Not enough VRAM: context length is the hidden killer

A colleague pasted the whole Service — about two thousand lines. The critical function sat in the middle. Head and tail entered; the middle often did not. It still occupies length. Length up, the local card dies. Weights were already loaded. Loaded only means parameters are in. Chat starts, KV cache grows with this turn’s tokens, and the top is OOM. The pack still says long context.