2025-05-15

Highlights

• FlexAI Inference - Preview Release: You can now create and manage Inference Endpoints using the flexai inference command!
• Hugging Face Storage Connector: Hugging Face is now part of the list of Remote Storage Providers you can leverage by creating a storage connection.
• TensorBoard Support: Training Jobs can now be tracked using TensorBoard, allowing for real-time visualization of training metrics and performance.

Added

• Preview release: Support for creating and managing Inference Endpoints via the flexai inference command, enabling users to serve, update, and monitor model inference endpoints directly from the CLI. You only need to specify pass a --hf-token-secret flag pointing to the name of the Secret containing your Hugging Face access token and the name of the vLLM supported model you'd like to serve. Refer to the inference CLI command documentation for more details.
• Hugging Face as a supported Remote Storage Provider. This addition enables you to easily push Datasets and model Checkpoints directly from the Hugging Face Hub — no need to download them locally to push them back again any more.
• Support for TensorBoard in Training Jobs has been added to FlexAI's suite of observability tools. TensorBoard is enabled by default and you only have to make your code write the files required by TensorBoard to the path specified in the TENSORBOARD_LOG_DIR environment variable. Visit https://tensorboard.flex.ai/ and log in using the same credentials you use to access the FlexAI CLI.

info

Contrary to what the name might indicate, TensorBoard is not limited to TensorFlow projects; it can also be used with PyTorch and other frameworks.