Laguna-XS-2.1-DFlash-GGUF (quantized drafter)
Quantized GGUF builds of the official poolside/Laguna-XS-2.1-DFlash speculator, packaged for the Lucebox inference engine's speculative decoding of poolside/Laguna-XS-2.1.
| file | scheme | size | notes |
|---|---|---|---|
laguna-xs21-dflash-q4.gguf |
Q4_0 projections, Q8_0 feature projection (fc), F32 norms | 271 MB | recommended |
laguna-xs21-dflash-q8.gguf |
Q8_0 projections, F32 norms | 492 MB | conservative |
Because speculative decoding verifies every draft against the target model, drafter quantization cannot change the target's greedy output quality; the only quantity at stake is acceptance length. Measured on an RTX 3090 with the Laguna XS 2.1 Q4_K_M target (HumanEval/GSM8K/Math screens): acceptance unchanged vs the BF16 drafter, end-to-end decode ~+3% (q4), gold-scored task accuracy identical.
Produced with server/scripts/requant_dflash_draft.py from the lucebox engine
tree (GGUF-to-GGUF requant; all metadata, gates and aux norms preserved).
License: OpenMDW-1.1, inherited from the base speculator. All credit for the DFlash speculator weights to poolside.