Status QueueRunner::Start(Session* sess, int wait_for) {
  counter_.reset(new BlockingCounter(runs_));
  for (const string& enqueue_op : enqueue_op_names_) {
    thread_pool_->Schedule(
        std::bind(&QueueRunner::Run, this, sess, enqueue_op));
  }
  if (coord_) {
    thread_pool_->Schedule(std::bind(&QueueRunner::Stop, this, sess));
  }
  // Wait for up to 'wait_for' milliseconds.
  if (wait_for > 0) {

      client_(client),
      transfer_manager_(client->backend().transfer_manager()),
      shape_determination_fns_(std::move(shape_determination_fns)),
      thread_pool_(thread_pool) {
  CHECK(host_to_device_stream_ != nullptr);
  CHECK(stream_ != nullptr);
}

void XlaDeviceContext::CopyTensorInSameDevice(const Tensor* input_tensor,
                                              Device* device,

  VLOG(1) << "XlaDevice options: use_multiple_streams: "
          << options.use_multiple_streams << " use_global_compute_stream: "
          << options.use_global_compute_stream;
  thread_pool_.reset(new thread::ThreadPool(session_options.env, "xla_device",
                                            /*num_threads=*/1));

  // We have multiple device to device streams to allow for some concurrency

#include "tensorflow/compiler/mlir/tfrt/transforms/passes.h"
#include "tensorflow/compiler/mlir/tfrt/utils/host_context.h"
#include "tensorflow/core/framework/tensor.h"
#include "tensorflow/core/platform/threadpool.h"
#include "tensorflow/core/platform/threadpool_interface.h"
#include "tensorflow/core/runtime_fallback/kernel/kernel_fallback_execute_compat_eager.h"
#include "tensorflow/core/runtime_fallback/runtime/kernel_utils.h"

  // TODO(b/113217601): This is needed for EagerContext::runner_ to use a
  // threadpool, so that we avoid the possibility of running the runner_ in the
  // threadpool of GPU event mgr, as that can trigger more callbacks to be
  // scheduled on that same threadpool, causing a deadlock in cases where the
  // caller of event_mgr->ThenExecute() blocks on the completion of the callback

### Path operation functions

When you declare a *path operation function* with normal `def` instead of `async def`, it is run in an external threadpool that is then awaited, instead of being called directly (as it would block the server).

### Path operation functions

Cuando declaras una *path operation function* con `def` normal en lugar de `async def`, se ejecuta en un threadpool externo que luego es "<abbr title="En español: esperado. Usando await.">awaited</abbr>", en lugar de ser llamado directamente (ya que bloquearía el servidor).

// Get-or-create thread pool for a given collective.
static thread::ThreadPool* GetOrCreateThreadPoolForCollective(
    const XlaCompilationResult::CollectiveInfo& collective_info) {
  static absl::Mutex m(absl::kConstInit);
  static auto& thread_pool_cache ABSL_GUARDED_BY(m) =
      *new absl::node_hash_map<XlaCompilationResult::CollectiveInfo,
                               thread::ThreadPool>();
  absl::MutexLock l(&m);

### Funções de operação de rota

Quando você declara uma *função de operação de rota* com `def` normal ao invés de `async def`, ela é rodada em uma threadpool externa que então é aguardada, ao invés de ser chamada diretamente (ela poderia bloquear o servidor).

  tensorflow::Status error_status;
  {
    // Start a threadpool to convert signatures, since signature conversion can
    // be time consuming especially for large models. Threadpool destructor
    // blocks until all work is done.
    thread::ThreadPool thread_pool(Env::Default(), "ConvertSignatures",
                                   kNumThreadToConvertSignatures);