Multi-gpu processing on tch branch of Modkit

[adbeggs]

Hi team

Thanks for modkit. I'm using it to predict open-chromatin. The tch branch of it is indeed very fast on our HPC compared to the CPU only version - but the limitation is only one GPU can be specified instead of using all. For instance - I have a gpu node with 4 x A100 available presented as CUDA devices 0,1,2,3. It would be great if it could use all devices similar to dorado.

For others looking to set this up on a HPC to overcome a lot of GCC issues and torch and having to recompile I downloaded a suitable apptainer image by:

apptainer pull docker://nvidia/cuda:12.6.1-cudnn-runtime-ubuntu24.04

and then ran modkit within the apptainer image which solved a lot of the issues:

apptainer exec --nv --env LD_LIBRARY_PATH=/rds/homes/b/beggsa/beggsa-clinicalnanopore/software/modkit-gpu/libtorch/lib /rds/projects/b/beggsa-clinicalnanopore/software/modkit-gpu/cuda_12.6.1-cudnn-runtime-ubuntu24.04.sif bash -c '/rds/homes/b/beggsa/beggsa-clinicalnanopore/software/modkit-gpu/modkit open-chromatin predict germline_temp6mA.bam --model /rds/projects/b/beggsa-clinicalnanopore/software/modkit/models/r1041_e82_400bps_hac_v5.2.0\@v0.1.0/ --log modkit_predict.log -o ./accessible_regions.bedgraph --threshold 0.8 --device 0'

[ArtRand]

Hello @adbeggs Thanks for your message. Let me look into it.