Skip to content

fix(trainer): skip controller-side CUDA sync in single-controller mode #1377

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
sitabulaixizawaluduo merged 2 commits into from
Jun 2, 2026
+13 −9
Merged

fix(trainer): skip controller-side CUDA sync in single-controller mode #1377

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6f44c9c
fix(trainer): skip controller-side CUDA sync in single-controller mode
Adiactive May 29, 2026
f960f90
fix(trainer): also skip controller-side barrier in single-controller …
Adiactive Jun 1, 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
22 changes: 13 additions & 9 deletions areal/trainer/rl_trainer.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1113,7 +1113,7 @@ def _save_hf(self, epoch: int, epoch_step: int, global_step: int):
name="critic",
)
# Async mode: synchronization handled by AsyncCheckpointManager
if not self.saver.is_async:
if not self.saver.is_async and not is_single_controller():
dist.barrier(group=self.actor.cpu_group)
Comment thread
Adiactive marked this conversation as resolved.
current_platform.synchronize()

Expand All @@ -1139,8 +1139,9 @@ def _save_recover_checkpoint(self, epoch: int, epoch_step: int, global_step: int
processor=self.processor,
)

dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()
if not is_single_controller():
dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()

def _evaluate_fn(
self,
Expand All @@ -1161,8 +1162,9 @@ def _evaluate_fn(
cnt += 1
self.eval_rollout.wait(cnt, timeout=None)

dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()
if not is_single_controller():
dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()

def _evaluate(
self,
Expand All @@ -1188,8 +1190,9 @@ def _evaluate(
epoch_step,
global_step,
)
dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()
if not is_single_controller():
dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()

def _export_and_commit_stats(self, epoch: int, epoch_step: int, global_step: int):
# Upload statistics to the logger (e.g., wandb)
Expand All @@ -1199,8 +1202,9 @@ def _export_and_commit_stats(self, epoch: int, epoch_step: int, global_step: int
stats.update(self.eval_rollout.export_stats())
self.stats_logger.commit(epoch, epoch_step, global_step, stats)

dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()
if not is_single_controller():
dist.barrier(group=self.actor.cpu_group)
current_platform.synchronize()

def _validate_cfg(self):
"""validate config for incompatible settings before weight initialization, to avoid wasted resources on spawning workers and loading models."""
Expand Down
Loading