NVIDIA
About nvidia driver version and cuda version
1. Issue or feature description
My host nvidia driver version is 384.183. According to https://docs.nvidia.com/cuda/cuda-toolkit-release-notes/index.html, the cuda versin which I could use is 9.0. As a test, I run nvidia/cuda:10.1-devel-ubuntu18.04 image and get error. It make sense. But I found that in one container I created the cuda version is 10.1 by 'cat /usr/local/cuda/version.txt'. So why? That container's image is build by 'FROMpytorch/pytorch:1.6.0-cuda10.1-cudnn7-devel' in dockerfile.
Sorry, I don't quite understand your question. Are you asking why a container containing code written against CUDA 10.1 is able to run on your older 384.183 driver?
The answer is most likely CUDA forward compatibility, as described here: https://docs.nvidia.com/pdf/CUDA_Compatibility.pdf
Sorry, I don't quite understand your question. Are you asking why a container containing code written against CUDA 10.1 is able to run on your older 384.183 driver?
The answer is most likely CUDA forward compatibility, as described here: https://docs.nvidia.com/pdf/CUDA_Compatibility.pdf
Basically, Yes.It is my question that why a container containing code written against CUDA 10.1 is able to run on your older 384.183 driver.The driver version is too old to support CUDA 10.1.