【Paddle.Fleet】Support local save sparse param (#30175)

* add save tensor support
Co-authored-by: NseiriosPlus <tangwei12@baidu.com>

paddle/fluid/distributed/fleet.cc

@@ -459,6 +459,16 @@ void FleetWrapper::SaveModelOneTable(const uint64_t table_id,
   }
 }
 
+void FleetWrapper::RecvAndSaveTable(const uint64_t table_id,
+                                    const std::string& path) {
+  auto* communicator = Communicator::GetInstance();
+  auto ret = communicator->_worker_ptr->recv_and_save_table(table_id, path);
+  if (ret != 0) {
+    LOG(ERROR) << "save model of table id: " << table_id
+               << ", to path: " << path << " failed";
+  }
+}
+
 void FleetWrapper::PrintTableStat(const uint64_t table_id) {
   auto* communicator = Communicator::GetInstance();
   auto ret = communicator->_worker_ptr->print_table_stat(table_id);

paddle/fluid/distributed/fleet.h

@@ -198,6 +198,10 @@ class FleetWrapper {
   // mode = 1, save delta feature, which means save diff
   void SaveModelOneTable(const uint64_t table_id, const std::string& path,
                          const int mode);
+
+  // recv table from server and save it in LodTensor
+  void RecvAndSaveTable(const uint64_t table_id, const std::string& path);
+
   // clear all models, release their memory
   void ClearModel();
   // clear one table

paddle/fluid/distributed/service/brpc_ps_client.cc

@@ -14,6 +14,7 @@
 
 #include <algorithm>
 #include <memory>
+#include <sstream>
 #include <string>
 #include <vector>
 
@@ -21,6 +22,7 @@
 #include "paddle/fluid/distributed/service/brpc_ps_client.h"
 #include "paddle/fluid/distributed/table/table.h"
 #include "paddle/fluid/framework/archive.h"
+#include "paddle/fluid/string/string_helper.h"
 
 const static int max_port = 65535;
 
@@ -55,6 +57,16 @@ DEFINE_int32(pserver_connect_timeout_ms, 10000,
 
 DEFINE_int32(pserver_sparse_merge_thread, 1, "pserver sparse merge thread num");
 
+namespace paddle {
+namespace framework {
+class Scope;
+class Variable;
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace distributed {
 
@@ -903,5 +915,72 @@ std::future<int32_t> BrpcPsClient::push_sparse_raw_gradient_partial(
   return fut;
 }
 
+int32_t BrpcPsClient::recv_and_save_table(const uint64_t table_id,
+                                          const std::string &path) {
+  // get var information
+  std::string var_name = "";
+  int64_t var_num = 0;
+  int64_t var_shape = 0;
+  const auto &worker_param = _config.worker_param().downpour_worker_param();
+  for (size_t i = 0; i < worker_param.downpour_table_param_size(); ++i) {
+    if (worker_param.downpour_table_param(i).table_id() == table_id) {
+      var_name = worker_param.downpour_table_param(i).common().table_name();
+      var_num = worker_param.downpour_table_param(i).accessor().fea_dim();
+      var_shape = worker_param.downpour_table_param(i).accessor().embedx_dim();
+      break;
+    }
+  }
+
+  PADDLE_ENFORCE_NE(
+      var_name, "",
+      platform::errors::InvalidArgument(
+          "Cannot find table id %d to save variables.", table_id));
+
+  std::string var_store = string::Sprintf("%s", path);
+  MkDirRecursively(var_store.c_str());
+
+  // pull sparse from server
+  std::vector<float> save_huge_vec(var_num * var_shape);
+  std::vector<uint64_t> save_key(var_num);
+  std::vector<float *> save_vec;
+  for (size_t i = 0; i < save_key.size(); ++i) {
+    save_key[i] = i;
+    save_vec.push_back(save_huge_vec.data() + i * var_shape);
+  }
+
+  auto status = pull_sparse((float **)save_vec.data(), table_id,
+                            save_key.data(), save_key.size());
+  status.wait();
+
+  // create lod tensor
+  std::shared_ptr<framework::Scope> scope;
+  scope.reset(new framework::Scope());
+  auto place = platform::CPUPlace();
+  platform::DeviceContextPool &pool = platform::DeviceContextPool::Instance();
+  auto &dev_ctx = *pool.Get(place);
+
+  framework::Variable *var = scope->Var(var_name);
+  framework::LoDTensor *var_tensor = var->GetMutable<framework::LoDTensor>();
+
+  std::vector<int64_t> vec_dim = {var_num, var_shape};
+  var_tensor->Resize(framework::make_ddim(vec_dim));
+
+  // copy and save
+  float *tensor_data = var_tensor->mutable_data<float>(place);
+  memcpy(tensor_data, save_huge_vec.data(),
+         var_num * var_shape * sizeof(float));
+
+  std::string file_name = string::Sprintf("%s/%s", var_store, var_name);
+  std::ofstream fout(file_name, std::ios::binary);
+  PADDLE_ENFORCE_EQ(static_cast<bool>(fout), true,
+                    platform::errors::Unavailable(
+                        "Cannot open %s to save variables.", file_name));
+
+  framework::SerializeToStream(fout, *var_tensor, dev_ctx);
+  fout.close();
+
+  return 0;
+}
+
 }  // namespace distributed
 }  // namespace paddle

paddle/fluid/distributed/service/brpc_ps_client.h

@@ -22,6 +22,9 @@
 #include "brpc/controller.h"
 #include "brpc/server.h"
 #include "paddle/fluid/distributed/service/ps_client.h"
+#include "paddle/fluid/framework/lod_tensor.h"
+#include "paddle/fluid/framework/scope.h"
+#include "paddle/fluid/framework/tensor_util.h"
 
 namespace paddle {
 namespace distributed {
@@ -148,6 +151,10 @@ class BrpcPsClient : public PSClient {
   virtual std::future<int32_t> send_client2client_msg(
       int msg_type, int to_client_id, const std::string &msg) override;
 
+  // for local save sparse
+  virtual int32_t recv_and_save_table(const uint64_t table_id,
+                                      const std::string &path);
+
  private:
   virtual int32_t initialize() override;
 

paddle/fluid/distributed/service/ps_client.h

@@ -134,6 +134,11 @@ class PSClient {
   virtual std::future<int32_t> push_global_step(int table_id,
                                                 int64_t *total_send_data,
                                                 void *done) = 0;
+
+  // recv table from server and save it in LodTensor
+  virtual int32_t recv_and_save_table(const uint64_t table_id,
+                                      const std::string &path) = 0;
+
   virtual void finalize_worker() = 0;
   // client to client, 消息发送
   virtual std::future<int32_t> send_client2client_msg(int msg_type,

paddle/fluid/distributed/table/common_sparse_table.cc

@@ -21,6 +21,7 @@
 #include "paddle/fluid/string/printf.h"
 #include "paddle/fluid/string/string_helper.h"
 
+#define PSERVER_SAVE_SUFFIX "_txt"
 namespace paddle {
 namespace distributed {
 
@@ -290,7 +291,8 @@ int32_t CommonSparseTable::save(const std::string& dirname,
   VLOG(0) << "sparse table save: " << dirname << " mode: " << mode;
 
   auto varname = _config.common().table_name();
-  std::string var_store = string::Sprintf("%s/%s", dirname, varname);
+  std::string var_store =
+      string::Sprintf("%s/%s%s", dirname, varname, PSERVER_SAVE_SUFFIX);
   MkDirRecursively(var_store.c_str());
 
   VLOG(3) << "save " << varname << " in dir: " << var_store << " begin";

paddle/fluid/pybind/fleet_py.cc

@@ -58,6 +58,7 @@ void BindDistFleetWrapper(py::module* m) {
       .def("pull_dense_params", &FleetWrapper::PullDenseVarsSync)
       .def("save_all_model", &FleetWrapper::SaveModel)
       .def("save_one_model", &FleetWrapper::SaveModelOneTable)
+      .def("recv_and_save_model", &FleetWrapper::RecvAndSaveTable)
       .def("sparse_table_stat", &FleetWrapper::PrintTableStat)
       .def("stop_server", &FleetWrapper::StopServer)
       .def("stop_worker", &FleetWrapper::FinalizeWorker)

python/paddle/distributed/fleet/base/fleet_base.py

@@ -545,7 +545,7 @@ class Fleet(object):
             executor, dirname, feeded_var_names, target_vars, main_program,
             export_for_deployment)
 
-    def save_persistables(self, executor, dirname, main_program=None, mode=1):
+    def save_persistables(self, executor, dirname, main_program=None, mode=0):
         """
 
         saves all persistable tensors from :code:`main_program` to

python/paddle/distributed/fleet/meta_optimizers/parameter_server_optimizer.py

@@ -64,12 +64,12 @@ class ParameterServerOptimizer(MetaOptimizerBase):
         _main = compiled_config.origin_main_program.clone()
         _startup = compiled_config.origin_startup_program.clone()
 
+        if not compiled_config.is_geo_mode():
             from paddle.fluid.incubate.fleet.parameter_server.ir.public import _add_lr_decay_table_pass
             _add_lr_decay_table_pass(
                 _main, compiled_config,
                 self.user_defined_strategy.a_sync_configs["lr_decay_steps"])
 
-        if not compiled_config.is_geo_mode():
             # for main program
             _main = worker.delete_optimizer_pass(_main, compiled_config)
             _main = worker.distributed_ops_pass(_main, compiled_config)

python/paddle/distributed/fleet/runtime/the_one_ps.py

@@ -851,15 +851,26 @@ class TheOnePSRuntime(RuntimeBase):
 
         return is_valid
 
-    def _save_sparse_params(self, executor, dirname, context, main_program):
+    def _save_sparse_params(self, executor, dirname, context, main_program,
+                            mode):
+        from paddle.fluid.incubate.fleet.parameter_server.ir.public import get_sparse_tablenames
+        distributed_varnames = get_sparse_tablenames(
+            self.compiled_strategy.origin_main_program, True)
         values = []
         for id, names in context.items():
+            if names not in distributed_varnames:
+                # only save sparse param to local
+                self._worker.recv_and_save_model(id, dirname)
+            # save sparse & distributed param on server
+            self._worker.save_one_model(id, dirname, mode)
             values.extend(names)
-            self._worker.save_one_model(id, dirname, 0)
         return values
 
-    def _save_distributed_persistables(self, executor, dirname, main_program,
-                                       mode):
+    def _save_distributed_persistables(self,
+                                       executor,
+                                       dirname,
+                                       main_program,
+                                       mode=0):
 
         denses = self.compiled_strategy.get_the_one_recv_context(
             is_dense=True,
@@ -870,14 +881,14 @@ class TheOnePSRuntime(RuntimeBase):
             split_dense_table=self.role_maker._is_heter_parameter_server_mode,
             use_origin_program=True)
 
-        recv_sparse_varnames = self._save_sparse_params(executor, dirname,
-                                                        sparses, main_program)
+        sparse_varnames = self._save_sparse_params(executor, dirname, sparses,
+                                                   main_program, mode)
 
         recv_dense_varnames = []
         for id, names in denses.items():
             recv_dense_varnames.extend(names)
 
-        saved_varnames = recv_sparse_varnames
+        saved_varnames = sparse_varnames
 
         remaining_vars = list(
             filter(
@@ -925,6 +936,7 @@ class TheOnePSRuntime(RuntimeBase):
                 "in fleet.save_persistables() function, main_program must be as Program type, CompiledProgram is not allowed"
             )
 
+        # Todo(MrChengmo): Save optimizer status
         self._save_distributed_persistables(executor, dirname, main_program,
                                             mode)
 
@@ -971,8 +983,7 @@ class TheOnePSRuntime(RuntimeBase):
 
             program = Program.parse_from_string(program_desc_str)
             program._copy_dist_param_info_from(fluid.default_main_program())
-            self._ps_inference_save_persistables(
-                executor, dirname, program, mode=0)
+            self._ps_inference_save_persistables(executor, dirname, program)
 
     def _save_inference_model(self, *args, **kwargs):
         self._ps_inference_save_inference_model(*args, **kwargs)