Ubuntu 16.04下NVIDIA驱动与CUDA崩溃求助：重启卡死且恢复模式无法联网

Hey, let's work through this issue together— I’ve dealt with similar NVIDIA driver/CUDA boot conflicts before, so here’s an actionable plan even without network access in recovery mode:

This is the most likely fix, since botched driver/CUDA reinstalls often cause boot hangs. Here's how to do it:

• Boot into recovery mode: Restart your machine, hold Shift (for Ubuntu) or Esc (for most Debian-based distros) as it boots to bring up the GRUB menu. Select "Advanced options" then "Recovery mode", and finally choose "root" to get a root shell.
• First, remount your root filesystem as writable (it’s read-only by default):

  mount -o remount,rw /
  

• Uninstall the drivers/CUDA based on how you installed them:
  • If you used the official NVIDIA .run installer:

    # Replace "xxx.xx" with your actual driver version number
    ./NVIDIA-Linux-x86_64-xxx.xx.run --uninstall
    

  • If you used apt (from NVIDIA’s repo or distro repos):

    apt purge nvidia* cuda*
    apt autoremove && apt clean
    

• Reboot your machine with:

  reboot
  

Sometimes leftover driver files get stuck in the initramfs (initial RAM filesystem), causing persistent boot issues. Run these commands in the recovery root shell (after remounting as writable):

update-initramfs -u
update-grub
reboot

To avoid repeating this issue when you reinstall drivers/CUDA:

• Match driver & CUDA versions: Stick to NVIDIA’s official compatibility matrix to pick a driver version certified for your target CUDA version—don’t just grab the latest driver.
• Skip GUI components (if using SSH only): If you never interact with the machine via a graphical interface, deselect "NVIDIA X Server" and other GUI-related tools during installation (especially critical for .run installers).
• Use CUDA’s bundled driver (optional): When installing CUDA, choose the option to install the bundled driver instead of using a separate package—this ensures compatibility out of the box.
• Verify post-install: After installation, confirm everything works with these commands:

  nvidia-smi
  nvcc -V
  

内容的提问来源于stack exchange，提问作者knk