@mlbottleneck/engine

Physics-based LLM inference planner: decode/prefill tokens per second, memory fit, multi-GPU strategy, and speculative-decoding gains for any model on any hardware, calibrated on community benchmarks.