Bugifx on random seed inconsistencies and keyword grad_max_norm

[yingkaisha]

Bugfix

• Some places inside credit uses random seed 42 as default and left untouched in train.py and train_multistep.py. This is fixed so that all random seed entries accept the random seed from YAML input.
• Keyword grad_max_norm did not apply gradient clipping properly. It is now fixed in trainerERA5_v2.py and trainerERA5_multistep_grad_accum.py

[yingkaisha]

ready

[jsschreck]

This looks good, thanks for finding the seed inconsistency. For grad norm, sometimes I use it, sometimes not (I have found it most useful with LSTMs/GRUs, have not tested extensively with transformers). Do we want to add an option not to use it (for instance maybe 0.0 is the flag to ignore it?)

[yingkaisha]

This looks good, thanks for finding the seed inconsistency. For grad norm, sometimes I use it, sometimes not (I have found it most useful with LSTMs/GRUs, have not tested extensively with transformers). Do we want to add an option not to use it (for instance maybe 0.0 is the flag to ignore it?)

I will add it

[djgagne]

Changes look good to me. Thank you for catching the issue with the random seed setting.