let description = [{
    This pass outlines the body of a `tf_device.cluster` into a function and
    replaces the `tf_device.cluster` op with an equivalent `tf_device.cluster_func`
    op. Implicit operands will be captured and materialized as explicit arguments to
    the newly created functions and associated `tf_device.cluster_func` ops.

    For example, the following:

    ```mlir

// Tests `tf_device.cluster_func` with unsupported operand type.

module attributes {tf.versions = {producer = 888 : i32}, tf.devices = ["/job:worker/replica:0/task:0/device:CPU:0", "/job:worker/replica:0/task:0/device:TPU_SYSTEM:0", "/job:worker/replica:0/task:0/device:TPU:0"]} {
  func.func @unsupported_operand_type(%arg0: tensor<?xi2>) {
    // expected-error@+1 {{failed to determine operand type at index 0: Converting i2 to DataType}}

		case ACELFBR:
			opcode = op_CELFBR
		case ACDLFBR:
			opcode = op_CDLFBR
		case ACELGBR:
			opcode = op_CELGBR
		case ACDLGBR:
			opcode = op_CDLGBR
		}
		// set immediate operand M3 to 0 to use the default BFP rounding mode
		// (usually round to nearest, ties to even)
		// TODO(mundaym): should this be fixed at round to nearest, ties to even?
		// M4 is reserved and must be 0

(Less32 x y) => (LessThan (CMPW x y))
(Less64 x y) => (LessThan (CMP  x y))

// Set condition flags for floating-point comparisons "x < y"
// and "x <= y". Because if either or both of the operands are
// NaNs, all three of (x < y), (x == y) and (x > y) are false,
// and ARM Manual says FCMP instruction sets PSTATE.<N,Z,C,V>
// of this case to (0, 0, 1, 1).
(Less32F x y) => (LessThanF (FCMPS x y))

  // expected-remark@above {{Sinks: {5}}}
}

// -----

// Tests that we create dependencies to a fetch op, even if the fetch op has
// known effects (in this case due to resource operand) and the resources of
// other side-effecting ops are independent.
func.func @fetch_with_resource_operand(
  // expected-remark@above {{ID: 9}}
  %arg0: tensor<!tf_type.string>,

        equation,
        y_shape,
        x_signature,
        y_signature,
        bias_shape,
        activation_fn=nn_ops.relu,
    )

    # Use constant y operand.
    signatures = {
        'serving_default': model.einsum_with_kernel.get_concrete_function(),
    }

    saved_model_save.save(model, self._input_saved_model_path, signatures)