void MarkAsResult(AbstractTensorHandle* gradient) const override;

  void DeleteGradient(AbstractTensorHandle* gradient) const override;

 private:
  // The context where the aggregation op `Add` is to be created.
  AbstractContext* ctx_;
};

// Returns the number of elements in the gradient tensor.
int64_t TapeVSpace::NumElements(AbstractTensorHandle* tensor) const {

  llvm::SmallVector<mlir::TF::SplitOp, 4> split_ops_for_tiled_input;
  split_ops_for_tiled_input.reserve(
      input_sharding.tile_assignment_devices_size());

  // Creates a tree of split nodes for sharding tiled inputs. Splits nodes
  // are created such that input data is sharded in row major order.
  // Split nodes at ith depth from the original input node represent nodes
  // that split the input data at i-th dimension.

  } 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,

  if (subtypes.size() != 1) {
    return std::nullopt;
  }
  return subtypes[0].getElementType();
}

// Create an `ConstBytesAttr` which encodes the options
// for the `tf.custom` tensor list op to be created. If the given
// op is not a `tf.TensorList*` op, return empty, although this case
// should never be trigged in practice since patterns are only applied
// on `tf.TensorList*` ops.

  }

  auto dataset_while = builder.create<TF::WhileRegionOp>(
      reduce_dataset.getLoc(), while_input_types, /*input=*/while_input_values,
      /*parallel_iterations=*/10, false,
      /*shape_invariant=*/false);

  // `_lower_using_switch_merge` is the default for While ops created
  // in TensorFlow and allows lowering to V1 control flow for loop
  // parallelization.

        xla_tensor->AllocateShapedBuffer(device_tensor->dtype(), shape, client_,
                                         stream_->parent()->device_ordinal()));

    // The cpu_tensor and literal that we created here hold the data of host
    // tensor in descending layout. The layout could be different from layout in
    // device_tensor (but the logical shape has to be the same). The

          attr_cst =
              builder.create<ConstOp>(op->getLoc(), output_type, attribute);
        } else {
          attr_cst = builder.create<mlir::arith::ConstantOp>(
              op->getLoc(), cast<TypedAttr>(attribute));
        }
        new_operands.push_back(attr_cst);
      }
    }

    // Create the TFR call op
    auto new_op = builder.create<CallOp>(

      absl::StrCat("device: ", DEVICE_TPU_SYSTEM, " device"));
  devices->push_back(std::make_unique<VirtualDevice>(options.env, attrs));
  VLOG(1) << "Created TPU_SYSTEM device. This host has " << device_count
          << " TPUs";

  return absl::OkStatus();
}

}  // namespace

void RegisterTpuDeviceToDeviceCopy() {
  static auto* const register_tpu_tpu_copy = new CopyTensor::Registration(

namespace mlir {
namespace TF {

namespace {

// Creates ConstOp for int32_t value.
ConstOp createI32ConstOp(int32_t value, Location loc,
                         PatternRewriter* rewriter) {
  auto int_attr = IntegerAttr::get(rewriter->getIntegerType(32), value);
  return rewriter->create<ConstOp>(loc, int_attr);
}

// Creates ConstantOp for array of int32_t.

}

// This is a basic test to make sure the C++ gradient code can handle while
// loops created by the C API (which calls the C++ API under the hood). There
// are more while loop gradient tests in cc/framework/while_gradients_test.cc.
TEST_F(CApiWhileLoopTest, Gradients) {
  Init(1);

  // Create loop: while (i < 10) i += 1
  TF_Operation* ten = ScalarConst(10, params_->cond_graph, s_);
  TF_Operation* less_than =