Gemma 4 QAT + MTP: max 33% speed increase in token generation, any ideas?

Hello, My setup is 2x RTX 3060 Ti 8GB, without the assistant model (MTP) I get around 75t/s, adding the assistant model as draft I manage to reach 100t/s peak. I tried puting the model on a single card with minimal context size, but still not helping in terms of t/s. I set the following draft parame