Vertex AI release notes

Vector Search hybrid search and sparse embeddings are generally available (GA)

Vector Search hybrid search and sparse embeddings are generally available (GA). Hybrid search uses both dense and sparse embeddings, which lets you search based on a combination of keyword search and semantic search. For more information about hybrid search, see About hybrid search.

You can copy tuned Gemini 1.5 Pro 002 and Gemini 1.5 Flash 002 adapter models across projects. For details, see Copy a model in Vertex AI Model Registry.

Preview: You can consume reservations of VMs that have GPUs attached with your custom training jobs or prediction jobs. Reservations of Compute Engine zonal resources help you gain a high level of assurance that your jobs have the necessary resources to run. For more information, see the following:

• Use reservations with training

• Use reservations with prediction

PSC-I Egress is supported for Ray clusters Vertex AI. PSC-I is recommended for private connectivity since it reduces the chance of IP exhaustion, and allows for transitive peering. Check out Private Service Connect interface for Ray on Vertex AI. This feature is available in Preview.

Private Service Connect interface (PSC-I) is now supported for ML pipeline runs in Vertex AI Pipelines. PSC-I is recommended for private connectivity, since it reduces the chance of IP exhaustion, and allows for transitive peering.

For more information, see Configure Private Service Connect interface for a pipeline. This feature is available in Preview.

Vector Search Private Service Connect automation

Deploying an index with Private Service Connect automation is generally available (GA). You can set up a service connection policy so that you don't have to manually create a compute address and forwarding rule after each index deployment.

For more information, see Set up Vector Search with Private Service Connect.

To ensure that VM resources are available when your custom training and prediction jobs need them, you can now use Compute Engine reservations. Reservations provide a high level of assurance in obtaining capacity for Compute Engine resources. This feature is available in Preview for A2 and A3 machine series reservations.

For more information, see Use reservations with training and Use reservations with prediction.

To reduce the cost of running your training and prediction jobs, you can now use Spot VMs. Spot VMs are virtual machine (VM) instances that are excess Compute Engine capacity. Spot VMs have significant discounts, but Compute Engine might preemptively stop or delete Spot VMs to reclaim the capacity at any time. This feature is available in Preview.

For more information, see Use Spot VMs with training and Use Spot VMs with prediction.

Schedule Vertex AI custom training jobs based on resource availability. For details, see the Vertex AI documentation.

Ray cluster's autoscaling feature is now supported. See Scale Ray clusters on Vertex AI

The Vertex AI Model Registry now offers Preview support for model copy across different projects. For information about how to copy your model projects and regions, see Copy models in Model Registry.

Model Monitoring v2 is in Preview, which centralizes model monitoring configuration and visualization on a model version and enables monitoring models being served outside of Vertex AI. For more information, see Vertex AI Model Monitoring overview.

Vertex AI custom training supports TPU v5e. For details, see Training with TPU accelerators.

Vertex AI Feature Store

The following features of Vertex AI Feature Store are now generally available (GA):

• Optimized online serving: Serve features at ultra-low latencies. For more information, see Optimized online serving.

• Search using embeddings: Perform vector similarity searches to retrieve semantically similar or related features for real-time serving. You can search using embeddings if your online store is configured to support embeddings. For more information, see Search using embeddings.

• Feature view sync: Refresh or synchronize the feature data in a feature view within an online store from the feature data source in BigQuery. For more information, see Sync feature data to online store.

Vertex AI Feature Store

The following features of Vertex AI Feature Store are now available in Preview:

• Integration of Vertex AI Feature Store with Dataplex: Online store instances, feature views, and feature groups are now automatically registered as data assets in Data Catalog, a Dataplex feature that catalogs metadata from these resources. You can use the metadata search capability of Dataplex to search for and view the metadata of these resources. For more information, see Search for resource metadata in Data Catalog.

• Service account configuration for feature views: You can configure a feature view to use a dedicated service account. By default, every feature view uses the service account configured for your project. For more information, see Configure the service account for a feature view.

• Multiple entity IDs for a feature view: While creating or updating a feature view, you can specify multiple entity ID columns. For more information, see Create a feature view.

Create an empty index with Vector Search

You can create an empty index in Vector Search for batch and for streaming. No embedding data is required at index creation time, which enables faster startup time. To learn more, see Manage indexes.

Vertex AI Prediction

You can now use A3 machine types to serve predictions.

Vector Search feature launch

Update streaming index metadata: With this launch, you can directly update restricts and numeric restricts of data points inside StreamUpdate indexes without the compaction cost of a full update. To learn more, see Update dynamic metadata.

Structured logging support for Vertex AI custom training. For details, see Write code to return container logs.

Ground Multimodal Models

Model grounding for gemini-pro is available in Preview. Use grounding to connect the gemini-pro model to unstructured text data stores in Vertex AI Search. Grounding lets models access and use the information in the data repositories to generate more enhanced and nuanced responses. For more information, see Ground multimodal models.

Gemma open models are available

Gemma models, a family of lightweight, open models built from the same research and technology used to create the Gemini models, are available to run on your hardware, mobile devices, or hosted services. To learn more, see Use Gemma open models and the Gemma Model Garden card.

Multimodal embeddings video support is generally available

Embeddings for video data is now generally available (GA) using the multimodal embedding model (multimodalembedding). For more information, see the product documentation.

This features incurs pricing based on the mode you use. For more information, see pricing.

The following models have been added to Model Garden:

• Stable Diffusion XL LCM: The Latent Consistency Model (LCM) enhances text-to-image generation in Latent Diffusion Models by enabling faster and high-quality image creation with fewer steps.
• LLaVA 1.5: Deploy LLaVA 1.5 models.
• PyTorch-ZipNeRF: The Pytorch-ZipNeRF model is a state-of-the-art implementation of the ZipNeRF algorithm in the Pytorch framework, designed for efficient and accurate 3D reconstruction from 2D images.
• LLaMA 2 (Quantized): A quantized version of Meta's Llama 2 models.
• WizardLM: WizardLM is a large language model (LLM) developed by Microsoft, fine-tuned on complex instructions by adapting the Evol-Instruct method.
• WizardCoder: WizardCoder is a large language model (LLM) developed by Microsoft, fine-tuned on complex instructions by adapting the Evol-Instruct method to the domain of code.
• AutoGluon: With AutoGluon you can train and deploy high-accuracy machine learning and deep learning models for tabular data.
• Lama (Large mask inpainting): Use Large Mask Inpainting with fast Fourier convolutions (FFCs), a high receptive field perceptual loss, and large training masks for resolution-robust image inpainting.

Query an index from the Vector Search console

Vector Search has launched an improved console experience for querying both private and public deployed indexes, now available in Preview. From the console, you can create an index and endpoint, deploy the index to the endpoint, and query the index for nearest neighbors. For more information, see Manage indexes.

Vertex Prediction

You can now customize more deployment parameters when uploading your models, such as shared memory allocation and custom startup and readiness probes. These parameters may be useful when deploying LLMs.

For more information, see Deploy generative AI models, Custom container requirements for prediction, and ModelContainerSpec.

Model tuning for the textembedding-gecko and textembedding-gecko-multilingual models is available in GA. You can use supervised fine-tuning to tune the textembedding-gecko model. For more information, see Tune text embeddings.

Vertex AI Prediction

You can now use Cloud TPU v5e to serve online predictions. For more information, see Use Cloud TPUs for online prediction.

Vertex AI Gemini models

Vertex AI Gemini Pro and Gemini Pro Vision multimodal language models are available in Preview. For more information, see the following topics:

• Overview of multimodal models
• Multimodal prompt design
• Vertex AI Gemini API reference
• Gemini Python SDK reference
• Migrate from PaLM API to Gemini API

Imagen 2 General Availability

The 005 version of Imagen's image generation model (imagegeneration@005) is now generally available (GA) for image generation tasks. This model version is now the default for image generation tasks. For more information, see the product documentation.

For general information about Imagen models and versions, see Imagen model versions and lifecycle.

Text embedding model 003 (textembedding-gecko@003) available

The updated stable version of the text embedding foundation model, textembedding-gecko@003, is available. textembedding-gecko@003 features improved quality compared to the previous stable versions, textembedding-gecko@001 and textembedding-gecko@002. For more information on model versions, see Model versions and lifecycle.

Version @002 of the models for text, chat, code, and code chat are available. The @002 model versions include improved prompt responses. The @002 models are:

• text-bison@002
• chat-bison@002
• code-bison@002
• codechat-bison@002

To ensure that you always use the stable model version, specify the model identifier with the version number. For example, text-bison@002. For more information, see Model versions and lifecycle.

Version 2 of the stable version of the Codey code completion foundation model, named code-gecko@002, is available. code-gecko@002 features improved quality and reduced latency compared to the previous stable version, code-gecko@001. These improvements can lead to a higher rate of acceptance.

Grounding with Vertex AI Search

Model grounding is available in (Preview). Use grounding to connect the text-bison and chat-bison models to unstructured data stores in Vertex AI Search. Grounding lets models access and use the information in the data repositories to generate more enhanced and nuanced responses. For more information, see the Grounding Overview.

The Unicorn model size for PaLM 2 for Text is generally available (GA). The text-unicorn model provides improved response quality and reasoning capability compared to the text-bison model. For details, see Model information.

ComputeToken API now available in Preview

The ComputeToken API is now available in (Preview). You can use this API to get a list of tokens for a given prompt. A token is a way to represent a common sequence of characters found in a text input. To learn more, see Get a list of tokens.

Vertex AI Feature Store

The following features of the new and improved Vertex AI Feature Store are now generally available (GA):

• Feature Registry: Register your feature data sources in BigQuery by creating feature groups and features. For more information, see Create a feature group and Create a feature.

• Cloud Bigtable online serving: Serve features from one or more BigQuery data sources. You can set up Cloud Bigtable online serving by defining online serving clusters called online store instances and creating feature views within the online store instances.

Note that the following features of Vertex AI Feature Store are still in Preview:

• Serve features at ultra-low latencies with Optimized online serving.
• Sync data in a feature view within an online store.
• Retrieve vector embeddings for real-time serving.

For more information, see About Vertex AI Feature Store.

Numerical filtering available in Vertex AI Vector Search

With Vector Search you can restrict results by "filtering" your index results. In addition to filtering by using categorical restrictions, you can now use numeric filtering. To learn more, see Filter vector matches.

Generative AI on Vertex AI

Security controls are available for additional Generative AI on Vertex AI features.

The following models have been added to Model Garden:

• ImageBind: Multimodal embedding model.
• Vicuna v1.5: LLM finetuned based on llama2.
• OWL-ViT v2: SoTA Open Vocabulary Object Detection model.
• DITO: SoTA Open Vocabulary Object Detection model.
• NLLB: Multi-language translation model.
• Mistral-7B: SoTA LLM at small size.
• BioGPT: LLM finetuned for biomedical domain.
• BiomedCILP: Multimodal foundational model finetuned for biomedical domain.

To see a list of all available models, see Explore models in Model Garden.

New Vertex AI Vector Search Console

Vector Search has launched a console experience in Google Cloud for creating and deploying indexes, now available in Preview. From the console, you can create indexes, and create public or VPC endpoints for your indexes, and deploy. For more information, see Manage indexes.

Ray on Vertex AI is now available in Preview

Ray is an open-source framework for scaling AI and Python applications. Ray provides the infrastructure to perform distributed computing and parallel processing for your machine learning workflow.

You can now create Ray clusters and develop your Ray applications on Vertex AI. This feature is in Preview. For more information, see Ray on Vertex AI overview.

Model tuning for the textembedding-gecko model is now available in Preview

You can now use supervised fine-tuning to tune the textembedding-gecko model. This feature is in (Preview). For more information, see Tune text embeddings.

Vertex AI Prediction

You can now use C3 machine types to serve predictions.

Vertex AI Feature Store

The new and improved Vertex AI Feature Store is now available in Preview. With the new Vertex AI Feature Store you can streamline your feature management in the following ways:

• Store and maintain your offline feature data in BigQuery, taking advantage of the data management capabilities of BigQuery. In the new Vertex AI Feature Store, BigQuery serves as the offline store. You don't need to copy or import feature data to an offline store in Vertex AI.

• Register your feature data sources in BigQuery by creating feature groups and features.

• Define online serving clusters called online store instances; and then serve features from one or more BigQuery data sources, by aggregating them in a feature view within an online store instance. Use Optimized online serving for ultra-low latency needs and Cloud Bigtable online serving for high data volumes.

• Retrieve vector embeddings stored in BigQuery for real-time serving.

For more information, see About Vertex AI Feature Store.

Vertex AI Prediction

You can now use A2 Ultra machines to serve predictions in us-central1, us-east4, europe-west4, and asia-southeast1. Each A2 Ultra machine has a fixed number of NVIDIA A100 80GB GPUs attached.

Experiment management: Google Cloud console now supports visualization of your model's performance changes over steps during training, and shows advanced run comparisons. To learn more, see Compare and analyze runs: Google Cloud console.

Imagen on Vertex AI now offers the following Generally Available (GA) features:

• Subject model tuning (standard)*
• Style model tuning*

* Restricted access feature.

For more information about Imagen or how to get access to restricted GA features, see the Imagen on Vertex AI overview.

Stream responses from Generative AI models

Generative AI model streaming support is now Generally Available (GA). After you send a prompt, the model returns response tokens as they're generated instead of waiting for the entire output to be available.

Supported models are:

• text-bison
• chat-bison
• code-bison
• codechat-bison

To learn more, see Stream responses from Generative AI models.

New Generative AI support on Vertex AI models and expanded language support

Generative AI support on Vertex AI has been updated to include new language model candidates (latest models), language models that support input and output tokens up to 32k, and more supported languages.

For details, see Available models and Model versions and lifecycle.

Model tuning for the text-bison model is now Generally Available (GA)

Tuning the text-bison model with supervised fine-tuning (SFT) is now Generally Available (GA) . For more information, see Tune text models.

Model tuning for the chat-bison model is now available in Preview

You can now use supervised fine-tuning to tune the chat-bison model. This feature is in (Preview). For more information, see Tune text models.

New embedding model now available in Preview

Generative AI support on Vertex AI users can now create embeddings using a new model trained on a wide range of non-English languages in (Preview).

• textembedding-gecko-multilingual

To learn more, see Get text embeddings.

Reinforcement learning from human feedback (RLHF) tuning for text-bison The Generative AI text generation foundation model (text-bison) now supports RLHF tuning. The RLHF tuning feature is in (Preview). For more information, see Use RLHF model tuning.

Vertex AI Codey APIs language support

Vertex AI Codey APIs now support additional programming languages. For more information, see Supported coding languages

Vertex AI Codey APIs now support supervised fine-tuning (SFT)

The code chat (codechat-bison) and code generation (code-bison) Vertex AI Codey APIs models now support supervised fine-tuning (SFT). The supervised-fine tuning for Vertex AI Codey APIs models feature is in (Preview). For more information, see Tune code models.

Metrics-based model evaluation

You can evaluate the performance of foundation models and tuned models against an evaluation dataset for classification, summarization, question answering, and general text generation. This feature is available in (Preview).

To learn more, see Evaluate model performance.

Vertex AI Model Registry Models and managed datasets are now synced to Dataplex's Data Catalog service. Data Catalog enables organization-wide search and discovery of data artifacts, while still maintaining IAM boundaries. The sync and search of these assets is available in Preview. For more information, see Data Catalog documentation.

CountToken API now available in Preview

The CountToken API is now available in (Preview). You can use this API to get the token count and the number of billable characters for a prompt. To learn more, see Get token count.

The Vertex AI Pipelines Template Gallery is now generally available (GA). The Template Gallery contains Google-authored pipeline and component templates to bootstrap your MLOps practice. Customize and run the templates as-is or embed them into your own pipelines. For more information, see Use a prebuilt template from the Template Gallery.

Tabular Workflow for Forecasting is available in Preview. For documentation, refer to Tabular Workflow for Forecasting.

Vertex AI custom training has launched persistent resources in Preview. A persistent resource is a long-running cluster of machines that you can use to run custom training jobs. Once created, the persistent resource remains available for future training jobs, so you don't have to wait for compute resources to be provisioned each time you want to train a model.

The Vertex AI Matching Engine public endpoint is now generally available (GA). For information about how to get started, see Matching Engine Setup.

Imagen Multimodal embeddings available in GA

Imagen on Vertex AI now offers the following GA feature:

• Multimodal embeddings

This feature incurs different pricing based on if you use image input or text input. For more information, see the multimodal embeddings feature page.

The schedules API for Vertex AI Pipelines is now generally available (GA). You can schedule recurring pipeline runs in Vertex AI by specifying a frequency in cron syntax, and optionally the start time and/or end time. Additionally, you can pause, resume, update, and delete schedules.For more information, see Schedule a pipeline run with scheduler API.

Model tuning updates for text-bison:

• Upgraded tuning pipeline now offers more efficient tuning and better performance on text-bison.
• New learning_rate parameter lets you adjust the step size at each iteration.

For details, see Tune language foundation models.

Imagen on Vertex AI now offers the following Generally Available (GA) features:

• Image generation (text-to-image generation)*
• Image editing*
• Image visual captioning
• Visual Question Answering (VQA)

* Restricted access feature.

For more information about Imagen or how to get access to restricted GA or Preview features, see the Imagen on Vertex AI overview.

Imagen now supports human face generation for the following features:

• Image generation (text-to-image generation)*
• Image editing*

* Restricted access feature.

Human face generation is enabled by default, except for images with children and/or celebrities. For more information, see the usage guidelines.

The Vertex AI PaLM API has added support for the following languages:

• Spanish (es)
• Korean (ko)
• Hindi (hi)
• Chinese (zh)

For the complete list of supported languages, see Supported languages.

Support for batch text (text-bison) requests is now available in (GA). You can review pricing for the chat-bison model at Vertex AI pricing page.

Support for PaLM 2 for Chat (chat-bison) is now available in (GA). You can review pricing for the chat-bison model at Vertex AI pricing page.

Vertex AI model evaluation is now generally available (GA) with the following new Preview features:

• Model evaluation with sliced metrics.
• Model evaluation with fairness and bias metrics.
• Vision error analysis for AutoML image classification models.

Vertex Explainable AI

Support for example-based explanations is now generally available (GA).

Vertex AI Codey APIs

The Vertex AI Codey APIs are now generally available (GA). Use the Codey APIs to create solutions with code generation, code completion, and code chat. Because the Vertex AI Codey APIs are GA, you incur usage costs if you use them. To learn about pricing, see the Generative AI support on Vertex AI pricing page.

The models in this release include:

• code-bison (code generation)
• codechat-bison (multi-turn code chat)
• code-gecko (code completion)

The maximum tokens for input was increased from 4,096 to 6,144 tokens for code-bison and codechat-bison to allow longer prompts and chat history. The maximum tokens for output was increased from 1,024 to 2,048 for code-bison and codechat-bison to allow for longer responses.

Additional programming languages are supported. For more information, see Supported coding languages.

Several fine-tuning datasets were removed from the code-bison and codechat-bison models to implement the following improvements:

• Excessive chattiness.
• Artifacting, such as NBSP (non-breaking space) characters.
• Low quality code responses.

To learn about cloud horizontals, please see Vertex AI certifications.

Vertex AI Pipeline task-level logs are now generally available (GA) in Cloud Logging. Additionally, from Cloud Logging you can route pipeline logs to a Pub/Sub sink to power your event-driven architecture. For more information, see View pipeline job logs.

A100 80GB accelerators are now generally available (GA) for custom training jobs in the following regions:

• asia-southeast1
• europe-west4
• us-central1
• us-east4

For more information, see Locations.

The chat-bison@001 model has been updated to better follow instructions in the context field. For details, on how to create chat prompts for chat-bison@001, see Design chat prompts.

PaLM Text and Embeddings APIs, and Generative AI Studio

The Generative AI support on Vertex AI is now generally available (GA). With this feature launch, you can leverage the PaLM API to generate AI models that you can test, tune, and deploy in your AI-powered applications. With the GA of these features, you will incur usage costs if you use the text-bison and textembedding-gecko PaLM APIs. To learn about pricing, see the Vertex AI pricing page.

Features and models in this release include:

• PaLM 2 for Text: text-bison
• Embedding for Text: textembedding-gecko
• Generative AI Studio for Language

Vertex AI Model Garden

The Vertex AI Model Garden is now generally available (GA). The Model Garden is a platform that helps you discover, test, customize, and deploy Vertex AI and select OSS models. These models range from tunable to task-specific - all available on the Model Garden page in the Google Cloud console.

To get started, see Explore AI models and APIs in Model Garden.

Vertex AI Codey APIs

The Vertex AI Codey APIs are now in Preview. With the Codey API, code generation, code completion, and code chat APIs can be used from any Google Cloud project without allowlisting. The APIs can be accessed from the us-central1 region. The Codey APIs can be used in the Generative AI studio or programmatically in REST commands.

To get started, see the Code models overview.

Vertex Prediction

You can now specify a multi-region BigQuery table as the input or output to a batch prediction request.

Vertex Prediction

You can now co-host models on the same VM from the Google Cloud Console. Previously, this capability was available only from the REST API. For more information, see Share resources across deployments.

Vertex AI custom training now supports deep integration with Vertex AI Experiments. You can submit training jobs with autologging enabled to automatically log parameters and model performance metrics. For more information, see Run training job with experiment tracking

The scheduler API for Vertex AI Pipelines is now available in Preview. You can schedule recurring pipeline runs in Vertex AI by specifying a frequency, start time (optional), and end time (optional). For more information, see Schedule a pipeline run with scheduler API.

Generative AI Support for Vertex AI

Generative AI Support for Vertex AI is now available in Preview. With this feature launch, you can leverage the Vertex AI PaLM API to generate AI models that you can test, tune, and deploy in your AI-powered applications.

Features and models in this release include:

• PaLM 2 for Text: text-bison@001
• PaLM 2 for Chat: chat-bison@001
• Embedding for Text: textembedding-gecko@001
• Generative AI Studio for Language
• Tuning for PaLM 2
• Vertex AI SDK v1.25, which includes new features such as TextGenerationModel(text-bison@001), ChatModel(chat-bison@001), TextEmbeddingModel(textembedding-gecko@001)

You can interact with the generative AI features on Vertex AI by using Generative AI Studio in the Google Cloud console, the Vertex AI API, and the Vertex AI SDK for Python.

• Learn more about Generative AI Support for Vertex AI
• See an Introduction to Generative AI Studio
• Get started with a Generative AI Studio quickstart

Vertex AI Model Garden

The Vertex AI Model Garden is now available in Preview. The Model Garden is a platform that helps you discover, test, customize, and deploy Vertex AI and select OSS models. These models range from tunable to task-specific - all available on the Model Garden page in the Google Cloud console.

• To get started, see Explore AI models and APIs in Model Garden.

Vertex AI Prediction

You can now use G2 accelerator-optimized machine types to serve predictions. Each G2 machine has a fixed number of NVIDIA L4 GPUs attached.

Vertex AI Prediction

You can now update some scaling and container logging configuration settings on a DeployedModel without undeploying and redeploying it to an endpoint.

For more information, see update the scaling configuration and container logging.

The Vertex AI Matching Engine service now offers Preview support for deploying an index to a public endpoint. For information about how to get started, see Matching Engine Setup.

Vertex AI Prediction

You can now view logs for Vertex AI Batch Prediction jobs in Cloud Logging.

Vertex AI Pipelines is now integrated with Cloud Asset Inventory service. You can use Cloud Asset Inventory to search, export, monitor, and analyze pipeline resources and metadata, and also view the resource history.

The Vertex AI Model Registry now offers Preview support for model copy between regions. For information about how to copy your model between regions, see Copy models in Model Registry.

Vertex AI Pipelines cost showback with billing labels is now generally available (GA). You can now use billing labels to review the cost of a pipeline run, along with the cost of individual resources generated from Google Cloud Pipeline Components in the pipeline run. For more information, see Understand pipeline run costs.

Vertex AI supports running Explainable AI on certain types of BQML models when they are added to the Vertex AI Model Registry (GA). To learn more, see Explainable AI for BigQuery ML models.

Vertex AI Feature Store

The ability to delete feature values from an entity type is now generally available (GA). The following features are available:

• Delete feature values from specified entities
• Delete feature values from specified features within a time range

Links to additional resources:

• Learn more about the delete_feature_values method in the Vertex AI SDK for Python.
• View the Python code sample on GitHub
• View the Python (Async) code sample on GitHub

Vertex AI Prediction

You can now use N2, N2D, C2, and C2D machine types to serve predictions.

Vertex AI Prediction

Pre-built PyTorch containers for serving predictions from PyTorch models is generally available (GA).

Vertex AI Matching Engine now supports Private Service Connect in Preview. To learn how to set up a a Private Service Connect instance, see Using Private Service Connect.

Support for resource-level IAM policies for Vertex AI featurestore and entityType resources is generally available (GA). For more information, see Control access to resources.

The Vertex AI Pipelines Template Gallery is now available in Preview. You can bootstrap your MLOps workflows with Google-authored pipeline and component templates. For more information, see Use a prebuilt template from the Template Gallery.

Tabular Workflow for End-to-End AutoML is generally available (GA). For documentation, refer to Tabular Workflow for End-to-End AutoML.

Vertex AI TensorFlow Profiler

Vertex AI TensorFlow Profiler is generally available GA. You can use TensorFlow Profiler to debug model training performance for your custom training jobs.

For details, see Profile model training performance using Profiler.

Vertex AI Matching Engine

Vertex AI Matching Engine now offers General Availability support for updating your indices using Streaming Update, which is real-time indexing for the Approximate Nearest Neighbor (ANN) service.

Vertex AI Feature Store streaming ingestion is now generally available (GA).

The Pipeline Templates feature is now generally available (GA). The Your Templates tab is supported by Artifact Registry and allows you to publish and curate pipeline and component templatess. For documentation, refer to Create, upload, and use a pipeline template.

AutoML image model updates

AutoML image classification and object detection now support a higher-accuracy model type. This model is available in Preview.

For information about how to train a model using the higher accuracy model type, see Begin AutoML model training.

Batch prediction is currently not supported for this model type.

Cloud Logging for Vertex AI Pipelines is now generally available (GA). For more information, see View pipeline job logs.

Vertex AI Prediction

You can now perform some simple filtering and transformation on the batch input in your BatchPredictionJob requests without having to write any code in the prediction container. This feature is in Preview. For more information, see Filter and transform input data.

The Vertex AI Pipelines email notification component is now generally available (GA). This component enables you to configure your pipeline to send up to three emails upon success or failure of a pipeline run. For more information, see Configure email notifications and the Email notification component.

Vertex AI has added support for the following regions:

• us-west3 (Salt Lake City)
• europe-central2 (Warsaw)
• asia-southeast2 (Jakarta)
• me-west1 (Tel aviv)

Some features of Vertex AI are not supported in these regions. Check feature availability for all regions on the Vertex AI Locations page.

AutoML Image Classification Error Analysis

Error analysis allows you to examine error cases after training a model from within the model evaluation page. This feature is available in Preview.

For each image you can inspect similar images from the training set to help identify the following:

• Label inconsistencies between visually similar images
• Outliers if a test sample has no visually similar images in the training set

After fixing any data issues, you can retrain the model to improve model performance.

Feature Transform Engine is available in Preview. For documentation, refer to Feature engineering.

Vertex AI Prediction

You can now use A2 machine types to serve predictions.

Vertex ML Metadata

You can now filter contexts, executions, and artifacts by association and attribution.

Custom training on Vertex AI now supports NVIDIA A100 80GB GPUs on a2-ultragpu-1g/2g/4g/8g machines. For details, see Configure compute resources for custom training.

Vertex AI Prediction

Custom prediction routines (CPR) are now Generally Available. CPR lets you easily build custom containers for prediction with pre/post processing support.

Vertex AI Prediction

You can now use E2 machine types to serve predictions.

Tabular Workflow for TabNet Training is available in Preview. For documentation, refer to Tabular Workflows for TabNet Training.

Tabular Workflow for Wide & Deep Training is available in Preview. For documentation, refer to Tabular Workflow for Wide & Deep Training.

Vertex AI Feature Store streaming ingestion is available in Preview.

The Vertex AI Model Registry is generally available (GA). Vertex AI Model Registry is a searchable repository where you can manage the lifecycle of your ML models. From the Vertex AI Model Registry, you can better organize your models, train new versions, and deploy directly to endpoints.

The Vertex AI Model Registry and BigQuery ML integration is generally available (GA). With this integration, BigQuery ML models can be managed alongside other ML models in Vertex AI to easily version, evaluate, and deploy for prediction.

Incrementally train an AutoML model

You can now incrementally train an AutoML image classification or object detection model by selecting a previously trained model. This feature is in Preview. For more information, see Train an AutoML image classification model.

Vertex AI Feature Store

The ability to delete feature values from an entity type is now available in Preview. The following features are available:

• Delete feature values from specified entities
• Delete feature values from specified features within a time range

Links to additional resources:

• Learn more about the delete_feature_values method in the Vertex AI SDK for Python.
• View the Python code sample on GitHub
• View the Python (Async) code sample on GitHub

Vertex AI model evaluation is now available in Preview. Model evaluation provides model evaluation metrics, such as precision and recall, to help you determine the performance of your models.

Vertex AI Model Monitoring

Vertex AI Model Monitoring now offers Preview support for batch prediction jobs. For more details, see Vertex AI Model Monitoring for batch predictions.

Vertex AI Feature Store

Feature value monitoring is now generally available (GA).

Vertex AI Matching Engine

Vertex AI Matching Engine now offers Preview support for updating your indices using Streaming Update, which is real-time indexing for the Approximate Nearest Neighbor (ANN) service.

The option to configure pipeline run caching (enable_caching) is now available in the Cloud console.

You can now limit the number of concurrent or parallel task runs in a pipeline run using dsl.ParallelFor. For more information, see the Kubeflow Pipelines SDK Documentation.

Vertex Explainable AI

Vertex Explainable AI now offers Preview support for example-based explanations. For more information, see Configure example-based explanations for custom training.

TensorFlow Profiler integration: Debug model training performance for your custom training jobs. For details, see Profile model training performance using Profiler.

Vertex AI now offers Preview support for Custom prediction routines (CPR). CPR lets you easily build custom containers for prediction with pre/post processing support.

NFS support for custom training is GA. For details, see Mount an NFS share for custom training.

The Pipeline Templates feature is available in Preview. For documentation, refer to Create, upload, and use a pipeline template.

The features supported by pipeline templates include the following:

• Create a template registry using Artifact Registry (AR).
• Compile and publish a pipeline template.
• Create a pipeline run using the template and filter the runs.
• Manage (create, update, or delete) the pipeline template resources.

Vertex AI Pipelines now lets you configure task-level retries. You can set the number of times a task is retried before it fails. For more information about this option, see the Kubeflow Pipelines SDK Documentation.

Tabular Workflows is available in Preview. For documentation, refer to Tabular Workflows on Vertex AI.

End-to-End AutoML workflow is available in Public Preview. For documentation, refer to End-to-End AutoML.

Feature: Vertex AI Experiments is generally available (GA). Vertex AI Experiments helps users track and compare multiple experiment runs and analyze key model metrics.

Features supported by Experiments include:

• Vary and track parameters and metrics.
• Compare parameters, metrics, and artifacts between pipeline runs.
• Track steps and artifacts to capture the lineage of experiments.
• Compare vertex pipelines against Notebook experiments.

Vertex AI Forecasting is available in GA. The following features are available:

• Forecasting with ARIMA+
• Explainable AI for Forecasting
• Hierarchical Forecasting
• Model training with AutoML or Seq2Seq+
• Forecasting window configuration during model training
• Holiday effect modeling during model training

Support for IAM resource-level policies for Vertex AI featurestore and entityType resources is available in Preview.

You can now configure the failure policy for a pipeline run.

The ability to configure Vertex AI private endpoints is now general available (GA). Vertex AI private endpoints provide a low-latency, secure connection to the Vertex AI online prediction service. You can configure Vertex AI private endpoints by using VPC Network Peering. For more information, see Use private endpoints for online prediction.

You can now train your custom models using Cloud TPU Architecture (TPU VMs).

You can now use a pre-built container to perform custom training with PyTorch 1.11.

Vertex AI Model Registry is available in Preview. Vertex AI Model Registry is a searchable repository where you can manage the lifecycle of your ML models. From the Vertex AI Model Registry, you can better organize your models, train new versions, and deploy directly to endpoints.

Vertex AI Feature Store online store autoscaling is available in Preview. The online store nodes automatically scale to balance performance and cost with different traffic patterns. The offline store already scales automatically.

You can now mount Network File System (NFS) shares to access remote files when you run a custom training job. For more information, see Mount an NFS share for custom training.

This feature is in Preview.

Google Cloud Pipeline Components SDK v1.0 is now generally available.

You can now use a pre-built container to perform custom training with TensorFlow 2.8.

You can now use a pre-built container to perform custom training with PyTorch 1.10.

You can now use a pre-built container to perform custom training with TensorFlow 2.7.

Vertex AI TensorBoard is generally available (GA).

The autopackaging feature of the gcloud ai custom-jobs create command is generally available (GA). Autopackaging lets you use a single command to run code on your local computer as a custom training job in Vertex AI.

The gcloud ai customs-jobs local-run command is generally available (GA). You can use this command to containerize and run training code locally.

Vertex AI Pipelines is generally available (GA).

Using interactive shells to inspect custom training jobs is generally available (GA).

You can use these interactive shells with VPC Service Controls.

Vertex ML Metadata is generally available (GA).

Vertex Feature Store is generally available (GA).

Vertex Matching Engine is generally available (GA).

Vertex AI Vizier is generally available (GA).

Vertex Explainable AI is generally available (GA).

You can use a pre-built container to serve predictions from TensorFlow 2.6 models.

Vertex Model Monitoring is generally available (GA).

When you perform custom training, you can access Cloud Storage buckets by reading and writing to the local filesystem. This feature, based on Cloud Storage Fuse, is available in Preview.

You can now use a pre-built container to perform custom training with TensorFlow 2.6 and PyTorch 1.9.

The following tools for creating embeddings to use with Vertex Matching Engine are available in Preview:

• the Two Tower built-in algorithm
• the Swivel pipeline template

You can use the Reduction Server algorithm (Preview) to increase throughput and reduce latency during distributed custom training.

The following features are generally available (GA):

• Access Transparency for Vertex AI
• Using a custom service account for custom training and prediction
• Using VPC Service Controls with Vertex AI
• Setting up VPC Network Peering with Vertex AI and using private IP for custom training (Using private IP for prediction and vector matching with Matching Engine remains in preview.)

Private endpoints for online prediction are now available in preview. After you set up VPC Network Peering with Vertex AI, you can create private endpoints for low-latency online prediction within your private network.

Additionally, the documentation for VPC Network Peering with custom training has moved. The general instructions for setting up VPC Network Peering with Vertex AI are available at the original link, https://cloud.google.com/vertex-ai/docs/general/vpc-peering. The documentation for custom training is now available here: Using private IP with custom training.

You can now use an interactive shell to inspect your custom training container while it runs. The interactive shell can be helpful for monitoring and debugging training.

This feature is available in preview.

You can now use the gcloud beta ai custom-jobs create command to build a Docker image based on local training code, push the image to Container Registry, and create a CustomJob resource.

You can now containerize and run your training code locally by using the new gcloud beta ai custom-jobs local-run command. This feature is available in preview.

You can now use NVIDIA A100 GPUs and several accelerator-optimized (A2) machine types for training. You must use A100 GPUs and A2 machine types together. Learn about their pricing.

You can now use a pre-built container to serve predictions from TensorFlow 2.5 models.

You can now use a pre-built container to serve predictions from XGBoost 1.4 models.

Vertex AI has added support for custom model training, custom model batch prediction, custom model online prediction, and a limited number of other services in the following regions:

• us-west1
• us-east1
• us-east4
• northamerica-northeast1
• europe-west2
• europe-west1
• asia-southeast1
• asia-northeast1
• australia-southeast1
• asia-northeast3

Vertex AI now supports forecasting with time series data for AutoML tabular models, in Preview. You can use forecasting to predict a series of numeric values that extend into the future.

Vertex Pipelines is now available in Preview. Vertex Pipelines helps you to automate, monitor, and govern your ML systems by orchestrating your ML workflow.

Vertex Model Monitoring is now available in Preview. Vertex Model Monitoring enables you to monitor model quality over time.

Vertex Feature Store is now available in Preview. Vertex Feature Store provides a centralized repository for organizing, storing, and serving ML features.

Vertex ML Metadata is now available in Preview. Vertex ML Metadata lets you record the metadata and artifacts produced by your ML system so you can analyze the performance of your ML system.

Vertex Matching Engine is now available in Preview. Vertex Matching Engine enables vector similarity search.

Vertex TensorBoard is now available in Preview. Vertex TensorBoard enables you to track, visualize, and compare ML experiments.

You can now use a pre-built container to serve predictions from TensorFlow 2.4 models.

You can now use a pre-built container to serve predictions from scikit-learn 0.24 models.

You can now use a pre-built container to serve predictions from XGBoost 1.3 models.

AI Platform Vizier is now available in preview. Vizier is a feature of AI Platform (Unified) that you can use to perform black-box optimization. You can use Vizier to tune hyperparameters or optimize any evaluable system.

The Python client library for AI Platform (Unified) is now called the AI Platform (Unified) SDK. With the release of version 0.7 (Preview), the AI Platform (Unified) SDK provides two levels of support. The high-level aiplatform library is designed to simplify common data science workflows by using wrapper classes and opinionated defaults. The lower-level aiplatform.gapic library remains available for those times when you need more flexibility or control. Learn more.

You can now use a pre-built container to perform custom training with PyTorch 1.7.

AI Platform (Unified) now supports Access Transparency in beta. Google Cloud organizations with certain support packages can use this feature. Learn more about using Access Transparency with AI Platform (Unified).

In addition to Docker images hosted on Container Registry, you can now use Docker images hosted on Artifact Registry and Docker Hub for custom container training on AI Platform.

You can now use a pre-built container to perform custom training with TensorFlow 2.4.

You can now use a pre-built container to serve predictions from TensorFlow 2.3 models.

You can now use a pre-built container to serve predictions from XGBoost 1.2 models.

You can now use a pre-built container to perform custom training with PyTorch 1.6.

Preview: Select AI Platform (Unified) resources can now be configured to use Customer-managed encryption keys (CMEK).

Currently you can only create resources with a CMEK key in the UI; this functionality is not currently available using the client libraries.

You can now use a pre-built container to perform custom training with TensorFlow 2.3.

AI Platform (Unified) now stores and processes your data only in the region you specify for most features. Learn more.

Preview release

AI Platform (Unified) is now available in Preview.

For more information, see the product documentation.