Fix ICEBERG forward prediction on pip-install / CPU / pytorch_lightning 2.x

[hugogontijomachado]

Summary

Three small, backward-compatible fixes so ICEBERG forward prediction
(iceberg_prediction() → dag_pred/predict_smis.py) runs end-to-end when ms_pred is
pip-installed, on CPU, and with pytorch_lightning ≥ 2.0. None of them change
behavior for an editable checkout / GPU / PL 1.x.

1. dag_pred/iceberg_elucidation.py — resolve predict_smis.py relative to the package

iceberg_prediction() builds the subprocess with a cwd-relative path:

cmd = f'''{python_path} src/ms_pred/dag_pred/predict_smis.py ...'''

That only resolves when run from the repo root. On a pip install there is no src/, so the
subprocess fails (can't open file '.../src/ms_pred/dag_pred/predict_smis.py') and
load_pred_spec then raises on the missing preds.hdf5. Fixed by locating the script via
Path(__file__).resolve().parent / "predict_smis.py" (the two files live side by side).

2. dag_pred/predict_smis.py — pl.seed_everything

pytorch_lightning.utilities.seed.seed_everything was removed in PL 2.0
(AttributeError: module 'pytorch_lightning.utilities.seed' has no attribute 'seed_everything').
pl.seed_everything is the public API and exists in both PL 1.x and 2.x.

3. dag_pred/predict_smis.py — guard torch.cuda.set_device on CPU

In producer_func, torch.cuda.set_device(gpu_id) runs unconditionally, but gpu_id is
only assigned in the GPU branch — on CPU this raises NameError (and there is no CUDA
device to select). Guarded with if gpu and avail_gpu_num > 0:.

How these surfaced

Predicting MS/MS from SMILES with the public MassSpecGym checkpoints on macOS (CPU) with
pytorch_lightning 2.6.5; each fix revealed the next.

Testing

After all three, iceberg_prediction(...) runs end-to-end on CPU and produces preds.hdf5
(verified by predicting glyphosate MS/MS for [M-H]- and [M+H]+).

[rogerwwww]

Hi Hugo,

First, my apologies that we weren’t able to review this PR sooner, and thank you for tracking down these issues, documenting them so clearly, and testing the fixes on CPU.

Since this PR was opened, the codebase has been reorganized, dag_pred was renamed to iceberg. The installed-package path fix and the PyTorch Lightning compatibility updates were subsequently merged through #41. The CPU-device issue has also been addressed on the current main branch using use_gpu = gpu and torch.cuda.device_count() > 0.

As a result, all three issues you identified have now been addressed. Thank you again for identifying and carefully reporting these problems. I’m closing this PR as superseded by #41.