WFGY/ProblemMap/GlobalFixMap/LocalDeploy_Inference/exllama.md

# ExLLaMA: Guardrails and Fix Patterns

<details>
  <summary><strong>🧭 Quick Return to Map</strong></summary>

<br>

  > You are in a sub-page of **LocalDeploy_Inference**.  
  > To reorient, go back here:  
  >
  > - [**LocalDeploy_Inference** — on-prem deployment and model inference](./README.md)  
  > - [**WFGY Global Fix Map** — main Emergency Room, 300+ structured fixes](../README.md)  
  > - [**WFGY Problem Map 1.0** — 16 reproducible failure modes](../../README.md)  
  >
  > 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.
</details>


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](https://github.com/onestardao/WFGY/blob/main/ProblemMap/rag-architecture-and-recovery.md)
* Retrieval and eval knobs: [Retrieval Playbook](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-playbook.md)
* Boot and ordering: [bootstrap-ordering.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/bootstrap-ordering.md), [deployment-deadlock.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/deployment-deadlock.md), [predeploy-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/predeploy-collapse.md)
* Snippet and trace schema: [retrieval-traceability.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/retrieval-traceability.md), [data-contracts.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/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](https://github.com/onestardao/WFGY/blob/main/ProblemMap/bootstrap-ordering.md)                                                                                                               |
| ΔS spikes when using quantized weights       | Tokenizer drift vs chunked embeddings          | [embedding-vs-semantic.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/embedding-vs-semantic.md), [chunking-checklist.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/chunking-checklist.md) |
| Memory corruption after long runs            | Fragmented KV cache, no eviction strategy      | [context-drift.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/context-drift.md), [entropy-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/entropy-collapse.md)                     |
| API or WebUI tool schema breaks              | JSON schema not enforced at inference layer    | [prompt-injection.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/prompt-injection.md), [logic-collapse.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/logic-collapse.md)                   |
| Multi-shard mismatch on large models         | Rank-order desync across GPUs                  | [deployment-deadlock.md](https://github.com/onestardao/WFGY/blob/main/ProblemMap/deployment-deadlock.md)                                                                                                             |

---

## Fix in 60 seconds

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

---

## Diagnostic prompt (copy-paste)

```txt
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](https://github.com/onestardao/WFGY/blob/main/I_am_not_lizardman/WFGY_All_Principles_Return_to_One_v1.0_PSBigBig_Public.pdf) | 1️⃣ Download · 2️⃣ Upload to your LLM · 3️⃣ Ask “Answer using WFGY + \<your question>”   |
| **TXT OS (plain-text OS)** | [TXTOS.txt](https://github.com/onestardao/WFGY/blob/main/OS/TXTOS.txt)                                                                     | 1️⃣ Download · 2️⃣ Paste into any LLM chat · 3️⃣ Type “hello world” — OS boots instantly |

---

<!-- WFGY_FOOTER_START -->

### Explore More

| Layer | Page | What it’s for |
| --- | --- | --- |
| ⭐ Proof | [WFGY Recognition Map](/recognition/README.md) | External citations, integrations, and ecosystem proof |
| ⚙️ Engine | [WFGY 1.0](/legacy/README.md) | Original PDF tension engine and early logic sketch (legacy reference) |
| ⚙️ Engine | [WFGY 2.0](/core/README.md) | Production tension kernel for RAG and agent systems |
| ⚙️ Engine | [WFGY 3.0](/TensionUniverse/EventHorizon/README.md) | TXT based Singularity tension engine (131 S class set) |
| 🗺️ Map | [Problem Map 1.0](/ProblemMap/README.md) | Flagship 16 problem RAG failure taxonomy and fix map |
| 🗺️ Map | [Problem Map 2.0](/ProblemMap/wfgy-rag-16-problem-map-global-debug-card.md) | Global Debug Card for RAG and agent pipeline diagnosis |
| 🗺️ Map | [Problem Map 3.0](/ProblemMap/wfgy-ai-problem-map-troubleshooting-atlas.md) | Global AI troubleshooting atlas and failure pattern map |
| 🧰 App | [TXT OS](/OS/README.md) | .txt semantic OS with fast bootstrap |
| 🧰 App | [Blah Blah Blah](/OS/BlahBlahBlah/README.md) | Abstract and paradox Q&A built on TXT OS |
| 🧰 App | [Blur Blur Blur](/OS/BlurBlurBlur/README.md) | Text to image generation with semantic control |
| 🏡 Onboarding | [Starter Village](/StarterVillage/README.md) | Guided entry point for new users |

If this repository helped, starring it improves discovery so more builders can find the docs and tools.  
[![GitHub Repo stars](https://img.shields.io/github/stars/onestardao/WFGY?style=social)](https://github.com/onestardao/WFGY)

<!-- WFGY_FOOTER_END -->