replace for GGML_NATIVE=OFF too

Author: Aman Gupta

ggml/src/ggml-cuda/CMakeLists.txt

@@ -45,23 +45,25 @@ if (CUDAToolkit_FOUND)
 
     enable_language(CUDA)
 
-    # Replace any 12x-real architectures with 12x{a}-real. FP4 ptx instructions are not available in just 12x
-    set(PROCESSED_ARCHITECTURES "")
     if (GGML_NATIVE)
-        foreach(ARCH ${CMAKE_CUDA_ARCHITECTURES_NATIVE})
-            if(ARCH MATCHES "^12[0-9af]+")
-                # Extract the base architecture number (e.g., "120" or "120f")
-                string(REGEX REPLACE "^(12[0-9]+).*$" "\\1" BASE_ARCH ${ARCH})
-                # Replace with 12xa-real
-                message(STATUS "Replacing ${ARCH} with ${BASE_ARCH}a-real")
-                list(APPEND PROCESSED_ARCHITECTURES "${BASE_ARCH}a-real")
-            else()
-                list(APPEND PROCESSED_ARCHITECTURES ${ARCH})
-            endif()
-        endforeach()
-        set(CMAKE_CUDA_ARCHITECTURES ${PROCESSED_ARCHITECTURES})
+        set(ARCH_LIST ${CMAKE_CUDA_ARCHITECTURES_NATIVE})
+    else()
+        set(ARCH_LIST ${CMAKE_CUDA_ARCHITECTURES})
     endif()
 
+    # Replace any 12x-real architectures with 12x{a}-real. FP4 ptx instructions are not available in just 12x
+    set(PROCESSED_ARCHITECTURES "")
+    foreach(ARCH ${ARCH_LIST})
+        if(ARCH MATCHES "^12[0-9]")
+            string(REGEX REPLACE "^(12[0-9]).*$" "\\1" BASE_ARCH ${ARCH})
+            message(STATUS "Replacing ${ARCH} with ${BASE_ARCH}a-real")
+            list(APPEND PROCESSED_ARCHITECTURES "${BASE_ARCH}a-real")
+        else()
+            list(APPEND PROCESSED_ARCHITECTURES ${ARCH})
+        endif()
+    endforeach()
+    set(CMAKE_CUDA_ARCHITECTURES ${PROCESSED_ARCHITECTURES})
+
     file(GLOB   GGML_HEADERS_CUDA "*.cuh")
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")