ggml-cuda: fix blackwell native builds (#18361)

* ggml-cuda: fix blackwell native builds

Replace 12x in native architectures by 12xa

* replace for GGML_NATIVE=OFF too

* only replace for native

* remove 120f-virtual for default compilation

---------

Co-authored-by: Aman Gupta <aman>

ggml/src/ggml-cuda/CMakeLists.txt

@@ -35,16 +35,33 @@ if (CUDAToolkit_FOUND)
             if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "11.8")
                 list(APPEND CMAKE_CUDA_ARCHITECTURES 89-real)
             endif()
-
-            if (CUDAToolkit_VERSION VERSION_GREATER_EQUAL "12.8")
-                list(APPEND CMAKE_CUDA_ARCHITECTURES 120f-virtual)
-            endif()
         endif()
     endif()
     message(STATUS "Using CUDA architectures: ${CMAKE_CUDA_ARCHITECTURES}")
 
     enable_language(CUDA)
 
+    # Replace any 12x-real architectures with 12x{a}-real. FP4 ptx instructions are not available in just 12x
+    if (GGML_NATIVE)
+        set(PROCESSED_ARCHITECTURES "")
+        foreach(ARCH ${CMAKE_CUDA_ARCHITECTURES_NATIVE})
+            if(ARCH MATCHES "^12[0-9]$")
+                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})
+    else()
+        foreach(ARCH ${CMAKE_CUDA_ARCHITECTURES})
+            if(ARCH MATCHES "^12[0-9]$")
+                message(FATAL_ERROR "Compute capability ${ARCH} used, use ${ARCH}a or ${ARCH}f for Blackwell specific optimizations")
+            endif()
+        endforeach()
+    endif()
+
     file(GLOB   GGML_HEADERS_CUDA "*.cuh")
     list(APPEND GGML_HEADERS_CUDA "../../include/ggml-cuda.h")