[Bug] Undefined Symbol cuDeviceGetName when loading libtvm_runtime.so

[petersktang]

python -c "import tilelang" failed

USE_CUDA=ON, USE_ROCM=ON, USE_LLVM=ON, TILELANG_USE_CUDA_STUBS=ON

• configure with above and pip install successfuly
• python -c "import tilelang"

Actual behavior

raceback (most recent call last):
  File "<string>", line 1, in <module>
    import tilelang
  File tilelang/tilelang/__init__.py", line 154, in <module>
    with _lazy_load_lib():
         ~~~~~~~~~~~~~~^^
  File ".conda/envs/tilelang/lib/python3.14/contextlib.py", line 141, in __enter__
    return next(self.gen)
  File "tilelang/tilelang/__init__.py", line 129, in _lazy_load_lib
    import torch  # preload torch to avoid dlopen errors
    ^^^^^^^^^^^^
  File ".conda/envs/tilelang/lib/python3.14/site-packages/torch/__init__.py", line 444, in <module>
    from torch._C import *  # noqa: F403
    ^^^^^^^^^^^^^^^^^^^^^^
ImportError: .conda/envs/tilelang/lib/python3.14/site-packages/torch/lib/libtorch_cuda.so: undefined symbol: ncclCommResume