Serverless GPU environment version 3 (beta)

This article outlines the system environment information for Serverless GPU environment version 3.

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 the environment version using the Environment side panel in your serverless notebooks. See Set up serverless GPU compute.

New features and improvements

Serverless GPU environment 3 is built on top of serverless environment 3 (CPU). See what's new in serverless environment 3 (CPU).

Known issues

Threadpoolctl library warning when importing transformers

When importing the transformers library, you may see the following error message. This is a known issue with threadpoolctl library in Serverless GPU environment 3.

Exception ignored on calling ctypes callback function: <function _ThreadpoolInfo._find_modules_with_dl_iterate_phdr.<locals>.match_module_callback at 0x7fe29c7da8e0>
Traceback (most recent call last):
  File "/databricks/python/lib/python3.12/site-packages/threadpoolctl.py", line 400, in match_module_callback
    self._make_module_from_path(filepath)
  File "/databricks/python/lib/python3.12/site-packages/threadpoolctl.py", line 515, in _make_module_from_path
    module = module_class(filepath, prefix, user_api, internal_api)
             ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/databricks/python/lib/python3.12/site-packages/threadpoolctl.py", line 606, in __init__
    self.version = self.get_version()
                   ^^^^^^^^^^^^^^^^^^
  File "/databricks/python/lib/python3.12/site-packages/threadpoolctl.py", line 646, in get_version
    config = get_config().split()
             ^^^^^^^^^^^^^^^^^^
AttributeError: 'NoneType' object has no attribute 'split'

To resolve this issue, upgrade to serverless GPU environment 4 or upgrade the threadpoolctl library:

pip install threadpoolctl==3.1.0

System environment

• Operating System: Ubuntu 24.04.2 LTS
• Python: 3.12.3
• Databricks Connect: 16.4.2
• NVIDIA CUDA Toolkit: 12.4

Installed Python libraries

In addition to what's included in serverless environment 3 (CPU), serverless GPU environment 3 includes the following:

• torch 2.6.0
• torchvision 0.21.0

To reproduce serverless GPU environment 3 in your local Python virtual environment, download the requirements-env-gpu-3.txt file and run pip install -r requirements-env-gpu-3.txt. This command installs all of the open source libraries from serverless GPU environment 3.