NVIDIA's Cross-Model KV Cache Transfer Speeds Up Inference by 25x

theomitsa · x · 2026-08-07

NVIDIA researchers introduced a cross-model KV cache transfer technique. When swapping between different-sized models in an LLM family (e.g., Qwen3 14B to 32B), this method allows the target model to reuse the source's KV cache, entirely skipping the prefill phase.

The researchers found substantial linear structure across matched KV pairs. They designed a closed-form ridge mapper that selects the top-k most predictive source layers to generate the target's cache.

Experiments show this conversion runs 2.7 to 25x faster than reprocessing the context, which is crucial for optimizing cost-quality cascading, mid-conversation routing, and reducing LLM serving costs.

Related event: NVIDIA Research Achieves 25x Inference Speedup with Cross-Model KV Cache Reuse(2 posts)→

Original post →

More from Infra

Infra channel →