Skip Hopper-only kernels in CI (#3184)

.github/actions/build-cuda-release/action.yml

@@ -20,7 +20,7 @@ runs:
       run: |
         pip install auditwheel build patchelf setuptools
         python setup.py clean --all
-        MLX_BUILD_STAGE=2 python -m build -w
+        MLX_DISABLE_SM90A_KERNELS=1 MLX_BUILD_STAGE=2 python -m build -w
 
         auditwheel repair dist/mlx_cuda*.whl \
           --plat manylinux_2_35_${{ inputs.arch }} \

mlx/backend/cuda/CMakeLists.txt

@@ -158,8 +158,10 @@ message(STATUS "CUDA architectures: ${MLX_CUDA_ARCHITECTURES}")
 set_target_properties(mlx PROPERTIES CUDA_ARCHITECTURES
                                      "${MLX_CUDA_ARCHITECTURES}")
 
-if(("90a" IN_LIST MLX_CUDA_ARCHITECTURES) OR ("90a-real" IN_LIST
-                                              MLX_CUDA_ARCHITECTURES))
+# Skip Hopper-only kernels when not building for sm90a.
+if(NOT DEFINED ENV{MLX_DISABLE_SM90A_KERNELS}
+   AND (("90a" IN_LIST MLX_CUDA_ARCHITECTURES) OR ("90a-real" IN_LIST
+                                                   MLX_CUDA_ARCHITECTURES)))
   target_compile_definitions(mlx PRIVATE MLX_CUDA_SM90A_ENABLED)
 endif()