open-source nanoSGLang Minimal high-performance LLM serving engine in Python and PyTorch. nanoMegatron Minimal distributed training framework covering DDP, ZeRO, tensor, sequence, pipeline, and expert parallelism.