Home › Evidence › Records › cfg-0164
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0164

backfilledfork
citable URL: https://halobench.com/records/cfg-0164/ — this address never moves; the anchor /records/#cfg-0164 keeps resolving
modelQwen3.8-27B-IU4-Kairic-Edge.gguf · IU4 · rev 360caf7381907c3eca7ac0afd1228efc016af747f3f38637fb1c7f94daabac2a
engineciru-ai/ROCmFPX e1da26bb8 · rocm · host aihydra (igpu)
flagsThe Kairic Edge TheRock served config at c32768 on this build (compat mode; prompts even fast-greedy forbids tool calls) used by HK-001-FULLBENCH. Generic flag set (per the fullbench launch script): -m Qwen3.8-27B-IU4-Kairic-Edge.gguf -dev ROCm0 -ngl 999 -c 32768 -b 2048 -ub 512 -fa on -ctk f16 -ctv f16 -t 16 -tb 32 -np 1 --no-mmap --kairic-edge --metrics -n 4096 --slots with the PROMPTFORGE_* env block (PROMPTFORGE_MODE=iu4_ffn, IU4/GDN/GDN-Output sidecar pfs paths, HADAMARD+SEGMENTED switches). PLAIN spec only — NO MTP (n_max>=4 cliff on the Qwen3.8 family, clm-0055). --kairic-edge and the sidecars are server-only; llama-bench (the engine path that produced the depth-ladder cells) can NOT carry the flag or the GDN row, so the depth cells are recorded as engine-path on this same build+sidecar-env, per protocol 1d.
templatenot recorded at test time
treefork — carrying con-0012
aged evidence — reconstructed from the archive. Fingerprint reconstructed from the HK-001-FULLBENCH launch script and build doc (hk001-results-r1.md + scratch/hk001-launch.sh) — exact start_server invocation not re-verified on-box (the runner's logs are on aihydra, build commit e1da26bb8 / binary e121a4d3; model sha 360caf73; sidecar hashes adcbb90a / 82f93131 / 3b07e7b1). Capability is measured (guard exercised 3/3 + isolation N/A), everything on the launched served config; the throughput/depth cells are engine-path (llama-bench). Single config record for this build; every HK-001-FULLBENCH run below points at it.

Capability basis

measured on this config

Runs on this config (4)

rundatekindsuitemetrics
run-06012026-08-23guardhk001-fullbench-guard@c32768tasks_passed=3 · tasks_total=3
run-06022026-08-23performancehk001-fullbench-fit@c32768decode_tps=11.73 · prefill_tps=279.6
run-06032026-08-23performancehk001-fullbench-depth-r3@llama-benchdecode_tps=13.14 · prefill_tps=612.3
run-06042026-08-23performancehk001-fullbench-depth-r3@llama-benchdecode_tps=11.66 · prefill_tps=272.1

Cited by — computed at build time, never stored

claims clm-0116