Config variables for insitutes to set GPU acceleration

[keiran-rowell-unsw]

Description of feature

Don't try to live detect in pipeline, a bit fiddly.
Assume institutional admins can set up for their cluster.

Just templating so that values can be passed to ext.args

[tlitfin]

Profile throughput of stream async loading vs all-in-memory. Part of bigger benchmarking exercise to profile performance on different hardware (consider cost v throughput).

[nbtm-sh]

Look into colabfoldsearch first. Should be set in the existing module; do not create a new module. Should be done with the ext.args variable. No need to make it call mmseqs externally

[keiran-rowell-unsw]

So because colabfold_batch in proteinfold always has 'process_gpu' label it makes sense to enable GPU accelerated mmseqs in colab_batch by default since a GPU is always request.

I agree it should be in a .config variable because while an efficient default, mmseqs-GPU and mmseqs are not functionally equivalent algorithms and we don't want to prevent the user from being able to tweak in the future

[nbtm-sh]

Opened draft RP in nf-core/proteinfold

nf-core#497

[nbtm-sh]

@keiran-rowell-unsw I'm moving more and more to the side of implementing this as as separate module. MMseqs GPU in colabfoldsearch requires GPU padded databases, so we will likely need to add another "prepare dbs" workflow to create the databases.

They could be created in the mmseqs alignment step itself, but this will likely negate any speedup by using the GPUs.

[keiran-rowell-unsw]

We definitely don't want to do it in the mmseqs alignment by itself.

Definitely pre-compute the databases in with makepaddeddb alongside the download databases, that's in alignment with the colabfold setup_databases.sh

[nbtm-sh]

Opened discussion in nf-core Slack if DBs should be made in the pipeline or provided by the administrator. Awaiting a response. It is my opinion that the admin should provide DBs as it will depend on the hardware.

[keiran-rowell-unsw]

I definitely think the admin should supply the DBs for exactly the hardware-dependent reason.

Issue #203 on proteinfold also proposes a downloading subworkflow, which would make this convenient