DSpark draft sidecar GGUF for mainline llama.cpp (DFlash/DSpark speculative decoding, merged in ggml-org/llama.cpp #25173). Converted from tugot17/LFM2.5-8B-A1B-DSpark-3L with convert_hf_to_gguf.py --target-model-dir <target>.

Pairs with the target model LiquidAI/LFM2.5-8B-A1B-GGUF — the sidecar carries only the drafter (block size 7, rank-256 Markov head, confidence head); token embeddings and lm_head are shared from the target at load time.

Greedy outputs are identical to the target alone; the timings field of each response reports draft_n / draft_n_accepted for acceptance-length tracking.