Evo2 Agent Skill

Evo 2 — DNA Language Model

Prerequisites

| Requirement | Minimum | Recommended |

| ----------- | ------- | ---------------- |

| Python | 3.11 | 3.12 (<3.13) |

| CUDA | 12.1+ | 12.4+ |

| GPU VRAM | 24 GB (7B bf16) | 80 GB (40B) |

| RAM | 32 GB | 128 GB |

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

| Name | Params | Context | VRAM (bf16) | Notes |

| ----------- | ------ | ------- | ----------- | ---------------------------------- |

| evo2_7b | 7 B | 1 M nt | ~22 GB | Default; fits on a single 24 GB+ GPU |

| evo2_40b | 40 B | 1 M nt | ~78 GB | H100 80 GB or multi-GPU |

| evo2_1b_base | 1 B | 8 K nt | ~6 GB | FP8 path requires sm_89+ (H100) |

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). Use a selected and

probed ssh:<alias> context and load remote-compute-ssh. Confirm that the

environment imports Evo 2 and that the desired weights are cached. Submit a

self-contained scoring script through one run_in_context call:

{
  "context_id": "ssh:gpu-box",
  "title": "Evo 2 variant scoring",
  "command": "source ~/miniforge3/etc/profile.d/conda.sh && conda activate evo2 && HF_HOME=/srv/model-cache HF_HUB_OFFLINE=1 python score_evo2.py --output /home/me/wisp-results/evo2/scores.json",
  "timeout_secs": 1800,
  "input_paths": ["runs/score_evo2.py"],
  "output_specs": [
    {
      "glob": "ssh://gpu-box/home/me/wisp-results/evo2/scores.json",
      "kind": "json",
      "residency": "remote"
    }
  ]
}

Replace context, environment, cache, and output paths with discovered values.

Call monitor_run once to wait, get_run once for a snapshot, or cancel_run

to stop. Set HF_HUB_OFFLINE=1 only after confirming the cache is complete, so

the loader does not try to write refs/ into a read-only mount. Weight footprint:

~15 GB (7B), ~80 GB (40B).

Typical performance

| Task | 7B on H100 | Notes |

| --------------------------- | ---------- | --------------------------- |

| Model load (cached) | ~5-7 min | First call hydrates weights |

| score_sequences, 200×200bp| ~10-20 s | After load |

| generate, 1×512 nt | ~15 s | |

Troubleshooting

| Symptom | Cause | Fix |

| ------------------------------------ | ------------------------------ | ------------------------------------------ |

| Transformer Engine not installed | No FP8 — falls back to bf16 | Informational only on non-H100; ignore |

| OOM on load | 40B on <80 GB GPU | Use evo2_7b or shard with device_map |

| HF tries to write refs/main | HF_HOME points at RO mount | Set HF_HUB_OFFLINE=1 |

| dtype mismatch in score_sequences| Passing tensors not strings | Pass list[str]; the API tokenises for you |

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

variants.