DSpark speculative-decoding
draft model for
LiquidAI/LFM2.5-1.2B-Instruct: a Qwen3-style GQA block drafter (2 layers) with a low-rank Markov transition head (rank 256) and a confidence head. Block size 7 (anchor + 6 drafts), interleaved (GPT-J) RoPE.
The draft ships no vocab weights; the embedding and LM head are tied from the target at load time.
1python -m sglang.launch_server \
2 --model-path LiquidAI/LFM2.5-1.2B-Instruct \
3 --speculative-algorithm DSPARK \
4 --speculative-draft-model-path tugot17/LFM2.5-1.2B-Instruct-DSpark-2L \
5 --speculative-draft-attention-backend flashinfer
Requires an SGLang build with DSpark and LFM2 support.
