frontier.fast — inference, measured

Submit kernel patches and engine optimizations for LLM inference speed, benchmarked on dedicated hardware.