WFGY/ProblemMap/GlobalFixMap/LocalDeploy_Inference/vllm.md

vLLM: Guardrails and Fix Patterns

🧭 Quick Return to Map

You are in a sub-page of LocalDeploy_Inference.
To reorient, go back here:

• LocalDeploy_Inference — on-prem deployment and model inference
• WFGY Global Fix Map — main Emergency Room, 300+ structured fixes
• WFGY Problem Map 1.0 — 16 reproducible failure modes

Think of this page as a desk within a ward.
If you need the full triage and all prescriptions, return to the Emergency Room lobby.

Field guide for stabilizing vLLM-based local inference pipelines. Use these checks when models serve correctly on API providers but fail under high-throughput GPU serving with vLLM.

---

Open these first

• Visual recovery: RAG Architecture & Recovery
• Retrieval tuning knobs: Retrieval Playbook
• Snippet traceability: retrieval-traceability.md
• Ordering & race fixes: bootstrap-ordering.md, deployment-deadlock.md, predeploy-collapse.md
• Embedding issues: embedding-vs-semantic.md

---

Core acceptance

• ΔS(question, retrieved) ≤ 0.45
• Coverage ≥ 0.70 for target section
• λ remains convergent across 3 paraphrases and 2 seeds
• Throughput scaling does not shift retrieved citations

---

Typical vLLM breakpoints and fix

Symptom	Likely cause	Fix
Works at batch=1 but fails at scale	Context window fragmentation / GPU memory swap	context-drift.md, entropy-collapse.md
Citations disappear at high load	Async batch merge drops offsets	retrieval-traceability.md, data-contracts.md
Different answers run-to-run	λ flips with batch ordering	logic-collapse.md, rerankers.md
Index correct but retrieval unstable	Embedding vs metric mismatch in store	embedding-vs-semantic.md, vectorstore-fragmentation.md
GPU OOM / crash at warm-up	Preload sequence too large, missing fence	bootstrap-ordering.md

---

Fix in 60 seconds

1. Measure ΔS at batch=1 and batch=32. If ΔS rises >0.60 only at scale → async batching issue.
2. Probe λ across 3 paraphrases. If flips, apply BBAM.
3. Enforce contracts: citations must include snippet_id, offsets.
4. GPU warm-up: preload with a dummy batch before first live call.
5. Verify throughput stability with replay test (2 seeds, same dataset).

---

Copy-paste test prompt

I am running vLLM locally.  
Models served with async batching.  
Question: "{user_question}"  

Please return:
1. ΔS at batch=1 and batch=32  
2. λ across 3 paraphrases  
3. Whether citations preserved (snippet_id, offsets)  
4. Minimal structural fix if ΔS ≥ 0.60  

---

🔗 Quick-Start Downloads (60 sec)

Tool	Link	3-Step Setup
WFGY 1.0 PDF	Engine Paper	1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + <your question>”
TXT OS (plain-text OS)	TXTOS.txt	1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly

---

Explore More

Module	Description	Link
WFGY Core	Canonical framework entry point	View
Problem Map	Diagnostic map and navigation hub	View
Tension Universe Experiments	MVP experiment field	View
Recognition	Where WFGY is referenced or adopted	View
AI Guide	Anti-hallucination reading protocol for tools	View

If this repository helps, starring it improves discovery for other builders.