GGUF conversions of RespectMathias/Laguna-XS-2.1-DSpark, a DSpark speculative decoding model for poolside/Laguna-XS-2.1.

This drafter requires Laguna DSpark support from RespectMathias/llama.cpp. It is loaded as a DFlash GGUF because llama.cpp implements DSpark as DFlash plus Markov and confidence heads. Select DSpark proposal behavior with --spec-type draft-dspark.

Q8_0 or Q6_K is recommended for draft quality. Lower-bit draft quantization can reduce speculative acceptance enough to offset memory savings.

The checkpoint uses block size 16 and proposes up to 15 draft tokens. It shares tokenizer, token embeddings, and output projection with target model.

Quantizations were generated from BF16 GGUF with llama-quantize.

Runtime validation completed with Laguna-XS-2.1 target and draft-dspark: model loading, fused QKV, auxiliary feature norms, causal draft attention, Laguna SWA RoPE, Markov head, and token acceptance all executed successfully.

Source checkpoint is an experimental small-data training run. Measured acceptance depends on prompt, target quantization, sampling settings, and draft quantization. See source model card for training details and limitations.

OpenMDW-1.1. Inherited from poolside/Laguna-XS-2.1.