Right now, the binary release requires that users have specific CUDA versions installed. v11.8 works, v12 did not work for me without removing the uint8_t typedef in utils.h.ch.
Ideally, users downloading the binary release should not need a CUDA installation. I guess we could have a release build that ships the precompiled CUDA kernels? Would also improve startup times since compiling the kernels takes a few seconds.
@JolifantoBambla can we do that in CMAKE? Have three builds like this?
- Debug build with the workdir set to the project root folder so that CUDA source files in ./modules are hot-reloaded.
- Release build with the workdir set to the project root folder so that CUDA source files in ./modules are hot-reloaded.
- Release-Publish build that ships with compiled CUDA kernels.
Right now, the binary release requires that users have specific CUDA versions installed. v11.8 works, v12 did not work for me without removing the uint8_t typedef in utils.h.ch.
Ideally, users downloading the binary release should not need a CUDA installation. I guess we could have a release build that ships the precompiled CUDA kernels? Would also improve startup times since compiling the kernels takes a few seconds.
@JolifantoBambla can we do that in CMAKE? Have three builds like this?