// Unformat op.
  auto unformat_op = builder.create<TF::TPUReshardVariablesOp>(
      while_op.getLoc(), llvm::ArrayRef<Type>{}, unformat_operands);
  WrapOpInLaunch(&builder, execute_launch.getLoc(), unformat_op,
                 execute_launch.getDevice());
  builder.create<tf_device::ReturnOp>(while_op.getLoc(), ArrayRef<Value>{});

  return true;
}

void TPUVariableRuntimeReformattingPass::runOnOperation() {

      block_arg.getLoc(),
      DenseIntElementsAttr::get(
          RankedTensorType::get({1, num_replicas}, builder.getIntegerType(32)),
          group_assignment_val));

  StringAttr reduce_op = builder.getStringAttr("Add");
  StringAttr mode = builder.getStringAttr("CrossReplica");
  return builder.create<XlaAllReduceOp>(block_arg.getLoc(), block_arg.getType(),

    auto yield = builder->create<tf_executor::YieldOp>(
        terminator->getLoc(), terminator->getOperands());
    terminator->erase();

    // Create new island for each region.
    builder->setInsertionPoint(island_op);
    auto execute_island = builder->create<tf_executor::IslandOp>(
        island_op.getLoc(), yield.getOperandTypes(),
        island_op.getControl().getType(), island_op.getControlInputs());

    auto *new_ewise_op =
        rewriter.create(op->getLoc(), op->getName().getIdentifier(),
                        {tpose_arg1.getOperand(0), tpose_arg2.getOperand(0)},
                        new_out_type, op->getAttrs());

    // Apply original tranpose to output of ewise op.
    auto out_tpose_op = rewriter.create<TFL::TransposeOp>(
        new_ewise_op->getLoc(), op->getResult(0).getType(),

    }
    Value broadcast_multiplier = rewriter.create<mhlo::BroadcastInDimOp>(
        mul_op.getLoc(), filter.getType(), multiplier, broadcast_dims);
    Value new_filter = rewriter.create<mhlo::MulOp>(
        mul_op.getLoc(), filter.getType(), filter, broadcast_multiplier);
    Value new_conv = rewriter.create<mhlo::ConvolutionOp>(
        mul_op.getLoc(), conv_op.getType(), conv_op.getLhs(), new_filter,

  OpBuilder builder(func_.getBody());
  auto op = builder.create<mlir::TFL::NonMaxSuppressionV4Op>(
      func_.getLoc(), output_type0, output_type1, boxes, scores,
      max_output_size, iou_threshold, score_threshold);

  builder.create<mlir::func::ReturnOp>(func_.getLoc(), op.getResults());
}

LogicalResult ConvertNMSPaddedFunc::VerifySignature() {
  // Verify high-level function signature.

      cluster_func.getLoc(), 1, output_types);
  cluster_func->remove();
  auto& block = parallel_execute.GetRegionBlockWithIndex(0);
  builder->setInsertionPointToEnd(&block);
  builder->insert(cluster_func);
  cluster_func.replaceAllUsesWith(parallel_execute);
  builder->create<tf_device::ReturnOp>(block.getParent()->getLoc(),
                                       cluster_func.getResults());

            const_op->getLoc(),
            DenseElementsAttr::get(
                tensorflow::GetTypeFromTFTensorShape(
                    {}, splat_elements_attr.getType().getElementType()),
                splat_elements_attr.getSplatValue<Attribute>()));
    TFL::FillOp fill = op_builder->create<TFL::FillOp>(
        const_op->getLoc(), splat_elements_attr.getType(), fill_shape,
        fill_value);

        CreateI8F32UniformQuantizedType(
            uniform_quantize_call_op.getLoc(), *rewriter.getContext(),
            input_scale_value, input_zero_point_value);
    auto input_uniform_quantize_op =
        rewriter.create<stablehlo::UniformQuantizeOp>(
            uniform_quantize_call_op.getLoc(),
            /*result=*/
            mlir::cast<TensorType>(input_value.getType())

  auto launch_op = builder.create<tf_device::LaunchOp>(
      op->getLoc(), builder.getStringAttr(device),
      /*result_types=*/op->getResultTypes());
  op->replaceAllUsesWith(launch_op);

  launch_op.getBody().push_back(new Block);
  builder.setInsertionPointToEnd(&launch_op.GetBody());
  auto* return_op =
      builder.create<tf_device::ReturnOp>(op->getLoc(), op->getResults())
          .getOperation();