expert-prefetch

Here are 2 public repositories matching this topic...

SuperMarioYL / xpref

v0.3.0: live attach against a real patched engine works — the bundled patch maps ring records from the 44-byte header the consumer reads (was 40, crashing every attach) and mid-stream attach syncs instead of waiting for token 0; madvise hints are page-rounded (Linux silently dropped unaligned ones); lazy predictor decay; recall@16 0.8154.

inference moe nvme llm deepseek ai-radar kimi-k3 expert-prefetch router-logit moe-runtime

Updated Oct 9, 2026
Python

andreyivan4enkov / moe-orbit-prefetch

Star

Object A: MoE orbit prefetch — source+math+215KB trajectories+plots+sparse DeepSeek runtime (Apache-2.0). Analyze without weights; credit on substantial use.

pytorch moe memory-efficient mixture-of-experts huggingface dynamic-inference safetensors llm-inference deepseek sparse-moe expert-prefetch expert-offload

Updated Jul 29, 2026
Python

Add this topic to your repo

To associate your repository with the expert-prefetch topic, visit your repo's landing page and select "manage topics."

Learn more