Numba 0.61.0 broken on Google Colab (Feb 25)

[jacobtomlinson]

As of Feb 13th 2025 the latest numba release 0.61.0 is not compatible with the Google Colab GPU runtime.

Reproducer

• Open this example notebook from nvidia/accelerated-computing-hub
• In the menu select Runtime > Change Runtime Type and choose a GPU like a T4
• Run all cells

The first numba cell fails

import numpy as np
from numba import cuda

data = np.asarray(range(10))
output = np.zeros(len(data))

@cuda.jit
def foo(input_array, output_array):
    i = cuda.grid(1)
    output_array[i] = input_array[i]
    
foo[1, len(data)](data, output)

output

/usr/local/lib/python3.11/dist-packages/numba_cuda/numba/cuda/dispatcher.py:536: NumbaPerformanceWarning: Grid size 1 will likely result in GPU under-utilization due to low occupancy.
  warn(NumbaPerformanceWarning(msg))
ERROR:numba.cuda.cudadrv.driver:Call to cuLinkAddData results in CUDA_ERROR_UNSUPPORTED_PTX_VERSION

---------------------------------------------------------------------------

CudaAPIError                              Traceback (most recent call last)

[/usr/local/lib/python3.11/dist-packages/numba_cuda/numba/cuda/cudadrv/driver.py](https://localhost:8080/#) in add_ptx(self, ptx, name)
   2806         try:
-> 2807             driver.cuLinkAddData(self.handle, enums.CU_JIT_INPUT_PTX,
   2808                                  ptxbuf, len(ptx), namebuf, 0, None, None)

10 frames

CudaAPIError: [222] Call to cuLinkAddData results in CUDA_ERROR_UNSUPPORTED_PTX_VERSION


During handling of the above exception, another exception occurred:

LinkerError                               Traceback (most recent call last)

[/usr/local/lib/python3.11/dist-packages/numba_cuda/numba/cuda/cudadrv/driver.py](https://localhost:8080/#) in add_ptx(self, ptx, name)
   2808                                  ptxbuf, len(ptx), namebuf, 0, None, None)
   2809         except CudaAPIError as e:
-> 2810             raise LinkerError("%s\n%s" % (e, self.error_log))
   2811 
   2812     def add_file(self, path, kind):

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'

[jacobtomlinson]

There is a workaround from googlecolab/colabtools#5081 where you upgrade the numba-cuda package and enable JIT link for compatibility.

!uv pip install -q --system numba-cuda==0.4.0
from numba import config
config.CUDA_ENABLE_PYNVJITLINK = 1

[ncclementi]

cc: @taureandyernv for visibility.