feat: add set_shared_memory_carveout

[kaeun97]

This adds set_shared_memory_carveout using cuFuncSetAttribute requested here.

Do want to confirm with @gmarkall about this design (and would be nice to chat about it). An alternative is to combine this withcache_config in some way. However, I do think separating the two makes more sense mainly because it makes the intent clearer for the user (cache_config sets preference hints while set_shared_memory_carveout sets explicit percentages).

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[gmarkall]

/ok to test

[copy-pr-bot]

/ok to test

@gmarkall, there was an error processing your request: E1

See the following link for more information: https://docs.gha-runners.nvidia.com/cpr/e/1/

[gmarkall]

/ok to test 1697fc7

[gmarkall]

Thanks for the PR! I'll circle back to this when I get a chance (hopefully soon) and come back to you. Happy to chat about the design as well.

[gmarkall]

/ok to test 93a149e

[gmarkall]

Do want to confirm with @gmarkall about this design (and would be nice to chat about it).

I think the design for the part of the problem that this PR solves looks great - I can think of no improvement.

An alternative is to combine this withcache_config in some way. However, I do think separating the two makes more sense mainly because it makes the intent clearer for the user (cache_config sets preference hints while set_shared_memory_carveout sets explicit percentages).

I agree, keeping it separate makes sense to me too.

I'm happy to chat if you have unresolved questions, but I'm also happy for this to be merged - let me know how you'd like to proceed.

[kaeun97]

@gmarkall Thank you for the review! Good to hear that the design makes sense. Just wanted to make sure that I am interpreting the description from the issue properly :) Would be nice to re-run CI and merge this.

[greptile-apps]

Greptile Overview

Greptile Summary

This PR adds set_shared_memory_carveout functionality to configure the shared memory carveout percentage for CUDA kernels using cuFuncSetAttribute. The implementation:

• Adds abstract method to Function base class and concrete implementations in both CtypesFunction and CudaPythonFunction
• Validates carveout values are in the range -1 to 100 (where -1 means default)
• Includes comprehensive test coverage for valid/invalid values and kernel execution

Key Issue Found:

• The CtypesFunction implementation calls driver.cuFuncSetAttribute(), but this function is not defined in API_PROTOTYPES in drvapi.py. While the default code path uses CudaPythonFunction (which accesses the API via cuda.bindings), the CtypesFunction path will fail if ever used. This should be addressed by either removing the dormant ctypes code path or adding the missing API prototype.

Confidence Score: 3/5

• This PR is generally safe to merge with one critical issue that needs attention
• Score reflects that while the main implementation path (CudaPythonFunction) appears correct and well-tested, there's a logical error in the CtypesFunction implementation that calls an undefined API function. This would cause runtime failures if the ctypes code path is ever used. The default code path uses CudaPythonFunction, so this may not impact most users immediately, but it's a correctness issue that should be addressed.
• Pay special attention to numba_cuda/numba/cuda/cudadrv/driver.py - the CtypesFunction implementation references an undefined API function

Important Files Changed

File Analysis

Filename	Score	Overview
numba_cuda/numba/cuda/cudadrv/driver.py	3/5	Added set_shared_memory_carveout method to Function classes with validation, but CtypesFunction implementation references undefined API function
numba_cuda/numba/cuda/tests/cudadrv/test_cuda_driver.py	5/5	Comprehensive test added covering valid/invalid carveout values and kernel execution after configuration

[greptile-apps]

Additional Comments (1)

1. numba_cuda/numba/cuda/cudadrv/driver.py, line 2400 (link)

   logic: cuFuncSetAttribute is not defined in API_PROTOTYPES in drvapi.py. If CtypesFunction is ever used, this will fail with CudaDriverError: Driver missing function: cuFuncSetAttribute. While the default code path uses CudaPythonFunction, the CtypesFunction path should either be removed or have the API prototype added to drvapi.py:

_{2 files reviewed, 1 comment}

_{Edit Code Review Agent Settings | Greptile}

[gmarkall]

Additional Comments (1)

1. `numba_cuda/numba/cuda/cudadrv/driver.py`, line 2400 ([link](/nvidia/numba-cuda/blob/683e774faebe7254a870818c648c556f24554cf5/numba_cuda/numba/cuda/cudadrv/driver.py#L2400))
   **logic:** `cuFuncSetAttribute` is not defined in `API_PROTOTYPES` in `drvapi.py`. If `CtypesFunction` is ever used, this will fail with `CudaDriverError: Driver missing function: cuFuncSetAttribute`. While the default code path uses `CudaPythonFunction`, the `CtypesFunction` path should either be removed or have the API prototype added to `drvapi.py`:

2 files reviewed, 1 comment

Edit Code Review Agent Settings | Greptile

This is dead code, so it won't be hit. We need to remove it at some point.

[gmarkall]

/ok to test 683e774

[gmarkall]

Many thanks @kaeun97!