build: fix CI CUDA build (#620)

Author: giladgd

docs/blog/v3.19-gemma-4.md

@@ -1,6 +1,6 @@
 ---
 title: Gemma 4 is here!
-date: 2026-06-29T18:00:00Z
+date: 2026-06-30T00:00:00Z
 lastUpdated: false
 author:
     name: Gilad S.

llama/CMakeLists.txt

@@ -102,7 +102,8 @@ if (NLC_CI_MODE AND GGML_CUDA AND NOT DEFINED CMAKE_CUDA_ARCHITECTURES AND NOT N
             list(APPEND CMAKE_CUDA_ARCHITECTURES 121a-real)
         endif()
 
-        # Remove `90-virtual` due to low usage and because it inflates the binary size too much
+        # Remove these architectures due to low usage in CUDA 12 and because it inflates the binary size too much
+        list(REMOVE_ITEM CMAKE_CUDA_ARCHITECTURES 89-real)
         list(REMOVE_ITEM CMAKE_CUDA_ARCHITECTURES 90-virtual)
         message(STATUS "CUDAToolkit version: ${CUDAToolkit_VERSION}, setting CMAKE_CUDA_ARCHITECTURES to ${CMAKE_CUDA_ARCHITECTURES}")
     endif()