[Fix] Eliminate unnecessary vkb CPU allocation on GPU path (#7296)

* [Fix] Skip unnecessary vkb CPU allocation on GPU path in pseudopot_cell_vnl

On GPU path, vkb.create(nkb, npwx) allocates CPU ComplexMatrix memory
that is never used — getvnl() writes directly to GPU buffers (c_vkb/z_vkb).
The only consumer of vkb.nc metadata is the leading dimension in gemm/gemv.

This wastes nkb*npwx*16 bytes of CPU memory (~3.2 GB for large systems).

Changes:
- Add vkbnc member to store column dimension independently
- Guard vkb.create() behind !use_gpu_ in init()
- Replace all ppcell->vkb.nc with ppcell->vkbnc (op_pw_nl.cpp, hamilt_pw.cpp)
- Add lazy-allocation guard in getgradq_vnl() for GPU Velocity path

Tested: GPU build + 28/28 kernel UTs + 38/40 GPU integration tests
(2 pre-existing failures: scf_bpcg, scf_out_wf)

* Update source/source_pw/module_pwdft/vnl_pw_grad.cpp

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Update source/source_pw/module_pwdft/vnl_pw_grad.cpp

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

* Revert "Update source/source_pw/module_pwdft/vnl_pw_grad.cpp"

This reverts commit b71e3fe.

* Revert "Update source/source_pw/module_pwdft/vnl_pw_grad.cpp"

This reverts commit a0f43dc.

---------

Co-authored-by: Copilot <175728472+Copilot@users.noreply.github.com>

Author: Cstandardlib

source/source_pw/module_pwdft/hamilt_pw.cpp

@@ -272,7 +272,7 @@ void HamiltPW<T, Device>::sPsi(const T* psi_in, // psi
                           this->ppcell->nkb,
                           &one,
                           this->vkb,
-                          this->ppcell->vkb.nc,
+                          this->ppcell->vkbnc,
                           psi_in,
                           inc,
                           &zero,
@@ -288,7 +288,7 @@ void HamiltPW<T, Device>::sPsi(const T* psi_in, // psi
                           npw,
                           &one,
                           this->vkb,
-                          this->ppcell->vkb.nc,
+                          this->ppcell->vkbnc,
                           psi_in,
                           nrow,
                           &zero,
@@ -360,7 +360,7 @@ void HamiltPW<T, Device>::sPsi(const T* psi_in, // psi
                           this->ppcell->nkb,
                           &one,
                           this->vkb,
-                          this->ppcell->vkb.nc,
+                          this->ppcell->vkbnc,
                           ps,
                           inc,
                           &one,
@@ -376,7 +376,7 @@ void HamiltPW<T, Device>::sPsi(const T* psi_in, // psi
                           this->ppcell->nkb,
                           &one,
                           this->vkb,
-                          this->ppcell->vkb.nc,
+                          this->ppcell->vkbnc,
                           ps,
                           this->ppcell->nkb,
                           &one,

source/source_pw/module_pwdft/op_pw_nl.cpp

@@ -172,7 +172,7 @@ void Nonlocal<OperatorPW<T, Device>>::add_nonlocal_pp(T *hpsi_in, const T *becp,
             this->ppcell->nkb,
             &this->one,
             this->vkb,
-            this->ppcell->vkb.nc,
+            this->ppcell->vkbnc,
             this->ps,
             inc,
             &this->one,
@@ -197,7 +197,7 @@ void Nonlocal<OperatorPW<T, Device>>::add_nonlocal_pp(T *hpsi_in, const T *becp,
             this->ppcell->nkb,
             &this->one,
             this->vkb,
-            this->ppcell->vkb.nc,
+            this->ppcell->vkbnc,
             this->ps,
             npm,
             &this->one,
@@ -251,7 +251,7 @@ void Nonlocal<OperatorPW<T, Device>>::act(
                 nkb,
                 &this->one,
                 this->vkb,
-                this->ppcell->vkb.nc,
+                this->ppcell->vkbnc,
                 tmpsi_in,
                 inc,
                 &this->zero,
@@ -276,7 +276,7 @@ void Nonlocal<OperatorPW<T, Device>>::act(
                 this->npw,
                 &this->one,
                 this->vkb,
-                this->ppcell->vkb.nc,
+                this->ppcell->vkbnc,
                 tmpsi_in,
                 max_npw,
                 &this->zero,

source/source_pw/module_pwdft/vnl_pw.cpp

@@ -214,10 +214,17 @@ void pseudopot_cell_vnl::init(const UnitCell& ucell,
     // dq+4)*cell_factor;
     this->lmaxq = 2 * this->lmaxkb + 1;
     int npwx = this->wfcpw->npwk_max;
+    this->vkbnc = npwx;
     if (nkb > 0 && allocate_vkb)
     {
-        vkb.create(nkb, npwx);
-        ModuleBase::Memory::record("VNL::vkb", nkb * npwx * sizeof(std::complex<double>));
+        if (!this->use_gpu_)
+        {
+            vkb.create(nkb, npwx);
+            ModuleBase::Memory::record("VNL::vkb", nkb * npwx * sizeof(std::complex<double>));
+        }
+        // GPU path: vkb ComplexMatrix is not allocated.
+        // Column dimension is stored in vkbnc for gemm/gemv leading dimension.
+        // Actual GPU buffers (c_vkb/z_vkb) are allocated below.
     }
 
     // this->nqx = 10000;		// calculted in allocate_nlpot.f90

source/source_pw/module_pwdft/vnl_pw.h

@@ -108,6 +108,10 @@ class pseudopot_cell_vnl
     std::complex<double>*** vkb_alpha;
     Structure_Factor* psf = nullptr;
 
+    // Column dimension of vkb matrix (= npwx), used as leading dimension in gemm/gemv.
+    // On GPU path vkb ComplexMatrix is not allocated to save CPU memory; this stores the dimension.
+    int vkbnc = 0;
+
     // other variables
     std::complex<double> Cal_C(int alpha, int lu, int mu, int L, int M);
 

source/source_pw/module_pwdft/vnl_pw_grad.cpp

@@ -91,6 +91,11 @@ void pseudopot_cell_vnl::getgradq_vnl(const UnitCell& ucell,
 
 	ModuleBase::YlmReal::grad_Ylm_Real(x1, npw, gk, ylm, dylm[0], dylm[1], dylm[2]);
 
+	// GPU path skips vkb allocation in init(); allocate now if needed
+	if (this->vkb.nc == 0 && this->nkb > 0 && this->vkbnc > 0) {
+		this->vkb.create(this->nkb, this->vkbnc);
+	}
+
 	int jkb = 0;
 	for(int it = 0;it < ucell.ntype;it++)
 	{