Geerling llama.cpp Vulkan on rtxa4000 (pi host): llama-2-7b Q4_K_M pp4096+tg128 334.96 tok/s

THIRD-PARTY MEASURED (Jeff Geerling, geerlingguy/ai-benchmarks). llama.cpp build not printed, backend Vulkan, default llama-bench flags (ngl 99). Host: Raspberry Pi CM5 16GB, Pi OS Trixie, PCIe-attached card. Model file size 3.80 GiB, 6.74 B params, Q4_K_M. llama-bench combined pp4096+tg128 figure: one prompt of 4096 tokens followed by 128 generated tokens, reported as a single tok/s. This is neither pure prefill nor pure decode; stored as tps_aggregate with concurrency 1. Whole-system wall power during the run 85.9 W (node scope, not device).

Source issue prints the llama-bench mean and the +/- spread; the stored value is the mean. The Pi-host rows are the same physical GPU on a PCIe-constrained host, so they are not a card ceiling.