FAIR ESM (Evolutionary Scale Modeling)

2.0.0 verified Mon Apr 27 auth: no python

FAIR ESM provides pretrained transformer language models for proteins, including ESM-2 and ESM-1b. Version 2.0.0 adds new models and enhancements. The library is actively maintained by Meta AI.

pip install fair-esm

Common errors

error RuntimeError: "LayerNormKernelImpl" not implemented for 'Half' ↓

cause Half precision not supported on CPU.

fix

Use torch.float32 or run on GPU with torch.cuda.amp if supported.

error KeyError: 'representations' ↓

cause Incorrect repr_layers argument; model does not have requested layer.

fix

Set repr_layers to a valid layer number, e.g., [33] for ESM-2 650M which has 33 layers.

error ModuleNotFoundError: No module named 'esm' ↓

cause Package not installed or wrong import name.

fix

Run: pip install fair-esm, then import as esm (not fair_esm).

error AttributeError: module 'esm' has no attribute 'pretrained' ↓

cause Old version of fair-esm (<2.0.0?) or import error.

fix

Upgrade to latest: pip install --upgrade fair-esm, or use torch.hub.load for older versions.

Warnings

breaking In v2.0.0, model loading through torch.hub is deprecated; use esm.pretrained instead. ↓

fix Use from esm.pretrained import ... instead of torch.hub.load('facebookresearch/esm', ...)

breaking The 'esm.pretrained' module no longer provides the old ESM-1b model alias 'esm1b_t33_650M_UR50S' directly; use 'esm1b_t33_650M_UR50S()' remains but check model list. ↓

fix Refer to the model zoo documentation for updated model names.

gotcha GPU memory is very high (e.g., ESM-2 15B requires ~300GB); smaller models like 650M are recommended for most users. ↓

fix Use esm2_t33_650M_UR50D (650M params) for typical usage.

gotcha The 'return_contacts' flag in forward() can cause OOM; only use if needed. ↓

fix Set return_contacts=False (default) unless you need attention map.

deprecated The 'esm.model' submodule is being reorganized; direct class imports may break in future versions. ↓

fix Use esm.pretrained for model loading and esm for Alphabet.

Imports

ESM2
```
from esm.pretrained import esm2_t48_15B_UR50D
```
Direct import of ESM2 class is not typical; use pretrained model aliases.
Alphabet
wrong
```
from fair_esm import Alphabet
```
correct
```
from esm import Alphabet
```
The package is imported as 'esm', not 'fair_esm'.

Quickstart

Load an ESM-2 model, tokenize sequences, and extract representations.

import torch
from esm.pretrained import esm2_t33_650M_UR50D
from esm import Alphabet

model, alphabet = esm2_t33_650M_UR50D()
batch_converter = alphabet.get_batch_converter()
model.eval()

data = [
    ("protein1", "MKTVRQERLKSIVRILERSKEPVSGAQLAEELSVSRQVIVQDIAYLRSLGYNIVATPRGYVLAGG"),
]
batch_labels, batch_strs, batch_tokens = batch_converter(data)

with torch.no_grad():
    results = model(batch_tokens, repr_layers=[33], return_contacts=True)
token_representations = results["representations"][33]
print(token_representations.shape)