Custom CUDA code runs Qwen3 30B A3B at 50 tok/s on RTX 5060 Ti

A Reddit user achieved 50-54 tok/s on Qwen3-30B-A3B using custom CUDA and C++ code on an RTX 5060 Ti (16GB VRAM) at float8. This represents a roughly 50% speedup over llama.cpp's 33-34 tok/s.