LoongForge-A high-performance training framework for LLMs, VLMs, diffusion, vlas (github.com)

Baidu has introduced LoongForge, a high-performance training framework for large language models (LLMs), vision-language models (VLMs), diffusion models, and embodied AI. This unified framework, part of Baidu Baige's open-source series, boasts up to a 5x training speedup compared to mainstream baselines, making it a significant advancement for the AI/ML community. Built on NVIDIA’s Megatron-LM, LoongForge features deep optimizations for improved model coverage and hardware support, including native compatibility with both NVIDIA GPUs and Baidu’s Kunlun XPUs. The framework's architecture supports flexible multi-modal configurations and various advanced training techniques such as heterogeneous parallelism, decoupled encoder-decoder training, and adaptive FP8 training. Users benefit from enhanced scalability options like load balancing and bespoke fused operators, ideal for accelerating the training of complex models across diverse industries such as education and computer vision. With its open-source release, LoongForge not only makes sophisticated AI training more accessible but also invites community contributions, further enriching the ecosystem of AI model development.