Closed ndryden closed 10 months ago
With CUDA 12.2, there are now additional default conversions for __half (and bfloat), so we need to provide an explicit specialization of operator<< for them, except for bfloat16 on ROCm platforms, where operator<< is provided.
__half
operator<<
With CUDA 12.2, there are now additional default conversions for
__half(and bfloat), so we need to provide an explicit specialization ofoperator<<for them, except for bfloat16 on ROCm platforms, whereoperator<<is provided.