I spent a week trying to make CUDA and Nvidia GPU work properly. With basic testing tensorflow and torch detect and see GPU ok:
Код: Выделить всё
$ python -c "import tensorflow as tf; print(tf.config.list_physical_devices('GPU'))"
I tensorflow/stream_executor/cuda/cuda_gpu_executor.cc:975] successful NUMA node read from SysFS had negative value (-1), but there must be at least one NUMA node, so returning NUMA node zero
[PhysicalDevice(name='/physical_device:GPU:0', device_type='GPU')]Код: Выделить всё
$ python -c 'import torch; print(torch.cuda.is_available())'
TruePython 3.7 is no longer officially supported by TensorFlow and Pytorch, but TensoFlow versions from 2.5 to 2.11 seem to work with Python 3.7. with the highest supported Cuda 11.2 and Cudnn 8.1.
Without GPU hardware, TensorFlow 2.9.0 and torch 1.13.1 work fine in the program.
Затрахался с Nvidia Cuda and GPU installation and configuration. Internet says that one can setup all the Cuda and GPU software entirely within Nvidia docker. They say I will only have to install Nvidia videocard driver (which I already have installed) and everything else related to Cuda, Cudnn and GPU will be set and run within docker.
Is it too good to be true? Какие подводные камни?
In the docker, I will need at least two different versions of Cuda: 11.2 for TensorFlow and Spacy and Cuda 11.6 (or 11.8 ) for torch.
Will I be able to run it all in one docker?
