Skip to content

[Feat] Support immediate rollout recovery by export hf asynchronously #1966

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

Open
YanhuiDua wants to merge 3 commits into
base: main
Choose a base branch
from
Open

[Feat] Support immediate rollout recovery by export hf asynchronously #1966

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1c01bce
[Feat] Support immediate rollout recovery by export hf asynchronously
YanhuiDua Jul 17, 2026
270a6a9
move immediate worker recovery into health manager
YanhuiDua Jul 22, 2026
b0bd6d9
add recovery hf ut and delete useless code
YanhuiDua Jul 23, 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
504 changes: 504 additions & 0 deletions .dev_scripts/test_qwen35_vl_moe_async_hf_recovery_e2e.py
View file
Open in desktop

Large diffs are not rendered by default.

1 change: 1 addition & 0 deletions tests/rl/test_rl_colocate_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -123,6 +123,7 @@ def _make_trainer(self, agent_loop_manager, *, total_train_steps: int = 1, sync_
trainer._enable_evaluate = False
trainer._enable_initial_evaluate = False
trainer._evaluate_step = 1
trainer._hf_export_executor = None
trainer._cpu_resource_manager = None
trainer._train_worker_cfg = SimpleNamespace(pack_max_length=16)
trainer._meta = SimpleNamespace(
Expand Down
13 changes: 6 additions & 7 deletions tests/rl/test_rl_disaggregated_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -113,6 +113,9 @@ def _make_trainer(self, agent_loop_manager):
trainer._enable_initial_evaluate = False
trainer._evaluate_step = 1
trainer._debug_rollout = False
trainer._enable_immediate_recovery = False
trainer._hf_export_executor = None
trainer._pending_hf_export = None
trainer._display_all_workers_log = False
trainer._num_workers = 1.0
trainer._rollout_num_workers = 1.0
Expand Down Expand Up @@ -150,7 +153,7 @@ def _make_trainer(self, agent_loop_manager):
check_and_shutdown_inactive_workers=SimpleNamespace(
remote=MagicMock(return_value="rollout_inactive_workers_shutdown")
),
restart_inactive_workers=SimpleNamespace(remote=MagicMock(return_value="rollout_restarted")),
restart_inactive_workers=SimpleNamespace(remote=AsyncMock(return_value=None)),
pause_generation=SimpleNamespace(remote=MagicMock(return_value="pause")),
continue_generation=SimpleNamespace(remote=MagicMock(return_value="continue")),
onload_weights=SimpleNamespace(remote=MagicMock(return_value="onload_weights")),
Expand All @@ -160,7 +163,7 @@ def _make_trainer(self, agent_loop_manager):
return trainer

def _run_fit(self, trainer):
with patch("xtuner.v1.train.rl_trainer.asyncio_run", side_effect=asyncio.run):
with patch("xtuner.v1.train.rl_trainer.ray.get", side_effect=lambda obj, timeout=None: obj):
trainer.fit()

def _minimal_train_info(self, *, training_samples: int, training_tokens: int, benchmark_end_time_s: float = 108.0):
Expand Down Expand Up @@ -202,7 +205,6 @@ def test_fit_persists_checkpoint_for_completed_model_step(self):
)

with (
patch("xtuner.v1.train.rl_trainer.asyncio_run", side_effect=asyncio.run),
patch("xtuner.v1.train.rl_trainer.ray.get", side_effect=lambda obj, timeout=None: obj),
patch("xtuner.v1.train.rl_trainer.bind_train_rollout"),
):
Expand Down Expand Up @@ -263,10 +265,7 @@ def test_fit_rebinds_weight_update_with_rollout_update_address(self):
trainer = self._make_trainer(manager)
trainer._rollout_config = SimpleNamespace(weight_update_host="10.0.0.1", weight_update_port=23456)

with (
patch("xtuner.v1.train.rl_trainer.asyncio_run", side_effect=asyncio.run),
patch("xtuner.v1.train.rl_trainer.bind_train_rollout") as bind_train_rollout_mock,
):
with patch("xtuner.v1.train.rl_trainer.bind_train_rollout") as bind_train_rollout_mock:
trainer.fit()

bind_train_rollout_mock.assert_called_once_with(
Expand Down
Loading
Loading