Fixed missing code generation for older archs

CMakelists.txt

@@ -20,11 +20,15 @@ endif()
 set(CMAKE_CXX_FLAGS "${CMAKE_CXX_FLAGS} /fp:fast")
 
 if( MAKE_CUDA_BUILD ) 
+
+    set(COMPUTE_CAPABILITY_NOCUBLAST 50 52 60 61 70 72)
+    set(COMPUTE_CAPABILITY_KEPLER 35 37)
+    set(COMPUTE_CAPABILITY_ADA_HOPPER 89 90)
+
     # Later versions of CUDA support the new architectures
     set(CC_CUDA10x 75)
     set(CC_CUDA110 75 80)
     set(CC_CUDA11x 75 80 86)
-    set(CC_CUDA12x 89 90)
 
     option(NO_CUBLASLT "Don't use CUBLAST" OFF)
     set(CUDA_TARGET_ARCH_FEATURE_LEVEL "11.x" CACHE STRING 
@@ -45,15 +49,28 @@ if( MAKE_CUDA_BUILD )
 
     if(${CUDA_TARGET_ARCH_FEATURE_LEVEL} STREQUAL "11.0")
         set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA110})
+        if(NO_CUBLASLT)
+            list(APPEND CMAKE_CUDA_ARCHITECTURES ${COMPUTE_CAPABILITY_NOCUBLAST} ${COMPUTE_CAPABILITY_KEPLER})
+        endif()
+
     elseif(${CUDA_VERSION_TARGET_FEATURE_MAJOR} STREQUAL "11")
         set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA11x})
+        if(NO_CUBLASLT)
+            list(APPEND CMAKE_CUDA_ARCHITECTURES ${COMPUTE_CAPABILITY_NOCUBLAST} ${COMPUTE_CAPABILITY_KEPLER})
+        endif()
     elseif(${CUDA_VERSION_TARGET_FEATURE_MAJOR} STREQUAL "12")
-        set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA12x})
+        set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA11x})
+        if(NO_CUBLASLT)
+            list(APPEND CMAKE_CUDA_ARCHITECTURES ${COMPUTE_CAPABILITY_NOCUBLAST} ${COMPUTE_CAPABILITY_ADA_HOPPER})
+        endif()
     else()
         set(CMAKE_CUDA_ARCHITECTURES ${CC_CUDA10x})
+        list(APPEND CMAKE_CUDA_ARCHITECTURES ${COMPUTE_CAPABILITY_NOCUBLAST} ${COMPUTE_CAPABILITY_KEPLER})
         set(NO_CUBLASLT ON)
     endif()
 
+
+
     message(CONFIGURE_LOG " CUDA Targeting feature level ${CUDA_TARGET_ARCH_FEATURE_LEVEL}, with architectures ${CMAKE_CUDA_ARCHITECTURES}")
 
     set (LIBBITSANDBYTESNAME "libbitsandbytes_cuda${CUDA_VERSION_DLLNAME}")