TF_SetAttrType(desc, "T", TF_INT32);
  // Set device to CPU since there is no version of split for int32 on GPU
  // TODO(iga): Convert all these helpers and tests to use floats because
  // they are usually available on GPUs. After doing this, remove TF_SetDevice
  // call in c_api_function_test.cc
  TF_SetDevice(desc, "/cpu:0");
  *op = TF_FinishOperation(desc, s);
  ASSERT_EQ(TF_OK, TF_GetCode(s)) << TF_Message(s);

		t.Run(testCase.name, func(t *testing.T) {
			// Hack cpuid to the CPU doesn't appear to support AVX2.
			// Restore whatever happens.
			if cpuid.CPU.Supports(cpuid.AVX2) {
				cpuid.CPU.Disable(cpuid.AVX2)
				defer cpuid.CPU.Enable(cpuid.AVX2)
			}
			if simdjson.SupportedCPU() {
				t.Fatal("setup error: expected cpu to be unsupported")
			}
			testReq := testCase.requestXML
			if len(testReq) == 0 {

---

Ejemplos comunes de operaciones limitadas por la CPU son cosas que requieren procesamiento matemático complejo.

Por ejemplo:

* **Procesamiento de audio** o **imágenes**.

Go 1.25 added a new `containermaxprocs` setting that controls whether the Go
runtime will consider cgroup CPU limits when setting the default GOMAXPROCS.
The default value `containermaxprocs=1` will use cgroup limits in addition to
the total logical CPU count and CPU affinity. `containermaxprocs=0` will
disable consideration of cgroup limits. This setting only affects Linux.

Et comme la plupart du temps d'exécution est pris par du "vrai" travail (et non de l'attente), et que le travail dans un ordinateur est fait par un <abbr title="Central Processing Unit">CPU</abbr>, ce sont des problèmes dits "CPU bound".

---

Des exemples communs d'opérations "CPU bounds" sont les procédés qui requièrent des traitements mathématiques complexes.

Par exemple :

* Traitements d'**audio** et d'**images**.

	systemDriveCollectorPath            collectorPath = "/system/drive"
	systemMemoryCollectorPath           collectorPath = "/system/memory"
	systemCPUCollectorPath              collectorPath = "/system/cpu"
	systemProcessCollectorPath          collectorPath = "/system/process"

	debugGoCollectorPath collectorPath = "/debug/go"

	clusterHealthCollectorPath       collectorPath = "/cluster/health"

        *   `tf.raw_ops.Bucketize` op on CPU.
        *   `tf.where` op for data types
            `tf.int32`/`tf.uint32`/`tf.int8`/`tf.uint8`/`tf.int64`.
        *   `tf.random.normal` op for output data type `tf.float32` on CPU.
        *   `tf.random.uniform` op for output data type `tf.float32` on CPU.
        *   `tf.random.categorical` op for output data type `tf.int64` on CPU.

*   `tensorflow.experimental.tensorrt`:

      return false;
    }

    if (!finalizeReference(firstReference, finalizeReferentMethod)) {
      return false;
    }

    /*
     * Loop as long as we have references available so as not to waste CPU looking up the Method
     * over and over again.
     */
    while (true) {
      Reference<?> furtherReference = queue.poll();
      if (furtherReference == null) {
        return true;
      }

각 프로세스의 **PID**를 확인할 수 있습니다. `27365`는 상위 프로세스(**프로세스 매니저**), 그리고 각각의 워커프로세스는 `27368`, `27369`, `27370`, 그리고 `27367`입니다.

## 배포 개념들

여기에서는 **유비콘 워커 프로세스**를 관리하는 **구니콘**(또는 유비콘)을 사용하여 애플리케이션을 **병렬화**하고, CPU **멀티 코어**의 장점을 활용하고, **더 많은 요청**을 처리할 수 있는 방법을 살펴보았습니다.

워커를 사용하는 것은 배포 개념 목록에서 주로 **복제본** 부분과 **재시작**에 약간 도움이 되지만 다른 배포 개념들도 다루어야 합니다:

* **보안 - HTTPS**
* **서버 시작과 동시에 실행하기**
* ***재시작***
* 복제본 (실행 중인 프로세스의 숫자)

  }

  auto* gpu_options = config.mutable_gpu_options();
  gpu_options->set_allow_growth(gpu_memory_allow_growth);

  (*config.mutable_device_count())["CPU"] = num_cpu_devices;

  // 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