file("input-nested.txt").createNewFile()
        file("input1.txt").createNewFile()
        file("input2.txt").createNewFile()
        file("inputs").createDir()
        file("inputs/inputA.txt").createNewFile()
        file("inputs/inputB.txt").createNewFile()
        file("inputs1").createDir()
        file("inputs2").createDir()

        buildFile """
            class MyConfig {

  // The first one will the function's output
  std::vector<TF_Output> outputs;

  // Add while loop to func_graph_
  {
    // The inputs to the while loop
    std::vector<TF_Output> inputs = {{feed1, 0}, {feed2, 0}};
    std::unique_ptr<TF_WhileParams> params(new TF_WhileParams(
        TF_NewWhile(func_graph_, &inputs[0], inputs.size(), s_)));

    }

    self._assertOpAndComposite([input_, filter_, bias],
                               tf.function(gen_mnist_ops.new_conv2d),
                               ops_defs._composite_conv_add_relu, kwargs)

  def test_new_conv2d_relu6(self):
    input_ = tf.random.uniform([1, 4, 4, 1])
    filter_ = tf.random.uniform([2, 2, 1, 8])
    bias = tf.zeros([8])
    kwargs = {
        'input_': input_,
        'filter_': filter_,

        private final ModelType<T> type;
        private final List<ModelReference<?>> inputs;

        private UnmanagedElementCreationAction(ModelRuleDescriptor descriptor, ModelReference<?> subject, List<ModelReference<?>> inputs, ModelType<T> type) {
            this.subject = subject;
            this.inputs = inputs;
            this.descriptor = descriptor;
            this.type = type;
        }

        @Override

    
    def input1
    def input2
    def input1Artifact
    def input2Artifact
    def output1
    def output2
    
    def setup() {
        applyPlugin()
        addSigningProperties()
        
        input1 = getResourceFile("1.txt")
        input1Artifact = new DefaultPublishArtifact(input1.name, "Text File", "txt", null, null, input1)

            this.inputs = inputs;
            this.ruleSourceType = ruleSourceType;
            this.ruleExtractor = ruleExtractor;
        }

        @Override
        public ModelReference<?> getSubject() {
            return targetReference;
        }

        @Override
        public List<? extends ModelReference<?>> getInputs() {
            return inputs;
        }

        @Override

  // CHECK: func private @_stablehlo_main_0(%arg0: tensor<i32>
  // CHECK-SAME: %[[INPUT1:.*]]: tensor<1024x3xf32>, %[[INPUT2:.*]]: tensor<1024x3xf32>
  // CHECK: %[[CONSTANT2:.*]] = stablehlo.constant dense<1.000000e+03> : tensor<1024x3xf32>
  // CHECK: %[[ADD:.*]] = stablehlo.add %[[INPUT1]], %[[CONSTANT2]] : tensor<1024x3xf32>
  // CHECK: %[[MUL:.*]] = stablehlo.multiply %[[INPUT1]], %[[INPUT2]] : tensor<1024x3xf32>
  // CHECK: return %[[ADD]], %[[MUL]]

          absl::StrCat("tensor parsing error: ", alias));
    }

    inputs.emplace_back(feed_name, tensor);
  }

  if (!request_inputs.empty() &&
      seen_request_inputs.size() != request_inputs.size()) {
    return absl::InvalidArgumentError(absl::StrCat(
        "Inputs contains invalid name. Used request inputs: ",
        absl::StrJoin(seen_request_inputs, ","),

            inputs.add(input);
        }
        return inputs.build();
    }

    private static DataSource packSample(String name, List<DataSource> inputs, Packer packer, DataAccessor accessor) throws IOException {
        long sumLength = 0;
        for (DataSource input : inputs) {
            sumLength += input.getLength();
        }

Another common, but less obvious source of unstable inputs is when a task consumes the output of another task which produces non-repeatable results, such as the example before of a code generator that embeds timestamps in its output.

A task can only be loaded from the cache if it has stable task inputs.
Unstable task inputs result in the task having a unique set of inputs for every build, which will always result in a cache miss.