Skip to content

[BugFix] [RL] remove shutdown_process_group/restart_process_group for RL #5433

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
Jiang-Jia-Jun merged 4 commits into from
Dec 9, 2025
Merged

[BugFix] [RL] remove shutdown_process_group/restart_process_group for RL #5433

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
816478f
[fix] remove shutdown_process_group/restart_process_group for RL
liyonghua0910 Dec 8, 2025
3cc6125
[chore] remove log
liyonghua0910 Dec 8, 2025
705464d
[chore] remove log
liyonghua0910 Dec 8, 2025
24b4efd
[chore] set log to debug level
liyonghua0910 Dec 8, 2025
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
20 changes: 10 additions & 10 deletions fastdeploy/rl/dynamic_weight_manager.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -68,11 +68,11 @@ def update_parameters(self, pid: int = 0) -> None:
paddle.device.cuda.empty_cache()

# step1 : restart paddle process group
if not self.first_load:
paddle.distributed.restart_process_group()
paddle.distributed.restart_process_group(self.parallel_config.tp_group)
if self.parallel_config.enable_expert_parallel:
paddle.distributed.restart_process_group(self.parallel_config.ep_group)
# if not self.first_load:
# paddle.distributed.restart_process_group()
# paddle.distributed.restart_process_group(self.parallel_config.tp_group)
# if self.parallel_config.enable_expert_parallel:
# paddle.distributed.restart_process_group(self.parallel_config.ep_group)

# step2 : recreat deepep buffer when enable expert parallel
if self.parallel_config.enable_expert_parallel and not self.first_load:
Expand Down Expand Up @@ -136,7 +136,7 @@ def clear_parameters(self, pid: int = 0) -> None:
# ep barrier
paddle.distributed.barrier(self.parallel_config.ep_group)
# shutdown ep group
paddle.distributed.shutdown_process_group(self.parallel_config.ep_group)
# paddle.distributed.shutdown_process_group(self.parallel_config.ep_group)

paddle.device.cuda.empty_cache()
# step2: release model weight
Expand All @@ -149,11 +149,11 @@ def clear_parameters(self, pid: int = 0) -> None:
if self.parallel_config.tensor_parallel_size > 1:
# tp barrier
paddle.distributed.barrier(self.parallel_config.tp_group)
paddle.distributed.shutdown_process_group(self.parallel_config.tp_group)
# paddle.distributed.shutdown_process_group(self.parallel_config.tp_group)
if self.parallel_config.enable_expert_parallel:
paddle.distributed.barrier(self.parallel_config.ep_group)
paddle.distributed.shutdown_process_group(self.parallel_config.ep_group)
paddle.distributed.shutdown_process_group()
# paddle.distributed.shutdown_process_group(self.parallel_config.ep_group)
# paddle.distributed.shutdown_process_group()
self._update_shared_status(pid, ModelWeightsStatus.CLEARED)

def _update_model_from_state(self, state_dict: Dict[str, paddle.Tensor], src_type: str):
Expand Down Expand Up @@ -257,7 +257,7 @@ def check_model_weights_status(model_weights_status, model_runner, pid):
"""
check model weights status
"""
logger.info(f"dynamic weight manager is check model weights status! {model_weights_status.value[0]}")
# logger.info(f"dynamic weight manager is check model weights status! {model_weights_status.value[0]}")
while (
model_weights_status.value[0] != ModelWeightsStatus.NORMAL
and model_weights_status.value[0] != ModelWeightsStatus.CLEARED
Expand Down
6 changes: 3 additions & 3 deletions fastdeploy/worker/worker_process.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -459,7 +459,7 @@ def event_loop_normal(self) -> None:
else:
paddle.distributed.barrier(self.parallel_config.tp_group)
if self.model_weights_signal[0] != ModelWeightsStatus.NORMAL:
logger.info(
logger.debug(
f"Rank: {self.local_rank} to update or clear parameters, signal is {self.model_weights_signal[0]}, [-1:clear, 1:update]"
)
from fastdeploy.rl.dynamic_weight_manager import (
Expand All @@ -473,10 +473,10 @@ def event_loop_normal(self) -> None:
self.worker.model_runner,
self.parallel_config.engine_worker_queue_port,
)
logger.info(f"current task queue data: {self.task_queue.num_tasks()}")
logger.debug(f"current task queue data: {self.task_queue.num_tasks()}")
self.task_queue.clear_data()
self.model_weights_signal[0] = ModelWeightsStatus.NORMAL
logger.info(f"Rank: {self.local_rank} has updated or cleared parameters.")
logger.debug(f"Rank: {self.local_rank} has updated or cleared parameters.")

if self.exist_task_signal.value[0] == ExistTaskStatus.EXIST or self.task_queue.read_finish_flag.get() == 1:
logger.info(f"Rank: {self.local_rank} Detected new requests.")
Expand Down
Loading