Merge pull request #16981 from luotao1/disable_runtime_context_default

disable runtime_context_cache pass by default

paddle/fluid/inference/api/paddle_pass_builder.cc

@@ -87,7 +87,6 @@ const std::vector<std::string> kAnakinSubgraphPasses({
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
     "infer_clean_graph_pass",  //
-        "runtime_context_cache_pass",  //
         //   "identity_scale_op_clean_pass",              //
         "conv_affine_channel_fuse_pass",             //
         "conv_eltwiseadd_affine_channel_fuse_pass",  //
@@ -118,10 +117,6 @@ CpuPassStrategy::CpuPassStrategy() : PassStrategy({}) {
   // not be damaged by smaller ones.
   passes_.assign({
       "infer_clean_graph_pass",         //
-      // TODO(luotao): runtime_context_cache_pass should be located in the
-      // front, see https://github.com/PaddlePaddle/Paddle/issues/16609,
-      // will enhance this pass later.
-      "runtime_context_cache_pass",     //
       "attention_lstm_fuse_pass",       //
       "seqconv_eltadd_relu_fuse_pass",  //
       // "seqpool_concat_fuse_pass",    //

paddle/fluid/inference/tests/api/analyzer_pyramid_dnn_tester.cc

@@ -110,6 +110,11 @@ void SetConfig(AnalysisConfig *cfg) {
   if (FLAGS_zero_copy) {
     cfg->SwitchUseFeedFetchOps(false);
   }
+  // Enable runtime_context_cache_pass, disabled by default since it doesn't
+  // cover all the cases.
+  // See detail: https://github.com/PaddlePaddle/Paddle/issues/16609
+  // https://github.com/PaddlePaddle/Paddle/issues/16841
+  cfg->pass_builder()->AppendPass("runtime_context_cache_pass");
 }
 
 void SetInput(std::vector<std::vector<PaddleTensor>> *inputs) {