Qwen3.6-27B start script: the 262,144 rung promised an 8-bit cache and its launch line expanded an array nothing ever assigned. Excerpts of two versions of the same script, with their line numbers. Paths, the bridge address and the settings-variable prefix are redacted as data/README.md states; em dashes in our own comment and message lines became colons (counted there). Nothing else is changed. The comment on line 77 (line 90 in part B) carries figures of its own, 27,947 MiB, 71.8 t/s, "healthy in 13 s" and a 16 GB f16 cache: they are the script's comment, no log in this package shows the run behind them, and the page does not use them. Backup copies are named for the day they were taken; a copy's file time is the last time the script had been edited before that day. So "bak-20260921-141932" (file time 2026-09-17 16:10) is the script as it stood from 17 September 16:10, and "bak-20260926-256k" (file time 2026-09-21 18:37) is the script as it stood until the 26 September edit that added the assignment. History of the two strings across the copies on disk (grep for KVARGS and the 262144 comment): start_server.sh.bak-20260917-256k file time 2026-09-16 18:37 (no KVARGS, no 262144 comment) lines assigning KVARGS: none start_server.sh.bak-20260921-141932-m27 file time 2026-09-17 16:10 line 77: # 262144 MEASURED 2026-09-17: 27,947 MiB VRAM, 71.8 t/s, healthy in 13 s: but ONLY with a line 195: ${KVARGS[@]+"${KVARGS[@]}"} \ lines assigning KVARGS: none start_server.sh.bak-20260921-batch file time 2026-09-21 14:19 line 77: # 262144 MEASURED 2026-09-17: 27,947 MiB VRAM, 71.8 t/s, healthy in 13 s: but ONLY with a line 196: ${KVARGS[@]+"${KVARGS[@]}"} \ lines assigning KVARGS: none start_server.sh.bak-20260926-256k file time 2026-09-21 18:37 line 77: # 262144 MEASURED 2026-09-17: 27,947 MiB VRAM, 71.8 t/s, healthy in 13 s: but ONLY with a line 207: ${KVARGS[@]+"${KVARGS[@]}"} \ lines assigning KVARGS: none start_server.sh file time 2026-09-26 22:28 line 90: # 262144 MEASURED 2026-09-17: 27,947 MiB VRAM, 71.8 t/s, healthy in 13 s: but ONLY with a line 191: KVARGS=() line 193: KVARGS=(--cache-type-k q8_0 --cache-type-v q8_0) line 234: ${KVARGS[@]+"${KVARGS[@]}"} \ lines assigning KVARGS: [191, 193] ==== A. The script as it stood from 17 September 16:10 to 26 September (copy bak-20260921-batch, file time 2026-09-21 14:19) ---- the settings file is sourced, then the window is read (lines 66 to 70) 66 set -euo pipefail 67 ENVF=/ 68 [ -f "$ENVF" ] && . "$ENVF" 69 CTX="${QWEN36_CTX:-131072}" 70 HOST="${QWEN36_HOST:-}"; PORT="${QWEN36_PORT:-}" ---- the promise, the window check that accepts 262144, and the error text that says it is not offered (lines 77 to 84; line 76, a heading comment, is not shipped) 77 # 262144 MEASURED 2026-09-17: 27,947 MiB VRAM, 71.8 t/s, healthy in 13 s: but ONLY with a 78 # q8_0 KV cache, which this rung sets for you below. At f16 the KV alone wants 16 GB and the 79 # context fails to allocate. 80 case "$CTX" in 32768|65536|131072|262144) ;; *) 81 echo "ERROR: QWEN36_CTX=$CTX is not a validated context (32768|65536|131072; default 131072)." >&2 82 echo " 262144 is the ARCHITECTURE ceiling from the GGUF and is deliberately not offered." >&2 83 exit 2 ;; 84 esac ---- the launch line expands KVARGS, which no line of this file assigns (lines 192 to 201) 192 exec "$BIN" \ 193 --model "$MODEL" \ 194 --host "${HOST}" --port "${PORT}" --alias qwen3.6-27b \ 195 --jinja --ctx-size "${CTX}" --parallel 1 \ 196 ${KVARGS[@]+"${KVARGS[@]}"} \ 197 ${THINKARGS[@]+"${THINKARGS[@]}"} \ 198 --n-gpu-layers 999 \ 199 --threads 24 --threads-batch 24 \ 200 --flash-attn on --fit off --cors-origins localhost --timeout 3600 \ 201 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0 ==== B. The script after the 26 September edit (start_server.sh, file time 2026-09-26 22:28; another session edited the file again that evening, so later copies may differ in other places) ---- lines 66 to 70 66 set -euo pipefail 67 ENVF=/ 68 [ -f "$ENVF" ] && . "$ENVF" 69 CTX="${QWEN36_CTX:-131072}" 70 HOST="${QWEN36_HOST:-}"; PORT="${QWEN36_PORT:-}" ---- the same promise and check (lines 90 to 97; line 89, a heading comment, is not shipped) 90 # 262144 MEASURED 2026-09-17: 27,947 MiB VRAM, 71.8 t/s, healthy in 13 s: but ONLY with a 91 # q8_0 KV cache, which this rung sets for you below. At f16 the KV alone wants 16 GB and the 92 # context fails to allocate. 93 case "$CTX" in 32768|65536|131072|262144) ;; *) 94 echo "ERROR: QWEN36_CTX=$CTX is not a validated context (262144 default since 2026-09-26 | 131072 | 65536 | 32768)." >&2 95 echo " 262144 is offered too (q8_0 KV, MTP off; measured 2026-09-17 and 2026-09-26)." >&2 96 exit 2 ;; 97 esac ---- the assignment that was missing, with the script's own account (lines 188 to 195) 188 # --- the KV block the 262144 rung always needed (it was referenced below but never defined, so 256K started 189 # with an f16 cache that cannot fit). MTP does not fit beside a 256K cache (measured 2026-09-26: "failed to 190 # create MTP context"), so the rung turns it off, as Qwen3.8-27B's script does. 191 KVARGS=() 192 if [ "$CTX" = 262144 ]; then 193 KVARGS=(--cache-type-k q8_0 --cache-type-v q8_0) 194 [ "${#SPEC[@]}" -gt 0 ] && { echo "NOTE: ctx 262144 forces MTP off: it does not fit beside a 256K cache." >&2; SPEC=(); } 195 fi ---- the launch line (lines 229 to 240) 229 exec "$BIN" \ 230 --model "$MODEL" \ 231 --host "${HOST}" --port "${PORT}" --alias qwen3.6-27b \ 232 --jinja --ctx-size "${CTX}" --parallel 1 \ 233 --batch-size "${QWEN36_BATCH}" --ubatch-size "${QWEN36_UBATCH}" \ 234 ${KVARGS[@]+"${KVARGS[@]}"} \ 235 ${SPEC[@]+"${SPEC[@]}"} \ 236 ${THINKARGS[@]+"${THINKARGS[@]}"} \ 237 --n-gpu-layers 999 \ 238 --threads 24 --threads-batch 24 \ 239 --flash-attn on --fit off --cors-origins localhost --timeout 3600 \ 240 --temp 1.0 --top-p 0.95 --top-k 20 --min-p 0.0