ b/source/source_io/test/for_testing_input_conv.h
@@ -12,7 +12,6 @@
 #include "source_estate/module_pot/gatefield.h"
 #include "source_lcao/FORCE_STRESS.h"
 #include "source_lcao/module_dftu/dftu.h"
-#include "source_lcao/module_rt/evolve_elec.h"
 #include "source_pw/module_pwdft/vnl_pw.h"
 #include "source_pw/module_pwdft/structure_factor.h"
 #include "source_hsolver/hsolver_lcao.h"
diff --git a/source/source_lcao/hamilt_lcao.cpp b/source/source_lcao/hamilt_lcao.cpp
index 192f3a5ad77..b6ea2b75990 100644
--- a/source/source_lcao/hamilt_lcao.cpp
+++ b/source/source_lcao/hamilt_lcao.cpp
@@ -4,6 +4,10 @@
 #include "source_base/memory_recorder.h"
 #include "source_base/timer.h"
 #include "source_lcao/module_dftu/dftu.h"
+#include "source_lcao/setup_exx.h"
+#include "source_lcao/setup_deepks.h"
+#include "source_estate/module_dm/density_matrix.h"
+#include "source_estate/module_pot/potential_new.h"
 #include "source_io/module_parameter/parameter.h"
 
 #include <vector>
diff --git a/source/source_lcao/hamilt_lcao.h b/source/source_lcao/hamilt_lcao.h
index 70f73dfaad6..c286da764d7 100644
--- a/source/source_lcao/hamilt_lcao.h
+++ b/source/source_lcao/hamilt_lcao.h
@@ -3,8 +3,7 @@
 
 #include "source_basis/module_nao/two_center_bundle.h"
 #include "source_cell/klist.h"
-#include "source_estate/module_dm/density_matrix.h"
-#include "source_estate/module_pot/potential_new.h"
+#include "source_cell/module_neighbor/sltk_grid_driver.h"
 #include "source_hamilt/hamilt.h"
 #include "source_lcao/hs_matrix_k.hpp"
 #include "source_lcao/module_hcontainer/hcontainer.h"
@@ -12,14 +11,24 @@
 #include <memory>
 #include <vector>
 
-#include "source_lcao/setup_deepks.h" // mohan add 20251008
+// elecstate::Potential forward declaration, full definition in potential_new.h (moved to .cpp)
+// mohan add 20260605
+namespace elecstate { class Potential; }
 
-#ifdef __EXX
-#include "source_lcao/module_ri/Exx_LRI.h"
-#endif
+// elecstate::DensityMatrix forward declaration, full definition in density_matrix.h (moved to .cpp)
+// mohan add 20260605
+namespace elecstate { template <typename TK, typename TR> class DensityMatrix; }
+
+// Setup_DeePKS forward declaration, full definition in setup_deepks.h (moved to .cpp)
+// mohan add 20260605
+template <typename TK> class Setup_DeePKS;
+// Plus_U forward declaration, full definition in module_dftu/dftu.h (moved to .cpp)
+// mohan add 20260605
+class Plus_U;
 
-#include "source_lcao/setup_exx.h" // for exx, mohan add 20251022
-#include "source_lcao/module_dftu/dftu.h" // mohan add 2025-11-05
+// Exx_NAO forward declaration, full definition in setup_exx.h (moved to .cpp)
+// mohan add 20260605
+template <typename TK> class Exx_NAO; 
 
 namespace hamilt
 {
@@ -43,16 +52,16 @@ class HamiltLCAO : public Hamilt<TK>
      */
     HamiltLCAO(const UnitCell& ucell,
                const Grid_Driver& grid_d,
-			   const Parallel_Orbitals* paraV,
-			   elecstate::Potential* pot_in,
-			   const K_Vectors& kv_in,
-			   const TwoCenterBundle& two_center_bundle,
+               const Parallel_Orbitals* paraV,
+               elecstate::Potential* pot_in,
+               const K_Vectors& kv_in,
+               const TwoCenterBundle& two_center_bundle,
                const LCAO_Orbitals& orb,
-			   elecstate::DensityMatrix<TK, double>* DM_in,
-			   Plus_U* p_dftu, // mohan add 2025-11-05
-			   Setup_DeePKS<TK> &deepks,
-			   const int istep, 
-			   Exx_NAO<TK> &exx_nao);
+               elecstate::DensityMatrix<TK, double>* DM_in,
+               Plus_U* p_dftu, // mohan add 2025-11-05
+               Setup_DeePKS<TK> &deepks,
+               const int istep, 
+               Exx_NAO<TK> &exx_nao);
 
     /**
      * @brief Constructor of vacuum Operators, only HR and SR will be initialed as empty HContainer
@@ -145,7 +154,7 @@ class HamiltLCAO : public Hamilt<TK>
      * @param hk_type 0: SK is row-major, 1: SK is collumn-major
      * @return void
      */
-	void updateSk(const int ik, const int hk_type = 0);
+    void updateSk(const int ik, const int hk_type = 0);
 
     // core function: return H(k) and S(k) matrixs for direct solving eigenvalues.
     // not used in PW base
diff --git a/source/source_lcao/module_deepks/LCAO_deepks_interface.cpp b/source/source_lcao/module_deepks/LCAO_deepks_interface.cpp
index 928617f4cdb..aaaa469477f 100644
--- a/source/source_lcao/module_deepks/LCAO_deepks_interface.cpp
+++ b/source/source_lcao/module_deepks/LCAO_deepks_interface.cpp
@@ -5,6 +5,7 @@
 #include "source_base/global_variable.h"
 #include "source_base/tool_title.h"
 #include "source_estate/cal_dm.h"
+#include "source_estate/module_dm/density_matrix.h"
 #include "source_io/module_parameter/parameter.h"
 #include "source_lcao/module_deepks/deepks_check.h"
 #include "source_lcao/module_deepks/deepks_descriptor.h"
diff --git a/source/source_lcao/module_gint/gint_interface.cpp b/source/source_lcao/module_gint/gint_interface.cpp
index f8b896fe255..d701fbef6e9 100644
--- a/source/source_lcao/module_gint/gint_interface.cpp
+++ b/source/source_lcao/module_gint/gint_interface.cpp
@@ -9,6 +9,7 @@
 #include "gint_fvl_meta.h"
 #include "gint_rho.h"
 #include "gint_tau.h"
+#include "gint_dvlocal.h"
 
 #ifdef __CUDA
 #include "gint_vl_gpu.h"
diff --git a/source/source_lcao/module_gint/gint_interface.h b/source/source_lcao/module_gint/gint_interface.h
index cb9112cf4f4..8710fc65131 100644
--- a/source/source_lcao/module_gint/gint_interface.h
+++ b/source/source_lcao/module_gint/gint_interface.h
@@ -2,7 +2,11 @@
 #include <vector>
 #include "source_lcao/module_hcontainer/hcontainer.h"
 #include "gint_type.h"
-#include "gint_dvlocal.h"
+
+class Parallel_Orbitals;
+class UnitCell;
+class Grid_Driver;
+class LCAO_HS_Arrays;
 
 namespace ModuleGint
 {
diff --git a/source/source_lcao/module_lr/esolver_lrtd_lcao.cpp b/source/source_lcao/module_lr/esolver_lrtd_lcao.cpp
index 71b9b961202..92c36d52ae3 100644
--- a/source/source_lcao/module_lr/esolver_lrtd_lcao.cpp
+++ b/source/source_lcao/module_lr/esolver_lrtd_lcao.cpp
@@ -5,6 +5,7 @@
 #include "source_lcao/module_lr/potentials/pot_hxc_lrtd.h"
 #include "source_lcao/module_lr/hsolver_lrtd.hpp"
 #include "source_lcao/module_lr/lr_spectrum.h"
+#include "source_lcao/module_gint/gint.h"
 #include <memory>
 #include "source_lcao/hamilt_lcao.h"
 #include "source_io/module_wf/read_wfc_nao.h"
diff --git a/source/source_lcao/module_rdmft/rdmft.cpp b/source/source_lcao/module_rdmft/rdmft.cpp
index c8c36cf9119..d62a3ee3716 100644
--- a/source/source_lcao/module_rdmft/rdmft.cpp
+++ b/source/source_lcao/module_rdmft/rdmft.cpp
@@ -9,6 +9,10 @@
 #include "source_base/parallel_reduce.h"
 #include "source_cell/module_symmetry/symmetry.h"
 
+#ifdef __EXX
+#include "source_lcao/module_ri/Exx_LRI.h"
+#endif
+
 #include <iostream>
 #include <cmath>
 #include <complex>
diff --git a/source/source_lcao/module_rdmft/rdmft.h b/source/source_lcao/module_rdmft/rdmft.h
index c11b504ebd6..e67ea66181a 100644
--- a/source/source_lcao/module_rdmft/rdmft.h
+++ b/source/source_lcao/module_rdmft/rdmft.h
@@ -19,9 +19,10 @@
 #include "source_lcao/hs_matrix_k.hpp"
 
 #ifdef __EXX
-#include "source_lcao/module_ri/Exx_LRI.h"
+// Exx_LRI forward declaration, full definition in Exx_LRI.h (moved to .cpp)
+// mohan add 20260605
+template <typename TR> class Exx_LRI;
 #include "source_lcao/module_ri/module_exx_symmetry/symmetry_rotation.h"
-// there are some operator reload to print data in different formats
 #endif
 
 #include "source_estate/elecstate.h"
diff --git a/source/source_lcao/module_rt/band_energy.cpp b/source/source_lcao/module_rt/band_energy.cpp
index 601840b2857..2a03a5f7dcc 100644
--- a/source/source_lcao/module_rt/band_energy.cpp
+++ b/source/source_lcao/module_rt/band_energy.cpp
@@ -3,6 +3,7 @@
 #include "evolve_elec.h"
 #include "source_base/module_container/ATen/kernels/blas.h"
 #include "source_base/module_external/scalapack_connector.h"
+#include "source_io/module_parameter/parameter.h"
 
 #ifdef __CUBLASMP
 #include "kernels/cuda/band_energy_kernel.cuh"
diff --git a/source/source_lcao/module_rt/evolve_elec.cpp b/source/source_lcao/module_rt/evolve_elec.cpp
index 491f5c63e0f..14aa91ca3a0 100644
--- a/source/source_lcao/module_rt/evolve_elec.cpp
+++ b/source/source_lcao/module_rt/evolve_elec.cpp
@@ -3,9 +3,6 @@
 #include "evolve_psi.h"
 #include "source_base/parallel_reduce.h"
 #include "source_base/timer.h"
-#include "source_estate/module_charge/symmetry_rho.h"
-#include "source_lcao/hamilt_lcao.h"
-#include "source_lcao/module_dftu/dftu.h"
 
 namespace module_rt
 {
diff --git a/source/source_lcao/module_rt/evolve_elec.h b/source/source_lcao/module_rt/evolve_elec.h
index bd6420f5ca2..f571027a6ee 100644
--- a/source/source_lcao/module_rt/evolve_elec.h
+++ b/source/source_lcao/module_rt/evolve_elec.h
@@ -6,14 +6,20 @@
 #include "source_base/module_container/ATen/core/tensor.h"     // ct::Tensor
 #include "source_base/module_device/device.h"                  // base_device
 #include "source_base/module_device/memory_op.h"               // memory operations
-#include "source_esolver/esolver_ks_lcao.h"
-#include "source_esolver/esolver_ks_lcao_tddft.h"
-#include "source_lcao/hamilt_lcao.h"
+#include "source_hamilt/hamilt.h"
 #include "source_lcao/module_rt/gather_mat.h" // MPI gathering and distributing functions
 #include "source_lcao/module_rt/kernels/cublasmp_context.h"
 #include "source_lcao/module_rt/td_moving_gauge.h"
 #include "source_psi/psi.h"
 
+namespace ModuleESolver
+{
+template <typename TK, typename TR>
+class ESolver_KS_LCAO;
+template <typename TR, typename Device>
+class ESolver_KS_LCAO_TDDFT;
+} // namespace ModuleESolver
+
 //-----------------------------------------------------------
 // mohan add 2021-02-09
 // This class is used to evolve the electronic wave functions
diff --git a/source/source_lcao/module_rt/evolve_psi.cpp b/source/source_lcao/module_rt/evolve_psi.cpp
index 162f64e4574..d40fd4305e5 100644
--- a/source/source_lcao/module_rt/evolve_psi.cpp
+++ b/source/source_lcao/module_rt/evolve_psi.cpp
@@ -5,11 +5,14 @@
 #include "norm_psi.h"
 #include "propagator.h"
 #include "solve_propagation.h"
+#include "source_base/module_external/blas_connector.h"
+#include "source_base/global_function.h"
 #include "source_base/module_container/ATen/kernels/blas.h"   // cuBLAS handle
 #include "source_base/module_container/ATen/kernels/lapack.h" // cuSOLVER handle
-#include "source_esolver/esolver_ks_lcao_tddft.h"             // use gatherMatrix
+#include "source_base/timer.h"
+#include "source_base/tool_title.h"
 #include "source_io/module_parameter/parameter.h"
-#include "source_lcao/hamilt_lcao.h"
+#include "gather_mat.h"
 #include "upsi.h"
 
 #include <complex>
diff --git a/source/source_lcao/module_rt/evolve_psi.h b/source/source_lcao/module_rt/evolve_psi.h
index c723b6fba75..ad76e74b33d 100644
--- a/source/source_lcao/module_rt/evolve_psi.h
+++ b/source/source_lcao/module_rt/evolve_psi.h
@@ -8,8 +8,7 @@
 
 #include "source_base/module_container/ATen/core/tensor.h"     // ct::Tensor
 #include "source_basis/module_ao/parallel_orbitals.h"
-#include "source_lcao/hamilt_lcao.h"
-#include "source_lcao/module_rt/evolve_elec.h"
+#include "source_hamilt/hamilt.h"
 #include "source_lcao/module_rt/kernels/cublasmp_context.h"
 
 namespace module_rt
diff --git a/source/source_lcao/module_rt/gather_mat.h b/source/source_lcao/module_rt/gather_mat.h
index 00df8620ad5..151ee066695 100644
--- a/source/source_lcao/module_rt/gather_mat.h
+++ b/source/source_lcao/module_rt/gather_mat.h
@@ -2,6 +2,7 @@
 #define GATHER_MAT_H
 
 #include "source_base/module_external/scalapack_connector.h" // Cpxgemr2d
+#include "source_basis/module_ao/parallel_orbitals.h"
 #include "source_hamilt/matrixblock.h"
 
 namespace module_rt
diff --git a/source/source_lcao/module_rt/test/band_energy_test.cpp b/source/source_lcao/module_rt/test/band_energy_test.cpp
index d35f25ebefe..800e2555c0a 100644
--- a/source/source_lcao/module_rt/test/band_energy_test.cpp
+++ b/source/source_lcao/module_rt/test/band_energy_test.cpp
@@ -5,7 +5,6 @@
 #include <mpi.h>
 
 #include "source_basis/module_ao/parallel_orbitals.h"
-#include "source_lcao/module_rt/evolve_elec.h"
 #include "tddft_test.h"
 
 /************************************************
diff --git a/source/source_lcao/setup_exx.cpp b/source/source_lcao/setup_exx.cpp
index f552ed06dff..841d3579707 100644
--- a/source/source_lcao/setup_exx.cpp
+++ b/source/source_lcao/setup_exx.cpp
@@ -1,5 +1,9 @@
 #include "source_lcao/setup_exx.h"
 
+#ifdef __EXX
+#include "source_lcao/module_ri/Exx_LRI_interface.h"
+#endif
+
 template <typename TK>
 Exx_NAO<TK>::Exx_NAO(){}
 
diff --git a/source/source_lcao/setup_exx.h b/source/source_lcao/setup_exx.h
index badf9d14c7c..faf232e3de9 100644
--- a/source/source_lcao/setup_exx.h
+++ b/source/source_lcao/setup_exx.h
@@ -10,7 +10,9 @@
 
 // for EXX
 #ifdef __EXX
-#include "source_lcao/module_ri/Exx_LRI_interface.h"
+// Exx_LRI_Interface forward declaration, full definition in Exx_LRI_interface.h (moved to .cpp)
+// mohan add 20260605
+template <typename TK, typename TR> class Exx_LRI_Interface;
 #endif
 
 template <typename TK>