Skip to content

fix: allow zero grad norm in dtensor policies for consistency with Megatron #1618

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

Closed
smahdavi4 wants to merge 4 commits into from
Closed

fix: allow zero grad norm in dtensor policies for consistency with Megatron #1618

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
4 commits
Select commit Hold shift + click to select a range
6ceb345
allow zero grad norm for consistency with megatron
smahdavi4 Dec 9, 2025
36874c6
fix
smahdavi4 Jan 13, 2026
de24b4b
fix
smahdavi4 Jan 13, 2026
030e6e0
linting fix
smahdavi4 Jan 14, 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
3 changes: 3 additions & 0 deletions nemo_rl/models/policy/workers/dtensor_policy_worker.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -191,6 +191,9 @@ def __init__(
self.cpu_offload = self.cfg["dtensor_cfg"]["cpu_offload"]
self.offload_optimizer_for_logprob = self.cfg["offload_optimizer_for_logprob"]
self.max_grad_norm = self.cfg["max_grad_norm"]
# allow zero grad norm for consistency with megatron
if self.max_grad_norm == 0.0:
self.max_grad_norm = None

if self.cfg["precision"] == "float32":
self.dtype = torch.float32
Expand Down
3 changes: 3 additions & 0 deletions nemo_rl/models/policy/workers/dtensor_policy_worker_v2.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,6 +235,9 @@ def __init__(
self.cpu_offload = self.cfg["dtensor_cfg"]["cpu_offload"]
self.offload_optimizer_for_logprob = self.cfg["offload_optimizer_for_logprob"]
self.max_grad_norm = self.cfg["max_grad_norm"]
# allow zero grad norm for consistency with megatron
if self.max_grad_norm == 0.0:
self.max_grad_norm = None

try:
self.dtype = STRING_TO_DTYPE[self.cfg["precision"]]
Expand Down
Loading