Method load

oneflow/core/framework/tensor.cpp:193–212

}

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

  // Load cpu to cuda.
  int64_t device_id = JUST(this->device())->device_id();
  std::shared_ptr<Tensor> cpu_tensor = std::make_shared<LocalTensor>(offloaded_impl_);
  auto loaded_tensor = JUST(functional::Copy(cpu_tensor, "cuda", device_id,
                                             /*pin_memory=*/JUST(cpu_tensor->is_pinned())));
  JUST(vm::CurrentRankSync());
  JUST(set_data(loaded_tensor));

  // Release cpu memory.
  cpu_tensor.reset();
  offloaded_impl_.reset();
  auto* vm = JUST(SingletonMaybe<VirtualMachine>());
  JUST(vm->ShrinkAllMem());

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

std::shared_ptr<Tensor> GlobalTensor::contiguous() const {
  std::shared_ptr<Tensor> tensor = std::const_pointer_cast<Tensor>(shared_from_this());