EvSpark

EvSpark — lossless speculative decoding for Evo2

Drafter checkpoints for EvSpark: a small distilled drafter that accelerates single-stream Evo2 7B (StripedHyena2) generation while staying distributionally lossless.

Paper (bioRxiv) · DOI · GitHub · ModelScope

Numbers: Evo2 7B bf16, single stream, RTX 4090, 48 genomic prompts × 1024 tokens, 3 training seeds. One checkpoint serves any decode-time draft length γ′ ≤ training γ (exact causal prefix; no retraining).

Beyond 7B: the same drafter recipe retrained for Evo2 20B / 40B reaches 2.18–2.46× (real-43) and 2.51–2.78× (all-48) on one H20 (SDPA), and a predictor-guided regulatory-DNA design workflow sees a median 1.57× complete-design speedup over a calibrated batched native baseline (see paper).

Checkpoints

All four files are single-layer L27 injection, d_model=1024, distilled offline from frozen Evo2 7B hidden states. The target model is never fine-tuned. Each .pt is self-contained (frozen embedding, Markov head, confidence head, metadata).

s1 / s2 are independent training seeds of the same cell.

SHA-256

Quickstart

These files are drafters, not a standalone DNA LM. You need Evo2 7B + the EvSpark engine (block verify + Hyena/attention state-slice rollback).

git clone https://github.com/dhnihaoya/EvSpark && cd EvSpark
# env: torch 2.7.1 + flash-attn 2.8.0.post2 + evo2  (see GitHub README)
python scripts/download_ckpt.py L27_g12_150M_s1          # HF first, ModelScope fallback
python scripts/demo.py --ckpt L27_g12_150M_s1 --n-tokens 1024

Library:

from evspark import EvSpark

with EvSpark.load("L27_g12_150M_s1") as es:
    out = es.generate("ACGTACGT...", n_tokens=1024)       # T=1.0, top_k=4
    print(out.text, f"{out.tok_s:.1f} tok/s, tau={out.mean_tau:.2f}")

    g = es.generate(prompt, greedy=True)
    nat = es.generate_native(prompt, greedy=True)
    assert g.ids.tolist() == nat.ids.tolist()             # token-for-token

Decode-time γ′ without retraining:

es.generate(prompt, n_tokens=1024, gamma=8)               # γ′=8 from a γ=12 ckpt

Manual download:

from huggingface_hub import hf_hub_download
path = hf_hub_download("dinghhhhhhhhhhhhhhh/EvSpark", "L27_g12_150M_s1.pt")

If huggingface.co is unreachable: HF_ENDPOINT=https://hf-mirror.com, or pull the same files from ModelScope.

Citation

@article{ding2026evspark,
  title   = {EvSpark: Lossless Speculative Decoding for Hybrid DNA Foundation Models},
  author  = {Ding, Hao and Wu, Nannan and Qiu, Tianyi},
  journal = {bioRxiv},
  year    = {2026},
  doi     = {10.64898/2026.09.02.749017},
  url     = {https://www.biorxiv.org/content/10.64898/2026.09.02.749017}
}

Paper: https://www.biorxiv.org/content/10.64898/2026.09.02.749017

License

Checkpoints and code are MIT. Evo2 / Vortex weights and runtime follow their upstream licenses.