pub fn require_cuda_compute_library(stem: &str) -> Result<(), String>Expand description
Require that the platform loader can open the named CUDA compute library
(cublas, cusolver, cusparse) from the one selected userspace stack.
cudarc 0.19 attempts to lazy-load these via its own generated
panic_no_lib_found helpers the first time CudaBlas::new /
DnHandle::new / cuSPARSE handle creation is invoked. On a host that
has only the CUDA driver (e.g. large-scale workbench images expose
libcuda.so.1 but no cuBLAS at all), those calls panic out of the
PyO3 FFI boundary instead of returning a typed error.
GpuRuntime::probe() calls this for every compute library it depends on;
failure retains the exact stack-selection or loader error in the typed GPU
refusal and keeps cudarc’s panic completely off the call path.