[QST] Replace binaries part of nvidia_cutlass_dsl package

Hi,

Is it possible to use latest llvm/mlir components to replace these shiped with nvidia_cutlass_dsl? I might want to do some modifications on the source code from llvm/mlir and use these built from source code.
Those binaries are listed as folllowing:

• nvidia_cutlass_dsl/lib/libmlir_c_runner_utils.so
• nvidia_cutlass_dsl/lib/libmlir_cuda_runtime.so
• nvidia_cutlass_dsl/lib/libmlir_float16_utils.so
• nvidia_cutlass_dsl/lib/libmlir_runner_utils.so
• nvidia_cutlass_dsl/python_packages/cutlass/_mlir/_mlir_libs/_mlir.cpython-313-x86_64-linux-gnu.so
• nvidia_cutlass_dsl/python_packages/cutlass/_mlir/_mlir_libs/_mlirDialectsGPU.cpython-313-x86_64-linux-gnu.so
• nvidia_cutlass_dsl/python_packages/cutlass/_mlir/_mlir_libs/_mlirDialectsLLVM.cpython-313-x86_64-linux-gnu.so
• nvidia_cutlass_dsl/python_packages/cutlass/_mlir/_mlir_libs/_mlirDialectsNVGPU.cpython-313-x86_64-linux-gnu.so
• nvidia_cutlass_dsl/python_packages/cutlass/_mlir/_mlir_libs/_mlirExecutionEngine.cpython-313-x86_64-linux-gnu.so
• nvidia_cutlass_dsl/python_packages/cutlass/_mlir/_mlir_libs/_mlirGPUPasses.cpython-313-x86_64-linux-gnu.so

Thanks,
Peter Han

Start by examining how the package produces and loads the listed files under nvidia_cutlass_dsl/lib and python_packages/cutlass/_mlir/_mlir_libs. Trace the packaging entry points for these LLVM/MLIR binaries and determine what would be required to use components built from an external source tree. Done means a documented, working path exists for replacing the shipped components while keeping the package usable.