absl::Span<AbstractTensorHandle* const> inputs,
                             std::vector<AbstractTensorHandle*>* params);

// A callable that takes tensor inputs and returns zero or more tensor outputs.
using Model = std::function<Status(AbstractContext*,
                                   absl::Span<AbstractTensorHandle* const>,
                                   absl::Span<AbstractTensorHandle*>)>;

    EXPECT_EQ(got_n, want_n);
    calls++;
    memset(buffer, 'x', got_n);
    TF_SetStatus(status, TF_OK, "");
    return got_n;
  };
  // If block_size, max_bytes, or both are zero, or want_n is larger than
  // max_bytes the cache is a pass-through.
  tf_gcs_filesystem::RamFileBlockCache cache1(1, 0, 0, fetcher);
  tf_gcs_filesystem::RamFileBlockCache cache2(0, 1, 0, fetcher);

    if (!status->status.ok()) return false;
  }
  return true;
}

// Create an empty tensor of type 'dtype'. 'shape' can be arbitrary, but has to
// result in a zero-sized tensor.
static TF_Tensor* EmptyTensor(TF_DataType dtype,
                              const tensorflow::TensorShape& shape) {
  static char empty;
  int64_t nelems = 1;
  std::vector<int64_t> dims;

typedef struct TF_Thread TF_Thread;

typedef struct TF_ThreadOptions {
  // Thread stack size to use (in bytes), zero implies that the system default
  // will be used.
  size_t stack_size;

  // Guard area size to use near thread stacks to use (in bytes), zero implies
  // that the system default will be used.
  size_t guard_size;

  return result;
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, indicating how
// ParallelTensors wrapped in TFE_TensorHandles should be cleaned up once their
// reference counts drop to zero.
void ParallelTensorDeallocator(void* data) {
  delete reinterpret_cast<ParallelTensor*>(data);
}

// Used as an argument to TFE_NewCustomDeviceTensorHandle, for computing the

//
// `num_return_outputs` must be the number of return outputs added (i.e. the
// result of TF_ImportGraphDefOptionsNumReturnOutputs()).  If
// `num_return_outputs` is non-zero, `return_outputs` must be of length
// `num_return_outputs`. Otherwise it can be null.
TF_CAPI_EXPORT extern void TF_GraphImportGraphDefWithReturnOutputs(
    TF_Graph* graph, const TF_Buffer* graph_def,

    VLOG(1) << "Calling gradient function for '" << trace.op_type << "'";
    std::vector<Gradient*> in_gradients(trace.input_tensor_id.size());
    DCHECK(build_default_zeros_grads || zero_indices.empty());
    if (any_gradient_nonzero) {
      for (const auto i : zero_indices) {
        out_gradients[i] = trace.output_tensor_info[i].ZerosLike();
      }
      Status s;
      s = vspace.CallBackwardFunction(trace.op_type, trace.backward_function,

-		  ? ({ void *__s = (s); __bzero (__s, n); __s; })	      \
-		  : memset (s, c, n)))
-#  endif
-# endif
-
-/* GCC < 3.0 optimizes memset(s, 0, n) but not bzero(s, n).
-   The optimization is broken before EGCS 1.1.
-   GCC 3.0+ has __builtin_bzero as well, but at least till GCC 3.4
-   if it decides to call the library function, it calls memset
-   and not bzero.  */
-# if __GNUC_PREREQ (2, 91)

    unsigned char should_enable);

// Create a serialized tensorflow.ConfigProto proto, where:
//
// a) ConfigProto.optimizer_options.global_jit_level is set to ON_1 if
// `enable_xla_compilation` is non-zero, and OFF otherwise.
// b) ConfigProto.gpu_options.allow_growth is set to `gpu_memory_allow_growth`.
// c) ConfigProto.device_count is set to `num_cpu_devices`.
TF_CAPI_EXPORT extern TF_Buffer* TF_CreateConfig(

def write_action_env_to_bazelrc(var_name, var):
  write_to_bazelrc('build --action_env {}="{}"'.format(var_name, str(var)))


def run_shell(cmd, allow_non_zero=False, stderr=None):
  if stderr is None:
    stderr = sys.stdout
  if allow_non_zero:
    try:
      output = subprocess.check_output(cmd, stderr=stderr)
    except subprocess.CalledProcessError as e:
      output = e.output
  else: