Skip to content

Fix issue #5242 grad_norm and loss is nan #7171

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
hwchen2017 merged 4 commits into from
Mar 29, 2025
Merged

Fix issue #5242 grad_norm and loss is nan #7171

Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
38952e6
Fix issue #5242 grad_norm and loss is nan
Glaceon-Hyy Mar 25, 2025
b6e78d9
Fix format
Glaceon-Hyy Mar 26, 2025
49f38a1
handle total_norm invalid value
Glaceon-Hyy Mar 26, 2025
409f9cc
Merge branch 'master' into fix_grad_norm
hwchen2017 Mar 27, 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
12 changes: 7 additions & 5 deletions deepspeed/runtime/zero/stage_1_and_2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1727,10 +1727,10 @@ def get_grad_norm_direct(self, gradients, params, norm_type=2):

norm_is_inf = total_norm.isinf()
norm_is_nan = total_norm.isnan()
inf_or_nan = norm_is_nan.logical_or(norm_is_inf)

err = torch.tensor(-1.0, device=self.device, dtype=torch.float)
total_norm = inf_or_nan * err + inf_or_nan.logical_not() * total_norm
if norm_is_inf or norm_is_nan:
total_norm = torch.tensor(-1.0, device=self.device, dtype=torch.float)

return total_norm

# creates a flat fused tensor from the tensor list starting at the first_offset
Expand Down Expand Up @@ -1987,8 +1987,10 @@ def unscale_and_clip_grads(self, grad_groups_flat, total_norm):
if self.clip_grad > 0.:
# norm is in fact norm*scale
clip = ((total_norm / self.loss_scale) + 1e-6) / self.clip_grad
clip = torch.clamp(clip, min=1.0)
combined_scale = clip * self.loss_scale

# handle total_norm invalid value -1
if clip > 1:
combined_scale = clip * self.loss_scale

for grad in grad_groups_flat:
if isinstance(grad, list):
Expand Down