Also always return false in fp16_mma_hardware_available when compiled for AMD and GGML_HIP_ROCWMMA_FATTN not enabled.

Author: hjc4869

ggml/src/ggml-cuda/common.cuh

@@ -237,8 +237,12 @@ static bool fp16_mma_available(const int cc) {
 
 // To be used for feature selection of external libraries, e.g. cuBLAS.
 static bool fp16_mma_hardware_available(const int cc) {
+#if defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
+    return false;
+#else
     return cc < GGML_CUDA_CC_OFFSET_AMD && cc >= GGML_CUDA_CC_VOLTA ||
            cc >= GGML_CUDA_CC_CDNA && cc < GGML_CUDA_CC_RDNA1 || cc >= GGML_CUDA_CC_RDNA3;
+#endif // defined(GGML_USE_HIP) && defined(__HIP_PLATFORM_AMD__) && !defined(GGML_HIP_ROCWMMA_FATTN)
 }
 
 // Volta technically had FP16 tensor cores but they work very differently compared to Turing and later.