Numerical Performance for CUDA Capability 7.x GPUs
Issue Description
When running AlphaFold 3 on CUDA Capability 7.x GPUs (such as the V100), the model produces poor quality predictions with:- Lots of clashing residues
- Ranking scores of -99 or lower
- Generally unusable output
Workaround
Set theXLA_FLAGS environment variable to include the following flag:
Affected Hardware
- NVIDIA V100 GPUs
- All other CUDA Capability 7.x GPUs
Incorrect Handling of Two-Letter Atoms in SMILES Ligands
This issue was fixed in commit 4e4023c.
Issue Description
Between commits f8df1c7 and 4e4023c, AlphaFold 3 handled incorrectly any two-letter atoms in ligands defined using SMILES strings.Affected Atoms
Two-letter atoms such as:- Chlorine (Cl)
- Bromine (Br)
- Other two-letter atomic symbols
Resolution
This issue has been resolved in the current version. If you are using an older version of AlphaFold 3, please update to the latest version or ensure you are using a commit after 4e4023c.Reporting New Issues
If you encounter an issue not listed here, please check the GitHub issues tracker first.
- Check if the issue already exists in the issues tracker
- If not found, create a new issue
- Provide detailed information about:
- Your hardware configuration
- AlphaFold 3 version
- Input JSON structure
- Error messages or unexpected behavior
- Steps to reproduce