noonghunnaandClaude Fable 5
6f674fa2ce
Two 5090 owners hit `--kv-cache-dtype nvfp4 requires sm100f` crashing mid-boot on the #246 A/B (disc #571). Root-caused from vLLM #43562 / TRT-LLM #10241: nvfp4 KV forces the trtllm-gen FP4 FMHA, built ONLY for datacenter Blackwell sm_100/sm_103. Consumer Blackwell (sm_120/121 — RTX 5090 / PRO 6000 Blackwell) is a HIGHER cc number but a different family with no FMHA build. NVFP4 *weights* work there; only the KV path doesn't. Our #246 gate had used a numeric ">=10.0" floor that wrongly passed sm_120 — a floor can't express "sm_100/103 but not the numerically-higher sm_120". - gates.py: new `_ARCH_KERNEL_SM_FAMILY` allowlist ({nvfp4: sm_100/103}); dropped nvfp4 from the numeric `_ARCH_KERNEL_SM` floor; family-membership reject with the FMHA reason + fp8_e4m3 fallback. - arch-ab.sh: nvfp4 arm now refuses on consumer Blackwell (not just <sm_10), naming the FMHA gap + the fp8_e4m3 path; dropped nvfp4 from the recommended arms in help. - hardware profiles: removed nvfp4 from rtx-5090 / rtx-6000-pro-blackwell KV lists (both sm_120) + added a why-not note. - docs (DTYPE_MATRIX / HARDWARE / KV_MATH / QUANTIZATION): corrected the "Blackwell sm >= 10.0" framing to "datacenter sm_100/103 only". - UPSTREAM.md: #43562 / TRT-LLM #10241 row + re-test trigger. - test-arch-ab: nvfp4 refuses on sm_86 AND sm_120, allowed on sm_100; the dual-5090 all-arms test drops nvfp4. Full scripts gate 66/66. Co-Authored-By: Claude Fable 5 <[email protected]> Claude-Session: https://claude.ai/code/session_01EfF565T9eSLaqGzidyJ1Pm
club-3090 docs index
Two tracks. Pick the one that matches what you're doing.
- User track — "I have GPUs and a model; how do I serve it?"
- Contributor / maintainer track — "I'm working on the v0.8.0 pull pipeline, patches, or the calibration loop."
Every link below resolves to a file in this repo.
User track
Start here if you want to run a model.
| Doc | What it is |
|---|---|
GETTING_STARTED.md |
Start here — 5-minute clone-to-curl path. No decisions, no menus. |
SINGLE_CARD.md |
1× RTX 3090 — workload → curated config → quick start. |
DUAL_CARD.md |
2× RTX 3090 (PCIe / NVLink auto-detected) — workload → config → quick start. |
MULTI_CARD.md |
3+ GPUs — TP scaling math, derivation from dual.yml, valid TP values. |
PULL.md |
Any HF safetensors repo — evaluate against the KV math, honest about confidence. |
BRING_YOUR_OWN.md |
Serve + tune + validate your own model/compose (any engine, single or dual) without touching the catalog. |
HARDWARE.md |
Card-class questions — 4090/5090, power caps, NVLink, laptop EC power. |
PCIE_P2P.md |
PCIe-only multi-GPU — read your topo -m, and optionally enable P2P over PCIe without NVLink. |
GLOSSARY.md |
TPS / KV / MTP / TP and the rest of the vocabulary. |
FAQ.md |
Common setup and operational questions. |
COMPARISONS.md |
Self-host vs cloud APIs — cost crossover and when each wins. |
EXAMPLES.md |
Worked end-to-end usage examples. |
ai-studio/ |
Club 3090 AI Studio — chat-driven, all-modality creative studio (text · image · video · audio) on 2× 3090. Start here for the architecture + the full 8-lane matrix. |
ai-studio/image.md |
Image lanes — HiDream-O1 (top-quality/photoreal) · Ideogram-4 (design/logo/text) · Chroma (uncensored). |
ai-studio/video.md |
Video — LTX-2.3 (video+audio) + Sulphur (uncensored), text/image→video, 60 s+ chaining. |
ai-studio/audio.md |
Audio — Step-Audio-EditX (voice clone+edit) · Kokoro (narration) · ACE-Step (music) · Stable Audio (SFX). |
Contributor / maintainer track
The v0.8.0 pull pipeline (in pipeline order)
A model slug flows through these stages. Read them in order to understand the whole.
| Stage | Doc | What it owns |
|---|---|---|
[D] |
COMPOSE_GENERATOR.md |
The #141 compose generator — the substrate that owns the arch→patches matrix. |
| Gate | PULL_GATE.md |
scripts/pull.sh — the locked 6-stratum abort taxonomy, [C0]/[C2a]/[B]/[C1] gates, §4.1 confidence×verdict table. |
[E] |
PULL_EMIT_DERIVED.md |
Download → boot → smoke for a download-eligible derived model; writes the §6 capture artifacts. |
[F] |
LOOP.md |
The calibration loop — reads the capture bundle, classifies, runs the inbound-trust pipeline, dedups failures into the tracker. |
Patch & model contribution
| Doc | What it is |
|---|---|
PATCH_POLICY.md |
When/how a patch ships, the local-overlay vs upstream rules. |
PATCH_ATTRIBUTION.md |
The Phase-A patch-attribution matrix — arch → engine-pin → required patches. |
ADDING_MODELS.md |
How a new model gets added to the curated catalog. |
KV_MATH.md |
The KV-cache math the [B] fit verdict is computed from. |
Stack reference & ops
| Doc | What it is |
|---|---|
ARCHITECTURE.md |
Repo/stack architecture overview. |
UPSTREAM.md |
Upstream PR / issue tracker for this stack. |
NIGHTLY_BUMP_RUNBOOK.md |
Procedure for bumping the vLLM nightly pin. |
CONTAINER_RUNTIMES.md |
Docker / container runtime notes. |
Reference matrices & deep dives
These are cross-cutting references both tracks reach for.
| Doc | What it is |
|---|---|
scripts/switch.sh --list (runtime command, not a doc) |
The authoritative compose × slug matrix. Registry-derived from scripts/lib/profiles/compose_registry.py, so it's always current — every launchable slug with its topology, model, engine, KV format, and max ctx. Run this rather than trusting any hand-maintained table; the static lists in the per-topology docs are illustrative, this is the source of truth. |
engines/ |
Per-engine deep dives — vLLM, llama.cpp, SGLang. |
INFERENCE_ENGINES.md |
Engine picker — which engine for which workload, and structural gaps. |
CLIFFS.md |
The accumulated-context / prefill failure modes (Cliff 2, Cliff 2b) and how to detect them. |
QUANTIZATION.md |
Quant field guide — GGUF ladder, the vLLM/safetensors schemes, KLD-by-fidelity (§4a), and the fast/balanced/max tier trade-space (§4b). The conceptual home the dtype/kernel matrices support. |
DTYPE_MATRIX.md |
Supported dtype × model × engine matrix. |
KERNEL_MATRIX.md |
Quant-kernel availability and alignment constraints. |
QUALITY_TEST.md |
The quality-test harness and what it measures. |
RESULTS_CARD.md |
The standard 3-panel format (Serving · Quality · Takeaways) for sharing a config's measured results. |
ANNOUNCEMENT_TEMPLATE.md |
The "we shipped X" Announcements-post skeleton that wraps a Results Card (intro+credits · Results Card · getting it · run it · credits). |
STRUCTURED_COT.md |
The bounded-thinking / structured-CoT compose path. |
TQ3_MTP_GENESIS.md |
TQ3 KV × MTP × Genesis-patch results and config. |