Method offload

oneflow/core/framework/tensor.cpp:167–191

  }

Maybe<void> LocalTensor::offload() {
  TENSOR_OFFLOAD_CHECK(false, "offloaded");

  // Offload to cpu mem with a cpu tensor implantation.
  int64_t device_id = JUST(this->device())->device_id();
  std::shared_ptr<Tensor> cuda_tensor = shared_from_this();
  auto offloaded_tensor =
      JUST(functional::Copy(cuda_tensor, "cpu", device_id, /*pin_memory=*/JUST(is_pinned())));
  JUST(vm::CurrentRankSync());

  const auto& detached_tensor =
      std::dynamic_pointer_cast<LocalTensor>(JUST(offloaded_tensor->detach()));
  CHECK_NOTNULL_OR_RETURN(detached_tensor) << " detached_tensor must be a local tensor.";
  offloaded_impl_ = detached_tensor->impl_;

  // Release cuda memory, but the meta data is valid.
  auto eager_blob_obj = JUST(JUST(impl_->mut_eager_local_tensor_impl())->eager_blob_object());
  JUST(eager_blob_obj->DeallocateBlobDataPtr());

  auto* vm = JUST(SingletonMaybe<VirtualMachine>());
  JUST(vm->ShrinkAllMem());

  is_offloaded_ = true;
  return Maybe<void>::Ok();
}

Maybe<void> LocalTensor::load() {
  TENSOR_OFFLOAD_CHECK(true, "loaded");