cfg-0109
backfilled
citable URL: https://halobench.com/records/cfg-0109/ — this address never moves; the anchor /records/#cfg-0109 keeps resolving
| model | laguna-s-2.1-Q4_K_M.gguf · Q4_K_M |
| engine | ggml-org/llama.cpp 3653e6d · rocm · host aihydra (igpu) |
| flags | -ngl 999 -fa 1 -dev ROCm0 -ctk f16 -ctv f16 -lm none |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. Part 2 deep-cell session (2026-08-17), same fingerprint as cfg-0090 (identical llama-bench invocation: build 3653e6d, ROCm, Q4_K_M, f16 KV, --load-mode none -- only depth differs). ROCm is this candidate's required backend at any useful depth (Vulkan lost the GPU device on both allowed attempts at d131072, clm-0065); this cell extends the ROCm-only series to d204800. A fresh config id is minted per house convention. Ingested from llama-bench, which measures throughput and not footprint.
Capability basis
unmeasured — performance numbers on this config stand on a guard alone, not an established capability
Runs on this config (2)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0391 | 2026-08-17 | performance | llama-bench@1 | prefill_tps=81.946523 |
| run-0392 | 2026-08-17 | performance | llama-bench@1 | decode_tps=4.914985 |
Cited by — computed at build time, never stored
model pages laguna-s-21