50% tg increase with offloading "hot" experts to VRAM

I got a 50% performance boost (20 t/s -> 30 t/s) in llama.cpp for MoE models that don’t fit entirely in VRAM—in my case, Qwen 3.8 Flash Next. The idea is simple: instead of offloading entire layers to the GPU, I offload only the “hot” experts. I found that certain groups of experts remain rel

中文全文翻译正在补齐，可先查看下方英文原文。

I got a 50% performance boost (20 t/s -> 30 t/s) in llama.cpp for MoE models that don’t fit entirely in VRAM—in my case, Qwen 3.8 Flash Next. The idea is simple: instead of offloading entire layers to the GPU, I offload only the “hot” experts. I found that certain groups of experts remain rel