cfg-0171
backfilledfork
citable URL: https://halobench.com/records/cfg-0171/ — this address never moves; the anchor /records/#cfg-0171 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-014 RUN Cell-3 q4_0/q4_0 KV arm at the production config b1024/ub512 d131072, f16 baseline comparison arm is cfg-0166. llama-bench pp512/tg256 N=5 d131072, -m Qwen3.6-35B-A3B-UD-Q4_K_M.gguf -ngl 999 -fa on -b 1024 -ub 512 --load-mode none -dev Vulkan0 -ctk q4_0 -ctv q4_0 -t 16 -p 512 -n 256 -d 131072 -r 5, build 2586f6edd Vulkan/RADV, plain mode (-rea off, no MTP, no spec). Identical to cfg-0166 EXCEPT -ctk q4_0 -ctv q4_0 (KV quant lever). The Cell-3A KLD (f16 logits, replay q4_0) was measured at capability context c32768, both arms identical, after the initial ctx=131072 arm hit an instrument logits-buffer memory wall (~96 GiB RSS, earlyoom — not model OOM). |
| template | not recorded at test time |
| tree | fork — carrying con-0002, con-0009 |
aged evidence — reconstructed from the archive. From HO-014-RUN reviewer-admitted Cell-3 evidence (cell3-verdict.md; G3 arch pre-probe PASS 11 quantisable-KV layers blk3..40 step4; bench pair + KLD). KV quant is a LOSSY capability lever, so capability is NOT inherited — this arm is a throughput/quality delta cell only. Guarded 4/4 (run-0615). No cross-model or equivalence claim; the small q4_0 delta is partly ARCHITECTURAL (hybrid: only 11/40 layers carry a quantisable full-attention KV).
Capability basis
unmeasured — performance numbers on this config stand on a guard alone, not an established capability
Runs on this config (1)
| run | date | kind | suite | metrics |
|---|---|---|---|---|
| run-0617 | 2026-08-23 | performance | ho014-run-kv-overlap.cell3b-bench-pair | decode_tps=46.15 · ttft_ms=1334.2 |
Cited by — computed at build time, never stored
claims clm-0117