deepmodeling · mohanchen · Jun 4, 2026 · Jun 5, 2026 · Jun 5, 2026 · Jun 5, 2026
diff --git 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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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
@@ -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"