Overview

What is BuiltinTrainer

The Kubeflow SDK train() API supports two types of trainers: BuiltinTrainer() and CustomTrainer().

These options allow you to specify how you want to configure the TrainJob:

1. CustomTrainer: Use this when you need full control over the training process. It requires you to define a self-contained Python function that includes the entire model training process.

2. BuiltinTrainer: Designed for configuration-driven TrainJobs using a predefined training script, often tailored for tasks like LLMs fine-tuning. The training script contains entire post-training logic for LLMs fine-tuning, and it allows you to adjust the configurations for dataset, LoRA parameters, learning rates, etc. The BuiltinTrainer is ideal for fast iteration without modifying the training loop.

Currently, Kubeflow SDK supports these configs for BuiltinTrainer:

1. TorchTuneConfig: Configuration to fine-tune LLMs with TorchTune.

Next Steps

• Learn how to use TorchTune BuiltinTrainer