: 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(

  // If the output is a scalar, then return the scalar output
  if (num_dims_out == 0) {
    outputs[0] = model_out.release();
    return absl::OkStatus();
  }

  // Else, reduce sum the output to get a scalar

  // Will sum all dimensions, so get a Tensor containing [0,...,num_dims_out-1].
  AbstractTensorHandlePtr sum_dims;
  {
    vector<int32_t> vals(num_dims_out);
    int64_t vals_shape[] = {num_dims_out};

// 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/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {
namespace parallel_device {

Variable* Variable::Create(TFE_Context* context, TF_DataType type,

#include "tensorflow/core/platform/test.h"

// NOTE(allenl): These tests currently go through TFE_Execute and so are
// integration testing rather than purely testing the parallel device. They
// correspond fairly well to the implementation, but testing the C++ directly is
// another option.

namespace tensorflow {
namespace parallel_device {

using ::testing::HasSubstr;

TEST(PARALLEL_DEVICE, TestBasicCPU) {

//     we are still warning users): the entries that core TensorFlow expects
//     but plugins didn't provide will be set to `nullptr` values and core
//     TensorFlow will know to not call these on behalf of users;
//   * increased security as plugins will not be able to alter function table
//     after loading up. Thus, malicious plugins can't alter functionality to

      gcs_file->buffer.resize(gcs_file->buffer_size);
      auto read_fill_buffer = gcs_file->read_fn(
          gcs_file->path, gcs_file->buffer_start, gcs_file->buffer_size,
          &(gcs_file->buffer[0]), status);
      gcs_file->buffer_end_is_past_eof =
          (TF_GetCode(status) == TF_OUT_OF_RANGE);
      if (read_fill_buffer >= 0) gcs_file->buffer.resize(read_fill_buffer);
      if (TF_GetCode(status) != TF_OK &&

  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);