* java.util.LinkedHashSet}. Generally speaking, this class reduces object allocation and memory
 * consumption at the price of moderately increased constant factors of CPU. Only use this class
 * when there is a specific reason to prioritize memory over CPU.
 *
 * @author Louis Wasserman
 */
@GwtIncompatible // not worth using in GWT for now
final class CompactLinkedHashSet<E extends @Nullable Object> extends CompactHashSet<E> {

async def sse_items_raw():
    yield ServerSentEvent(raw_data="plain text without quotes")
    yield ServerSentEvent(raw_data="<div>html fragment</div>", event="html")
    yield ServerSentEvent(raw_data="cpu,87.3,1709145600", event="csv")


router = APIRouter()


@router.get("/events", response_class=EventSourceResponse)
async def stream_events():
    yield {"msg": "hello"}
    yield {"msg": "world"}

 * java.util.LinkedHashMap}. Generally speaking, this class reduces object allocation and memory
 * consumption at the price of moderately increased constant factors of CPU. Only use this class
 * when there is a specific reason to prioritize memory over CPU.
 *
 * @author Louis Wasserman
 */
@J2ktIncompatible // no support for access-order mode in LinkedHashMap delegate
@GwtIncompatible // not worth using in GWT for now

TFCI_BAZEL_TARGET_SELECTING_CONFIG_PREFIX=windows_x86_ml_actions
TFCI_BUILD_PIP_PACKAGE_WHEEL_NAME_ARG="--repo_env=WHEEL_NAME=tensorflow"
TFCI_OUTPUT_DIR=build_output
TFCI_FIND_BIN=C:/tools/msys64/usr/bin/find.exe
TFCI_LIB_SUFFIX="-cpu-windows-x86_64"
# auditwheel is not supported for Windows
TFCI_WHL_AUDIT_ENABLE=0
TFCI_WHL_AUDIT_PLAT=0
# Tests are extremely slow at the moment
TFCI_WHL_BAZEL_TEST_ENABLE=0
TFCI_WHL_SIZE_LIMIT=450M

                    return "3";
                case "crawler.hotthread.timeout":
                    return "30s";
                case "crawler.hotthread.type":
                    return "cpu";

                default:
                    throw new ConfigPropertyNotFoundException(propertyKey);
                }
            }

            @Override

                            // the ID of the CPU socket
                            if (name.equals("physical id")) {
                                currentID = value;
                            }
                            // Number of cores not including hyper-threading
                            if (name.equals("cpu cores")) {
                                assert currentID.isEmpty() == false;

        ctx, name, TF_FLOAT, dims, 2, data, size, &Deleter, copy, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TFE_TensorHandle* on_host =
        TFE_TensorHandleCopyToDevice(copy_aliased, ctx, "CPU:0", status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);
    TF_Tensor* resolved = TFE_TensorHandleResolve(on_host, status);
    CHECK_EQ(TF_OK, TF_GetCode(status)) << TF_Message(status);

 * java.util.LinkedHashMap}. Generally speaking, this class reduces object allocation and memory
 * consumption at the price of moderately increased constant factors of CPU. Only use this class
 * when there is a specific reason to prioritize memory over CPU.
 *
 * @author Louis Wasserman
 */
@J2ktIncompatible // no support for access-order mode in LinkedHashMap delegate
@GwtIncompatible // not worth using in GWT for now

這裡唯一新增的選項是 `--workers`，告訴 Uvicorn 要啟動 4 個工作處理序。

你也會看到它顯示每個處理序的 **PID**，`27365` 是父處理序（這是**處理序管理器**），另外每個工作處理序各有一個：`27368`、`27369`、`27370`、`27367`。

## 部署概念 { #deployment-concepts }

你已經看到如何使用多個 **workers** 來將應用的執行進行**平行化**，善用 CPU 的**多核心**，並能服務**更多請求**。

在上面的部署概念清單中，使用 workers 主要能幫助到**副本**這一塊，並對**重啟**也有一點幫助，但你仍需要處理其他部分：

* **安全 - HTTPS**
* **系統啟動時執行**
* ***重啟***
* 副本（正在執行的處理序數量）
* **記憶體**
* **啟動前的前置作業**

  std::string device_type = parsed_name.type;
  int device_id = 0;
  if (parsed_name.has_id) {
    device_id = parsed_name.id;
  }

  ctx.device_id = device_id;
  if (device_type == "CPU") {
    ctx.device_type = DLDeviceType::kDLCPU;
  } else if (device_type == "GPU") {
#if TENSORFLOW_USE_ROCM
    ctx.device_type = DLDeviceType::kDLROCM;
#else
    ctx.device_type = DLDeviceType::kDLCUDA;
#endif