Is there a way to provide an installation-default runner_yaml file?

[roccomoretti]

Is your feature request related to a problem? Please describe.

My system is running afoul of the missing evoformer_attn.so issues described in #17 & #23
The fix recommended is to provide a custom -runner_yaml file which turns off use_deepspeed_evo_attention. However, it's a bit inconvenient to have to remember to pass this command line parameter each time. It would be more convenient if I can set these defaults once at the install level, and have them take effect automatically.

Describe the solution you'd like
Provide a location (probably the $OPENFOLD_CACHE directory) where you can place a central runner yaml file that would take effect for each run of that install. (And which could be overridden by the command line -runner_yaml, if needed.)

Describe alternatives you've considered
I guess you could manually edit the values in openfold3/entry_points/validator.py, but manually editing the official install files seems hacky to me, and would likely not persist if OpenFold was updated. It would also be much more user friendly to have the default values be set using the same syntax as the standard configuration takes.

[jnwei]

I agree that it would be helpful to have some additional personal settings stored, perhaps in $OPENFOLD_CACHE

I think I would implement this solution in the following way

1. Add the settings for cuequivariance only as a model preset, e.g. add the following section to the model presets for the of3_all_atom project

cuequivariance_only:
   settings:
      memory:
        eval:
          use_cueq_triangle_kernels: true
          use_deepspeed_evo_attention: false

This will create the cuequivariance_only model preset, which can be used in conjunction with the other presets like low_mem, pae_enabled

2. Add some logic to read runner.yml files saved in $OPENFOLD_CACHE. One option could be to make an additional constructor for InferenceExperimentConfig that
   i. Checks if there is a default runner.yml in $OPENFOLD_CACHE
   ii. If there is default runner.yml loads this runner.yml as a dictionary (using config_utils.load_yaml or something similar)
   iii. Updates these user-default settings with any provided runner.yml inputs provided by the user at the command line.

Would this make sense?

Between some infrastructure tasks and vacation I have in the next few weeks, it will take me some time to implement this change. I'd appreciate any external contributions.

---

As a workaround, one option would be to add a line that loads a user-default runner.yml in run_openfold predict before the command line runner.yml is loaded, so possibly here.