cuda : only use native when supported by cmake (llama/10389)

diff --git a/src/ggml-cuda/CMakeLists.txt b/src/ggml-cuda/CMakeLists.txt
index e592f79896db2d7574cffd1abcce53425b33eb92..25858ece8ddd7774aab85b0e1025ae54ecb7d156 100644 (file)
--- a/src/ggml-cuda/CMakeLists.txt
+++ b/src/ggml-cuda/CMakeLists.txt
@@ -12,7 +12,7 @@ if (CUDAToolkit_FOUND)
         # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
         # 70     == V100, FP16 tensor cores
         # 75     == Turing, int8 tensor cores
-        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6")
+        if (GGML_NATIVE AND CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.6" AND CMAKE_VERSION VERSION_GREATER_EQUAL "3.24")
             set(CMAKE_CUDA_ARCHITECTURES "native")
         elseif(GGML_CUDA_F16 OR GGML_CUDA_DMMV_F16)
             set(CMAKE_CUDA_ARCHITECTURES "60;61;70;75")