Remove RunFromCinn in PE because We Will Call CinnRunner in Compute of SubgraphOp (#36385)

Remove RunFromCinn method in PE because We Will Call CinnRunner in Compute method of SubgraphOp

paddle/fluid/framework/parallel_executor.cc

@@ -34,7 +34,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/ir/memory_optimize_pass/memory_optimization_var_info.h"
 #include "paddle/fluid/framework/ir/memory_optimize_pass/reference_count_pass_helper.h"
 #include "paddle/fluid/framework/ir/multi_devices_graph_pass/set_reader_device_info_utils.h"
-#include "paddle/fluid/framework/paddle2cinn/cinn_runner.h"
 #include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/platform/cuda_graph_with_memory_pool.h"
 #include "paddle/fluid/platform/event.h"
@@ -44,7 +43,6 @@ limitations under the License. */
 #include "paddle/fluid/platform/cuda_device_guard.h"
 #endif
 
-DECLARE_bool(use_cinn);
 DECLARE_double(eager_delete_tensor_gb);
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
@@ -945,40 +943,6 @@ void ParallelExecutor::RunWithoutFetch(
   member_->executor_->Run(/*fetch_tensors*/ {}, /*return_merged*/ false);
 }
 
-FetchResultType ParallelExecutor::RunFromCinn(
-    const std::unordered_map<std::string, LoDTensor> &feed_tensors,
-    const std::vector<std::string> &fetch_names) {
-  // Feed tensor to scope, now only support 1 scope
-  // TODO(zhhsplendid): handle multiple scope
-  size_t scope_id = 0;
-  std::map<std::string, const LoDTensor *> cinn_input_tensors;
-  for (auto &name_tensor_pair : feed_tensors) {
-    bool is_persistable = member_->IsPersistable(name_tensor_pair.first);
-    if (!is_persistable) {
-      member_->SetSkipMemoryReuse(scope_id, name_tensor_pair.first);
-    }
-    Scope *feed_scope = is_persistable ? member_->local_scopes_[scope_id]
-                                       : member_->local_exec_scopes_[scope_id];
-    Variable *feed_var = feed_scope->Var(name_tensor_pair.first);
-    LoDTensor *trg = feed_var->GetMutable<LoDTensor>();
-    trg->ShareDataWith(name_tensor_pair.second);
-    trg->set_lod(name_tensor_pair.second.lod());
-
-    cinn_input_tensors[name_tensor_pair.first] = trg;
-  }
-
-  // TODO(zhhsplendid): get correct API after CINN API is ready
-  // now only return empty fetch result;
-  std::shared_ptr<paddle2cinn::CinnRunner> cinn_runner =
-      paddle2cinn::CinnRunner::GetInstance();
-
-  cinn_runner->Run(Graph(), member_->local_exec_scopes_[scope_id],
-                   &cinn_input_tensors);
-
-  paddle::framework::FetchResultType fetches = FetchList(fetch_names.size());
-  return fetches;
-}
-
 void ParallelExecutor::SkipMemoryReuse(
     size_t scope_idx, const std::vector<std::string> &skip_vars) {
   for (auto &var_name : skip_vars) {

paddle/fluid/framework/parallel_executor.h

@@ -14,7 +14,6 @@ limitations under the License. */
 
 #pragma once
 
-#include <map>
 #include <memory>
 #include <string>
 #include <unordered_map>
@@ -93,10 +92,6 @@ class ParallelExecutor {
 
   void RunWithoutFetch(const std::vector<std::string> &skip_eager_vars);
 
-  FetchResultType RunFromCinn(
-      const std::unordered_map<std::string, LoDTensor> &feed_tensors,
-      const std::vector<std::string> &fetch_names);
-
   void ResetOpHandleScopeMapOfGraphs(
       const std::unordered_map<Scope *, Scope *> &scope_map);
 

paddle/fluid/pybind/pybind.cc

@@ -3293,18 +3293,6 @@ All parameter, weight, gradient are variables in Paddle.
                    BOOST_GET(paddle::framework::FetchUnmergedList, ret)));
              }
            })
-      .def("run_from_cinn",
-           [](ParallelExecutor &self,
-              const std::unordered_map<std::string, LoDTensor> &feed_tensors,
-              const std::vector<std::string> &fetch_names) -> py::object {
-             paddle::framework::FetchResultType ret;
-             {
-               pybind11::gil_scoped_release release;
-               ret = self.RunFromCinn(feed_tensors, fetch_names);
-             }
-             return py::cast(
-                 std::move(BOOST_GET(paddle::framework::FetchList, ret)));
-           })
       .def("device_count", &ParallelExecutor::DeviceCount);
 
   BindFleetWrapper(&m);

python/paddle/fluid/executor.py

@@ -24,7 +24,7 @@ from .wrapped_decorator import signature_safe_contextmanager
 import six
 from .data_feeder import convert_dtype
 from .framework import Program, default_main_program, Variable, Operator
-from .framework import convert_np_dtype_to_dtype_, get_flags
+from .framework import convert_np_dtype_to_dtype_
 from . import core
 from . import unique_name
 from . import compiler
@@ -1016,17 +1016,8 @@ class Executor(object):
                 if need_check_feed:
                     check_feed_shape_type(var, feed_tensor, exe.device_count())
                 feed_tensor_dict[feed_name] = feed_tensor
-
-            #TODO(zhhsplendid): handle other feed data format case for CINN
-            use_cinn = get_flags("FLAGS_use_cinn")["FLAGS_use_cinn"]
-            if use_cinn:
-                fetch_var_names = list(map(_to_name_str, fetch_list))
-                fetch_tensors = exe.run_from_cinn(
-                    feed_tensor_dict, fetch_var_names)._move_to_list()
-                return as_numpy(
-                    fetch_tensors) if return_numpy else fetch_tensors
-            else:
             exe.feed_and_split_tensor_into_local_scopes(feed_tensor_dict)
+
         elif isinstance(feed, list) or isinstance(feed, tuple):
             res = list()
             for i, each in enumerate(feed):
@@ -1047,7 +1038,6 @@ class Executor(object):
                     res_dict[feed_name] = tensor
                 res.append(res_dict)
 
-            use_cinn = get_flags("FLAGS_use_cinn")["FLAGS_use_cinn"]
             exe.feed_tensors_into_local_scopes(res)
 
         if hasattr(program._program, 'lr_sheduler'):