Home › Evidence › Records › cfg-0175
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0175

backfilledfork
citable URL: https://halobench.com/records/cfg-0175/ — this address never moves; the anchor /records/#cfg-0175 keeps resolving
modelQwen3.8-27B-ROCMFPX-MQ-Q4.gguf + Qwen3.8-27B-DFlash2-Q8_0.gguf · ROCmFPX-MQ-Q4 + DFlash2-Q8_0
engineLaurentZuijdwijk/llama.cpp 16f0799a6 · Vulkan0/RADV Mesa 26.0.3 · host aihydra (igpu)
flagsOperator-directed autonomous cache-reuse trace: target ROCmFPX-MQ-Q4 with DFlash2 Q8_0, -dev Vulkan0 -ngl 99 -c 98304 -fa on -t 16 -np 1 --jinja --spec-type draft-dflash --spec-draft-n-max 7 -ngld 99. Each completion request used temperature 0, n_predict 80, and cache_prompt true while appending a new synthetic section.
templatenot recorded at test time
treefork — carrying con-0013
aged evidence — reconstructed from the archive. The trace provides no artifact SHA256s, chat-template hash, standard house-guard result, or independent answer grading. The script's displayed depth column is the newly-prefilled-token count, not a machine-recorded cumulative-context field.

Capability basis

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

Runs on this config (1)

rundatekindsuitemetrics
run-06282026-08-24performanceqwen38-prod-tuning.cache-reuse-trace@45-turnturns_before_degradation=45

Cited by — computed at build time, never stored

claims clm-0121