Skip to content

Run frontier MoE models on hardware you already own — pure C, zero deps, experts streamed from disk. Tiny engine, immense model. 🐦

C

docs(v4): track dynamic DSpark tuning

DrewZdb1463ee85aa
diff --git a/docs/deepseek-v4.md b/docs/deepseek-v4.md
index 0476e91a..5092fc41 100644
--- a/docs/deepseek-v4.md
+++ b/docs/deepseek-v4.md
@@ -269,4 +269,6 @@ Default non-thinking encoding:
- [ ] **Model quantization**: more aggressive weight/activation paths for footprint and bandwidth
- [ ] **Server**: HTTP / OpenAI-compatible API for multi-client use
- [ ] **CUDA**: optional GPU backend alongside the CPU path
+- [ ] **Dynamic DSpark for long outputs**: adapt the verify window between K=2–4
+ from recent prefix survival / confidence instead of keeping a fixed K
- [ ] Chat cross-turn KV cache reuse (avoid full re-prefill every turn)
diff --git a/docs/deepseek-v4.zh-CN.md b/docs/deepseek-v4.zh-CN.md
index f8961280..5d01cc13 100644
--- a/docs/deepseek-v4.zh-CN.md
+++ b/docs/deepseek-v4.zh-CN.md
@@ -262,4 +262,6 @@ Chat 每轮都会重新 prefill 历史记录;尚未实现跨轮 KV 复用。
- [ ] **模型量化**:采用更激进的权重/激活路径,降低占用和带宽需求
- [ ] **服务器**:面向多客户端的 HTTP/OpenAI 兼容 API
- [ ] **CUDA**:在 CPU 路径之外增加可选 GPU 后端
+- [ ] **长输出动态 DSpark 调节**:根据近期 prefix survival/confidence,
+ 在 K=2–4 之间动态选择验证窗口,不再始终使用固定 K
- [ ] Chat 跨轮 KV cache 复用,避免每轮重新 prefill 全部历史