Learn how to optimize local LLM inference on 8GB GPUs using KV-cache quantization, Flash Attention, batching, and practical latency benchmarks.