catalog: add beellama/qwopus-coder (Qwopus3.6-27B-Coder, KVarN-4) — first KVarN compose (#391)
* catalog: add beellama/qwopus-coder (Qwopus3.6-27B-Coder, KVarN-4) — first KVarN compose
Wires the Qwopus3.6-27B-Coder Q5_K_M GGUF (Jackrong, embedded MTP head) into
the curated catalog on beellama + the new KVarN-4 KV format. Single 3090,
kvarn4/kvarn4 @ 160K + MTP (230K no-MTP env opt-in). 🧪 experimental (--force),
gated on the v0.3.2-preview KVarN engine build (digest-pinned, merged earlier).
- engines/beellama-local.yml + hardware/rtx-3090.yml: register `kvarn4` KV format
- drafters/qwopus-mtp-gguf.yml: embedded-MTP-GGUF drafter (Jackrong)
- models/qwen3.6-27b.yml: qwopus-coder-mtp-q5km weights entry (Jackrong GGUF)
- compose_registry.py: `beellama/qwopus-coder` slug (kvcalc_key SKIP, port 8067)
- compose mtp.yml: git-tracked; image fallback → the KVarN digest (the v0.3.0
multiarch fallback rejects kvarn*)
- setup.sh: `WEIGHTS=qwopus-coder` fetch knob
- bump stale count asserts (drafters 9→10, registry 45→46, disk 46→47)
Validated 2026-06-12 (sm_86): embedded MTP loads, verify-full all-pass, NIAH
@72K = q5_0/q4_1 control, bench ~46/58 TPS (KVarN decode-neutral), 8-pack
104/103 ≈ q5_0/q4_1 102/107 (quality-neutral; disc #329). scripts/tests/*.sh
green in a clean tree (the lone test-compose-registry-disk red is pre-existing
untracked nex-n2-mini WIP, #473 — not in this commit). pending soak.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
* qwopus-coder: soak PASS + launcher-path validated — drop pending-soak note
Launcher path (switch.sh beellama/qwopus-coder --force → KVarN digest injected →
kvarn4 boots → serves :8067) + verify-full all-pass + soak-continuous PASS
(0 MiB growth, 0/25 silent-empty, 100% TPS retention). Stays 🧪 (pre-release
engine); full verify-stress NIAH ladder is the only gate left for ⚠️ promotion.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
---------
Co-authored-by: noonghunna <[email protected]>
Co-authored-by: Claude Opus 4.8 <[email protected]>
