Summary

• detect the loaded CUDA driver and NVRTC versions during CUDA server initialization
• advertise CUDA 12.8-only compiler and tensor-map capabilities only when both runtime components are at least 12.8
• reject 12.8-only tensor-map requests before cudarc attempts to resolve unavailable driver symbols

Why

The CUDA backend previously used cudarc's compile-time CUDA_VERSION as if it described the target machine. With CUDA 12.8 bindings, a machine running a CUDA 12.4 driver and NVRTC could therefore advertise and call 12.8-only functionality.

cudarc resolves dynamic driver symbols lazily, so retaining the 12.8 binding set is compatible with older runtimes as long as newer capabilities are gated using the actual loaded driver and NVRTC. This keeps the baseline path usable on CUDA 12.4 while preserving the complete hardware-supported feature set on CUDA 12.8 and newer environments.

Validation

• cargo fmt --all --check
• cargo test -p t4a-cubecl-cuda --test cudarc_feature_contract
• cargo test -p t4a-cubecl-cuda runtime::tests::cuda_12_8_capabilities_require_both_driver_and_nvrtc
• cargo test -p t4a-cubecl-cuda --no-run