CUDA: mix virt/real CUDA archs for GGML_NATIVE=OFF

Author: JohannesGaessler

ggml/src/ggml-cuda/CMakeLists.txt

@@ -12,12 +12,22 @@ if (CUDAToolkit_FOUND)
         # 61     == Pascal, __dp4a instruction (per-byte integer dot product)
         # 70     == V100, FP16 tensor cores
         # 75     == Turing, int8 tensor cores
+        # 80     == Ampere, asynchronous data loading, faster tensor core instructions
+        # 86     == RTX 3000
+        # 89     == RTX 4000
+        #
+        # XX-virtual == compile CUDA code as PTX, do JIT compilation to binary code on first run
+        # XX-real    == compile CUDA code as device code for this specific architecture
+        # no suffix  == compile as both PTX and device code
+        #
+        # The default behavior for a non-native is to build virtual architectures as needed to cover all features needed
+        #     for best performance and to also build real architectures for the most commonly used GPUs.
         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;80")
+            set(CMAKE_CUDA_ARCHITECTURES "60-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
         else()
-            set(CMAKE_CUDA_ARCHITECTURES "50;61;70;75;80")
+            set(CMAKE_CUDA_ARCHITECTURES "50-virtual;61-virtual;70-virtual;75-virtual;80-virtual;86-real;89-real")
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")