GitLab

Changes resume_model_weights_only flag in train_openfold to reload weights only, without parsing a time step.

- much faster due to the use of threading and mp
- also supports sharding

Adds mkl version to environment.yml

Fix resolution field in mmcif_parsing

Fix distributed seeding behavior

Without the break statement this function will fall back to setting the resolution to _reflns.d_resolution_high for most structures, which reflects the resolution limit in the raw data and not the final structure (see https://mmcif.wwpdb.org/dictionaries/mmcif_pdbx_v50.dic/Items/_reflns.d_resolution_high.html and https://mmcif.wwpdb.org/dictionaries/mmcif_pdbx_v50.dic/Items/_refine.ls_d_res_high.html).

This adds workers=True to the Lightning seed_everything function which guarantees different random states across all processes in distributed training. Prior to that some processes on different GPUs with the same worker ID could share the same random state.

Note that this will break reproducibility between runs prior to and after this change.

Also removes the seed and supress_output modules that were not used anymore in OpenFold.