trunk/132535f990621819aa7950f6ba0c048d7f6a34e4: Fix CUDA driver linking for Inductor CUTLASS builds (#186538)
- PyTorch: 1329 events in the last 90 days
- PyTorch: 1314th Release in the last 90 days
- Previous: earlier the same day · trunk/1979db84b6ee9e51e4ea6c331756a1c8a4b789ca: Add guangyey/etaf to XPU merge rule (#196947)
What happened
Inductor's CUDA/CUTLASS compile flags linked the CUDA driver with -lcuda. That requires an unversioned libcuda.so linker symlink, but common NVIDIA driver installs and runtime container setups expose only the runtime soname libcuda.so.1. In those environments torch.compile can fail at link time with /usr/bin/ld: cannot find -lcuda even though the CUDA driver is installed and usable at runtime. Teach the CUDA compile…
Summary assembled by rule from the sources below