cfg-0129
backfilled
citable URL: https://halobench.com/records/cfg-0129/ — this address never moves; the anchor /records/#cfg-0129 keeps resolving
| model | Ornith-1.0-35B-UD-Q4_K_XL.gguf · UD-Q4_K_XL |
| engine | ggml-org/llama.cpp 3653e6d6d · vulkan · host aihydra (igpu) |
| flags | Serving/capability boundary: -ngl 999 -fa 1 -b 2048 -ub 512 -c 32768 -dev Vulkan0 -ctk f16 -ctv f16 -t 16 --load-mode none --parallel 1 --jinja --reasoning-format deepseek -n 4096 --slots. Bounded throughput rows using this same artifact/runtime/backend/f16-KV boundary add their llama-bench-specific -p/-n/-d/-o json flags on the run records. |
| template | not recorded at test time |
| tree | upstream — stock |
aged evidence — reconstructed from the archive. Bounded HO-005 r3 preflight/smoke ingestion only, admitted by hbreviewer t_2cabe4fa from /home/aihydra/bench-results/ho005-ornith-q4kxl-vulkan-preflight-r3, then a successor admitted full-26 tau2 capability run added run-0498 and run-0499 on the same artifact/runtime/backend/f16-KV serving fingerprint. Artifact identity: /home/aihydra/models/ornith-35b/Ornith-1.0-35B-UD-Q4_K_XL.gguf, 22,324,804,000 bytes, sha256 67081ae4a1a291bd6c72834094ea056332cb3cb5fa15e88536ec7f233a475b71, sidecar/local sha match, source unsloth/Ornith-1.0-35B-GGUF per candidate record. Build/backend fingerprint: /home/aihydra/src/llama.cpp-vk3653e6d HEAD 3653e6d6d547ec763317d9ecd0ace334a7e21359, build_commit 3653e6d6d, Vulkan0 Radeon 8060S, f16 KV, --load-mode none, tau2 head 668d3bcd135c02aa3438f987ef45735b7c163ee3. FIT/header/load proof was healthy at c32768 (2026-08-20T00:17:59Z..00:18:15Z, rc=0, gtt_used_bytes 22444122112). The later full-26 arm used the exact serving flags recorded above, passed a fresh 4/4 guard, and measured 22/26 with mean_reward 0.8461538461538461 at Q4_K_XL. capability_basis is measured for this Q4_K_XL boundary only; it is not inherited from Q8_0 and does not establish Q4/Q8 equivalence.
HO-005 reviewer-admitted production-depth performance matrix r2 later reused this same UD-Q4_K_XL / Vulkan0 / f16/f16 KV / build_commit 3653e6d6d throughput boundary for llama-bench pp1024/tg256 cells at depths 0/32768/131072/204800/262144 (run-0507 through run-0516) after a fresh schema-visible guard (run-0500). These are performance records only, not a new tau2 capability claim, not Q4/Q8 equivalence, and not a broad production recommendation; energy is structured-unjoined on the new records because the hbrunner ingestion profile lacked Home Assistant history credentials.
Capability basis
measured on this config
Runs on this config (17)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0494 | 2026-08-20 | guard | ho005-ornith-q4kxl-vulkan-preflight-r3 | tasks_passed=4 · tasks_total=4 |
| run-0495 | 2026-08-20 | capability | ho005-ornith-q4kxl-vulkan-preflight-r3 | tool_call_success=1 · empty_arg_calls=0 · tasks_passed=5 · tasks_total=5 · mean_score=1 |
| run-0496 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=86.158933 |
| run-0497 | 2026-08-20 | performance | llama-bench@1 | decode_tps=24.312862 |
| run-0498 | 2026-08-20 | guard | guard-c32768-depth8000 | tasks_passed=4 · tasks_total=4 |
| run-0499 | 2026-08-20 | capability | tau2-bench-airline@668d3bc | tasks_passed=22 · tasks_total=26 · mean_score=0.8461538461538461 |
| run-0500 | 2026-08-20 | guard | guard-c32768-depth8000 | tasks_passed=4 · tasks_total=4 |
| run-0507 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=1098.256763 |
| run-0508 | 2026-08-20 | performance | llama-bench@1 | decode_tps=61.803053 |
| run-0509 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=695.045118 |
| run-0510 | 2026-08-20 | performance | llama-bench@1 | decode_tps=50.592308 |
| run-0511 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=277.8588 |
| run-0512 | 2026-08-20 | performance | llama-bench@1 | decode_tps=34.330286 |
| run-0513 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=126.770673 |
| run-0514 | 2026-08-20 | performance | llama-bench@1 | decode_tps=27.868642 |
| run-0515 | 2026-08-20 | performance | llama-bench@1 | prefill_tps=85.675048 |
| run-0516 | 2026-08-20 | performance | llama-bench@1 | decode_tps=24.286389 |
Cited by — computed at build time, never stored
model pages ornith-35b