Google Colab With Cuda 12.4 and Numba gives - CUDA_ERROR_UNSUPPORTED_PTX_VERSION

[mnielsm]

Describe the current behavior
I am using NUMBA on the GPU and as of a couple of days it no longer works in google colab.

The following code worked fine before and now it doesn't any more.

from numba import cuda
import numpy as np

@cuda.jit
def increment_by_one(an_array):
    pos = cuda.grid(1)
    if pos < an_array.size:
        an_array[pos] += 1

an_array = np.zeros(10)
increment_by_one[16,16](an_array)

Describe the expected behavior
It shows an error about CUDA_ERROR_UNSUPPORTED_PTX_VERSION which seems to come from the CUDA driver.

LinkerError: [222] Call to cuLinkAddData results in CUDA_ERROR_UNSUPPORTED_PTX_VERSION
ptxas application ptx input, line 9; fatal   : Unsupported .version 8.5; current version is '8.4'

What web browser you are using
Doesn't matter for this bug

Additional context
Crossposted to the NVIDIA NUMBA CUDA github as bug (link)

[mnielsm]

When looking a bit futher I expect the issue to be that the Nvidia T4 Driver (550.54.15) only supports CUDA 12.4 as far as I can find on the NVIDIA page (link)
And the current installed CUDA Runtime Version is 12.5.

After updating cuda in google colab to 12.8 the error changes slightly (from unsupported version 8.5 to 8.7 so it has to do something with the version of CUDA installed in google colab:

LinkerError: [222] Call to cuLinkAddData results in CUDA_ERROR_UNSUPPORTED_PTX_VERSION
ptxas application ptx input, line 9; fatal   : Unsupported .version 8.7; current version is '8.4'

[lizhieffe]

Short-term workaround:

In the colab, switch to the fallback version in the Command palette. It changes the CUDA runtime to 12.2

(I met the same issue today and this works for me)

[metrizable]

@mnielsm Thanks for filing the issue and thanks for using Colab.

If you'd like to use numba with CUDA, you'll want to install the latest version of the numba-cuda package, since the CUDA target built into Numba was deprecated and further development has been moved to the NVIDIA numba-cuda package.

!uv pip install -q --system numba-cuda==0.4.0

And to ensure that numba-cuda uses the JIT link for compatibility, specify the following:

from numba import config
config.CUDA_ENABLE_PYNVJITLINK = 1

Example:

Also, as aside, for small examples, you can hide the low occupancy warnings, if desired, by specifying:

config.CUDA_LOW_OCCUPANCY_WARNINGS = 0

We will consider pre-installing numba-cuda==0.4.0 in the runtime in the future.

[mnielsm]

@mnielsm Thanks for filing the issue and thanks for using Colab.

If you'd like to use numba with CUDA, you'll want to install the latest version of the numba-cuda package, since the CUDA target built into Numba was deprecated and further development has been moved to the NVIDIA numba-cuda package.

!uv pip install -q --system numba-cuda==0.4.0
And to ensure that numba-cuda uses the JIT link for compatibility, specify the following:

from numba import config
config.CUDA_ENABLE_PYNVJITLINK = 1
Example:

Also, as aside, for small examples, you can hide the low occupancy warnings, if desired, by specifying:

config.CUDA_LOW_OCCUPANCY_WARNINGS = 0
We will consider pre-installing numba-cuda==0.4.0 in the runtime in the future.

Well I was doubtful this would do it but it seems to work fine indeed with the numba-cuda package.

[metrizable]

Re: #5061, great to hear!

[oginito]

it worked for me also. many thanks!!

[HadasRavikovitch]

Hi,
I have the same problem but this commands didnt solve it for me :(
Is there any other way you know?
Maybe something different in my case?

[Yaroslav-Muravev]

Add '--force-reinstall' after the command. It helped me

[HadasRavikovitch]

It worked!
Thank you so much :)