esm

1# ESM: protein generation, embeddings, and folding
2 
3## When to use
4 
5Use for Biohub's `esm` SDK: ESM3 masked multimodal generation, ESMC sequence
6representations, and ESMFold2 all-atom prediction. The former Forge platform
7migrated to `https://biohub.ai`, including ESM3; SDK class names still contain
8`Forge`. This skill targets the released **esm 3.4.1.post1**, verified against
9its wheel and current official documentation.
10 
11`fair-esm` is Meta's separate, older ESM2/ESMFold/ESM-IF distribution. Both
12packages import as `esm`; install them in separate environments. The legacy
13`esm.pretrained.esm2_*` interface is not the Biohub ESMC interface. Hugging Face
14Transformers' native ESMC implementation is another API: do not interchange its
15version requirements or output types with this SDK.
16 
17## Setup and model choice
18 
19```bash
20uv venv --python 3.12 .venv-esm
21uv pip install --python .venv-esm/bin/python "esm==3.4.1.post1"
22```
23 
24The release declares Python >=3.12, Torch >=2.11,<2.12 and Transformers
25>=4.57.6,<5. Python 3.12, Torch 2.11.0 and Transformers 4.57.6 were tested on
26CPU. Linux x86_64 installs include GPU-specific dependencies; use a supported
27inference machine and budget disk space. Flash Attention is optional; it is
28unnecessary for the tiny CPU tests. Do not install this stack into a shared
29environment containing incompatible Transformers or Torch pins.
30 
31| Task | Local model/API | Hosted client/model |
32| --- | --- | --- |
33…