Was '**v0.8.2 surfaces (current):**' — stale at v0.8.7. The four pull-gate bullets are still the current surface, so reword to '**Pull-gate surfaces (current):**' (no pinned version → can't go stale on the next release). Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
10 KiB
Architecture — how this stack thinks about LLM serving on 24 GB
A short orientation to the design choices in this repo. Not a deep technical doc — just the mental model.
What the stack assumes
You have 1 or 2 RTX 3090s (or compatible 24 GB Ampere-class cards). You want to serve a modern LLM locally for chat / coding / agents / RAG. You're OK with a bit of setup but you don't want to fork engines or write CUDA kernels.
Two ways in: curated catalog + universal pull
The stack has two entry paths, and the mental model is "the curated catalog is the measured backbone; pull is the general front door."
- Curated catalog (
models/<model-name>/) — models we have measured: real benchmark numbers, validated composes, KV-math calibration anchors, per-model gotchas. This is the high-confidence path and the source of truth the math is calibrated against. - Universal
pull(scripts/pull.sh <hf-repo> --profile-like <catalog-key>) — point it at any safetensors HF repo. It derives a spec from the repo's own files, runs it through the gate, and only proceeds for a vLLM-loadable, supported model whose fit math passes (or where you accept an explicit override). It is honest about confidence and never silently gate-passes — every non-pass outcome hard-stops with a precise reason. Headline: "evaluate any safetensors HF repo; pull only vLLM-loadable supported ones, and only when the gates pass (or an explicit override is accepted)."
The pipeline behind pull:
pull → derive (read repo config/safetensors)
→ gate ([C0] engine-support · [C2a] disk · [B] kv-calc fit verdict
· [C1] confidence × verdict → terminal action)
→ emit (generate a minimal derived compose for an in-scope model)
→ boot (gated download → derived-compose boot → capture artifacts)
→ loop (classify failures · trust pipeline · dedup · feed the
calibration backbone)
Pull-gate surfaces (current):
scripts/pull.sh … --recommend— appends an honest one-line fit verdict (FITS / FITS-but-not-yet-accepted / DOES-NOT-FIT) over the same gate result; presentation only, never changes the verdict, carries the boot-fit≠runtime caveat.- Failure on-ramp — every hard-block leaves a redacted
.pull-captures/<slug>/<ts>/bundle;scripts/pull.sh --submit-last(or--submit <dir>) is a separate, consented, user-invoked step that submits it (works with or withoutgh; no telemetry, no auto-send). - Broadened arch registry — materially more safetensors architectures pass
[C0]without--experimental-arch; native built-ins reach a clean serve verdict, per-repo remote-code stays fail-closed (zero false-pass). - Optional non-NVIDIA
hwdetect— bounded subprocess augmenting hardware detection wherenvidia-smidoesn't apply (AMD/Apple); absent → graceful degrade, never feeds kv-calc.
You don't need to know any of these stage names to use it — you run one command (scripts/pull.sh); the [C0]/[B]/[C1]/… taxonomy above is internal flow for contributors. Start at the user guide: docs/PULL.md (Quickstart at the top). Contributor depth, in pipeline order: COMPOSE_GENERATOR.md (emit) → PULL_GATE.md (gate) → PULL_EMIT_DERIVED.md (boot+capture) → LOOP.md (classify+trust+dedup). The curated path still works exactly as before — pull is additive, the front door for anything not in the catalog.
A boot-fit pass is a static check. It is necessary-not-sufficient: a fits-clean config can still degrade under accumulated-context agent workloads (the Cliff 2 / prefill-cliff failure modes — see CLIFFS.md). The gate verdict says so explicitly and points at soak-continuous validation; trust the caveat, not just the green.
How the repo is organized
The mental model: engines are general; models are specific; hardware is fixed.
docs/ engine + hardware docs (general, model-agnostic)
engines/ vLLM / llama.cpp / SGLang — comparison + deep dives
HARDWARE.md Ampere SM 8.6+, 24 GB, no NVLink
GLOSSARY.md plain-language definitions
img/ chart sources + PNG exports (performance, vram-budget)
models/<model-name>/ everything specific to a model
README.md model overview + quants + Genesis surface
INTERNALS.md this model's quirks (architecture, bugs, fixes)
CHANGELOG.md this model's dated history
vllm/ vLLM-specific configs for this model
compose/ docker-compose files
patches/ model+engine patches
README.md "vLLM recipes for this model"
llama-cpp/ llama.cpp recipes for this model
recipes/ shell scripts
README.md "llama.cpp recipes for this model"
sglang/ SGLang status / TBD recipes
scripts/ shared, model-aware
setup.sh <model> downloads + verifies model + clones patches
verify.sh quick smoke (~10 sec)
verify-full.sh fast functional test, 8 checks (~1-2 min)
verify-stress.sh boundary-case stress test, 2 checks (~5-10 min)
bench.sh canonical TPS bench
pull.sh <hf-repo> universal evaluate→gate→emit→boot front door
(+ --recommend verdict; --submit-last/--submit on-ramp)
generate-compose.sh emit a minimal compose for an in-scope profile
lib/profiles/ the gate/derive/classify/trust pipeline (engine)
tests/ executable specs (test-pull, test-classifier, …)
.pull-captures/<slug>/<ts>/ runtime: per-pull capture bundle the loop consumes
(pt1–5 + manifest; redacted; created by pull)
Why this layout
Why "models/" isn't at the top
If "qwen3.6-27b" were the top-level partition, every cross-model concept (engines, hardware, scripts) would either be duplicated or live awkwardly in some shared subdir. By putting models inside models/, the top-level becomes infrastructure (engines, hardware, glossary, scripts) and models/<m>/ becomes content. This scales: when we add Qwen3.5-27B / GLM-4.6 / Llama-3.x, they get a new subdir under models/ with the same internal pattern, and the top-level docs stay relevant.
Why engines are general docs
vLLM behaves the same way regardless of whether you're serving Qwen, GLM, or Llama. The tuning levers (mem-util, KV type, spec-decode config, power cap) are model-agnostic. So docs/engines/VLLM.md covers vLLM-the-engine, not vLLM-on-Qwen. Per-model engine recipes live under models/<m>/<engine>/.
Why patches are per-model-per-engine
A patch like patch_tolist_cudagraph.py fixes a bug that hits Qwen3-Next + vLLM + TurboQuant + spec-decode together. It wouldn't apply to a different model with different attention layout. So patches live at the most specific level: models/<m>/<engine>/patches/.
If a patch is general (across engines or models), it bubbles up to docs/engines/<engine>.md notes or — rarely — into a top-level patches/ (none today).
Why scripts are top-level but model-aware
bash scripts/setup.sh qwen3.6-27b is the model-aware form. The script reads the model name and does the right downloads / SHA verification / patch fetching. We keep the script set in one place because the operation (download, verify, boot, test, bench) is the same shape across models.
How information flows
A user comes in cold:
- Lands on top-level README → understands what the stack is, picks their model.
- Goes to
models/<m>/README.md→ sees recommended config + quick start for their card count. - Boots; tests with
verify-full.sh(fast, 8 checks); for boundary cases (KV-cache pressure, prefill OOM) runsverify-stress.sh; benches withbench.sh.
A user hits a problem:
- Checks
docs/SINGLE_CARD.mdordocs/DUAL_CARD.mdfor workload-specific gotchas matching their hardware. - Checks
docs/FAQ.md"Troubleshooting" section for the specific failure mode. - If still stuck:
models/<m>/INTERNALS.mdfor engineering depth. - If engine-related:
docs/engines/<engine>.mdfor general engine tuning. - Files an issue with logs.
A power user wants to push limits:
docs/engines/<engine>.md— engine tuning levers.models/<m>/INTERNALS.md— model-specific knobs.models/<m>/<engine>/README.md— recipe-specific tips.
Design rules
A few principles the repo follows:
- No tutorials disguised as configs. Composes are working configs, not pedagogy. Configs reference docs for the "why."
- Honest framing always. If a config has a known cliff, the cliff is documented at the top of the relevant doc, not buried in a footnote. Users discovering issues at boot should already have read the warning.
- Cross-rig data welcome. TPS numbers are run-to-run variable; we publish ours and welcome PRs adding "your rig" rows.
- Patches stay surgical. We don't fork engines. Disk-edits at boot, runtime monkey-patches, or volume-mounts of patched source. When upstream lands a fix, the patch becomes a no-op (anchor doesn't match) and we drop it cleanly.
- Verification gates production.
verify-full.shruns 8 fast functional checks;verify-stress.shruns the heavy boundary-case tests (long-context needle ladder, ~25K-token tool-response prefill OOM detection). We don't claim a config works until both are green. - Document the negative results too. Probes that didn't pan out (PR #40798 backport,
--enforce-eagermode) are documented so future-us doesn't redo the experiments.
Things this stack is NOT
- A vLLM fork. All vLLM patches are mounted at boot, not forked into a custom build.
- A model card / training recipe. We use pre-quantized weights as-is. For training/quantization details, see the model authors' (Lorbus, Qwen) repos.
- A general benchmarking suite.
bench.shis the minimum needed to verify your setup matches ours. For rigorous A/B comparisons use vllm-project/bench or similar. - A cloud-replacement service. It's a recipe for running locally. Wrap it in your own auth/queueing/quota/etc. for production.