A user discovered that applying a patch to llama.cpp for AMD hardware significantly increases available context length by reducing MTP buffer overhead. By preventing the overestimation of memory required for MTP compute-buffer and scheduler allocation during auto-fit, the Qwen 27B model saw context capacity increase from 64K to 149K on a 16GB + 12GB configuration. This optimization ensures more VRAM is allocated to the user context rather than being wasted on overestimated buffers.