Home › Evidence › Records › cfg-0179
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0179

fork
citable URL: https://halobench.com/records/cfg-0179/ — this address never moves; the anchor /records/#cfg-0179 keeps resolving
modelQwen3.8-Flash-Next-Q4_0-ROCmFP4-STRIX.gguf · Q4_0_ROCmFP4_STRIX · rev c6770d7442a06bf1d78edf28cec83e1ec93afdd34664c23ff898807b6b9349fa
enginekingjones30/ROCmFPX 36e9acd40e10a87cd3c3ef8ec734668757dc8520 · HIP/gfx1151 · host aihydra (igpu)
flagsNo model command executed. The fixed intended uninstrumented serving boundary was one slot, --n-gpu-layers 999, --flash-attn on, --fit off, --jinja, runtime-default mmap, and no MTP, draft or speculation. The separately required receipt-only capture variant attempted a Release HIP/gfx1151/native targeted llama-server build and stopped rc=2 because examples/gguf-hash/deps/sha256/sha256.c could not include rotate-bits/rotate-bits.h, despite the exact tracked header being present in the source tree. Resolved template, thinking, KV, batch, ubatch, cache and thread values were never produced because the campaign stopped before server launch.
templatenot recorded at test time
treefork — carrying con-0017

Capability basis

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

Runs on this config (1)

rundatekindsuitemetrics
run-06372026-08-29guardq38fn-kingjones-r2-targeted-receipt-build-gate-v1runner_invalid=true

Cited by — computed at build time, never stored

claims clm-0126