Installing cuda, tensorflow and miniconda on Pop!_OS 20.04

install_cuda_tensorflow_miniconda_Pop!Os20_04.md

Installing cuda, tensorflow and miniconda on Pop!_OS 20.04

Cuda and Tensorflow

• Install tensorflow: sudo apt install tensorflow-cuda-latest
• Install cuda: sudo apt install system76-cuda-latest

Miniconda

• Download latest miniconda installer from https://docs.conda.io/en/latest/miniconda.html
• Install it: bash Miniconda3-latest-Linux-x86_64.sh
• Create conda environment with tensorflow-gpu support:
  • v2: conda create --name ENV_NAME_HERE tensorflow-gpu
  • v1: conda create --name ENV_NAME_HERE tensorflow-gpu==1.15
• Activate new environment: conda activate ENV_NAME_HERE

Extra

Fix for error CUPTI_ERROR_INSUFFICIENT_PRIVILEGES

(Source: tensorflow/tensorflow#35860 (comment))

• Add options nvidia "NVreg_RestrictProfilingToAdminUsers=0" to /etc/modprobe.d/nvidia-kernel-common.conf
• Reboot

Fix for error ./bin/ptxas not found

For some reason tensorflow 2 tries to run ptxas based on a relative path, and not from the system path. So the workaround to fix this is to just create a symlink to it inside the project directory:

mkdir bin
ln -s `which ptxas` bin/

Fix for error successful NUMA node read from SysFS had negative value (-1)

(Sources: https://dmitry.ai/t/topic/50/2 and https://stackoverflow.com/questions/44232898/memoryerror-in-tensorflow-and-successful-numa-node-read-from-sysfs-had-negativ/44233285#44233285)

Add to /etc/crontab:

@reboot root for a in /sys/bus/pci/devices/*; do echo 0 | tee -a $a/numa_node; done > /dev/null