Home › Evidence › Records › cfg-0158
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0158

backfilled
citable URL: https://halobench.com/records/cfg-0158/ — this address never moves; the anchor /records/#cfg-0158 keeps resolving
modelQwen3.5-122B-A10B-UD-Q4_K_M-00001-of-00003.gguf · UD-Q4_K_M
engineggml-org/llama.cpp ce7689f · ROCm0/gfx1151 · host aihydra (igpu)
flagsHO-012 Cell-3 KV q4_0 arm:llama-server build ce7689f (llama.cpp-kvfix), -ngl 999 -fa 1 -c 32768 --parallel 1 --load-mode none --jinja -rea off -ctk q4_0 -ctv q4_0 -b 2048 -ub 512 -t 16 -n 4096, ROCm0 Radeon 8060S gfx1151 122880 MiB. Identical to cfg-0151 EXCEPT -ctk q4_0 -ctv q4_0 (KV quant lever). From server-c3-q4_0.command.txt.
templatenot recorded at test time
treeupstream — stock
aged evidence — reconstructed from the archive. From reviewer-admitted HO-012 evidence. kv_quant is a LOSSY capability lever, so capability is NOT inherited; this arm is a throughput/delta cell only, guarded 4/4, no capability claim. q4_0/q4_0 KV, b2048/ub512. Intra-suite only; the q4_0 decode regression vs f16 is recorded truthfully (see clm-0111), never as equivalence.

Capability basis

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

Runs on this config (2)

rundatekindsuitemetrics
run-05872026-08-23performanceho012-122b-prod-matrix-r2.cell3-q4_0decode_tps=21.8276 · ttft_ms=1418.99
run-05902026-08-23guardguard-c32768-depth8000tasks_passed=4 · tasks_total=4

Cited by — computed at build time, never stored

claims clm-0111