#### Prerequisite > This directory contains only the runtime scripts for **Dataset Preparation** and **Model** (model inference/generation). > **Please first follow [`../torchx_gpu/README.md`](../torchx_gpu/README.md) and install from `../torchx_gpu/torchx`**, including: > - Conda environment creation and installation of torch, etc. > - Torchx installation, chemical data building > > After completing those steps, return here for the actual design run. Recommended runtime: **Linux x86_64 + Python 3.11 + PyTorch 2.8.0 (CUDA 12 / cu128)**. `src/checkpoint_reentrant_function.cpython-311-x86_64-linux-gnu.so` is required for gradient checkpointing and only matches this Python/ABI. After the torchx env is active: ```shell cd src python -c "from checkpoint_reentrant_function import CheckpointReentrantFunction; print('ok')" ``` #### Optional: cueq acceleration Default backends in `task/config/base.yaml` are `"torch"` and do not need extra packages. To enable cuEquivariance triangle kernels, install the pinned CUDA 12 packages **after** torchx is installed. Do not run an unpinned `pip install cuequivariance-torch` — that can replace torch with a CUDA 13 build. ```shell pip install -r requirements-cueq.txt python -c "import torch; print(torch.__version__)" # still 2.8.0 ``` Then set `TRIANGLE_MULTIPLICATIVE` and `TRIANGLE_ATTENTION` to `"cuequivariance"` in `task/config/base.yaml`. #### Design Before running, modify paths in the design script and in `task/config/base.yaml`. In `task/config/base.yaml`, set **exactly one** of: - `model_dir`: directory of official AlphaFold 3 parameters - `checkpoint_path`: TorchFold trained checkpoint file (`.pt`) If both are set, `checkpoint_path` takes priority. Comment out the unused one. ```shell cd task/monomer_unconditional bash batch_submit.sh ``` VHH and mini-binder design entry points: `task/vhh/vhh_design.sh` and `task/mini_binder/mini_binder_design.sh`.