Serverless GPU environment version 5 (Preview)

This page outlines the system environment information for Serverless GPU environment version 5. This compute offering is part of AI Runtime, which is designed for modern AI and deep learning workloads.

Serverless GPU environment 5 is built on top of serverless environment 5 (CPU). See what's new in serverless environment 5 (CPU). It includes the following environment:

• Default base environment: a minimal environment with stable client API.

To ensure compatibility for the application, Serverless GPU workloads use a versioned API, known as the environment version, which remains compatible with newer server versions.

You can select a base environment that includes this environment version using the Environment side panel in your serverless notebooks. See Interactive (Notebooks).

New features and improvements

In version 5, Databricks is providing a minimal environment that supports only serverless GPU API, the Databricks dependencies, and MLflow. Users can customize the environment to their needs. The default base environment in version 5 does not include certain packages present in the CPU environment, in order to provide a minimal functional GPU environment. Users wanting a complete environment with comprehensive AI libraries should use the Databricks AI environment.

API updates

April 14, 2026

Serverless GPU Python API upgraded to 0.5.15

The Serverless GPU Python API 0.5.15 includes the following API updates:

• New features:
  • Added log propagation to MLflow for local Serverless GPU API runs. Training logs from local runs are now automatically uploaded to MLflow.
• Bug fixes:
  • When PyTorch is not installed, the error message now correctly indicates that the Standard v5 base environment does not include torch.
• Security:
  • Pinned all Python dependency versions to exact versions to prevent supply chain attacks from newly published malicious package versions.

Mar 17, 2026

Serverless GPU Python API upgraded to 0.5.14

The Serverless GPU Python API 0.5.14 includes the following API updates:

• Bug fixes:
  • Fixed an issue where the Serverless GPU API failed on control planes without MAPI available. The API now gracefully falls back to on-demand compute when MAPI is unreachable.

Mar 2, 2026

Serverless GPU Python API upgraded to 0.5.13

The Serverless GPU Python API 0.5.13 includes the following API updates:

• Bug fixes:
  • Fixed Azure A10 GPU device name parsing in local mode. The API now handles Azure-specific formats such as "A10-24Q".
  • Fixed an issue where the MLflow active run was not closed after local execution completed.
  • Fixed MLflow log streaming compatibility with MLflow 3.x. A change in artifact download path behavior in MLflow 3.x caused log streaming to break. Log streaming now works correctly with both MLflow 2.x and 3.x.
  • Fixed an issue causing training runs to hang. The cross-rank MLflow status synchronization feature introduced in a previous version has been reverted.

Packages not included in base environment

The following packages are not included in the base environment but remain available in the CPU environment:

• scipy
• seaborn
• scikit-learn

Based on customer feedback requesting control over PyTorch versioning, torch is not included in the base environment for version 5. This allows you to install the version that best fits your workload. To use a Databricks-tested version, run %pip install torch==2.9.0 in a notebook cell or install torch==2.9.0 using the Dependencies section of the Environment side panel.

System environment

• Operating System: Ubuntu 24.04.2 LTS
• Python: 3.12.3
• Databricks Connect: 18.0.0
• NVIDIA CUDA Toolkit: 12.9

Installed Python libraries

Default base environment

To reproduce serverless GPU environment 5 in your local Python virtual environment, download the requirements-env-gpu-5.txt file and run pip install -r requirements-env-gpu-5.txt. This command installs all of the open source libraries from serverless environment 5.