Program Listing for File graph_group_sync.h

↰ Return to documentation for file (src/training/graph_group_sync.h)

#pragma once

#include "optimizers/quantizer.h"
#include "training/graph_group.h"

namespace marian {

class SyncGraphGroup : public GraphGroup {
  using Base = GraphGroup;
  const double delay_{1.}; // optimizer-delay parameter. Fractional means to use a fraction of whatever the MB size is

  // @TODO: instead, create an array of ExponentialSmoothing objects, and don't use ExponentialSmoothing as a base class
  std::vector<Ptr<TensorAllocator>> paramsAllocs_; // [deviceIndex] we must hold a reference to the memory until this class dies

  // model quantizer
  std::vector<Ptr<ModelQuantizer>> quantizers_;

  // state for update()
  bool first_{ true };                           // gets interpreted and cleared by update()
  std::vector<Ptr<data::Batch>> pendingBatches_; // in case of dynamic MB-size scaling, we temporarly buffer up batches across update() calls until enough
  double updateMultiplier_{1};                  // multiplier not applied in collectStats() (no multiplier if not mini-batch-fit)

  void initialize(const Ptr<data::Batch>& exampleBatch);

  bool tryGetSubBatches(Ptr<data::Batch> newBatch, std::vector<Ptr<data::Batch>>& subBatches, size_t& numReadBatches);
  void update(std::vector<Ptr<data::Batch>> subBatches, size_t numReadBatches);

public:
  SyncGraphGroup(Ptr<Options> config, Ptr<IMPIWrapper> mpi);

  void setScheduler(Ptr<Scheduler> scheduler) override;

  void update(Ptr<data::Batch> batch) override;

  Ptr<data::BatchStats> collectStats(const std::vector<Ptr<Vocab>>&) override;
  void finalize() override;

  // @TODO: consider to make this a virtual as well? Currently it is a template dispatch
};
}  // namespace marian