llama.cpp MTP patch boosts Qwen 27B context from 64K to 149K on AMD

A user patch reducing multi-token prediction (MTP) buffer overhead in llama.cpp increases Qwen 27B context from 64,256 to 149,248 tokens on AMD ROCm. A single-GPU IQ4_XS config jumps from 19,456 to 76,032 tokens.