Home › Evidence › Records › cfg-0103
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0103

backfilled
citable URL: https://halobench.com/records/cfg-0103/ — this address never moves; the anchor /records/#cfg-0103 keeps resolving
modelQwen3.8-27B-Q8_0.gguf · Q8_0
engineggml-org/llama.cpp 3653e6d · rocm · host aihydra (igpu)
flags-ngl 999 -fa 1 -b 2048 -ub 512 -ctk f16 -ctv f16 -t 16 --load-mode none
templatenot recorded at test time
treeupstream — stock
aged evidence — reconstructed from the archive. Depth-ladder backfill session (2026-08-17), same fingerprint as cfg-0062 (identical llama-bench invocation: build 3653e6d, ROCm, Q8_0, f16 KV, -fa 1, --load-mode none — only depth differs, which is not a fingerprint field). A fresh config id is minted per house convention (each backfill invocation gets its own config record rather than silently appending to a prior session's), matching the qwen36-27b-mtp d204800 backfill precedent (cfg-0100 vs its own cfg-0097). Ingested from llama-bench, which measures throughput and not footprint: the memory block, chat-template hash and model revision are absent by construction. Everything present here — build commit, backend, KV quant, batch sizes, flash attention, load-mode — is machine-reported at launch.

Capability basis

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

Runs on this config (4)

rundatekindsuitemetrics
run-03732026-08-17performancellama-bench@1prefill_tps=61.002739
run-03742026-08-17performancellama-bench@1decode_tps=5.449447
run-03752026-08-17performancellama-bench@1prefill_tps=51.225787
run-03762026-08-17performancellama-bench@1decode_tps=5.025283

Cited by — computed at build time, never stored

model pages qwen38-27b