Google 會運用 AI 技術將內容翻譯成你偏好的語言，但可能會出錯。

管理 TPU 資源

本頁說明如何列出、停止、啟動、刪除及連線至 TPU VM。

必要條件

在執行這些步驟之前，請完成以下步驟：

依照為 TPU 設定 Google Cloud 項目中的描述，為您的 TPU 建立一個 Google Cloud 項目。
依照規劃 Cloud TPU 資源中的描述來決定您的 TPU 需求。
依照建立 TPU VM中的描述建立 TPU VM。
如果您使用其中一個 Cloud 用戶端程式庫，請按照所用語言的設定說明操作：
- Python
- Java
- Node.js

設定環境變數。

  export TPU_NAME=your-tpu-name
  export ZONE=your-zone

連接到 Cloud TPU

您可以使用 SSH 連線到 Cloud TPU。

如果無法使用 SSH 連線至 TPU VM，可能是因為 TPU VM 沒有外部 IP 位址。若要存取沒有外部 IP 位址的 TPU VM，請依照連接到沒有公用 IP 位址的 TPU VM 中的說明進行操作。

gcloud

使用 SSH 連線至 Cloud TPU：

$ gcloud compute tpus tpu-vm ssh $TPU_NAME --zone=$ZONE

當您要求的切片大於單一主機時，Cloud TPU 會為每個主機建立一個 TPU VM。每個主機上的 TPU 晶片數量取決於 TPU 版本。

若要安裝二進位檔案或執行程式碼，請使用 tpu-vm ssh 指令連接到每個 TPU VM。

$ gcloud compute tpus tpu-vm ssh $TPU_NAME --zone=$ZONE

若要使用 SSH 連接到特定的 TPU VM，請使用 --worker 旗標和從 0 開始的索引：

$ gcloud compute tpus tpu-vm ssh $TPU_NAME --zone=$ZONE --worker=1

若要在所有 TPU 虛擬機器上執行命令，請使用 --worker=all 和 --command 標誌：

$ gcloud compute tpus tpu-vm ssh $TPU_NAME \
  --zone=$ZONE \
  --worker=all \
  --command='pip install "jax[tpu]==0.4.20" -f https://storage.googleapis.com/jax-releases/libtpu_releases.html'

如果是 Multislice，您可以使用列舉的 TPU 名稱，在單一 VM 上執行指令，每個切片前置字串和附加的數字都會附加到該名稱。如要在所有切片的所有 TPU VM 上執行指令，請使用 --node=all、--worker=all 和 --command 標記，並視需要使用 --batch-size 標記。

$ gcloud compute tpus queued-resources ssh your-queued-resource-id \
  --zone=$ZONE \
  --node=all \
  --worker=all \
  --command='pip install jax[tpu] -f https://storage.googleapis.com/jax-releases/libtpu_releases.html' \
  --batch-size=4

當您使用 Google Cloud CLI 連線到虛擬機器時，Compute Engine 會建立一個持久的 SSH 金鑰。

控制台

若要在 Google Cloud 控制台中連接到您的 TPU，請在瀏覽器中使用 SSH：

前往 Google Cloud 控制台的「TPUs」頁面：

前往 TPU
在 TPU VM 清單中，找到要連線的 TPU VM，然後按一下該列中的「SSH」SSH。

透過 Google Cloud 控制台連線至 TPU VM 時，Compute Engine 會建立暫時性 SSH 金鑰。

列出 Cloud TPU 資源

您可以列出指定區域內的所有 Cloud TPU 資源。

gcloud

$ gcloud compute tpus tpu-vm list --zone=$ZONE

控制台

前往 Google Cloud 控制台的「TPUs」頁面：

前往 TPU

Java

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.tpu.v2.ListNodesRequest;
import com.google.cloud.tpu.v2.TpuClient;
import java.io.IOException;

public class ListTpuVms {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    // Project ID or project number of the Google Cloud project you want to use.
    String projectId = "YOUR_PROJECT_ID";
    // The zone where the TPUs are located.
    // For more information about supported TPU types for specific zones,
    // see https://cloud.google.com/tpu/docs/regions-zones
    String zone = "us-central1-f";

    listTpuVms(projectId, zone);
  }

  // Lists TPU VMs in the specified zone.
  public static TpuClient.ListNodesPage listTpuVms(String projectId, String zone)
      throws IOException {
    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests.
    try (TpuClient tpuClient = TpuClient.create()) {
      String parent = String.format("projects/%s/locations/%s", projectId, zone);

      ListNodesRequest request = ListNodesRequest.newBuilder().setParent(parent).build();

      return tpuClient.listNodes(request).getPage();
    }
  }
}

Node.js

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

// Import the TPUClient
// TODO(developer): Uncomment below line before running the sample.
// const {TpuClient} = require('@google-cloud/tpu').v2;

// Instantiate a tpuClient
// TODO(developer): Uncomment below line before running the sample.
// tpuClient = new TpuClient();

// TODO(developer): Update these variables before running the sample.
// Project ID or project number of the Google Cloud project you want to retrive a list of TPU nodes.
const projectId = await tpuClient.getProjectId();

// The zone from which the TPUs are retrived.
const zone = 'europe-west4-a';

async function callTpuVMList() {
  const request = {
    parent: `projects/${projectId}/locations/${zone}`,
  };

  const [response] = await tpuClient.listNodes(request);

  return response;
}

return await callTpuVMList();

Python

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

from google.cloud import tpu_v2

# TODO(developer): Update and un-comment below lines
# project_id = "your-project-id"
# zone = "us-central1-b"

client = tpu_v2.TpuClient()

nodes = client.list_nodes(parent=f"projects/{project_id}/locations/{zone}")
for node in nodes:
    print(node.name)
    print(node.state)
    print(node.accelerator_type)
# Example response:
# projects/[project_id]/locations/[zone]/nodes/node-name
# State.READY
# v2-8

取得 Cloud TPU 資訊

您可以檢索有關特定雲端 TPU 的資訊。

gcloud

$ gcloud compute tpus tpu-vm describe $TPU_NAME \
  --zone=$ZONE

控制台

前往 Google Cloud 控制台的「TPUs」頁面：

前往 TPU
點選您的 Cloud TPU 名稱。控制台顯示 Cloud TPU 詳細資料頁面。

Java

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.tpu.v2.GetNodeRequest;
import com.google.cloud.tpu.v2.Node;
import com.google.cloud.tpu.v2.NodeName;
import com.google.cloud.tpu.v2.TpuClient;
import java.io.IOException;

public class GetTpuVm {

  public static void main(String[] args) throws IOException {
    // TODO(developer): Replace these variables before running the sample.
    // Project ID or project number of the Google Cloud project you want to use.
    String projectId = "YOUR_PROJECT_ID";
    // The zone in which to create the TPU.
    // For more information about supported TPU types for specific zones,
    // see https://cloud.google.com/tpu/docs/regions-zones
    String zone = "europe-west4-a";
    // The name for your TPU.
    String nodeName = "YOUR_TPU_NAME";

    getTpuVm(projectId, zone, nodeName);
  }

  // Describes a TPU VM with the specified name in the given project and zone.
  public static Node getTpuVm(String projectId, String zone, String nodeName)
      throws IOException {
    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests.
    try (TpuClient tpuClient = TpuClient.create()) {
      String name = NodeName.of(projectId, zone, nodeName).toString();

      GetNodeRequest request = GetNodeRequest.newBuilder().setName(name).build();

      return tpuClient.getNode(request);
    }
  }
}

Node.js

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

// Import the TPUClient
// TODO(developer): Uncomment below line before running the sample.
// const {TpuClient} = require('@google-cloud/tpu').v2;

// Instantiate a tpuClient
// TODO(developer): Uncomment below line before running the sample.
// tpuClient = new TpuClient();

// TODO(developer): Update these variables before running the sample.
// Project ID or project number of the Google Cloud project you want to retrive a node.
const projectId = await tpuClient.getProjectId();

// The name of TPU to retrive.
const nodeName = 'node-name-1';

// The zone, where the TPU is created.
const zone = 'europe-west4-a';

async function callGetTpuVM() {
  const request = {
    name: `projects/${projectId}/locations/${zone}/nodes/${nodeName}`,
  };

  const [response] = await tpuClient.getNode(request);

  console.log(`Node: ${nodeName} retrived.`);
  return response;
}

return await callGetTpuVM();

Python

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

from google.cloud import tpu_v2

# TODO(developer): Update and un-comment below lines
# project_id = "your-project-id"
# zone = "us-central1-b"
# tpu_name = "tpu-name"

client = tpu_v2.TpuClient()
node = client.get_node(
    name=f"projects/{project_id}/locations/{zone}/nodes/{tpu_name}"
)

print(node)
# Example response:
# name: "projects/[project_id]/locations/[zone]/nodes/tpu-name"
# state: "READY"
# runtime_version: ...

停止 Cloud TPU 資源

您可以停止單一 Cloud TPU 以避免產生費用，而不會遺失其 VM 配置和軟體。

已排隊資源 API 不支援停止 TPU 切片或 TPU。如要停止支付透過佇列資源 API 分配的 TPU 費用，請刪除 TPU。

gcloud

$ gcloud compute tpus tpu-vm stop $TPU_NAME \
  --zone=$ZONE

控制台

前往 Google Cloud 控制台的「TPUs」頁面：

前往 TPU
選取雲端 TPU 旁的複選框。
按一下「停止」。

Java

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.tpu.v2.Node;
import com.google.cloud.tpu.v2.NodeName;
import com.google.cloud.tpu.v2.StopNodeRequest;
import com.google.cloud.tpu.v2.TpuClient;
import java.io.IOException;
import java.util.concurrent.ExecutionException;

public class StopTpuVm {

  public static void main(String[] args)
      throws IOException, ExecutionException, InterruptedException {
    // TODO(developer): Replace these variables before running the sample.
    // Project ID or project number of the Google Cloud project you want to use.
    String projectId = "YOUR_PROJECT_ID";
    // The zone where the TPU is located.
    // For more information about supported TPU types for specific zones,
    // see https://cloud.google.com/tpu/docs/regions-zones
    String zone = "us-central1-f";
    // The name for your TPU.
    String nodeName = "YOUR_TPU_NAME";

    stopTpuVm(projectId, zone, nodeName);
  }

  // Stops a TPU VM with the specified name in the given project and zone.
  public static Node stopTpuVm(String projectId, String zone, String nodeName)
      throws IOException, ExecutionException, InterruptedException {
    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests.
    try (TpuClient tpuClient = TpuClient.create()) {
      String name = NodeName.of(projectId, zone, nodeName).toString();

      StopNodeRequest request = StopNodeRequest.newBuilder().setName(name).build();

      return tpuClient.stopNodeAsync(request).get();
    }
  }
}

Node.js

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

// Import the TPUClient
// TODO(developer): Uncomment below line before running the sample.
// const {TpuClient} = require('@google-cloud/tpu').v2;

// Instantiate a tpuClient
// TODO(developer): Uncomment below line before running the sample.
// tpuClient = new TpuClient();

// TODO(developer): Update these variables before running the sample.
// Project ID or project number of the Google Cloud project you want to stop a node.
const projectId = await tpuClient.getProjectId();

// The name of TPU to stop.
const nodeName = 'node-name-1';

// The zone, where the TPU is created.
const zone = 'europe-west4-a';

async function callStopTpuVM() {
  const request = {
    name: `projects/${projectId}/locations/${zone}/nodes/${nodeName}`,
  };

  const [operation] = await tpuClient.stopNode(request);
  // Wait for the operation to complete.
  const [response] = await operation.promise();

  console.log(`Node: ${nodeName} stopped.`);
  return response;
}

return await callStopTpuVM();

Python

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

from google.cloud import tpu_v2

# TODO(developer): Update and un-comment below lines
# project_id = "your-project-id"
# zone = "us-central1-a"
# tpu_name = "tpu-name"

client = tpu_v2.TpuClient()

request = tpu_v2.StopNodeRequest(
    name=f"projects/{project_id}/locations/{zone}/nodes/{tpu_name}",
)
try:
    operation = client.stop_node(request=request)
    print("Waiting for stop operation to complete...")
    response = operation.result()
    print(f"This TPU {tpu_name} has been stopped")
    print(response.state)
    # Example response:
    # State.STOPPED

except Exception as e:
    print(e)
    raise e

啟動 Cloud TPU 資源

您可以啟動已停止的 Cloud TPU。

已排隊資源 API 不支援啟動 TPU Pod 或 TPU。

gcloud

$ gcloud compute tpus tpu-vm start $TPU_NAME \
  --zone=$ZONE

控制台

前往 Google Cloud 控制台的「TPUs」頁面：

前往 TPU
選取雲端 TPU 旁的複選框。
按一下「開始」。

Java

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.cloud.tpu.v2.Node;
import com.google.cloud.tpu.v2.NodeName;
import com.google.cloud.tpu.v2.StartNodeRequest;
import com.google.cloud.tpu.v2.TpuClient;
import java.io.IOException;
import java.util.concurrent.ExecutionException;

public class StartTpuVm {

  public static void main(String[] args)
      throws IOException, ExecutionException, InterruptedException {
    // TODO(developer): Replace these variables before running the sample.
    // Project ID or project number of the Google Cloud project you want to use.
    String projectId = "YOUR_PROJECT_ID";
    // The zone where the TPU is located.
    // For more information about supported TPU types for specific zones,
    // see https://cloud.google.com/tpu/docs/regions-zones
    String zone = "us-central1-f";
    // The name for your TPU.
    String nodeName = "YOUR_TPU_NAME";

    startTpuVm(projectId, zone, nodeName);
  }

  // Starts a TPU VM with the specified name in the given project and zone.
  public static Node startTpuVm(String projectId, String zone, String nodeName)
      throws IOException, ExecutionException, InterruptedException {
    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests.
    try (TpuClient tpuClient = TpuClient.create()) {
      String name = NodeName.of(projectId, zone, nodeName).toString();

      StartNodeRequest request = StartNodeRequest.newBuilder().setName(name).build();

      return tpuClient.startNodeAsync(request).get();
    }
  }
}

Node.js

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

// Import the TPUClient
// TODO(developer): Uncomment below line before running the sample.
// const {TpuClient} = require('@google-cloud/tpu').v2;

// Instantiate a tpuClient
// TODO(developer): Uncomment below line before running the sample.
// tpuClient = new TpuClient();

// TODO(developer): Update these variables before running the sample.
// Project ID or project number of the Google Cloud project you want to start a node.
const projectId = await tpuClient.getProjectId();

// The name of TPU to start.
const nodeName = 'node-name-1';

// The zone, where the TPU is created.
const zone = 'europe-west4-a';

async function callStartTpuVM() {
  const request = {
    name: `projects/${projectId}/locations/${zone}/nodes/${nodeName}`,
  };

  const [operation] = await tpuClient.startNode(request);

  // Wait for the operation to complete.
  const [response] = await operation.promise();

  console.log(`Node: ${nodeName} started.`);
  return response;
}

return await callStartTpuVM();

Python

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

from google.cloud import tpu_v2

# TODO(developer): Update and un-comment below lines
# project_id = "your-project-id"
# zone = "us-central1-a"
# tpu_name = "tpu-name"

client = tpu_v2.TpuClient()

request = tpu_v2.StartNodeRequest(
    name=f"projects/{project_id}/locations/{zone}/nodes/{tpu_name}",
)
try:
    operation = client.start_node(request=request)
    print("Waiting for start operation to complete...")
    response = operation.result()
    print(f"TPU {tpu_name} has been started")
    print(response.state)
    # Example response:
    # State.READY

except Exception as e:
    print(e)
    raise e

刪除 Cloud TPU

工作階段結束後，請刪除 TPU VM 節點。

gcloud

$ gcloud compute tpus tpu-vm delete $TPU_NAME \
  --zone=$ZONE \
  --quiet

指令旗標說明

zone: 您計劃刪除 Cloud TPU 的 zone。
quiet: 執行 gcloud CLI 指令時停用所有互動式提示。

控制台

前往 Google Cloud 控制台的「TPUs」頁面：

前往 TPU
選取 Cloud TPU 旁的核取方塊。
按一下「刪除」圖示。

Java

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

import com.google.api.gax.longrunning.OperationTimedPollAlgorithm;
import com.google.api.gax.retrying.RetrySettings;
import com.google.cloud.tpu.v2.DeleteNodeRequest;
import com.google.cloud.tpu.v2.NodeName;
import com.google.cloud.tpu.v2.TpuClient;
import com.google.cloud.tpu.v2.TpuSettings;
import java.io.IOException;
import java.util.concurrent.ExecutionException;
import org.threeten.bp.Duration;

public class DeleteTpuVm {

  public static void main(String[] args)
      throws IOException, ExecutionException, InterruptedException {
    // TODO(developer): Replace these variables before running the sample.
    // Project ID or project number of the Google Cloud project you want to create a node.
    String projectId = "YOUR_PROJECT_ID";
    // The zone in which to create the TPU.
    // For more information about supported TPU types for specific zones,
    // see https://cloud.google.com/tpu/docs/regions-zones
    String zone = "europe-west4-a";
    // The name for your TPU.
    String nodeName = "YOUR_TPU_NAME";

    deleteTpuVm(projectId, zone, nodeName);
  }

  // Deletes a TPU VM with the specified name in the given project and zone.
  public static void deleteTpuVm(String projectId, String zone, String nodeName)
      throws IOException, ExecutionException, InterruptedException {
    // With these settings the client library handles the Operation's polling mechanism
    // and prevent CancellationException error
    TpuSettings.Builder clientSettings =
        TpuSettings.newBuilder();
    clientSettings
        .deleteNodeOperationSettings()
        .setPollingAlgorithm(
            OperationTimedPollAlgorithm.create(
                RetrySettings.newBuilder()
                    .setInitialRetryDelay(Duration.ofMillis(5000L))
                    .setRetryDelayMultiplier(1.5)
                    .setMaxRetryDelay(Duration.ofMillis(45000L))
                    .setInitialRpcTimeout(Duration.ZERO)
                    .setRpcTimeoutMultiplier(1.0)
                    .setMaxRpcTimeout(Duration.ZERO)
                    .setTotalTimeout(Duration.ofHours(24L))
                    .build()));

    // Initialize client that will be used to send requests. This client only needs to be created
    // once, and can be reused for multiple requests.
    try (TpuClient tpuClient = TpuClient.create(clientSettings.build())) {
      String name = NodeName.of(projectId, zone, nodeName).toString();

      DeleteNodeRequest request = DeleteNodeRequest.newBuilder().setName(name).build();

      tpuClient.deleteNodeAsync(request).get();
      System.out.println("TPU VM deleted");
    }
  }
}

Node.js

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

// Import the TPUClient
// TODO(developer): Uncomment below line before running the sample.
// const {TpuClient} = require('@google-cloud/tpu').v2;

// Instantiate a tpuClient
// TODO(developer): Uncomment below line before running the sample.
// tpuClient = new TpuClient();

// TODO(developer): Update these variables before running the sample.
// Project ID or project number of the Google Cloud project you want to delete a node.
const projectId = await tpuClient.getProjectId();

// The name of TPU to delete.
const nodeName = 'node-name-1';

// The zone, where the TPU is created.
const zone = 'europe-west4-a';

async function callDeleteTpuVM() {
  const request = {
    name: `projects/${projectId}/locations/${zone}/nodes/${nodeName}`,
  };

  const [operation] = await tpuClient.deleteNode(request);

  // Wait for the delete operation to complete.
  const [response] = await operation.promise();

  console.log(`Node: ${nodeName} deleted.`);
  return response;
}

return await callDeleteTpuVM();

Python

如要向 Cloud TPU 進行驗證，請設定應用程式預設憑證。詳情請參閱「為本機開發環境設定驗證機制」。

from google.cloud import tpu_v2

# TODO(developer): Update and un-comment below lines
# project_id = "your-project-id"
# zone = "us-central1-b"
# tpu_name = "tpu-name"

client = tpu_v2.TpuClient()
try:
    client.delete_node(
        name=f"projects/{project_id}/locations/{zone}/nodes/{tpu_name}"
    )
    print("The TPU node was deleted.")
except Exception as e:
    print(e)

後續步驟

瞭解如何設定網路並存取 Cloud TPU
瞭解已排隊的資源。
瞭解 GKE 中的 TPU。
瞭解如何在 TPU VM 上執行 JAX 程式碼。
學習如何在 TPU VM 上執行 PyTorch 程式碼。
瞭解如何在 TPU 上執行 ML 工作負載：在 TPU 上使用 vLLM 服務 Qwen2-72B-Instruct。