May I ask for a club-5090？

[li-yifei]

Sry to bother

[noonghunna]

@li-yifei — never a bother, you are very welcome here. 5090 owners absolutely belong in this club.

I want to address the name directly, because you're not the first to feel this way (see #27 and others in the last 24 hours):

"club-3090" is a tribute, not a hardware filter

The 3090 holds a unique place in 2026: it's the last GPU class that's actually still affordable for someone getting started with AI (~$700-900 used in most markets vs $2-3K for a 4090, $3-4K for a 5090, $30K+ for an H100). The name honors that. It's the card you can buy to learn this stack on without taking a loan.

But this repo is about squeezing the absolute maximum out of whatever GPU you happen to have. The methodology — TQ3 KV / MTP K=3 / Genesis patches / Cliff 1+2 closures / verify-stress as ground truth — applies on every CUDA card from SM 8.0 (A100) through SM 12.0 (5090) and Blackwell datacenter. We just ship default configs tuned for what most people own.

What's interesting about 5090 specifically

5090 is Blackwell consumer SM 12.0 — a regime where most of our shipped patches will need re-tuning, and several blockers we work around on Ampere simply disappear:

• 32 GB VRAM vs our 24 GB target — Cliff 1 (TQ tool-prefill at 50K) and Cliff 2 (DeltaNet GDN forward at 50-60K single-prompt) both probably close on a single card without TP=2 or our GDN chunked-prefill workarounds. Worth verifying.
• Native FP8 compute — entire FP8 KV path can use real FP8 GEMM instead of decompressed FP16. Should change the FP8-vs-TQ3 trade for tools-text.yml significantly.
• FA3 + FA4 — vLLM auto-enables on SM ≥ 9.0. Sparse-V kernels (PN26b) become irrelevant; FA3's structured sparsity does the same job with less code.
• Genesis patches need re-vetting — we don't know which Genesis env vars are no-ops vs counter-productive on Blackwell. Sander has zero Blackwell hardware (per his comment in #27), so the only path to characterize this is from owners like you.

What we'd love from any 5090 owner

The fastest path to a useful data point is ~15 min:

git clone https://github.com/noonghunna/club-3090
cd club-3090
bash scripts/setup.sh qwen3.6-27b
# Try the dual-turbo or default compose:
docker compose -f models/qwen3.6-27b/vllm/compose/docker-compose.yml up -d
# Then:
bash scripts/verify-stress.sh

Post the boot log + verify-stress output as a discussion. We'll turn it into:

• The first SM 12.0 row in BENCHMARKS.md (with credit)
• HARDWARE.md guidance specific to 5090
• A 5090.yml compose variant if your config diverges from defaults
• Genesis bug reports back to Sander when something doesn't transfer

If a 5090.yml (or a sister club-blackwell umbrella) ends up making sense to ship, we'll absolutely do it. The architectural break is real — SM 12.0 has its own optimal config that SM 8.6 can't hit. But that doesn't mean we split repos preemptively; it means we add a tested compose for your hardware and credit the contributor who validated it.

The diversity push is in #27 if you want context — same applies here. Welcome aboard.