2. I encountered a build error that seems to be related to the library size (failed to convert GOTPCREL relocation against 'relink with --no-relax'). It seems to be a particular problem on Debian, perhaps because it is being built with debug symbols. As a workaround, I've added -DOPTIMAL=OFF to the rocSOLVER cmake arguments. When OPTIMAL is OFF, the build will omit all specialized kernels for small matrix sizes. This omission has no effect on the API or ABI of the library, since more general kernels can implement the same functionality, but it will affect the performance.
For ROCm 6.0, I plan to split the small matrix kernels out into separate code object files (like rocBLAS does for its Tensile kernels). Unlike rocBLAS, however, the rocSOLVER library will be fully functional without the optimized code objects, so they can be packaged separately. If you measure by binary size, something like 85% of the rocsolver library is just small matrix kernels, so that should end the rocsolver library size woes once and for all.