显存VRAM
VRAM
解决什么问题What problem it solves
显卡上的内存。装不下模型就是硬装不下，跟别的无关。Memory on the graphics card. If the model does not fit, it simply does not run — nothing else matters.
这是什么What it is
GPU 上的高速内存。模型权重和 KV 缓存都要放进去，超了就无法运行。
显存是本地部署的第一道硬门槛，不是性能问题而是「能不能跑」问题。
High-speed memory on the GPU. Model weights and the KV cache both live there, and overflowing means it cannot run.
VRAM is the first hard threshold for local deployment — not a performance question but a can-it-run-at-all question.
为什么需要知道Why it matters
这是本地跑模型唯一需要先算清楚的账。选量化档位本质上就是为了塞进显存。It is the only budget you must settle before running a model locally. Choosing a quantisation tier is essentially about fitting inside VRAM.