Skip to content

Rescale the load balancing loss. #14

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
MasterJH5574 merged 1 commit into from
Apr 5, 2026
Merged

Rescale the load balancing loss. #14

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
5 changes: 4 additions & 1 deletion pithtrain/tasks/pretrain_language_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -422,12 +422,15 @@ def train_step(cfg: PretrainLanguageModelCfg, ctx: PretrainLanguageModelCtx) ->
torch.distributed.all_reduce(peak_gpu_mem, op=torch.distributed.ReduceOp.MAX)

# Collect the mean load balance loss (reduced across all ranks).
# The tracked values include the coefficient: lb_coef * E * dot(f, p).
# We divide it out so the logged metric is E * dot(f, p), where 1.0
# represents perfect balance (matches Megatron-LM convention).
moe_load_balance_coef = cfg.training.moe_load_balance_coef
lb_total, lb_count = MoELoadBalanceLossTracker.get_total_count_and_clear()
if moe_load_balance_coef > 0:
lb_stats = torch.tensor([lb_total, lb_count], device=device)
torch.distributed.all_reduce(lb_stats, op=torch.distributed.ReduceOp.SUM)
lb_loss = (lb_stats[0] / lb_stats[1]).item() if lb_stats[1] > 0 else 0.0
lb_loss = (lb_stats[0] / lb_stats[1]).item() / moe_load_balance_coef
else:
lb_loss = 0.0

Expand Down
Loading