diff --git a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
index ff86bdb8fa86f..6d9611ebd1393 100644
--- a/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
+++ b/paddle/fluid/framework/ir/conv_elementwise_add2_act_fuse_pass.cc
@@ -105,6 +105,22 @@ ConvElementwiseAdd2ActFusePass::ConvElementwiseAdd2ActFusePass() {
       .AddOutput("Out")
       .IsTensor()
       .End();
+
+  AddOpCompat(OpCompat("sigmoid"))
+      .AddInput("X")
+      .IsTensor()
+      .End()
+      .AddOutput("Out")
+      .IsTensor()
+      .End();
+
+  AddOpCompat(OpCompat("tanh"))
+      .AddInput("X")
+      .IsTensor()
+      .End()
+      .AddOutput("Out")
+      .IsTensor()
+      .End();
 }
 
 void ConvElementwiseAdd2ActFusePass::ApplyImpl(ir::Graph* graph) const {
@@ -188,4 +204,6 @@ REGISTER_PASS_CAPABILITY(conv_elementwise_add2_act_fuse_pass)
             .LE("conv2d", 1)
             .LE("elementwise_add", 1)
             .EQ("relu", 0)
+            .EQ("sigmoid", 0)
+            .EQ("tanh", 0)
             .EQ("identity", 0));
diff --git a/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc
index f67e83bc10171..47e2c5e380bcb 100644
--- a/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc
+++ b/paddle/fluid/framework/ir/conv_elementwise_add_act_fuse_pass.cc
@@ -102,6 +102,22 @@ ConvElementwiseAddActFusePass::ConvElementwiseAddActFusePass() {
       .AddOutput("Out")
       .IsTensor()
       .End();
+
+  AddOpCompat(OpCompat("sigmoid"))
+      .AddInput("X")
+      .IsTensor()
+      .End()
+      .AddOutput("Out")
+      .IsTensor()
+      .End();
+
+  AddOpCompat(OpCompat("tanh"))
+      .AddInput("X")
+      .IsTensor()
+      .End()
+      .AddOutput("Out")
+      .IsTensor()
+      .End();
 }
 
 void ConvElementwiseAddActFusePass::ApplyImpl(ir::Graph* graph) const {
@@ -170,4 +186,6 @@ REGISTER_PASS_CAPABILITY(conv_elementwise_add_act_fuse_pass)
             .LE("conv2d", 1)
             .LE("elementwise_add", 1)
             .EQ("relu", 0)
+            .EQ("sigmoid", 0)
+            .EQ("tanh", 0)
             .EQ("identity", 0));
diff --git a/paddle/fluid/framework/ir/graph_pattern_detector.cc b/paddle/fluid/framework/ir/graph_pattern_detector.cc
index 154df498e7d13..f0949cb9dfbd2 100644
--- a/paddle/fluid/framework/ir/graph_pattern_detector.cc
+++ b/paddle/fluid/framework/ir/graph_pattern_detector.cc
@@ -2324,7 +2324,8 @@ PDNode *patterns::PriorBox::operator()() {
   return boxes_var;
 }
 
-std::unordered_set<std::string> conv_act_set({"identity", "relu"});
+std::unordered_set<std::string> conv_act_set(
+    {"identity", "relu", "sigmoid", "tanh"});
 
 PDNode *patterns::ConvElementwiseaddAct::operator()(PDNode *conv_in) {
   conv_in->AsInput();
diff --git a/paddle/fluid/operators/fused/conv_fusion_op.cu b/paddle/fluid/operators/fused/conv_fusion_op.cu
index 5e96ca140274d..2ee63c9364221 100644
--- a/paddle/fluid/operators/fused/conv_fusion_op.cu
+++ b/paddle/fluid/operators/fused/conv_fusion_op.cu
@@ -544,9 +544,11 @@ class CUDNNConvFusionOpKernel : public framework::OpKernel<T> {
 
 namespace ops = paddle::operators;
 #if CUDNN_VERSION >= 7100
-REGISTER_OP_CUDA_KERNEL(conv2d_fusion,
-                        ops::CUDNNConvFusionOpKernel<float>,
-                        ops::CUDNNConvFusionOpKernel<double>);
+REGISTER_OP_CUDA_KERNEL(
+    conv2d_fusion,
+    ops::CUDNNConvFusionOpKernel<float>,
+    ops::CUDNNConvFusionOpKernel<double>,
+    ops::CUDNNConvFusionOpKernel<paddle::platform::float16>);
 #endif
 #ifdef PADDLE_WITH_HIP
 REGISTER_OP_CUDA_KERNEL(conv2d_fusion, ops::CUDNNConvFusionOpKernel<float>);