We open-sourced LoongForge: Up to 2.67x faster training throughput for VLA & world-action models

If you’ve been training Vision-Language-Action (VLA) or world-action models, you’ve probably noticed that the training infrastructure feels a bit unoptimized compared to the mature stack we have for LLMs. Most researchers end up relying on each model's official reference repository, which is great for reproducibility, but throughput was rarely the main priority.

We open-sourced LoongForge to bridge that gap. 

It comes with ready-to-run, high-throughput training setups for Pi0.5, GR00T N1.6 & N1.7, X-VLA, FastWAM, DreamZero, Lingbot-VA, and several others. Across these models, we’ve measured significant training speedups (in samples/sec) compared to the official implementations:

• DreamZero: 2.67x speedup

• GR00T N1.6: 2.31x speedup

• Pi0.5: 2.23x speedup

• Lingbot-VA：1.80x speedup

• ...

All benchmark numbers, hardware configurations, and reproduction scripts are documented in the repo.

Why we’re posting: We want to build what the community actually needs next. If you train these types of models, what are the biggest bottlenecks, memory issues, or pipeline breaks you run into? We'd love to hear your pain points and suggestions!

Github：https://github.com/baidu-baige/LoongForge
Web：https://baidu-baige.github.io/LoongForge/