Defer PTX file load to runtime

python/cudf/cudf/core/udf/utils.py

@@ -1,5 +1,6 @@
 # Copyright (c) 2020-2023, NVIDIA CORPORATION.
 
+import functools
 import os
 from typing import Any, Callable, Dict
 
@@ -63,7 +64,10 @@
 precompiled: cachetools.LRUCache = cachetools.LRUCache(maxsize=32)
 launch_arg_getters: Dict[Any, Any] = {}
 
-_PTX_FILE = _get_ptx_file(os.path.dirname(__file__), "shim_")
+
+@functools.cache
+def _ptx_file():
+    return _get_ptx_file(os.path.dirname(__file__), "shim_")
 
 
 @_cudf_nvtx_annotate
@@ -286,7 +290,7 @@ def _get_kernel(kernel_string, globals_, sig, func):
     exec(kernel_string, globals_)
     _kernel = globals_["_kernel"]
     kernel = cuda.jit(
-        sig, link=[_PTX_FILE], extensions=[str_view_arg_handler]
+        sig, link=[_ptx_file()], extensions=[str_view_arg_handler]
     )(_kernel)
 
     return kernel

python/cudf/cudf/tests/test_no_cuinit.py

@@ -109,3 +109,15 @@ def test_cudf_create_series_cuinit(cuda_gdb):
     print(output.stderr)
     assert output.returncode == 0
     assert cuInit_called >= 0
+
+
+def test_cudf_import_no_device():
+    env = os.environ.copy()
+    env["CUDA_VISIBLE_DEVICES"] = "-1"
+    output = subprocess.run(
+        ["python", "-c", "import cudf"],
+        env=env,
+        capture_output=True,
+        text=True,
+    )
+    assert output.returncode == 0

python/cudf/cudf/tests/test_string_udfs.py

@@ -20,10 +20,12 @@
     string_view,
     udf_string,
 )
-from cudf.core.udf.utils import _PTX_FILE, _get_extensionty_size
+from cudf.core.udf.utils import _get_extensionty_size, _ptx_file
 from cudf.testing._utils import assert_eq, sv_to_udf_str
 from cudf.utils._numba import _CUDFNumbaConfig
 
+_PTX_FILE = _ptx_file()
+
 
 def get_kernels(func, dtype, size):
     """