"//tensorflow/c/experimental/filesystem:modular_filesystem",
            "//tensorflow/cc:grad_ops",
            "//tensorflow/cc:gradients",
            "//tensorflow/cc:ops",
            "//tensorflow/cc:scope_internal",
            "//tensorflow/cc:while_loop",
            "//tensorflow/cc/saved_model:loader_lite",
            "//tensorflow/compiler/mlir/tfr:graph_decompose_pass",

// efficiency of forward-mode autodiff. Instead, language bindings should call
// DeleteGradient as soon as a Tensor which was `Watch`ed or was an output
// Tensor passed to Accumulate goes out of scope.
//
// Not thread-safe.
template <typename Gradient, typename BackwardFunction, typename TapeTensor>
class ForwardAccumulator {
 public:
  // Does not take ownership of `vspace`, which must outlive the

  auto* rewrite_options =
      config->mutable_graph_options()->mutable_rewrite_options();
  rewrite_options->set_scoped_allocator_optimization(
      tensorflow::RewriterConfig::ON);
  rewrite_options->mutable_scoped_allocator_opts()->add_enable_op(
      "CollectiveReduce");

  if ((tensorflow::IsGoogleCudaEnabled() || tensorflow::IsBuiltWithROCm()) &&
      num_virtual_gpus > 0) {