diff --git a/paddle/fluid/distributed/store/tcp_store.cc b/paddle/fluid/distributed/store/tcp_store.cc
index 25b66718e4a92ea33693375164d29f34af7816a7..0ecfcef42458d18a482860579111031f9b3ab787 100644
--- a/paddle/fluid/distributed/store/tcp_store.cc
+++ b/paddle/fluid/distributed/store/tcp_store.cc
@@ -20,7 +20,7 @@
 
 #include "paddle/fluid/distributed/store/tcp_utils.h"
 #include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 namespace paddle {
 namespace distributed {
diff --git a/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cpu.cc b/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cpu.cc
index 273d0555288888f53e12932d940db7d7f061f40e..0474f253dece30a5c77490b162278d5a05407f35 100644
--- a/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cpu.cc
+++ b/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cpu.cc
@@ -25,7 +25,7 @@
 #include "paddle/fluid/eager/tests/performance_tests/benchmark_utils.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
 #include "paddle/fluid/imperative/tracer.h"
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 #ifdef WITH_GPERFTOOLS
 #include "gperftools/profiler.h"
diff --git a/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cuda.cc b/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cuda.cc
index 93260a315b32e30d158d09aec43d67845097cbfa..72dd1cc8ff56a74fd6e893ab704c8af6b6aca367 100644
--- a/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cuda.cc
+++ b/paddle/fluid/eager/tests/performance_tests/benchmark_eager_cuda.cc
@@ -24,7 +24,7 @@
 #include "paddle/fluid/eager/tests/performance_tests/benchmark_utils.h"
 #include "paddle/fluid/eager/tests/test_utils.h"
 #include "paddle/fluid/imperative/tracer.h"
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 #ifdef WITH_GPERFTOOLS
 #include "gperftools/profiler.h"
diff --git a/paddle/fluid/framework/new_executor/executor_statistics.cc b/paddle/fluid/framework/new_executor/executor_statistics.cc
index c1ba3b193f1deac0e3060c15be9b537e4d505872..f773a386af47358493aba88b62d8359a50e91a82 100644
--- a/paddle/fluid/framework/new_executor/executor_statistics.cc
+++ b/paddle/fluid/framework/new_executor/executor_statistics.cc
@@ -24,9 +24,9 @@
 #include <vector>
 
 #include "glog/logging.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/fluid/platform/os_info.h"
 #include "paddle/fluid/platform/profiler/utils.h"
+#include "paddle/phi/core/flags.h"
 
 DECLARE_bool(use_stream_safe_cuda_allocator);
 PADDLE_DEFINE_EXPORTED_string(static_executor_perfstat_filepath,
diff --git a/paddle/fluid/imperative/flags.cc b/paddle/fluid/imperative/flags.cc
index f66aacc89ec506e4054eef9864f938c930b28cfe..06c2719ebfa6399511817792e7a808349ffb9641 100644
--- a/paddle/fluid/imperative/flags.cc
+++ b/paddle/fluid/imperative/flags.cc
@@ -14,7 +14,7 @@
 
 #include "paddle/fluid/imperative/flags.h"
 
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 PADDLE_DEFINE_EXPORTED_uint64(dygraph_debug,
                               0,
diff --git a/paddle/fluid/imperative/profiler.cc b/paddle/fluid/imperative/profiler.cc
index 3365d5239da96b1f351d3991ad7680eb1defc5d3..fc3066051d8533618130c9c5c7eb8929564e4a58 100644
--- a/paddle/fluid/imperative/profiler.cc
+++ b/paddle/fluid/imperative/profiler.cc
@@ -21,7 +21,7 @@
 
 #include <mutex>  // NOLINT
 
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 PADDLE_DEFINE_EXPORTED_string(
     tracer_profile_fname,
diff --git a/paddle/fluid/inference/api/analysis_predictor.cc b/paddle/fluid/inference/api/analysis_predictor.cc
index bff18b330081d76c1e1892a04d56a3f5ee30133e..af4d83f55a6ee2fef289e9e693fb7190cfcb2c4a 100644
--- a/paddle/fluid/inference/api/analysis_predictor.cc
+++ b/paddle/fluid/inference/api/analysis_predictor.cc
@@ -1422,8 +1422,7 @@ CreatePaddlePredictor<AnalysisConfig, PaddleEngineKind::kAnalysis>(
       }
 
       // support set flags from enviorment.
-      const platform::ExportedFlagInfoMap &env_map =
-          platform::GetExportedFlagInfoMap();
+      const phi::ExportedFlagInfoMap &env_map = phi::GetExportedFlagInfoMap();
       std::ostringstream os;
       os << "--tryfromenv=";
       for (auto &pair : env_map) {
diff --git a/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc b/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
index 972665562c30fa98ab88b933ec40fe5b8b68ddb9..309f6c3fdd124043ee582e682502c513f1b5283f 100644
--- a/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
+++ b/paddle/fluid/memory/allocation/auto_growth_best_fit_allocator.cc
@@ -18,8 +18,8 @@
 #include <mutex>  // NOLINT
 
 #include "paddle/fluid/memory/allocation/aligned_allocator.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/fluid/platform/profiler/event_tracing.h"
+#include "paddle/phi/core/flags.h"
 
 PADDLE_DEFINE_EXPORTED_READONLY_bool(
     free_idle_chunk,
diff --git a/paddle/fluid/operators/controlflow/conditional_block_op.cc b/paddle/fluid/operators/controlflow/conditional_block_op.cc
index 1efc5085165776a276c145d23e8f4c22841cb5e0..f11bf6612c8cab0131a20d1e08aaa1776ffddacf 100644
--- a/paddle/fluid/operators/controlflow/conditional_block_op.cc
+++ b/paddle/fluid/operators/controlflow/conditional_block_op.cc
@@ -17,7 +17,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/new_executor/standalone_executor.h"
 #include "paddle/fluid/operators/assign_op.h"
 #include "paddle/fluid/operators/controlflow/control_flow_op_helper.h"
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 #include "paddle/phi/kernels/funcs/math_function.h"
 
 #ifdef PADDLE_WITH_MKLDNN
diff --git a/paddle/fluid/operators/conv_op.cc b/paddle/fluid/operators/conv_op.cc
index 107e3b5a3de495016ded58c45a835cea836ba999..0262c74923e1e80b8c5e4f5b61106682c26853fb 100644
--- a/paddle/fluid/operators/conv_op.cc
+++ b/paddle/fluid/operators/conv_op.cc
@@ -25,7 +25,6 @@ limitations under the License. */
 #include "paddle/fluid/platform/mkldnn_helper.h"
 #endif
 #include "paddle/fluid/framework/infershape_utils.h"
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
 #include "paddle/phi/infermeta/binary.h"
 
 namespace paddle {
diff --git a/paddle/fluid/operators/conv_transpose_op.cc b/paddle/fluid/operators/conv_transpose_op.cc
index 99ec8d0a8b11dcc048041228fdd2cd2b5fe69df0..ebc9f8afdb0b7f84644d718bd3f3427ac8d9938c 100644
--- a/paddle/fluid/operators/conv_transpose_op.cc
+++ b/paddle/fluid/operators/conv_transpose_op.cc
@@ -21,7 +21,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/infershape_utils.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/op_version_registry.h"
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
 #include "paddle/phi/core/infermeta_utils.h"
 #include "paddle/phi/infermeta/backward.h"
 #include "paddle/phi/infermeta/binary.h"
diff --git a/paddle/fluid/operators/fused/fusion_conv_inception_op.cc b/paddle/fluid/operators/fused/fusion_conv_inception_op.cc
index 4dd0b964838b777ddf1c7241ec68259cbc6fad27..9df22199106726bc1f3922459bf6e9066b4d6ef5 100644
--- a/paddle/fluid/operators/fused/fusion_conv_inception_op.cc
+++ b/paddle/fluid/operators/fused/fusion_conv_inception_op.cc
@@ -16,7 +16,7 @@ limitations under the License. */
 #include <vector>
 
 #include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 
 namespace paddle {
 namespace operators {
@@ -113,7 +113,7 @@ class ConvInceptionFusionOpMaker : public framework::OpProtoAndCheckerMaker {
                  "allocated/freed each time the operator runs, larger "
                  "workspace size can increase performance but also requires "
                  "better hardware. This size should be chosen carefully.")
-        .SetDefault(platform::GetDefaultConvWorkspaceSizeLimitMB());
+        .SetDefault(phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB());
     AddComment(R"DOC(
 )DOC");
   }
diff --git a/paddle/fluid/platform/CMakeLists.txt b/paddle/fluid/platform/CMakeLists.txt
index 2db144f423fc7cbfb0e2144cb349aad3875284cd..1dc762b9e185466c39ee7b6388ab34aea499c152 100644
--- a/paddle/fluid/platform/CMakeLists.txt
+++ b/paddle/fluid/platform/CMakeLists.txt
@@ -31,10 +31,6 @@ if(WITH_PYTHON)
   endif()
 endif()
 
-cc_library(
-  flags
-  SRCS flags.cc
-  DEPS gflags)
 cc_library(
   denormal
   SRCS denormal.cc
@@ -178,11 +174,6 @@ if(WITH_GLOO)
     DEPS framework_proto gloo_wrapper enforce)
 endif()
 
-cc_library(
-  cudnn_workspace_helper
-  SRCS cudnn_workspace_helper.cc
-  DEPS)
-
 # separate init from device_context to avoid cycle dependencies
 cc_library(
   init
diff --git a/paddle/fluid/platform/cpu_info.cc b/paddle/fluid/platform/cpu_info.cc
index 8ce9dc54b021d4a6033f9189c2e5bb6bb4f4c409..7edc322a90f5aef6d37cdd0ce8edd4dfee9613cf 100644
--- a/paddle/fluid/platform/cpu_info.cc
+++ b/paddle/fluid/platform/cpu_info.cc
@@ -32,7 +32,7 @@ limitations under the License. */
 
 #include <algorithm>
 
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 DECLARE_double(fraction_of_cpu_memory_to_use);
 DECLARE_uint64(initial_cpu_memory_in_mb);
diff --git a/paddle/fluid/platform/device/gpu/gpu_info.cc b/paddle/fluid/platform/device/gpu/gpu_info.cc
index b25f15d688730f182e6d195bc8c9ae052eb1354d..6952ce33a9318626fd2d9675b264bc77819c7a74 100644
--- a/paddle/fluid/platform/device/gpu/gpu_info.cc
+++ b/paddle/fluid/platform/device/gpu/gpu_info.cc
@@ -24,7 +24,6 @@ limitations under the License. */
 #include "paddle/fluid/memory/memory.h"
 #include "paddle/fluid/platform/cuda_device_guard.h"
 #include "paddle/fluid/platform/enforce.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/fluid/platform/lock_guard_ptr.h"
 #include "paddle/fluid/platform/macros.h"
 #include "paddle/fluid/platform/monitor.h"
@@ -32,6 +31,7 @@ limitations under the License. */
 #include "paddle/fluid/platform/profiler/mem_tracing.h"
 #include "paddle/fluid/string/split.h"
 #include "paddle/phi/backends/gpu/gpu_info.h"
+#include "paddle/phi/core/flags.h"
 
 #ifdef PADDLE_WITH_HIP
 #include "paddle/fluid/platform/dynload/miopen.h"
diff --git a/paddle/fluid/platform/enforce.h b/paddle/fluid/platform/enforce.h
index 4be05f24bc70adbafb903c9620cb929c8cac00d8..eff252b7d79701e46ad42def45385ee4bb85e6b2 100644
--- a/paddle/fluid/platform/enforce.h
+++ b/paddle/fluid/platform/enforce.h
@@ -101,7 +101,7 @@ limitations under the License. */
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 #include "paddle/fluid/platform/device/gpu/gpu_types.h"
 #endif
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 namespace phi {
 class ErrorSummary;
diff --git a/paddle/fluid/platform/profiler/host_tracer.cc b/paddle/fluid/platform/profiler/host_tracer.cc
index db452b3427fd1a68280ca62a2496b273c7392cf7..185792f6750cdba1fdcdcb4ea617a82a3ca08ded 100644
--- a/paddle/fluid/platform/profiler/host_tracer.cc
+++ b/paddle/fluid/platform/profiler/host_tracer.cc
@@ -17,9 +17,9 @@
 
 #include "glog/logging.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/fluid/platform/profiler/common_event.h"
 #include "paddle/fluid/platform/profiler/host_event_recorder.h"
+#include "paddle/phi/core/flags.h"
 
 // Used to filter events, works like glog VLOG(level).
 // RecordEvent will works if host_trace_level >= level.
diff --git a/paddle/fluid/pybind/global_value_getter_setter.cc b/paddle/fluid/pybind/global_value_getter_setter.cc
index c45566ba356732bd9ada901c4e92aee4bbe5085c..e6935b577d737705bffabb8f47c43b6cd5d21ce0 100644
--- a/paddle/fluid/pybind/global_value_getter_setter.cc
+++ b/paddle/fluid/pybind/global_value_getter_setter.cc
@@ -252,7 +252,7 @@ static void RegisterGlobalVarGetterSetter() {
   REGISTER_PUBLIC_GLOBAL_VAR(FLAGS_rpc_prefetch_thread_num);
 #endif
 
-  const auto &flag_map = platform::GetExportedFlagInfoMap();
+  const auto &flag_map = phi::GetExportedFlagInfoMap();
   for (const auto &pair : flag_map) {
     const std::string &name = pair.second.name;
     bool is_writable = pair.second.is_writable;
diff --git a/paddle/phi/api/yaml/generator/ops_extra_info_gen.py b/paddle/phi/api/yaml/generator/ops_extra_info_gen.py
index 34b1f76338110257b93425635bc10c43cc924b03..9109399f9276ba4247eeeedd63dd07e591ab9669 100644
--- a/paddle/phi/api/yaml/generator/ops_extra_info_gen.py
+++ b/paddle/phi/api/yaml/generator/ops_extra_info_gen.py
@@ -22,7 +22,7 @@ def map_code_template(attrs_str, attrs_checker_str):
     return f"""// This file is generated by paddle/phi/api/yaml/generator/ops_extra_info_gen.py
 #include "paddle/fluid/operators/ops_extra_info.h"
 
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 
 namespace paddle {{
 namespace operators {{
diff --git a/paddle/phi/api/yaml/op_compat.yaml b/paddle/phi/api/yaml/op_compat.yaml
index 843ff811f911424be50855df7088a603a8eae10d..7d69e45638beb02b536fa667294485dc5687460d 100644
--- a/paddle/phi/api/yaml/op_compat.yaml
+++ b/paddle/phi/api/yaml/op_compat.yaml
@@ -202,7 +202,7 @@
              str fuse_activation = "", float fuse_alpha = 0.0f, float fuse_beta = 0.0f, bool use_addto = false,
              bool fuse_residual_connection = false, float Scale_in = 1.0f, float Scale_out = 1.0f,
              float Scale_in_eltwise = 1.0f, 'float[] Scale_weights = {1.0f}', bool force_fp32_output = false,
-             int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
+             int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
 
 - op : conv2d_fusion
   extra :
@@ -211,7 +211,7 @@
              str fuse_activation = "", float fuse_alpha = 0.0f, float fuse_beta = 0.0f, bool use_addto = false,
              bool fuse_residual_connection = false, float Scale_in = 1.0f, float Scale_out = 1.0f,
              float Scale_in_eltwise = 1.0f, 'float[] Scale_weights = {1.0f}', bool force_fp32_output = false,
-             int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
+             int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
 
 - op : conv2d_transpose
   backward : conv2d_transpose_grad
@@ -219,7 +219,7 @@
     attrs : [bool is_test = false, bool use_cudnn = true, bool use_mkldnn = false, bool force_fp32_output = false,
              str mkldnn_data_type = "float32", bool fuse_relu = false,
              str fuse_activation = "", float fuse_alpha = 0.0f, float fuse_beta = 0.0f,
-             int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB()]
+             int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB()]
 
 - op : conv3d
   backward : conv3d_grad
@@ -227,12 +227,12 @@
     attrs : [bool is_test = false, bool use_cudnn = true, bool use_mkldnn = false, str mkldnn_data_type = "float32", bool fuse_relu = false,
              str fuse_activation = "", float fuse_alpha = 0.0f, float fuse_beta = 0.0f,
              bool use_addto = false, bool fuse_residual_connection = false, bool force_fp32_output = false,
-             int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
+             int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
 
 - op : conv3d_transpose
   backward : conv3d_transpose_grad
   extra :
-    attrs : [bool use_cudnn = true, bool use_mkldnn = false, int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB()]
+    attrs : [bool use_cudnn = true, bool use_mkldnn = false, int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB()]
 
 - op : cos
   backward : cos_grad, cos_double_grad, cos_triple_grad
@@ -273,7 +273,7 @@
              str fuse_activation = "", float fuse_alpha = 0.0f, float fuse_beta = 0.0f, bool use_addto = false,
              bool fuse_residual_connection = false, float Scale_in = 1.0f, float Scale_out = 1.0f,
              float Scale_in_eltwise = 1.0f, 'float[] Scale_weights = {1.0f}', bool force_fp32_output = false,
-             int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
+             int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB(), bool exhaustive_search = false]
 
 - op : depthwise_conv2d_transpose
   backward : depthwise_conv2d_transpose_grad
@@ -281,7 +281,7 @@
     attrs : [bool is_test = false, bool use_cudnn = false, bool use_mkldnn = false, bool force_fp32_output = false,
              str mkldnn_data_type = "float32", bool fuse_relu = false,
              str fuse_activation = "", float fuse_alpha = 0.0f, float fuse_beta = 0.0f,
-             int workspace_size_MB = platform::GetDefaultConvWorkspaceSizeLimitMB()]
+             int workspace_size_MB = phi::backends::gpu::GetDefaultConvWorkspaceSizeLimitMB()]
 
 - op : dequantize_linear
   extra :
diff --git a/paddle/phi/backends/CMakeLists.txt b/paddle/phi/backends/CMakeLists.txt
index f8a6b2174a830acf212e26bb289204ede08943e6..ddb7adaa92b83897158fc91db10bfae798103355 100644
--- a/paddle/phi/backends/CMakeLists.txt
+++ b/paddle/phi/backends/CMakeLists.txt
@@ -1,4 +1,5 @@
 add_subdirectory(dynload)
+add_subdirectory(gpu)
 
 set(BACKENDS_SRCS all_context.cc cpu/cpu_context.cc)
 set(BACKENDS_DEPS enforce place flags eigen3 phi_device_context)
diff --git a/paddle/phi/backends/gpu/CMakeLists.txt b/paddle/phi/backends/gpu/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..85d57c9b14482bbeaaf8476acaee9a50efb2daa0
--- /dev/null
+++ b/paddle/phi/backends/gpu/CMakeLists.txt
@@ -0,0 +1 @@
+add_subdirectory(cuda)
diff --git a/paddle/phi/backends/gpu/cuda/CMakeLists.txt b/paddle/phi/backends/gpu/cuda/CMakeLists.txt
new file mode 100644
index 0000000000000000000000000000000000000000..7768cdd116122a140d67246b0d7fc8450b4cae22
--- /dev/null
+++ b/paddle/phi/backends/gpu/cuda/CMakeLists.txt
@@ -0,0 +1 @@
+cc_library(cudnn_workspace_helper SRCS cudnn_workspace_helper.cc)
diff --git a/paddle/fluid/platform/cudnn_workspace_helper.cc b/paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.cc
similarity index 79%
rename from paddle/fluid/platform/cudnn_workspace_helper.cc
rename to paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.cc
index bb0e9a226d15001d8f19eff136cb29152e1906fb..d5b484c8eeb561a9bbd652e8e39f887c41bc4a27 100644
--- a/paddle/fluid/platform/cudnn_workspace_helper.cc
+++ b/paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.cc
@@ -1,4 +1,4 @@
-// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -12,13 +12,14 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 
 #include <cstdlib>
 #include <string>
 
-namespace paddle {
-namespace platform {
+namespace phi {
+namespace backends {
+namespace gpu {
 
 static int GetDefaultConvWorkspaceSizeLimitMBImpl() {
   const char *env_str = std::getenv("FLAGS_conv_workspace_size_limit");
@@ -30,6 +31,6 @@ int GetDefaultConvWorkspaceSizeLimitMB() {
   static auto workspace_size = GetDefaultConvWorkspaceSizeLimitMBImpl();
   return workspace_size;
 }
-
-}  // namespace platform
-}  // namespace paddle
+}  // namespace gpu
+}  // namespace backends
+}  // namespace phi
diff --git a/paddle/fluid/platform/cudnn_workspace_helper.h b/paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h
similarity index 78%
rename from paddle/fluid/platform/cudnn_workspace_helper.h
rename to paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h
index 766dbe113c33705d54ce9e25aae5c20ecaf2a068..6b7a1f6eec7bb1bb170594a8d58a21b300346ef9 100644
--- a/paddle/fluid/platform/cudnn_workspace_helper.h
+++ b/paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h
@@ -1,4 +1,4 @@
-// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
+// Copyright (c) 2022 PaddlePaddle Authors. All Rights Reserved.
 //
 // Licensed under the Apache License, Version 2.0 (the "License");
 // you may not use this file except in compliance with the License.
@@ -14,12 +14,13 @@
 
 #pragma once
 
-namespace paddle {
-namespace platform {
+namespace phi {
+namespace backends {
+namespace gpu {
 
 static constexpr int kDefaultConvWorkspaceSizeLimitMB = 512;
 
 int GetDefaultConvWorkspaceSizeLimitMB();
-
-}  // namespace platform
-}  // namespace paddle
+}  // namespace gpu
+}  // namespace backends
+}  // namespace phi
diff --git a/paddle/phi/backends/xpu/xpu_info.cc b/paddle/phi/backends/xpu/xpu_info.cc
index 89ebce438a46492db369345434d8ac577fc634d5..997d3cd9d11bb4df39c0691f70cb30e38fc569cb 100644
--- a/paddle/phi/backends/xpu/xpu_info.cc
+++ b/paddle/phi/backends/xpu/xpu_info.cc
@@ -21,7 +21,7 @@ limitations under the License. */
 
 // TODO(wilber): The phi computing library requires a component to manage
 // flags.
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 
 PADDLE_DEFINE_EXPORTED_string(
     selected_xpus,
diff --git a/paddle/phi/core/CMakeLists.txt b/paddle/phi/core/CMakeLists.txt
index 6dc43ff633f1910d89c70d05b23f22260fc32be5..92d87cf79bd7045abc7b336d54f914960dd464ce 100644
--- a/paddle/phi/core/CMakeLists.txt
+++ b/paddle/phi/core/CMakeLists.txt
@@ -5,6 +5,11 @@ if(WITH_GPU)
   proto_library(external_error_proto SRCS external_error.proto)
 endif()
 
+cc_library(
+  flags
+  SRCS flags.cc
+  DEPS gflags)
+
 cc_library(errors SRCS errors.cc)
 set(phi_enforce_deps errors flags)
 if(WITH_GPU)
diff --git a/paddle/fluid/platform/flags.cc b/paddle/phi/core/flags.cc
similarity index 98%
rename from paddle/fluid/platform/flags.cc
rename to paddle/phi/core/flags.cc
index b809e026544bb8715508e9582087c4bd6d272f06..f11e09cf890e892745b2bd903e28112e85d635f9 100644
--- a/paddle/fluid/platform/flags.cc
+++ b/paddle/phi/core/flags.cc
@@ -13,13 +13,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include "paddle/fluid/platform/flags.h"
+#include "paddle/phi/core/flags.h"
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 #endif
 
-namespace paddle {
-namespace platform {
+namespace phi {
 
 const ExportedFlagInfoMap &GetExportedFlagInfoMap() {
   return *GetMutableExportedFlagInfoMap();
@@ -30,8 +29,7 @@ ExportedFlagInfoMap *GetMutableExportedFlagInfoMap() {
   return &g_exported_flag_info_map;
 }
 
-}  // namespace platform
-}  // namespace paddle
+}  // namespace phi
 
 PADDLE_DEFINE_EXPORTED_int32(inner_op_parallelism,
                              0,
@@ -261,9 +259,10 @@ PADDLE_DEFINE_EXPORTED_bool(
  * increased.
  *       Users need to balance memory and speed.
  */
-PADDLE_DEFINE_EXPORTED_int64(conv_workspace_size_limit,
-                             paddle::platform::kDefaultConvWorkspaceSizeLimitMB,
-                             "cuDNN convolution workspace limit in MB unit.");
+PADDLE_DEFINE_EXPORTED_int64(
+    conv_workspace_size_limit,
+    phi::backends::gpu::kDefaultConvWorkspaceSizeLimitMB,
+    "cuDNN convolution workspace limit in MB unit.");
 
 /**
  * CUDNN related FLAG
diff --git a/paddle/fluid/platform/flags.h b/paddle/phi/core/flags.h
similarity index 95%
rename from paddle/fluid/platform/flags.h
rename to paddle/phi/core/flags.h
index 6db5e710b8dc83f6614f7cc6e284f046e4d74908..e9ace6206255d4992ecee894dde4a97bbd034c09 100644
--- a/paddle/fluid/platform/flags.h
+++ b/paddle/phi/core/flags.h
@@ -20,12 +20,11 @@
 #include <type_traits>
 
 #include "gflags/gflags.h"
-#include "paddle/fluid/platform/macros.h"
+#include "paddle/phi/core/macros.h"
 
 #include "paddle/utils/variant.h"
 
-namespace paddle {
-namespace platform {
+namespace phi {
 
 struct FlagInfo {
   using ValueType =
@@ -51,7 +50,7 @@ ExportedFlagInfoMap *GetMutableExportedFlagInfoMap();
       static_assert(std::is_same<FlagDeclaredType, ::std::string>::value ||   \
                         std::is_arithmetic<FlagDeclaredType>::value,          \
                     "FLAGS should be std::string or arithmetic type");        \
-      auto *instance = ::paddle::platform::GetMutableExportedFlagInfoMap();   \
+      auto *instance = ::phi::GetMutableExportedFlagInfoMap();                \
       auto &info = (*instance)[#__name];                                      \
       info.name = #__name;                                                    \
       info.value_ptr = &(FLAGS_##__name);                                     \
@@ -96,5 +95,4 @@ ExportedFlagInfoMap *GetMutableExportedFlagInfoMap();
   __PADDLE_DEFINE_EXPORTED_FLAG(                                \
       name, true, ::std::string, string, default_value, doc)
 
-}  // namespace platform
-}  // namespace paddle
+}  // namespace phi
diff --git a/paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu b/paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
index fd6e92b2ffe06df67056874c545b9264b76a743d..252fbe7d21b74406fa2a89d10f1ed77230555f09 100644
--- a/paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
+++ b/paddle/phi/kernels/gpu/batch_norm_grad_kernel.cu
@@ -14,11 +14,11 @@
 
 #include "paddle/fluid/operators/layout_utils.h"
 #include "paddle/fluid/operators/norm_utils.cu.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_dnn.h"
 #include "paddle/phi/common/layout.h"
 #include "paddle/phi/core/enforce.h"
+#include "paddle/phi/core/flags.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/batch_norm_kernel.h"
 #include "paddle/phi/kernels/empty_kernel.h"
diff --git a/paddle/phi/kernels/gpu/batch_norm_kernel.cu b/paddle/phi/kernels/gpu/batch_norm_kernel.cu
index cd1665ba9032e6b9c19c4629268ab7dfce5a5615..d01397c1fa0665a2059ed0e11a1c306acddfb98e 100644
--- a/paddle/phi/kernels/gpu/batch_norm_kernel.cu
+++ b/paddle/phi/kernels/gpu/batch_norm_kernel.cu
@@ -22,11 +22,11 @@ namespace cub = hipcub;
 
 #include "paddle/fluid/operators/layout_utils.h"
 #include "paddle/fluid/operators/norm_utils.cu.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/backends/gpu/gpu_dnn.h"
 #include "paddle/phi/common/layout.h"
 #include "paddle/phi/core/enforce.h"
+#include "paddle/phi/core/flags.h"
 #include "paddle/phi/core/kernel_registry.h"
 #include "paddle/phi/kernels/batch_norm_kernel.h"
 #include "paddle/phi/kernels/funcs/batch_norm_utils.h"
diff --git a/paddle/phi/kernels/gpu/gelu_funcs.h b/paddle/phi/kernels/gpu/gelu_funcs.h
index 81f5354d42c5666590e4cc9e284c3ad3615512ae..42b5322426f70f85d4fcff283d344f138cd0f634 100644
--- a/paddle/phi/kernels/gpu/gelu_funcs.h
+++ b/paddle/phi/kernels/gpu/gelu_funcs.h
@@ -14,10 +14,10 @@
 
 #pragma once
 
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/phi/backends/gpu/gpu_context.h"
 #include "paddle/phi/common/amp_type_traits.h"
 #include "paddle/phi/common/place.h"
+#include "paddle/phi/core/flags.h"
 #include "paddle/phi/kernels/funcs/aligned_vector.h"
 
 DECLARE_bool(use_fast_math);
diff --git a/paddle/phi/kernels/gpudnn/conv_grad_kernel.cu b/paddle/phi/kernels/gpudnn/conv_grad_kernel.cu
index f1114e59bad75454e63e1a8d4aba1d7a40482f2a..166403454a4176be8610dce0d26085de55ab8814 100644
--- a/paddle/phi/kernels/gpudnn/conv_grad_kernel.cu
+++ b/paddle/phi/kernels/gpudnn/conv_grad_kernel.cu
@@ -23,8 +23,8 @@
 #include "paddle/phi/kernels/gpudnn/conv_cudnn_v7.h"
 #endif
 
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
 #include "paddle/fluid/platform/profiler.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/cpu/conv_util.h"
diff --git a/paddle/phi/kernels/gpudnn/conv_kernel.cu b/paddle/phi/kernels/gpudnn/conv_kernel.cu
index 04ac84d606effd425df0aaf645bc7bce1d0003d0..40d75ffcdf1f94eaacfe65e06ecd568d94a882bc 100644
--- a/paddle/phi/kernels/gpudnn/conv_kernel.cu
+++ b/paddle/phi/kernels/gpudnn/conv_kernel.cu
@@ -24,8 +24,8 @@
 #include "paddle/phi/kernels/gpudnn/conv_cudnn_v7.h"
 #endif
 
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
 #include "paddle/fluid/platform/profiler.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 #include "paddle/phi/common/bfloat16.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/cpu/conv_util.h"
diff --git a/paddle/phi/kernels/impl/conv_cudnn_impl.h b/paddle/phi/kernels/impl/conv_cudnn_impl.h
index cf419cd1fd4b38a26e8888b401899ff26b9c1405..300976fbef6b13f2acb57de81c7f4671ad540f90 100644
--- a/paddle/phi/kernels/impl/conv_cudnn_impl.h
+++ b/paddle/phi/kernels/impl/conv_cudnn_impl.h
@@ -23,9 +23,9 @@
 #include "paddle/phi/kernels/gpudnn/conv_cudnn_v7.h"
 #endif
 
-#include "paddle/fluid/platform/cudnn_workspace_helper.h"
 #include "paddle/fluid/platform/profiler.h"
 #include "paddle/phi/backends/dynload/cudnn.h"
+#include "paddle/phi/backends/gpu/cuda/cudnn_workspace_helper.h"
 #include "paddle/phi/common/float16.h"
 #include "paddle/phi/kernels/cpu/conv_util.h"
 #include "paddle/phi/kernels/funcs/batch_norm_utils.h"
diff --git a/paddle/testing/paddle_gtest_main.cc b/paddle/testing/paddle_gtest_main.cc
index 16c683e39fa8cd4b0e3ed469d6ea46ba221e5ee9..6a57737ac4127da6a30b635f0c72932ba228d091 100644
--- a/paddle/testing/paddle_gtest_main.cc
+++ b/paddle/testing/paddle_gtest_main.cc
@@ -17,8 +17,8 @@ limitations under the License. */
 #include "paddle/fluid/framework/phi_utils.h"
 #include "paddle/fluid/memory/allocation/allocator_strategy.h"
 #include "paddle/fluid/platform/device/npu/npu_info.h"
-#include "paddle/fluid/platform/flags.h"
 #include "paddle/fluid/platform/init.h"
+#include "paddle/phi/core/flags.h"
 
 #if defined(PADDLE_WITH_CUDA) || defined(PADDLE_WITH_HIP)
 DECLARE_bool(enable_gpu_memory_usage_log);
@@ -43,7 +43,7 @@ int main(int argc, char** argv) {
   }
 #endif
 
-  const auto& flag_map = paddle::platform::GetExportedFlagInfoMap();
+  const auto& flag_map = phi::GetExportedFlagInfoMap();
   for (const auto& pair : flag_map) {
     const std::string& name = pair.second.name;
     // NOTE(zhiqiu): some names may not linked in some tests, so add to