AMD developer - 'llama.cpp on AMD EPYC: ZenDNN backend' (EPYC 9755 128c benchmarks)

AMD vendor blog (ZD-066/067/068). EPYC 9755 128c single socket pinned (one NUMA node), ZenDNN 5.2.2 vs native CPU backend on llama.cpp b9326. llama-batched-bench 512in/128out/128seq, 3-run mean, flash-attn on, AOCL-DLP. Up to 4.5x prefill Mixtral-8x7B Q8_0, ~2x dense. Decode falls back to CPU path. WikiText-2 PPL delta <=0.0004.