rx7900xtx-local-llm
victoralcazardev/rx7900xtx-local-llm
Python
262K-context Qwen3.8-27B on a single RX 7900 XTX: llama.cpp ROCm b11846 + native q8_0/q5_1 TILE FA, IQ3_S+MTP n=3 + n-gram map, 272 W. 22-33 tok/s at 240K fill (55 copy), 76 tok/s empty, 68/68 recall.