10%+ performance improvement on MoE ssd-streaming with expert-lookahead

I implemented expert lookahead, a trick that gets a 10%+ performance improvement on MoE models running on low-memory macs (Qwen 3.8 flash in this case) while using expert-offloading/ssd-streaming (using slotstream). This is 10% on top of several optimizations. I initially trained a small model to pr

中文全文翻译正在补齐，可先查看下方英文原文。

I implemented expert lookahead, a trick that gets a 10%+ performance improvement on MoE models running on low-memory macs (Qwen 3.8 flash in this case) while using expert-offloading/ssd-streaming (using slotstream). This is 10% on top of several optimizations. I initially trained a small model to pr