diff --git a/paddle/phi/kernels/cpu/cholesky_kernel.cc b/paddle/phi/kernels/cpu/cholesky_kernel.cc
index aaa312995d2826..6580724975c42d 100644
--- a/paddle/phi/kernels/cpu/cholesky_kernel.cc
+++ b/paddle/phi/kernels/cpu/cholesky_kernel.cc
@@ -32,12 +32,12 @@ void CholeskyKernel(const Context& dev_ctx,
   using InputMatrixMap = Eigen::Map<const EigenMatrix>;
   using OutputMatrixMap = Eigen::Map<EigenMatrix>;
 
-  auto& dims = x.dims();
-  if (x.numel() == 0) {
-    out->Resize(dims);
+  if (out->numel() == 0) {
     dev_ctx.template Alloc<T>(out);
     return;
   }
+
+  auto& dims = x.dims();
   int batch_count = 1;
   for (int i = 0; i < dims.size() - 2; i++) {
     batch_count *= static_cast<int>(dims[i]);
diff --git a/paddle/phi/kernels/gpu/cholesky_kernel.cu b/paddle/phi/kernels/gpu/cholesky_kernel.cu
index 755adb96e21a44..8a831d2ee299a8 100644
--- a/paddle/phi/kernels/gpu/cholesky_kernel.cu
+++ b/paddle/phi/kernels/gpu/cholesky_kernel.cu
@@ -114,12 +114,12 @@ void CholeskyKernel(const Context& dev_ctx,
                     const DenseTensor& x,
                     bool upper,
                     DenseTensor* out) {
-  auto& dims = x.dims();
-  if (x.numel() == 0) {
-    out->Resize(dims);
+  if (out->numel() == 0) {
     dev_ctx.template Alloc<T>(out);
     return;
   }
+
+  auto& dims = x.dims();
   int batch_count = 1;
   for (int i = 0; i < dims.size() - 2; i++) {
     batch_count *= dims[i];
diff --git a/paddle/phi/kernels/impl/cholesky_grad_kernel_impl.h b/paddle/phi/kernels/impl/cholesky_grad_kernel_impl.h
index 8a2e1484302f0a..5d146dae8d53e2 100644
--- a/paddle/phi/kernels/impl/cholesky_grad_kernel_impl.h
+++ b/paddle/phi/kernels/impl/cholesky_grad_kernel_impl.h
@@ -242,14 +242,13 @@ void CholeskyGradKernel(const Context& dev_ctx,
                         const DenseTensor& out_grad,
                         bool upper,
                         DenseTensor* x_grad) {
-  auto* x_grad_data = dev_ctx.template Alloc<T>(x_grad);
-
-  auto& dims = out.dims();
-  if (out.numel() == 0) {
-    x_grad->Resize(dims);
+  if (x_grad->numel() == 0) {
     dev_ctx.template Alloc<T>(x_grad);
     return;
   }
+
+  auto* x_grad_data = dev_ctx.template Alloc<T>(x_grad);
+  auto& dims = out.dims();
   int batch_count = 1;
   for (int i = 0; i < dims.size() - 2; i++) {
     batch_count *= dims[i];