Checkpoint tests should fail, not skip, when required cuda.bindings symbols are missing #2149

The CUDA checkpoint test availability guard currently masks a class of cuda-bindings regressions as unsupported-environment skips.

This mattered for the CUDA 13.3.0 CUcheckpointRestoreArgs regression fixed by PR #2144. In that release, cuda.bindings.driver.CUcheckpointRestoreArgs was missing because a CUDA 13.3 header layout change caused generation to silently omit the restore-argument binding. cuda.core.checkpoint._get_driver() correctly detects this as a missing required binding attribute and raises:

CUDA checkpointing requires cuda.bindings with CUDA checkpoint API support. Missing: CUcheckpointRestoreArgs

However, the checkpoint tests use _checkpoint_available() as a broad skip gate. That helper catches any RuntimeError from checkpoint._get_driver() and returns False, so a missing required binding symbol is treated the same as a truly unsupported driver/platform:

• Skip gate:

  cuda-python/cuda_core/tests/test_checkpoint.py

  Lines 31 to 43 in cc50515

  	def _checkpoint_available():
  	"""Return True if the checkpoint API is usable on this system."""
  	try:
  	checkpoint._get_driver()
  	return True
  	except RuntimeError:
  	return False
  	needs_checkpoint = pytest.mark.skipif(
  	sys.platform != "linux" or not _checkpoint_available(),
  	reason="CUDA checkpoint API requires Linux and a supported driver/bindings",
  	)

• Checkpoint tests guarded by that skip marker:

  cuda-python/cuda_core/tests/test_checkpoint.py

  Lines 397 to 423 in cc50515

  	@needs_checkpoint
  	class TestCheckpointLifecycle:
  	def test_initial_state_is_running(self):
  	_run_checkpoint_scenario_or_skip("initial_state_is_running")
  	def test_restore_thread_id_is_positive(self):
  	_run_checkpoint_scenario_or_skip("restore_thread_id_is_positive")
  	def test_lock_unlock(self):
  	_run_checkpoint_scenario_or_skip("lock_unlock")
  	def test_lock_default_timeout(self):
  	"""lock() with the default timeout_ms=0 (no timeout)."""
  	_run_checkpoint_scenario_or_skip("lock_default_timeout")
  	def test_lock_with_timeout(self):
  	_run_checkpoint_scenario_or_skip("lock_with_timeout")
  	def test_full_cycle_no_migration(self):
  	"""lock -> checkpoint -> restore -> unlock, verify state at each step."""
  	_run_checkpoint_scenario_or_skip("full_cycle_no_migration")
  	# -- GPU migration (>= 2 same-chip GPUs, real driver) ---------------------
  	@needs_checkpoint

• Missing required binding detection in checkpoint._get_driver():

  cuda-python/cuda_core/cuda/core/checkpoint.py

  Lines 133 to 149 in cc50515

  	def _get_driver():
  	global _driver_capability_checked
  	if _driver_capability_checked:
  	return _driver
  	binding_ver = _binding_version()
  	if not _binding_version_supports_checkpoint(binding_ver):
  	raise RuntimeError(
  	"CUDA checkpointing requires cuda.bindings with CUDA checkpoint API support. "
  	f"Found cuda.bindings {'.'.join(str(part) for part in binding_ver[:3])}."
  	)
  	missing = [name for name in _REQUIRED_BINDING_ATTRS if not hasattr(_driver, name)]
  	if missing:
  	raise RuntimeError(
  	f"CUDA checkpointing requires cuda.bindings with CUDA checkpoint API support. Missing: {', '.join(missing)}"
  	)

As a result, the tests that would have exercised checkpoint restore were skipped instead of failing, even though the installed cuda-bindings version was expected to provide checkpoint API support.

Suggested fix:

• Keep skipping true unsupported environments, such as non-Linux platforms, drivers older than checkpoint API support, or hardware/driver combinations where runtime checkpoint scenarios cannot complete.
• Fail when checkpoint._get_driver() reports missing required cuda-bindings attributes for a cuda-bindings version that is expected to support checkpointing.
• Consider adding a lower-level cuda-bindings API completeness test that asserts the required checkpoint symbols are present when the binding version advertises checkpoint support. This would catch omissions before cuda-core runtime tests depend on them.

Related:

• PR Fix CUDA checkpoint restore args and generated API docs #2144 fixes the CUcheckpointRestoreArgs generation/layout issue.
• cybind MR 396 is the related generator-side follow-up.
• See also: nvbug 6097301 comment 34.