openfold3

1# OpenFold3 Structure Prediction
2 
3## Prerequisites
4 
5| Requirement    | Minimum | Recommended |
6| -------------- | ------- | ----------- |
7| Python         | 3.10+   | 3.11        |
8| CUDA           | 12.1+   | 12.4+       |
9| GPU VRAM       | 24GB    | 80GB (H100) |
10| RAM            | 32GB    | 64GB        |
11| Disk (weights) | 3GB     | -           |
12 
13## How to run
14 
15### Installation
16 
17```bash
18pip install 'openfold3[cuequivariance]==0.4.1'
19```
20 
21The default attention kernel is DeepSpeed `DS4Sci_EvoformerAttention`. If
22DeepSpeed is unavailable, switch to the cuEquivariance triangle kernels (no
23build-from-source) by overriding the eval memory settings in
24`model_config.py` (`use_deepspeed_evo_attention: False`,
25`use_cueq_triangle_kernels: True`). Some pre-built environments already ship
26this override; check before re-patching.
27 
28### Weights
29 
30Apache-2.0, ~2.3 GB from HF `OpenFold/OpenFold3`. The repo is **gated** (auto-approval) — accept the access form on the HF model page and authenticate (`huggingface-cli login` or `HF_TOKEN`) before downloading:
31 
32```bash
33export OPENFOLD_CACHE=~/.openfold3
34huggingface-cli download OpenFold/OpenFold3 checkpoints/of3-p2-155k.pt \
35  --local-dir "$OPENFOLD_CACHE"
36```
37 
38`run_openfold` will also auto-download to `$OPENFOLD_CACHE` on first run if
39egress is open and HF credentials are available (either `HF_TOKEN` or a prior
40`huggingface-cli login`) with repo access granted. The interactive
41…