Add Carnice-V2-27B beellama single-card compose (beellama/carnice-v2-single-q5km-mtp) (#406)
* Add Carnice-V2-27B beellama single-card compose (beellama/carnice-v2-single-q5km-mtp)
stuchapin Carnice-V2-27B Q5_K_M GGUF (kai-os/Carnice-V2-27b, a Hermes-style agentic
SFT of Qwen3.6-27B) with embedded MTP head, on beellama v0.3.2-preview KVarN — single
3090, kvarn4/kvarn4 KV, MTP n=1, reasoning-on default.
Validated 2026-06-14 (rebench-full, reasoning-on): engine-compat PASS (beellama loads
the PR#22673-fused GGUF — the card's "mainline fails to load" does not apply),
verify-stress 8/8 (NIAH clean to 150K), soak PASS (0-growth, 0/100 silent-empty,
100.5% retention), bench 46.8/50.5 TPS narr/code, MTP accept ~94%. 8-pack reasoning-on
110/150 — beats sibling beellama/qwopus-coder 103/150 (edge is agentic/instruct).
- compose: models/qwen3.6-27b/beellama/compose/single/carnice-v2-q5km/mtp-kvarn4.yml
- drafter profile: scripts/lib/profiles/drafters/carnice-mtp-gguf.yml (n_default/n_max=1
— mtp_num_hidden_layers=1; author warns n=3 is wrong)
- weights_variant carnice-v2-q5km + compose_registry entry (port 8068, kvcalc SKIP)
- bump test count assertions (registry 47, disk 48, drafters 11)
Stays experimental (beellama v0.3.2 is a rolling pre-release, the #455 engine gate),
not for any quality reason. n=1 is the card-faithful default; n=2 is a documented
+12%-TPS opt-in (DRAFT_N_MAX=2, does not crash on our single-card kvarn4) pending a
dedicated soak.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
* setup.sh: add WEIGHTS=carnice-v2 fetch knob for beellama/carnice-v2-single-q5km-mtp
UX parity with WEIGHTS=qwopus-coder — maps to qwen3.6-27b:carnice-v2-q5km so
`WEIGHTS=carnice-v2 bash scripts/setup.sh qwen3.6-27b` fetches the GGUF.
Co-Authored-By: Claude Opus 4.8 <[email protected]>
---------
Co-authored-by: noonghunna <[email protected]>
Co-authored-by: Claude Opus 4.8 <[email protected]>
