diff --git a/benchmarks/cuda_bindings/runner/main.py b/benchmarks/cuda_bindings/runner/main.py
index 9c984c340d6..eb2bcdacaf0 100644
--- a/benchmarks/cuda_bindings/runner/main.py
+++ b/benchmarks/cuda_bindings/runner/main.py
@@ -232,11 +232,18 @@ def parse_args(argv: list[str], default_output: Path = DEFAULT_OUTPUT) -> tuple[
 
 def main(
     *,
-    bench_dir: Path = BENCH_DIR,
-    default_output: Path = DEFAULT_OUTPUT,
+    bench_dir: Path | None = None,
+    default_output: Path | None = None,
     module_name_prefix: str = DEFAULT_MODULE_NAME_PREFIX,
     bench_filter_env_var: str = DEFAULT_BENCH_FILTER_ENV_VAR,
 ) -> None:
+    # Resolve the defaults inside the call, for the same reason
+    # discover_benchmarks() does: a literal default would be bound at def-time
+    # and would ignore a later monkeypatch of the module-level constant.
+    if bench_dir is None:
+        bench_dir = BENCH_DIR
+    if default_output is None:
+        default_output = DEFAULT_OUTPUT
     parsed, remaining_argv = parse_args(sys.argv[1:], default_output=default_output)
 
     registry = discover_benchmarks(bench_dir=bench_dir, module_name_prefix=module_name_prefix)
diff --git a/benchmarks/cuda_bindings/tests/test_runner.py b/benchmarks/cuda_bindings/tests/test_runner.py
index 56d88444c9e..836653522a1 100644
--- a/benchmarks/cuda_bindings/tests/test_runner.py
+++ b/benchmarks/cuda_bindings/tests/test_runner.py
@@ -164,3 +164,24 @@ def test_bench_launch_initializes_on_first_use(monkeypatch):
 
     assert len(compile_calls) == 1
     assert len(launch_calls) == 2
+
+
+def test_main_honors_a_monkeypatched_bench_dir(monkeypatch, tmp_path, capsys):
+    """main() must resolve BENCH_DIR at call time, like discover_benchmarks() does.
+
+    A literal default would be bound at def-time and would silently ignore a
+    later patch of the module-level constant.
+    """
+    runner_main = load_runner_main(monkeypatch)
+
+    (tmp_path / "bench_patched.py").write_text(
+        "def bench_only_here(loops: int) -> float:\n    return loops + 0.5\n",
+        encoding="utf-8",
+    )
+    monkeypatch.setattr(runner_main, "BENCH_DIR", tmp_path)
+    runner_main._MODULE_CACHE.clear()
+    monkeypatch.setattr(sys, "argv", ["run_pyperf.py", "--list"])
+
+    runner_main.main()
+
+    assert capsys.readouterr().out.split() == ["patched.only_here"]