Developer open-sources CUDA megakernel for Ant Group's Ling-mini-2.0

A developer has open-sourced a single-launch CUDA megakernel for Ant Group's Ling-mini-2.0 model, achieving 843 tokens per second on a single RTX 5090 Laptop GPU. The megakernel fuses the expert router and down-projection steps to reduce synchronization barriers and speeds up vocabulary table searches. Open article →