Home › Evidence › Records › cfg-0059

cfg-0059

backfilled
citable URL: https://halobench.com/records/cfg-0059/ — this address never moves; the anchor /records/#cfg-0059 keeps resolving
modelNVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_M-00001-of-00003.gguf · UD-Q4_K_M
engineggml-org/llama.cpp 3653e6d · rocm · host aihydra (igpu)
flags-ngl 999 -fa 1 -b 2048 -ub 512 -ctk f16 -ctv f16 -t 16 --load-mode none
templatenot recorded at test time
treeupstream — stock
aged evidence — reconstructed from the archive. Same arm as cfg-0057 (nemotron3-super-rocm-kvf16-faon-d32768), re-run 2026-08-14 with the one flag cfg-0057 was missing: --load-mode none. cfg-0057 and its predecessor perf-matrix cell both OOM'd (run-0234; status.tsv on aihydra for the original 2026-08-13 cell) because the perf-matrix/perf-matrix-gtt120 llama-bench queue scripts never carried --load-mode none, so ROCm was benchmarked under mmap on a model within ~40 GiB of the box's full 122 GiB — see clm-0053. On the identical binary with mmap off, this cell runs cleanly: run-0236, run-0237. No memory block recorded — ingested from llama-bench, which measures throughput and not footprint; everything present here (build commit, backend, KV quant, batch sizes, flash attention, load-mode) is machine-reported at launch.

Capability basis

unmeasured — performance numbers on this config stand on a guard alone, not an established capability

Runs on this config (4)

rundatekindsuitemetrics
run-02362026-08-14performancellama-bench@1prefill_tps=244.051835
run-02372026-08-14performancellama-bench@1decode_tps=16.434903
run-02432026-08-14performancellama-bench@1prefill_tps=248.788416
run-02442026-08-14performancellama-bench@1decode_tps=16.390674