Single-HPU Training

Training on a single device is as simple as in Transformers:

• You need to replace the Transformers’ Trainer class with the GaudiTrainer class,
• You need to replace the Transformers’ TrainingArguments class with the GaudiTrainingArguments class and add the following arguments:
  • use_habana to execute your script on an HPU,
  • use_lazy_mode to use lazy mode or not (i.e. eager mode),
  • gaudi_config_name to give the name of (Hub) or the path to (local) your Gaudi configuration file.

You can find several examples in the official repository for the following tasks:

• language modeling,
• question answering,
• summarization,
• text classification,
• translation,
• image classification,
• audio classification,
• speech recognition,
• contrastive image-text training.

To go further, we invite you to read our guides about accelerating training and pretraining.