[NewIR] fix bug: feed_with_place should consider variable existence (#55756)

* fix bug: feed_with_place should consider variable existence

* fix

* fix build scope

* change method to set feed var name

* fix

paddle/fluid/framework/executor_cache.cc

@@ -368,11 +368,15 @@ std::unique_ptr<::ir::Program> ConstructFowardIrProgram(
   }
 
   // add fetch with place op to program
+  auto *block = local_program.MutableBlock(0);
   for (auto &in_t : x) {
     auto name = in_t.name();
+    if (block->FindVarRecursive(name) == nullptr) {
+      continue;
+    }
     auto place = in_t.place().GetType();
 
-    auto op_desc = local_program.MutableBlock(0)->PrependOp();
+    auto op_desc = block->PrependOp();
     op_desc->SetType("feed_with_place");
     op_desc->SetAttr("index", 0);
     // TODO(phlrain) : using tensor dtype
@@ -429,13 +433,17 @@ std::unique_ptr<::ir::Program> ConstructBackwardIrProgram(
   auto local_program =
       paddle::framework::ProgramDesc(*(backward_global_block->Program()));
   // add feed kernel
+  auto *block = local_program.MutableBlock(0);
   for (auto &out_grad_t : out_grad) {
     auto name = out_grad_t.name();
+    if (block->FindVarRecursive(name) == nullptr) {
+      continue;
+    }
     auto place = out_grad_t.place().GetType();
     if (name == "@EMPTY@") {
       continue;
     }
-    auto op_desc = local_program.MutableBlock(0)->PrependOp();
+    auto op_desc = block->PrependOp();
     op_desc->SetType("feed_with_place");
     op_desc->SetAttr("index", 0);
     // TODO(phlrain) : using tensor dtype

paddle/fluid/framework/feed_fetch_method.cc

@@ -38,10 +38,9 @@ void SetFeedVariable(Scope* scope,
   VLOG(3) << "SetFeedVariable name=" << var_name << " index=" << index;
   if (FLAGS_enable_new_ir_in_executor) {
     // shared data with input tensor
-    auto inner_var_name = var_name + "_" + std::to_string(index);
-    auto feed_ele = scope->Var(inner_var_name);
+    auto feed_ele = scope->Var(var_name);
     if (!feed_ele->IsType<phi::DenseTensor>()) {
-      VLOG(3) << "Reset " << inner_var_name << " to phi::DenseTensor";
+      VLOG(3) << "Reset " << var_name << " to phi::DenseTensor";
       feed_ele->Clear();
     }
     auto val = feed_ele->GetMutable<phi::DenseTensor>();

paddle/fluid/framework/new_executor/new_ir_interpreter.cc

@@ -1628,9 +1628,18 @@ std::string NewIRInterpreter::DebugValueInfo() {
      << "value -> var_name -> id -> variable*"
      << "\n";
   for (auto kv : value_2_var_name_) {
+    PADDLE_ENFORCE((bool)kv.first,
+                   platform::errors::PreconditionNotMet(
+                       "vlaue(%s) should not be nullptr", kv.second));
+    PADDLE_ENFORCE(var_name_2_id_.count(kv.second) > 0,
+                   platform::errors::PreconditionNotMet(
+                       "var(%s) should exist in var_name_2_id_", kv.second));
+    auto* var = InnerScope()->FindVar(kv.second);
+    PADDLE_ENFORCE(var != nullptr,
+                   platform::errors::PreconditionNotMet(
+                       "var(%s) should exist in var_name_2_id_", kv.second));
     os << kv.first.impl() << " -> " << kv.second << " -> "
-       << var_name_2_id_.at(kv.second) << " -> "
-       << InnerScope()->FindVar(kv.second) << "\n";
+       << var_name_2_id_.at(kv.second) << " -> " << var << "\n";
   }
   return os.str();
 }

paddle/fluid/ir/phi_kernel_adaptor/phi_kernel_util.cc

@@ -200,17 +200,17 @@ void HandleForSpecialOp(
     auto value = op->result(0);
     VLOG(6) << "link feed output to feed in variable" << inner_scope;
 
-    int index =
-        op->attributes().at("col").dyn_cast<ir::Int32Attribute>().data();
     std::string name =
         op->attributes().at("name").dyn_cast<ir::StrAttribute>().AsString();
     paddle::framework::Variable* var = inner_scope->FindVar(name);
+    PADDLE_ENFORCE(var,
+                   paddle::platform::errors::InvalidArgument(
+                       "The variable %s shoud exist", name));
 
-    auto feed_var_name = "feed_" + std::to_string(index);
-    value_2_var_name->emplace(value, feed_var_name);
-    variable_2_var_name->emplace(var, feed_var_name);
+    value_2_var_name->emplace(value, name);
+    variable_2_var_name->emplace(var, name);
     auto id = var_name_2_id->size();
-    var_name_2_id->emplace(feed_var_name, id);
+    var_name_2_id->emplace(name, id);
     variable_list->push_back(var);
     PADDLE_ENFORCE_EQ(
         variable_list->size(),
@@ -226,6 +226,21 @@ void HandleForSpecialOp(
 
     auto value = op->result(0);
     value_2_var_name->emplace(value, var_name);
+
+    paddle::framework::Variable* var = inner_scope->FindVar(var_name);
+    PADDLE_ENFORCE(var,
+                   paddle::platform::errors::InvalidArgument(
+                       "The variable %s shoud exist", var_name));
+
+    variable_2_var_name->emplace(var, var_name);
+    auto id = var_name_2_id->size();
+    var_name_2_id->emplace(var_name, id);
+    variable_list->push_back(var);
+    PADDLE_ENFORCE_EQ(
+        variable_list->size(),
+        var_name_2_id->size(),
+        paddle::platform::errors::InvalidArgument(
+            "The size of variable_list and var_name_2_id map should be equal"));
   }
 
   if (op_name == "builtin.combine") {

paddle/fluid/ir_adaptor/translator/op_translator.cc

@@ -475,6 +475,9 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
     std::string legacy_output_name =
         op_normalizer.GetLegacyArgName(op_desc.Type(), info.name);
 
+    VLOG(10) << "[op:" << op_desc.Type() << "][output]" << info.name << " "
+             << legacy_output_name;
+
     // return empty type if this arg is optional and not shown in OpDesc
     if (!op_desc.HasOutput(legacy_output_name)) {
       VLOG(10) << "[output translating]"
@@ -491,11 +494,19 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
     const auto& legacy_output_vars = op_desc.Output(legacy_output_name);
     bool is_vector = (info.type_name.find("VectorType") != std::string::npos);
 
+    VLOG(10) << "[op:" << op_desc.Type() << "][output]" << info.name << " "
+             << legacy_output_name << " " << legacy_output_vars.size() << " "
+             << is_vector;
+
     // Specially process TensorArray, this because we cannot distinguish it with
     // Vector<DenseTensor> by other conditions but we cannot support it like
     // Vector<DenseTensor>
     if (legacy_output_vars.size() == 1) {
       VarDesc* var = block->FindVarRecursive(legacy_output_vars[0]);
+      IR_ENFORCE(var != nullptr,
+                 "[op:%s] Output %s should not be null",
+                 op_desc.Type(),
+                 legacy_output_vars[0]);
       if (var->GetType() ==
           paddle::framework::proto::VarType::LOD_TENSOR_ARRAY) {
         ir::Type translated_var_type =
@@ -519,6 +530,10 @@ OpTranscriber::GenerateOperationOutput(ir::IrContext* ctx,
 
       auto& var_name = legacy_output_vars[0];
       VarDesc* var = block->FindVarRecursive(var_name);
+      IR_ENFORCE(var != nullptr,
+                 "[op:%s] Output %s should not be null",
+                 op_desc.Type(),
+                 var_name);
       VLOG(10) << "[output translating]"
                << "[" << op_desc.Type() << "]" << info.name << " " << var_name
                << " " << var->GetType();

python/paddle/fluid/executor.py

@@ -26,7 +26,7 @@ from .framework import convert_np_dtype_to_dtype_, _apply_pass
 from . import core
 from . import unique_name
 from . import compiler
-from . import set_flags
+from . import set_flags, get_flags
 from .trainer_factory import TrainerFactory
 from .trainer_factory import FetchHandlerMonitor
 import copy
@@ -1071,7 +1071,13 @@ class Executor:
                         )
                     check_feed_shape_type(var, cur_feed)
                 idx = op.desc.attr('col')
-                core.set_feed_variable(scope, cur_feed, feed_var_name, idx)
+                new_ir_flag_name = 'FLAGS_enable_new_ir_in_executor'
+                if get_flags(new_ir_flag_name)[new_ir_flag_name]:
+                    core.set_feed_variable(
+                        scope, cur_feed, feed_target_name, idx
+                    )
+                else:
+                    core.set_feed_variable(scope, cur_feed, feed_var_name, idx)
             else:
                 break