liangsu9988/the-context-dividend

The Context Dividend

Hot-attaching compiled weight structures to a stock serving engine (vLLM / SGLang) returns the memory it saves as KV cache — roughly tripling usable context for Qwen3.8-27B (NVFP4) on a single RTX 5090, with decode speed held at depth.

Measured with paired stock baselines, with and without speculative decoding (MTP, DSpark, DFlash2). This is a fairness-framed hot-plug experiment, not an engine benchmark.