cfg-0057
backfilled
citable URL: https://halobench.com/records/cfg-0057/ — this address never moves; the anchor /records/#cfg-0057 keeps resolving
| model | NVIDIA-Nemotron-3-Super-120B-A12B-UD-Q4_K_M-00001-of-00003.gguf · UD-Q4_K_M |
| engine | ggml-org/llama.cpp 3653e6d · rocm · host aihydra (igpu) |
| flags | -ngl 999 -fa 1 -b 2048 -ub 512 -ctk f16 -ctv f16 -t 16 |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. perf-matrix-gtt120 re-test: the same nemotron3-super ROCm f16 cell that OOM'd in the main perf-matrix run (2026-08-13, default GTT ceiling), re-run 2026-08-14 with the kernel boot param amdgpu.gttsize=122880 (122 GiB GTT, vs the fleet default ~84 GiB — see strix-halo-host-config-reference) to test whether more GTT headroom rescues the cell. It did not: see run-0234. No memory block recorded — the process never reached a state where weights+KV footprint could be measured; it thrashed for 55 minutes before the kernel refused further SVM mapping. Everything present here is machine-asserted at launch (protocol_check_build), not measured mid-run.
Capability basis
unmeasured — performance numbers on this config stand on a guard alone, not an established capability
Runs on this config (1)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0234 | 2026-08-14 | performance | perf-matrix-gtt120@1 | oom=true |