search

google-deepmind / alphafold3

AlphaFold 3 inference pipeline.
Other
5.13k stars 576 forks source link

MemoryError - jackhammer #62

Closed hegelab closed 6 days ago

hegelab commented 1 week ago

Hi,

Conserved proteins may result in very large alignments, pulled directly into memory. This was solved in AF2: See: https://github.com/google-deepmind/alphafold/pull/296

Now, AF3 running with 256GB RAM stops with this error:

 File "/alphafold3_venv/lib/python3.11/site-packages/alphafold3/data/msa.py", line 340, in get_msa
    a3m=get_msa_tool(run_config.config).query(target_sequence).a3m,
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/alphafold3_venv/lib/python3.11/site-packages/alphafold3/data/tools/jackhmmer.py", line 129, in query
    output_sto_str = f.read()
                     ^^^^^^^^
  File "<frozen codecs>", line 322, in decode
MemoryError

Best regards, Tamas

Augustin-Zidek commented 1 week ago

Ah yes, sorry, this slipped my mind when working on AlphaFold 3. I will port this feature from AlphaFold 2.

Also, could you voice your support on the following issues:

https://github.com/EddyRivasLab/hmmer/issues/177 https://github.com/EddyRivasLab/hmmer/issues/323

If these two features existed in HMMER, the expensive Stockholm -> a3m conversion would not even be necessary.

Augustin-Zidek commented 6 days ago

First part of the fix landed in https://github.com/google-deepmind/alphafold3/commit/e60f750b6289129f5aed5e57a40ce9d3651b024e.

Augustin-Zidek commented 6 days ago

Fix finalized in https://github.com/google-deepmind/alphafold3/commit/cb2133d0fdaa66505051acc077f7b17617c2e608.

Thanks for reporting.

Could you let me know whether these fixes solved the issue?

hegelab commented 4 days ago

Thanks for the prompt action - it solved the issue.