build : pass all warning flags to nvcc via -Xcompiler

[cebtenzzre]

Ref #5543

Fixes #5542

[cebtenzzre]

Looks like PR #3952 broke the Makefile. I promise, this stuff all worked when I wrote it :)

[slaren]

This seems to fix it.

[bartowski1182]

For the record this made my builds stop working with this error:

35.96 /usr/include/c++/11/type_traits:79:52: error: redefinition of 'constexpr const _Tp std::integral_constant<_Tp, __v>::value'
35.96    79 |   template<typename _Tp, _Tp __v>
35.96       |                                                    ^                           
35.96 /usr/include/c++/11/type_traits:67:29: note: 'constexpr const _Tp value' previously declared here
35.96    67 |       static constexpr _Tp                  value = __v;
35.96       |                             ^~~~~
36.46 make: *** [Makefile:447: ggml-cuda.o] Error 1

Reverting this change made it go through, though I imagine it's more likely that something else is causing a true error and your change is making it be accurately reported?

[slaren]

I noticed that with make builds, NDEBUG is no longer defined in CUDA builds, even without LLAMA_DEBUG. I think that passing the preprocessor defines to nvcc with -Xcompiler doesn't work, and should be passed directly to nvcc instead. What would be the best way to do this? From what I understand, the preprocessor defined are already separately defined in CPPFLAGS, so this should be doable without major changes.

[cebtenzzre]

I think that passing the preprocessor defines to nvcc with -Xcompiler doesn't work, and should be passed directly to nvcc instead.

#5598