DeepGEMM already has a CMake path intended for IDE indexing, but clangd setup is currently not straightforward for CUDA headers.
In my local setup I needed:
- a dedicated
build-clangd directory with CMAKE_EXPORT_COMPILE_COMMANDS=ON
pybind11/Torch CMake prefixes passed explicitly.clangd CUDA flags for .cuh files, e.g. -x cuda, --cuda-gpu-arch=sm_100, --cuda-path=...- removal of some nvcc-only flags
- a few diagnostic suppressions for CUDA 13.x / SM100 false positives
Would the project be open to adding either:
- a short clangd setup section in the README, or
- a helper script such as
scripts/configure_clangd.sh, or
- a checked-in template
.clangd.example?
The goal would be reliable navigation/indexing for kernel development, not necessarily zero clang diagnostics for all CUDA intrinsics.
DeepGEMM already has a CMake path intended for IDE indexing, but clangd setup is currently not straightforward for CUDA headers.
In my local setup I needed:
build-clangddirectory withCMAKE_EXPORT_COMPILE_COMMANDS=ONpybind11/Torch CMake prefixes passed explicitly.clangdCUDA flags for.cuhfiles, e.g.-x cuda,--cuda-gpu-arch=sm_100,--cuda-path=...Would the project be open to adding either:
scripts/configure_clangd.sh, or.clangd.example?The goal would be reliable navigation/indexing for kernel development, not necessarily zero clang diagnostics for all CUDA intrinsics.