diff --git a/paddle/fluid/platform/device/xpu/xpu2_op_list.h b/paddle/fluid/platform/device/xpu/xpu2_op_list.h
index f8b15d4d4ee280ce8fb561b302a091496502207b..b1838a0f714c86429f73b60d8ee5a99bb30664f0 100644
--- a/paddle/fluid/platform/device/xpu/xpu2_op_list.h
+++ b/paddle/fluid/platform/device/xpu/xpu2_op_list.h
@@ -568,7 +568,9 @@ XPUOpMap& get_kl2_ops() {
       {"sqrt", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"sqrt_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
       {"square_grad", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
-      {"square", XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace())})},
+      {"square",
+       XPUKernelSet({pOpKernelType(vartype::FP32, XPUPlace()),
+                     pOpKernelType(vartype::FP16, XPUPlace())})},
       {"squeeze2_grad",
        XPUKernelSet({pOpKernelType(vartype::FP64, XPUPlace()),
                      pOpKernelType(vartype::INT64, XPUPlace()),
diff --git a/paddle/phi/kernels/xpu/activation_kernel.cc b/paddle/phi/kernels/xpu/activation_kernel.cc
index 51f74bd34750a91a14d089212ff7c4ac5a7c705f..2425f304a3a98c150f0dbed90430d5e0918c28db 100644
--- a/paddle/phi/kernels/xpu/activation_kernel.cc
+++ b/paddle/phi/kernels/xpu/activation_kernel.cc
@@ -479,6 +479,9 @@ PD_REGISTER_KERNEL(
 PD_REGISTER_KERNEL(
     tanh, XPU, ALL_LAYOUT, phi::TanhKernel, float, phi::dtype::float16) {}
 
+PD_REGISTER_KERNEL(
+    square, XPU, ALL_LAYOUT, phi::SquareKernel, float, phi::dtype::float16) {}
+
 PD_REGISTER_ACTIVATION_KERNEL(exp, ExpKernel)  // no grad
 PD_REGISTER_ACTIVATION_KERNEL(log, LogKernel)
 PD_REGISTER_ACTIVATION_KERNEL(leaky_relu, LeakyReluKernel)
@@ -492,4 +495,3 @@ PD_REGISTER_ACTIVATION_KERNEL(sigmoid, SigmoidKernel)
 PD_REGISTER_ACTIVATION_KERNEL(sqrt, SqrtKernel)
 PD_REGISTER_ACTIVATION_KERNEL(swish_raw, SwishRawKernel)
 PD_REGISTER_ACTIVATION_KERNEL(softplus, SoftplusKernel)
-PD_REGISTER_ACTIVATION_KERNEL(square, SquareKernel)