cfg-0150
backfilledfork
citable URL: https://halobench.com/records/cfg-0150/ — this address never moves; the anchor /records/#cfg-0150 keeps resolving
| model | Qwen3.6-35B-A3B-MTP-UD-Q4_K_M.gguf · UD-Q4_K_M · rev 0b21525e972670ed59e1812e170b27c26355381f0656ecc4e25617ece7dac58b |
| engine | ggml-org/llama.cpp 2586f6eddae19bef3dd21a1a0b109cca7bcf1c32 · Vulkan0/Strix-Halo (strix-halo-v0610 fork via RADV, NOT ggml-org ROCm0) · host aihydra (igpu) |
| flags | HO-009-v0610 n2/MTP (reasoning-OFF) capability fingerprint on the v0.6.10 fork/Vulkan runtime, mirroring cfg-0145 (same model/quant/flags as cfg-0149 PLUS --spec-type draft-mtp --spec-draft-n-max 2 -ctkd f16 -ctvd f16; draft KV stays f16 = no KV quant). tau2-protocol agent settings, -rea off on the tau2 path, f16/f16 target KV, -ngl 999 -fa on -c 32768 --load-mode none --parallel 1 --jinja -b 2048 -ub 512 -t 16, temp 0 seed 42 max_tokens 4096. n_max=2 ONLY; no n_max>=4, no KV quant, no alt drafter. Built on the v0.6.10 fork under Vulkan/RADV vs upstream 7077abb ROCm0 — repo + backend + build all changed together (stack effect). |
| template | not recorded at test time |
| tree | fork — carrying con-0002, con-0009 |
aged evidence — reconstructed from the archive. Recorded from the raw tau2 artifacts and the v0.6.10 fork MANIFEST (payload source 2586f6edd, Vulkan/RADV). The -rea off delta is the deliberate DIAG-recommended fix from cfg-0145. Scoped to the v0.6.10-fork-on-Vulkan vs 7077abb-upstream-on-ROCm0 capability recovery question only; no DFlash2/stock-MTP/n_max>=4/reasoning-on/alt-model/production-throughput claim.
Capability basis
measured on this config
Runs on this config (2)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0577 | 2026-08-23 | guard | guard-c32768-depth8000 | tasks_passed=4 · tasks_total=4 |
| run-0578 | 2026-08-23 | capability | tau2-bench-airline@668d3bc | tasks_passed=17 · tasks_total=26 · mean_score=0.6538461538461539 |
Cited by — computed at build time, never stored
claims clm-0110