fair-esm2

1# fair-esm2 — ESM-2 (Meta AI)
2 
3ESM-2 code and weights are MIT (Meta AI, github.com/facebookresearch/esm).
4 
5> **Package disambiguation.** `pip install fair-esm` gives you `import esm`
6> with `esm.pretrained.*` (ESM-1/2). Biohub's github.com/Biohub/esm fork
7> (MIT) gives you `from esm.models.esmfold2 import ESMFold2InputBuilder` —
8> see the **`esmfold2`** skill. Both share the `esm` namespace but are
9> different libraries. This skill covers **fair-esm** (the Meta package).
10 
11## Prerequisites
12 
13| Requirement | Minimum                 | Recommended        |
14| ----------- | ----------------------- | ------------------ |
15| Python      | 3.8+                    | 3.11               |
16| CUDA        | 11.7+                   | 12.x               |
17| GPU VRAM    | 8 GB (8M), 16 GB (650M) | 24 GB+ (650M / 3B) |
18 
19## How to run
20 
21### Embeddings
22 
23```python
24import torch, esm
25 
26model, alphabet = esm.pretrained.esm2_t33_650M_UR50D()
27model = model.eval().cuda()
28bc = alphabet.get_batch_converter()
29 
30_, _, toks = bc([("ubq", "MQIFVKTLTGKTITLEVEPSDTIENVK")])
31with torch.no_grad():
32    out = model(toks.cuda(), repr_layers=[33])
33emb = out["representations"][33]      # (1, L+2, 1280) — includes BOS/EOS
34seq_emb = emb[0, 1:-1].mean(0)        # per-sequence mean
35```
36 
37### Masked-LM scoring
38 
39```python
40with torch.no_grad():
41    out = model(toks.cuda(), repr_layers=[33])
42logits = out["logits"][0, 1:-1]       # (L, |vocab|)
43# WT marginal log-likelihood; for mutation scoring, mask the position and
44…