Note
Access to this page requires authorization. You can try signing in or changing directories.
Access to this page requires authorization. You can try changing directories.
Important
This feature is in Public Preview.
AI Runtime is a serverless GPU compute offering at Databricks intended for deep learning workloads. You can use AI Runtime to train and fine-tune custom models using your favorite frameworks and get state-of-the-art efficiency, performance, and quality.
Get started
Run your first workload in minutes using our quickstart guides.
| Get started | Description |
|---|---|
| 🚀 Databricks CLI quickstart | Coming from a Slurm or Kubernetes cluster? Train on AI Runtime from your terminal in minutes. |
| 📓 Notebook quickstart | Attach your notebook to a GPU in seconds and develop interactively. |
| âš¡ Ray | Coming from a Ray cluster? Run Ray on AI Runtime with dashboard support. |
| 🧠Overview | Learn the key things about AI Runtime in two minutes: available GPUs, how it works, and limitations. |
Features
AI Runtime is a full-stack GPU platform, with many ways to connect to GPUs, built-in observability and debugging tools, and pre-built environments.
Connect to Serverless GPUs: Reach serverless GPUs from wherever you work.
| Feature | Description |
|---|---|
| Notebooks | Attach a notebook to serverless GPU compute and develop interactively, with no cluster setup. |
| IDE over SSH | Connect from your IDE or terminal over an SSH tunnel to work on a GPU node directly. |
| Databricks CLI | Submit and manage distributed GPU training jobs from your laptop using YAML job configs. |
| Ray | Run Ray Core, Ray Data, Ray Train, and Ray Tune on serverless GPU compute. |
Managing Dependencies: Control the Python and system libraries your workload runs with.
| Feature | Description |
|---|---|
| Pre-built environments | Start from a minimal Standard environment or a Databricks AI environment pre-loaded with ML frameworks. |
Loading data into GPUs: Feed training data to your GPUs efficiently.
| Feature | Description |
|---|---|
| Efficient data loaders | Stream data from Unity Catalog volumes with fault-tolerant loaders built for high GPU utilization. |
Debugging and Observability: Track experiments, inspect output, and diagnose failures.
| Feature | Description |
|---|---|
| MLflow for deep learning | Track experiments, metrics, and runs with MLflow, and save model checkpoints to Unity Catalog volumes. |
| Logs viewer | View training output and monitor GPU resource usage while your code runs. |
| Debugging with agents | Use Genie Code to generate training code, resolve environment issues, and debug GPU failures. |
Scheduling and real-time serving: Move from interactive development to scheduled jobs and endpoints.
| Feature | Description |
|---|---|
| Productionize your jobs | Deploy training code with Declarative Automation Bundles, schedule runs, and build multi-task GPU and CPU workflows. |
| Serve your models | Deploy your trained models behind a scalable endpoint with Model Serving. |
Examples
Clone end-to-end examples and run them on AI Runtime in minutes, from the CLI or in notebooks.
| Example | Description |
|---|---|
| Classic machine learning | GPU-accelerated XGBoost, zero-shot tabular prediction, and time series forecasting. |
| Recommendation systems | Train deep-learning recommender models, such as two-tower architectures. |
| Computer vision | Object detection and image classification workloads on GPUs. |
| Post-training OSS models (LLMs) | Fine-tune and post-train open-source LLMs with LoRA, full fine-tuning, and libraries like TRL, Unsloth, and Axolotl. |
| Batch inference | Run large-scale batch inference with Ray Data and vLLM across multiple GPUs. |
| Multi-GPU distributed training | Scale training across multiple GPUs and nodes with DDP, FSDP, and DeepSpeed. |
Guides
Task-focused guides for migrating workloads to AI Runtime and getting the most out of your GPUs.
| Guide | Description |
|---|---|
| Migrate from Slurm | Map your Slurm batch scripts, distributed launchers, and shared storage onto the Databricks CLI. |
| Migrate classic GPU workloads | Move an existing deep learning workload from a classic Databricks cluster to AI Runtime. |
| Performance and resiliency | Improve training throughput and make long-running jobs more resilient to failures. |
Learn more
Find helpful content: the latest product updates and how AI Runtime works under the hood.
| Resource | Description |
|---|---|
| Product updates | See the latest AI Runtime product updates and announcements. |
| How AI Runtime works under the hood | Read how Databricks keeps GPUs reliable across AI Runtime. |