The intuition an LLM engineer needs, without the hardware manual. By the end, techniques like quantization, speculative decoding, and continuous batching should stop looking like a list of tricks to memorize.
