/home/michael/code/ScaleLLM/src/kernels/quantization/marlin/./memory.h:17:41: error: there are no arguments to ‘__cvta_generic_to_shared’ that depend on a template parameter, so a declaration of ‘__cvta_generic_to_shared’ must be available [-fpermissive]
[build]    17 |   uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
[build]       |                                         ^~~~~~~~~~~~~~~~~~~~~~~~
[build] /home/michael/code/ScaleLLM/src/kernels/quantization/marlin/./memory.h:17:41: note: (if you use ‘-fpermissive’, G++ will accept your code, but allowing the use of an undeclared name is deprecated)
[build] /home/michael/code/ScaleLLM/src/kernels/quantization/marlin/./memory.h: In function ‘void marlin::cp_async4_pred(void*, const void*, bool)’:
[build] /home/michael/code/ScaleLLM/src/kernels/quantization/marlin/./memory.h:50:41: error: ‘__cvta_generic_to_shared’ was not declared in this scope
[build]    50 |   uint32_t smem = static_cast<uint32_t>(__cvta_generic_to_shared(smem_ptr));
[build]       |                                         ^~~~~~~~~~~~~~~~~~~~~~~~