diff --git a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
index a9d91b6e854898a5e037078aa6165beaea37f541..cd88139af23a75f1864e3515dd996cf85915dc87 100644
--- a/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
+++ b/paddle/fluid/prim/api/composite_backward/composite_backward_api.h
@@ -60,9 +60,12 @@ void leaky_relu_grad(const Tensor& out,
 }
 
 template <typename T>
-void silu_grad(const Tensor& x, const Tensor& out_grad, Tensor* x_grad) {
+void silu_grad(const Tensor& x,
+               const Tensor& out,
+               const Tensor& out_grad,
+               Tensor* x_grad) {
   if (x_grad) {
-    auto sigmoid = 1.0 / (1.0 + exp<T>(-x));
+    auto sigmoid = out / x;
     auto res = out_grad * sigmoid * (1.0 + x * (1.0 - sigmoid));
     set_output<T>(res, x_grad);
   }
diff --git a/paddle/phi/api/yaml/backward.yaml b/paddle/phi/api/yaml/backward.yaml
index 6ae8c19006410e37a508a547ded5a5e8eb360135..27bdf904eee6e1df7ca0654232750592d370568e 100644
--- a/paddle/phi/api/yaml/backward.yaml
+++ b/paddle/phi/api/yaml/backward.yaml
@@ -1599,14 +1599,14 @@
 
 - backward_op : silu_grad
   forward : silu (Tensor x) -> Tensor(out)
-  args : (Tensor x, Tensor out_grad)
+  args : (Tensor x, Tensor out, Tensor out_grad)
   output : Tensor(x_grad)
   infer_meta :
     func : UnchangedInferMeta
     param : [x]
   kernel :
     func : silu_grad
-  composite : silu_grad(x, out_grad, x_grad)
+  composite : silu_grad(x, out, out_grad, x_grad)
   inplace : (out_grad -> x_grad)
 
 - backward_op : sin_double_grad
diff --git a/paddle/phi/kernels/activation_grad_kernel.h b/paddle/phi/kernels/activation_grad_kernel.h
index ca75a6e0b24a48d99fcc057735dfee333b7eeddd..302e0f6ba9080afe6e196e060af7f3d306ffc0a7 100644
--- a/paddle/phi/kernels/activation_grad_kernel.h
+++ b/paddle/phi/kernels/activation_grad_kernel.h
@@ -75,6 +75,12 @@ namespace phi {
                         DenseTensor* dx);
 
 template <typename T, typename Context>
+void SiluGradKernel(const Context& dev_ctx,
+                    const DenseTensor& x,
+                    const DenseTensor& out,
+                    const DenseTensor& dout,
+                    DenseTensor* dx);
+template <typename T, typename Context>
 void ReluDoubleGradKernel(const Context& dev_ctx,
                           const DenseTensor& out,
                           const DenseTensor& ddx,
@@ -277,7 +283,6 @@ DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(Asinh);
 DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(Acosh);
 DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(Atanh);
 DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(TanhShrink);
-DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(Silu);
 DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(Square);
 DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(Softsign);
 DECLARE_ACTIVATION_GRAD_KERNEL_DEPX(LogSigmoid);
diff --git a/paddle/phi/kernels/cpu/activation_grad_kernel.cc b/paddle/phi/kernels/cpu/activation_grad_kernel.cc
index 9273f8393b5b35923001d6fe1f239e8f95ed85eb..7b9074ffa92f341d49b2aac0a44849f84e958a6b 100644
--- a/paddle/phi/kernels/cpu/activation_grad_kernel.cc
+++ b/paddle/phi/kernels/cpu/activation_grad_kernel.cc
@@ -128,7 +128,6 @@ DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPX(Asinh, AsinhGradFunctor);
 DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPX(Acosh, AcoshGradFunctor);
 DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPX(Atanh, AtanhGradFunctor);
 DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPX(TanhShrink, TanhShrinkGradFunctor);
-DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPX(Silu, SiluGradFunctor);
 DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPX(Square, SquareGradFunctor);
 
 DEFINE_CPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Exp, ExpGradFunctor);
@@ -190,6 +189,17 @@ DEFINE_CPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT(HardSigmoid,
                                                  slope,
                                                  offset);
 
+template <typename T, typename Context>
+void SiluGradKernel(const Context& dev_ctx,
+                    const DenseTensor& x,
+                    const DenseTensor& out,
+                    const DenseTensor& dout,
+                    DenseTensor* dx) {
+  funcs::SiluGradFunctor<T> functor;
+  ActivationGradImpl<T, Context, funcs::SiluGradFunctor<T>>(
+      dev_ctx, &x, &out, &dout, dx, functor);
+}
+
 template <typename T, typename Context>
 void EluGradKernel(const Context& dev_ctx,
                    const DenseTensor& x,
diff --git a/paddle/phi/kernels/gpu/activation_grad_kernel.cu b/paddle/phi/kernels/gpu/activation_grad_kernel.cu
index c0fb7342a80b65bfae189fe91135362d661f552a..aa703ede3bad6600cc0a54354eed6f225e6e65c6 100644
--- a/paddle/phi/kernels/gpu/activation_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/activation_grad_kernel.cu
@@ -190,7 +190,6 @@ DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Asinh, CudaAsinhGradFunctor);
 DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Acosh, CudaAcoshGradFunctor);
 DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Atanh, CudaAtanhGradFunctor);
 DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(TanhShrink, CudaTanhShrinkGradFunctor);
-DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Silu, CudaSiluGradFunctor);
 DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPX(Square, CudaSquareGradFunctor);
 
 DEFINE_GPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Exp, CudaExpGradFunctor);
@@ -249,6 +248,16 @@ DEFINE_GPU_ACT_GRAD_KERNEL_WITH_TWO_ATTRS_DEPOUT(HardSigmoid,
                                                  slope,
                                                  offset);
 
+template <typename T, typename Context>
+void SiluGradKernel(const Context& dev_ctx,
+                    const DenseTensor& x,
+                    const DenseTensor& out,
+                    const DenseTensor& dout,
+                    DenseTensor* dx) {
+  funcs::CudaSiluGradFunctor<T> functor;
+  ActivationGradGPUImpl<T, Context, funcs::CudaSiluGradFunctor<T>>(
+      dev_ctx, &x, &out, &dout, dx, functor);
+}
 template <typename T, typename Context>
 void EluGradKernel(const Context& dev_ctx,
                    const DenseTensor& x,
diff --git a/paddle/phi/kernels/xpu/activation_grad_kernel.cc b/paddle/phi/kernels/xpu/activation_grad_kernel.cc
index 02243215f95886dbf5f0d4dbf2e1c85ab4a74d63..8aae57b1e498ba5f41240570773dcf3d472c5feb 100644
--- a/paddle/phi/kernels/xpu/activation_grad_kernel.cc
+++ b/paddle/phi/kernels/xpu/activation_grad_kernel.cc
@@ -567,7 +567,6 @@ DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Tanh, XPUTanhGradFunctor);
 DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Relu, XPUReluGradFunctor);
 DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPOUT(Relu6, XPURelu6GradFunctor);
 
-DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(Silu, XPUSiluGradFunctor);
 DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(Log, XPULogGradFunctor);
 DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(Square, XPUSquareGradFunctor);
 DEFINE_XPU_ACTIVATION_GRAD_KERNEL_DEPX(Swish, XPUSwishGradFunctor);
@@ -605,6 +604,16 @@ void HardSwishGradKernel(const Context& dev_ctx,
       dev_ctx, &x, nullptr, &dout, dx, functor);
 }
 
+template <typename T, typename Context>
+void SiluGradKernel(const Context& dev_ctx,
+                    const DenseTensor& x,
+                    const DenseTensor& out,
+                    const DenseTensor& dout,
+                    DenseTensor* dx) {
+  XPUSiluGradFunctor<T> functor;
+  ActivationGradXPUImpl<T, Context, XPUSiluGradFunctor<T>>(
+      dev_ctx, &x, &out, &dout, dx, functor);
+}
 }  // namespace phi
 
 PD_REGISTER_KERNEL(relu_grad,