Training-free geometric KV routing cuts Qwen memory traffic by 16x
Electrical_Offer5667 · reddit · 2026-08-22
A training-free geometric KV routing method for frozen Qwen models assigns expired KV entries to centroid-represented regions, computing attention only over {sinks + recent window + selected regions}. Tests on Qwen3.5-2B (32k) and Qwen3-4B (8k) show 3.7x to 31x fewer physical KV reads while maintaining retrieval capabilities. Although wall-clock speed doesn't yet beat optimized dense attention due to GPU brute force and routing overhead, it highlights potential for reducing memory bandwidth and capacity in long-context scenarios. A GitHub demo is available.
More from Infra
- Jensen Huang joins BG2 podcast: does AI revenue justify capex, '27 bottlenecks — elonmusk · 2026-10-16
- Mistral's on-device infra rack revealed: DGX Spark, Mac Studio and more — gregmushen · 2026-10-16
- NVIDIA and Sweden's Wallenberg launch Sferical One, a shared AI supercomputer — nvidia · 2026-10-16
- Next infra frontier: self-evolving systems that adapt and self-verify — _shreya_s · 2026-10-16
- Xiaomi's MiMo RL dashboard reveals $3.47M spent on just 30 training steps — bycloud · 2026-10-16
- Cerebras pushes disaggregated inference: heterogeneous prefill pools feeding wafer-scale decode — ai · 2026-10-16