├── logo.ico # Multi-size icon for shortcuts
We have one horrible disjuncture, between layers 6 → 2. I have one more hypothesis: A little bit of fine-tuning on those two layers is all we really need. Fine-tuned RYS models dominate the Leaderboard. I suspect this junction is exactly what the fine-tuning fixes. And there’s a great reason to do this: this method does not use extra VRAM! For all these experiments, I duplicated layers via pointers; the layers are repeated without using more GPU memory. Of course, we do need more compute and more KV cache, but that’s a small price to pay for a verifiably better model. We can just ‘fix’ an actual copies of layers 2 and 6, and repeat layers 3-4-5 as virtual copies. If we fine-tune all layer, we turn virtual copies into real copies, and use up more VRAM.
，推荐阅读迅雷下载获取更多信息
Россиянам раскрыли простые способы выглядеть дорожеMichel Kataná: Шоколадный и бежевый цвета делают образ визуально дороже
Discoverability: How does an LLM know it should read a specific document? Will it be referenced at the right time, during planning and generation?