LFM2.5-8B-A1B-DSpark-5L

DSpark speculative-decoding draft model for LiquidAI/LFM2.5-8B-A1B: a Qwen3-style GQA block drafter (5 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.

Usage (SGLang)

python -m sglang.launch_server \
  --model-path LiquidAI/LFM2.5-8B-A1B \
  --speculative-algorithm DSPARK \
  --speculative-draft-model-path tugot17/LFM2.5-8B-A1B-DSpark-5L \
  --speculative-draft-attention-backend flashinfer

Requires an SGLang build with DSpark and LFM2 support.