Skip to content

WIP: fix two DDP wandb logging bugs (duplicate runs + dropped eval metrics) #468

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
eugenevinitsky merged 1 commit into from
Jun 3, 2026
Merged

WIP: fix two DDP wandb logging bugs (duplicate runs + dropped eval metrics) #468

Changes from all commits
Commits
Show all changes
1 commit
Select commit
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
33 changes: 20 additions & 13 deletions pufferlib/pufferl.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -292,6 +292,7 @@ def __init__(self, config, vecenv, policy, logger=None):
self.vecenv = vecenv
self.epoch = 0
self.global_step = 0
self.agent_steps = 0
self.last_log_step = 0
self.last_log_time = time.time()
self.start_time = time.time()
Expand Down Expand Up @@ -548,7 +549,7 @@ def train(self):
policy=self.uncompiled_policy,
env_name=self.config["env"],
logger=self.logger,
global_step=self.global_step,
global_step=self.agent_steps,
)

return logs
Expand Down Expand Up @@ -814,6 +815,7 @@ def mean_and_log(self):

device = config["device"]
agent_steps = int(dist_sum(self.global_step, device))
self.agent_steps = agent_steps
logs = {
"SPS": dist_sum(self.sps, device),
"agent_steps": agent_steps,
Expand Down Expand Up @@ -1437,17 +1439,22 @@ def train(env_name, args=None, vecenv=None, policy=None, logger=None, early_stop
model.forward_eval = policy.forward_eval
policy = model.to(local_rank)

if args["neptune"]:
logger = NeptuneLogger(args)
elif args["wandb"]:
logger = WandbLogger(args)
elif args["tb"]:
date_time = datetime.now().strftime("%Y%m%d-%H%M%S")
experiment_dir = os.path.join(args["train"]["data_dir"], rf"{env_name}_" + date_time)
logger = TensorBoardLogger(
run_id=date_time,
experiment_dir=experiment_dir,
)
# Under DDP only rank 0 owns the run logger; other ranks keep logger=None,
# which PuffeRL wraps in a NoLogger. Without this gate every rank calls
# wandb.init()/NeptuneLogger and you get world_size duplicate runs.
is_rank0 = (not torch.distributed.is_initialized()) or torch.distributed.get_rank() == 0
if is_rank0:
if args["neptune"]:
logger = NeptuneLogger(args)
elif args["wandb"]:
logger = WandbLogger(args)
elif args["tb"]:
date_time = datetime.now().strftime("%Y%m%d-%H%M%S")
experiment_dir = os.path.join(args["train"]["data_dir"], rf"{env_name}_" + date_time)
logger = TensorBoardLogger(
run_id=date_time,
experiment_dir=experiment_dir,
)

train_config = dict(**args["train"], env=env_name, eval=args.get("eval", {}))
pufferl = PuffeRL(train_config, vecenv, policy, logger)
Expand Down Expand Up @@ -1546,7 +1553,7 @@ def train(env_name, args=None, vecenv=None, policy=None, logger=None, early_stop
policy=pufferl.uncompiled_policy,
env_name=pufferl.config["env"],
logger=pufferl.logger,
global_step=pufferl.global_step,
global_step=pufferl.agent_steps,
force=True,
)

Expand Down
Loading