# 2026-09-15 07:4xZ — GPU 0 (idle), llama-main 69320fe, Qwen3-8B Q4_K_M, llama-bench -fa 1 -ngl 99 -p 2048 -n 0 -r 1
# nvidia-smi utilization sampled at 1 s during the run. CPU fallback shows as an idle GPU.
K q5_1  V q5_1   pp2048  181.7 tok/s   run 25 s   GPU util samples: 4 4 4 2 4 2 4 23
K q8_0  V q4_0   pp2048  275.2 tok/s   run 17 s   GPU util samples: 4 25 10 8 6 6 4 4
K q4_0  V q4_0   pp2048 4661.7 tok/s   run  2 s   GPU util samples: 72 0 100
# Phase-1 (GPU 1) context: q5_1 pp8192 43.3 / pp32768 10.9 tok/s (2h41m for the type); K q8_0/V q4_0 pp8192 62.7 (run cut).
# Reading: with these K/V types the batched (prefill) flash-attention op is not executed on the GPU on this build;
# q8_0/q8_0 and q4_0/q4_0 are. Decode (tg128) still ran on the GPU for all types (q5_1 tg128 91.6).
