how to use the gpu mode?

[WUD2018]

Hey developer,

I have added a RTX4090 (24G) in my workstation. Could you please let me know how to use the gpu to run PhaBOX? I remember there was an option in (maybe) the early versions, but now I used the 2.1.11, and could not find any parameters to use.

[KennthShang]

Hi @WUD2018 ,

If there is a GPU, the program will automatically detect and use the cards. You do not need to specify any other parameters.

Best,
Jiayu

[WUD2018]

Hi @WUD2018 ,

If there is a GPU, the program will automatically detect and use the cards. You do not need to specify any other parameters.

Best, Jiayu

Thanks for your reply. I just did a test run. The input is a metagenomic assembly of environmental samples (280Mb). I used the 'end_to_end' model and the log information is listed below. Could you please help me check which step it calls the usage of GPU?
It seems like the first step (PhaMer) and the last step (PhaTYP) both used cpu rather tan gpu.

PhaBOX2 is running with: 80 threads!
Running program: PhaMer (virus identification)
[1/7] filtering the length of contigs...
[2/7] calling genes with prodigal...
[3/7] running all-against-all alignment...
[4/7] converting sequences to sentences for language model...
[5/7] Predicting the viruses...
100%|███████████████████████████████████████████████████████████████████████████████| 17510/17510 [00:57<00:00, 303.95it/s]
[6/7] summarizing the results...
[7/7] writing the results...
Run time: 239.22 seconds

PhaMer finished! please check the results in Phabox2_output_folder/final_prediction
Running program: PhaGCN (taxonomy classification)
[1/8] reusing existing filtered contigs...
[2/8] reusing existing protein file...
[3/8] using existing all-against-all alignment results...
[4/8] generating phagcn networks...
[5/8] predicting the taxonomy...
100%|█████████████████████████████████████████████████████████████████████████████| 15739/15739 [00:00<00:00, 20172.98it/s]
[6/8] summarizing the results...
[7/8] generating genus level clusters...
[8/8] writing the results...
Run time: 160.37 seconds

PhaGCN finished! please check the results in Phabox2_output_folder/final_prediction
Running program: CHERRY (Host prediction)
[1/10] reusing existing filtered contigs...
[2/10] no MAGs provided, skipping MAG CRISPRs prediction...
[3/10] will continue running CHERRY with database bersion...
[4/10] predicting DB CRISPRs...
[5/10] reusing existing protein file...
[6/10] reusing all-against-all alignment from PhaGCN...
[7/10] generating cherry networks...
100%|█████████████████████████████████████████████████████████████████████████████| 15739/15739 [00:00<00:00, 15887.77it/s]
[8/10] predicting the host...
[9/10] summarizing the results...
[10/10] writing the results...
Run time: 179.28 seconds

Cherry finished! please check the results in Phabox2_output_folder/final_prediction
Running program: PhaTYP (Lifestyle prediction)
running with cpu
[1/7] reusing existing filtered contigs...
[2/7] reusing existing protein file...
[3/7] using existing all-against-all alignment results...
[4/7] converting sequences to sentences for language model...
[5/7] Predicting the lifestyle...
Map: 100%|█████████████████████████████████████████████████████████████████| 14159/14159 [00:00<00:00, 21317.75 examples/s]
[6/7] summarizing the results...
[7/7] writing the results...
PhaTYP finished! please check the results in Phabox2_output_folder/final_prediction
Summarized finished! please check the results in Phabox2_output_folder/final_prediction/final_prediction_summary.tsv

[KennthShang]

Then, it might be because you have not installed the CUDA on your system, or the CUDA version does not match the version of PyTorch. You can first check whether you have installed the CUDA that meets the requirements of your system. Then, reinstall PyTorch as recommended via: https://pytorch.org/get-started/locally/

BTW, since the program is reoptimized on CPU, the GPU will not help much. In my test, at most 10 minutes for a 1Gb file. The most time-consuming part is the feature extraction, which only depends on the CPUs/threads you can supply.

Best,
Jiayu