Home › Evidence › Records › cfg-0063

cfg-0063

backfilled
citable URL: https://halobench.com/records/cfg-0063/ — this address never moves; the anchor /records/#cfg-0063 keeps resolving
modelQwen3.8-27B-UD-Q4_K_XL.gguf · UD-Q4_K_XL
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. Ingested from llama-bench, which measures throughput and not footprint: the memory block, chat-template hash and model revision are absent by construction. Peak memory comes from a Tier 0 capability probe, not from this tool. Everything present here — build commit, backend, KV quant, batch sizes, flash attention — is machine-reported at launch. --load-mode none comes from the queue invocation itself (bench/queue-qwen38-screen.sh passes -lm none on every rep), added per clm-0053's rule that benchmark flags must match the serving standard on unified memory.

Capability basis

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

Runs on this config (6)

rundatekindsuitemetrics
run-02512026-08-14performancellama-bench@1prefill_tps=346.49
run-02522026-08-14performancellama-bench@1decode_tps=11.4415
run-02532026-08-14performancellama-bench@1prefill_tps=207.606
run-02542026-08-14performancellama-bench@1decode_tps=10.3626
run-02552026-08-14performancellama-bench@1prefill_tps=95.5937
run-02562026-08-14performancellama-bench@1decode_tps=8.13583

Cited by — computed at build time, never stored

model pages qwen38-27b