Linux nvidia "GPU screens are not yet supported "

Graphics / Linux Linux
,
7

Hello,

Finally, I was able to install nvidia on my debian bullseye amd 64, through the nvidia-driver debian package (it also worked with the Nvidia installation script) and with and without secure boot.

1. Installing nvidia driver

With recent kernels, Secure boot impose to create / enroll a MOK key + sign the nvidia driver. This is explained in the Wiki Debian / Secure boot. Below, I provide two script I adapted / wrote to achieve the corresponding steps.

  1. (secure boot only) Create the following files:

/root/secure_boot/enroll.sh

#!/bin/sh
#https://wiki.debian.org/SecureBoot

echo "Creating MOK.priv and MOK.der"
openssl req -new -x509 -newkey rsa:2048 -keyout MOK.priv -outform DER -out MOK.der -days 36500 -subj "/CN=My Name/" -nodes

echo "Importing MOK.der, please enter your one-time password"
mokutil --import MOK.der 
echo "Check that your key is listed below"
mokutil --list-new 

echo "Now, reboot the machine. Then, enters MOK manager EFI utility"
echo "enroll MOK, continue, confirm, enter password, reboot>"
echo "Finally, then check it is loaded with dmesg | grep cert"
exit 0

/root/secure_boot/sign.sh

#!/bin/sh
PRIV=/root/secure_boot/MOK.priv
DER=/root/secure_boot/MOK.der

for filename in $PRIV $DER
do
  (test -f $filename && echo "$filename found :-)") || (echo "$filename not found" && exit 1)
done

KBUILD_VER=$(uname -r | cut -d"." -f1,2)
echo "Kbuild version $KBUILD_VER"
cd /lib/modules/$(uname -r)/updates/dkms
for ko in $(ls -1 *.ko)
do
  echo "Signing $ko"
  /usr/lib/linux-kbuild-$KBUILD_VER/scripts/sign-file sha256 $PRIV $DER $ko
done
exit 0
  1. (secure boot only) Create MOK keys (/root/secure_boot/MOK.der and /root/secure_boot/MOK.der) by running /root/secure_boot/enroll.sh then enroll they key. This is required only once.
  2. Install nvidia-driver:
apt update
apt install nvidia-driver
  1. (secure boot only) Sign the module by running /root/secure_boot/sign.sh. This is required whenever the kernel or the nvidia driver is updated.
  2. In my case, I also have to create /etc/X11/xorg.conf.d/nvidia.conf(see previous messages):
Section "ServerLayout"
  Identifier "layout"
  Option "AllowNVIDIAGPUScreens"
EndSection
  1. Reboot. Then, ensure that nvidia driver is loaded.
lsmod | grep nvidia
nvidia-smi

2. Benchmark CPU vs GPU

This section presents steps to run a python3 script demonstrating the gap between CPU and GPU.

  1. Install Nvidia driver (see previous section).
  2. Install torch & CUDA requirements:
apt update
apt install nvidia-cuda-dev nvidia-cuda-toolkit python3 python3-pip
pip3 install torch
  1. Create and run the following python script:
#!/usr/bin/env python3

import functools, time
import torch

def timer(f):
    """Print the runtime of the decorated function"""
    @functools.wraps(f)
    def wrapper_timer(*args, **kwargs):
        start_time = time.perf_counter()
        value = f(*args, **kwargs)
        end_time = time.perf_counter()
        run_time = end_time - start_time
        print(f"Finished {f.__name__!r} in {run_time:.4f} secs")
        return value
    return wrapper_timer

@timer
def cpu(x, y, num_times):
    for _ in range(num_times):
        z = torch.matmul(x, y)

@timer
def gpu(x, y, num_times):
    for _ in range(num_times):
        z = torch.matmul(x, y)

torch.cuda.init()
assert torch.cuda.is_available() # Fail if nvidia and CUDA dependencies are not installed 

X = torch.randn(10, 300, 400)
Y = torch.randn(10, 400, 500)
X_cuda = X.cuda()
Y_cuda = Y.cuda()
N = 1000
cpu(X, Y, N)
gpu(X_cuda, Y_cuda, N)

Results: (note that matrices have to be large enough to make GPU usage worth)

Finished 'cpu' in 5.3031 secs
Finished 'gpu' in 0.0213 secs

Once again, I’d like to thank to @generix and @leigh123linux for their help.

Best regards,
mando