R9700 1-GPU ROCm Qwen3.8-27B UD-Q4_K_XL depth-0 decode

Donato Capitella, local-llm-benchmarks.dev speed dataset (results.json, run run-3fc4ea643a41c841, completed 2026-10-02). Protocol deterministic-prime-extend-completions-v1, runner local-llm-speed v1, measured at the API endpoint of an OpenAI-completions server, concurrency 1, 2 repeats after 1 warmup. Depth 0: no primed prefix (actual_cached_depth_tokens 0, recomputed_history_tokens 0), so the whole 2048-token prompt is computed. Model file Qwen3.8-27B-UD-Q4_K_XL.gguf from unsloth/Qwen3.8-27B-GGUF, KV cache type default, speculative decoding off. Single GPU (execution_label '1 GPU'). llama.cpp version 0.5.0-dev build 11347 commit 5fc4f3c8c. ROCm 10.0, container image r9700-llama-rocm-10.0. Decode figure is client-observed generation throughput: the 256 generated tokens after the first, divided by the first-to-last token interval, so TTFT is excluded. Per-request, not aggregate. Repeat means 30.826 and 30.485 tok/s (stddev 0.241).

Same host and model file as the paired prefill/decode record; the two are one run split by phase. Not comparable with the repo's llama-bench pp512/tg128 rows for this card: different prompt length, different measurement point (API endpoint, TTFT excluded from decode), and a server config rather than llama-bench.