for (ClosingFuture<?> input : inputs) {
        input.becomeSubsumedInto(closeables);
      }
    }

    /**
     * Returns a new {@code ClosingFuture} pipeline step derived from the inputs by applying a
     * combining function to their values. The function can use a {@link DeferredCloser} to capture
     * objects to be closed when the pipeline is done.
     *

      func.func @main(%arg_0: tensor<!tf_type.string> {tf_saved_model.bound_input = @__tf_saved_model_asset0}) -> () attributes {tf.entry_function = {inputs = "arg_0:0", outputs = ""}} {
        return
      }
    }
  )mlir");

  FailureOr<SmallVector<AssetFileDef>> asset_file_defs =
      ConvertAssetArgs(*module_op);

  EXPECT_TRUE(succeeded(asset_file_defs));

        // Once we migrate to all-scheduled transforms we should consider if we can avoid having this optimization and use only normalized inputs.
        //
        // Note that we are not capturing this value in the actual inputs of the work; doing so would cause unnecessary cache misses.
        // This is why the hash is captured here and not in visitIdentityInputs().

                compile 'test:test1:1.0'
                compile 'test:test2:1.0'
                compile 'test:test3:1.0'
                compile 'test:test4:1.0'
            }
            task resolve {
                inputs.files configurations.compile
                def files = $expression
                doLast {
                    println files as List
                }
            }
"""

        given:

// Compiles a Graph from TF to HLO and adds the resulting HLO to the
// XlaBuilder. This function adds HLO to a larger HLO computation, so
// HLO-level inputs are supplied, and HLO-level outputs are produced.
// xla_params is the HLO-level inputs and returns is the HLO-level outputs.
// If unconditionally_use_output_shapes is true then the unregistered
// attribute _output_shapes is always used to set the output shapes of the ops.

                                PatternRewriter& rewriter) const override {
    MLIRContext* ctx = reduce_op.getContext();
    ValueRange inputs = reduce_op.getInputs();
    // Only handle binary reduce ops for now
    if (inputs.size() != 1) return failure();
    Value reduce_input = inputs[0];

    RankedTensorType reduce_type =
        reduce_op.getResultTypes()[0].cast<RankedTensorType>();

      TF_AddGradientsWithPrefix(graph_, prefix, outputs, noutputs, inputs,
                                ninputs, grad_inputs, s_, grad_outputs);
    } else {
      TF_AddGradientsWithPrefix(graph_, prefix, outputs, noutputs, inputs,
                                ninputs, nullptr, s_, grad_outputs);
    }
  }

  void BuildErrorGraph(TF_Output* inputs, TF_Output* outputs) {

    $metadata is a serialized TPUCompileMetadataProto describing the shapes and types of the inputs to the computation, as well as a mapping onto the TPU pod topology.
    $constant_operand_indices are the indices of the inputs that are constant to the TPU program (e.g. weights in inference), the rest of the inputs are input tensors.
    constant_operand_indices is sorted in ascending order.

  Output a = ops::_Arg(root.WithOpName("A"), DT_RESOURCE, 0);
  std::vector<NodeBuilder::NodeOut> inputs({NodeBuilder::NodeOut(a.node())});

  Node* call;
  NameAttrList f_name_attr;
  f_name_attr.set_name(fd.signature().name());
  TF_ASSERT_OK(
      NodeBuilder("B", "StatefulPartitionedCall", &root.graph()->flib_def())
          .Input(inputs)
          .Attr("Tin", {DT_RESOURCE})
          .Attr("Tout", {DT_RESOURCE})

            TaskInputs inputs = task.getInputs();
            inputs.files(extension.getCssFiles())
                .withPropertyName("manual")
                .withPathSensitivity(PathSensitivity.RELATIVE);
            inputs.dir("src/main/resources")
                .withPropertyName("resources")
                .withPathSensitivity(PathSensitivity.RELATIVE);
            inputs.dir(extension.getUserManual().getSnippets())