Skip to content

fix(mtp): support multi-head MTP loss logging (mtp-num-layers > 1) #2132

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

Draft
ZiyiTsang wants to merge 1 commit into
base: main
Choose a base branch
from
+6 −4
Draft

fix(mtp): support multi-head MTP loss logging (mtp-num-layers > 1) #2132

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
10 changes: 6 additions & 4 deletions slime/backends/megatron_utils/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -854,15 +854,15 @@ def train(
torch.distributed.all_reduce(values, group=tracker.get("reduce_group"))
if tracker.get("avg_group") is not None:
torch.distributed.all_reduce(values, group=tracker["avg_group"], op=torch.distributed.ReduceOp.AVG)
# here we assume only one mtp layer
mtp_losses = (tracker["values"] * mtp_loss_scale).item()
# Multi-head MTP: tracker["values"] is [num_mtp_layers]; aggregate below.
mtp_losses = tracker["values"] * mtp_loss_scale
MTPLossLoggingHelper.clean_loss_in_tracker()

# CI check: verify MTP loss is within expected bounds
if args.ci_test:
from slime.backends.megatron_utils.ci_utils import check_mtp_loss

check_mtp_loss(mtp_losses)
check_mtp_loss(mtp_losses.sum().item())

# per train step log.
if (
Expand All @@ -879,7 +879,9 @@ def train(
}
log_dict[f"train/{role_tag}grad_norm"] = grad_norm
if args.enable_mtp_training:
log_dict[f"train/{role_tag}mtp_loss"] = mtp_losses
for _i in range(mtp_losses.shape[0]):
log_dict[f"train/{role_tag}mtp_{_i + 1}_loss"] = mtp_losses[_i].item()
log_dict[f"train/{role_tag}mtp_loss"] = mtp_losses.sum().item()

for param_group_id, param_group in enumerate(optimizer.param_groups):
log_dict[f"train/{role_tag}lr-pg_{param_group_id}"] = opt_param_scheduler.get_lr(param_group)
Expand Down
Loading