I built a PyTorch MoE/MoD training framework with custom CUDA kernels [Apache 2.0]

In three linesOpen-source PyTorch framework for training LLMs with MoE/MoD architectures, custom CUDA kernels (2-7x faster than vanilla PyTorch on T4), up to 64 experts, adaptive training orchestrator monitoring 20+ metrics, configs from 500K to 300B parameters. Apache 2.0, free Colab demo included.

Summary generated by Claude — human-verified