https://harshitmalik.dev Harshit Malik is an AI engineer working on LLM inference and multi-agent systems in production: KV cache arithmetic, serving economics, and agent orchestration under real traffic. en https://harshitmalik.dev/blog/inside-the-kv-cache https://harshitmalik.dev/blog/inside-the-kv-cache Tue, 29 Sep 2026 00:00:00 GMT From gpu_memory_utilization and max_model_len to weights, activations, CUDA graphs and the KV pool: where every gigabyte on the card actually goes.