Merge remote-tracking branch 'upstream/develop' into winters006

PaddlePaddle · Sep 30, 2023 · 5360925 · 5360925
2 parents 1635319 + a46df40
commit 5360925
Show file tree

Hide file tree

Showing 51 changed files with 1,559 additions and 744 deletions.
diff --git a/paddle/fluid/distributed/auto_parallel/spmd_rules/rules.h b/paddle/fluid/distributed/auto_parallel/spmd_rules/rules.h
@@ -17,7 +17,6 @@
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/common.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/cross_entropy_with_softmax_spmd_rule.h"
 #include "paddle/fluid/distributed/auto_parallel/spmd_rules/replicated_spmd_rule.h"
-#include "paddle/fluid/distributed/auto_parallel/spmd_rules/softmax_spmd_rule.h"
 
 // TODO(ljz) Automatic this process in cmake file.
 namespace paddle {
@@ -27,10 +26,6 @@ namespace auto_parallel {
 // replicated rule
 REGISTER_SPMD_RULE(replicated, ReplicatedSPMDRule);
 
-// softmax rule
-REGISTER_SPMD_RULE(softmax, SoftmaxSPMDRule);
-REGISTER_SPMD_RULE(log_softmax, SoftmaxSPMDRule);
-
 // cross_entropy_with_softmax
 REGISTER_SPMD_RULE(cross_entropy_with_softmax, CrossEntropyWithSoftmaxSPMDRule);
 REGISTER_SPMD_RULE(softmax_with_cross_entropy, CrossEntropyWithSoftmaxSPMDRule);

diff --git a/paddle/fluid/distributed/auto_parallel/spmd_rules/softmax_spmd_rule.cc b/paddle/fluid/distributed/auto_parallel/spmd_rules/softmax_spmd_rule.cc
diff --git a/paddle/fluid/eager/to_static/run_program_op_func.h b/paddle/fluid/eager/to_static/run_program_op_func.h
@@ -291,9 +291,7 @@ inline void newir_run_program_ad_func(
     // Set TensorWrappers
     grad_node->SetFwdX(filter_x);
 
-    auto filter_params =
-        newir_filter_unused_input_var_in_backward(params, "bp", attrs);
-    grad_node->SetFwdParams(filter_params);
+    grad_node->SetFwdParams(params);
 
     grad_node->SetStepScope(step_scope);  // just for set useable.
 

diff --git a/paddle/fluid/eager/to_static/run_program_op_node.h b/paddle/fluid/eager/to_static/run_program_op_node.h
@@ -175,22 +175,23 @@ static void ShareTensorsIntoScopeWithName(
 }
 
 static auto GetNameFromValue(const ::pir::Block *block,
-                             const std::vector<::pir::Value> &values) {
+                             const std::vector<::pir::Value> &values,
+                             bool is_input) {
   // we use name here, later value is used directly.
   std::unordered_map<::pir::Value, std::string> value2name;
   for (auto *op : *block) {
     std::string name;
-    if (op->name() == "pd_op.data") {
+    if (is_input && op->name() == "pd_op.data") {
       name =
           op->attributes().at("name").dyn_cast<pir::StrAttribute>().AsString();
       value2name[op->results()[0].Value::impl()] = name;
-    } else if (op->name() == "builtin.set_parameter") {
+    } else if (!is_input && op->name() == "builtin.set_parameter") {
       name = op->attributes()
                  .at("parameter_name")
                  .dyn_cast<pir::StrAttribute>()
                  .AsString();
       value2name[op->operand(0).source()] = name;
-    } else if (op->name() == "builtin.get_parameter") {
+    } else if (is_input && op->name() == "builtin.get_parameter") {
       name = op->attributes()
                  .at("parameter_name")
                  .dyn_cast<pir::StrAttribute>()
@@ -256,7 +257,7 @@ static void ShareTensorsIntoScopeByValue(
     const std::vector<Tensor> &tensors,
     const std::vector<::pir::Value> &values,
     paddle::framework::Scope *scope) {
-  auto names = GetNameFromValue(block, values);
+  auto names = GetNameFromValue(block, values, true);
   if (VLOG_IS_ON(4)) {
     for (auto &s : names) {
       VLOG(4) << "ShareTensorIntoScopeByValue name: " << s;
@@ -270,7 +271,7 @@ static void ShareTensorsFromScopeByValue(
     const std::vector<Tensor *> &tensors,
     const std::vector<::pir::Value> &values,
     paddle::framework::Scope *scope) {
-  auto names = GetNameFromValue(block, values);
+  auto names = GetNameFromValue(block, values, false);
   for (size_t i = 0; i < tensors.size(); ++i) {
     auto &name = names[i];
     auto &value = values[i];
@@ -506,15 +507,16 @@ inline void NewIRRunProgramAPI(
 
     // update interpretercore skip_gc_var
     auto skip_names =
-        details::GetNameFromValue(forward_global_block, middle_values);
+        details::GetNameFromValue(forward_global_block, middle_values, false);
     auto skip_names_set =
         std::set<std::string>(skip_names.begin(), skip_names.end());
-    skip_names = details::GetNameFromValue(forward_global_block, output_values);
+    skip_names =
+        details::GetNameFromValue(forward_global_block, output_values, false);
     skip_names_set.insert(skip_names.begin(), skip_names.end());
     auto no_need_buffer_values = PADDLE_GET_CONST(std::vector<::pir::Value>,
                                                   attrs.at("no_need_buffers"));
-    auto no_need_buffer_names =
-        details::GetNameFromValue(forward_global_block, no_need_buffer_values);
+    auto no_need_buffer_names = details::GetNameFromValue(
+        forward_global_block, no_need_buffer_values, false);
     VLOG(4) << "start skip no need buffer vars with name:";
     for (auto &name : no_need_buffer_names) {
       VLOG(4) << "Skip no need buffer vars with name:" << name;
@@ -1053,10 +1055,10 @@ inline void NewIRRunProgramGradAPI(
     // get all eager gc vars
     std::set<std::string> skip_eager_delete_vars;
     auto skip_names =
-        details::GetNameFromValue(backward_global_block, x_grad_values);
+        details::GetNameFromValue(backward_global_block, x_grad_values, false);
     skip_eager_delete_vars.insert(skip_names.begin(), skip_names.end());
     skip_names =
-        details::GetNameFromValue(backward_global_block, p_grad_values);
+        details::GetNameFromValue(backward_global_block, p_grad_values, false);
     skip_eager_delete_vars.insert(skip_names.begin(), skip_names.end());
     interpreter_core->SetSkipGcVars(skip_eager_delete_vars);
     interpretercore_info_cache.UpdateSkipEagerDeleteVars(