Home › Evidence › Records › cfg-0166
⚠ This notebook has stopped taking notes: newest record is 23 days old (eng-0280), against an expected cadence of 14 days.

cfg-0166

backfilledfork
citable URL: https://halobench.com/records/cfg-0166/ — this address never moves; the anchor /records/#cfg-0166 keeps resolving
modelQwen3.6-35B-A3B-MTP-UD-Q4_K_M.gguf · UD-Q4_K_M · rev 0b21525e972670ed59e1812e170b27c26355381f0656ecc4e25617ece7dac58b
engineggml-org/llama.cpp 2586f6eddae19bef3dd21a1a0b109cca7bcf1c32 · Vulkan0/Strix-Halo (strix-halo-v0610 fork via RADV, NOT ggml-org ROCm0) · host aihydra (igpu)
flagsHO-014 PRODUCTION config b1024/ub512 f16/f16 KV — the Cell-1 best-TTFT arm (Cell-1 tdepths), and the base for the Cell-2 depth probe (d131072→d200000→ d262144) and the RUN Cell-3 f16 arm. llama-bench pp512/tg256 N=5 d131072 for the Cell-1 latency arm (command c1-b1024-ub512.command.txt); pp16384/tg256 N=3 for the Cell-2 depth ladder. -m Qwen3.6-35B-A3B-UD-Q4_K_M.gguf -ngl 999 -fa on -b 1024 -ub 512 --load-mode none -dev Vulkan0 -ctk f16 -ctv f16 -t 16, build 2586f6edd Vulkan/RADV, plain mode (-rea off, no MTP). Model sha/bytes re-verified by RUN gate G0. Same model/quant/flags as cfg-0149 but the production b/ub choice for interactive serving.
templatenot recorded at test time
treefork — carrying con-0002, con-0009
aged evidence — reconstructed from the archive. From HO-014 reviewer-admitted FILL + RUN evidence (t_dc235dcd ADMIT: summary.tsv best-ttft b1024/ub512 1353.1 ms; cell2-summary.tsv; RUN cell3 f16 arm at this config). batch/ubatch + depth are perf levers; the house guard (4/4, RUN + FILL both) clears binary/hazard levers only and no new capability claim is made or inherited. Intra-matrix/production-optimisation only; no cross-model/build/backend claim.

Capability basis

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

Runs on this config (6)

rundatekindsuitemetrics
run-06072026-08-23performanceho014a-fill-r1.cell1-latencydecode_tps=35.1 · ttft_ms=1353.1
run-06122026-08-23performanceho014a-fill-r1.cell2-depthdecode_tps=35 · prefill_tps=356.8
run-06132026-08-23performanceho014a-fill-r1.cell2-depthdecode_tps=28.8 · prefill_tps=238.8
run-06142026-08-23performanceho014a-fill-r1.cell2-depthdecode_tps=24.9 · prefill_tps=167.4
run-06152026-08-23guardho014-run-kv-overlap.guard-c32768-depth8000tasks_passed=4 · tasks_total=4
run-06162026-08-23performanceho014-run-kv-overlap.cell3b-bench-pairdecode_tps=35.08 · ttft_ms=1357.2

Cited by — computed at build time, never stored

claims clm-0117