cfg-0060
backfilled
citable URL: https://halobench.com/records/cfg-0060/ — this address never moves; the anchor /records/#cfg-0060 keeps resolving
| model | Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf · UD-Q4_K_XL |
| 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 --load-mode none |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. cfg-0032 (the backend-decider ROCm f16 d32768 arm behind clm-0050's Qwen3.6-35B numbers, benchmarked under llama-bench's mmap default) with --load-mode none added and nothing else changed — the matched no-mmap companion built 2026-08-14 to test whether the mmap default is a throughput confound for a model that comfortably fits in memory (clm-0053's Qwen3.6-35B control pair; contrast Nemotron-3-Super, which does not fit and could not run under mmap at all — cfg-0057/run-0234). Same recording gaps as cfg-0032: ingested from llama-bench, no memory block, machine- reported build fingerprint only.
Capability basis
unmeasured — performance numbers on this config stand on a guard alone, not an established capability