[XPU] default no autotune (PaddlePaddle#62636)

hanhaowen-mt · Mar 12, 2024 · fde63d1 · fde63d1
1 parent d527fb5
commit fde63d1
Show file tree

Hide file tree

Showing 3 changed files with 8 additions and 2 deletions.
diff --git a/paddle/fluid/inference/api/analysis_config.cc b/paddle/fluid/inference/api/analysis_config.cc
@@ -180,6 +180,11 @@ void AnalysisConfig::EnableXpu(int l3_size,
                                bool transformer_encoder_adaptive_seqlen,
                                bool enable_multi_stream) {
 #if defined(PADDLE_WITH_XPU) || defined(LITE_SUBGRAPH_WITH_XPU)
+  LOG_FIRST_N(WARNING, 1)
+      << "Parameters in EnableXpu/enable_xpu is deprecated since version "
+         "2.6.1, and will be removed in version 3.0! Please use "
+         "EnableXpu/enable_xpu without parameters, and use "
+         "SetXpuConfig/set_xpu_config to set options.";
   use_xpu_ = true;
   xpu_config_.l3_size = l3_size;
   xpu_config_.conv_autotune_level = conv_autotune;

diff --git a/paddle/fluid/inference/api/paddle_analysis_config.h b/paddle/fluid/inference/api/paddle_analysis_config.h
@@ -111,6 +111,7 @@ struct PD_INFER_DECL XpuConfig {
   bool conv_autotune_file_writeback{false};
 
   // Fc autotune level. The Optional values are 0-9. Default 0 means no
+  // autotune.
   int fc_autotune_level{0};
   // Base fc autotune info is read from fc_autotune_file.
   std::string fc_autotune_file;
@@ -367,7 +368,7 @@ struct PD_INFER_DECL AnalysisConfig {
   ///
   void EnableXpu(int l3_size = 0xfffc00,
                  bool l3_locked = false,
-                 bool conv_autotune = true,
+                 bool conv_autotune = false,
                  const std::string& conv_autotune_file = "",
                  const std::string& transformer_encoder_precision = "int16",
                  bool transformer_encoder_adaptive_seqlen = false,

diff --git a/paddle/fluid/pybind/inference_api.cc b/paddle/fluid/pybind/inference_api.cc
@@ -800,7 +800,7 @@ void BindAnalysisConfig(py::module *m) {
            &AnalysisConfig::EnableXpu,
            py::arg("l3_size") = 16 * 1024 * 1024,
            py::arg("l3_locked") = false,
-           py::arg("conv_autotune") = true,
+           py::arg("conv_autotune") = false,
            py::arg("conv_autotune_file") = "",
            py::arg("transformer_encoder_precision") = "int16",
            py::arg("transformer_encoder_adaptive_seqlen") = false,