Fix conv2d+dequantize squash for residual fusion (#19545)

* Fix conv2d+dequantize squash for residual fusion

test=develop

* Change condition

test=develop

paddle/fluid/framework/ir/mkldnn/cpu_quantize_squash_pass.cc

@@ -177,7 +177,10 @@ void CPUQuantizeSquashPass::ConvDequantSquash(Graph* graph) const {
     GET_IR_NODE_FROM_SUBGRAPH(dequant_out, dequant_out, conv_dequant_pattern);
 
     // if conv2d has one output
-    if (conv_out->outputs.size() == 1) {
+    // and there is no fuse residual connection
+    // because residual fusion does not support force output with fp32
+    if (conv_out->outputs.size() == 1 &&
+        !(conv_op->Op()->GetAttrIfExists<bool>("fuse_residual_connection"))) {
       conv_op->Op()->SetAttr("force_fp32_output", true);
       conv_op->Op()->SetOutput("Output",
                                std::vector<std::string>({dequant_out->Name()}));