4. Run the builds scripts for the build, applying any requested project plugins.
5. Calculate the task graph for the requested tasks, running any deferred configuration actions.

Following the configuration phase, Gradle writes a snapshot of the task graph to a new configuration cache entry, for later Gradle invocations.

  }
  return %graph#0, %graph#1, %graph#2, %graph#3, %graph#4, %graph#5 : !tf_res, !tf_res, tensor<f32>, tensor<f32>, tensor<i32>, tensor<i32>
}
```
### `-tf-executor-graph-pruning`

_Prunes unreachable ops in a tf_executor.graph_

This pass removes ops from a `tf_executor.graph` that are not transitively, via

       subgraph_idx++) {
    const auto graph = model.subgraphs[subgraph_idx].get();
    const auto expected_graph = expected_model.subgraphs[subgraph_idx].get();
    for (auto& op : graph->operators) {
      for (int idx = 0; idx < op->inputs.size(); idx++) {
        if (op->inputs[idx] < 0) {
          continue;
        }
        const auto& tensor = graph->tensors[op->inputs[idx]];

  ASSERT_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

  // Add a config_proto attr, to trigger grappler graph rewrites in the current
  // eager runtime.
  if (enable_grappler) {
    tensorflow::ConfigProto config;
    // Do not skip grappler optimization even for small graphs.
    config.mutable_graph_options()
        ->mutable_rewrite_options()
        ->set_min_graph_nodes(-1);

// masks will complicate the strided_slice computation logic, we can simplify
// the logic by inserting a reshape op to pad the inputs so strided_slice can
// be easier to handle.
//
// So the graph may looks like below:
//   original_input -> strided_slice -> output
//      (transforms)
//   original_input -> reshape -> strided_slice -> output
//
// And the new shape is computed based on the masks.
//

	//
	// spec: "Implementation restriction: A compiler may restrict
	// ImportPaths to non-empty strings using only characters belonging
	// to Unicode's L, M, N, P, and S general categories (the Graphic
	// characters without spaces) and may also exclude the characters
	// !"#$%&'()*,:;<=>?[\]^`{|} and the Unicode replacement character
	// U+FFFD."
	if s == "" {
		return lit
	}

        and:
        outputDoesNotContain("Cannot set attributes for constraint \"org:test:1.0\": it was probably created by a plugin using internal APIs")
    }

    def "attributes declared on constraints contribute to graph selection"() {
        given:
        settingsFile << """
            include 'producer'
        """
        file("producer/build.gradle") << """
            plugins {
                id("base")

// files, so they may be corrupt. We are going to emit them unquoted,
// so while they don't need to be valid symbol names (and in some cases,
// involving symbol versions, they won't be) they must contain only
// graphic characters and must not contain Go comments.
func checkImportSymName(s string) {
	for _, c := range s {
		if !unicode.IsGraphic(c) || unicode.IsSpace(c) {
			fatalf("dynamic symbol %q contains unsupported character", s)

  ModuleOp module = getOperation();

  // When allow_tensorlist_pass_through_ == false the target dynamic legal
  // checks will cause this pass to fail if there are any variant type tensors
  // in the graph. The graph should still be treated as valid as long as those
  // variant tensors are not from tf-dialect tensorlist ops. Rather than change
  // the dynamic legalization predicates

The difference is that Gradle _may_ discover more dependencies and artifacts depending on the tasks you execute.
As a matter of fact, Gradle cannot automatically discover _detached configurations_, which are basically dependency graphs resolved as an internal implementation detail of the execution of a task: they are not, in particular, declared as an input of the task because they effectively depend on the configuration of the task at execution time.