AMD llama.cpp patch cuts MTP overhead, boosts Qwen 27B context to 149K

Reducing MTP buffer overhead in llama.cpp on AMD expanded Qwen 27B context from 64,256 to 149,248 tokens on a 16GB+12GB setup. The patch also lifted ROCm context on a single 16GB GPU from 19,456 to 76,032 tokens.