replicas taking part in sync training.
    *   Performance improvements for GPU multi-worker distributed training using
        `tf.distribute.experimental.MultiWorkerMirroredStrategy`
    *   Update NVIDIA `NCCL` to `2.5.7-1` for better performance and performance
        tuning. Please see

    default_cc_opt_flags = '/arch:AVX'
  else:
    # On all other platforms, no longer use `-march=native` as this can result
    # in instructions that are too modern being generated. Users that want
    # maximum performance should compile TF in their environment and can pass
    # `-march=native` there.
    # See https://github.com/tensorflow/tensorflow/issues/45744 and duplicates
    default_cc_opt_flags = '-Wno-sign-compare'

  if (!s.ok()) {
    return s;
  }

  std::unordered_map<int64_t, int64_t> gradients_size;
  // TODO(apassos) multiple threads could be dequeuing from op_stack at the same
  // time, for better CPU backprop performance.
  VLOG(1) << "Initial stack:";
  if (VLOG_IS_ON(1)) {
    for (auto t : op_stack) {
      VLOG(1) << "  " << t;
    }
  }
  while (!op_stack.empty()) {
    const int64_t op = op_stack.back();

# Linux x86 so that we can use RBE. Since tests still need to run on the single
# host Arm64 machine, the build becomes too slow (~30 min) to be a presubmit.
# For testing purposes, we want to see the runtime performance of an
# experimental job that is build-only, i.e, we only build the test targets and
# do not run them. By prefixing the configs with "build", we can run both

                                    fn_name, "'");
    output_tensors->emplace_back(node, idx);
  }
  return absl::OkStatus();
}

// Populates `body_nodes` with the nodes that will become function's body.
// Performs various checks.
Status ComputeBodyNodes(
    const TF_Graph* fn_body, const char* fn_name, int num_opers,
    const TF_Operation* const* opers,
    const std::unordered_map<const Node*, std::vector<int>>& input_nodes,