CreateLoop(
      [](const Scope& s, const std::vector<Output>& inputs, Output* output) {
        *output = ops::Less(s, inputs[0], 10);
        return s.status();
      },
      [](const Scope& s, const std::vector<Output>& inputs,
         std::vector<Output>* outputs) {
        outputs->push_back(ops::AddN(s, {inputs[0], inputs[1]}));
        outputs->push_back(ops::AddN(s, {inputs[1], 1}));
        outputs->push_back(inputs[2]);

[.screenshot]
image::build-cache/overlapping-outputs-input-comparison.png[]

== Achieving stable task inputs

It is crucial to have <<build_cache_concepts.adoc#stable_task_inputs,stable task inputs>> for every cacheable task.
In the following section you will learn about different situations which violate stable task inputs and look at possible solutions.

Input normalization is used by Gradle to determine if two task inputs are _essentially_ the same.
Gradle uses normalized inputs when doing up-to-date checks and when determining if a cached result can be re-used instead of executing the task.
As input normalization is declared by the task _consuming_ the data as input, different tasks can define different ways to normalize the same data.

    }

    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, ","),
        ", request input: ", absl::StrJoin(GetMapKeys(request_inputs), ",")));
  }

        }
        if (TF::TPUReplicatedInputOp input =
                llvm::dyn_cast<TF::TPUReplicatedInputOp>(user)) {
          if (!input.getIsPacked()) {
            input.emitOpError() << "unexpected variable input, not packed";
            return LogicalResult::failure();
          }

          if (is_variable) {
            input.emitOpError() << "unexpected multiple TPUReplicatedInputOp "

        }
        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();
        }
        DataTarget target = accessor.createTarget(name, Level.Trial);
        packer.pack(inputs, target);

    }

    @Override
    public void pack(List<DataSource> inputs, DataTarget output) throws IOException {
        TarOutputStream tarOutput = new TarOutputStream(output.openOutput());
        for (DataSource input : inputs) {
            @SuppressWarnings("OctalInteger")
            TarHeader header = TarHeader.createHeader(input.getName(), input.getLength(), 0, false, 0644);
            TarEntry entry = new TarEntry(header);

  [(IsInt32ElementType $einsum),
   // Constraint to check:
   // 1. The einsum has two inputs and one output.
   // 2. The einsum is not created by the convert function itself.
   // 3. Both inputs are int32 tensor.
   // 4. Both inputs have the graph ancestor of either const-(sub), or cast-sub.
   // 5. The type of the const tensor (or input of the cast operation) is int8.
   (IsEinsumOpSupported $einsum, $args, $equation)],

     *
     * @return The input files. Returns an empty collection if this task has no input files.
     */
    FileCollection getFiles();

    /**
     * Registers some input files for this task.
     *
     * @param paths The input files. The given paths are evaluated as per {@link org.gradle.api.Project#files(Object...)}.
     * @return a property builder to further configure the property.
     */

    }

    @Override
    public void pack(List<DataSource> inputs, DataTarget output) throws IOException {
        TarArchiveOutputStream tarOutput = new TarArchiveOutputStream(output.openOutput());
        for (DataSource input : inputs) {
            TarArchiveEntry entry = new TarArchiveEntry(input.getName());
            entry.setSize(input.getLength());
            tarOutput.putArchiveEntry(entry);