Laguna-S-2.1-DFlash-GGUF

Laguna S 2.1 DFlash GGUF

This repository contains a Q8_0 GGUF conversion of Poolside's DFlash drafter for Laguna S 2.1. It is an auxiliary draft model. It is not a standalone language model and does not include the Laguna S 2.1 target weights.

Artifact

The GGUF has six DFlash layers and target taps [2, 11, 20, 30, 39, 48]. Its 49 matrix tensors use Q8_0. Its 27 normalization and bias tensors remain F32.

Download

hf download dev7a/Laguna-S-2.1-DFlash-GGUF Laguna-S-2.1-DFlash-Q8_0.gguf

Use this drafter only with a compatible Laguna S 2.1 target and a runtime that supports the standardized llama.cpp dflash GGUF schema. In NS4, use catalog coordinate dev7a/laguna-s:dflash.

Provenance and reproduction

The source weights are Poolside's SafeTensors at revision b0486d1586daa0d56435c508108171fc1c8daff9. The target tokenizer comes from Laguna S 2.1 revision 00af5a51782109b587a3b3bbf11875e566036fa7. Conversion uses Poolside's llama.cpp revision 06f8cebd7fe728687be3d19f8bdedb70d75883af.

The source manifest records every downloaded file, byte size, and SHA-256. The dependency locks contain hashes and use binary packages only. The released file and a clean repeated conversion were byte-identical.

Reproduce on Linux AArch64 with Python 3.12:

python3.12 scripts/download_sources.py --destination sources
python3.12 scripts/reproduce.py --sources sources --repeat-check

Verify the release without third-party Python packages:

python3.12 scripts/verify.py Laguna-S-2.1-DFlash-Q8_0.gguf

This is a community conversion, not an official Poolside release. The weights remain under OpenMDW-1.1. The unmodified upstream license is in LICENSE. The conversion-code license and third-party notices are in LICENSE.code and THIRD_PARTY_NOTICES.md.