Skip to content

fix: Fix fsdp1 grad clipping and log grad norm #251

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
terrykong merged 2 commits into from
Apr 24, 2025
Merged

fix: Fix fsdp1 grad clipping and log grad norm #251

Show file tree
Hide file tree
Changes from 1 commit
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
385649b
fix fsdp1 grad clipping and log grad norm
ashors1 Apr 23, 2025
80a8cae
fixes, add comment
ashors1 Apr 24, 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
1 change: 1 addition & 0 deletions nemo_reinforcer/algorithms/dpo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -482,6 +482,7 @@ def dpo_train(
losses = train_results["loss"]
metrics = {
"loss": train_results["loss"].numpy(),
"grad_norm": train_results["grad_norm"],
}
metrics.update(train_results["all_mb_metrics"])
for k, v in metrics.items():
Expand Down
1 change: 1 addition & 0 deletions nemo_reinforcer/algorithms/grpo.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -576,6 +576,7 @@ def grpo_train(
metrics = {
"loss": train_results["loss"].numpy(),
"reward": rewards.numpy(),
"grad_norm": train_results["grad_norm"],
}
metrics.update(train_results["all_mb_metrics"])
for k, v in metrics.items():
Expand Down
1 change: 1 addition & 0 deletions nemo_reinforcer/algorithms/sft.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -486,6 +486,7 @@ def sft_train(
losses = train_results["loss"]
metrics = {
"loss": train_results["loss"].numpy(),
"grad_norm": train_results["grad_norm"],
}
metrics.update(train_results["all_mb_metrics"])
for k, v in metrics.items():
Expand Down
13 changes: 10 additions & 3 deletions nemo_reinforcer/models/policy/fsdp1_policy_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -310,10 +310,16 @@ def train(
all_mb_metrics.append(loss_metrics)

# Clip gradients
grad_norm = None
if not eval_mode:
torch.nn.utils.clip_grad_norm_(
self.model.parameters(), max_norm=self.cfg["max_grad_norm"]
)
if isinstance(self.model, FullyShardedDataParallel):
grad_norm = self.model.clip_grad_norm_(
Comment thread
ashors1 marked this conversation as resolved.
max_norm=self.cfg["max_grad_norm"]
)
else:
grad_norm = torch.nn.utils.clip_grad_norm_(
self.model.parameters(), max_norm=self.cfg["max_grad_norm"]
)

# Update parameters
self.optimizer.step()
Expand All @@ -336,6 +342,7 @@ def train(
metrics = {
"global_loss": global_loss.cpu(),
"local_loss": local_loss.cpu(),
"grad_norm": grad_norm,
"rank": torch.distributed.get_rank(),
"all_mb_metrics": dict(mb_metrics),
}
Expand Down
6 changes: 4 additions & 2 deletions nemo_reinforcer/models/policy/hf_policy.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -177,8 +177,10 @@ def train(
results = self.worker_group.get_all_worker_results(futures)

# Aggregate the results
aggregated_results = {}
aggregated_results["loss"] = results[0]["global_loss"]
aggregated_results = {
"loss": results[0]["global_loss"],
"grad_norm": results[0]["grad_norm"],
}

# Aggregate metrics across all workers
all_mb_metrics = defaultdict(list)
Expand Down