[Python][GPU] numba-cuda dependency is maintenance mode only

Describe the enhancement requested

The numba-cuda project added a maintenance notice:

Numba-CUDA is in maintenance mode. Moving forward, we intend to support only security issues and critical bug fixes through the lifetime of CUDA 13. New feature development is targeted towards Numba-CUDA-MLIR, and we recommend upgrading to Numba-CUDA-MLIR as soon as practical. For migration guidance, see Migration from Numba / Numba-CUDA.

We should probably deprecate our usage of numba-cuda and migrate to numba-cuda-mlir. Not an expert on this area but using a deprecated maintenance only dependency doesn't sound great.

Component(s)

GPU, Python

@kkraus14

It's not really a dependency, though, its a third-party lib that we provide integration with :-)

So we can also provide an integration with the new thing if some contributor is interested.

@kkraus14 @leofang @tpn

Fair enough, should we remove the current third-party integration code? The CI jobs have been failing for months at this point.

@kkraus14

Perhaps NVidia can task someone to fix CI? The failing builds that needs to be fixed AFAIK is https://github.com/apache/arrow/actions/runs/34091957769/job/101647306765

Ping @kkraus14 @leofang Again :-)

@raulcd

@raulcd Note that the crashing code on the CUDA Python CI is not Numba-related.

(and it's inherently crashy, though it used to raise an error before for some reason: I think we could simply remove it)

arrow/python/pyarrow/tests/test_cuda.py

Lines 300 to 306 in e235bdd

    
           # test deallocating the host buffer memory making it inaccessible 
        
           hbuf = cuda.new_host_buffer(size * dtype.itemsize) 
        
           fbuf = ctx.foreign_buffer(hbuf.address, hbuf.size) 
        
           del hbuf 
        
           with pytest.raises(pa.ArrowIOError, 
        
                              match=('Cuda error ')): 
        
               fbuf.copy_to_host()

The error it used to raise (and that I still get locally) is Cuda error 1 in function 'cuMemcpyDtoH': [CUDA_ERROR_INVALID_VALUE] invalid argument. But the code is really trying to access an invalid pointer.

It originally started failing with:

[Python][C++][GPU] Python Cuda jobs fail with 'cuda.bindings.driver.CUcontext' object has no attribute 'value' #49437

It originally started failing with:

We're talking about different failing tests here :)

Q: Isn't this fix already merged? #51206

@leofang

@leofang Right, sorry, that was a subthread but this issue is still valid.

Got it. Now that the CI is fixed, is there anything else you need help with?

#49437 is not fixed (there's an attached PR that deserves reviewing and perhaps helping along).

Seems like it'd be fixed by #50923. CUDA 11 support has been dropped since numba-cuda v0.19.0 (released in Aug, 2025).

Navigation Menu

Repository navigation

[Python][GPU] numba-cuda dependency is maintenance mode only #51204

Describe the enhancement requested

Component(s)

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

	# test deallocating the host buffer memory making it inaccessible
	hbuf = cuda.new_host_buffer(size * dtype.itemsize)
	fbuf = ctx.foreign_buffer(hbuf.address, hbuf.size)
	del hbuf
	with pytest.raises(pa.ArrowIOError,
	match=('Cuda error ')):
	fbuf.copy_to_host()

[Python][GPU] numba-cuda dependency is maintenance mode only #51204

Description

Describe the enhancement requested

Component(s)

Activity

pitrou commented on Sep 7, 2026

raulcd commented on Sep 7, 2026

pitrou commented on Sep 7, 2026

pitrou commented on Sep 7, 2026

pitrou commented on Sep 7, 2026

raulcd commented on Sep 7, 2026

pitrou commented on Sep 7, 2026

leofang commented on Sep 8, 2026

pitrou commented on Sep 8, 2026

leofang commented on Sep 8, 2026

pitrou commented on Sep 8, 2026

leofang commented on Sep 8, 2026

Metadata

Metadata

Assignees

Labels

Type

Projects

Milestone

Relationships

Development

Issue actions