fair-esm2 — ESM-2 (Meta AI)

ESM-2 code and weights are MIT (Meta AI, github.com/facebookresearch/esm).

Package disambiguation. pip install fair-esm gives you import esm with esm.pretrained.* (ESM-1/2). Biohub's github.com/Biohub/esm fork (MIT) gives you from esm.models.esmfold2 import ESMFold2InputBuilder — see the esmfold2 skill. Both share the esm namespace but are different libraries. This skill covers fair-esm (the Meta package).

Prerequisites

| Requirement | Minimum | Recommended | | ----------- | ------- | ----------- | | Python | 3.8+ | 3.11 | | CUDA | 11.7+ | 12.x | | GPU VRAM | 8 GB (8M), 16 GB (650M) | 24 GB+ (650M / 3B) |

How to run

Embeddings

import torch, esm

model, alphabet = esm.pretrained.esm2_t33_650M_UR50D()
model = model.eval().cuda()
bc = alphabet.get_batch_converter()

_, _, toks = bc([("ubq", "MQIFVKTLTGKTITLEVEPSDTIENVK")])
with torch.no_grad():
    out = model(toks.cuda(), repr_layers=[33])
emb = out["representations"][33]      # (1, L+2, 1280) — includes BOS/EOS
seq_emb = emb[0, 1:-1].mean(0)        # per-sequence mean

Masked-LM scoring

with torch.no_grad():
    out = model(toks.cuda(), repr_layers=[33])
logits = out["logits"][0, 1:-1]       # (L, |vocab|)
# WT marginal log-likelihood; for mutation scoring, mask the position and
# compare logit[mut] − logit[wt].

Contact prediction

with torch.no_grad():
    out = model(toks.cuda(), repr_layers=[33], return_contacts=True)
contacts = out["contacts"][0]         # (L, L)

Models

| Name | Layers | Dim | Params | Use | | -------------------------- | ------ | ---- | ------ | -------------------------- | | esm2_t6_8M_UR50D | 6 | 320 | 8 M | Fast smoke / tiny embeddings | | esm2_t33_650M_UR50D | 33 | 1280 | 650 M | Default embedding model | | esm2_t36_3B_UR50D | 36 | 2560 | 3 B | Best embeddings, 24 GB+ |

Output format

out["representations"][layer] is (B, L+2, D); slice [ :, 1:-1, : ] to drop BOS/EOS. out["contacts"] (when return_contacts=True) is (B, L, L).

Remote compute

Needs ≥16 GB VRAM (650M model) and either pre-cached .pt checkpoints or egress to dl.fbaipublicfiles.com. Read compute_details({provider, mode:'read'}) for an environment with fair-esm and a torch-hub weight cache, then:

c = host.compute.create(provider)
job = c.submit_job(
    intent="ESM-2 650M embeddings for 200 sequences — 1×GPU, ~2 min",
    inputs=[
        {"src": "seqs.fasta", "dst_filename": "seqs.fasta"},
        {"src": "embed_esm2.py", "dst_filename": "embed_esm2.py"},
    ],
    command="python3 embed_esm2.py",
    environment=...,   # env name from compute_details
    outputs=["embeddings.pt"],
    timeout_seconds=1800,
)
print(job.job_id)   # cell ends here — kernel never blocks on compute

Then call the wait_for_notification brain-tool. When the compute_done notification arrives, act on its payload:

save_artifacts(payload["featured_files"])   # paths under hpc/<job_id>/

For the full result dict (output_files, remote_workdir, …), re-enter the kernel: c.attach_job(job_id).result() then c.close(). See the remote-compute-ssh / remote-compute-modal skill for the orchestration details.

Inside embed_esm2.py, set TORCH_HOME to the provider's torch-hub cache mount (path is in compute_details) so esm.pretrained.* resolves locally.

Troubleshooting

| Symptom | Cause | Fix | | --------------------------------------------- | ---------------------------------- | ------------------------------------- | | ModuleNotFoundError: No module named 'esm.models' | You want Biohub's esm fork, not fair-esm | See esmfold2 skill; this skill uses esm.pretrained.* | | Slow first call | Downloading weights via torch.hub | Set TORCH_HOME to a cached location |

---

Next: feed embeddings to a classifier. For structure prediction, use esmfold2.