GPU with higher CUDA version can not be used within old pytorch

[sixvable]

Hi！ Thanks for this wonderful tools! 👍 However we use RTX3090 which was coupled with a newer CUDA version 11.4 so it cant be used with pytorch 1.7.1. Could you fix this problem? That would be really helpful.

sixvable

[dawnmy]

Hi sixvable,

Thank you for reporting this issue. I will update the package in pip to support newer version of PyTorch and CUDA. Will let you know when it is updated.

Best Zhi-Luo

[dawnmy]

@sixvable Are you able to install a newer version of pytorch in the environment with pip?

[sixvable]

@sixvable Are you able to install a newer version of pytorch in the environment with pip?

Nope.

ERROR: pip's dependency resolver does not currently take into account all the packages that are installed. This behaviour is the source of the following dependency conflicts.
ribodetector 0.2.3 requires torch==1.7.1, but you have torch 1.8.1 which is incompatible.

[dawnmy]

I updated the dependency versions in the repo but have't update it in pip. Could you install it with:

conda create -n ribodetector_0.2.4 python=3.8
conda activate ribodetector_0.2.4
git clone https://github.com/hzi-bifo/RiboDetector.git
cd RiboDetector
python setup.py install

And then test if it works for you.

[sixvable]

Traceback (most recent call last):
  File "/share/home/limu/.conda/envs/ribodetector/bin/ribodetector", line 33, in <module>
    sys.exit(load_entry_point('ribodetector==0.2.4', 'console_scripts', 'ribodetector')())
  File "/share/home/limu/.conda/envs/ribodetector/lib/python3.8/site-packages/ribodetector-0.2.4-py3.8.egg/ribodetector/detect.py", line 754, in main
  File "/share/home/limu/.conda/envs/ribodetector/lib/python3.8/site-packages/ribodetector-0.2.4-py3.8.egg/ribodetector/parse_config.py", line 38, in from_json
  File "/share/home/limu/.conda/envs/ribodetector/lib/python3.8/site-packages/ribodetector-0.2.4-py3.8.egg/ribodetector/utils/util.py", line 15, in read_json
  File "/share/home/limu/.conda/envs/ribodetector/lib/python3.8/pathlib.py", line 1200, in open
    return io.open(self, mode, buffering, encoding, errors, newline,
  File "/share/home/limu/.conda/envs/ribodetector/lib/python3.8/pathlib.py", line 1054, in _opener
    return self._accessor.open(self, flags, mode)
NotADirectoryError: [Errno 20] Not a directory: '/share/home/limu/.conda/envs/ribodetector/lib/python3.8/site-packages/ribodetector-0.2.4-py3.8.egg/ribodetector/config.json'

Not work 😕 . Am I wrong with some options?

ribodetector -t 20 -m 24 -d 3 -i rawdata/read_1.fq.gz rawdata/read_2.fq.gz -o reads.nonrrna.{1,2}.fq.gz -e norrna -l 150

[dawnmy]

I just did a test, python setup.py install could not install RiboDetector into a right path, pip install . works. Could you install it with

pip install .

instead? The command line you used to run ribodetector is correct.

[sixvable]

Test with CUDA 11.1 and Pytorch 1.8.2, works fine! Really thanks!

[sixvable]

Test with CUDA 11.1 and Pytorch 1.8.2, works fine! Really thanks!

But seems like it does not work with pytorch 1.10.2

[dawnmy]

Could you post the error message when using pytorch 1.10.2. I was able to run it with pytorch 1.10.2. The Python is v3.8.12 and CUDA v10.2. Below are the package versions of the env:

Package           Version
----------------- --------
biopython         1.79
flatbuffers       2.0
numpy             1.22.2
onnxruntime       1.10.0
pandas            1.4.1
pip               22.0.3
protobuf          4.0.0rc2
python-dateutil   2.8.2
pytz              2021.3
ribodetector      0.2.4
setuptools        60.9.3
six               1.16.0
torch             1.10.2
tqdm              4.63.0
typing_extensions 4.1.1
wheel             0.37.1

[sixvable]

It just said that the RTX3090 is not compatible with pytorch 1.10.2. I guess it is because of the CUDA version, not ribodetector itself.

[dawnmy]

Thank you for the clarification. I think the new version PyTorch (v1.10.2) supports only CUDA v10.2 and v11.3 (https://pytorch.org/get-started/locally/). Since you have a different version of CUDA installed. They will be incompatible.

You can install PyTorch v1.10.2 together with CUDA v11.3 as following:

pip3 install torch==1.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

You will have RiboDetector working with PyTorch v1.10.2

[dawnmy]

I think this issue has been solved by installing torch together with the compatible CUDA with:

pip3 install torch==1.10.2+cu113 -f https://download.pytorch.org/whl/cu113/torch_stable.html

So I will close this issue.