esm2 with .csv files

[nayeemmz]

Hi, I am trying to run python predict.py \ --model-location esm1v_t33_650M_UR90S_1 esm1v_t33_650M_UR90S_2 esm1v_t33_650M_UR90S_3 esm1v_t33_650M_UR90S_4 esm1v_t33_650M_UR90S_5 \ --sequence HPETLVKVKDAEDQLGARVGYIELDLNSGKILESFRPEERFPMMSTFKVLLCGAVLSRVDAGQEQLGRRIHYSQNDLVEYSPVTEKHLTDGMTVRELCSAAITMSDNTAANLLLTTIGGPKELTAFLHNMGDHVTRLDRWEPELNEAIPNDERDTTMPAAMATTLRKLLTGELLTLASRQQLIDWMEADKVAGPLLRSALPAGWFIADKSGAGERGSRGIIAALGPDGKPSRIVVIYTTGSQATMDERNRQIAEIGASLIKHW \ --dms-input ./data/BLAT_ECOLX_Ranganathan2015.csv \ --mutation-col mutant \ --dms-output ./data/BLAT_ECOLX_Ranganathan2015_labeled.csv \ --offset-idx 24 \ --scoring-strategy wt-marginals

code using esm2_t48_15B_UR50D() and other esm2 models instead of the esm1 models. However, the program terminates without finishing. I should mention that the code runs fine with esm1 models and I have downloaded esm2 models. Even cloned esm repository locally to see if that would help. Is it that esm2 needs to be run only with fasta and .pdb files?

[mdlakic]

In model names, third part of the string (ends in M or B) tells you about the number of parameters. In simple terms, it correlates with model size. All those esm1v_t33_650M_UR90S models have 650 million parameters, while esm2_t48_15B_UR50D has 15 billion parameters. I don't know the exact GPU memory requirement for that model, but it might be 40Gb+. You may be able to run esm2_t33_650M_UR50D, and it isn't a significant difference in terms of accuracy from that model to esm2_t36_3B_UR50D or esm2_t48_15B_UR50D.

[nayeemmz]

Thanks for the response. I knew about the number of parameters in each model, didn't pay attention to the possibility of that being the issue of concern here. Thanks for pointing it out. I think that is probably the case. I will test it out and update here.

[nayeemmz]

You're right I confirmed that it is a GPU memory issue.