CUDA code compile error with clang: function template partial specialization is not allowed

[zhiweij1]

Branch/Tag/Commit

main

Docker Image Version

N/A

GPU name

N/A

CUDA Driver

N/A

Reproduced Steps

Smaller reproducer:


template<class T1, class T2>
struct A {
  template <class T3>
  void foo(T3 x);
};

template<typename T1, typename T2>
template<typename T3>
void A<T1, T2>::foo<T3>(T3 x) {

}

https://godbolt.org/z/xn7vre9db

The original code comes from

FasterTransformer/src/fastertransformer/kernels/cutlass_kernels/fpA_intB_gemm/fpA_intB_gemm_template.h

Line 426 in afdf9a9

void CutlassFpAIntBGemmRunner<T, WeightType>::dispatch_to_arch<EpilogueTag>(const T* A,

The syntax of definition of “foo” is incorrect according to https://en.cppreference.com/w/cpp/language/member_template

LLVM community think it is not a bug in clang: llvm/llvm-project#68772

I think FasterTransformer should fix the usage in the source code.