From e3e5027c89db567bcdc2470c63f391a3cac2baac Mon Sep 17 00:00:00 2001
From: dzzz2001 <dzcpku2001@163.com>
Date: Thu, 22 Jan 2026 15:25:10 +0800
Subject: [PATCH] change col_major to row_major

---
 source/source_hsolver/kernels/cuda/diag_cusolvermp.cu | 3 ++-
 1 file changed, 2 insertions(+), 1 deletion(-)

diff --git a/source/source_hsolver/kernels/cuda/diag_cusolvermp.cu b/source/source_hsolver/kernels/cuda/diag_cusolvermp.cu
index 032efd797b..003fa4f483 100644
--- a/source/source_hsolver/kernels/cuda/diag_cusolvermp.cu
+++ b/source/source_hsolver/kernels/cuda/diag_cusolvermp.cu
@@ -113,12 +113,13 @@ Diag_CusolverMP_gvd<inputT>::Diag_CusolverMP_gvd(const MPI_Comm mpi_comm,
     // This is because the current implementation of the cusolvermp library is ONE process ONE GPU.
     // So, when we use cusolvermp, we must ensure that the number of processes is equal to the number of GPUs.
     // In a sense, the MPI usage strategy of ABACUS must be subject to the cusolvermp.
+    // Use ROW_MAJOR to match BLACS grid initialization (order='R' in parallel_2d.cpp)
     CUSOLVER_CHECK(cusolverMpCreateDeviceGrid(cusolverMpHandle,
                                                    &this->grid,
                                                    this->cusolverCalComm,
                                                    this->nprows,
                                                    this->npcols,
-                                                   CUSOLVERMP_GRID_MAPPING_COL_MAJOR));
+                                                   CUSOLVERMP_GRID_MAPPING_ROW_MAJOR));
 
     // 20240529 zhanghaochong
     // Actually, there should be three matrix descriptors, A matrix, B matrix, and output eigenvector matrix.