Monitor and debug training with an interactive shell

This page shows you how to use an interactive shell to inspect the container where your training code is running. You can browse the file system and run debugging utilities in each prebuilt container or custom container running on Vertex AI.

Using an interactive shell to inspect your training container can help you debug problems with your training code or your Vertex AI configuration. For example, you can use an interactive shell to do the following:

  • Run tracing and profiling tools.
  • Analyze GPU usage.
  • Check Google Cloud permissions available to the container.

Vertex AI is also integrated with TensorFlow Profiler, which you can use to debug model training performance for your custom training jobs. For details, see Profile model training performance using Profiler.

Before you begin

You can use an interactive shell when you perform custom training with a CustomJob resource, a HyperparameterTuningJob resource, or a custom TrainingPipeline resource. As you prepare your training code and configure the custom training resource of your choice, make sure to meet the following requirements:

  • Ensure that your training container has bash installed.

    All prebuilt training containers have bash installed. If you create a custom container for training, use a base container that includes bash or install bash in your Dockerfile.

  • Perform custom training in a region that supports interactive shells.

  • Ensure that anyone who wants to access an interactive shell has the following permissions for the Google Cloud project where custom training is running:

    • aiplatform.customJobs.create
    • aiplatform.customJobs.get
    • aiplatform.customJobs.cancel

    If you initiate custom training yourself, then you most likely already have these permissions and can access an interactive shell. However, if you want to use an interactive shell to inspect a custom training resource created by someone else in your organization, then you might need to obtain these permissions.

    One way to obtain these permissions is to ask an administrator of your organization to grant you the Vertex AI User role (roles/aiplatform.user).

Requirements for advanced cases

If you are using certain advanced features, meet the following additional requirements:

  • If you attach a custom service account to your custom training resource, then make sure that any user who wants to access an interactive shell has the iam.serviceAccounts.actAs permission for the attached service account.

    The guide to custom service accounts notes that you must have this permission to attach a service account. You also need this permission to view an interactive shell during custom training.

    For example, to create a CustomJob with a service account attached, you must have the iam.serviceAccounts.actAs permission for the service account. If one of your colleagues then wants to view an interactive shell for this CustomJob, they must also have the same iam.serviceAccounts.actAs permission.

  • If you have configured your project to use VPC Service Controls with Vertex AI, then account for the following additional limitations:

    • You can't use private IP for custom training.

    • From within an interactive shell, you can't access the public internet or Google Cloud resources outside your service perimeter.

    • To secure access to interactive shells, you must add notebooks.googleapis.com as a restricted service in your service perimeter, in addition to aiplatform.googleapis.com. If you only restrict aiplatform.googleapis.com and not notebooks.googleapis.com, then users can access interactive shells from machines outside the service perimeter, which reduces the security benefit of using VPC Service Controls.

Enable interactive shells

To enable interactive shells for a custom training resource, set the enableWebAccess API field to true when you create a CustomJob, HyperparameterTuningJob, or custom TrainingPipeline.

The following examples show how to do this using several different tools:

Console

Follow the guide to creating a custom TrainingPipeline in the Google Cloud console. In the Train new model pane, when you reach the Model details step, do the following:

  1. Click Advanced options.

  2. Select the Enable training debugging checkbox.

Then, complete the rest of the Train new model workflow.

gcloud

To learn how to use these commands, see the guide to creating a CustomJob and the guide to creating a HyperparameterTuningJob.

API

The following partial REST request bodies show where to specify the enableWebAccess field for each type of custom training resource:

CustomJob

The following example is a partial request body for the projects.locations.customJobs.create API method:

{
  ...
  "jobSpec": {
    ...
    "enableWebAccess": true
  }
  ...
}

For an example of sending an API request to create a CustomJob, see Creating custom training jobs.

HyperparameterTuningJob

The following example is a partial request body for the projects.locations.hyperparameterTuningJobs.create API method:

{
  ...
  "trialJobSpec": {
    ...
    "enableWebAccess": true
  }
  ...
}

For an example of sending an API request to create a HyperparameterTuningJob, see Using hyperparameter tuning.

Custom TrainingPipeline

The following examples show partial request bodies for the projects.locations.trainingPipelines.create API method. Select one of the following tabs, depending on whether you are using hyperparameter tuning:

Without hyperparameter tuning

{
  ...
  "trainingTaskInputs": {
    ...
    "enableWebAccess": true
  }
  ...
}

With hyperparameter tuning

{
  ...
  "trainingTaskInputs": {
    ...
    "trialJobSpec": {
      ...
      "enableWebAccess": true
    }
  }
  ...
}

For an example of sending an API request to create a custom TrainingPipeline, see Creating training pipelines.

Python

To learn how to install or update the Python, see Install the Vertex AI SDK for Python. For more information, see the Python API reference documentation.

Set the enable_web_access parameter to true when you run one of the following methods:

After you have initiated custom training according to the guidance in the preceding section, Vertex AI generates one or more URIs that you can use to access interactive shells. Vertex AI generates a unique URI for each training node in your job.

You can navigate to an interactive shell in one of the following ways:

  • Click a link in the Google Cloud console
  • Use the Vertex AI API to get the shell's web access URI
  1. In the Google Cloud console, in the Vertex AI section, go to one of the following pages:

  2. Click the name of your custom training resource.

    If you created a TrainingPipeline for custom training, click the name of the CustomJob or HyperparameterTuningJob that was created by your TrainingPipeline. For example, if your pipeline has the name PIPELINE_NAME, this might be called PIPELINE_NAME-custom-job or PIPELINE_NAME-hyperparameter-tuning-job.

  3. On the page for your job, click Launch web terminal. If your job uses multiple nodes, click Launch web terminal next to the node for which you want an interactive shell.

    Note that you can only access an interactive shell while the job is running. If you don't see Launch web terminal, this might be because Vertex AI hasn't started running your job yet, or because the job has already finished or failed. If the job's Status is Queued or Pending, wait a minute; then try refreshing the page.

    If you are using hyperparameter tuning, there are separate Launch web terminal links for each trial.

Get the web access URI from the API

Use the projects.locations.customJobs.get API method or the projects.locations.hyperparameterTuningJobs.get API method to see the URIs that you can use to access interactive shells.

Depending on which type of custom training resource you are using, select one of the following tabs to see examples of how to find the webAccessUris API field, which contains an interactive shell URI for each node in your job:

CustomJob

The following tabs show different ways to send a projects.locations.customJobs.get request:

gcloud

Run the gcloud ai custom-jobs describe command:

gcloud ai custom-jobs describe JOB_ID \
  --region=LOCATION \
  --format=json

Replace the following:

  • JOB_ID: The numerical ID of your job. This ID is the last last part of the job's name field. You might have seen the ID when you created the job. (If you don't know your job's ID, you can run the gcloud ai custom-jobs list command and look for the appropriate job.)

  • LOCATION: The region where you created the job.

REST

Before using any of the request data, make the following replacements:

  • LOCATION: The region where you created the job.

  • PROJECT_ID: Your project ID.

  • JOB_ID: The numerical ID of your job. This ID is the last last part of the job's name field. You might have seen the ID when you created the job.

HTTP method and URL:

GET https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/customJobs/JOB_ID

To send your request, expand one of these options:

 

In the output, look for the following:

{
  ...
  "state": "JOB_STATE_RUNNING",
  ...
  "webAccessUris": {
    "workerpool0-0": "INTERACTIVE_SHELL_URI"
  }
}

If you don't see the webAccessUris field, this might be because Vertex AI hasn't started running your job yet. Verify that you see JOB_STATE_RUNNING in the state field. If the state is JOB_STATE_QUEUED or JOB_STATE_PENDING, wait a minute; then try getting the project info again.

HyperparameterTuningJob

The following tabs show different ways to send a projects.locations.hyperparameterTuningJobs.get request:

gcloud

Run the gcloud ai hp-tuning-jobs describe command:

gcloud ai hp-tuning-jobs describe JOB_ID \
  --region=LOCATION \
  --format=json

Replace the following:

  • JOB_ID: The numerical ID of your job. This ID is the last last part of the job's name field. You might have seen the ID when you created the job. (If you don't know your job's ID, you can run the gcloud ai hp-tuning-jobs list command and look for the appropriate job.)

  • LOCATION: The region where you created the job.

REST

Before using any of the request data, make the following replacements:

  • LOCATION: The region where you created the job.

  • PROJECT_ID: Your project ID.

  • JOB_ID: The numerical ID of your job. This ID is the last last part of the job's name field. You might have seen the ID when you created the job.

HTTP method and URL:

GET https://LOCATION-aiplatform.googleapis.com/v1/projects/PROJECT_ID/locations/LOCATION/hyperparameterTuningJobs/JOB_ID

To send your request, expand one of these options:

 

In the output, look for the following:

{
  ...
  "state": "JOB_STATE_RUNNING",
  ...
  "trials": [
    ...
    {
      ...
      "state": "ACTIVE",
      ...
      "webAccessUris": {
        "workerpool0-0": "INTERACTIVE_SHELL_URI"
      }
    }
  ],
}

If you don't see the webAccessUris field, this might be because Vertex AI hasn't started running your job yet. Verify that you see JOB_STATE_RUNNING in the state field. If the state is JOB_STATE_QUEUED or JOB_STATE_PENDING, wait a minute; then try getting the project info again.

Vertex AI provides a set of interactive shell URIs for each hyperparameter tuning trial as the trial enters the ACTIVE state. If you want to get interactive shell URIs for later trials, get the job info again after those trials start.

The preceding example shows the expected output for single-replica training: one URI for the primary training node. If you are performing distributed training, the output contains one URI for each training node, identified by worker pool.

For example, if your job has a primary worker pool with one replica and a secondary worker pool with two replicas, then the webAccessUris field looks similar to the following:

{
  "workerpool0-0": "URI_FOR_PRIMARY",
  "workerpool1-0": "URI_FOR_FIRST_SECONDARY",
  "workerpool1-1": "URI_FOR_SECOND_SECONDARY"
}

Use an interactive shell

To use the interactive shell for a training node, navigate to one of the URIs that you found in the preceding section. A Bash shell appears in your browser, giving you access to the file system of the container where Vertex AI is running your training code.

The following sections describe some things to consider as you use the shell and provide some examples of monitoring tools you might use in the shell.

Prevent the job from ending

When Vertex AI finishes running your job or trial, you will immediately lose access to your interactive shell. If this happens, you might see the message command terminated with exit code 137 or the shell might stop responding. If you created any files in the container's file system, they will not persist after the job ends.

In some cases, you might want to purposefully make your job run longer in order to debug with an interactive shell. For example, you can add code like the following to your training code in order to make the job keep running for at least an hour after an exception occurs:

import time
import traceback

try:
    # Replace with a function that runs your training code
    train_model()
except Exception as e:
    traceback.print_exc()
    time.sleep(60 * 60)  # 1 hour

However, note that you incur Vertex AI training charges as long as the job keeps running.

Check permissions issues

The interactive shell environment is authenticated using application default credentials (ADC) for the service account that Vertex AI uses to run your training code. You can run gcloud auth list in the shell for more details.

In the shell, you can use gsutil, bq, and other tools that support ADC. This can help you verify that the job is able to access a particular Cloud Storage bucket, BigQuery table, or other Google Cloud resource that your training code needs.

Visualize Python execution with py-spy

py-spy lets you profile an executing Python program, without modifying it. To use py-spy in an interactive shell, do the following:

  1. Install py-spy:

    pip3 install py-spy
    
  2. Run ps aux in the shell, and look for the PID of the Python training program.

  3. Run any of the subcommands described in the py-spy documentation, using the PID that you found in the preceding step.

  4. If you use py-spy record to create an SVG file, copy this file to a Cloud Storage bucket so you can view it later on your local computer. For example:

    gsutil cp profile.svg gs://BUCKET
    

    Replace BUCKET with the name of a bucket you have access to.

Analyze performance with perf

perf lets you analyze the performance of your training node. To install the version of perf appropriate for your node's Linux kernel, run the following commands:

apt-get update
apt-get install -y linux-tools-generic
rm /usr/bin/perf
LINUX_TOOLS_VERSION=$(ls /usr/lib/linux-tools | tail -n 1)
ln -s "/usr/lib/linux-tools/${LINUX_TOOLS_VERSION}/perf" /usr/bin/perf

After this, you can run any of the subcommands described in the perf documentation.

Retrieve information about GPU usage

GPU-enabled containers running on nodes with GPUs typically have several command-line tools preinstalled that can help you monitor GPU usage. For example:

  • Use nvidia-smi to monitor GPU utilization of various processes.

  • Use nvprof to collect a variety of GPU profiling information. Since nvprof can't attach to an existing process, you might want to use the tool to start an additional process running your training code. (This means your training code will run twice on the node.) For example:

    nvprof -o prof.nvvp python3 -m MODULE_NAME
    

    Replace MODULE_NAME with the fully-qualified name of your training application's entry point module; for example, trainer.task.

    Then transfer the output file to a Cloud Storage bucket so you can analyze it later on your local computer. For example:

    gsutil cp prof.nvvp gs://BUCKET
    

    Replace BUCKET with the name of a bucket you have access to.

  • If you encounter a GPU error (not a problem with your configuration or with Vertex AI), use nvidia-bug-report.sh to create a bug report.

    Then transfer the report to a Cloud Storage bucket so you can analyze it later on your local computer or send it to NVIDIA. For example:

    gsutil cp nvidia-bug-report.log.gz gs://BUCKET
    

    Replace BUCKET with the name of a bucket you have access to.

If bash can't find any of these NVIDIA commands, try adding /usr/local/nvidia/bin and /usr/local/cuda/bin to the shell's PATH:

export PATH="/usr/local/nvidia/bin:/usr/local/cuda/bin:${PATH}"

What's next