Fix/torch cuda

[Felipedino]

This pull request introduces several updates to the DistilBertTransformerSchema class and associated methods in the DashAI backend, as well as a modification to the requirements.in file to specify the PyTorch package source. The key changes include adjustments to placeholders for training parameters, improved handling of text column validation, and updates to the dataset processing logic.

Updates to DistilBertTransformerSchema:

• Training Parameters:
  • Changed the placeholder for num_train_epochs from 2 to 1 to reflect a default single-epoch training configuration.
  • Updated the placeholder for learning_rate from 5e-5 to 3e-5 to adjust the default optimizer configuration.

Improvements to Dataset Handling:

• Tokenization:

  • Added validation to ensure that the dataset contains exactly one text column during tokenization. This prevents errors when processing datasets with unexpected column structures.

• Prediction:

  • Incorporated the same text column validation logic into the predict method and adjusted the column removal process to dynamically handle the identified text column instead of hardcoding "text".

Dependency Management:

• PyTorch Installation Source:
  • Updated requirements.in to include the --extra-index-url for downloading PyTorch wheels compatible with CUDA 11.8, ensuring proper GPU support.