PaddlePaddle
diff --git a/‎paddle/fluid/framework/ir/CMakeLists.txt
Copy file name to clipboard
+1-2Lines changed: 1 addition & 2 deletions b/‎paddle/fluid/framework/ir/CMakeLists.txt
Copy file name to clipboard
+1-2Lines changed: 1 addition & 2 deletions
diff --git a/‎paddle/fluid/framework/ir/delete_c_identity_op_pass.cc
Copy file name to clipboard
-129Lines changed: 0 additions & 129 deletions b/‎paddle/fluid/framework/ir/delete_c_identity_op_pass.cc
Copy file name to clipboard
-129Lines changed: 0 additions & 129 deletions
diff --git a/‎paddle/fluid/framework/ir/delete_c_identity_op_pass.h
Copy file name to clipboard
-52Lines changed: 0 additions & 52 deletions b/‎paddle/fluid/framework/ir/delete_c_identity_op_pass.h
Copy file name to clipboard
-52Lines changed: 0 additions & 52 deletions
diff --git a/‎paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc renamed to ‎paddle/fluid/framework/ir/identity_op_clean_pass.cc
Copy file name to clipboard
+54-30Lines changed: 54 additions & 30 deletions b/‎paddle/fluid/framework/ir/identity_scale_op_clean_pass.cc renamed to ‎paddle/fluid/framework/ir/identity_op_clean_pass.cc
Copy file name to clipboard
+54-30Lines changed: 54 additions & 30 deletions
diff --git a/‎paddle/fluid/framework/ir/identity_scale_op_clean_pass.h renamed to ‎paddle/fluid/framework/ir/identity_op_clean_pass.h
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions b/‎paddle/fluid/framework/ir/identity_scale_op_clean_pass.h renamed to ‎paddle/fluid/framework/ir/identity_op_clean_pass.h
Copy file name to clipboard
+2-2Lines changed: 2 additions & 2 deletions
diff --git a/‎paddle/fluid/framework/ir/pass.cc
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion b/‎paddle/fluid/framework/ir/pass.cc
Copy file name to clipboard
+1-1Lines changed: 1 addition & 1 deletion
diff --git a/‎paddle/fluid/inference/api/paddle_pass_builder.cc
Copy file name to clipboard
+4-5Lines changed: 4 additions & 5 deletions b/‎paddle/fluid/inference/api/paddle_pass_builder.cc
Copy file name to clipboard
+4-5Lines changed: 4 additions & 5 deletions
@@ -89,7 +89,7 @@ pass_library(conv_elementwise_add2_act_fuse_pass inference)
 pass_library(conv_elementwise_add_fuse_pass inference)
 pass_library(transpose_flatten_concat_fuse_pass inference)
 pass_library(inplace_op_var_pass inference)
-pass_library(identity_scale_op_clean_pass base)
+pass_library(identity_op_clean_pass base)
 pass_library(sync_batch_norm_pass base)
 pass_library(runtime_context_cache_pass base)
 pass_library(quant_conv2d_dequant_fuse_pass inference)
@@ -102,7 +102,6 @@ pass_library(delete_weight_dequant_linear_op_pass inference)
 pass_library(delete_quant_dequant_linear_op_pass inference)
 pass_library(delete_dropout_op_pass inference)
 pass_library(delete_concat_op_pass inference)
-pass_library(delete_c_identity_op_pass inference)
 pass_library(preln_residual_bias_fuse_pass inference)
 pass_library(constant_folding_pass inference)
 pass_library(auto_mixed_precision_pass inference)
 
@@ -12,7 +12,7 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/framework/ir/identity_scale_op_clean_pass.h"
+#include "paddle/fluid/framework/ir/identity_op_clean_pass.h"
 
 #include "paddle/fluid/framework/ir/graph_pattern_detector.h"
 #include "paddle/fluid/framework/op_version_registry.h"
@@ -23,17 +23,16 @@ namespace ir {
 
 class Graph;
 
-void IdentityScaleOpCleanPass::ApplyImpl(ir::Graph* graph) const {
+void IdentityOpCleanPass::ApplyImpl(ir::Graph* graph) const {
   FusePassBase::Init("identity_scale_op_clean", graph);
 
-  // pre_op -> scale_in -> scale_op -> scale_out
+  // pre_op -> useless_op_in -> useless_op -> useless_op_out
   // ->
-  // pre_op -> scale_out
+  // pre_op -> useless_op_out
   GraphPatternDetector detector;
-  auto scale_in =
+  auto useless_op_in =
       detector.mutable_pattern()
-          ->NewNode("scale_in")
-          ->assert_is_op_input("scale")
+          ->NewNode("useless_op_in")
           ->assert_has_n_outputs(1)
           ->assert_var_not_persistable()
           ->assert_more([](Node* x) {
@@ -45,27 +44,51 @@ void IdentityScaleOpCleanPass::ApplyImpl(ir::Graph* graph) const {
             }
             return true;
           });
-  auto scale_op = detector.mutable_pattern()
-                      ->NewNode("scale_fuse")
-                      ->assert_is_op("scale")
-                      ->assert_op_attr<float>("scale", 1.)
-                      ->assert_op_attr<float>("bias", 0.);
-  auto scale_out = detector.mutable_pattern()
-                       ->NewNode("scale_out")
-                       ->assert_is_op_output("scale");
 
-  scale_op->LinksFrom({scale_in}).LinksTo({scale_out});
+  // This useless_op must have only one input and one output!
+  auto useless_op =
+      detector.mutable_pattern()
+          ->NewNode("useless_op")
+          ->assert_has_n_inputs(1)
+          ->assert_has_n_outputs(1)
+          ->assert_more([](Node* x) {
+            if (!x->IsOp()) {
+              return false;
+            }
+            if (x->Op()->Type() == "scale") {
+              auto scale = x->Op()->GetAttrIfExists<float>("scale");
+              auto bias = x->Op()->GetAttrIfExists<float>("bias");
+              if (std::abs(bias) <= 1e-6 && std::abs(scale - 1) <= 1e-6) {
+                return true;
+              }
+            }
+            if (x->Op()->Type() == "cast") {
+              auto in_dtype = x->Op()->GetAttrIfExists<int>("in_dtype");
+              auto out_dtype = x->Op()->GetAttrIfExists<int>("out_dtype");
+              if (in_dtype == out_dtype) {
+                return true;
+              }
+            }
+            if (x->Op()->Type() == "c_identity") {
+              return true;
+            }
+            // you can add more cases here.
+            return false;
+          });
+  auto useless_op_out = detector.mutable_pattern()->NewNode("useless_op_out");
+
+  useless_op->LinksFrom({useless_op_in}).LinksTo({useless_op_out});
 
   int found_subgraph_count = 0;
   GraphPatternDetector::handle_t handler =
       [&](const GraphPatternDetector::subgraph_t& subgraph, Graph* graph) {
-        Node* scale_op_var = subgraph.at(scale_op);
-        Node* scale_in_var = subgraph.at(scale_in);
-        Node* scale_out_var = subgraph.at(scale_out);
-        const std::string scale_in_name = scale_in_var->Name();
-        const std::string scale_out_name = scale_out_var->Name();
+        Node* useless_op_var = subgraph.at(useless_op);
+        Node* useless_op_in_var = subgraph.at(useless_op_in);
+        Node* useless_op_out_var = subgraph.at(useless_op_out);
+        const std::string useless_op_in_name = useless_op_in_var->Name();
+        const std::string useless_op_out_name = useless_op_out_var->Name();
         // Remove links in graph
-        GraphSafeRemoveNodes(graph, {scale_in_var, scale_op_var});
+        GraphSafeRemoveNodes(graph, {useless_op_in_var, useless_op_var});
         // Modify pre_op_desc
         // Link pre_op directly to scale_out
         for (auto& node : graph->Nodes()) {
@@ -76,16 +99,16 @@ void IdentityScaleOpCleanPass::ApplyImpl(ir::Graph* graph) const {
               auto names = out_var_map.second;
               bool reset = false;
               for (size_t i = 0; i < names.size(); i++) {
-                if (names[i] == scale_in_name) {
+                if (names[i] == useless_op_in_name) {
                   reset = true;
-                  names[i] = scale_out_name;
+                  names[i] = useless_op_out_name;
                   break;
                 }
               }
               if (reset) {
                 op_desc->SetOutput(out_var_map.first, names);
                 op_desc->Flush();
-                IR_NODE_LINK_TO(node, scale_out_var);
+                IR_NODE_LINK_TO(node, useless_op_out_var);
                 break;
               }
             }
@@ -102,9 +125,10 @@ void IdentityScaleOpCleanPass::ApplyImpl(ir::Graph* graph) const {
 }  // namespace framework
 }  // namespace paddle
 
-REGISTER_PASS(identity_scale_op_clean_pass,
-              paddle::framework::ir::IdentityScaleOpCleanPass);
-REGISTER_PASS_CAPABILITY(identity_scale_op_clean_pass)
+REGISTER_PASS(identity_op_clean_pass,
+              paddle::framework::ir::IdentityOpCleanPass);
+REGISTER_PASS_CAPABILITY(identity_op_clean_pass)
     .AddCombination(
-        paddle::framework::compatible::OpVersionComparatorCombination().EQ(
-            "scale", 0));
+        paddle::framework::compatible::OpVersionComparatorCombination()
+            .EQ("scale", 0)
+            .LE("c_identity", 1));
@@ -22,12 +22,12 @@ namespace ir {
 
 class Graph;
 
-class IdentityScaleOpCleanPass : public FusePassBase {
+class IdentityOpCleanPass : public FusePassBase {
  protected:
   void ApplyImpl(ir::Graph* graph) const override;
 
  private:
-  virtual ~IdentityScaleOpCleanPass() = default;
+  virtual ~IdentityOpCleanPass() = default;
 };
 
 }  // namespace ir
 
@@ -54,7 +54,7 @@ static const std::vector<std::string> support_subgraph_passes = {
 static const std::vector<std::string> xpu_support_subgraph_passes = {
     "delete_dropout_op_pass",
     "delete_concat_op_pass",
-    "identity_scale_op_clean_pass",
+    "identity_op_clean_pass",
     "delete_op_device_pass",
     "constant_folding_pass",
     "delete_elementwise_mul_op_pass",
 
@@ -94,12 +94,11 @@ const std::vector<std::string> kTRTSubgraphPasses({
       "delete_quant_dequant_filter_op_pass",          //
       "trt_delete_weight_dequant_linear_op_pass",     //
       "delete_quant_dequant_linear_op_pass",          //
-      "identity_scale_op_clean_pass",                 //
+      "identity_op_clean_pass",                       //
       "add_support_int8_pass",                        //
       "simplify_with_basic_ops_pass",                 //
       "trt_embedding_eltwise_layernorm_fuse_pass",    //
       "preln_embedding_eltwise_layernorm_fuse_pass",  //
-      "delete_c_identity_op_pass",                    //
       "trt_multihead_matmul_fuse_pass_v2",            //
       "trt_multihead_matmul_fuse_pass_v3",            //
       "multihead_matmul_roformer_fuse_pass",          //
@@ -175,7 +174,7 @@ const std::vector<std::string> kLiteSubgraphPasses({
 // running errors. After fusion operator supports low precision, delete this.
 const std::vector<std::string> kGpuLowerPrecisionPasses{
     "map_op_to_another_pass",
-    "identity_scale_op_clean_pass",
+    "identity_op_clean_pass",
     "simplify_with_basic_ops_pass",
     "silu_fuse_pass",
     "delete_quant_dequant_linear_op_pass",
@@ -222,7 +221,7 @@ const std::vector<std::string> kCINNCompilerPasses{
 GpuPassStrategy::GpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
     "map_op_to_another_pass",                                           //
-        "identity_scale_op_clean_pass",                                 //
+        "identity_op_clean_pass",                                       //
         "is_test_pass",                                                 //
         "simplify_with_basic_ops_pass",                                 //
         "delete_quant_dequant_linear_op_pass",                          //
@@ -511,7 +510,7 @@ XpuPassStrategy::XpuPassStrategy() : PassStrategy({}) {
   passes_.assign({
       "delete_dropout_op_pass",
       "delete_concat_op_pass",
-      "identity_scale_op_clean_pass",
+      "identity_op_clean_pass",
       "delete_repeated_ops_pass",
       "delete_op_device_pass",
       "constant_folding_pass",