3 ms·
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
by a_e_k 5mo ago
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.
- coder543 5mo agoThat's great news. That has not been the case with other MTP implementations like Qwen3.5, but I see the section in the article saying Google introduced some architectural optimizations to make this possible.