noonghunna/club-3090

⚠️ Heads-up — defaults & pins changed.

• Single-card defaults are now beellama.cpp (DFlash): Qwen3.6-27B → beellama/dflash, Gemma-4-31B → beellama/gemma-dflash (previously vLLM/Genesis paths).
• vLLM composes now pin stable v0.22.0 instead of nightlies (which Docker Hub purged, causing "image not found" boot failures). The custom vllm-club3090 image is retired — we run stock vLLM with patches mounted at boot.
• Genesis vLLM composes deprecated; vLLM single-card default → vllm/minimal. Gemma-4 dual slugs renamed/pruned (9 → 3; default vllm/gemma-int8-mtp). vLLM dual-DFlash composes deprecated → use beellama.
  Use the registry --variant keys (bash scripts/launch.sh vllm/dual, beellama/dflash, ik-llama/byteshape-iq4xs-mtp, …); switch.sh --list shows the current set.

v0.8.7 — 2026-06-03

Highlights

• New serving engine — beellama.cpp (DFlash). A llama.cpp fork with external-drafter speculative decoding, now a first-class engine and the single-card default for both Qwen3.6-27B and Gemma-4-31B. Experimental Q8_K_XL dual composes included.
• Qwen3.6-35B-A3B is now Production on dual 3090 — 262K context + vision (vllm/qwen-35b-a3b-dual), plus new single-card ik-llama presets (apex-fit-q8q5, community byteshape-iq4xs-mtp).
• Gemma-4 moved to stable vLLM v0.22.0 — off the purged nightlies, reasoning parser wired, sampling aligned to the model card.
• switch.sh overhaul — --list groups by model · topology, filters to your GPU count, and shows max context + health per config; you can now pin your own default.

✨ Features

• beellama.cpp DFlash as a first-class engine + promoted to single-card default for Qwen3.6-27B and Gemma-4-31B (#268, #271, #272); experimental v0.3.0 Q8_K_XL dual composes (#296).
• Qwen3.6-35B-A3B → Production: dual 262K + vision (#259); single-card ik-llama apex-fit-q8q5 (#242) and community byteshape-iq4xs-mtp (#299).
• Gemma-4 duals → vLLM v0.22.0 with the gemma4 reasoning parser (#287, #289).
• switch.sh: group --list by model · topology with per-slug max-context + health, GPU-count filtering (--all), and user-pinnable defaults (#264, #265, #266, #267, #283).
• NVLINK_MODE=pcie_p2p for boards with PCIe peer-to-peer but no NVLink (#290/#291).
• Operational robustness: orphan-safe switch.sh, reboot-surviving vLLM containers, multi-GPU power sweep (#285).
• Quality/bench tooling: rebench-full reworked (think-OFF + think-ON 8-packs, fail-fast verify-full preflight, live progress — #303, #306); quality-test live progress on by default + smarter timeouts (#248, #245); bench measurement-record producer (#249).

🐛 Bug fixes

• kv-calc: shard Qwen3-Next-MoE weights by TP — fixes a false long-context "won't fit" verdict (#261).
• NVLink config clobber on PCIe-P2P boards (#290).
• preflight / report.sh now detect beellama containers + check the spec-draft GGUF.
• setup.sh rejects unknown positional args instead of silently ignoring them (#273).
• soak-test auto-detects the 35B-A3B container (#244).

📝 Documentation

• New guide: "Bring your own model or compose" — serve, tune, and validate any model on your rig without the curated catalog.
• Contribution rules: one model — or one feature/concern — per PR; mandatory compose Profile header.
• FAQ: agent stops mid-task → set client temperature 0.6 (#232); which KV-cache quant to pick; how to switch models; 4090/5090 cross-rig notes.
• Results Card — a standard format for sharing a config's measured results.
• Refreshed the add-a-model workflow; AGENTS.md is now one source with CLAUDE.md.

🧹 Maintenance

• vLLM composes pinned to stable v0.22.0 (off nightlies); custom vllm-club3090 image retired — stock vLLM + mounted patches (#269, #277).
• Genesis vLLM composes deprecated; vLLM single default → vllm/minimal (#276).
• Pruned dual vLLM composes; renamed/laddered Gemma-4 slugs (#278, #279, #286); litellm route cleanup (#263).

[Pin: git checkout v0.8.7] · Full diff