Skip to content

Refine rollout health manager lifecycle handling #1943

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
YanhuiDua merged 2 commits into from
Jul 8, 2026
+617 −414
Merged

Refine rollout health manager lifecycle handling #1943

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
2e7a4c0
Refine rollout health manager lifecycle handling
YanhuiDua Jul 6, 2026
4d0cbd0
fix comments
YanhuiDua Jul 8, 2026
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
173 changes: 145 additions & 28 deletions tests/rl/test_rollout_logic.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -631,6 +631,9 @@ def test_registry_filters_entrypoints_and_tracks_lifecycle(self):
self.assertEqual(unhealthy_groups[0].ranks, (0, 1))
self.assertEqual(tuple(worker.rank for worker in registry.inactive_workers()), (0, 1))
self.assertEqual(registry.active_entrypoints(), ())
inactive_groups = registry.inactive_worker_groups()
self.assertEqual(inactive_groups[0].ranks, (0, 1))
self.assertEqual(self._worker_by_rank(registry, 0).lifecycle_state, WorkerLifecycleState.INACTIVE)
claimed_groups = registry.claim_inactive_groups_for_recovery()
self.assertEqual(claimed_groups[0].ranks, (0, 1))
self.assertEqual(self._worker_by_rank(registry, 0).lifecycle_state, WorkerLifecycleState.RECOVERING)
Expand Down Expand Up @@ -1109,30 +1112,30 @@ def test_marks_worker_inactive_after_consecutive_health_failures(self):
worker_lifecycle_listeners=[listener],
)

manager._check_and_deactivate_failed_worker_groups()
manager.run_once()

self.assertTrue(self._worker_by_rank(registry, 0).is_active())
self.assertEqual(actor.check_health.calls, [()])
self.assertEqual(inactive_groups, [])

manager._check_and_deactivate_failed_worker_groups()
manager.run_once()

self.assertFalse(self._worker_by_rank(registry, 0).is_active())
self.assertEqual(self._worker_by_rank(registry, 0).lifecycle_state, WorkerLifecycleState.INACTIVE)
self.assertEqual(actor.check_health.calls, [(), ()])
self.assertEqual([group.ranks for group in inactive_groups], [(0,)])

def test_inactive_listener_runs_under_operation_lock(self):
def test_inactive_listener_runs_outside_lifecycle_operation_lock(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(False))
worker_info = WorkerSnapshot(rank=0, actor=actor, url="http://worker-0")
lock_acquired_by_listener = []
manager, _ = self._build_manager({0: worker_info}, failure_threshold=1)

def on_worker_group_inactive(group):
acquired = manager._operation_lock.acquire(blocking=False)
acquired = manager._lifecycle_operation_lock.acquire(blocking=False)
lock_acquired_by_listener.append(acquired)
if acquired:
manager._operation_lock.release()
manager._lifecycle_operation_lock.release()

manager._worker_lifecycle_listeners = (
SimpleNamespace(
Expand All @@ -1141,12 +1144,12 @@ def on_worker_group_inactive(group):
),
)

manager._check_and_deactivate_failed_worker_groups()
manager.run_once()

self.assertEqual(lock_acquired_by_listener, [False])
self.assertEqual(lock_acquired_by_listener, [True])

def test_inactive_worker_is_not_cleaned_up_again(self):
# 已 inactive 的 worker 不再重复健康检查。
# 已 inactive 的 worker 不再重复健康检查，也不再重复触发 inactive 通知。
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
workers_info = {
0: WorkerSnapshot(
Expand All @@ -1156,33 +1159,75 @@ def test_inactive_worker_is_not_cleaned_up_again(self):
lifecycle_state=WorkerLifecycleState.INACTIVE,
)
}
manager, _ = self._build_manager(workers_info)
inactive_groups = []
listener = SimpleNamespace(
on_worker_group_inactive=inactive_groups.append,
on_worker_group_recovered=MagicMock(),
)
manager, _ = self._build_manager(workers_info, worker_lifecycle_listeners=[listener])

checked_count = manager._check_and_deactivate_failed_worker_groups()
manager.run_once()

self.assertEqual(checked_count, 0)
self.assertEqual(actor.check_health.calls, [])
self.assertEqual(inactive_groups, [])

def test_health_check_threshold_zero_disables_periodic_health_check(self):
# threshold <= 0 表示关闭周期健康监测，不应把 active worker 直接判 inactive。
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(False))
worker_info = WorkerSnapshot(rank=0, actor=actor, url="http://worker-0")
manager, registry = self._build_manager({0: worker_info}, failure_threshold=0)

checked_count = manager._check_and_deactivate_failed_worker_groups()
with patch("xtuner.v1.rl.rollout.health_manager.threading.Thread") as thread_cls:
manager.start()

self.assertEqual(checked_count, 0)
thread_cls.assert_not_called()
self.assertIsNone(manager._thread)
self.assertTrue(self._worker_by_rank(registry, 0).is_active())
self.assertEqual(actor.check_health.calls, [])

manager.run_once()

self.assertTrue(self._worker_by_rank(registry, 0).is_active())
self.assertEqual(actor.check_health.calls, [])

def test_run_once_does_not_log_error_when_no_active_workers(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
workers_info = {
0: WorkerSnapshot(
rank=0,
actor=actor,
url="http://worker-0",
lifecycle_state=WorkerLifecycleState.INACTIVE,
)
}
manager, _ = self._build_manager(workers_info)

with patch("xtuner.v1.rl.rollout.health_manager.logger.error") as log_error:
manager.run_once()

log_error.assert_not_called()
self.assertEqual(actor.check_health.calls, [])

def test_run_once_does_not_log_error_when_last_active_worker_becomes_inactive(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(False))
worker_info = WorkerSnapshot(rank=0, actor=actor, url="http://worker-0")
manager, registry = self._build_manager({0: worker_info}, failure_threshold=1)

with patch("xtuner.v1.rl.rollout.health_manager.logger.error") as log_error:
manager.run_once()

log_error.assert_not_called()
self.assertFalse(self._worker_by_rank(registry, 0).is_active())
self.assertEqual(actor.check_health.calls, [()])

def test_fail_fast_health_check_still_runs_when_periodic_health_check_is_disabled(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(False))
worker_info = WorkerSnapshot(rank=0, actor=actor, url="http://worker-0")
manager, registry = self._build_manager({0: worker_info}, failure_threshold=0)

checked_count = manager._check_and_deactivate_failed_worker_groups(fail_fast=True)
with patch.object(manager, "_shutdown_worker_group", return_value=True):
manager.check_and_shutdown_inactive_workers()

self.assertEqual(checked_count, 1)
self.assertFalse(self._worker_by_rank(registry, 0).is_active())
self.assertEqual(actor.check_health.calls, [()])

Expand All @@ -1197,10 +1242,40 @@ async def fake_wait_for(awaitable, timeout):
return await awaitable

with patch("xtuner.v1.rl.rollout.health_manager.asyncio.wait_for", side_effect=fake_wait_for):
manager._check_and_deactivate_failed_worker_groups()
manager.run_once()

self.assertEqual(observed_timeouts, [2.5])

def test_wait_until_next_check_waits_for_resume_when_paused_during_interval(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
worker_info = WorkerSnapshot(rank=0, actor=actor, url="http://worker-0")
manager, _ = self._build_manager({0: worker_info})
manager._check_interval = 0.01
manager._pause_event = threading.Event()

class _FakeStopEvent:
def __init__(self):
self.wait_calls = []
self._paused_once = False

def is_set(self):
return False

def wait(self, timeout=None):
self.wait_calls.append(timeout)
if timeout == manager._check_interval and not self._paused_once:
self._paused_once = True
manager._pause_event.set()
elif timeout == 0.5:
manager._pause_event.clear()
return False

stop_event = _FakeStopEvent()
manager._stop_event = stop_event

self.assertTrue(manager._wait_until_next_check())
self.assertEqual(stop_event.wait_calls, [manager._check_interval, 0.5, manager._check_interval])

def test_shutdown_barrier_keeps_failed_shutdown_group_inactive(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
worker_info = WorkerSnapshot(
Expand All @@ -1211,17 +1286,10 @@ def test_shutdown_barrier_keeps_failed_shutdown_group_inactive(self):
)
manager, registry = self._build_manager({0: worker_info})

with (
patch.object(manager, "_shutdown_worker_group", return_value=False),
patch("xtuner.v1.rl.rollout.health_manager.logger.error") as log_error,
):
with patch.object(manager, "_shutdown_worker_group", return_value=False):
manager.check_and_shutdown_inactive_workers()

self.assertEqual(self._worker_by_rank(registry, 0).lifecycle_state, WorkerLifecycleState.INACTIVE)
self.assertTrue(
any("training can continue" in call.args[0] for call in log_error.call_args_list),
f"Expected shutdown failure log to explain why it is non-fatal, got: {log_error.call_args_list}",
)

def test_restart_barrier_keeps_failed_recovery_group_inactive(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
Expand Down Expand Up @@ -1271,6 +1339,55 @@ def test_restart_barrier_notifies_recovered_group_after_success(self):
self.assertEqual([group.ranks for group in recovered_groups], [(0,)])
self.assertTrue(all(worker.is_active() for worker in recovered_groups[0].workers))

def test_restart_barrier_cleans_claimed_groups_when_stopping(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
worker_info = WorkerSnapshot(
rank=0,
actor=actor,
url="http://worker-0",
lifecycle_state=WorkerLifecycleState.INACTIVE,
)
manager, registry = self._build_manager({0: worker_info})

def stop_after_restart(groups):
manager._stop_event.set()
return {group.ranks: False for group in groups}

with (
patch.object(manager, "_restart_worker_groups", side_effect=stop_after_restart),
patch.object(manager, "_shutdown_worker_group", return_value=True) as shutdown_group,
):
manager.restart_inactive_workers()

shutdown_group.assert_called_once()
self.assertEqual(shutdown_group.call_args.args[0].ranks, (0,))
self.assertEqual(shutdown_group.call_args.kwargs, {"wait_server_down": False})
self.assertEqual(self._worker_by_rank(registry, 0).lifecycle_state, WorkerLifecycleState.INACTIVE)

def test_shutdown_without_waiting_server_down_does_not_probe_worker_server(self):
actor = SimpleNamespace(
shutdown=_FakeAsyncRemoteMethod(None),
check_health=_FakeAsyncRemoteMethod(True),
)
worker_info = WorkerSnapshot(
rank=0,
actor=actor,
url="http://worker-0",
lifecycle_state=WorkerLifecycleState.INACTIVE,
)
manager, registry = self._build_manager({0: worker_info})
group = registry.claim_inactive_groups_for_recovery()[0]

def fake_ray_get(ref, timeout=None):
del timeout
return asyncio.run(ref)

with patch("xtuner.v1.rl.rollout.health_manager.ray.get", side_effect=fake_ray_get):
self.assertTrue(manager._shutdown_worker_group(group, wait_server_down=False))

self.assertEqual(actor.shutdown.calls, [()])
self.assertEqual(actor.check_health.calls, [])

def test_restart_worker_group_uses_reinit(self):
init_result = RolloutWorkerInitResult(
rank=0,
Expand Down Expand Up @@ -1313,7 +1430,7 @@ def fake_ray_get(refs, timeout=None):
self.assertEqual(actor.offload.calls, [()])
self.assertEqual(actor.restore_skip_load_weights.calls, [()])

def test_recovered_listener_runs_under_operation_lock(self):
def test_recovered_listener_runs_outside_lifecycle_operation_lock(self):
actor = SimpleNamespace(check_health=_FakeAsyncRemoteMethod(True))
worker_info = WorkerSnapshot(
rank=0,
Expand All @@ -1325,10 +1442,10 @@ def test_recovered_listener_runs_under_operation_lock(self):
manager, _ = self._build_manager({0: worker_info})

def on_worker_group_recovered(group):
acquired = manager._operation_lock.acquire(blocking=False)
acquired = manager._lifecycle_operation_lock.acquire(blocking=False)
lock_acquired_by_listener.append(acquired)
if acquired:
manager._operation_lock.release()
manager._lifecycle_operation_lock.release()

manager._worker_lifecycle_listeners = (
SimpleNamespace(
Expand All @@ -1340,7 +1457,7 @@ def on_worker_group_recovered(group):
with patch.object(manager, "_restart_worker_group", return_value=True):
manager.restart_inactive_workers()

self.assertEqual(lock_acquired_by_listener, [False])
self.assertEqual(lock_acquired_by_listener, [True])


class TestPartialRolloutHandler(unittest.IsolatedAsyncioTestCase):
Expand Down
Loading
Loading