Evo2

Evo 2 — DNA Language Model

Prerequisites

How to run

Installation

pip install evo2
# Weights pulled from Hugging Face on first model load.

Loading and scoring

from evo2 import Evo2

model = Evo2("evo2_7b")        # or "evo2_40b" — see model table
seqs = ["ATCG" * 50, "GGGCTTAA" * 25]
ll = model.score_sequences(seqs)   # → list[float], mean per-token log-likelihood
print(ll)

Generation

out = model.generate(
    prompt_seqs=["ATGAAAGCT"],
    n_tokens=256,
    temperature=0.7,
)
print(out.sequences[0])

Models

Output format

score_sequences returns a list[float] (or np.ndarray) of mean log-likelihoods, one per input sequence. More negative ⇒ less likely under the model. For variant effect, compute Δll = ll_alt - ll_ref over a fixed window.

generate returns a GenerationOutput with .sequences (list[str]), .logits (list[Tensor]), and .logprobs_mean (list[float]) — always populated, no flag required.

Decision tree

Need a DNA model?
│
├─ Per-base/per-sequence likelihood, generation → Evo 2 ✓
├─ Predict experimental tracks (expression, accessibility) → borzoi
└─ Protein, not DNA → fair-esm2 / esmfold2

Remote compute

7B/40B inference is GPU-bound (≥24 GB / 80 GB VRAM). Read compute_details({provider, mode:'read'}) for an environment with evo2 + flash-attn and a pre-cached HF weight mount, then submit:

c = host.compute.create(provider)
job = c.submit_job(
    intent="Evo2-7B score 200bp variant window — 1×GPU, ~2 min",
    inputs=[{"src": "score_evo2.py", "dst_filename": "score_evo2.py"}],
    command="python3 score_evo2.py",   # env selection is host-specific — see compute_details for your provider
    outputs=["scores.json"],
    timeout_seconds=1800,
)
print(job.job_id)   # cell ends here — kernel never blocks on compute

Then poll from a later cell. .result() is one non-blocking probe of the remote and is what harvests the outputs once the job is terminal — nothing runs in the background, so a job you never poll is never harvested. While the job is still running it returns {"status": "running", …}; end the cell and call it again later:

r = c.attach_job(job_id).result()   # {status, exit_code, output_files,
                                    #  featured_files, remote_workdir, …}
if r["status"] == "succeeded":
    for path in r["featured_files"]:      # paths under hpc/<job_id>/
        host.save_artifact(path)
    c.close()
# `unknown` is not a finished job — poll again rather than closing over it.

See the remote-compute-ssh / remote-compute-nvidia skill for the orchestration details.

Inside score_evo2.py, point HF_HOME at the provider's weight-cache mount (path is in compute_details) and set HF_HUB_OFFLINE=1 so the loader doesn't try to write refs/ into a read-only mount. Weight footprint: ~15 GB (7B), ~80 GB (40B).

Typical performance

Troubleshooting

Next: pair with borzoi to predict track-level effects of the same variants.