RuntimeError: InternalServerError: 502 Bad Gateway during Unsupervised Data Generation (Client Backend)

[florentiino]

Problem Description:
When attempting to run the 'Unsupervised Learning with TabPFN' section of the demo notebook, specifically the 'Data Generation' experiment using tabpfn_extensions.unsupervised.experiments.GenerateSyntheticDataExperiment, a RuntimeError with an HTTPError: 502 Server Error: Bad Gateway is encountered.

This error occurs when tabpfn_extensions.unsupervised.unsupervised.TabPFNUnsupervisedModel.generate_synthetic_data internally calls model.predict on the TabPFN client, suggesting a communication issue with the backend inference service.

Steps to Reproduce:

1. Ensure backend is set to 'client' in the 'Backend Selection' section.
2. Navigate to the 'Unsupervised Learning with TabPFN' section.
3. Execute the code cell that initializes unsupervised.TabPFNUnsupervisedModel and runs unsupervised.experiments.GenerateSyntheticDataExperiment (cell npJnFdtTa_nS in the provided notebook state).

Expected Behavior:
Synthetic data should be generated successfully, and the experiment should complete without errors.

Actual Behavior:
An InternalServerError with a 502 Bad Gateway HTTP status is raised, preventing the completion of the synthetic data generation.

RuntimeError: InternalServerError: An unexpected error occurred on our server, sorry about that! Please report this error on GitHub (https://github.com/automl/tabpfn-client) or Discord (https://discord.com/invite/VJRuU3bSxt). Please include the following error message: HTTPError: 502 Server Error: Bad Gateway for url: https://a56a1e89df0acc73.europe-west4-668127571160.prediction.vertexai.goog/v1/projects/668127571160/locations/europe-west4/endpoints/a56a1e89df0acc73:rawPredict

Environment:

• Google Colab notebook (current state shows Python 3.12.12)
• tabpfn-client (version installed via !uv pip install tabpfn-client)
• tabpfn-extensions (version installed via !uv pip install tabpfn-extensions[all])
• tabpfn (version installed via !uv pip install tabpfn)
• Backend selected: client

[bejaeger]

Hi @florentiino , could you test the demo again? This looks like it might have been a temporary issue with our servers.