Nvidia uses linear algebra to fix cross-model KV cache transfer, 25x faster
bendee983 · x · 2026-08-22
Transferring conversations between multiple LLMs in agentic systems invalidates the KV cache, forcing expensive and slow recomputation of the prefill stage. Nvidia researchers studied cross-model KV cache transfer to avoid this.
Their main finding is that for within-family transfers (e.g., different Qwen3 sizes), simple algebra tricks (no gradient-based methods) can map KV caches accurately. They designed a mapper with three key components, including one using simple linear regression. This technique is up to 25x faster than recomputing from scratch.
More from Infra
- Nvidia's developer base hits 11 million, up ~3.7x from 3 million in 2021 — Beth_Kindig · 2026-10-16
- Muse Spark Open runs locally on a Mac M3 Ultra via llama.cpp: frontier model on desktop — pcuenq · 2026-10-16
- Muse-Spark open-weights drop: 816B-A42B MoE model gets 359GB MLX 3-bit build — ns123abc · 2026-10-16
- Meta launches Muse Spark Open, says it runs on two M5 Ultra Mac Studios with Apple's help — The Verge AI · 2026-10-16
- vLLM ships Day-0 LoRA RL for Muse Spark Open, lifting AIME from 0.56 to 0.85 — vllm_project · 2026-10-16
- DSpark draft model for Muse Spark Open proposes 7 tokens per pass with 23-42% acceptance — vllm_project · 2026-10-16