MACE interface, vol. 2

[danielhollas]

This is a continuation from #224, with the following changes

• remove &mace section from the abin input file, instead the path to the model and other configuration is passed via command line parameter to mace_server.py. This is specified at the top of utils/run.mace_abin.sh
• Fix coordinates received from ABIN (without this the interface was simply not working)
• Use MACECalculator instead of calling pytorch directly
• Fix float32 evaluation
• removed install_mace.sh and the MACE-OFF model file
• test mace_server.py in MACE test directly
• lot's of error-handling improvements
• add tests for error paths

TODO:

• Test run.mace_abin.sh
• Let Veronika test everything
• Check size of incoming coordinates from ABIN in mace_server.py
• Use logging module?

[danielhollas]

Okay, I think this is mostly ready now for merging! We'll do some internal testing. @FelyCZ would be great if you could have a look. I summarized the changes I made in the PR description.

@JanosJiri up to you how much you want to review this. Perhaps at least the README files and mace_server.py for clarity. I'll ask people here for review as well.

[JanosJiri]

Hi @danielhollas, I'm sorry but I struggle with time now and I can't make a useful review in a short time. I checked README.md though and it looks clean and understandable to me. Sorry.

[FelyCZ]

Hello, thanks for revising the MACE interface!
I think it is now ready to be released. I have left you some comments and implemented the logging module (see PR #257).

[FelyCZ]

I think we need to include torch.cuda.synchnronize() when the computation is running on CUDA. Otherwise is won't be measuring the time correctly.

if config.device == "cuda":
    import torch
    torch.cuda.synchronize() 

end = perf_counter()

[danielhollas]

Is it really needed? Surely mace_model.evaluate cannot return the final result until CUDA kernels finished?

[FelyCZ]

Move natqm to global scope (as it is used in both subroutines)?

use mod_qmmm, only: natqm

[danielhollas]

Yeah, we could, but it also is nice to have it imported locally as it's closer to its use. I'll leave it like that for now.

[danielhollas]

@FelyCZ thanks for taking a look! Let's merge this, and we can continue with tweaks in #257.

We've tested this locally with our own MACE model. We might need to do further testing with various foundational model but that can be a follow-up as well.

Thanks for your original work on this @FelyCZ! I think this is going to be super-useful going forward.