EvSpark Drafter Checkpoints

Lossless speculative decoding for Evo2 (StripedHyena2 hybrid DNA foundation models).

Code & docs: https://github.com/dhnihaoya/EvSpark Mirror (China): https://www.modelscope.cn/models/dinghao1120/EvSpark

EvSpark accelerates single-stream Evo2 7B generation by 2.9–3.2× with a small distilled drafter, while staying provably lossless (rejection-sampling verified; greedy output token-for-token identical, 0 non-tie divergences on 24 prompts × 4 checkpoints). Speedup holds from 1k to 262k context. Training a competitive drafter costs ~1 GPU-hour on a single RTX 4090.

Checkpoints

File	γ	Distill budget	Suite speedup	Note
L27_g12_80M_s1.pt / _s2.pt	12	80M	3.16× (mean of seeds)	flagship
L27_g12_30M_s1.pt / _s2.pt	12	30M	3.03×	"~1 GPU-hour" cell
L27_final15_150M_s1.pt / _s2.pt	7	150M	2.89× / 2.87×	main-table model

All drafters: single injection layer L27, d_model=1024, distilled offline from frozen Evo2 7B hidden states. The target model is never fine-tuned. Each checkpoint is self-contained (frozen embedding, Markov head, confidence head, scheme/d_model/γ metadata) and loads directly via specdec.block.neural_draft.NeuralDraftModel.from_checkpoint.

Usage

# 1. environment: torch 2.7.1 + flash-attn 2.8.0.post2 + evo2  (see GitHub README)
# 2. fetch a checkpoint (HF first, ModelScope fallback)
python scripts/download_ckpt.py L27_g12_80M_s1
# 3. run the demo
python scripts/demo.py --ckpt L27_g12_80M_s1 --n-tokens 1024

Or directly:

from huggingface_hub import hf_hub_download
hf_hub_download("dinghhhhhhhhhhhhhhh/EvSpark", "L27_g12_80M_s1.pt", local_dir="checkpoints")

License

Checkpoint weights: CC-BY-NC-4.0. Code: MIT (GitHub). Evo2 / Vortex upstream assets are under their own licenses.