cfg-0148
backfilledfork
citable URL: https://halobench.com/records/cfg-0148/ — this address never moves; the anchor /records/#cfg-0148 keeps resolving
| model | Qwen3.8-27B-Q8_0.gguf + Qwen3.8-27B-DFlash2-Q8_0.gguf · Q8_0 + DFlash2-Q8_0 · rev 57484e3196aaff8dbbd71c666158b00206baa2bc134e4f16ae90fc8cbdfbeea2 + 7f1c9a31a6ed40044c69f6508b50fd63b87abd8e1fb7fe4290303df549153751 |
| engine | ggml-org/llama.cpp 2586f6eddae19bef3dd21a1a0b109cca7bcf1c32 · Vulkan0/Strix-Halo · host aihydra (igpu) |
| flags | HO-004-DIAG v0.6.10 DFlash2 Stage-A sentinel fingerprint: identical target/runtime/backend/KV/batch/load-mode/-c 32768/--temp 0/--seed 4242 as cfg-0146 plus the external DFlash2 drafter /home/aihydra/models/qwen38-27b-dflash2/Qwen3.8-27B-DFlash2-Q8_0.gguf with -md DFlash2-Q8_0 -ngld 999 --spec-type draft-dflash --spec-draft-n-max 3 (card forbids n_max>=4; prior matched-r1 used 4). No stock MTP, no ngram/proxy/llama-swap/ROCm, no KV quant, no alternate target quant. |
| template | not recorded at test time |
| tree | fork — carrying con-0002, con-0009, con-0003 |
aged evidence — reconstructed from the archive. Recorded from the runner command line at review time. Sentinal outcome: DFlash2 Q8_0 n_max=3 PASS Stage-A (empty-assistant 0, empty-tool-call 0, single-token-EOS 0, code-valid 3/3, tool-correct 3/3) plus the 4/4 house guard, with log-backed DFlash2 draft activation (draft acceptance 0.8505). This config is ONLY a Stage-A sentinel/gate record: no throughput, no Stage-B production-depth performance, and no DFlash2 quality-equivalence or throughput-improvement claim (G2 varied-prompt invariance is NEGATIVE).
Capability basis
measured on this config
Runs on this config (1)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0574 | 2026-08-23 | guard | stageA-c32768-varied-prompts | tasks_passed=4 · tasks_total=4 |
Cited by — computed at build time, never stored
claims clm-0109