[fleet executor] add comm init for dist model inf (#39012)

paddle/fluid/distributed/fleet_executor/dist_model.cc

@@ -15,7 +15,9 @@
 #include <glog/logging.h>
 
 #include "paddle/fluid/distributed/fleet_executor/dist_model.h"
+#include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/naive_executor.h"
+#include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/tensor.h"
@@ -37,24 +39,173 @@ bool IsPersistable(const framework::VarDesc *var) {
 
 bool DistModel::Init() {
   /* TODO(fleet exe dev): implement this funct */
-  place_ = paddle::platform::CUDAPlace(config_.device_id);
+  bool init_method = (!config_.model_dir.empty() || config_.program_desc);
+  PADDLE_ENFORCE_EQ(init_method, true,
+                    platform::errors::InvalidArgument(
+                        "One of model dir or program desc must be provided to "
+                        "dist model inference."));
+  if (config_.program_desc) {
+    PADDLE_ENFORCE_NOT_NULL(
+        config_.scope, platform::errors::InvalidArgument(
+                           "Scope must be provided to dist model inference if "
+                           "program desc has been provided."));
+  }
+  if (!PreparePlace()) {
+    return false;
+  }
+  if (!config_.program_desc) {
+    if (config_.scope) {
+      LOG(WARNING) << "The provided scope will be ignored if model dir has "
+                      "also been provided.";
+    }
     if (!PrepareScope()) {
       return false;
     }
     if (!PrepareProgram()) {
       return false;
     }
+  } else {
+    program_.reset(config_.program_desc);
+    scope_.reset(config_.scope);
+  }
   if (!CommInit()) {
     return false;
   }
   return true;
 }
 
+bool DistModel::PreparePlace() {
+  if (config_.place == "GPU") {
+    place_ = paddle::platform::CUDAPlace(config_.device_id);
+  } else if (config_.place == "CPU") {
+    place_ = paddle::platform::CPUPlace();
+  } else {
+    PADDLE_THROW(platform::errors::InvalidArgument(
+        "Place must be choosen from GPU or CPU, but got %s.", config_.place));
+  }
+  return true;
+}
+
 bool DistModel::CommInit() {
-  // TODO(fleet executor): init the comm
+  // NOTE (Yuang Liu): The peer endpoints will be obtained with the assumption
+  // that mp part is always on inner side and pp part is always on outer side.
+  // TODO(fleet exe dev): The peer endpoints could be configured by users.
+  PADDLE_ENFORCE_EQ(
+      config_.pp_degree * config_.mp_degree, config_.nranks,
+      platform::errors::InvalidArgument(
+          "The mp_degree multiplies pp_degree is not equal with nranks"));
+  std::unique_ptr<framework::ProgramDesc> comm_init_program(
+      new framework::ProgramDesc());
+  framework::BlockDesc *comm_init_block = comm_init_program->MutableBlock(0);
+  if (config_.mp_degree > 1) {
+    PADDLE_ENFORCE_GE(
+        config_.mp_ring_id, 0,
+        platform::errors::InvalidArgument(
+            "mp ring id must be provided for inference under mp."));
+    VLOG(3) << "Init comm group for mp.";
+    std::vector<std::string> peer_endpoints;
+    for (int64_t
+             idx = (config_.local_rank / config_.mp_degree) * config_.mp_degree,
+             i = 0;
+         i < config_.mp_degree; ++idx, ++i) {
+      if (config_.trainer_endpoints[idx] == config_.current_endpoint) {
+        continue;
+      }
+      peer_endpoints.emplace_back(config_.trainer_endpoints[idx]);
+    }
+    // get nranks in a mp group and inner group rank for local rank
+    int64_t mp_group_nranks = config_.nranks / config_.pp_degree;
+    int64_t mp_group_rank = config_.local_rank % config_.mp_degree;
+    InsertCommOp("mp_comm_id", mp_group_nranks, mp_group_rank, peer_endpoints,
+                 comm_init_block, config_.mp_ring_id);
+  }
+  if (config_.pp_degree) {
+    // NOTE: the last pp stage doesn't need init pp comm
+    VLOG(3) << "Init comm group for pp.";
+    if (config_.local_rank - config_.mp_degree >= 0) {
+      PADDLE_ENFORCE_EQ(config_.pp_upstream_ring_id >= 0, true,
+                        platform::errors::InvalidArgument(
+                            "pp upstream ring id must be provided for "
+                            "non-first pp stage if inference under pp."));
+      // not the first pp stage, has upstream
+      std::vector<std::string> upstream_peer_endpoints;
+      upstream_peer_endpoints.emplace_back(
+          config_.trainer_endpoints[config_.local_rank - config_.mp_degree]);
+      InsertCommOp("pp_upstream_comm_id", 2, 1, upstream_peer_endpoints,
+                   comm_init_block, config_.pp_upstream_ring_id);
+    }
+
+    if (config_.local_rank + config_.mp_degree < config_.nranks) {
+      PADDLE_ENFORCE_EQ(config_.pp_downstream_ring_id >= 0, true,
+                        platform::errors::InvalidArgument(
+                            "pp downstream ring id must be provided for "
+                            "non-last pp stage if inference under pp."));
+      // not the last pp stage, has downstream
+      std::vector<std::string> downstream_peer_endpoints;
+      downstream_peer_endpoints.emplace_back(
+          config_.trainer_endpoints[config_.local_rank + config_.mp_degree]);
+      InsertCommOp("pp_downstream_comm_id", 2, 0, downstream_peer_endpoints,
+                   comm_init_block, config_.pp_downstream_ring_id);
+    }
+  }
+  framework::NaiveExecutor e(place_);
+  e.CreateVariables(*comm_init_program, 0, true, scope_.get());
+  e.Prepare(scope_.get(), *comm_init_program, 0, false);
+  e.Run();
+  VLOG(3) << "Comm init successful.";
   return true;
 }
 
+void DistModel::InsertCommOp(std::string tmp_var_name, int nranks, int rank,
+                             const std::vector<std::string> &peer_endpoints,
+                             framework::BlockDesc *block, int ring_id) {
+  /*
+   * tmp_var_name: the var name for var comm_id
+   * nranks: number of total ranks
+   * rank: the rank of local rank in the comm group
+   * peer_endpoints: peer's endpoints
+   * block: the block where to insert the comm ops
+   * ring_id: the ring_id to be inited
+   */
+  std::string &endpoint = config_.current_endpoint;
+  std::stringstream ss;
+  ss << "Init comm with tmp var: " << tmp_var_name
+     << ". The ring id is: " << ring_id << ". The group has: " << nranks
+     << " ranks. Current rank in the group is: " << rank
+     << ". The endpoint is: " << endpoint << ". Peer endpoints are: ";
+  for (auto ep : peer_endpoints) {
+    ss << ep << ", ";
+  }
+  VLOG(3) << ss.str();
+  if (config_.place == "GPU") {
+    framework::VarDesc *new_var = block->Var(tmp_var_name);
+    new_var->SetType(framework::proto::VarType::RAW);
+    new_var->SetPersistable(true);
+    framework::OpDesc *gen_nccl_id_op = block->AppendOp();
+    gen_nccl_id_op->SetType("c_gen_nccl_id");
+    gen_nccl_id_op->SetOutput("Out", {tmp_var_name});
+    gen_nccl_id_op->SetAttr("rank", rank);
+    gen_nccl_id_op->SetAttr("endpoint", config_.current_endpoint);
+    gen_nccl_id_op->SetAttr("other_endpoints", peer_endpoints);
+    gen_nccl_id_op->SetAttr("ring_id", ring_id);
+    gen_nccl_id_op->SetAttr("op_role",
+                            static_cast<int>(framework::OpRole::kForward));
+    gen_nccl_id_op->CheckAttrs();
+    framework::OpDesc *comm_init_op = block->AppendOp();
+    comm_init_op->SetType("c_comm_init");
+    comm_init_op->SetInput("X", {tmp_var_name});
+    comm_init_op->SetAttr("rank", rank);
+    comm_init_op->SetAttr("nranks", nranks);
+    comm_init_op->SetAttr("ring_id", ring_id);
+    comm_init_op->SetAttr("op_role",
+                          static_cast<int>(framework::OpRole::kForward));
+    comm_init_op->CheckAttrs();
+  } else {
+    LOG(WARNING) << "DistModelInf doesn't init comm.";
+    // TODO(fleet exe dev): comm init for more devices
+  }
+}
+
 bool DistModel::PrepareScope() {
   scope_.reset(new framework::Scope());
   return true;
@@ -119,6 +270,8 @@ bool DistModel::LoadParameters() {
       new_var->SetLoDLevel(var->GetLoDLevel());
       new_var->SetPersistable(true);
       params.push_back(new_var->Name());
+      // NOTE: if the params are stored in different files, 'load' op should be
+      // added here
     }
   }
 

paddle/fluid/distributed/fleet_executor/dist_model.h

@@ -23,22 +23,30 @@
 #include "paddle/fluid/platform/place.h"
 
 namespace paddle {
+
 namespace framework {
 class ProgramDesc;
 class Scope;
+class BlockDesc;
 }
 
 namespace distributed {
 
 struct DistModelConfig {
   std::string model_dir{};
+  framework::ProgramDesc* program_desc{nullptr};
+  framework::Scope* scope{nullptr};
+  std::string place{};
+  int64_t device_id{0};
   std::vector<std::string> trainer_endpoints{};
   std::string current_endpoint{};
   int64_t nranks{1};
   int64_t local_rank{0};
-  int64_t device_id{0};
   int64_t mp_degree{1};
   int64_t pp_degree{1};
+  int64_t mp_ring_id{-1};
+  int64_t pp_upstream_ring_id{-1};
+  int64_t pp_downstream_ring_id{-1};
 };
 
 class DistModel {
@@ -56,12 +64,16 @@ class DistModel {
   bool PrepareProgram();
   bool LoadProgram();
   bool LoadParameters();
+  bool PreparePlace();
   bool CommInit();
+  void InsertCommOp(std::string tmp_var_name, int nranks, int rank,
+                    const std::vector<std::string>& peer_endpoints,
+                    framework::BlockDesc* block, int ring_id);
 
   DistModelConfig config_;
   FleetExecutorDesc executor_desc_;
-  platform::Place place_;
   std::shared_ptr<framework::Scope> scope_;
+  paddle::platform::Place place_;
   std::shared_ptr<framework::ProgramDesc> program_;
 };
 

paddle/fluid/pybind/bind_fleet_executor.cc

@@ -58,13 +58,21 @@ void BindFleetExecutor(py::module* m) {
   py::class_<DistModelConfig>(*m, "DistModelConfig")
       .def(py::init<>())
       .def_readwrite("model_dir", &DistModelConfig::model_dir)
+      .def_readwrite("program_desc", &DistModelConfig::program_desc)
+      .def_readwrite("scope", &DistModelConfig::scope)
+      .def_readwrite("place", &DistModelConfig::place)
+      .def_readwrite("device_id", &DistModelConfig::device_id)
       .def_readwrite("trainer_endpoints", &DistModelConfig::trainer_endpoints)
       .def_readwrite("current_endpoint", &DistModelConfig::current_endpoint)
       .def_readwrite("nranks", &DistModelConfig::nranks)
       .def_readwrite("local_rank", &DistModelConfig::local_rank)
-      .def_readwrite("device_id", &DistModelConfig::device_id)
       .def_readwrite("mp_degree", &DistModelConfig::mp_degree)
-      .def_readwrite("pp_degree", &DistModelConfig::pp_degree);
+      .def_readwrite("pp_degree", &DistModelConfig::pp_degree)
+      .def_readwrite("mp_ring_id", &DistModelConfig::mp_ring_id)
+      .def_readwrite("pp_upstream_ring_id",
+                     &DistModelConfig::pp_upstream_ring_id)
+      .def_readwrite("pp_downstream_ring_id",
+                     &DistModelConfig::pp_downstream_ring_id);
 
   py::class_<DistModel>(*m, "DistModel")
       .def(py::init<const DistModelConfig&>())