cfg-0176
backfilled
citable URL: https://halobench.com/records/cfg-0176/ — this address never moves; the anchor /records/#cfg-0176 keeps resolving
| model | Qwen3.8-27B-UD-Q4_K_XL.gguf · UD-Q4_K_XL |
| engine | ggml-org/llama.cpp 7077abb · ROCm0 / TheRock ROCm 7.14 · host aihydra (igpu) |
| flags | Operator-directed autonomous served-path depth samples: -dev ROCm0 -ngl 999 -c 204800 -fa on -ctk q8_0 -ctv q8_0 -t 16 -tb 32 -np 1 --load-mode none --jinja --reasoning on --reasoning-effort medium --reasoning-format deepseek --spec-type draft-mtp --spec-draft-n-max 3. |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. The archived speed JSON has one repetition per target depth and reports the actual prompt token count, which is materially below the requested target at both retained deep samples. It has no artifact SHA256, template hash, house guard, plain/spec-off paired floor, or energy window; it cannot establish a production recommendation.
Capability basis
unmeasured — performance numbers on this config stand on a guard alone, not an established capability