From 0152800f6eeb007b6a1611566dffa3137d41ff2f Mon Sep 17 00:00:00 2001
From: Critsium <tsfxwbbzxy@163.com>
Date: Fri, 31 Jul 2026 15:55:48 +0800
Subject: [PATCH 1/2] Refactor: inject nlocal/nbands into the LCAO dense
 diagonalizers

Group D of the source_hsolver PARAM removal, and the largest single cluster:
the six dense LCAO diagonalizers all sized their eigenvalue buffers from
PARAM.globalv.nlocal and copied out PARAM.inp.nbands eigenvalues.

Each of DiagoLapack, DiagoScalapack, DiagoElpa, DiagoElpaNative, DiagoCusolver
and DiagoCusolverMP now takes (nlocal, nbands) through its constructor and
stores them as members. DiagoElpaNative additionally takes use_gpu, replacing
its PARAM.inp.device == "gpu" test.

HSolverLCAO gains nbands and use_gpu (it already received nlocal in #7711) and
supplies all eleven construction sites across hamiltSolvePsiK, parakSolve and
parakSolve_cusolver. The three call contexts agree on both values: parakSolve
distributes the same nlocal x nlocal matrix over a smaller pool grid, and
parakSolve_cusolver gathers it per k-point, so a single pair covers all of them.

nbands is injected rather than read from ParaV->get_nbands(). The two are equal
for the LCAO path (LCAO_init_basis.cpp seeds ParaV from PARAM.inp.nbands), but
module_lr sets paraMat_.nbands to nocc + nvirt, so sourcing it from ParaV would
plant a trap for any future caller. The existing ParaV->get_nbands() reads
inside parakSolve are left alone; they are not PARAM reads and are out of scope.

In diago_scalapack.cpp, four call sites pass the address of the value to the
Fortran routines. PARAM.globalv returns a const reference, so &PARAM.globalv.nlocal
was already a const int*, matching pdsygvx_/pzhegvx_'s const int* n; &this->nlocal
has the same type.

Also resolves the two error-message strings in diago_lapack.cpp and
diago_scalapack.cpp that embedded "PARAM.inp.nbands = ". These were deferred
from #7706 because removing them needed exactly the nbands member added here.

Two dead commented-out ELPA_Solver calls in diago_elpa.cpp are removed rather
than updated; each duplicated the live statement two lines below it.

The three LCAO diagonalizer tests supply their own nlocal/nbands, which they
already read from the H/S input files, so their PARAM.sys.nlocal /
PARAM.input.nbands writes and the now-dead `#define private public` include
blocks are dropped.

PARAM occurrences in source_hsolver production code: 60 -> 5. What remains is
DiagoIterAssist's basis_type/calculation switches (group E) and one comment in
simple_pexsi.cpp.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---
 source/source_esolver/esolver_ks_lcao.cpp     |  4 +-
 .../source_esolver/esolver_ks_lcao_tddft.cpp  |  4 +-
 source/source_hsolver/diago_cusolver.cpp      |  7 ++--
 source/source_hsolver/diago_cusolver.h        |  9 ++++-
 source/source_hsolver/diago_cusolvermp.cpp    |  5 +--
 source/source_hsolver/diago_cusolvermp.h      |  8 +++-
 source/source_hsolver/diago_elpa.cpp          | 29 ++++++---------
 source/source_hsolver/diago_elpa.h            |  7 ++++
 source/source_hsolver/diago_elpa_native.cpp   |  9 ++---
 source/source_hsolver/diago_elpa_native.h     | 10 +++++
 source/source_hsolver/diago_lapack.cpp        | 33 ++++++++---------
 source/source_hsolver/diago_lapack.h          |  8 ++++
 source/source_hsolver/diago_scalapack.cpp     | 37 +++++++++----------
 source/source_hsolver/diago_scalapack.h       |  7 ++++
 source/source_hsolver/hsolver_lcao.cpp        | 22 +++++------
 source/source_hsolver/hsolver_lcao.h          | 11 ++++--
 .../source_hsolver/test/diago_lapack_test.cpp | 12 +-----
 .../test/diago_lcao_cusolver_test.cpp         |  9 +----
 .../source_hsolver/test/diago_lcao_test.cpp   | 11 ++----
 source/source_lcao/LCAO_set.cpp               |  4 +-
 .../module_deltaspin/cal_mw_from_lambda.cpp   |  4 +-
 21 files changed, 137 insertions(+), 113 deletions(-)

diff --git a/source/source_esolver/esolver_ks_lcao.cpp b/source/source_esolver/esolver_ks_lcao.cpp
index 7b44a2758c..7ebeda0f4c 100644
--- a/source/source_esolver/esolver_ks_lcao.cpp
+++ b/source/source_esolver/esolver_ks_lcao.cpp
@@ -438,7 +438,9 @@ void ESolver_KS_LCAO<TK, TR>::hamilt2rho_single(UnitCell& ucell, int istep, int
                                                   PARAM.inp.ks_solver,
                                                   PARAM.globalv.kpar_lcao,
                                                   PARAM.globalv.nlocal,
-                                                  PARAM.inp.nelec);
+                                                  PARAM.inp.nbands,
+                                                  PARAM.inp.nelec,
+                                                  PARAM.inp.device == "gpu");
         hsolver_lcao_obj.solve(static_cast<hamilt::Hamilt<TK>*>(this->p_hamilt), this->psi[0], this->pelec, *this->dmat.dm, 
           this->chr, PARAM.inp.nspin, skip_charge);
     }
diff --git a/source/source_esolver/esolver_ks_lcao_tddft.cpp b/source/source_esolver/esolver_ks_lcao_tddft.cpp
index de35a40cf8..4ba1c5fb87 100644
--- a/source/source_esolver/esolver_ks_lcao_tddft.cpp
+++ b/source/source_esolver/esolver_ks_lcao_tddft.cpp
@@ -333,7 +333,9 @@ void ESolver_KS_LCAO_TDDFT<TR, Device>::hamilt2rho_single(UnitCell& ucell,
                                                                         PARAM.inp.ks_solver,
                                                                         PARAM.globalv.kpar_lcao,
                                                                         PARAM.globalv.nlocal,
-                                                                        PARAM.inp.nelec);
+                                                                        PARAM.inp.nbands,
+                                                                        PARAM.inp.nelec,
+                                                                        PARAM.inp.device == "gpu");
             hsolver_lcao_obj.solve(static_cast<hamilt::Hamilt<std::complex<double>>*>(this->p_hamilt),
                                    this->psi[0],
                                    this->pelec,
diff --git a/source/source_hsolver/diago_cusolver.cpp b/source/source_hsolver/diago_cusolver.cpp
index b02c413257..34ced44cb8 100644
--- a/source/source_hsolver/diago_cusolver.cpp
+++ b/source/source_hsolver/diago_cusolver.cpp
@@ -6,7 +6,6 @@
 #include "source_base/module_external/scalapack_connector.h"
 #include "source_base/tool_title.h"
 #include "source_base/timer.h"
-#include "source_io/module_parameter/parameter.h"
 
 #include <memory>
 #include <type_traits>
@@ -22,7 +21,7 @@ template <typename T>
 int DiagoCusolver<T>::DecomposedState = 0;
 
 template <typename T>
-DiagoCusolver<T>::DiagoCusolver()
+DiagoCusolver<T>::DiagoCusolver(const int nlocal_in, const int nbands_in) : nlocal(nlocal_in), nbands(nbands_in)
 {
 }
 
@@ -42,13 +41,13 @@ void DiagoCusolver<T>::diag(
     ModuleBase::TITLE("DiagoCusolver", "diag");
     ModuleBase::timer::start("DiagoCusolver", "cusolver");
     // Allocate memory for eigenvalues
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     std::vector<T> eigenvectors(h_mat.row * h_mat.col);
     this->dc.Dngvd(h_mat.row, h_mat.col, h_mat.p, s_mat.p, eigen.data(), eigenvectors.data());
     const int size = psi.get_nbands() * psi.get_nbasis();
     BlasConnector::copy(size, eigenvectors.data(), 1, psi.get_pointer(), 1);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
     ModuleBase::timer::end("DiagoCusolver", "cusolver");
 }
 
diff --git a/source/source_hsolver/diago_cusolver.h b/source/source_hsolver/diago_cusolver.h
index 28ebdc852c..50d9a393c6 100644
--- a/source/source_hsolver/diago_cusolver.h
+++ b/source/source_hsolver/diago_cusolver.h
@@ -19,9 +19,11 @@ class DiagoCusolver
 
   public:
 
-    DiagoCusolver();
+    /// @param nlocal_in global dimension of the NAO Hamiltonian
+    /// @param nbands_in number of lowest eigenpairs to compute
+    DiagoCusolver(const int nlocal_in, const int nbands_in);
     ~DiagoCusolver();
-    
+
     // Override the diag function for CUSOLVER diagonalization
     void diag(
       hamilt::MatrixBlock<T>& h_mat,
@@ -40,6 +42,9 @@ class DiagoCusolver
     // Function to check if ELPA handle needs to be created or reused in MPI settings
     bool ifElpaHandle(const bool& newIteration, const bool& ifNSCF) const;
 #endif
+
+    const int nlocal;
+    const int nbands;
 };
 
 } // namespace hsolver
diff --git a/source/source_hsolver/diago_cusolvermp.cpp b/source/source_hsolver/diago_cusolvermp.cpp
index b5bb80eb5f..395add83d2 100644
--- a/source/source_hsolver/diago_cusolvermp.cpp
+++ b/source/source_hsolver/diago_cusolvermp.cpp
@@ -1,6 +1,5 @@
 #ifdef __CUSOLVERMP
 
-#include "source_io/module_parameter/parameter.h"
 #include "diago_cusolvermp.h"
 
 #include "source_base/module_external/blas_connector.h"
@@ -18,7 +17,7 @@ void DiagoCusolverMP<T>::diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real*
     hamilt::MatrixBlock<T> h_mat, s_mat;
     phm_in->matrix(h_mat, s_mat);
 
-    std::vector<Real> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<Real> eigen(this->nlocal, 0.0);
     std::vector<T> eigenvectors(h_mat.row * h_mat.col);
 
     MPI_Comm COMM_DIAG = MPI_COMM_WORLD; // use all processes
@@ -30,7 +29,7 @@ void DiagoCusolverMP<T>::diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real*
         ModuleBase::timer::end("DiagoCusolverMP", "Diag_CusolverMP_gvd");
     }
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
     const int size = psi.get_nbands() * psi.get_nbasis();
     BlasConnector::copy(size, eigenvectors.data(), inc, psi.get_pointer(), inc);
 }
diff --git a/source/source_hsolver/diago_cusolvermp.h b/source/source_hsolver/diago_cusolvermp.h
index d173315e31..155a4bf94d 100644
--- a/source/source_hsolver/diago_cusolvermp.h
+++ b/source/source_hsolver/diago_cusolvermp.h
@@ -16,11 +16,17 @@ class DiagoCusolverMP
     using Real = typename GetTypeReal<T>::type;
 
   public:
-    DiagoCusolverMP()
+    /// @param nlocal_in global dimension of the NAO Hamiltonian
+    /// @param nbands_in number of lowest eigenpairs to compute
+    DiagoCusolverMP(const int nlocal_in, const int nbands_in) : nlocal(nlocal_in), nbands(nbands_in)
     {
     }
     // the diag function for CUSOLVERMP diagonalization
     void diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real* eigenvalue_in);
+
+  private:
+    const int nlocal;
+    const int nbands;
 };
 } // namespace hsolver
 #endif // __CUSOLVERMP
diff --git a/source/source_hsolver/diago_elpa.cpp b/source/source_hsolver/diago_elpa.cpp
index a264a41b9c..55ca982539 100644
--- a/source/source_hsolver/diago_elpa.cpp
+++ b/source/source_hsolver/diago_elpa.cpp
@@ -2,7 +2,6 @@
 #include "source_base/global_function.h"
 #include "source_base/module_external/blas_connector.h"
 
-#include "source_io/module_parameter/parameter.h"
 #include "module_genelpa/elpa_solver.h"
 #include "source_base/module_external/blacs_connector.h"
 #include "source_base/global_variable.h"
@@ -75,13 +74,13 @@ void DiagoElpa<std::complex<double>>::diag(
     matcd h_mat, s_mat;
     phm_in->matrix(h_mat, s_mat);
 
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
 
     bool isReal = false;
     MPI_Comm COMM_DIAG = setmpicomm(); // set mpi_comm needed
     ELPA_Solver es((const bool)isReal,
                    COMM_DIAG,
-                   (const int)PARAM.inp.nbands,
+                   (const int)this->nbands,
                    (const int)h_mat.row,
                    (const int)h_mat.col,
                    (const int*)h_mat.desc);
@@ -97,7 +96,7 @@ void DiagoElpa<std::complex<double>>::diag(
     es.exit();
 
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 #else
     ModuleBase::WARNING_QUIT("DiagoElpa",
                              "DiagoElpa only can be used with macro __MPI");
@@ -113,15 +112,13 @@ void DiagoElpa<double>::diag(hamilt::Hamilt<double>* phm_in,
     matd h_mat, s_mat;
     phm_in->matrix(h_mat, s_mat);
 
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
 
     bool isReal = true;
     MPI_Comm COMM_DIAG = setmpicomm(); // set mpi_comm needed
-    // ELPA_Solver es(isReal, COMM_DIAG, PARAM.inp.nbands, h_mat.row, h_mat.col,
-    // h_mat.desc);
     ELPA_Solver es((const bool)isReal,
                    COMM_DIAG,
-                   (const int)PARAM.inp.nbands,
+                   (const int)this->nbands,
                    (const int)h_mat.row,
                    (const int)h_mat.col,
                    (const int*)h_mat.desc);
@@ -135,7 +132,7 @@ void DiagoElpa<double>::diag(hamilt::Hamilt<double>* phm_in,
     es.exit();
 
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 #else
     ModuleBase::WARNING_QUIT("DiagoElpa",
                              "DiagoElpa only can be used with macro __MPI");
@@ -151,11 +148,11 @@ void DiagoElpa<std::complex<double>>::diag_pool(hamilt::MatrixBlock<std::complex
     Real* eigenvalue_in,
     MPI_Comm& comm)
 {
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     bool isReal = false;
     ELPA_Solver es((const bool)isReal,
                    comm,
-                   (const int)PARAM.inp.nbands,
+                   (const int)this->nbands,
                    (const int)h_mat.row,
                    (const int)h_mat.col,
                    (const int*)h_mat.desc);
@@ -170,7 +167,7 @@ void DiagoElpa<std::complex<double>>::diag_pool(hamilt::MatrixBlock<std::complex
     ModuleBase::timer::end("DiagoElpa", "elpa_solve");
     es.exit();
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
 
 template <>
@@ -180,14 +177,12 @@ void DiagoElpa<double>::diag_pool(hamilt::MatrixBlock<double>& h_mat,
     Real* eigenvalue_in,
     MPI_Comm& comm)
 {
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
 
     bool isReal = true;
-    // ELPA_Solver es(isReal, COMM_DIAG, PARAM.inp.nbands, h_mat.row, h_mat.col,
-    // h_mat.desc);
     ELPA_Solver es((const bool)isReal,
                    comm,
-                   (const int)PARAM.inp.nbands,
+                   (const int)this->nbands,
                    (const int)h_mat.row,
                    (const int)h_mat.col,
                    (const int*)h_mat.desc);
@@ -203,7 +198,7 @@ void DiagoElpa<double>::diag_pool(hamilt::MatrixBlock<double>& h_mat,
     const int inc = 1;
     ModuleBase::GlobalFunc::OUT(GlobalV::ofs_running,
                                 "K-S equation was solved by genelpa2");
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
     ModuleBase::GlobalFunc::OUT(GlobalV::ofs_running,
                                 "eigenvalues were copied to ekb");
 }
diff --git a/source/source_hsolver/diago_elpa.h b/source/source_hsolver/diago_elpa.h
index c6050af224..2bee7a14fa 100644
--- a/source/source_hsolver/diago_elpa.h
+++ b/source/source_hsolver/diago_elpa.h
@@ -15,6 +15,10 @@ class DiagoElpa
     using Real = typename GetTypeReal<T>::type;
 
   public:
+    /// @param nlocal_in global dimension of the NAO Hamiltonian
+    /// @param nbands_in number of lowest eigenpairs to compute
+    DiagoElpa(const int nlocal_in, const int nbands_in) : nlocal(nlocal_in), nbands(nbands_in) {};
+
     void diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real* eigenvalue_in);
 #ifdef __MPI
     // diagnolization used in parallel-k case
@@ -30,6 +34,9 @@ class DiagoElpa
     bool ifElpaHandle(const bool& newIteration, const bool& ifNSCF) const;
     static int lastmpinum; // last using mpi;
 #endif
+
+    const int nlocal;
+    const int nbands;
 };
 
 template <typename T>
diff --git a/source/source_hsolver/diago_elpa_native.cpp b/source/source_hsolver/diago_elpa_native.cpp
index ee003a08e0..fb4cfbc448 100644
--- a/source/source_hsolver/diago_elpa_native.cpp
+++ b/source/source_hsolver/diago_elpa_native.cpp
@@ -4,7 +4,6 @@
 #include "source_base/module_external/blas_connector.h"
 #include "source_base/module_external/blacs_connector.h"
 #include "source_base/global_variable.h"
-#include "source_io/module_parameter/parameter.h"
 #include "source_base/timer.h"
 #include "source_base/tool_quit.h"
 #include "source_hsolver/module_genelpa/elpa_new.h"
@@ -59,7 +58,7 @@ void DiagoElpaNative<T>::diag_pool(hamilt::MatrixBlock<T>& h_mat,
 
     ModuleBase::timer::start("DiagoElpaNative", "elpa_solve");
 
-    int nev = PARAM.inp.nbands;
+    int nev = this->nbands;
     int narows = h_mat.row;
     int nacols = h_mat.col;
 
@@ -70,7 +69,7 @@ void DiagoElpaNative<T>::diag_pool(hamilt::MatrixBlock<T>& h_mat,
     int nprows, npcols, myprow, mypcol;
 
     Cblacs_gridinfo(cblacs_ctxt, &nprows, &npcols, &myprow, &mypcol);
-    std::vector<Real> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<Real> eigen(this->nlocal, 0.0);
     std::vector<T> eigenvectors(narows * nacols);
 
     if (elpa_init(20210430) != ELPA_OK)
@@ -107,7 +106,7 @@ void DiagoElpaNative<T>::diag_pool(hamilt::MatrixBlock<T>& h_mat,
     #define ELPA_WITH_SYCL_GPU_VERSION 0
  */
 #if ELPA_WITH_NVIDIA_GPU_VERSION
-    if (PARAM.inp.device == "gpu")
+    if (this->use_gpu)
     {
         elpa_set(handle, "nvidia-gpu", 1, &success);
         elpa_set(handle, "real_kernel", ELPA_2STAGE_REAL_NVIDIA_GPU, &success);
@@ -138,7 +137,7 @@ void DiagoElpaNative<T>::diag_pool(hamilt::MatrixBlock<T>& h_mat,
     }
 
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
     const int size = psi.get_nbands() * psi.get_nbasis();
     BlasConnector::copy(size, eigenvectors.data(), inc, psi.get_pointer(), inc);
 }
diff --git a/source/source_hsolver/diago_elpa_native.h b/source/source_hsolver/diago_elpa_native.h
index 2c556254e8..2cbd2c27d6 100644
--- a/source/source_hsolver/diago_elpa_native.h
+++ b/source/source_hsolver/diago_elpa_native.h
@@ -15,6 +15,12 @@ class DiagoElpaNative
     using Real = typename GetTypeReal<T>::type;
 
   public:
+    /// @param nlocal_in global dimension of the NAO Hamiltonian
+    /// @param nbands_in number of lowest eigenpairs to compute
+    /// @param use_gpu_in offload to the NVIDIA-GPU ELPA kernels when ELPA was built with GPU support
+    DiagoElpaNative(const int nlocal_in, const int nbands_in, const bool use_gpu_in)
+        : nlocal(nlocal_in), nbands(nbands_in), use_gpu(use_gpu_in) {};
+
     void diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real* eigenvalue_in);
 #ifdef __MPI
     // diagnolization used in parallel-k case
@@ -27,6 +33,10 @@ class DiagoElpaNative
 
     static int DecomposedState;
 
+  private:
+    const int nlocal;
+    const int nbands;
+    const bool use_gpu;
 };
 
 template <typename T>
diff --git a/source/source_hsolver/diago_lapack.cpp b/source/source_hsolver/diago_lapack.cpp
index 6bdba41ae9..0020efc1e8 100644
--- a/source/source_hsolver/diago_lapack.cpp
+++ b/source/source_hsolver/diago_lapack.cpp
@@ -1,7 +1,5 @@
 // Refactored according to diago_scalapack
 // This code will be futher refactored to remove the dependency of psi and hamilt
-#include "source_io/module_parameter/parameter.h"
-
 #include "diago_lapack.h"
 
 #include "source_base/global_variable.h"
@@ -42,14 +40,14 @@ void DiagoLapack<double>::diag(hamilt::Hamilt<double>* phm_in, psi::Psi<double>&
     phm_in->matrix(h_mat, s_mat);
 
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     check_lapack_layout(h_mat, s_mat, eigen.size());
 
     // Diag
     this->dsygvx_diag(h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     // Copy result
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
 
 template <>
@@ -61,11 +59,11 @@ void DiagoLapack<std::complex<double>>::diag(hamilt::Hamilt<std::complex<double>
     matcd h_mat, s_mat;
     phm_in->matrix(h_mat, s_mat);
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     check_lapack_layout(h_mat, s_mat, eigen.size());
     this->zhegvx_diag(h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
 
 #ifdef __MPI
@@ -78,11 +76,11 @@ void DiagoLapack<std::complex<double>>::diag(hamilt::Hamilt<std::complex<double>
 {
     ModuleBase::TITLE("DiagoLapack", "diag_pool");
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     check_lapack_layout(h_mat, s_mat, eigen.size());
     this->dsygvx_diag(h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
     template<>
     void DiagoLapack<std::complex<double>>::diag_pool(hamilt::MatrixBlock<std::complex<double>>& h_mat,
@@ -93,11 +91,11 @@ void DiagoLapack<std::complex<double>>::diag(hamilt::Hamilt<std::complex<double>
 {
     ModuleBase::TITLE("DiagoLapack", "diag_pool");
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     check_lapack_layout(h_mat, s_mat, eigen.size());
     this->zhegvx_diag(h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
 #endif
 
@@ -115,20 +113,20 @@ std::pair<int, std::vector<int>> DiagoLapack<T>::dsygvx_once(const int ncol,
     memcpy(s_tmp.c, s_mat, sizeof(double) * ncol * nrow);
 
     const char jobz = 'V', range = 'I', uplo = 'U';
-    const int itype = 1, il = 1, iu = PARAM.inp.nbands, one = 1;
+    const int itype = 1, il = 1, iu = this->nbands, one = 1;
     int M = 0, NZ = 0, lwork = -1, liwork = -1, info = 0;
     double vl = 0, vu = 0;
     const double abstol = LAPACK_ABSTOL, orfac = LAPACK_ORFAC;
     std::vector<double> work(3, 0);
     std::vector<int> iwork(1, 0);
-    std::vector<int> ifail(PARAM.globalv.nlocal, 0);
+    std::vector<int> ifail(this->nlocal, 0);
     std::vector<int> iclustr(2 * GlobalV::DSIZE);
     std::vector<double> gap(GlobalV::DSIZE);
 
     // LAPACK dsygvx signature:
     // (ITYPE, JOBZ, RANGE, UPLO, N, A, LDA, B, LDB, VL, VU, IL, IU,
     //  ABSTOL, M, W, Z, LDZ, WORK, LWORK, IWORK, IFAIL, INFO)
-    int n = PARAM.globalv.nlocal;
+    int n = this->nlocal;
     int lda = n, ldb = n, ldz = n;
     dsygvx_(&itype,
         &jobz,
@@ -224,7 +222,7 @@ std::pair<int, std::vector<int>> DiagoLapack<T>::zhegvx_once(const int ncol,
     memcpy(s_tmp.c, s_mat, sizeof(std::complex<double>) * ncol * nrow);
 
     const char jobz = 'V', range = 'I', uplo = 'U';
-    const int itype = 1, il = 1, iu = PARAM.inp.nbands, one = 1;
+    const int itype = 1, il = 1, iu = this->nbands, one = 1;
     int M = 0, NZ = 0, lwork = -1, lrwork = -1, liwork = -1, info = 0;
     const double abstol = LAPACK_ABSTOL, orfac = LAPACK_ORFAC;
     
@@ -232,14 +230,14 @@ std::pair<int, std::vector<int>> DiagoLapack<T>::zhegvx_once(const int ncol,
     std::vector<std::complex<double>> work(1, 0);
     std::vector<double> rwork(3, 0);
     std::vector<int> iwork(1, 0);
-    std::vector<int> ifail(PARAM.globalv.nlocal, 0);
+    std::vector<int> ifail(this->nlocal, 0);
     std::vector<int> iclustr(2 * GlobalV::DSIZE);
     std::vector<double> gap(GlobalV::DSIZE);
 
     // LAPACK zhegvx signature:
     // (ITYPE, JOBZ, RANGE, UPLO, N, A, LDA, B, LDB, VL, VU, IL, IU,
     //  ABSTOL, M, W, Z, LDZ, WORK, LWORK, RWORK, IWORK, IFAIL, INFO)
-    int n = PARAM.globalv.nlocal;
+    int n = this->nlocal;
     int lda = n, ldb = n, ldz = n;
     zhegvx_(&itype,
         &jobz,
@@ -416,8 +414,7 @@ void DiagoLapack<T>::post_processing(const int info, const std::vector<int>& vec
     {
         const std::string str_M = "M = " + ModuleBase::GlobalFunc::TO_STRING(vec[0]) + ".\n";
         const std::string str_NZ = "NZ = " + ModuleBase::GlobalFunc::TO_STRING(vec[1]) + ".\n";
-        const std::string str_NBANDS
-            = "PARAM.inp.nbands = " + ModuleBase::GlobalFunc::TO_STRING(PARAM.inp.nbands) + ".\n";
+        const std::string str_NBANDS = "nbands = " + ModuleBase::GlobalFunc::TO_STRING(this->nbands) + ".\n";
         throw std::runtime_error(str_info_FILE + str_M + str_NZ + str_NBANDS);
     }
     else if (info / 16 % 2)
diff --git a/source/source_hsolver/diago_lapack.h b/source/source_hsolver/diago_lapack.h
index bfdf78ac34..9f674fc20f 100644
--- a/source/source_hsolver/diago_lapack.h
+++ b/source/source_hsolver/diago_lapack.h
@@ -26,6 +26,10 @@ class DiagoLapack
     using Real = typename GetTypeReal<T>::type;
 
   public:
+    /// @param nlocal_in global dimension of the NAO Hamiltonian
+    /// @param nbands_in number of lowest eigenpairs to compute
+    DiagoLapack(const int nlocal_in, const int nbands_in) : nlocal(nlocal_in), nbands(nbands_in) {};
+
     void diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real* eigenvalue_in);
   #ifdef __MPI
     // diagnolization used in parallel-k case
@@ -61,6 +65,10 @@ class DiagoLapack
     int degeneracy_max = 12; // For reorthogonalized memory. 12 followes siesta.
 
     void post_processing(const int info, const std::vector<int>& vec);
+
+  private:
+    const int nlocal;
+    const int nbands;
 };
 
 } // namespace hsolver
diff --git a/source/source_hsolver/diago_scalapack.cpp b/source/source_hsolver/diago_scalapack.cpp
index 366478fdc2..f2ddb8289b 100644
--- a/source/source_hsolver/diago_scalapack.cpp
+++ b/source/source_hsolver/diago_scalapack.cpp
@@ -1,6 +1,5 @@
 //=====================
 // AUTHOR : Peize Lin
-#include "source_io/module_parameter/parameter.h"
 // DATE : 2021-11-02
 // REFACTORING AUTHOR : Daye Zheng
 // DATE : 2022-04-14
@@ -28,10 +27,10 @@ namespace hsolver
     matd h_mat, s_mat;
     phm_in->matrix(h_mat, s_mat);
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     this->pdsygvx_diag(h_mat.desc, h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
     template<>
     void DiagoScalapack<std::complex<double>>::diag(hamilt::Hamilt<std::complex<double>>* phm_in, psi::Psi<std::complex<double>>& psi, Real* eigenvalue_in)
@@ -40,10 +39,10 @@ namespace hsolver
     matcd h_mat, s_mat;
     phm_in->matrix(h_mat, s_mat);
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     this->pzhegvx_diag(h_mat.desc, h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
 
 #ifdef __MPI
@@ -56,10 +55,10 @@ namespace hsolver
 {
     ModuleBase::TITLE("DiagoScalapack", "diag_pool");
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     this->pdsygvx_diag(h_mat.desc, h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
     template<>
     void DiagoScalapack<std::complex<double>>::diag_pool(hamilt::MatrixBlock<std::complex<double>>& h_mat,
@@ -70,10 +69,10 @@ namespace hsolver
 {
     ModuleBase::TITLE("DiagoScalapack", "diag_pool");
     assert(h_mat.col == s_mat.col && h_mat.row == s_mat.row && h_mat.desc == s_mat.desc);
-    std::vector<double> eigen(PARAM.globalv.nlocal, 0.0);
+    std::vector<double> eigen(this->nlocal, 0.0);
     this->pzhegvx_diag(h_mat.desc, h_mat.col, h_mat.row, h_mat.p, s_mat.p, eigen.data(), psi);
     const int inc = 1;
-    BlasConnector::copy(PARAM.inp.nbands, eigen.data(), inc, eigenvalue_in, inc);
+    BlasConnector::copy(this->nbands, eigen.data(), inc, eigenvalue_in, inc);
 }
 #endif
 
@@ -92,13 +91,13 @@ namespace hsolver
     memcpy(s_tmp.c, s_mat, sizeof(double) * ncol * nrow);
 
     const char jobz = 'V', range = 'I', uplo = 'U';
-    const int itype = 1, il = 1, iu = PARAM.inp.nbands, one = 1;
+    const int itype = 1, il = 1, iu = this->nbands, one = 1;
     int M = 0, NZ = 0, lwork = -1, liwork = -1, info = 0;
     double vl = 0, vu = 0;
     const double abstol = SCALAPACK_ABSTOL, orfac = SCALAPACK_ORFAC;
     std::vector<double> work(3, 0);
     std::vector<int> iwork(1, 0);
-    std::vector<int> ifail(PARAM.globalv.nlocal, 0);
+    std::vector<int> ifail(this->nlocal, 0);
     std::vector<int> iclustr(2 * GlobalV::DSIZE);
     std::vector<double> gap(GlobalV::DSIZE);
 
@@ -106,7 +105,7 @@ namespace hsolver
              &jobz,
              &range,
              &uplo,
-             &PARAM.globalv.nlocal,
+             &this->nlocal,
              h_tmp.c,
              &one,
              &one,
@@ -152,7 +151,7 @@ namespace hsolver
              &jobz,
              &range,
              &uplo,
-             &PARAM.globalv.nlocal,
+             &this->nlocal,
              h_tmp.c,
              &one,
              &one,
@@ -217,7 +216,7 @@ namespace hsolver
     memcpy(s_tmp.c, s_mat, sizeof(std::complex<double>) * ncol * nrow);
 
     const char jobz = 'V', range = 'I', uplo = 'U';
-    const int itype = 1, il = 1, iu = PARAM.inp.nbands, one = 1;
+    const int itype = 1, il = 1, iu = this->nbands, one = 1;
     int M = 0, NZ = 0, lwork = -1, lrwork = -1, liwork = -1, info = 0;
     const double abstol = SCALAPACK_ABSTOL, orfac = SCALAPACK_ORFAC;
     //Note: pzhegvx_ has a bug
@@ -227,7 +226,7 @@ namespace hsolver
     std::vector<std::complex<double>> work(1, 0);
     std::vector<double> rwork(3, 0);
     std::vector<int> iwork(1, 0);
-    std::vector<int> ifail(PARAM.globalv.nlocal, 0);
+    std::vector<int> ifail(this->nlocal, 0);
     std::vector<int> iclustr(2 * GlobalV::DSIZE);
     std::vector<double> gap(GlobalV::DSIZE);
 
@@ -235,7 +234,7 @@ namespace hsolver
              &jobz,
              &range,
              &uplo,
-             &PARAM.globalv.nlocal,
+             &this->nlocal,
              h_tmp.c,
              &one,
              &one,
@@ -276,7 +275,7 @@ namespace hsolver
     //	GlobalV::ofs_running<<"lwork="<<work[0]<<"\t"<<"lrwork="<<rwork[0]<<"\t"<<"liwork="<<iwork[0]<<std::endl;
     lwork = work[0].real();
     work.resize(lwork, 0);
-    lrwork = rwork[0] + this->degeneracy_max * PARAM.globalv.nlocal;
+    lrwork = rwork[0] + this->degeneracy_max * this->nlocal;
     int maxlrwork = std::max(lrwork,3);
     rwork.resize(maxlrwork, 0);
     liwork = iwork[0];
@@ -286,7 +285,7 @@ namespace hsolver
              &jobz,
              &range,
              &uplo,
-             &PARAM.globalv.nlocal,
+             &this->nlocal,
              h_tmp.c,
              &one,
              &one,
@@ -431,7 +430,7 @@ namespace hsolver
         const std::string str_M = "M = " + ModuleBase::GlobalFunc::TO_STRING(vec[0]) + ".\n";
         const std::string str_NZ = "NZ = " + ModuleBase::GlobalFunc::TO_STRING(vec[1]) + ".\n";
         const std::string str_NBANDS
-            = "PARAM.inp.nbands = " + ModuleBase::GlobalFunc::TO_STRING(PARAM.inp.nbands) + ".\n";
+            = "nbands = " + ModuleBase::GlobalFunc::TO_STRING(this->nbands) + ".\n";
         throw std::runtime_error(str_info_FILE + str_M + str_NZ + str_NBANDS);
     }
     else if (info / 16 % 2)
diff --git a/source/source_hsolver/diago_scalapack.h b/source/source_hsolver/diago_scalapack.h
index 05b1144836..7dce030616 100644
--- a/source/source_hsolver/diago_scalapack.h
+++ b/source/source_hsolver/diago_scalapack.h
@@ -27,6 +27,10 @@ namespace hsolver
 private:
     using Real = typename GetTypeReal<T>::type;
   public:
+    /// @param nlocal_in global dimension of the NAO Hamiltonian
+    /// @param nbands_in number of lowest eigenpairs to compute
+    DiagoScalapack(const int nlocal_in, const int nbands_in) : nlocal(nlocal_in), nbands(nbands_in) {};
+
     void diag(hamilt::Hamilt<T>* phm_in, psi::Psi<T>& psi, Real* eigenvalue_in);
 #ifdef __MPI
     // diagnolization used in parallel-k case
@@ -67,6 +71,9 @@ namespace hsolver
     int degeneracy_max = 12; // For reorthogonalized memory. 12 followes siesta.
 
     void post_processing(const int info, const std::vector<int> &vec);
+
+    const int nlocal;
+    const int nbands;
 };
 
 } // namespace hsolver
diff --git a/source/source_hsolver/hsolver_lcao.cpp b/source/source_hsolver/hsolver_lcao.cpp
index d31658d8ad..3c8bcc3ff0 100644
--- a/source/source_hsolver/hsolver_lcao.cpp
+++ b/source/source_hsolver/hsolver_lcao.cpp
@@ -141,19 +141,19 @@ void HSolverLCAO<T, Device>::hamiltSolvePsiK(hamilt::Hamilt<T>* hm, psi::Psi<T>&
     if (this->method == "scalapack_gvx")
     {
 #ifdef __MPI
-        DiagoScalapack<T> sa;
+        DiagoScalapack<T> sa(this->nlocal, this->nbands);
         sa.diag(hm, psi, eigenvalue);
 #endif
     }
 #ifdef __ELPA
     else if (this->method == "genelpa")
     {
-        DiagoElpa<T> el;
+        DiagoElpa<T> el(this->nlocal, this->nbands);
         el.diag(hm, psi, eigenvalue);
     }
     else if (this->method == "elpa")
     {
-        DiagoElpaNative<T> el;
+        DiagoElpaNative<T> el(this->nlocal, this->nbands, this->use_gpu);
         el.diag(hm, psi, eigenvalue);
     }
 #endif
@@ -161,7 +161,7 @@ void HSolverLCAO<T, Device>::hamiltSolvePsiK(hamilt::Hamilt<T>* hm, psi::Psi<T>&
     else if (this->method == "cusolver")
     {
         // Note: This branch will only be executed in the single-process case
-        DiagoCusolver<T> cu;
+        DiagoCusolver<T> cu(this->nlocal, this->nbands);
         hamilt::MatrixBlock<T> hk, sk;
         hm->matrix(hk, sk);
         cu.diag(hk, sk, psi, eigenvalue);
@@ -169,14 +169,14 @@ void HSolverLCAO<T, Device>::hamiltSolvePsiK(hamilt::Hamilt<T>* hm, psi::Psi<T>&
 #ifdef __CUSOLVERMP
     else if (this->method == "cusolvermp")
     {
-        DiagoCusolverMP<T> cm;
+        DiagoCusolverMP<T> cm(this->nlocal, this->nbands);
         cm.diag(hm, psi, eigenvalue);
     }
 #endif
 #endif
     else if (this->method == "lapack") // only for single core
     {
-        DiagoLapack<T> la;
+        DiagoLapack<T> la(this->nlocal, this->nbands);
         la.diag(hm, psi, eigenvalue);
     }
     else
@@ -253,23 +253,23 @@ void HSolverLCAO<T, Device>::parakSolve(hamilt::Hamilt<T>* pHamilt,
             /// solve eigenvector and eigenvalue for H(k)
             if (this->method == "scalapack_gvx")
             {
-                DiagoScalapack<T> sa;
+                DiagoScalapack<T> sa(this->nlocal, this->nbands);
                 sa.diag_pool(hk_pool, sk_pool, psi_pool, &(pes->ekb(ik_global, 0)), k2d.POOL_WORLD_K2D);
             }
             else if (this->method == "lapack")
             {
-                DiagoLapack<T> la;
+                DiagoLapack<T> la(this->nlocal, this->nbands);
                 la.diag_pool(hk_pool, sk_pool, psi_pool, &(pes->ekb(ik_global, 0)), k2d.POOL_WORLD_K2D);
             }
 #ifdef __ELPA
             else if (this->method == "genelpa")
             {
-                DiagoElpa<T> el;
+                DiagoElpa<T> el(this->nlocal, this->nbands);
                 el.diag_pool(hk_pool, sk_pool, psi_pool, &(pes->ekb(ik_global, 0)), k2d.POOL_WORLD_K2D);
             }
             else if (this->method == "elpa")
             {
-                DiagoElpaNative<T> el;
+                DiagoElpaNative<T> el(this->nlocal, this->nbands, this->use_gpu);
                 el.diag_pool(hk_pool, sk_pool, psi_pool, &(pes->ekb(ik_global, 0)), k2d.POOL_WORLD_K2D);
             }
 #endif
@@ -427,7 +427,7 @@ void HSolverLCAO<T, Device>::parakSolve_cusolver(hamilt::Hamilt<T>* pHamilt,
         if(kpt_assigned != -1)
         {
             psi_local.resize(1, ncol, nrow);
-            DiagoCusolver<T> cu{};
+            DiagoCusolver<T> cu(this->nlocal, this->nbands);
             hamilt::MatrixBlock<T> hk_local = hamilt::MatrixBlock<T>{
                     hk_mat.data(), (size_t)nrow, (size_t)ncol,
                     mat_para_local.desc};
diff --git a/source/source_hsolver/hsolver_lcao.h b/source/source_hsolver/hsolver_lcao.h
index 60a777e3a4..3b54d92d1f 100644
--- a/source/source_hsolver/hsolver_lcao.h
+++ b/source/source_hsolver/hsolver_lcao.h
@@ -19,8 +19,11 @@ class HSolverLCAO
                 const std::string method_in,
                 const int kpar_lcao_in,
                 const int nlocal_in,
-                const double nelec_in)
-        : ParaV(ParaV_in), method(method_in), kpar_lcao(kpar_lcao_in), nlocal(nlocal_in), nelec(nelec_in) {};
+                const int nbands_in,
+                const double nelec_in,
+                const bool use_gpu_in)
+        : ParaV(ParaV_in), method(method_in), kpar_lcao(kpar_lcao_in), nlocal(nlocal_in), nbands(nbands_in),
+          nelec(nelec_in), use_gpu(use_gpu_in) {};
 
     void solve(hamilt::Hamilt<TK>* pHamilt,
                psi::Psi<TK>& psi,
@@ -49,8 +52,10 @@ class HSolverLCAO
     const std::string method;
 
     const int kpar_lcao; // number of pools for LCAO diagonalization
-    const int nlocal;    // global dimension of the NAO Hamiltonian, only used by the pexsi branch
+    const int nlocal;    // global dimension of the NAO Hamiltonian
+    const int nbands;    // number of bands to be solved for
     const double nelec;  // total number of electrons, only used by the pexsi branch
+    const bool use_gpu;  // true if running on GPU, only used by the native-ELPA branch
 };
 
 } // namespace hsolver
diff --git a/source/source_hsolver/test/diago_lapack_test.cpp b/source/source_hsolver/test/diago_lapack_test.cpp
index 8620eaafa4..e3d358052e 100644
--- a/source/source_hsolver/test/diago_lapack_test.cpp
+++ b/source/source_hsolver/test/diago_lapack_test.cpp
@@ -1,8 +1,5 @@
 // Author: Zhang Xiaoyang
 // A modified version of diago_lcao_test.cpp
-#define private public
-#include "source_io/module_parameter/parameter.h"
-#undef private
 // Remove some useless functions and dependencies. Serialized the full code
 // and refactored some function.
 
@@ -184,21 +181,14 @@ class DiagoLapackPrepare
         hmtest.ncol = nlocal;
     }
 
-    void set_env()
-    {
-        PARAM.sys.nlocal = nlocal;
-        PARAM.input.nbands = nbands;
-    }
-
     void diago()
     {
         this->pb2d();
         this->print_hs();
-        this->set_env();
 
         for (int i = 0; i < REPEATRUN; i++)
         {
-            hsolver::DiagoLapack<T> dh;
+            hsolver::DiagoLapack<T> dh(nlocal, nbands);
             dh.diag(&hmtest, psi, e_solver.data());
             // dh->diag(&hmtest, psi, e_solver.data());
         }
diff --git a/source/source_hsolver/test/diago_lcao_cusolver_test.cpp b/source/source_hsolver/test/diago_lcao_cusolver_test.cpp
index 9f340419b7..3a5154af10 100644
--- a/source/source_hsolver/test/diago_lcao_cusolver_test.cpp
+++ b/source/source_hsolver/test/diago_lcao_cusolver_test.cpp
@@ -1,8 +1,5 @@
 #include "source_hsolver/diago_scalapack.h"
 #include "source_hsolver/test/diago_elpa_utils.h"
-#define private public
-#include "source_io/module_parameter/parameter.h"
-#undef private
 #include "mpi.h"
 #include "string.h"
 
@@ -204,8 +201,6 @@ class DiagoPrepare
 
     void set_env()
     {
-        PARAM.sys.nlocal = nlocal;
-        PARAM.input.nbands = nbands;
         GlobalV::DSIZE = dsize;
     }
 
@@ -225,13 +220,13 @@ class DiagoPrepare
             hmtest.s_local = this->s_local;
             if (ks_solver == "scalapack_gvx")
             {
-                hsolver::DiagoScalapack<T> dh;
+                hsolver::DiagoScalapack<T> dh(nlocal, nbands);
                 dh.diag(&hmtest, psi, e_solver.data());
             }
     #ifdef __CUDA
             else if (ks_solver == "cusolver")
                 {
-                    hsolver::DiagoCusolver<T> dh;
+                    hsolver::DiagoCusolver<T> dh(nlocal, nbands);
                     hamilt::MatrixBlock<T> h_mat, s_mat;
                     hmtest.matrix(h_mat, s_mat);
                     dh.diag(h_mat, s_mat, psi, e_solver.data());
diff --git a/source/source_hsolver/test/diago_lcao_test.cpp b/source/source_hsolver/test/diago_lcao_test.cpp
index d0ed0fdae6..6d66269415 100644
--- a/source/source_hsolver/test/diago_lcao_test.cpp
+++ b/source/source_hsolver/test/diago_lcao_test.cpp
@@ -1,9 +1,6 @@
 #include "source_hsolver/diago_scalapack.h"
 #include "source_hsolver/diago_lapack.h"
 #include "source_hsolver/test/diago_elpa_utils.h"
-#define private public
-#include "source_io/module_parameter/parameter.h"
-#undef private
 #include "mpi.h"
 #include "string.h"
 
@@ -205,8 +202,6 @@ class DiagoPrepare
 
     void set_env()
     {
-        PARAM.sys.nlocal = nlocal;
-        PARAM.input.nbands = nbands;
         GlobalV::DSIZE = dsize;
     }
 
@@ -226,18 +221,18 @@ class DiagoPrepare
             hmtest.s_local = this->s_local;
             if (ks_solver == "scalapack_gvx")
             {
-                hsolver::DiagoScalapack<T> dh;
+                hsolver::DiagoScalapack<T> dh(nlocal, nbands);
                 dh.diag(&hmtest, psi, e_solver.data());
             }
             else if (ks_solver == "lapack")
             {
-                hsolver::DiagoLapack<T> la;
+                hsolver::DiagoLapack<T> la(nlocal, nbands);
                 la.diag(&hmtest, psi, e_solver.data());
             }
     #ifdef __ELPA
             else if (ks_solver == "genelpa")
             {
-                hsolver::DiagoElpa<T> dh;
+                hsolver::DiagoElpa<T> dh(nlocal, nbands);
                 dh.diag(&hmtest, psi, e_solver.data());
             }
     #endif
diff --git a/source/source_lcao/LCAO_set.cpp b/source/source_lcao/LCAO_set.cpp
index 584bfc6cb2..79ec359171 100644
--- a/source/source_lcao/LCAO_set.cpp
+++ b/source/source_lcao/LCAO_set.cpp
@@ -234,7 +234,9 @@ void LCAO_domain::init_chg_hr(
                                               ks_solver,
                                               PARAM.globalv.kpar_lcao,
                                               PARAM.globalv.nlocal,
-                                              PARAM.inp.nelec);
+                                              PARAM.inp.nbands,
+                                              PARAM.inp.nelec,
+                                              PARAM.inp.device == "gpu");
     hsolver_lcao_obj.solve(p_hamilt, psi, pelec, dm, chr, nspin, 0);
 }
 
diff --git a/source/source_lcao/module_deltaspin/cal_mw_from_lambda.cpp b/source/source_lcao/module_deltaspin/cal_mw_from_lambda.cpp
index 65cbd88608..fc2fdc06b3 100644
--- a/source/source_lcao/module_deltaspin/cal_mw_from_lambda.cpp
+++ b/source/source_lcao/module_deltaspin/cal_mw_from_lambda.cpp
@@ -525,7 +525,9 @@ void spinconstrain::SpinConstrain<std::complex<double>>::cal_mw_from_lambda(
                                                             PARAM.inp.ks_solver,
                                                             PARAM.globalv.kpar_lcao,
                                                             PARAM.globalv.nlocal,
-                                                            PARAM.inp.nelec);
+                                                            PARAM.inp.nbands,
+                                                            PARAM.inp.nelec,
+                                                            PARAM.inp.device == "gpu");
         if (this->nspin_ == 2)
         {
             dynamic_cast<hamilt::DeltaSpin<hamilt::OperatorLCAO<std::complex<double>, double>>*>(this->p_operator)

From 53da439609d8a79759df1223aad77e379dfc1f76 Mon Sep 17 00:00:00 2001
From: Critsium <tsfxwbbzxy@163.com>
Date: Mon, 3 Aug 2026 13:31:29 +0800
Subject: [PATCH 2/2] Refactor: pass basis_type/calculation into
 diag_subspace_init, drop dead test_exit_cond

Group E, the last PARAM cluster in source_hsolver. Both remaining reads were in
DiagoIterAssist, which is an all-static class.

diag_subspace_init used PARAM.inp.basis_type and PARAM.inp.calculation to pick
how the rotation matrix is applied to psi (skip the gemm entirely, apply it over
dmax rows, or over dmin rows). Both are now parameters. They are threaded rather
than added as static members: the class already carries mutable static config
(PW_DIAG_THR, SCF_ITER, ...), and coding rule 2 in AGENTS.md discourages adding
more workflow switches that can be set from several places. There are only two
call sites, so parameters are cheap here.

- HSolverLIP gains basis_type and calculation, mirroring HSolverPW which already
  receives both.
- PSIPrepare already holds basis_type as a member; calculation is taken from
  PARAM at that call site. source_psi is outside this refactor and already reads
  PARAM in the same function, so this introduces no new dependency there -- it
  just stops source_hsolver from reaching for the global itself.

test_exit_cond is deleted rather than converted. It has no caller anywhere in
the tree: DiagoCG and Diago_DavSubspace each define their own independent
test_exit_cond, and Diago_DavSubspace's already takes scf as an argument, which
is the shape this one would have needed. Its only apparent user is
test/test_diago_assist.cpp, which no CMakeLists references and which could not
compile if it were built (it includes a misspelled "diago_iter_assis.h", places
a bare statement in a class body, and calls EXPECT_EQ with one argument). That
file is left untouched -- it was already dead before this change, and removing
it is a separate call for the maintainers.

PARAM occurrences in source_hsolver production code: 5 -> 0. The only remaining
match in non-test code is a comment in simple_pexsi.cpp recording that a
hardcoded 2 should eventually become nspin.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>
---
 ...20\351\200\237\346\226\271\346\241\210.md" | 307 ++++++++++++++++++
 source/source_esolver/esolver_ks_lcaopw.cpp   |   5 +-
 source/source_hsolver/diago_iter_assist.cpp   |  36 +-
 source/source_hsolver/diago_iter_assist.h     |  10 +-
 source/source_hsolver/hsolver_lcaopw.cpp      |   6 +-
 source/source_hsolver/hsolver_lcaopw.h        |  10 +-
 .../source_hsolver/test/test_hsolver_pw.cpp   |  10 +-
 source/source_psi/psi_prepare.cpp             |   4 +-
 8 files changed, 347 insertions(+), 41 deletions(-)
 create mode 100644 "CI\346\217\220\351\200\237\346\226\271\346\241\210.md"

diff --git "a/CI\346\217\220\351\200\237\346\226\271\346\241\210.md" "b/CI\346\217\220\351\200\237\346\226\271\346\241\210.md"
new file mode 100644
index 0000000000..2d643ea937
--- /dev/null
+++ "b/CI\346\217\220\351\200\237\346\226\271\346\241\210.md"
@@ -0,0 +1,307 @@
+# ABACUS CI/CD 提速方案（可落地版）
+
+> 配套文档：《CI速度差异分析.md》（同目录，含全部实测数据）
+> 原则：每个方案给出**具体文件改动 + 预期收益 + 风险 + 验证方法**，可直接拆成 PR。
+> 已验证的关键事实：
+> - ccache 已通过 `CMakeLists.txt:299-303` 接入全部语言（含 CUDA），装了就用，无需改 CMake
+> - CMake 只用 `git log -1` 取 commit 信息（`CMakeLists.txt:163-174`），**不需要 fetch-depth: 0**
+> - `Autotest.sh -n` 是每个用例的 MPI 进程数，不是用例并行数
+> - 上游 cuda.yml 仍是 `-j4` 单 job；你的 f3e844ff7（nproc+matrix）只在本地/已关闭的 #7713
+
+---
+
+## 方案总览（按 ROI 排序）
+
+| 方案 | 改动量 | 预期收益 | 风险 | 建议批次 |
+|---|---|---|---|---|
+| B. CUDA 编译提速 | cuda.yml 2 行 | Build 33min → 8~15min | 极低 | **PR-1（立即）** |
+| D3. checkout 浅克隆 | test.yml 1 行 | 冷节点省 ~12min | 极低 | **PR-1（立即）** |
+| C. ccache 可靠化 | 2 个 yml 数行 + 1 个新 yml | 消除 2.7 vs 37.5min 波动 | 低 | PR-2 |
+| A. 镜像稳定化 | 2 个 Dockerfile 删段落 | init 8~78min → ~1min | 中（需维护者共识） | PR-3 |
+| E. toolchain 归零 | 随 A 一起做 | 冷节点省 3~10min | 低 | 随 PR-3 |
+| D. 测试并行化 | test.yml 重构 / cuda.yml | CPU 18min→~6min | 中 | PR-4 |
+
+预期总效果：**test.yml 24~100min → 12~20min；cuda.yml 53~131min → 15~25min。**
+
+---
+
+## 方案 B：CUDA 编译提速（PR-1，最先做）
+
+### B1. 裁剪 CUDA 架构（7 种 → 1~2 种）
+
+**问题**：`CMakeLists.txt:445-476` 未指定时，CUDA 12.2 下默认编 60/70/75/80/86/89/90 共 7 种架构，61 个 `.cu` 每个编 7 遍。CI 只需要能在 CI 的那块 GPU 上跑。
+
+**改动**（`.github/workflows/cuda.yml` Configure & Build 步骤）：
+
+```yaml
+      - name: Configure & Build
+        run: |
+          nvidia-smi
+          source toolchain/install/setup
+          rm -rf build
+          cmake -B build -G Ninja \
+            -DUSE_CUDA=ON \
+            -DBUILD_TESTING=ON \
+            -DENABLE_FLOAT_FFTW=ON \
+            -DCMAKE_CUDA_ARCHITECTURES=80 \   # ← 新增：按 CI 实际 GPU 填
+            -DCMAKE_INSTALL_PREFIX="${GITHUB_WORKSPACE}/install"
+          cmake --build build -j $(nproc)      # ← B2：-j4 改掉
+          cmake --install build
+```
+
+**架构怎么填**：打开最近一次 CUDA run 的 Configure & Build 日志，第一行 `nvidia-smi` 输出就是卡型：
+
+| GPU | Compute Capability | 填 |
+|---|---|---|
+| V100 | 7.0 | `70` |
+| T4 | 7.5 | `75` |
+| A100 | 8.0 | `80` |
+| A10 / RTX 30 系 | 8.6 | `86` |
+| L40 / RTX 40 系 | 8.9 | `89` |
+| H100 | 9.0 | `90` |
+
+池子里若有多种卡，填 2 个（如 `80;90`），仍比 7 种省 ~70% nvcc 工作量。**不要**用 `native`——你本地 matrix 版把 build 和 test 拆成了不同 job，可能落在不同型号的机器上。
+
+**预期收益**：33min → 8~15min（-70%~-85% nvcc 时间）。测试有效性不受影响（二进制只需在 CI GPU 上可运行）。
+
+### B2. 放开并行度 `-j4` → `-j $(nproc)`
+
+**问题**：`-j4` 自 #4032 时代写死。GPU 节点通常 16+ 核。
+
+**注意**：必须和 B1 一起做。7 架构 × 高并行 → nvcc 单进程 2~4GB 内存，容易 OOM（你之前关心的 OOM 问题就在这里）；架构裁到 1~2 种后放开并行是安全的。你本地 commit f3e844ff7 已包含此项，可直接摘出来。
+
+---
+
+## 方案 D3：checkout 浅克隆（并入 PR-1）
+
+**问题**：test.yml `fetch-depth: 0` 全历史克隆，冷节点实测 11.9min。已验证 CMake 只执行 `git log -1`，不需要历史。
+
+**改动**（`.github/workflows/test.yml`）：
+
+```yaml
+      - name: Checkout repository
+        uses: actions/checkout@v7
+        with:
+          fetch-depth: 1        # ← 0 改 1
+          submodules: 'false'
+```
+
+**风险排查**：`git submodule update --init --recursive` 不受 depth 影响；`version_check.yml` 是独立 workflow，如它依赖 tag 历史则保持它自己的 depth 不变（与本 PR 无关）。
+
+---
+
+## 方案 C：ccache 可靠化（PR-2，消除最大波动源）
+
+### C1. 显式容量 + 可观测（改动极小，先做）
+
+**问题**：两个 workflow 只挂载 `/tmp/ccache`，从未设置容量（ccache 4.x 默认约 10GB，7 架构 CUDA 产物 + 日常 churn 有驱逐压力），也从不打印命中率——慢的时候无法自证。
+
+**改动**（test.yml 和 cuda.yml 的 Build 步骤前后各加一步）：
+
+```yaml
+      - name: Setup ccache
+        run: |
+          ccache --max-size=30G
+          ccache --zero-stats
+          ccache -s
+
+      # ... Configure / Build ...
+
+      - name: ccache statistics
+        if: always()
+        run: ccache -s
+```
+
+**收益**：①容量兜底；②每次运行的日志里直接看到 `cache hit rate`，下次再出现"2min vs 37min"一眼定位是缓存问题还是机器问题。这也是给维护者讲故事的证据。
+
+### C2. 确认动态池的缓存持久化（需要 runner 管理员配合）
+
+**问题**：6/17 #7476 切到动态 runner 池（K8s Pod），`/tmp/ccache` 是宿主机目录还是 emptyDir 决定了缓存能不能跨 run 存活。实测同一 PR 越推越快，说明**部分节点**是能保住的，但新扩的节点是冷的。
+
+**行动**（在 issue/PR 里 @ 维护者确认两件事）：
+1. `/tmp/ccache` 是否 hostPath/PV？多节点池是否每节点独立一份？（独立 → 命中率随节点数稀释）
+2. 若是 emptyDir，改为 hostPath 或 PVC；
+3. 终极方案：ccache 4.x 支持二级远程缓存，池内共享一份：
+   ```yaml
+   env:
+     CCACHE_REMOTE_STORAGE: "http://内网缓存服务|layout=bazel"   # 或 redis:/s3
+     CCACHE_REMOTE_ONLY: "false"   # 本地+远程双层
+   ```
+   这样新扩的冷节点也能命中远程缓存，**直接消除"节点运气"因子**。
+
+### C3. cache-warmer：给池子预热（消除"合并风暴后第一批 PR 变慢"）
+
+**问题**：大重构合入 develop 后，全池缓存失效，之后第一批 PR 的 CI 全部变慢（7/29-31 的 347 文件/125 头文件风暴就是案例）。
+
+**方案**：新增 `.github/workflows/cache_warmer.yml`，在 develop 有合并时自动"只编不测"，让池子缓存始终贴着最新 develop：
+
+```yaml
+name: Cache Warmer
+on:
+  push:
+    branches: [develop]
+  workflow_dispatch:
+
+concurrency:
+  group: cache-warmer
+  cancel-in-progress: true
+
+jobs:
+  warm:
+    runs-on: X64
+    if: github.repository_owner == 'deepmodeling'
+    container:
+      image: ghcr.io/deepmodeling/abacus-gnu   # 方案 A 落地后换稳定 tag
+      volumes:
+        - /tmp/ccache:/github/home/.ccache
+    steps:
+      - uses: actions/checkout@v7
+        with:
+          fetch-depth: 1
+          submodules: recursive
+      - name: Build only (prime ccache)
+        run: |
+          sudo apt-get update && sudo apt-get install -y ccache ninja-build gfortran
+          cmake -B build -G Ninja -DBUILD_TESTING=ON
+          cmake --build build -j $(nproc)
+          ccache -s
+```
+
+GPU 池同理可加一条 `runs-on: gpu` 的 warm job（USE_CUDA=ON）。成本是 runner 时间，收益是所有 PR 的 Build 稳定在"只差 PR 自己改动"的水平。
+
+---
+
+## 方案 A：镜像稳定化（PR-3，收益最大，需维护者共识）
+
+### 问题回顾
+
+- `devcontainer.yml`：**每次 push 到 develop** 都重建并推送 `abacus-{gnu,intel,cuda}:latest`；
+- `Dockerfile.gnu:31-41` / `Dockerfile.cuda:27-36`：cache-bust（`ADD .../develop /dev/null`）+ **在镜像里完整编译一份 ABACUS**；
+- 结果：镜像几乎天天变 → 动态 runner 天天重拉（init 实测 8~78min）；
+- 而 CI 从不用镜像里那份二进制（每个 PR 都 checkout 自己重编）——**镜像内编译对 CI 是纯浪费**。
+
+### 改法 A1（推荐：删，不加新东西，最容易过审）
+
+从 `Dockerfile.gnu` 删 31-41 行、`Dockerfile.cuda` 删 27-36 行（cache-bust + 镜像内编译段落）。效果：
+
+- 镜像内容只在**依赖本身变化**（改 Dockerfile）时才变 → devcontainer 每次重建产出的镜像层摘要不变 → runner 端 `docker pull` 变 no-op → **Initialize containers 稳定 ~0-1min**；
+- 镜像体积：gnu 减 ~1GB、cuda 减数 GB（abacus 构建产物 + 源码）→ 真需要重拉时也更快；
+- 符合 AGENTS.md「复用现有 Docker 资产、不新增容器」——这是做减法。
+
+**需要处理的一个副作用**：这两个镜像的定位本来是"给用户体验 ABACUS"（Dockerfile 注释写明 aimed for evaluating ABACUS），删掉内置二进制后，想直接 `docker run` 用 abacus 的用户会受影响。PR 里提供两个选项让维护者选：
+- 选项 1：接受——evaluating 用户改用 release 二进制/自行编译，README/文档更新一句；
+- 选项 2：把"含 abacus 的 eval 镜像"拆成单独的 `Dockerfile.eval`（CI 用纯净 deps 镜像，eval 镜像照旧天天新）——代价是多一个 Dockerfile，需要在 PR 里按 AGENTS.md 要求写明理由。
+
+### 改法 A2（可叠加：拉镜像换国内源）
+
+`Dockerfile.gnu` 注释里已写明镜像同时发布在 `registry.dp.tech/deepmodeling/abacus-*`，devcontainer.yml 也推 `dp-harbor-registry.us-east-1.cr.aliyuncs.com`。runner 在国内，却让它们从 ghcr.io 拉。
+
+```yaml
+    container:
+      image: registry.dp.tech/deepmodeling/abacus-cuda   # 原: ghcr.io/deepmodeling/abacus-cuda
+```
+
+一行改动。先验证匿名可拉（找台机器 `docker pull registry.dp.tech/deepmodeling/abacus-gnu:latest` 试试），拉不通就用 dp-harbor 地址（需配 docker login，workflow 里用 secrets）。
+
+---
+
+## 方案 E：toolchain 时间归零（随 PR-3 的镜像一起改）
+
+**问题**：7/1 #7449 起每次运行跑完整 `install_stage4.sh`（dftd4/cereal/rapidjson/**libtorch ~200MB 下载**/libnpy/libri/libcomm/nep）。热节点 20 秒，冷节点数分钟；且 **cuda.yml 的构建根本没开 MLALGO，libtorch 白下载**。
+
+**改法**：
+1. 把 stage4 产物预装进 deps 镜像（Dockerfile 里加一段 stage4 安装，libtorch gnu 镜像已有）；
+2. CI 的 toolchain 步骤保留（幂等，检测到已装会秒过）——热路径不变，冷路径归零；
+3. cuda.yml 如确认不需要 libtorch/LibRI，toolchain 步骤可改为只装必需项（`install_dftd4.sh` 等单独调用），或 configure 时显式关掉对应 feature 避免误装。
+
+---
+
+## 方案 D：测试提速（PR-4，改动最大放最后）
+
+### D1. test.yml：26 个串行步骤 → 分组并行
+
+**现状**：单 job 里 16 个 Module_* 单测 + 10 个集成套件串行跑 ~18min（实测：01_PW 4.4、08_EXX 3.0、03_NAO_multik 2.6、09_DeePKS 2.1…）。
+
+**轻量版（diff 小，推荐先上）**：16 个 Module_* 步骤合并成一个并行调用：
+
+```yaml
+      - name: All Module Unittests
+        env:
+          GTEST_COLOR: 'yes'
+          OMP_NUM_THREADS: '2'
+        run: |
+          ctest --test-dir build -j $(nproc) --timeout 1700 \
+            -R 'MODULE_' -E PERF_MODULE_HSOLVER_KERNELS
+```
+
+集成套件（每个内部 Autotest.sh 串行跑用例）保持分步即可。**收益：单测 ~4.5min → ~1min。**
+
+**完整版（参照你 cuda.yml 的 matrix 改法）**：build job 产出 `install/` artifact → 10 个集成套件 + 1 个单测 job 并行下载运行。18min → ~5min，但 yml 改动大，建议等 PR-1~3 合入后再提。
+
+### D2. 01_PW GPU 套件瘦身（cuda.yml）
+
+**问题**：#7690 加的 73 个用例 `-n 1` 串行，+4.8min/次，且用例是为 CPU 写的。
+
+**改法**：
+1. `-n 1` → `-n 2`（与其他 GPU 套件一致）；
+2. `CASES_GPU.txt` 拆两份：`CASES_GPU_SMOKE.txt`（~20 个代表用例，PR 跑）+ 全量留给定时任务：
+
+```yaml
+on:
+  pull_request:
+  schedule:
+    - cron: '0 18 * * *'   # 每晚全量（UTC）
+  workflow_dispatch:
+
+jobs:
+  gpu-test:
+    steps:
+      - name: Select case list
+        run: |
+          if [ "${{ github.event_name }}" = "pull_request" ]; then
+            echo "CASES=CASES_GPU_SMOKE.txt" >> "$GITHUB_ENV"
+          else
+            echo "CASES=CASES_GPU.txt" >> "$GITHUB_ENV"
+          fi
+      - name: Test 01_PW on GPU
+        run: |
+          cd tests/01_PW && bash ../integrate/Autotest.sh -n 2 -a abacus -f "$CASES"
+```
+
+---
+
+## 落地路线图（给导师汇报版）
+
+| 批次 | 内容 | 文件 | 预期收益 | 过审难度 |
+|---|---|---|---|---|
+| **PR-1** | B1 架构裁剪 + B2 nproc + D3 浅克隆 | cuda.yml、test.yml | CUDA Build 33→8~15min；冷节点 checkout 省 12min | ★ 极易（几行配置） |
+| **PR-2** | C1 ccache 容量+统计；C3 cache-warmer | 2 个 yml + 1 个新 yml | 消除 2.7 vs 37.5min 波动；风暴后自动回暖 | ★★ 易（纯新增步骤） |
+| **PR-3** | A1 Dockerfile 删 cache-bust+镜像内编译（+E stage4 预装 +A2 国内源） | 2 个 Dockerfile、devcontainer.yml、2 个 yml | init 8~78→~1min；toolchain 归零 | ★★★ 中（改镜像定位，需共识） |
+| **PR-4** | D1 单测合并并行 + D2 GPU 套件瘦身（之后视情况上完整 matrix） | test.yml、cuda.yml、CASES_GPU | CPU 测试 18→~6min；GPU 测试省 ~3min | ★★ 易~中 |
+
+**验证方法（每个 PR 必做，写进 PR 描述，符合 AGENTS.md「报告确切验证」要求）**：
+1. 从 fork 提交后观察本 PR 的 CI 运行时长对比（注意：fork PR 改 workflow 文件首次需维护者 approve，先发 PR 后 @ 维护者）；
+2. 贴 `ccache -s` 的 hit rate 前后对比；
+3. 贴 Initialize containers / Build 步骤耗时的前后截图；
+4. 确认 GPU 测试全部通过（架构裁剪后二进制必须能在 CI 卡上跑——这也是 nvidia-smi 要先看的原因）。
+
+**风险提示（写进 PR）**：
+- B1 若填错架构 → GPU 测试全挂（立即可见，回退一行即可）；
+- B2 不做 B1 直接加并行 → nvcc 内存 ×7 架构可能 OOM；
+- A1 影响"用镜像体验 abacus"的用户路径，PR 里给维护者两个选项；
+- C3 cache-warmer 占用 runner 时间，建议 `concurrency` 防堆积。
+
+---
+
+## 附：为什么这些方案能成立（事实锚点）
+
+| 方案依赖的事实 | 出处 |
+|---|---|
+| ccache 自动接入（含 CUDA） | CMakeLists.txt:299-303 |
+| 构建只需 git log -1 | CMakeLists.txt:163-174 |
+| 7 架构默认列表 | CMakeLists.txt:445-476 |
+| -j4 写死 | 上游 cuda.yml `cmake --build build -j4` |
+| 镜像每次合并重建 + cache-bust + 镜像内编译 | devcontainer.yml；Dockerfile.gnu:31-41；Dockerfile.cuda:27-36 |
+| 国内镜像源已存在 | Dockerfile.gnu:4 注释；devcontainer.yml:39 |
+| 01_PW GPU -n1 串行 73 例 | #7690（4b6ce1c63） |
+| init 8~78min、Build 2.7 vs 37.5min 实测 | GitHub API runs/jobs 数据（见分析文档） |
diff --git a/source/source_esolver/esolver_ks_lcaopw.cpp b/source/source_esolver/esolver_ks_lcaopw.cpp
index 036648394b..95165d262b 100644
--- a/source/source_esolver/esolver_ks_lcaopw.cpp
+++ b/source/source_esolver/esolver_ks_lcaopw.cpp
@@ -135,7 +135,10 @@ namespace ModuleESolver
         hsolver::DiagoIterAssist<T>::PW_DIAG_NMAX = PARAM.inp.pw_diag_nmax;
         bool skip_charge = PARAM.inp.calculation == "nscf" ? true : false;
 
-        hsolver::HSolverLIP<T> hsolver_lip_obj(this->pw_wfc, PARAM.globalv.use_uspp);
+        hsolver::HSolverLIP<T> hsolver_lip_obj(this->pw_wfc,
+                                               PARAM.globalv.use_uspp,
+                                               PARAM.inp.basis_type,
+                                               PARAM.inp.calculation);
         hsolver_lip_obj.solve(static_cast<hamilt::Hamilt<T>*>(this->p_hamilt), *this->stp.template get_psi_t<T, base_device::DEVICE_CPU>(), this->pelec, 
           *this->psi_local, skip_charge,ucell.tpiba,ucell.nat);
 
diff --git a/source/source_hsolver/diago_iter_assist.cpp b/source/source_hsolver/diago_iter_assist.cpp
index b547febefe..8436e4a26b 100644
--- a/source/source_hsolver/diago_iter_assist.cpp
+++ b/source/source_hsolver/diago_iter_assist.cpp
@@ -1,5 +1,4 @@
 #include "diago_iter_assist.h"
-#include "source_io/module_parameter/parameter.h"
 #include "source_base/complexmatrix.h"
 #include "source_base/constants.h"
 #include "source_base/global_function.h"
@@ -177,6 +176,8 @@ void DiagoIterAssist<T, Device>::diag_subspace_init(hamilt::Hamilt<T, Device>* p
     int psi_nc,
     psi::Psi<T, Device>& evc,
     Real* en,
+    const std::string& basis_type,
+    const std::string& calculation,
     const std::function<void(T*, const int)>& add_to_hcc,
     const std::function<void(const T* const, const int, const int)>& export_vcc)
 {
@@ -330,13 +331,13 @@ void DiagoIterAssist<T, Device>::diag_subspace_init(hamilt::Hamilt<T, Device>* p
     //=======================
     // diagonize the H-matrix
     //=======================
-    if ((PARAM.inp.basis_type == "lcao" || PARAM.inp.basis_type == "lcao_in_pw") && PARAM.inp.calculation == "nscf")
+    if ((basis_type == "lcao" || basis_type == "lcao_in_pw") && calculation == "nscf")
     {
         GlobalV::ofs_running << " Not do zgemm to get evc." << std::endl;
     }
-    else if ((PARAM.inp.basis_type == "lcao" || PARAM.inp.basis_type == "lcao_in_pw" || PARAM.inp.basis_type == "pw")
-             && (PARAM.inp.calculation == "scf" || PARAM.inp.calculation == "md"
-                 || PARAM.inp.calculation == "relax")) // pengfei 2014-10-13
+    else if ((basis_type == "lcao" || basis_type == "lcao_in_pw" || basis_type == "pw")
+             && (calculation == "scf" || calculation == "md"
+                 || calculation == "relax")) // pengfei 2014-10-13
     {
         // because psi and evc are different here,
         // I think if psi and evc are the same,
@@ -637,31 +638,6 @@ void DiagoIterAssist<T, Device>::diag_subspace_psi(const T* hcc,
     ModuleBase::timer::end("DiagoIterAssist", "diag_subspace_psi");
 }
 
-template <typename T, typename Device>
-bool DiagoIterAssist<T, Device>::test_exit_cond(const int& ntry, const int& notconv)
-{
-    //================================================================
-    // If this logical function is true, need to do diag_subspace
-    // and cg again.
-    //================================================================
-
-    bool scf = true;
-    if (PARAM.inp.calculation == "nscf") {
-        scf = false;
-}
-
-    // If ntry <=5, try to do it better, if ntry > 5, exit.
-    const bool f1 = (ntry <= 5);
-
-    // In non-self consistent calculation, do until totally converged.
-    const bool f2 = ((!scf && (notconv > 0)));
-
-    // if self consistent calculation, if not converged > 5,
-    // using diag_subspace and cg method again. ntry++
-    const bool f3 = ((scf && (notconv > 5)));
-    return (f1 && (f2 || f3));
-}
-
 template class DiagoIterAssist<std::complex<float>, base_device::DEVICE_CPU>;
 template class DiagoIterAssist<std::complex<double>, base_device::DEVICE_CPU>;
 #if ((defined __CUDA) || (defined __ROCM))
diff --git a/source/source_hsolver/diago_iter_assist.h b/source/source_hsolver/diago_iter_assist.h
index e9c4c00520..8302933840 100644
--- a/source/source_hsolver/diago_iter_assist.h
+++ b/source/source_hsolver/diago_iter_assist.h
@@ -7,6 +7,7 @@
 #include "source_psi/psi.h"
 
 #include <functional>
+#include <string>
 
 namespace hsolver
 {
@@ -61,7 +62,10 @@ class DiagoIterAssist
     /// @param psi_nc number of columns (nbasis)
     /// @param evc new wavefunction
     /// @param en eigenenergies
-    /// @note exception handle: if there is no operator initialized in Hamilt, will directly copy value from psi to evc, 
+    /// @param basis_type "lcao", "lcao_in_pw" or "pw"; together with calculation it selects
+    /// how the rotation matrix is applied to psi
+    /// @param calculation "scf", "nscf", "md", "relax", ...
+    /// @note exception handle: if there is no operator initialized in Hamilt, will directly copy value from psi to evc,
     /// and return all - zero eigenenergies.
     static void diag_subspace_init(
             hamilt::Hamilt<T, Device>* pHamilt,
@@ -70,6 +74,8 @@ class DiagoIterAssist
             int psi_nc,
             psi::Psi<T, Device> &evc,
             Real* en,
+            const std::string& basis_type,
+            const std::string& calculation,
             const std::function<void(T*, const int)>& add_to_hcc = [](T* null, const int n) {},
             const std::function<void(const T* const, const int, const int)>& export_vcc = [](const T* null, const int n, const int m) {});
 
@@ -120,8 +126,6 @@ class DiagoIterAssist
                               psi::Psi<T, Device>& evc,
                               Real* en);
 
-    static bool test_exit_cond(const int& ntry, const int& notconv);
-
   private:
     constexpr static const Device* ctx = {};
 
diff --git a/source/source_hsolver/hsolver_lcaopw.cpp b/source/source_hsolver/hsolver_lcaopw.cpp
index f61035ef6a..66a14263ee 100644
--- a/source/source_hsolver/hsolver_lcaopw.cpp
+++ b/source/source_hsolver/hsolver_lcaopw.cpp
@@ -71,8 +71,10 @@ void HSolverLIP<T>::solve(hamilt::Hamilt<T>* pHamilt, // ESolver_KS_PW::p_hamilt
             transform.get_pointer(), // transform matrix between lcao and pw
             transform.get_nbands(),
             transform.get_nbasis(),
-            psi,                                  // psi in pw basis
-            eigenvalues.data() + ik * pes->ekb.nc // eigenvalues
+            psi,                                   // psi in pw basis
+            eigenvalues.data() + ik * pes->ekb.nc, // eigenvalues
+            this->basis_type,
+            this->calculation
 #ifdef __EXX
             ,
             add_exx_to_subspace_hamilt,
diff --git a/source/source_hsolver/hsolver_lcaopw.h b/source/source_hsolver/hsolver_lcaopw.h
index 49bbe63cb9..c226376b0c 100644
--- a/source/source_hsolver/hsolver_lcaopw.h
+++ b/source/source_hsolver/hsolver_lcaopw.h
@@ -18,8 +18,11 @@ class HSolverLIP
     using Real = typename GetTypeReal<T>::type;
 
   public:
-    HSolverLIP(ModulePW::PW_Basis_K* wfc_basis_in, const bool use_uspp_in)
-        : wfc_basis(wfc_basis_in), use_uspp(use_uspp_in) {};
+    HSolverLIP(ModulePW::PW_Basis_K* wfc_basis_in,
+               const bool use_uspp_in,
+               const std::string basis_type_in,
+               const std::string calculation_in)
+        : wfc_basis(wfc_basis_in), use_uspp(use_uspp_in), basis_type(basis_type_in), calculation(calculation_in) {};
 
     /// @brief solve function for lcao_in_pw
     /// @param pHamilt interface to hamilt
@@ -39,6 +42,9 @@ class HSolverLIP
     ModulePW::PW_Basis_K* wfc_basis = nullptr;
 
     const bool use_uspp; // true if ultrasoft pseudopotentials are in use
+
+    const std::string basis_type;  // "lcao_in_pw" for this solver
+    const std::string calculation; // "scf", "nscf", "md", "relax", ...
 };
 
 } // namespace hsolver
diff --git a/source/source_hsolver/test/test_hsolver_pw.cpp b/source/source_hsolver/test/test_hsolver_pw.cpp
index d484a5ed92..2e6a07fffb 100644
--- a/source/source_hsolver/test/test_hsolver_pw.cpp
+++ b/source/source_hsolver/test/test_hsolver_pw.cpp
@@ -375,9 +375,15 @@ TEST_F(TestHSolverPW, SolveLcaoInPW) {
     elecstate_test.ekb.c[1] = 2.0;
     
     hsolver::HSolverLIP<std::complex<float>> hs_f_lip
-        = hsolver::HSolverLIP<std::complex<float>>(&pwbk, PARAM.sys.use_uspp);
+        = hsolver::HSolverLIP<std::complex<float>>(&pwbk,
+                                                   PARAM.sys.use_uspp,
+                                                   PARAM.input.basis_type,
+                                                   PARAM.input.calculation);
     hsolver::HSolverLIP<std::complex<double>> hs_d_lip
-        = hsolver::HSolverLIP<std::complex<double>>(&pwbk, PARAM.sys.use_uspp);
+        = hsolver::HSolverLIP<std::complex<double>>(&pwbk,
+                                                    PARAM.sys.use_uspp,
+                                                    PARAM.input.basis_type,
+                                                    PARAM.input.calculation);
     hs_f_lip.solve(&hamilt_test_f, psi_test_cf, &elecstate_test,transform_test_cf, true,0.0,0);
     EXPECT_DOUBLE_EQ(hsolver::DiagoIterAssist<std::complex<float>>::avg_iter, 0.0);
     for (int i = 0; i < psi_test_cf.size(); i++)
diff --git a/source/source_psi/psi_prepare.cpp b/source/source_psi/psi_prepare.cpp
index eada5f13d1..5458d58992 100644
--- a/source/source_psi/psi_prepare.cpp
+++ b/source/source_psi/psi_prepare.cpp
@@ -198,7 +198,9 @@ void PSIPrepare<T, Device>::initialize_psi(Psi<std::complex<double>>* psi,
                                                                              nbands_start,
                                                                              nbasis,
                                                                              *(kspw_psi),
-                                                                             etatom.data());
+                                                                             etatom.data(),
+                                                                             this->basis_type,
+                                                                             PARAM.inp.calculation);
                 }
                 else
                 {