cfg-0061
backfilled
citable URL: https://halobench.com/records/cfg-0061/ — this address never moves; the anchor /records/#cfg-0061 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. Same llama-bench invocation flags as cfg-0057 (mmap default, no --load-mode) — the difference this config exists to isolate is a KERNEL BOOT parameter, not a benchmark flag: this arm was measured on a boot with amdgpu.no_system_mem_limit=1 added to the cmdline (queue.log, memgate-test STAGE=postboot: "no_system_mem_limit: Y"), specifically to test whether that flag changes the mmap-under-memory-pressure failure mode. It did not prevent the failure, but it changed its SHAPE: see run-0242. Ingested from nemotron-postboot-mmap-rep1.stderr/queue.log/journalctl (aihydra ~/bench-results/memgate-test/) rather than a completed llama-bench JSON, because the process never produced one. No memory block recorded — same gaps as cfg-0057.
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-0242 | 2026-08-14 | performance | memgate-test@1 | oom=true |