};

using ExecutorPtr = std::unique_ptr<TFE_Executor, ExecutorDeleter>;

}  // namespace

// Allows a single op at a time to be launched without blocking.
//
// DeviceThread itself is thread-safe, in that StartExecute will block if there
// is a pending execution. Since StartExecute is equivalent to grabbing a lock,
// multiple DeviceThreads should always be accessed in the same order to avoid
// deadlocks.
class DeviceThread {
 public:

// -------------------------  VariableInfo  ------------------------------------
struct TF_VariableInfo {
  TF_VariableInfo() = delete;
  // TF_VariableInfo is constructed here by TensorFlow, and will be passed to
  // plugin as a opaque pointer. Plugin will need to call C APIs below to
  // operate on TF_VariableInfo (such as allocate temp tensor for the `var` held
  // by the underlying tensorflow::VariableInfo.

             : absl::OkStatus();
}

// Converts `ninputs` and `inputs` into `inputs_tensors` and `input_nodes` and
// does various checks while doing so. `input_nodes` will contain the same
// information as input_tensors just in a different structure to make
// following processing easier. TODO(iga): Simplify this nested structure.
Status ProcessInputs(

// span the local and remote workers. If the graph optimization pass is executed
// on both the main function side and the component function side, an error will
// be thrown in the registered graph optimization pass.
TEST(CAPI, DistributedFunctionGraphPassOnlyOnce) {
  // Register graph pass that will raise error if called more than once.
  tensorflow::optimization_registration::OptimizationPassRegistration

  // The resulting TensorHandle owns an opaque pointer to "device memory", which
  // for a ParallelDevice is really a ParallelTensor. When the TensorHandle is
  // deleted, it will call ParallelTensorDeallocator to free the struct.
  ParallelTensor* t_released = t.release();
  TFE_CustomDeviceTensorHandleMethods handle_methods;
  handle_methods.num_dims = &ParallelTensorNumDims;

#define ASSERT_TF_OK(x) ASSERT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)
#define EXPECT_TF_OK(x) EXPECT_EQ(TF_OK, TF_GetCode(x)) << TF_Message(x)

static const char* content = "abcdefghijklmnopqrstuvwxyz1234567890";
// We will work with content_view instead of content.
static const absl::string_view content_view = content;

namespace gcs = google::cloud::storage;

static std::string InitializeTmpDir() {

#include "tensorflow/c/tf_file_statistics.h"
#include "tensorflow/c/tf_status.h"

// Implementation of a filesystem for POSIX environments.
// This filesystem will support `file://` and empty (local) URI schemes.

static void* plugin_memory_allocate(size_t size) { return calloc(1, size); }
static void plugin_memory_free(void* ptr) { free(ptr); }

struct TF_AttrBuilder : public tensorflow::AttrBuilder {
  using tensorflow::AttrBuilder::AttrBuilder;
  // The string buffers to make sure that any `attr_name` we pass into
  // `builder->Set()` will outlive the subsequent
  // `TF_AttrBuilderCheckCanRunOnDevice()` call(s) on the same `builder`.
  std::set<std::string> attr_names;
};

TF_AttrBuilder* TF_NewAttrBuilder(const char* op_name) {

  TF_EXPECT_OK(ReadCache(&cache, "", 0, block_size, &out));
  EXPECT_TRUE(first_block);
  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);

  /* Want to test simple MatMul example:
    [[0,0],    *   [[0,0],    =   [[0,0],
     [0,0]]         [0,0]]         [0,0]]
  */

  // Build an abstract input tensor.
  int64_t dims[] = {2, 2};  // Matrices will be 2 x 2
  int num_dims = sizeof(dims) / sizeof(dims[0]);

  float vals[] = {0.0f, 0.0f, 0.0f, 0.0f};
  TFE_Context* eager_ctx = TF_ExecutionContextGetTFEContext(ctx, status.get());
  TFE_TensorHandle* t =