Anyone involved in model training can likely relate to these challenges: the number of tasks keeps growing, experiment queues are getting longer, and computing bills are becoming increasingly terrifying.
Today, I’d like to share our recent open-source project: https://github.com/baidu-baige/LoongForge. What we are doing is simple: accelerating training speed while keeping the output exactly the same.
Currently, the LoongForge acceleration framework covers a batch of mainstream embodied AI models (such as pi05, GR00T, xVLA, FastWAM, Lingbot-VA, Cosmos, Dreamzero, Wall-oss-0.5, etc.). It works out of the box and generally delivers a near 200% acceleration effect—effectively cutting training time in half.
Below is a demonstration using DreamZero (an embodied AI model). Please note that this is just one example among the many models currently supported.
📊 In the demonstration video, you will notice:
✅ A 4x increase in training throughput.
✅ Loss curves that are perfectly aligned with the official baseline—not merely approximated, but exactly overlapping.
Furthermore, we welcome everyone to join us in discussing technical challenges and co-creating this open-source community. Stars, contributions, and discussions are highly welcome!
Top comments (2)
Great overview of how Loongforge tackles the bottlenecks in training embodied models. Could you share how the framework handles synchronization when scaling across multiple GPUs for complex task pipelines?
Some comments may only be visible to logged-in visitors. Sign in to view all comments.