cfg-0130
backfilled
citable URL: https://halobench.com/records/cfg-0130/ — this address never moves; the anchor /records/#cfg-0130 keeps resolving
| model | Ornith-1.0-35B-UD-Q4_K_XL.gguf · UD-Q4_K_XL |
| engine | ggml-org/llama.cpp 3653e6d · rocm · host aihydra (igpu) |
| flags | Throughput/guard boundary: -ngl 999 -fa 1 -b 2048 -ub 512 --load-mode none -ctk f16 -ctv f16 -t 16; llama-bench rows add pp1024/tg256 depth flags. Guard used c32768/depth8000 with raw llama-server, --parallel 1 / -np 1 semantics, no speculation, no MTP, no drafter, no proxy. |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. HO-005 reviewer-admitted production-depth performance matrix from /home/aihydra/bench-results/ho005-ornith-q4q8-tuning-matrix-r2. Artifact identity is the unsloth UD-Q4_K_XL file already recorded on ornith-35b (22,324,804,000 bytes, sha256 67081ae4a1a291bd6c72834094ea056332cb3cb5fa15e88536ec7f233a475b71). Bench JSON reports build_commit 3653e6d/build_number 1 on ROCm, while runtime source-path inspection during review found /home/aihydra/src/llama.cpp clean at HEAD ce7689f875f724ece08b1a9fbb2a83971120b4d2; preserve both rather than rewriting the source-tree identity. This config records guarded performance only, not tau2 capability and not a production recommendation.
Capability basis
measured on this config
Runs on this config (5)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0501 | 2026-08-20 | guard | guard-c32768-depth8000 | tasks_passed=4 · tasks_total=4 |
| run-0503 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=177.252767 |
| run-0504 | 2026-08-20 | performance | llama-bench@1 | decode_tps=22.967537 |
| run-0505 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=143.866662 |
| run-0506 | 2026-08-20 | performance | llama-bench@1 | decode_tps=19.943102 |