llvm::SmallVector<Value, 2> new_operands;
    FilterOutBlockArgControlDep(op->getOperands().drop_front(), new_operands);

    auto source_op = builder.create<tf_executor::NextIterationSourceOp>(
        op->getLoc(), op->getOperand(0).getType());
    builder.create<tf_executor::NextIterationSinkOp>(
        op->getLoc(), source_op.getToken(), /*input=*/op->getOperand(0),
        /*controlInputs=*/new_operands);

    indices[i] = dimension_numbers.start_index_map()[i];
  }

  // Fill elements from start_indices with start_index_map
  Value scattered_start_indices = builder.create<TF::TensorScatterUpdateOp>(
      loc, empty_start_indices,
      /*indices=*/
      builder.create<TF::ReshapeOp>(
          loc, RankedTensorType::get({index_map_size, 1}, builder.getI64Type()),
          Create1DConstValue<int64_t>(builder, loc, indices),

  Type resultType = RankedTensorType::get(shape, element_type);
  auto constant_attr = DenseElementsAttr::get(shape_spec_type, shape);
  auto shape_tensor =
      rewriter.create<TF::ConstOp>(loc, shape_spec_type, constant_attr);
  return rewriter.create<TF::ReshapeOp>(loc, resultType, /*tensor=*/value,
                                        /*shape=*/shape_tensor);
}

template <typename BatchMatMulOpType>

  func->setAttr(kTFImplements, attr);
  OpBuilder builder(func.getBody());
  std::string empty_option_buffer;
  auto op = builder.create<CustomOp>(
      func.getLoc(), func.getFunctionType().getResults(), func.getArguments(),
      api, CustomOption(&builder, empty_option_buffer));
  builder.create<func::ReturnOp>(func.getLoc(), op.getResults());
  return success();
}

LogicalResult VerifyNgrams(func::FuncOp func) {

  } else {
    return false;
  }
  return true;
}

// Create a dummy zero to be fed locally from the host to the TPUExecute.
Value CreateZeroInput(Location loc, OpBuilder& builder, Attribute zero_attr,
                      DenseIntElementsAttr shape_attr) {
  ConstOp zero = builder.create<ConstOp>(loc, zero_attr);
  zero->setAttr(kICIWeightDistributionMlirBridgeMarker,

                                    /*allow_partial_conversion=*/false);

  auto pass_instrumentors = mlir::GetPassInstrumentors();
  for (const auto& creator : pass_instrumentors) {
    tf2xla.addInstrumentation(creator());
  }
  if (DEBUG_DATA_DUMPER()->ShouldDump(module_name.str(), kDebugGroupMain) ||
      VLOG_IS_ON(1)) {
    tensorflow::DumpMlirOpToFile(

  for (int i = 0; i < y_num; i++) {
    dy_datas[i] = Tensor(ys[i].type(), y_shapes[i]);
    auto dy_data_flat = dy_datas[i].flat<Y_T>();
    dy_data_flat.setZero();
  }

  // Create the feed list.
  ClientSession::FeedType feed_list;
  for (int i = 0; i < x_num; i++) {
    feed_list.insert({xs[i], x_datas[i]});
  }
  for (int i = 0; i < y_num; i++) {
    feed_list.insert({dys[i], dy_datas[i]});

    compilation_successful.scalar<bool>()() = false;
    ctx->set_output(0, compilation_key);
    ctx->set_output(1, compilation_successful);
    return;
  }

  // Each execution of an XlaCompile op creates a new ExecutableClosure, even
  // if it didn't have to compile the cluster because of a compilation-cache
  // hit.  This is because we at least need new snapshots of the resource
  // variables.

                          std::vector<Output>* grad_outputs);

  // Returns a list mapping whether each node in the graph is reachable
  // from outputs_. Keyed by node id.
  std::vector<bool> GetReachableNodes();

  // Creates the gradient subgraph for a while loop (or just stores
  // `summed_grads` if not all incoming gradients are available yet). All exit
  // nodes (which are the first nodes of a loop encountered in the backwards

  const std::string& profiler_name =
      GetPjRtDeviceCompilationProfilerResourceName(device_type);
  bool deleted_old_device_compiler = false;

  // Lookup the DeviceCompiler, create one if not found.
  Status s = rm->Lookup<PjRtDeviceCompiler>(
      rm->default_container(), compiler_name, pjrt_device_compiler);
  if (s.ok() && device_type == DEVICE_TPU) {