Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into…

… remove_set_value_infershape

paddle/fluid/eager/auto_code_generator/eager_generator.cc

@@ -56,23 +56,29 @@ static std::string LegalizeVariableName(const std::string& var_name) {
   return ret;
 }
 
-static bool IgnoreGradAttribute(const std::string& op_type,
-                                const std::string& attr_name) {
-  // Attributes in operators_with_attrs are created manually during code
-  // generation
-  // We should ignore these arbitrary attrs when setting up grad attribute map
-  if (operators_with_attrs.count(op_type)) {
-    if (operators_with_attrs[op_type].count(attr_name)) {
-      return true;
-    }
-  }
+static std::string HandleDynamicGradAttributes(const std::string& fwd_op_type,
+                                               const std::string& attrs_name) {
+  std::string additional_grad_attrs_str = "";
+
+  if (fwd_op_type == "sum") {
+    const char* GRAD_ATTRS_TEMPLATE = "  %s[\"%s\"] = %s;\n";
+    additional_grad_attrs_str = paddle::string::Sprintf(
+        GRAD_ATTRS_TEMPLATE, attrs_name, "scale", "float(1.0)");
+    additional_grad_attrs_str += paddle::string::Sprintf(
+        GRAD_ATTRS_TEMPLATE, attrs_name, "bias", "float(0.0f)");
+    additional_grad_attrs_str += paddle::string::Sprintf(
+        GRAD_ATTRS_TEMPLATE, attrs_name, "bias_after_scale", "bool(true)");
 
-  // Only allow SumOp
-  if (op_type != "sum") {
-    return true;
+  } else if (fwd_op_type == "scale") {
+    const char* GRAD_ATTRS_TEMPLATE = "  %s[\"%s\"] = %s;\n";
+
+    additional_grad_attrs_str += paddle::string::Sprintf(
+        GRAD_ATTRS_TEMPLATE, attrs_name, "bias", "float(0.0f)");
+    additional_grad_attrs_str += paddle::string::Sprintf(
+        GRAD_ATTRS_TEMPLATE, attrs_name, "bias_after_scale", "bool(true)");
   }
 
-  return false;
+  return additional_grad_attrs_str;
 }
 
 static void PrepareAttrMapForOps() {
@@ -1866,18 +1872,9 @@ static std::string GenerateSingleOpBase(
   const char* ATTRS_TEMPLATE = "  auto& %s = this->attr_map_;\n";
   std::string grad_attrs_str =
       paddle::string::Sprintf(ATTRS_TEMPLATE, attrs_name);
-  for (const auto& iter : grad_attrs) {
-    if (IgnoreGradAttribute(fwd_op_type, iter.first)) continue;
-    std::pair<std::string, std::string> type_val =
-        GetAttrType(iter.second, false /*is_arg*/);
-    const char* GRAD_ATTRS_TEMPLATE =
-        "  %s %s = %s;\n"
-        "  %s[\"%s\"] = %s;\n";
-    std::string var_name = iter.first + std::to_string(*outs_size);
-    grad_attrs_str += paddle::string::Sprintf(
-        GRAD_ATTRS_TEMPLATE, type_val.first, var_name, type_val.second,
-        attrs_name, iter.first, var_name);
-  }
+
+  // Handle dynamic grad attributes
+  grad_attrs_str += HandleDynamicGradAttributes(fwd_op_type, attrs_name);
   generated_grad_function_body += grad_attrs_str;
 
   const char* TRACE_OP_TEMPLATE =

paddle/fluid/eager/auto_code_generator/final_state_generator/eager_gen.py

@@ -28,6 +28,7 @@
 yaml_types_mapping = {
     'int' : 'int', 'int32' : 'int32_t', 'int64' : 'int64_t',  'size_t' : 'size_t', \
     'float' : 'float', 'double' : 'double', 'bool' : 'bool', \
+    'str' : 'std::string', \
     'Backend' : 'paddle::experimental::Backend', 'DataLayout' : 'paddle::experimental::DataLayout', 'DataType' : 'paddle::experimental::DataType', \
     'int64[]' : 'std::vector<int64_t>', 'int[]' : 'std::vector<int>',
     'Tensor' : 'Tensor',
@@ -212,7 +213,8 @@ def ParseYamlArgs(string):
         default_value = m.group(3).split("=")[1].strip() if len(
             m.group(3).split("=")) > 1 else None
 
-        assert arg_type in yaml_types_mapping.keys()
+        assert arg_type in yaml_types_mapping.keys(
+        ), f"The argument type {arg_type} in yaml config is not supported in yaml_types_mapping."
         arg_type = yaml_types_mapping[arg_type]
 
         arg_name = RemoveSpecialSymbolsInName(arg_name)
@@ -247,7 +249,8 @@ def ParseYamlReturns(string):
         else:
             ret_type = ret.strip()
 
-        assert ret_type in yaml_types_mapping.keys()
+        assert ret_type in yaml_types_mapping.keys(
+        ), f"The return type {ret_type} in yaml config is not supported in yaml_types_mapping."
         ret_type = yaml_types_mapping[ret_type]
 
         assert "Tensor" in ret_type
@@ -1245,7 +1248,7 @@ def GenerateForwardHFile(filepath, forward_function_declaration_str):
             # Node Definition Generation
             definition_declaration_pair = GenerateForwardDefinition(
                 fwd_api_name, bwd_api_name, forward_inputs_position_map,
-                forward_outputs_position_map, forward_attrs_list,
+                forward_outputs_position_map, orig_forward_attrs_list,
                 backward_fwd_input_map, backward_grad_input_map,
                 backward_grad_output_map, backward_attrs_list, optional_inputs,
                 intermediate_outputs)
@@ -1257,7 +1260,7 @@ def GenerateForwardHFile(filepath, forward_function_declaration_str):
             # For python-level API dispatch
             CollectCoreOpsInformation(fwd_api_name, forward_inputs_position_map,
                                       forward_outputs_position_map,
-                                      forward_attrs_list)
+                                      orig_forward_attrs_list)
 
         if len(namespace) > 0:
             forward_definition_str += f"""namespace {namespace} {{

paddle/fluid/eager/auto_code_generator/final_state_generator/python_c_gen.py

@@ -24,7 +24,7 @@
     "long": "CastPyArg2Long",
     "int64_t": "CastPyArg2Long",
     "float": "CastPyArg2Float",
-    "string": "CastPyArg2String",
+    "std::string": "CastPyArg2String",
     "std::vector<bool>": "CastPyArg2Booleans",
     "std::vector<int>": "CastPyArg2Ints",
     "std::vector<long>": "CastPyArg2Longs",

paddle/fluid/framework/data_device_transform.cc

@@ -34,6 +34,14 @@ void TransDataDevice(const Tensor &in, const platform::Place &dst_place,
     return;
   }
 
+  // NOTE(hqp): Special case for CPU->MLU, avoid stream sync.
+  if (platform::is_cpu_place(in.place()) && platform::is_mlu_place(dst_place)) {
+    paddle::framework::TensorCopy(
+        in, dst_place, *platform::DeviceContextPool::Instance().Get(dst_place),
+        out);
+    return;
+  }
+
   // NOTE(yy): TransDataDevice should wait for computation of input.
   if (!platform::is_cuda_pinned_place(in.place())) {
     platform::DeviceContextPool::Instance().Get(in.place())->Wait();

paddle/fluid/framework/executor.cc

@@ -174,10 +174,11 @@ void Executor::Run(const ProgramDesc& pdesc, Scope* scope, int block_id,
                    bool force_disable_gc, bool keep_kid_scopes) {
   platform::RecordBlock b(block_id);
   if (FLAGS_use_mkldnn) EnableMKLDNN(pdesc);
+  auto ctx = Prepare(pdesc, block_id, skip_ref_cnt_vars, force_disable_gc);
 #ifdef PADDLE_WITH_MKLDNN
   platform::AttachPointerHashToMKLDNNKey(this, place_);
+  platform::RegisterModelLayout(ctx->ops_, place_);
 #endif
-  auto ctx = Prepare(pdesc, block_id, skip_ref_cnt_vars, force_disable_gc);
   RunPreparedContext(ctx.get(), scope, create_local_scope, create_vars,
                      keep_kid_scopes);
 }

paddle/fluid/framework/fleet/ps_gpu_wrapper.cc

@@ -148,7 +148,7 @@ void PSGPUWrapper::PreBuildTask(std::shared_ptr<HeterContext> gpu_task) {
       t.join();
     }
     timeline.Pause();
-    VLOG(1) << "GpuPs build task cost " << timeline.ElapsedSec() << " seconds.";
+    VLOG(0) << "GpuPs build task cost " << timeline.ElapsedSec() << " seconds.";
   } else {
     CHECK(data_set_name.find("MultiSlotDataset") != std::string::npos);
     VLOG(0) << "ps_gpu_wrapper use MultiSlotDataset";
@@ -182,7 +182,7 @@ void PSGPUWrapper::PreBuildTask(std::shared_ptr<HeterContext> gpu_task) {
       t.join();
     }
     timeline.Pause();
-    VLOG(1) << "GpuPs build task cost " << timeline.ElapsedSec() << " seconds.";
+    VLOG(0) << "GpuPs build task cost " << timeline.ElapsedSec() << " seconds.";
   }
 
   timeline.Start();
@@ -300,7 +300,7 @@ void PSGPUWrapper::BuildPull(std::shared_ptr<HeterContext> gpu_task) {
     int32_t cnt = 0;
     while (true) {
       auto tt = fleet_ptr->pslib_ptr_->_worker_ptr->pull_sparse_ptr(
-          reinterpret_cast<char**>(local_ptr[i].data()), this->table_id_,
+          i, reinterpret_cast<char**>(local_ptr[i].data()), this->table_id_,
           local_keys[i].data(), key_size);
       bool flag = true;
 
@@ -378,8 +378,8 @@ void PSGPUWrapper::BuildPull(std::shared_ptr<HeterContext> gpu_task) {
     int32_t cnt = 0;
     while (true) {
       auto tt = fleet_ptr->pslib_ptr_->_worker_ptr->pull_sparse_ptr(
-          reinterpret_cast<char**>(local_dim_ptr[i][j].data()), this->table_id_,
-          local_dim_keys[i][j].data(), key_size);
+          i, reinterpret_cast<char**>(local_dim_ptr[i][j].data()),
+          this->table_id_, local_dim_keys[i][j].data(), key_size);
       bool flag = true;
 
       tt.wait();
@@ -431,7 +431,7 @@ void PSGPUWrapper::BuildPull(std::shared_ptr<HeterContext> gpu_task) {
     t.join();
   }
   timeline.Pause();
-  VLOG(1) << "pull sparse from CpuPS into GpuPS cost " << timeline.ElapsedSec()
+  VLOG(0) << "pull sparse from CpuPS into GpuPS cost " << timeline.ElapsedSec()
           << " seconds.";
   if (multi_node_) {
     auto gloo_wrapper = paddle::framework::GlooWrapper::GetInstance();
@@ -603,7 +603,7 @@ void PSGPUWrapper::BuildPull(std::shared_ptr<HeterContext> gpu_task) {
     t.join();
   }
   timeline.Pause();
-  VLOG(1) << "GpuPs prepare for build hbm cost " << timeline.ElapsedSec()
+  VLOG(0) << "GpuPs prepare for build hbm cost " << timeline.ElapsedSec()
           << " seconds.";
 }
 
@@ -746,7 +746,7 @@ void PSGPUWrapper::BeginPass() {
         "[BeginPass] after build_task, current task is not null."));
   }
 
-  VLOG(1) << "BeginPass end, cost time: " << timer.ElapsedSec() << "s";
+  VLOG(0) << "BeginPass end, cost time: " << timer.ElapsedSec() << "s";
 }
 
 void PSGPUWrapper::EndPass() {
@@ -769,7 +769,7 @@ void PSGPUWrapper::EndPass() {
   current_task_ = nullptr;
   gpu_free_channel_->Put(current_task_);
   timer.Pause();
-  VLOG(1) << "EndPass end, cost time: " << timer.ElapsedSec() << "s";
+  VLOG(0) << "EndPass end, cost time: " << timer.ElapsedSec() << "s";
 }
 
 void PSGPUWrapper::PullSparse(const paddle::platform::Place& place,

paddle/fluid/framework/ir/graph.cc

@@ -95,6 +95,7 @@ std::map<std::string, std::vector<ir::Node *>> Graph::InitFromBlock(
   std::unordered_map<std::string, std::pair<VarDesc *, int>>
       name_to_desc_block_id;
 
+  block_id_ = block.ID();
   const BlockDesc *block_var_visible = &block;
   while (block_var_visible != nullptr) {
     for (auto *var : block_var_visible->AllVars()) {

paddle/fluid/framework/ir/graph.h

@@ -230,6 +230,7 @@ class Graph {
     auto *x =
         AddNode(new ir::Node(var_desc, block_id == -1 ? block_id_ : block_id));
     x->SetId(num_node_created_++);
+    x->SetGraphId(block_id_);
     return x;
   }
 
@@ -245,6 +246,7 @@ class Graph {
                      "The OpDesc used to create operator node is null."));
     auto *x = AddNode(new ir::Node(op_desc));
     x->SetId(num_node_created_++);
+    x->SetGraphId(block_id_);
     return x;
   }
 
@@ -263,6 +265,7 @@ class Graph {
         num_node_created_);
     auto *x = AddNode(new ir::Node(name, ir::Node::Type::kVariable, block_id_));
     x->SetId(num_node_created_++);
+    x->SetGraphId(block_id_);
     return x;
   }
 
@@ -276,6 +279,7 @@ class Graph {
     }
     auto *x = AddNode(new ir::Node(name, type, block_id_));
     x->SetId(num_node_created_++);
+    x->SetGraphId(block_id_);
     return x;
   }
 

paddle/fluid/framework/ir/graph_pattern_detector.cc

@@ -2052,18 +2052,19 @@ PDNode *patterns::Pool::operator()() {
   return output_var;
 }
 
-PDNode *patterns::ElementwiseAdd::operator()(PDNode *x_var, PDNode *y_var) {
-  auto elementwise_add_op = pattern->NewNode(elementwise_add_op_repr())
-                                ->assert_is_op("elementwise_add");
-
-  x_var->AsInput()->assert_is_op_input("elementwise_add", "X");
-  y_var->AsInput()->assert_is_op_input("elementwise_add", "Y");
-  auto out_var = pattern->NewNode(elementwise_add_out_repr())
+PDNode *patterns::Elementwise::operator()(PDNode *x_var, PDNode *y_var,
+                                          const std::string elementwise_type) {
+  auto elementwise_op =
+      pattern->NewNode(elementwise_op_repr())->assert_is_op(elementwise_type);
+
+  x_var->AsInput()->assert_is_op_input(elementwise_type, "X");
+  y_var->AsInput()->assert_is_op_input(elementwise_type, "Y");
+  auto out_var = pattern->NewNode(elementwise_out_repr())
                      ->AsOutput()
-                     ->assert_is_op_output("elementwise_add", "Out");
+                     ->assert_is_op_output(elementwise_type, "Out");
 
-  elementwise_add_op->LinksFrom({x_var, y_var});
-  elementwise_add_op->LinksTo({out_var});
+  elementwise_op->LinksFrom({x_var, y_var});
+  elementwise_op->LinksTo({out_var});
 
   return out_var;
 }

paddle/fluid/framework/ir/graph_pattern_detector.h

@@ -1016,20 +1016,20 @@ struct Pool : public PatternBase {
   PATTERN_DECL_NODE(pool_output);
 };
 
-// ElementwiseAdd used in residual connections.
-// y_var is used and convolution output.
-// The operator is removed, when residual
-// connection fusion is on.
-struct ElementwiseAdd : public PatternBase {
-  ElementwiseAdd(PDPattern* pattern, const std::string& name_scope)
-      : PatternBase(pattern, name_scope, "elementwise_add") {}
-
-  PDNode* operator()(PDNode* x_var, PDNode* y_var);
-
-  PATTERN_DECL_NODE(elementwise_add_op);
-  PATTERN_DECL_NODE(elementwise_add_x);
-  PATTERN_DECL_NODE(elementwise_add_y);
-  PATTERN_DECL_NODE(elementwise_add_out);
+// Elementwise ops
+// Forward pass for element-wise operators (add, mul)
+// elementwise_mul_out is the result of the operator
+struct Elementwise : public PatternBase {
+  Elementwise(PDPattern* pattern, const std::string& name_scope)
+      : PatternBase(pattern, name_scope, "elementwise") {}
+
+  PDNode* operator()(PDNode* x_var, PDNode* y_var,
+                     const std::string elementwise_type);
+
+  PATTERN_DECL_NODE(elementwise_op);
+  PATTERN_DECL_NODE(elementwise_x);
+  PATTERN_DECL_NODE(elementwise_y);
+  PATTERN_DECL_NODE(elementwise_out);
 };
 
 // Transpose op