sokrypton
Keep Jax compiled
First of all thanks for the great work! I really appreciate colabfold!
I am running colabfold (localcolabfold) on a GPU cluster. Using the 'turbo' mode works great when applied to models after the first model is compiled (First model around 2 min for a 100aa protein, following 4 models only 1-5s each) which is awesome, but I was wondering if there is a way to 'keep Jax compiled' for following predictions of proteins the same length to keep all models running at 1-5s.
Just deleting the 'done.txt' and predicting the same .fasta file in the output directory doesnt work, so I was wondering if this is not possible or if I am missing something.
Our colab_batch (installable through pip) can run batch jobs and keeps the model compiled. It sorts your input sequences by length and pads the input to optimize the amount of compilations needed.
To add a bit more context, we currently can't save the compiled models as jax doesn't support it (https://github.com/google/jax/issues/476), so you need to have all proteins in the same run.
