cfg-0007
citable URL: https://halobench.com/records/cfg-0007/ — this address never moves; the anchor /records/#cfg-0007 keeps resolving
| model | unsloth/Qwen3.5-122B-A10B-MTP-GGUF · UD-Q4_K_M · rev main |
| engine | ggml-org/llama.cpp 3653e6d6d547ec763317d9ecd0ace334a7e21359 · rocm · host aihydra (igpu) |
| flags | -ngl 999 -fa on -c 16384 --parallel 1 --load-mode none --jinja -ctk f16 -ctv f16 --spec-type draft-mtp --spec-draft-n-max 3 |
| template | not recorded at test time |
| tree | upstream — stock |
Memory — static estimate vs observed peak
context 16,384 × 1 slot(s) = 16,384 tokens
⌁ total 76.98 GiB of 105 GiB pool · ⌁ headroom 28.02 GiB
⌁ risk utilisation 73.3% — basis estimate (the estimate has never predicted an OOM — clm-0001)
Capability basis
inherited — chain: cfg-0007 (speculation) ← cfg-0006 · inheritance is legal across neutral levers only (protocol §1a)
Runs on this config (1)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0007 | 2026-08-08 | performance | spec-ab@v1 | decode_tps=31.71 · decode_tps_no_speculation=21.97 |