#include "tensorflow/core/platform/types.h"
#include "tensorflow/core/public/session.h"
#include "tensorflow/core/public/version.h"

// The implementation below is at the top level instead of the
// brain namespace because we are defining 'extern "C"' functions.
using tensorflow::AttrValueMap;
using tensorflow::DataType;
using tensorflow::ExtendSessionGraphHelper;
using tensorflow::FullTypeDef;
using tensorflow::Graph;

  EXPECT_EQ(out.size(), block_size);
  // Reading at offset file_size + 4 will read the second block (since the read
  // at file_size + 4 = 28 will be aligned to an offset of 16) but will return
  // OutOfRange because the offset is past the end of the 24-byte file.
  Status status = ReadCache(&cache, "", file_size + 4, 4, &out);
  EXPECT_EQ(status.code(), error::OUT_OF_RANGE);
  EXPECT_TRUE(second_block);

      "/job:localhost/replica:0/task:1/device:CPU:0";
  CheckRemoteMatMulExecutesOK(ctx, remote_device_name, local_device_name);

  // Adding a non-existent remote worker to cluster def. This should cause the
  // UpdateServerDef call to fail.
  tensorflow::ClusterDef* cluster_def = server_def.mutable_cluster();
  tensorflow::JobDef* job_def = cluster_def->mutable_job(0);
  int port = tensorflow::testing::PickUnusedPortOrDie();

        *GetTmpDir(),
        ::testing::UnitTest::GetInstance()->current_test_info()->name());
    status_ = TF_NewStatus();
    filesystem_ = new TF_Filesystem;
    filesystem_->plugin_filesystem = nullptr;
    // Because different tests requires different setup for filesystem. We
    // initialize filesystem in each testcase.
  }
  void TearDown() override {
    TF_DeleteStatus(status_);
    if (filesystem_->plugin_filesystem != nullptr)

  bool UseMlir() const { return strcmp(std::get<0>(GetParam()), "mlir") == 0; }
  bool UseFunction() const { return std::get<2>(GetParam()); }
};

TEST_P(CppGradients, TestIdentityNGrad) {
  // This test is interesting because the current implementation of GradientTape
  // would return [0, 1] whereas we use build_default_zeros_grads=false here
  // so we get back [nullptr, 1].

  AbstractTensorHandlePtr x1;
  {

                        const int in_flight_nodes_limit)
      : status_(TF_NewStatus()),
        // If the context's default exector is set to async, re-using that in
        // each thread would cause collectives to deadlock. For consistency we
        // create a new sync executor for every thread.
        //
        // TODO(allenl): We should have an async API that works with the
        // parallel device.

  TF_SetAttrInt(desc, "num_split", 3);
  TF_SetAttrType(desc, "T", TF_INT32);
  // Set device to CPU since there is no version of split for int32 on GPU
  // TODO(iga): Convert all these helpers and tests to use floats because
  // they are usually available on GPUs. After doing this, remove TF_SetDevice
  // call in c_api_function_test.cc
  TF_SetDevice(desc, "/cpu:0");
  *op = TF_FinishOperation(desc, s);

// default as the multiple API calls required add a risk of stranding temporary
// objects.
constexpr char kComposeAppend[] = "compose";

// We can cast `google::cloud::StatusCode` to `TF_Code` because they have the
// same integer values. See
// https://github.com/googleapis/google-cloud-cpp/blob/6c09cbfa0160bc046e5509b4dd2ab4b872648b4a/google/cloud/status.h#L32-L52
static inline void TF_SetStatusFromGCSStatus(

    TF_OperationDescription* desc =
        TF_NewOperation(host_graph_, func_name_, func_node_name_);
    for (auto input : inputs) {
      TF_AddInput(desc, input);
    }
    // Set device to CPU because some ops inside the function might not be
    // available on GPU.
    TF_SetDevice(desc, "/cpu:0");
    *op = TF_FinishOperation(desc, s_);
    ASSERT_EQ(TF_OK, TF_GetCode(s_)) << TF_Message(s_);