From the linked post, it didn't read like a separate KV cache was needed:

> The draft models seamlessly utilize the target model's activations and share its KV cache, meaning they don't have to waste time recalculating context the larger model has already figured out.