diff --git a/python/cudf_polars/cudf_polars/__init__.py b/python/cudf_polars/cudf_polars/__init__.py
index 41d06f8631b..b6b7e59fc22 100644
--- a/python/cudf_polars/cudf_polars/__init__.py
+++ b/python/cudf_polars/cudf_polars/__init__.py
@@ -10,9 +10,31 @@
 
 from __future__ import annotations
 
-from cudf_polars._version import __git_commit__, __version__
-from cudf_polars.callback import execute_with_cudf
-from cudf_polars.dsl.translate import translate_ir
+
+# We want to avoid initialising the GPU on import. Unfortunately,
+# while we still depend on cudf, the default mode is to check things.
+# This ugly hack patches away the import-time calls in cudf that
+# require initialising the GPU.
+# TODO: Remove this when we only depend on a pylibcudf package.
+def _patch_away_import_gpu_init():
+    import sys
+    import types
+
+    fake_mod = types.ModuleType("_numba")
+    fake_mod.__dict__["_setup_numba"] = lambda: None
+    fake_mod.__dict__["_CUDFNumbaConfig"] = lambda: None
+    fake_mod.__dict__["_get_ptx_file"] = lambda: None
+    fake_mod.__dict__["validate_setup"] = lambda: None
+    sys.modules["cudf.utils._numba"] = fake_mod
+    sys.modules["cudf.utils.gpu_utils"] = fake_mod
+
+
+_patch_away_import_gpu_init()
+del _patch_away_import_gpu_init
+
+from cudf_polars._version import __git_commit__, __version__  # noqa: E402
+from cudf_polars.callback import execute_with_cudf  # noqa: E402
+from cudf_polars.dsl.translate import translate_ir  # noqa: E402
 
 __all__: list[str] = [
     "execute_with_cudf",