https://github.com/ExpHP/gpaw-raman-script/blob/main/ep_script/leffers.py#L369-L431

A lot of the _distribute_bands_by_k function in leffers.py is silly, as is the similar code near the beginning of get_dipole_transitions, because it's effectively sending data from the other nodes to the root node (in the calc.wfs.collect_ methods) and then sending it back.

We should do what GPAW does and have each node iterate over calc.wfs.kpt_u, which already has a list of kpoint objects assigned to the current node.

This is blocked on #3 since the objects in calc.wfs.kpt_u also have their own spin.

This is still based on code that didn't support spin, even though spin is very simple to add. Mostly it just changes the shape of some output arrays.

when I run the raman code with multiple nodes on the cluster, it gets stuck at some point. force-set-eq.npy and raman-eq.ex.gz are generated but after that it keeps running with no output. It runs normally when I use one core though. Do you have an idea what could be the problem?

I can confirm this happens if I make a copy of the expected output directory, delete *.npz *.npy *.gz and run PYTHONPATH=.. gpaw -P2 python -m script ch4.

(note: there seems to also be another problem that's making it not generate phonopy_disp.yaml, which is why I start from a copy of the expected output directory. For that one I think the non-root processes might be checking too early and we need a barrier)