Missing Carnice 2x3090 YAML

[mtdphn]

The following link is missing : https://github.com/noonghunna/club-3090/blob/master/models/qwen3.6-27b/vllm/compose/dual/carnice-bf16mtp/bf16-mtp.yml

[noonghunna]

Thanks for reporting this @mtdphn . we don't formally support Carnice just yet. I previously added it but couldn't finish the work.
I do have some spare time now, i'll try to revive it tonight.

[noonghunna]

Carnice is back ✅ — revived, validated end-to-end, and merged in #406 (now on master).

One heads-up on the link: it pointed at the old, archived dual-vLLM AutoRound attempt, which is why it 404s. The revived path is single-card beellama — beellama/carnice-v2-single-q5km-mtp. For the new Carnice-V2 the Q5_K_M GGUF (with an embedded MTP head) fits comfortably on one 24 GB card at 160K ctx (176K ceiling) + KVarN-4 KV, so single-card is its natural home — no second GPU needed.

Quick start:

WEIGHTS=carnice-v2 bash scripts/setup.sh qwen3.6-27b
bash scripts/switch.sh beellama/carnice-v2-single-q5km-mtp --force   # :8068

Full results + the n=1-vs-n=2 finding are in the announcement → #407 (it edges the Qwopus-Coder sibling on the agentic packs, 110 vs 103 /150).

If you specifically want a 2×3090 Carnice for a bigger window or a higher quant (e.g. Q8), say the word and I'll look at adding a dual config. Thanks for the nudge, @mtdphn! 🙏

[mtdphn]

Thank you! I was looking at 2x3090, but a single is a very good option as well.

[steamEngineer]

I would also be interested in a 2x 3090 Carnice config for testing! Ideally using vllm but whichever works. :)

[noonghunna]

Thanks for the flag, @mtdphn — that link pointed at the old vLLM BF16-MTP path that got restructured. Rather than just repair the link, we built the dual Carnice config folks have been asking for, and went up a quant tier while at it.

It's now shipped as beellama/carnice-v2-dual-q8-mtp — Carnice-V2-27B Q8_0 + embedded MTP head on beellama v0.3.2, dual layer-split, q8_0 KV over the full 262K (PR #416). On the 8-pack it's within sampling noise of base Qwen3.6-27B at the max-fidelity dual tier (103/105 vs 110/150) — no real capability cost for the agentic/Hermes tuning.

Full Results Card (serving + quality + how to run) is in the announcement: #417 → #417

bash scripts/switch.sh --owui beellama/carnice-v2-dual-q8-mtp   # :8070 + Open WebUI

[mtdphn]

Thank you so much!