Build & Train VLA Models

From MoLe-VLA: Dynamic Layer-skipping Vision Language Action Model via Mixture-of-Layers for Efficient Robot Manipulation:

A 7B VLA model running on a commercial-grade GPU like the RTX 4090 generally achieves an inference frequency of approximately 5-12 Hz.

Training Parameters

Data Collator

Core functionality:

1. Batch creation.
2. Padding. Ensuring all samples in a batch (such as variable-length text inputs) have the same dimensions.
3. Convert Python objects to PyTorch.
4. Create attention masks or add special tokens.