Question - who here is regularly dealing with GPUs...
# team-maintainers
j
Question - who here is regularly dealing with GPUs/GPU-using tools? It came up here: https://github.com/nf-core/configs/pull/876#issuecomment-2776745070 I assume(?) we are increasing the number of GPU related tools now, so we should come up with a standardised way of distinguishing these types of modules so
configs
can send them to suitable queues, right?
j
We are using them in proteinfold, but our tools could be either run with gpus or cpus so labelling the tool by default wonโ€™t work, we rely on the
use_gpu
parameter. For the
accelerator
directive we use this in
modules.config
and for make it work on the different container technologies we have this (for singularity/apptainer) and this for docker. Then as the configuration of the gpus differs on the different institutional HPC, we also take advantage of the
use_gpu
and have a respective pipeline config for each institution e.g. CRG (that by the way needs to be updated)
๐Ÿซ  2
j
So there isn't an easy way to standardise this?
shakefist 1
j
We could have a dynamic labelling of the modules with a closure depending o the
use_gpu
parameter:
Copy code
label { params.use_gpu ? 'process_gpu' : 'process_high' }
Then we may also want to have several gpu labels
process_gpu_low
like in the PR discussion
๐Ÿ‘ 1
j
OK so you vote more for option 1 then
j
Not sure after a second though maybe I will more for 2 but would be nice to know others ideas
๐Ÿ‘ 1
f
@James Fellows Yates I literally asked about this like two times before the hackathon ๐Ÿ™„ meow cry https://nfcore.slack.com/archives/C043UU89KKQ/p1742501846858559?thread_ts=1741960877.303939&cid=C043UU89KKQ
shakefist 2
e
Maybe we should set them at the pipeline/institute level for now. I'm guessing the number of people using it will be low, and a pattern might emerge after we use it for a while.
f
But how do I let people for my module know that they can use it with GPU and tag a GPU container? sadpanda