Home › Evidence › Records › cfg-0101

cfg-0101

citable URL: https://halobench.com/records/cfg-0101/ — this address never moves; the anchor /records/#cfg-0101 keeps resolving
modelalways42-universal.gguf · F16 · rev main
engineggml-org/llama.cpp 3653e6d · rocm · host aihydra (igpu)
flags-ngl 999 -fa 1 -b 2048 -ub 512 -ctk f16 -ctv f16 -t 16
templatenot recorded at test time
treeupstream — stock

Memory — static estimate vs observed peak

context 8,192 × 1 slot(s) = 8,192 tokens
⌁ total 1.15 GiB of 120 GiB pool · ⌁ headroom 118.85 GiB
⌁ risk utilisation 1.0% — basis observed

Capability basis

measured on this config

Runs on this config (6)

rundatekindsuitemetrics
run-03632026-08-17guarddeep-thought-posttrain-fullbenchtasks_passed=2 · tasks_total=4
run-03642026-08-17performancellama-bench@1prefill_tps=16740.778
run-03652026-08-17performancellama-bench@1decode_tps=187.53
run-03662026-08-17performancellama-bench@1prefill_tps=9276.909
run-03672026-08-17performancellama-bench@1decode_tps=164.871
run-03722026-08-17capabilitydeep-thought-posttrain-fullbenchtasks_passed=5 · tasks_total=14 · mean_score=0.3571

Cited by — computed at build time, never stored