jo>SrSSKJr SSKrSSKrSSKJr SrSSjrg)u Eagerly preload the CUDA 12 runtime shared libraries bundled by the `nvidia-*-cu12` wheels so that `onnxruntime-gpu`'s `CUDAExecutionProvider` can resolve them via `dlopen`. Why this exists --------------- `onnxruntime-gpu` (>=1.26.0) is built against CUDA 12. On hosts running CUDA 13 (driver 580+) the system loader cannot find `libcublasLt.so.12`, `libcudnn.so.9`, etc. — and `CUDAExecutionProvider` silently falls back to CPU. The `nvidia-*-cu12` pip wheels install the libraries under `/nvidia//lib/`, but the dynamic linker does not search those paths by default and the wheels do not register them. `ctypes.CDLL(path, RTLD_GLOBAL)` at import time makes the libraries' symbols available to subsequent `dlopen` calls from onnxruntime, without requiring callers to export `LD_LIBRARY_PATH`. This is best-effort: a CPU-only environment without the `nvidia-*-cu12` wheels installed will skip the preload silently. RapidOCR will then run on CPU as before — extraction still works, just slower. ) annotationsN)Path)zlibcublasLt.so.12zlibcublas.so.12zlibcurand.so.10zlibcufft.so.11zlibcudart.so.12z libcudnn.so.9c[[R5S- S[RRS[RR 3- S- nUS- nUR 5(d/$/n[HZnURU5HBn[R"[U5[RS9 URU5 MX M\ U$![a MYf=f)zPreload CUDA 12 sonames from the `nvidia-*-cu12` pip wheels. Returns the list of soname strings that were successfully loaded. Silent on missing wheels (CPU-only environments) and on individual `OSError`s during `dlopen`. libpython.z site-packagesnvidia)mode)rsysprefix version_infomajorminoris_dir_CUDA_LIB_SONAMESrglobctypesCDLLstr RTLD_GLOBALappendOSError) site_packages nvidia_rootloadedsonamepaths 6/home/mande/repo/quber/src/quber/utils/cuda_runtime.pypreload_cuda_runtime_libsr(s SZZ   3##))*!C,<,<,B,B+C D E    (*K      F#%%f-D  CIF,>,>? f% .$ M  s=C C+*C+)returnz list[str]) __doc__ __future__rrr pathlibrrrrr&s%.# r%