Reading up on TensorRT-LLM

LLM inference works by tokenizing input, computing embeddings through transformer layers, then generating tokens autoregressively with KV caching and quantization optimizations.