OpenFold

OpenFold at the time of the release of our original model parameters and training database. Adds countless improvements over the previous beta release, including, but not limited to:

• Many bugfixes contribute to stabler, more correct, and more versatile training
• Options to run OpenFold using our original weights
• Custom attention kernels and alternative attention implementations that greatly reduce peak memory usage
• A vastly superior Colab notebook that runs inference many times faster than the original
• Efficient scripts for computation of alignments, including the option to run MMSeqs2's alignment pipeline
• Vastly improved logging during training & inference
• Careful optimizations for significantly improved speeds & memory usage during both inference and training
• Opportunistic optimizations that dynamically speed up inference on short (< ~1500 residues) chains
• Certain changes borrowed from updates made to the AlphaFold repo, including bugfixes, GPU relaxation, etc.
• "AlphaFold-Gap" support allows inference on complexes using OpenFold and AlphaFold weights
• WIP OpenFold-Multimer implementation on the multimer branch
• Improved testing for the data pipeline
• Partial CPU offloading extends the upper limit on inference sequence lengths
• Docker support
• Missing features from the original release, including learning rate schedulers, distillation set support, etc.