for (int i = 0; i < ninputs; ++i) body_outputs[i] = {nullptr, -1};
  const char* name = nullptr;

  for (int i = 0; i < ninputs; ++i) {
    // TODO(skyewm): prefix names with underscore (requires some plumbing)
    if (!CreateInput(inputs[i], cond_graph, StrCat("cond_input", i).c_str(),
                     &cond_inputs[i], status)) {
      break;
    }

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;

                    char* buffer, TF_Status* status) {
  auto posix_file = static_cast<PosixFile*>(file->plugin_file);
  char* dst = buffer;
  int64_t read = 0;

  while (n > 0) {
    // Some platforms, notably macs, throw `EINVAL` if `pread` is asked to read
    // more than fits in a 32-bit integer.
    size_t requested_read_length;
    if (n > INT32_MAX)
      requested_read_length = INT32_MAX;
    else

  TFE_TensorDebugInfo* debug_info = TFE_TensorHandleTensorDebugInfo(h, status);
  CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  ASSERT_EQ(2, TFE_TensorDebugInfoOnDeviceNumDims(debug_info));
  // Shape is the same for CPU tensors.
  EXPECT_EQ(3, TFE_TensorDebugInfoOnDeviceDim(debug_info, 0));
  EXPECT_EQ(2, TFE_TensorDebugInfoOnDeviceDim(debug_info, 1));

  TFE_DeleteTensorDebugInfo(debug_info);
  TFE_DeleteTensorHandle(h);

int TransferFileContents(const char* src, const char* dst, mode_t mode,
                         off_t size) {
  int src_fd = open(src, O_RDONLY);
  if (src_fd < 0) return -1;

  // When creating file, use the same permissions as original
  mode_t open_mode = mode & (S_IRWXU | S_IRWXG | S_IRWXO);

  // O_WRONLY | O_CREAT | O_TRUNC:
  //   Open file for write and if file does not exist, create the file.

      ASSERT_EQ(errors::OK, status_.code()) << status_.message();
      immediate_execution_ctx_.reset(ctx_raw);
    }

    // Computing numerical gradients with TensorFloat-32 is numerically
    // unstable. Some forward pass tests also fail with TensorFloat-32 due to
    // low tolerances
    enable_tensor_float_32_execution(false);
  }

  AbstractContextPtr immediate_execution_ctx_;
  GradientRegistry registry_;

  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback
  // (as in the case of ConstOp kernel creation on GPU, which involves copying a

  EXPECT_EQ(TF_GetCode(status.get()), TF_INVALID_ARGUMENT);
  EXPECT_THAT(TF_Message(status.get()), HasSubstr("assertion failed"));

  // Note that future collectives with the same context do not work at the
  // moment; once canceled, the collective executor requires the program to be
  // restarted / context to be reset.
}

TEST(PARALLEL_DEVICE_LIB, TestDifferentShapes) {

                             serialized_update.size(), status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // The device of var_handle0 is local device which is the same before and
  // after cluster update. Remove resource with valid device should succeed.
  TFE_Op* op = TFE_NewOp(ctx, "DestroyResourceOp", status);
  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // 1) nullptr indicates tensor is compact and row-majored.
  // 2) fill in the strides array as the real case for compact row-major data.
  // Here we choose option 2, since some frameworks didn't handle the strides
  // argument properly.
  dlm_tensor->dl_tensor.strides = stride_arr->data();

  dlm_tensor->dl_tensor.byte_offset =
      0;  // TF doesn't handle the strides and byte_offsets here