Tensor* host_tensor = CreateHostTensor<T>(shape, data);
    Tensor* device_tensor =
        new Tensor(device_allocator_, DataTypeToEnum<T>::v(), shape);
    TF_EXPECT_OK(device_context_->CopyCPUTensorToDeviceSync(
        host_tensor, device_, device_tensor));

    tensors_.push_back(device_tensor);
    return device_tensor;
  }

  // Gets the `output_index`-th output set in the context_

// GPU MIG.
def TransferToDeviceOp: Gpu_Op<"transfer_to_device"> {
  let summary = "Transfer a CPU tensor to device.";

  let description = [{
    Transfer a CPU tensor to device.

    Example:
      %device_tensor = gpurt.transfer_to_device %cpu_tensor
  }];

  let arguments = (ins TFTensorType);
  let results = (outs TFTensorType);
  let assemblyFormat = "operands attr-dict";
}

    Tensor* host_tensor = CreateHostTensor<T>(shape, data);
    Tensor* device_tensor =
        new Tensor(device_allocator_, DataTypeToEnum<T>::v(), shape);
    TF_EXPECT_OK(device_context_->CopyCPUTensorToDeviceSync(
        host_tensor, device_, device_tensor));

    tensors_.push_back(device_tensor);
    return device_tensor;
  }

  // Compiles the op set in the context_ to a PjRtLoadedExecutable

      thread::ThreadPool* thread_pool);

  void CopyCPUTensorToDevice(const Tensor* cpu_tensor, Device* device,
                             Tensor* device_tensor, StatusCallback done,
                             bool sync_dst_compute) const override;
  void CopyDeviceTensorToCPU(const Tensor* device_tensor,
                             absl::string_view tensor_name, Device* device,
                             Tensor* cpu_tensor, StatusCallback done) override;

#include "tensorflow/compiler/tf2xla/type_util.h"

namespace tensorflow {

void XlaHostSendDeviceContext::CopyCPUTensorToDevice(
    const Tensor* cpu_tensor, Device* device, Tensor* device_tensor,
    StatusCallback done, bool sync_dst_compute) const {
  auto status = stream_->Memcpy(device_memory_base_, cpu_tensor->data(),
                                device_memory_base_->size());
  if (!status.ok()) {

#include "tensorflow/compiler/tf2xla/shape_util.h"
#include "tensorflow/compiler/tf2xla/type_util.h"

namespace tensorflow {

void XlaHostRecvDeviceContext::CopyDeviceTensorToCPU(
    const Tensor* device_tensor, StringPiece tensor_name, Device* device,
    Tensor* cpu_tensor, StatusCallback done) {
  DataType dtype = EncodePrimitiveTypeAsDataType(shape_.element_type()).value();
  TensorShape tensor_shape;

        args.device_context = new XlaHostSendDeviceContext(
            stream, device_memory_base, shape, done_event);

        Tensor device_tensor;
        bool is_dead;
        TF_RETURN_IF_ERROR(ctx->rendezvous()->Recv(
            parsed_key, args, &device_tensor, /*is_dead=*/&is_dead));

        return std::move(done_event);
      };
}

absl::StatusOr<xla::ExecutionOutput> RunExecutable(