rocWMMA 2.1.0 fails to compile with current MSVC STL: specializes std::is_arithmetic/is_floating_point/is_signed (N5014 [meta.rqmts]/4)

[Dixon-Cider]

Environment

ROCm 7.1.1 clang 21.0.0git + MSVC 14.51.36231 STL, Windows. (Compiles fine on Linux/libstdc++, which lacks the annotation.) rocWMMA 2.1.0.

Problem

library/include/rocwmma/internal/float8.hpp and rocwmma_xfloat32.hpp specialize standard <type_traits> templates for rocWMMA's fp8/xfloat32 types, e.g.:

template <> struct is_arithmetic<hip_fp8_e4m3>     : true_type {};
template <> struct is_floating_point<hip_fp8_e4m3> : true_type {};
template <> struct is_signed<hip_fp8_e4m3>         : true_type {};

The current MSVC STL marks these traits [[clang::no_specializations]] (per N5014 [meta.rqmts]/4, which forbids user specialization of them), so clang errors:

error: 'is_arithmetic' cannot be specialized: Specializing this standard library template is forbidden by N5014 [meta.rqmts]/4 [-Winvalid-specialization]

This makes rocWMMA fail to compile on Windows with a current MSVC toolset (e.g. downstream llama.cpp -DGGML_HIP_ROCWMMA_FATTN=ON).

Minimal repro

#include <type_traits>
struct hip_fp8_e4m3 {};
namespace std { template <> struct is_arithmetic<hip_fp8_e4m3> : true_type {}; }
int main() { return 0; }

clang++ -x c++ -std=gnu++17 -c repro.cpp (with MSVC 14.51 STL) → error. Adding -Wno-invalid-specialization compiles.

Suggested fix

Avoid specializing standard <type_traits> templates for rocWMMA types; use a rocWMMA-internal trait (e.g. rocwmma::is_arithmetic) instead.

Workaround for consumers: build with -Wno-invalid-specialization.

Also

The Windows ROCm 7.1 HIP SDK does not appear to ship rocWMMA at all — it must be vendored from this repo. A note in the docs about Windows + current-MSVC support would help.