Run 35B MoE models in ~2.5 GB of RAM. Weights stream from SSD; a trained router predicts loads a step ahead. Mac, iPhone, Android, Windows supported
Python
README: drop numerical-guard bullet from public feature list
linyubupa5f70676cdf41
diff --git a/README.md b/README.mdindex 7f3e699..d9ca5de 100644--- a/README.md+++ b/README.md@@ -76,10 +76,7 @@ are co-located with each checkpoint and load automatically, sowith model size, routed width (K), and memory pressure;- **Parallel LoRA**: adapters are applied as a side path at forwardtime instead of being merged — the base stays a read-only mmap and- multiple adapter sets share one base;-- **Numerical guard**: per-layer hidden clipping (`LING_HIDDEN_CLIP`,- default 1000) breaks the fp16 overflow → all-NaN logits → token-0- death-spiral collapse chain.+ multiple adapter sets share one base.## Quick startdiff --git a/README_zh.md b/README_zh.mdindex f2502a8..8216f54 100644--- a/README_zh.md+++ b/README_zh.md@@ -58,9 +58,7 @@**存储越慢,收益越大**:该机制消除的正是专家工作集超出常驻容量时占主导的冷读等待,因此收益随模型规模、路由宽度(K)与内存压力增大;- **并行 LoRA**:适配器不合并进基模,前向时旁路叠加,- 基模保持只读 mmap、多套适配器共享同一份基模;-- **数值防护**:per-layer hidden clip(`LING_HIDDEN_CLIP`,默认 1000)- 阻断 fp16 溢出 → 全 NaN logits → token-0 死循环的塌缩链。+ 基模保持只读 mmap、多套适配器共享同一份基模。## 快速开始