AMD llama.cpp: reducing MTP buffer overhead gave me 64K → 149K context for Qwen 27B

Collecting trend data

The percentage is based on available heat signal, not comment count or independent people.