ColabFold 1.6.3

ColabFold 1.6.3

Highlights: ColabFold can now be fully installed from pip. Turing/Volta generation GPUs are now much faster with --use-fast-kernels. ColabFold now computes and displays ipSAE/pDockQ2.

Bug fixes

• --use-pallas produced wrong results in 1.6.2 Head dimensions below 16 (the extra MSA stack) computed a wrong dot product on NVIDIA
  GPUs, and the key mask was read from the first row only, leaving template pointwise
  attention unmasked. (9b67373)
• Fix --msa-mode single_sequence using one additional randomly picked homolog per seed when the
  input is an a3m file or a cached MSA, instead of the query alone (8d4aefe)
• Fix colabfold_search leaving per-chain a3m stubs behind in paired mode (99a9868)
• Fix --initial-guess compiling the model a second time (9b67373)

New features and enhancements

• Fused kernels for Turing and Volta GPUs New --kernel-backend flag (auto (default),
  pallas, cuda_legacy) selects the fused Evoformer implementation (~2x faster on V100; 343b72d, 9b67373, 97ca370, 785b4fa)
• --use-fast-kernels replaces --use-pallas. --use-pallas keeps working (f7e9d49)
• ipSAE and pDockQ2 interface scores are reported for complexes (0c788a0) Thanks @Gyuuul2
• Lower Evoformer memory Transition subbatching is sized to a 256 MiB intermediate
  rather than 4 rows, and head-8 attention no longer pads q/k/v to 16 (9b67373)
• kalign, hhsearch and openmm from pip No conda needed for the local install
  (334057d, a2ce769)