diff --git a/source/source_base/module_container/ATen/kernels/cuda/lapack.cu b/source/source_base/module_container/ATen/kernels/cuda/lapack.cu
index 6300df953f..32188c3e8b 100644
--- a/source/source_base/module_container/ATen/kernels/cuda/lapack.cu
+++ b/source/source_base/module_container/ATen/kernels/cuda/lapack.cu
@@ -70,8 +70,10 @@ struct lapack_trtri<T, DEVICE_GPU> {
     {
         // TODO: trtri is not implemented in this method yet
         // Cause the trtri in cuSolver is not stable for ABACUS!
-        //cuSolverConnector::trtri(cusolver_handle, uplo, diag, dim, Mat, lda);
-        cuSolverConnector::potri(cusolver_handle, uplo, diag, dim, Mat, lda);
+        // But why?! trtri and potri are different routines for different job! 
+        // How can BPCG work without using a proper routine? 
+        cuSolverConnector::trtri(cusolver_handle, uplo, diag, dim, Mat, lda);
+        // cuSolverConnector::potri(cusolver_handle, uplo, diag, dim, Mat, lda);
     }
 };
 
@@ -201,4 +203,4 @@ template struct lapack_getrs<std::complex<float>,  DEVICE_GPU>;
 template struct lapack_getrs<std::complex<double>, DEVICE_GPU>;
 
 } // namespace kernels
-} // namespace container
\ No newline at end of file
+} // namespace container
diff --git a/tests/16_SDFT_GPU/187_PW_SDFT_MALL_BPCG_GPU/result.ref b/tests/16_SDFT_GPU/187_PW_SDFT_MALL_BPCG_GPU/result.ref
index 077de5b358..cd6c5029d4 100644
--- a/tests/16_SDFT_GPU/187_PW_SDFT_MALL_BPCG_GPU/result.ref
+++ b/tests/16_SDFT_GPU/187_PW_SDFT_MALL_BPCG_GPU/result.ref
@@ -1,5 +1,5 @@
-etotref -96.9361190965003630
-etotperatomref -48.4680595483
-totalforceref 248.979444
-totalstressref 230453.604050
+etotref -96.9361125708192191
+etotperatomref -48.4680562854
+totalforceref 248.979476
+totalstressref 230453.662470
 totaltimeref 6.44