---

license: other

license_name: openmdw-1.1

license_link: LICENSE.md

base_model:

- poolside/Laguna-S-2.1

library_name: llama.cpp

tags:

- gguf

- layer-library

- ik_llama.cpp

- dreamcatcher

- xautonomics

---

Laguna-S-2.1-UD-IQ4_NL.LAYR.GGUF

A per-layer GGUF library of Unsloth's unsloth/Laguna-S-2.1-GGUF (UD-IQ4_NL). These are the same quantized weights, sliced once into one file per transformer block plus the shared parts. It's built for dreamcatcher, xAutonomics' fork of ik_llama.cpp.

| | |

|---|---|

| base model | poolside/Laguna-S-2.1 |

| source quant | unsloth/Laguna-S-2.1-GGUF · UD-IQ4_NL · pinned revision 750f92f90c · 3 shards |

| architecture | laguna · 48 blocks |

| files | 50 (embd, layer, output) · 54.9 GiB |

| integrity | manifest.json: blake2b-128 per whole file and per tensor |

| sliced with | layer_distribution.slice from tools/layer-distribution (commit in SLICER_COMMIT); nothing was re-quantized |

Support status

laguna — this library has not been run on dreamcatcher yet. The fork includes the laguna architecture, but it has no stage-ring hooks for it, so this library is for single-process serving only (llama-server --model-dir); multi-host stage rings are not supported for this architecture. This card will be updated when it is measured.

UD-IQ4_NL was chosen over UD-IQ4_XS because the fork's CPU kernels for iq4_xs (and iq4_kss/iq5_ks) currently lose precision at small batch, which matters when experts run on the CPU at decode (see the fork's known issues). This quant contains none of those types: its tensors are Q8_0 (385), F32 (287), IQ3_S (92), IQ4_NL (47) and Q6_K (3).

Only model files from huggingface.co/xautonomics are supported by dreamcatcher. Do not rename the files: the loader expects blk-NNNNN.gguf / parts-*.gguf / manifest.json exactly.

Layout

blk-00000.gguf … blk-00047.gguf   one transformer block each
parts-embd.gguf                          token embedding
parts-output.gguf                        output norm + head
manifest.json                            file list, sizes, per-file + per-tensor hashes, source provenance

Use

llama-server --model-dir /models/Laguna-S-2.1-UD-IQ4_NL.LAYR.GGUF

Verify the download before launching:

python3 -c "from layer_distribution import verify; r = verify('/models/Laguna-S-2.1-UD-IQ4_NL.LAYR.GGUF', '/models/Laguna-S-2.1-UD-IQ4_NL.LAYR.GGUF/manifest.json'); print(r.passed, r.hash_verified)"

Provenance, license and notices

Weights are byte-for-byte the tensors of unsloth/Laguna-S-2.1-GGUF at revision 750f92f90cf54159c4d7a610cb7b3e74498e75c6 (Unsloth's dynamic quantization of poolside/Laguna-S-2.1). Source shard names and sizes are recorded in manifest.json (source.shards[], source.content_hash).

Laguna S 2.1 is developed by Poolside (poolside/Laguna-S-2.1) and released under the OpenMDW License Agreement, version 1.1. A copy of the agreement is included in this repository as LICENSE.md, as the license requires for redistribution. Quantization by Unsloth. We made no changes to the weights; this repository only splits them into per-layer files.

About

Published by xAutonomics for dreamcatcher.