Draftfit · LFM2.5-8B-A1B DSpark
A validation checkpoint for Draftfit — train DFlash-family drafts, convert compatible families, and continue training existing drafts.
Performance
141.3 tok/s overall · +8.2% over target-only.
End-to-end request throughput in tok/s (including prefill and target verification); the second table reports decode-only throughput and draft-token acceptance.
| Draft / mode | Overall | OpenCode | CodeAlpaca | SciQ | PerfectBlend |
|---|---|---|---|---|---|
| Without Dspark | 130.5 | 113.3 | 144.7 | 138.3 | 132.9 |
| Official DSpark | 141.2 | 122.0 | 141.8 | 149.1 | 138.9 |
| Draftfit DSpark | 141.3 | 125.2 | 148.2 | 149.2 | 142.5 |
| Mode | Decode-only tok/s | Accepted / proposed draft tokens |
|---|---|---|
| Without Dspark | 160.2 | — |
| Official DSpark | 166.0 | 71.9% |
| Draftfit continual | 166.2 | 75.0% |
MLX measurements reported by the maintainer; values preserved from the current model card. Historical results in evaluation.md are from a separate run.
How to use
SGLang · NVIDIA GPU
With the DSpark-enabled SGLang build:
SGLANG_RAGGED_VERIFY_MODE=static python -m sglang.launch_server \
--model-path LiquidAI/LFM2.5-8B-A1B \
--speculative-algorithm DSPARK --speculative-draft-model-path jayyun98/LFM2.5-8B-A1B-Draftfit-DSpark \
--speculative-draft-attention-backend flashinfer --disable-radix-cache --port 30000
vLLM · NVIDIA GPU
Not ready on stock vLLM: the LFM2-MoE target lacks the auxiliary-state interface required by DSpark in the inspected source. Use SGLang for this model; details.
MLX · Apple silicon
After one-time MLX setup:
hf download jayyun98/LFM2.5-8B-A1B-Draftfit-DSpark run_mlx.py --local-dir ./lfm-draft
python ./lfm-draft/run_mlx.py --target LiquidAI/LFM2.5-8B-A1B-MLX-4bit \
--draft jayyun98/LFM2.5-8B-A1B-Draftfit-DSpark --prompt "Hello!"
Validation: real draft reload passed for both models; Taichu MLX passed a fresh 8-token target-only comparison. CUDA server generation is not yet verified. Setup and compatibility.