tlitfin-unsw
tlitfin-unsw
Looks to be available at the new download site: https://opendata.mmseqs.org/colabfold
I think the hosted database is missing the mapping file neccessary for pairing. I generated an ad hoc mapping file using sample-level ids (also creating dummy dmp files) which is...
OK, thanks Martin, that is good to know! Reproducing mapping based on contig ID is much easier 😄. My intution was that in the case of low paired depth there...
Better to use the official file from @martin-steinegger if it is available but I sent you the file I have been using via FileSender. I assigned each contig an arbitrary...
The memory bottleneck is the pairformer [transition](https://github.com/jwohlwend/boltz/blob/330cb969a8f01d04d259ec3a54189c283e1a8562/src/boltz/model/modules/trunk.py#L621) layer. Adding chunking to this operation (as implemented in #235) should enable inference for 1400 tokens with 24GB VRAM. I ran an octamer...
Closing this stale PR. I implemented these changes and added several other inference-only memory optimizations in [this](https://github.com/tlitfin/lmi4boltz) fork.
FYI, this will also be affected by #293 which can be closed by #289
This is because esmfold does not natively support a multimer mode and instead approximates multimers by predicting in monomer mode with a large gap in residue indices between chains. You...
Maybe a compressed archive is a better solution to preserve the original data format.