只有12g显存， 能否使用mini使用内存进行加速。

mini都是使用cpu进行推理的， 是否可以使用gpt进行加速。 比如我用的3060-12G . 是否能够充分发挥下显存。加快mini的推理速度。

Start by locating the mini inference entry point and checking how it selects CPU execution. Then determine whether the current runtime and model support the requested GPU path on a 12 GB RTX 3060, and define completion as a reproducible GPU-accelerated inference path with measured results. No files or tests are named in the issue, so repository research and design decisions are required.

Written by the indexing model from the issue text.