Home › Evidence › Records › cfg-0174
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0174

backfilledfork
citable URL: https://halobench.com/records/cfg-0174/ — this address never moves; the anchor /records/#cfg-0174 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 tuning served-path screen: target ROCmFPX-MQ-Q4 with DFlash2 Q8_0 drafter, -dev Vulkan0 -ngl 99 -c 32768 -fa on -t 16 -np 1 --jinja --spec-type draft-dflash --spec-draft-n-max 7 -ngld 99. The separate competence smoke used native MTP n_max 4 with reasoning medium, temperature 0.1, max_tokens 8192, tasks 0-4, seed 42, and the pinned Haiku simulator.
templatenot recorded at test time
treefork — carrying con-0013
aged evidence — reconstructed from the archive. The archived handoff does not provide target/drafter SHA256s, a chat-template hash, a standard house-guard artifact, a plain/spec-off paired served-path floor, or a llama-bench floor. The server samples are retained as screening evidence, not a production-performance or quant-equivalence admission.

Capability basis

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

Runs on this config (3)

rundatekindsuitemetrics
run-06262026-08-24performanceqwen38-prod-tuning.server-screen-code@1decode_tps=52.2
run-06272026-08-24capabilitytau2-airline-smoke@tasks-0-4tasks_passed=5 · tasks_total=5 · mean_score=1
run-06312026-08-24performanceqwen38-prod-tuning.server-screen-prose@1decode_tps=31.5

Cited by — computed at build time, never stored

claims clm-0120