shape_string = "<error computing shape>";
  } else {
    shape_string = shape.DebugString();
  }
  return absl::StrCat("TensorHandle(shape=", shape_string,
                      ", dtype=", DataType_Name(DataType()),
                      ", type=", FullType().DebugString(), ")");
}

Status AbstractTensorHandle::TensorHandleStatus() const {
  // Tensor handles in current runtime don't carry error info and this method

		// 2. If it is an empty interface, the pointer points to a _type.
		//   a. If it is a compile-time-allocated type, it points into
		//      the read-only data section.
		//   b. If it is a reflect-allocated type, it points into the Go heap.
		//      Reflect is responsible for keeping a reference to
		//      the underlying type so it won't be GCd.
		// If we ever have a moving GC, we need to change this for 2b (as

func.func @main() {
  tf_executor.graph {
  // CHECK:      node {
  // CHECK-NEXT:   name: "Const"
  // CHECK-NEXT:   op: "Const"
    %0:2 = tf_executor.island wraps "tf.Const"() {device = "", dtype = "tfdtype$DT_FLOAT", value = dense<2.500000e-01> : tensor<f32>} : () -> tensor<f32> loc("Const")

  // CHECK:      node {
  // CHECK-NEXT:   name: "tf.PartitionedCall"
  // CHECK-NEXT:   op: "PartitionedCall"

// CHECK-NEXT: [[o_chain_0:%.*]], [[o1:%.*]] = tfrt_fallback_async.executeop.seq([[in_chain]]) key(1) cost({{.*}}) device("/job:localhost/replica:0/task:0/device:CPU:0") "tf.ReadVariableOp"([[o]]) {dtype = f32} : 1
// CHECK-NEXT: [[out_ch:%.*]] = tfrt.merge.chains [[o_chain]], [[o_chain_0]]
// CHECK-NEXT: tfrt.return [[out_ch]], [[o1]] : !tfrt.chain, !tfrt_fallback.tf_tensor

</blockquote>

## Investigação

Ao usar todas as alternativas anteriores, eu tive a chance de aprender com todas elas, aproveitar ideias e combiná-las da melhor maneira que encontrei para mim e para os times de desenvolvedores com os quais trabalhava.

Por exemplo, estava claro que idealmente ele deveria ser baseado nos _type hints_ padrões do Python.

// CHECK-NEXT:       shape: [  ],
// CHECK-NEXT:       type: VARIANT,
// CHECK-NEXT:       buffer: 1,
// CHECK-NEXT:       name: "tf.Const",
// CHECK-NEXT:       quantization: {
// CHECK-EMPTY:
// CHECK-NEXT:       },
// CHECK-NEXT:       has_rank: true
// CHECK-NEXT:       variant_tensors: [ {
// CHECK-NEXT:         shape: [ 2 ],
// CHECK-NEXT:         type: INT32,
// CHECK-NEXT:         has_rank: true
// CHECK-NEXT:       } ]

  const char* device_name = DeviceName(&s);
  if (!s.ok()) {
    device_name = "<error fetching device name>";
  }
  return absl::StrCat("TensorHandle(", value_string, ", shape=", shape_string,
                      ", dtype=", DataType_Name(DataType()), ", device=\"",
                      device_name, "\")");
}

Status ImmediateExecutionTensorHandle::SummarizeValue(
    std::string& summary) const {
  Status status;

//
Status VarHandleOp(AbstractContext* ctx, AbstractTensorHandle** resource, DataType dtype, const PartialTensorShape shape, const char* container = "", const char* shared_name = "", const char* debug_name = "", absl::Span<string const> allowed_devices = {}, const char* name = nullptr, const char* raw_device_name = nullptr);

//

  TF_DeleteExecutionContext(ctx);
  return wrap(func);
}

TF_AbstractTensor* TF_AddFunctionParameter(TF_ExecutionContext* func,
                                           TF_DataType dtype, TF_Shape shape,
                                           TF_Status* s) {
  DCHECK_GE(shape.num_dims, -1);
  TracingTensorHandle* t;
  TracingContext* tracing_ctx = dyn_cast<TracingContext>(unwrap(func));

namespace {

class WhileLoopTest : public ::testing::Test {
 protected:
  WhileLoopTest() : scope_(Scope::NewRootScope()) {}

  void Init(int num_inputs, DataType dtype = DT_INT32) {
    for (int i = 0; i < num_inputs; ++i) {
      inputs_.push_back(ops::Placeholder(scope_, dtype));
    }
  }

  void CreateLoop(const ops::CondGraphBuilderFn& cond,
                  const ops::BodyGraphBuilderFn& body,