CUDA: no FP16 arithmetic for vector FA kernel (#17558)

2025-11-28 10:29:09 +01:00
parent 35cf8887e1
commit 73955f7d2a
3 changed files with 24 additions and 20 deletions
@@ -55,11 +55,11 @@ static __device__ __forceinline__ float vec_dot_fattn_vec_KQ_f16(
        ggml_cuda_memcpy_1<sizeof(tmp)>(tmp, K_h2 + k_KQ_0 + (threadIdx.x % nthreads)*cpy_ne);
 #pragma unroll
        for (int k_KQ_1 = 0; k_KQ_1 < cpy_ne; ++k_KQ_1) {
-#ifdef FAST_FP16_AVAILABLE
+#ifdef V_DOT2_F32_F16_AVAILABLE
            ggml_cuda_mad(sum,                tmp[k_KQ_1] , ((const half2  *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
 #else
            ggml_cuda_mad(sum, __half22float2(tmp[k_KQ_1]), ((const float2 *) Q_v)[k_KQ_0/nthreads + k_KQ_1]);
-#endif // FP16_AVAILABLE
+#endif // V_DOT2_F32_F16_AVAILABLE
        }
    }