Skip to content

[fix] dont use ep enabled variant of clip gradnorm #2144

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
Jackmin801 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

[fix] dont use ep enabled variant of clip gradnorm #2144

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
7e082d6
dont use ep enabled variant of clip gradnorm
Jackmin801 Mar 31, 2026
501b237
ruff
Jackmin801 Mar 31, 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
4 changes: 1 addition & 3 deletions src/prime_rl/trainer/sft/train.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -389,9 +389,7 @@ def run_validation(step: int) -> None:
nan_loss_count = nan_loss_count.item()

logger.debug(f"Clipping gradients with max norm {config.optim.max_norm}")
grad_norm = clip_grad_norm_(
model.parameters(), max_norm=config.optim.max_norm, ep_enabled=parallel_dims.ep_enabled
)
grad_norm = clip_grad_norm_(model.parameters(), max_norm=config.optim.max_norm)
Copy link
Copy Markdown

@cursor cursor bot Mar 31, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Incomplete fix: RL trainer still uses broken ep_enabled

Medium Severity

The ep_enabled=parallel_dims.ep_enabled parameter was removed from clip_grad_norm_ in the SFT trainer because it "doesn't actually work," but the identical call in src/prime_rl/trainer/rl/train.py still passes ep_enabled=parallel_dims.ep_enabled. If the EP-enabled variant of clip_grad_norm_ is broken, the RL trainer has the same problem and likely needs the same fix.

Fix in Cursor Fix in Web

Copy link
Copy Markdown
Collaborator

@S1ro1 S1ro1 Mar 31, 2026

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Imo this will be broken with standard EP, so it should be just en_enabled=use_deepep and the same should be added to rl? Though I'd swear it worked before with deepep

if grad_norm.device.type == "cpu":
grad_norm = grad_norm.to(torch.device("cuda"))
zero_grad_ratio = get_zero_gradient_ratio(model.parameters(), parallel_dims.dp_replicate)
Expand Down
Loading