CentOS 7下NVIDIA驱动已安装但nvidia-smi无法调用问题求助

Hey there, let's work through your issues one by one:

a) Fixing the nvidia-smi: command not found error #

Since you've confirmed the NVIDIA kernel module is installed (via cat /proc/driver/nvidia/version), the most likely issue is that the directory containing nvidia-smi isn't in your system's PATH environment variable. Here's how to fix it:

1. Locate the nvidia-smi executable
   Run this command to find where the binary is stored (it might take a minute):

   find / -name nvidia-smi 2>/dev/null
   

   Common locations on CentOS 7 are /usr/bin/nvidia-smi or /usr/local/nvidia/bin/nvidia-smi.

2. Add the directory to your PATH

   • Temporary fix (applies only to your current session):
     If the path was /usr/local/nvidia/bin, run:

     export PATH=$PATH:/usr/local/nvidia/bin
     

     Test it by running nvidia-smi right after.
   • Permanent fix (persists across reboots):
     Edit your bash profile to add the path permanently:

     echo 'export PATH=$PATH:/usr/local/nvidia/bin' >> ~/.bashrc
     

     Then reload the profile:

     source ~/.bashrc
     

     Now nvidia-smi should work every time you log in.

If the find command doesn't return any results, it might mean the full NVIDIA driver package wasn't installed (only the kernel module). In that case, you'll need to reinstall the driver package that includes user-space tools like nvidia-smi.

b) Installing CUDA on CentOS 7 #

First, note the driver-CUDA compatibility: Your NVIDIA driver version is 390.30, which supports CUDA 9.0 (minimum required driver for CUDA 9.0 is 384.81) or older versions. CUDA 9.1 and newer require a driver version ≥390.46, which is higher than your current 390.30, so stick with CUDA 9.0 for compatibility.

Follow these steps:

1. Install required dependencies
   Ensure you have the necessary development tools and kernel headers:

   sudo yum install -y gcc gcc-c++ kernel-devel kernel-headers
   

   Your existing GCC 4.8.5 is compatible with CUDA 9.0, so no need to upgrade it.

2. Run the CUDA installer
   Grab the CentOS 7 runfile installer (avoid the RPM version if you want to skip driver installation, since you already have a working driver). Make it executable and run with sudo:

   chmod +x cuda_9.0.176_384.81_linux.run
   sudo sh cuda_9.0.176_384.81_linux.run
   

   When prompted:

   • Do you accept the EULA? Type accept
   • Install NVIDIA Accelerated Graphics Driver for Linux-x86_64 384.81? Select no (you already have a newer driver installed)
   • Install the CUDA 9.0 Toolkit? Select yes
   • Install the CUDA 9.0 Samples? Select yes (optional but useful for testing)
   • Choose your preferred installation paths (defaults are usually fine)

3. Configure CUDA environment variables
   Add CUDA's binary and library paths to your profile:

   echo 'export PATH=$PATH:/usr/local/cuda-9.0/bin' >> ~/.bashrc
   echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/usr/local/cuda-9.0/lib64' >> ~/.bashrc
   source ~/.bashrc
   

4. Verify the installation
   Check the CUDA compiler version:

   nvcc -V
   

   You can also run the device query sample to confirm CUDA recognizes your GPU:

   cd /usr/local/cuda-9.0/samples/1_Utilities/deviceQuery
   make
   ./deviceQuery
   

   If it outputs a Result = PASS, your CUDA installation is working correctly.

内容的提问来源于stack exchange，提问作者AruniRC