Optimiza params sync between CPU and GPU. (#45805)

* enable memory optimize when fp16.

* optimiza params sync between cpu and gpu.

paddle/fluid/inference/analysis/passes/ir_params_sync_among_devices_pass.cc

@@ -21,9 +21,12 @@
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 #include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/tensor_util.h"
+#include "paddle/fluid/inference/api/helper.h"
 #include "paddle/fluid/platform/bfloat16.h"
 #include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/platform/place.h"
 #include "paddle/phi/common/data_type.h"
 
 namespace paddle {
@@ -114,6 +117,28 @@ void IrParamsSyncAmongDevicesPass::CopyParamsToGpu(Argument *argument) {
     reserve_cpu_weights = true;
   }
 
+  int64_t params_total_bytes{0};
+  for (auto *node : paddle::framework::ir::TopologySortOperations(graph)) {
+    if (!node->IsOp()) continue;
+    if (node->Op()->Type() == "feed" || node->Op()->Type() == "fetch") continue;
+    for (auto *var_node : node->inputs) {
+      if (!var_node->Var()->Persistable()) continue;
+      auto var_name = var_node->Var()->Name();
+      auto *var = scope->FindLocalVar(var_name);
+      if (var->IsType<framework::LoDTensor>() ||
+          var->IsType<framework::Tensor>()) {
+        auto *t = var->GetMutable<framework::LoDTensor>();
+        params_total_bytes += t->numel() * experimental::SizeOf(t->dtype());
+      }
+    }
+  }
+
+  {
+    // Alloc memory in pool to store all parameters.
+    framework::Tensor ts;
+    ts.mutable_data(place, params_total_bytes);
+  }
+
   std::unordered_set<std::string> visited;
   for (auto *node : paddle::framework::ir::TopologySortOperations(graph)) {
     if (!node->IsOp()) continue;