Cannot step into CUDA kernel while debugging with VSCode on WSL2 Ubuntu 22.04

[lhbran316]

Hi,

I’m developing a CUDA application on Windows 11 using WSL2 with Ubuntu 22.04, and debugging with VS Code.

My current setup allows me to:

• Successfully run CUDA programs (nvcc and cuda-gdb are working),
• Launch and attach the debugger from VS Code,
• Set breakpoints in both host code and device (kernel) code.

However, although breakpoints can be set in the kernel code, when I try to step into a kernel function (e.g., at a cudaMalloc or kernel launch like myKernel<<<>>>), the debugger immediately skips over the kernel and does not enter or step through the device code. It just resumes execution on the host side.

I have already:

• Verified that /usr/local/cuda-12.6 is installed correctly in WSL2,
• Added cuda-gdb to my environment path,
• Enabled the registry key HKEY_LOCAL_MACHINE\SOFTWARE\NVIDIA Corporation\GPUDebugger\EnableInterface with DWORD 1 on the Windows side,
• Ensured the GPU is visible to WSL2 with nvidia-smi.

Still, stepping into the kernel doesn’t work.

How can I get CUDA-GDB (via VSCode or otherwise) to properly debug inside kernel code?

Thanks in advance!

[veraj]

Have you tried using cuda-gdb directly ？ Is it the same behavior ？
Have you tried other CUDA sample ? You can use GitHub - NVIDIA/cuda-samples: Samples for CUDA Developers which demonstrates features in CUDA Toolkit

If you still repro, please provide the src code.

[veraj]

CUDA 12.9 Update 1 is recently released. Can you try with it ? Thanks !

[veraj]

Any update with the latest release ?
Are you ssh remote to the WSL ?

[pyxsqbs1]

I have the same problem? How to solve it.

[veraj]

Sorry, this is a known issue for now. Please don’t use cuda-gdb by SSH into WSL.
This is supposed to be fixed in future release.