bonito models vs. guppy 4.5.2+

[Sumsarium]

I get this error when I run the (latest) bonito model (res_dna_r941_min_crf_v032) in guppy 4.5.2 and 4.5.3:

terminate called after throwing an instance of 'std::runtime_error'
what(): Could not allocate shared buffer for CUDA basecaller.
Aborted (core dumped)

It works fine in e.g. guppy 4.4.2 with RTX3070. Just wanted to inform.

[tmassingham-ont]

Thanks for the report, I'll pass it on to the Guppy developers.

[xlin-1]

Hi,
We got the same issue when running guppy 4.5.3 with bonito models, also run with RTX 3070.

[tmassingham-ont]

Thanks for the confirmation, the Guppy team are investigating.

[Runsheng]

Just for your info. I also encountered this issue.
Changing the "--gpu_runners_per_device" from 8 to 4 solved my issue.

[tmassingham-ont]

Thanks, that's quite an illuminating clue. The configuration of the models is set for GPUs in our platforms, 16GB whereas the RTX 3070 only has 8GB.

[GeoMicroSoares]

Hi there, having the same issue, having ran guppy_basecaller (4.5.4+66c1a7753) previously with --chunk_size 2000 --chunks_per_runner 768 --gpu_runners_per_device 8 on guppy model dna_r9.4.1_450bps_hac.cfg. Can't run bonito on res_dna_r941_min_crf_v031.cfg or res_dna_r941_min_crf_v032.cfg with --gpu_runners_per_device set to 8, 4 or 2. Running on a GeForce RTX 2080. Any ideas? Thanks in advance!

[lauralwd]

@GeoMicroSoares I have a similar GPU like yours, and I solved a similar issue by also lowering chunk_size and chunks per runner. I suggest starting with some low settings, and then start optimising on a small sample set you may have lying around.

~/bin/ont-guppy/bin/guppy_basecaller -i ~/foo/bar.fast5 -s ./test_sup1 --compress_fastq -x "cuda:0" -c ~/bin/ont-guppy/data/dna_r9.4.1_450bps_sup_prom.cfg --gpu_runners_per_device 2 --chunk_size 1000
ONT Guppy basecalling software version 5.0.11+2b6dbff
config file:        /home/laura/bin/ont-guppy/data/dna_r9.4.1_450bps_sup_prom.cfg
model file:         /home/laura/bin/ont-guppy/data/template_r9.4.1_450bps_sup_prom.jsn
input path:         /home/laura/lambda2_fast5
save path:          ./test_sup1
chunk size:         1000
chunks per runner:  256
minimum qscore:     10
records per file:   4000
fastq compression:  ON
num basecallers:    4
gpu device:         cuda:0
kernel path:        
runners per device: 2

Hope this helps.

[francoissabot]

I have the same issue with a Quadro RTX 4000, the HAC and Fast models work but not the SUP. Reducing the chunk size to 500 allows it to run however