llama.cpp Mi50 benchmarks pit ROCm, Vulkan

This looks less like a universal backend verdict than a reminder that local inference performance is shaped by context length and model topology.

• –For dense models, Vulkan wins the interactive end: Qwen 3.5 9B prompt processing is 871.17 t/s on Vulkan at 512 tokens vs 708.58 on ROCm, and 27B is 252.68 vs 209.06.
• –The crossover shows up fast: at 32k context, 9B prompt processing flips to 593.8 t/s on ROCm vs 447.76 on Vulkan, and 27B flips to 176.69 vs 128.72.
• –Generation stays more Vulkan-friendly on dense models, so the ROCm win mostly comes from prompt processing amortizing the whole session rather than raw token-by-token speed.
• –The 122B run with 28 layers offloaded to CPU is where ROCm really earns its keep: at 32k, tg is 24.65 t/s on ROCm vs 18.41 on Vulkan, while pp is 153.16 vs 113.16.
• –Nightly ROCm is still a risk tradeoff: the reported llama-server prompt-cache OOM and earlier leak-like behavior make these results useful, but not production-safe.