Laguna S 2.1 Q4_K_M / ROCm b10008 without DFlash / native PP512 and separate TG128
AMD Radeon 8060S integrated GPU (Strix Halo; installed RAM unreported) · poolsideai/llama.cpp laguna b10008 / 04b2b72cb + local include fix / ROCm
- generation:
- 18.4 tokens/s
- prompt processing (prefill):
- 305.5 tokens/s
- quant:
- Q4_K_M (GGUF)
- flash attention:
- off
Reported by the source; GPU count, offloading and concurrent requests can change these figures. Prompt-processing speed is input throughput, not time to first token. Check the full setup before comparing.
User reports Laguna S 2.1 at 18.4 t/s generation and 305.5 t/s prompt processing on a Nimo Strix Halo box with an AMD Radeon 8060S Graphics GPU. Setup is poolsideai/llama.cpp laguna branch with Q4_K_M, -ngl 99, -fa 0 and --mmap 0; the printed model is laguna 118B.A8B Q4_K - Medium, 117.56 B parameters, 70.01 GiB. The figures are separate synthetic llama-bench phases, PP512 and TG128, with no DFlash active; KV cache type, threads, batch and repeats are not printed.