---
license: apache-2.0
base_model: poolside/Laguna-XS.2
tags:
- gguf
- llama.cpp
- moe
- code
- quantized
- laguna
- pflash
- dflash
---

# Laguna-XS.2 GGUF (BF16 + Q4_K_M)

GGUF conversions of [poolside/Laguna-XS.2](https://huggingface.co/poolside/Laguna-XS.2), a 33B-A3B (3B active) MoE coding model from Poolside under Apache 2.0. Built for use with [lucebox-hub](https://github.com/Luce-Org/lucebox-hub) (dflash + PFlash) on consumer GPUs.

## Files

| File | Quant | Size | BPW | Notes |
|------|-------|------|-----|-------|
| `laguna-xs2-bf16.gguf` | BF16 | 66.9 GB | 16.01 | reference, identical math to HF transformers fp/bf16 |
| `laguna-xs2-Q4_K_M.gguf` | Q4_K_M | 20.3 GB | 4.85 | imatrix-calibrated, fits a single 24 GB GPU |
| `laguna-xs2.imatrix` | imatrix | 188 MB | — | Bartowski calibration_datav3 (134 chunks, 68608 tokens) |

## Architecture

- 40 layers, n_embd 2048, n_head_kv 8, head_dim 128
- Per-layer head count [48, 64, 64, 64] × 10 (4-layer SWA pattern: full, sw, sw, sw)
- 256 experts, top-8 routing, 1 always-on shared expert
- Sigmoid router, expert weights scale 2.5
- Sliding window 512, partial RoPE with YaRN (orig ctx 4096, factor 32)
- Vocab 100,352, BOS=2, EOS=2, PAD=9

## Quality

| Metric | BF16 | Q4_K_M | Δ |
|--------|------|--------|---|
| Perplexity (Bartowski v3, 20×512) | 10.7594 ± 0.522 | 11.2854 ± 0.553 | +4.9% |

Imatrix calibration uses Bartowski `calibration_datav3.txt` (multilingual + code mix), the same corpus Unsloth-distributed quants use.

Verified vs the official Poolside HF reference (BF16, eager attention, greedy decoding): logits match exactly for the first 30+ tokens on a B-tree explanation prompt; subsequent divergence is fp precision drift, not a graph bug.

## Performance (RTX 3090 24 GB, Q4_K_M)

Measured with `bench_laguna_generate` from lucebox-hub (dflash autoregressive forward, no spec-decode draft yet):

| Workload | Throughput | Notes |
|----------|-----------|-------|
| Decode @ ctx=128 (greedy) | **113 tok/s** | n_gen=128 |
| Decode @ ctx=1K | 104 tok/s | |
| Decode @ ctx=4K | 65 tok/s | |
| 128K TTFT via dflash + PFlash | **15.91 s** | 5.4× faster than llama.cpp pp131072 (86.60 s) |
| Loader VRAM | 18.77 GiB | + 110 MiB tok_embd kept on CPU |

## Usage

### lucebox-hub (dflash + PFlash, recommended for 128K)

```bash
# clone
git clone https://github.com/Luce-Org/lucebox-hub
cd lucebox-hub/dflash

# build with sm_86 (3090 / A6000)
cmake -B build -DCMAKE_CUDA_ARCHITECTURES=86
cmake --build build -j

# fetch the Q4_K_M GGUF + Poolside tokenizer
hf download Lucebox/Laguna-XS.2-GGUF laguna-xs2-Q4_K_M.gguf --local-dir models/
hf download poolside/Laguna-XS.2 chat_template.jinja tokenizer.json tokenizer_config.json \
   special_tokens_map.json config.json --local-dir models/Laguna-XS-2

# run the OpenAI server (same server.py as qwen35, arch auto-detected from GGUF).
# -ctk/-ctv q4_0 keeps the 131K KV cache under ~6 GB so weights + KV fit on 24 GB.
python3 scripts/server.py \
  --target models/laguna-xs2-Q4_K_M.gguf \
  --tokenizer models/Laguna-XS-2 \
  --port 8000 --max-ctx 131072 \
  -ctk q4_0 -ctv q4_0

# chat
curl http://localhost:8000/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model":"luce-dflash","messages":[{"role":"user","content":"hello"}],"stream":true}'
```

## License

Apache 2.0, inherited from upstream `poolside/Laguna-XS.2`.

## See also

- [PFlash 128K speedup blog post](https://lucebox.com/blog/pflash)
- [DFlash on ggml blog post](https://lucebox.com/blog/dflash27b)
- [lucebox-hub on GitHub](https://github.com/Luce-Org/lucebox-hub)