gcloud-ai-platform

name	gcloud-ai-platform
description	AI Platform (legacy) via gcloud (`gcloud ai-platform`). Manage AI Platform jobs and models — jobs, local, models, operations, versions.

gcloud ai-platform — AI Platform (legacy)

Overview

gcloud ai-platform manages AI Platform (formerly Cloud ML Engine): a managed service for training ML models and serving online/batch predictions for TensorFlow, scikit-learn, XGBoost, and custom containers. This is a legacy service that has been largely superseded by Vertex AI (gcloud ai). Google has consolidated training and prediction into Vertex AI; existing AI Platform workloads continue to run, but new projects should use Vertex AI. Reach for gcloud ai-platform only to operate or migrate pre-existing AI Platform models, versions, and jobs — see the migration guide.

The underlying API is ml.googleapis.com. Enable it once per project:

gcloud services enable ml.googleapis.com

Quick reference — common workflows

1. Submit a custom training job and watch it

gcloud ai-platform jobs submit training my_training_job \
    --module-name trainer.task \
    --package-path ./trainer \
    --staging-bucket gs://my-bucket/staging \
    --region us-central1 \
    --runtime-version 2.11 \
    --python-version 3.7 \
    --scale-tier basic

# Stream logs while the job runs
gcloud ai-platform jobs stream-logs my_training_job

# Get a human-readable summary of the finished job
gcloud ai-platform jobs describe my_training_job --summarize

2. Submit a GPU training job (custom scale tier)

gcloud ai-platform jobs submit training my_gpu_job \
    --module-name trainer.task \
    --package-path ./trainer \
    --staging-bucket gs://my-bucket/staging \
    --region us-central1 \
    --runtime-version 2.11 \
    --python-version 3.7 \
    --scale-tier custom \
    --master-machine-type n1-standard-8 \
    --master-accelerator type=nvidia-tesla-v100,count=1

3. Deploy a model for online prediction

# Create the model resource
gcloud ai-platform models create my_model \
    --regions us-central1 \
    --enable-logging

# Create a version pointing to a SavedModel in Cloud Storage
gcloud ai-platform versions create v1 \
    --model my_model \
    --origin gs://my-bucket/model-dir \
    --runtime-version 2.11 \
    --python-version 3.7 \
    --framework tensorflow

# Promote it to the model's default version
gcloud ai-platform versions set-default v1 --model my_model

# Send an online prediction request
gcloud ai-platform predict \
    --model my_model \
    --version v1 \
    --json-instances instances.json \
    --region us-central1

4. Submit a batch prediction job

gcloud ai-platform jobs submit prediction my_batch_job \
    --model my_model \
    --version v1 \
    --data-format text \
    --input-paths gs://my-bucket/input/*.json \
    --output-path gs://my-bucket/output \
    --region us-central1 \
    --runtime-version 2.11

# Monitor it
gcloud ai-platform jobs describe my_batch_job
gcloud ai-platform jobs list --filter="state=RUNNING"

5. Test prediction and training locally (no cloud resources)

# Local online prediction against a SavedModel directory
gcloud ai-platform local predict \
    --model-dir ./model-dir \
    --json-instances instances.json \
    --framework tensorflow

# Local distributed training run
gcloud ai-platform local train \
    --module-name trainer.task \
    --package-path ./trainer \
    --distributed \
    --parameter-server-count 2 \
    --worker-count 4

6. Manage versions and job labels

# List and inspect versions of a model
gcloud ai-platform versions list --model my_model
gcloud ai-platform versions describe v1 --model my_model

# Update a version's description, then retire an old one
gcloud ai-platform versions update v1 --model my_model \
    --description "Production model v1"
gcloud ai-platform versions delete v1 --model my_model

# Label a job, then cancel it if needed
gcloud ai-platform jobs update my_training_job --update-labels env=prod,team=ml
gcloud ai-platform jobs cancel my_training_job

Command groups

Group	Reference file	Commands	Description
ai-platform jobs	jobs.md	7	Submit, list, describe, update, cancel, and stream logs for training and batch-prediction jobs
ai-platform local	local.md	2	Run training and prediction locally for testing/debugging
ai-platform models	models.md	9	Create/list/describe/update/delete models and manage their IAM policy
ai-platform operations	operations.md	4	Describe, list, wait on, and cancel long-running operations
ai-platform versions	versions.md	6	Create, list, describe, update, delete, and set-default model versions

Top-level command (see _commands.md):

• gcloud ai-platform predict — run an online prediction request (reads up to 100 instances; use batch prediction for more).

See index.md for a one-line index of all 29 GA commands.

Common flags & tips

• --region — Models, versions, prediction, and operations accept a regional endpoint. For models/versions/predict, the choices are global plus asia-east1, asia-northeast1, asia-southeast1, australia-southeast1, europe-west1/2/3/4, northamerica-northeast1, us-central1, us-east1, us-east4, us-west1. For operations the same list applies but without global (omitting --region uses the global endpoint). Training jobs instead take a Compute Engine --region and respect gcloud config set compute/region.
• --runtime-version / --python-version — Required for most TensorFlow/ scikit-learn/XGBoost work. They must be a compatible pair; the supported matrix lives at the runtime-version list.
• --scale-tier — Training tier; valid values are basic, basic-gpu, basic-tpu, custom, premium-1, standard-1. With custom you must set --master-machine-type (and may add --worker-count / --parameter-server-count and --*-accelerator).
• --framework — On versions create and local predict, one of scikit-learn, tensorflow (default), xgboost.
• --model is required on every versions subcommand and on predict.
• Custom containers: pass --master-image-uri (and optionally --worker-image-uri, --parameter-server-image-uri) to jobs submit training instead of --runtime-version.
• Async vs. blocking: jobs submit training defaults to async — add --stream-logs to follow it, or --async on versions create to return immediately. Use operations wait OPERATION to block on a long-running operation.
• Filtering/formatting: list commands support --filter, --sort-by, --limit, and --uri. Useful patterns:
  • gcloud ai-platform jobs list --filter="state=RUNNING"
  • gcloud ai-platform models list --uri --filter='name ~ vision[0-9]+' (pipe to xargs gcloud ai-platform models delete for bulk cleanup)

beta / alpha

• gcloud beta ai-platform adds explain — run an explanation request (feature attributions via Explainable AI). All GA subgroups are also present in beta.
• gcloud alpha ai-platform adds both explain (as in beta) and locations — query AI Platform location/region capabilities (not available in GA or beta).

Use GA for all standard workflows; reach for beta only for explain, and alpha only to query location capabilities via locations.

Official documentation

• AI Platform docs home — product landing page for legacy AI Platform (now redirects to Vertex AI, reflecting that it has been superseded).
• Migrating to Vertex AI — official guide for moving AI Platform workloads to gcloud ai.
• AI Platform Training overview — concepts for submitting and configuring training jobs.
• AI Platform Prediction overview — concepts for online and batch prediction.
• Deploying models — how to upload a SavedModel and create a serving version.
• Managing models and jobs — managing models/versions via gcloud and REST.
• Runtime version list — supported runtime + Python compatibility matrix.
• Regional endpoints — the global and regional endpoints accepted by --region.
• AI Platform IAM roles — predefined roles/ml.* roles (admin, developer, viewer, jobOwner, modelOwner, …).
• gcloud ai-platform CLI reference — full command/flag reference for all GA commands.