ExLLaMA: 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.

ExLLaMA (and its fork ExLLaMA2/ExLLaMA-HF) is a highly optimized CUDA inference backend used under TextGen WebUI and custom pipelines. It can run very large models (65B+) on limited VRAM, but often shows instability when sharded, quantized, or paired with retrieval layers. This guide stabilizes ExLLaMA with structural guardrails.

Open these first

Visual recovery map: RAG Architecture & Recovery
Retrieval and eval knobs: Retrieval Playbook
Boot and ordering: bootstrap-ordering.md, deployment-deadlock.md, predeploy-collapse.md
Snippet and trace schema: retrieval-traceability.md, data-contracts.md

Core acceptance

ΔS(question, retrieved) ≤ 0.45
Coverage ≥ 0.70 against anchor snippet
λ convergent across 3 paraphrases × 2 seeds
E_resonance flat across quantization modes (int4, int8)

Common ExLLaMA breakpoints

Symptom	Cause	Fix
First run slower or unstable than warm cache	Lazy CUDA graph compile, missing warm-up fence	bootstrap-ordering.md
ΔS spikes when using quantized weights	Tokenizer drift vs chunked embeddings	embedding-vs-semantic.md, chunking-checklist.md
Memory corruption after long runs	Fragmented KV cache, no eviction strategy	context-drift.md, entropy-collapse.md
API or WebUI tool schema breaks	JSON schema not enforced at inference layer	prompt-injection.md, logic-collapse.md
Multi-shard mismatch on large models	Rank-order desync across GPUs	deployment-deadlock.md

Fix in 60 seconds

Always warm-up: run a 10-token dummy batch before production queries.
Schema lock: enforce snippet_id, section_id, tokens in every trace.
λ probe: measure stability under 2 quant modes (int4 vs int8).
Cache rotation: reset KV cache every N tokens (e.g., 8192) to prevent drift.
Verify: coverage ≥ 0.70, ΔS ≤ 0.45 across three paraphrase probes.

Diagnostic prompt (copy-paste)

I am running ExLLaMA backend with quant={mode}, shards={n}, extensions={list}.  
Question: "{user_question}"  

Please output:
- ΔS vs retrieved snippet
- λ over 3 paraphrases × 2 seeds
- Quantization impact (int4 vs int8)
- Cache stability (tokens until drift)
- Minimal WFGY fix page 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.

7 KiB Raw Blame History Unescape Escape