search

dptech-corp / Uni-Fold

An open-source platform for developing protein models beyond AlphaFold.
https://doi.org/10.1101/2022.08.04.502811
Apache License 2.0
380 stars 74 forks source link

questions regarding cuda kernel and inference argument `model_name` #63

Closed pengzhangzhi closed 1 year ago

pengzhangzhi commented 2 years ago
  1. If I want to have the cuda kernels installed, I have to use your docker image? If I don't want the cuda kernels, can I skip the docker part and directly do pip install -e . ( I already have a cuda-torch compatible environment)?
  2. what does the argument model_name mean? The doc specify model name, must be consistent with model parameters confuses me. I find many so-called models and don't know what they represent and how to be consistent. I download the unifold ckpt from your URL https://uni-fold.dp.tech/unifold_params_2022-08-01.tar.gz. There are two ckpts, one for monomer, one for multimer. I want to use monomer, which model_name does monomer ckpt corresponds to?

Best, Zhangzhi

pengzhangzhi commented 2 years ago

BTW, I wonder how these cuda kernels are installed. Normally they are installed by the setup.py. But your setup.py have not code related to the installation of cuda kernels. Maybe the cuda kernels are installed by docker? If so, the annoying hardware compatibility problem would disappear.

guolinke commented 2 years ago

model_2_ft.

The CUDA kernels are installed in the docker, you can also install the uni-core by yourself.