Repository navigation
Allow user-provided annotation for kernels generated from Program andObjectCode #1281
Description
We have examples like this today
kernel = module.get_kernel("vectorAdd<float>")which does not tell us what args are expected on the device side, and so when launch(s, config, kernel, ...) is called we cannot do type checking for users.
Activity
cuda.core.launch()improvements #150leofang commentedon Nov 21, 2025
During the meeting the preliminary idea is to allow users to (optionally) register
similar to what cffi does today, ex:
cuda-python/cuda_core/examples/strided_memory_view_cpu.py
Line 98 in b347255
We still want to be able to keep the fast path that Python
intcan be interpreted as pointer addresses.leofang commentedon May 1, 2026
After last November a lot has changed. In particular, as I mentioned in the design meeting yesterday, there is actually a way to retrieve the function signature from CUBINs via driver APIs + CUDA C++ demangler. This is now implemented in cupy/cupy#9801. I hope to reuse/migrate the infrastructure I built there back to cuda-core.