Skip to content

Accept --fast_kernels for AlphaFold2 - #58

Merged
DimaMolod merged 2 commits into
mainfrom
exp/af2-fused-kernels
Oct 6, 2026
Merged

DimaMolod merged 2 commits into
mainfrom
exp/af2-fused-kernels

Conversation

@DimaMolod

Copy link
Copy Markdown
Collaborator

Summary

Accept AlphaPulldown's new AF2-only flag --fast_kernels (KosinskiLab/AlphaPulldown#645): ColabFold's fused Pallas kernels, about 2× faster AF2-Multimer on NVIDIA GPUs of compute capability 8.0+.

  • Flag table: the copied table in workflow/rules/common.smk gains _AF2_EXTRA_FLAGS = {"fast_kernels"}, for alphafold2 only, so the parse-time check doesn't warn about it. AF3 and AlphaLink still reject it.
  • README: the AlphaFold2 flag list shows it with a quoted value. YAML 1.1 (PyYAML) reads a bare on/off as a boolean; AlphaPulldown accepts true/false too, but quoting avoids the surprise.

Set it in structure_inference_arguments, e.g. --fast_kernels: "auto".

Also relevant to this workflow: AlphaPulldown#645 fixes padded AF2 predictions reporting scores over the padding. Every AF2-multimer batch with batch_size > 1 is padded to its largest fold, so it is affected.

Validation

  • CI-equivalent environment (Python 3.12, Snakemake 9.27): 158 passed, 2 skipped (the Slurm-plugin test), including the AF2/AF3 DAG dry-runs and a new test pinning the flag to alphafold2.

🤖 Generated with Claude Code

DimaMolod and others added 2 commits October 5, 2026 15:41
AlphaPulldown's new AF2-only flag (ColabFold's fused Pallas kernels for
AF2-Multimer, off|on|auto). The copied flag table gains it for alphafold2
only, so the parse-time check does not warn about it; the README flag list
shows it.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
@DimaMolod
DimaMolod merged commit 936c04c into main Oct 6, 2026
2 checks passed
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant