j[dZddlZddlZddlmZ ddlmZejn#e e f$rdZYnwxYwej dZ dS)zNOptional ahead-of-time WGPKernels CUDA operations, separate from INT8 linears.N)dispatch)opscdtduo'|jdkotj|jdkS)a#The library is built for compute capability 8.0+, but its operations only beat the PyTorch / Triton paths on 12.x GPUs where they were measured (its conv3d is 0.7-1.3x slower than cuDNN on 8.9): a performance choice. Eligibility before launch; never hide kernel execution/ABI errors.Ncuda) r)implementationtyper capabilityindex)devices AC:\Users\johnn\OneDrive\Documents\AI\shared\kernels\native_ops.py availablers5  % p&+*? pHDWX^XdDeDeipDpp) __doc__ functoolstorchr wgpkernelsrr load_library ImportErrorOSErrorcacherrrrsTT 000000N!!!!WNNN qqqqqs ( 44