[PIR AMP]Split tracer and amp attr (#61297)

0x45f · web-flow · commit 250d6a017571 · 2024-01-31T16:22:20.000+08:00
diff --git a/paddle/fluid/eager/api/manual/eager_manual/forwards/add_n_fwd_func.cc b/paddle/fluid/eager/api/manual/eager_manual/forwards/add_n_fwd_func.cc
@@ -42,7 +42,7 @@ paddle::Tensor add_n_ad_func(const std::vector<paddle::Tensor>& x) {
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return add_n_ad_func(NEW_x);
     }
diff --git a/paddle/fluid/eager/api/manual/eager_manual/forwards/conv2d_fwd_function.cc b/paddle/fluid/eager/api/manual/eager_manual/forwards/conv2d_fwd_function.cc
@@ -53,7 +53,7 @@ paddle::Tensor conv2d_ad_func(const paddle::Tensor& input,
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return conv2d_ad_func(new_input,
                             new_filter,
diff --git a/paddle/fluid/eager/api/manual/eager_manual/forwards/multiply_fwd_func.cc b/paddle/fluid/eager/api/manual/eager_manual/forwards/multiply_fwd_func.cc
@@ -52,7 +52,7 @@ paddle::Tensor multiply_ad_func(const paddle::Tensor& x,
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return multiply_ad_func(new_x, new_y);
     }
@@ -399,7 +399,7 @@ paddle::Tensor multiply_ad_func(const paddle::Tensor& x,
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return multiply_ad_func(new_x, new_y);
     }
diff --git a/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_attention_fwd_func.cc b/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_attention_fwd_func.cc
@@ -128,7 +128,7 @@ fused_attention_dygraph_function(
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return fused_attention_dygraph_function(NEW_X,
                                               NEW_LnScale,
diff --git a/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_bias_dropout_residual_layer_norm_fwd_func.cc b/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_bias_dropout_residual_layer_norm_fwd_func.cc
@@ -83,7 +83,7 @@ fused_bias_dropout_residual_layer_norm_dygraph_function(
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return fused_bias_dropout_residual_layer_norm_dygraph_function(
           NEW_X, NEW_Residual, NEW_Bias, NEW_LnScale, NEW_LnBias, attr_map);
diff --git a/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_feedforward_fwd_func.cc b/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_feedforward_fwd_func.cc
@@ -122,7 +122,7 @@ fused_feedforward_dygraph_function(
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return fused_feedforward_dygraph_function(NEW_X,
                                                 NEW_Dropout1Seed,
diff --git a/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_gate_attention_fwd_func.cc b/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_gate_attention_fwd_func.cc
@@ -127,7 +127,7 @@ fused_gate_attention_dygraph_function(
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return fused_gate_attention_dygraph_function(NEW_Query,
                                                    NEW_Key,
diff --git a/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_gemm_epilogue_fwd_func.cc b/paddle/fluid/eager/api/manual/fluid_manual/forwards/fused_gemm_epilogue_fwd_func.cc
@@ -49,7 +49,7 @@ paddle::Tensor fused_gemm_epilogue_dygraph_function(
 
     {
       paddle::imperative::AutoCastGuard guard(
-          egr::Controller::Instance().GetCurrentTracer(),
+          egr::Controller::Instance().GetCurrentAMPState(),
           paddle::imperative::AmpLevel::O0);
       return fused_gemm_epilogue_dygraph_function(
           NEW_X, NEW_Y, NEW_Bias, attr_map);
diff --git a/paddle/fluid/eager/api/utils/global_utils.h b/paddle/fluid/eager/api/utils/global_utils.h
@@ -74,19 +74,27 @@ class Controller {
   std::string GetPythonStack() { return tracer_->GetPythonStack(); }
 
   bool HasGrad() const { return tracer_->HasGrad(); }
+
   void SetHasGrad(bool has_grad) { tracer_->SetHasGrad(has_grad); }
+
   std::string GenerateUniqueName(std::string key = "eager_in_tmp") {
     return tracer_->GenerateUniqueName(key);
   }
+
   const std::shared_ptr<paddle::imperative::Tracer>& GetCurrentTracer() {
     return tracer_;
   }
+
   void SetCurrentTracer(
       const std::shared_ptr<paddle::imperative::Tracer>& tracer) {
     tracer_ = tracer;
     VLOG(6) << "Set current tracer for Controller: " << tracer_;
   }
 
+  const std::shared_ptr<paddle::imperative::AMPState>& GetCurrentAMPState() {
+    return paddle::imperative::GetCurrentAMPState();
+  }
+
   const std::unordered_map<std::string, std::vector<paddle::OpMetaInfo>>&
   GetOpMetaInfoMap() {
     return op_meta_info_map_;
diff --git a/paddle/fluid/eager/auto_code_generator/eager_generator.cc b/paddle/fluid/eager/auto_code_generator/eager_generator.cc
@@ -1763,7 +1763,7 @@ static std::pair<std::string, std::string> GenerateForwardFunctionContents(
     const char* CALL_BACK_TEMPLATE =
         "    {\n"
         "      paddle::imperative::AutoCastGuard "
-        "guard(egr::Controller::Instance().GetCurrentTracer(), "
+        "guard(egr::Controller::Instance().GetCurrentAMPState(), "
         "paddle::imperative::AmpLevel::O0);\n"
         "      return %s_dygraph_function(%s);\n"
         "    }";
diff --git a/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py b/paddle/fluid/eager/auto_code_generator/generator/eager_gen.py
@@ -520,7 +520,7 @@ class {} : public egr::GradNodeBase {{
     {}
     {}
     {{
-      paddle::imperative::AutoCastGuard guard(egr::Controller::Instance().GetCurrentTracer(), paddle::imperative::AmpLevel::O0);
+      paddle::imperative::AutoCastGuard guard(egr::Controller::Instance().GetCurrentAMPState(), paddle::imperative::AmpLevel::O0);
       {}
     }}
   }}
diff --git a/paddle/fluid/imperative/amp_auto_cast.cc b/paddle/fluid/imperative/amp_auto_cast.cc
@@ -119,17 +119,17 @@ OpSupportedInfos(const std::string& place,
       std::move(all_ops), std::move(supported_ops), std::move(unsupported_ops));
 }
 
-AutoCastGuard::AutoCastGuard(std::shared_ptr<Tracer> tracer, AmpLevel level)
-    : tracer_(tracer) {
-  pre_amp_level_ = tracer_->GetAmpLevel();
+AutoCastGuard::AutoCastGuard(std::shared_ptr<AMPState> state, AmpLevel level)
+    : state_(state) {
+  pre_amp_level_ = state_->GetAmpLevel();
 
   if (pre_amp_level_ != level) {
-    tracer_->SetAmpLevel(level);
+    state_->SetAmpLevel(level);
   }
 }
 
 AutoCastGuard::~AutoCastGuard() {  // NOLINT
-  tracer_->SetAmpLevel(pre_amp_level_);
+  state_->SetAmpLevel(pre_amp_level_);
 }
 
 AmpOperators::AmpOperators()
@@ -231,6 +231,46 @@ std::ostream& operator<<(std::ostream& os, AmpOperators& ops) {
   return os;
 }
 
+thread_local bool AMPState::use_promote_ = false;
+
+thread_local AmpLevel AMPState::amp_level_ = AmpLevel::O0;
+
+thread_local phi::DataType AMPState::amp_dtype_ = phi::DataType::FLOAT32;
+
+AMPState::AMPState() {}
+
+AMPState::~AMPState() = default;
+
+bool AMPState::GetUsePromote() const { return use_promote_; }
+
+void AMPState::SetUsePromote(bool use_promote) { use_promote_ = use_promote; }
+
+AmpLevel AMPState::GetAmpLevel() const { return amp_level_; }
+
+void AMPState::SetAmpLevel(AmpLevel level) { amp_level_ = level; }
+
+std::string AMPState::GetAmpDtype() const {
+  if (amp_dtype_ == phi::DataType::FLOAT16) {
+    return std::string("float16");
+  } else if (amp_dtype_ == phi::DataType::BFLOAT16) {
+    return std::string("bfloat16");
+  } else {
+    return std::string("float32");
+  }
+}
+
+void AMPState::SetAmpDtype(std::string amp_dtype) {
+  if (amp_dtype == "float16") {
+    amp_dtype_ = phi::DataType::FLOAT16;
+  } else if (amp_dtype == "bfloat16") {
+    amp_dtype_ = phi::DataType::BFLOAT16;
+  } else {
+    amp_dtype_ = phi::DataType::FLOAT32;
+  }
+}
+
+phi::DataType AMPState::GetAmpPhiDtype() const { return amp_dtype_; }
+
 template <typename VarType>
 inline std::string GetDtypeStr(const std::shared_ptr<VarType>& var) {
   return framework::DataTypeToString(GetDataType<VarType>(var));
@@ -268,7 +308,7 @@ static inline std::shared_ptr<VarType> CastToType(
   imperative::NameVarMap<VarType> outs = {{"Out", {out}}};
 
   {
-    AutoCastGuard guard(tracer, AmpLevel::O0);
+    AutoCastGuard guard(imperative::GetCurrentAMPState(), AmpLevel::O0);
     tracer->TraceOp("cast", ins, outs, std::move(attrs));
   }
 
diff --git a/paddle/fluid/imperative/amp_auto_cast.h b/paddle/fluid/imperative/amp_auto_cast.h
@@ -84,10 +84,29 @@ class AmpOperators {
 
 std::ostream& operator<<(std::ostream& os, AmpOperators& ops);
 
+class AMPState {
+ public:
+  AMPState();
+  ~AMPState();
+  bool GetUsePromote() const;
+  void SetUsePromote(bool use_promote);
+  AmpLevel GetAmpLevel() const;
+  void SetAmpLevel(AmpLevel level);
+  std::string GetAmpDtype() const;
+  void SetAmpDtype(std::string amp_dtype);
+  phi::DataType GetAmpPhiDtype() const;
+  // void Reset();
+
+ private:
+  static thread_local bool use_promote_;
+  static thread_local AmpLevel amp_level_;
+  static thread_local phi::DataType amp_dtype_;
+};
+
 // NOTE(zhiqiu): AutoCastGuard is used for RAII.
 class AutoCastGuard {
  public:
-  AutoCastGuard(std::shared_ptr<Tracer> tracer, AmpLevel guard_level);
+  AutoCastGuard(std::shared_ptr<AMPState> state, AmpLevel guard_level);
 
   ~AutoCastGuard();
 
@@ -96,7 +115,7 @@ class AutoCastGuard {
   AutoCastGuard& operator=(const AutoCastGuard& guard) = delete;
 
  private:
-  std::shared_ptr<Tracer> tracer_;
+  std::shared_ptr<AMPState> state_;
   AmpLevel pre_amp_level_;
 };
 
diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc
@@ -48,16 +48,13 @@ thread_local bool Tracer::enable_program_desc_tracing_ = false;
 
 thread_local bool Tracer::has_grad_ = true;
 
-thread_local bool Tracer::use_promote_ = true;
-
 thread_local bool Tracer::use_layout_autotune_ = false;
 
-thread_local AmpLevel Tracer::amp_level_ = AmpLevel::O0;
-
-thread_local phi::DataType Tracer::amp_dtype_ = phi::DataType::FLOAT32;
-
 static thread_local std::shared_ptr<Tracer> g_current_tracer(nullptr);
 
+static thread_local std::shared_ptr<AMPState> g_current_amp_state =
+    std::make_shared<AMPState>();
+
 TEST_API void Tracer::DisableLayoutAutoTune() { use_layout_autotune_ = false; }
 TEST_API void Tracer::EnableLayoutAutoTune() {
   use_layout_autotune_ = true;
@@ -92,6 +89,10 @@ TEST_API void SetCurrentTracer(const std::shared_ptr<Tracer>& tracer) {
   VLOG(6) << "Set current tracer: " << g_current_tracer;
 }
 
+const std::shared_ptr<AMPState>& GetCurrentAMPState() {
+  return g_current_amp_state;
+}
+
 void PassStopGradient(const NameVarBaseMap& outs, bool generate_grad) {
   for (const auto& pair : outs) {
     for (const auto& var : pair.second) {
@@ -275,22 +276,24 @@ void Tracer::TraceOpImpl(const std::string& type,
                               : attr_checker->GetDefaultAttrMap();
 
   std::unique_ptr<NameVarMap<VarType>> ins_amp = nullptr;
-  if (amp_level_ == AmpLevel::O1) {
-    if (amp_dtype_ == phi::DataType::FLOAT16) {
+  if (GetCurrentAMPState()->GetAmpLevel() == AmpLevel::O1) {
+    if (GetCurrentAMPState()->GetAmpPhiDtype() == phi::DataType::FLOAT16) {
       VLOG(5) << "Float16 Auto Mixed Precision O1 run operator: " << type;
       ins_amp = std::make_unique<NameVarMap<VarType>>(
           AutoCastInputs<VarType>(type, ins));
-    } else if (amp_dtype_ == phi::DataType::BFLOAT16) {
+    } else if (GetCurrentAMPState()->GetAmpPhiDtype() ==
+               phi::DataType::BFLOAT16) {
       VLOG(5) << "BFloat16 Auto Mixed Precision O1 run operator: " << type;
       ins_amp = std::make_unique<NameVarMap<VarType>>(
           AutoCastBF16Inputs<VarType>(type, ins));
     }
-  } else if (amp_level_ == AmpLevel::O2) {
-    if (amp_dtype_ == phi::DataType::FLOAT16) {
+  } else if (GetCurrentAMPState()->GetAmpLevel() == AmpLevel::O2) {
+    if (GetCurrentAMPState()->GetAmpPhiDtype() == phi::DataType::FLOAT16) {
       VLOG(5) << "Float16 Auto Mixed Precision O2 run operator: " << type;
       ins_amp = std::make_unique<NameVarMap<VarType>>(
           CastPureFp16Inputs<VarType>(type, ins));
-    } else if (amp_dtype_ == phi::DataType::BFLOAT16) {
+    } else if (GetCurrentAMPState()->GetAmpPhiDtype() ==
+               phi::DataType::BFLOAT16) {
       VLOG(5) << "BFloat16 Auto Mixed Precision O2 run operator: " << type;
       ins_amp = std::make_unique<NameVarMap<VarType>>(
           CastPureBf16Inputs<VarType>(type, ins));
@@ -557,17 +560,21 @@ TEST_API void Tracer::SetHasGrad(bool has_grad) { has_grad_ = has_grad; }
 
 TEST_API void Tracer::SetUsePromote(bool use_promote) {
   VLOG(4) << "set use_promote to " << use_promote;
-  use_promote_ = use_promote;
+  g_current_amp_state->SetUsePromote(use_promote);
 }
 
-TEST_API bool Tracer::GetUsePromote() const { return use_promote_; }
+TEST_API bool Tracer::GetUsePromote() const {
+  return g_current_amp_state->GetUsePromote();
+}
 
 TEST_API void Tracer::SetAmpLevel(AmpLevel level) {
   VLOG(4) << "set amp_level to " << static_cast<unsigned int>(level);
-  amp_level_ = level;
+  g_current_amp_state->SetAmpLevel(level);
 }
 
-TEST_API AmpLevel Tracer::GetAmpLevel() const { return amp_level_; }
+TEST_API AmpLevel Tracer::GetAmpLevel() const {
+  return g_current_amp_state->GetAmpLevel();
+}
 
 bool Tracer::ComputeRequiredGrad(const NameVarBaseMap& ins,
                                  const NameVarBaseMap& outs,
@@ -597,26 +604,17 @@ bool Tracer::IsProgramDescTracingEnabled() const {
 
 void Tracer::SetAmpDtype(std::string amp_dtype) {
   VLOG(4) << "set amp_dtype to " << amp_dtype;
-  if (amp_dtype == "float16") {
-    amp_dtype_ = phi::DataType::FLOAT16;
-  } else if (amp_dtype == "bfloat16") {
-    amp_dtype_ = phi::DataType::BFLOAT16;
-  } else {
-    amp_dtype_ = phi::DataType::FLOAT32;
-  }
+  g_current_amp_state->SetAmpDtype(amp_dtype);
 }
 
 std::string Tracer::GetAmpDtype() const {
-  if (amp_dtype_ == phi::DataType::FLOAT16) {
-    return std::string("float16");
-  } else if (amp_dtype_ == phi::DataType::BFLOAT16) {
-    return std::string("bfloat16");
-  } else {
-    return std::string("float32");
-  }
+  return g_current_amp_state->GetAmpDtype();
+}
+
+phi::DataType Tracer::GetAmpPhiDtype() const {
+  return g_current_amp_state->GetAmpPhiDtype();
 }
 
-phi::DataType Tracer::GetAmpPhiDtype() const { return amp_dtype_; }
 bool Tracer::ComputeRequiredGrad(const NameTensorMap& ins,
                                  const NameTensorMap& outs,
                                  bool trace_backward) {
diff --git a/paddle/fluid/imperative/tracer.h b/paddle/fluid/imperative/tracer.h
@@ -195,14 +195,12 @@ class Tracer {
   static thread_local bool enable_program_desc_tracing_;
   static thread_local bool use_layout_autotune_;
   static thread_local bool has_grad_;
-  static thread_local bool use_promote_;
-  static thread_local AmpLevel amp_level_;
-  static thread_local phi::DataType amp_dtype_;
 };
 
 // To access static variable current_tracer
 const std::shared_ptr<Tracer>& GetCurrentTracer();
 TEST_API void SetCurrentTracer(const std::shared_ptr<Tracer>& tracer_);
+const std::shared_ptr<AMPState>& GetCurrentAMPState();
 void IncreaseVarbaseReferenceCountUntilCopyComplete(
     const std::shared_ptr<imperative::VarBase>& var,
     const platform::Place& place);
diff --git a/paddle/fluid/pybind/imperative.cc b/paddle/fluid/pybind/imperative.cc
diff --git a/python/paddle/base/core.py b/python/paddle/base/core.py
diff --git a/test/amp/test_amp_api.py b/test/amp/test_amp_api.py

Original file line number	Diff line number	Diff line change
`@@ -42,7 +42,7 @@ paddle::Tensor add_n_ad_func(const std::vector<paddle::Tensor>& x) {`
`42`	`42`
`43`	`43`	`{`
`44`	`44`	`paddle::imperative::AutoCastGuard guard(`
`45`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`45`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`46`	`46`	`paddle::imperative::AmpLevel::O0);`
`47`	`47`	`return add_n_ad_func(NEW_x);`
`48`	`48`	`}`
Original file line number	Diff line number	Diff line change
`@@ -53,7 +53,7 @@ paddle::Tensor conv2d_ad_func(const paddle::Tensor& input,`
`53`	`53`
`54`	`54`	`{`
`55`	`55`	`paddle::imperative::AutoCastGuard guard(`
`56`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`56`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`57`	`57`	`paddle::imperative::AmpLevel::O0);`
`58`	`58`	`return conv2d_ad_func(new_input,`
`59`	`59`	`new_filter,`
Original file line number	Diff line number	Diff line change
`@@ -52,7 +52,7 @@ paddle::Tensor multiply_ad_func(const paddle::Tensor& x,`
`52`	`52`
`53`	`53`	`{`
`54`	`54`	`paddle::imperative::AutoCastGuard guard(`
`55`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`55`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`56`	`56`	`paddle::imperative::AmpLevel::O0);`
`57`	`57`	`return multiply_ad_func(new_x, new_y);`
`58`	`58`	`}`
`@@ -399,7 +399,7 @@ paddle::Tensor multiply_ad_func(const paddle::Tensor& x,`
`399`	`399`
`400`	`400`	`{`
`401`	`401`	`paddle::imperative::AutoCastGuard guard(`
`402`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`402`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`403`	`403`	`paddle::imperative::AmpLevel::O0);`
`404`	`404`	`return multiply_ad_func(new_x, new_y);`
`405`	`405`	`}`
Original file line number	Diff line number	Diff line change
`@@ -128,7 +128,7 @@ fused_attention_dygraph_function(`
`128`	`128`
`129`	`129`	`{`
`130`	`130`	`paddle::imperative::AutoCastGuard guard(`
`131`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`131`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`132`	`132`	`paddle::imperative::AmpLevel::O0);`
`133`	`133`	`return fused_attention_dygraph_function(NEW_X,`
`134`	`134`	`NEW_LnScale,`
Original file line number	Diff line number	Diff line change
`@@ -83,7 +83,7 @@ fused_bias_dropout_residual_layer_norm_dygraph_function(`
`83`	`83`
`84`	`84`	`{`
`85`	`85`	`paddle::imperative::AutoCastGuard guard(`
`86`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`86`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`87`	`87`	`paddle::imperative::AmpLevel::O0);`
`88`	`88`	`return fused_bias_dropout_residual_layer_norm_dygraph_function(`
`89`	`89`	`NEW_X, NEW_Residual, NEW_Bias, NEW_LnScale, NEW_LnBias, attr_map);`
Original file line number	Diff line number	Diff line change
`@@ -122,7 +122,7 @@ fused_feedforward_dygraph_function(`
`122`	`122`
`123`	`123`	`{`
`124`	`124`	`paddle::imperative::AutoCastGuard guard(`
`125`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`125`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`126`	`126`	`paddle::imperative::AmpLevel::O0);`
`127`	`127`	`return fused_feedforward_dygraph_function(NEW_X,`
`128`	`128`	`NEW_Dropout1Seed,`
Original file line number	Diff line number	Diff line change
`@@ -127,7 +127,7 @@ fused_gate_attention_dygraph_function(`
`127`	`127`
`128`	`128`	`{`
`129`	`129`	`paddle::imperative::AutoCastGuard guard(`
`130`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`130`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`131`	`131`	`paddle::imperative::AmpLevel::O0);`
`132`	`132`	`return fused_gate_attention_dygraph_function(NEW_Query,`
`133`	`133`	`NEW_Key,`
Original file line number	Diff line number	Diff line change
`@@ -49,7 +49,7 @@ paddle::Tensor fused_gemm_epilogue_dygraph_function(`
`49`	`49`
`50`	`50`	`{`
`51`	`51`	`paddle::imperative::AutoCastGuard guard(`
`52`		`- egr::Controller::Instance().GetCurrentTracer(),`
	`52`	`+ egr::Controller::Instance().GetCurrentAMPState(),`
`53`	`53`	`paddle::imperative::AmpLevel::O0);`
`54`	`54`	`return fused_gemm_epilogue_dygraph_function(`
`55`	`55`	`NEW_X, NEW_Y, NEW_Bias, attr_map);`