Skip to content

Cherry-pick: Fix mtp_num_layers and clip_qk issues (#2581, #2776) #3075

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
BestJuly merged 3 commits into from
Feb 5, 2026
Merged

Cherry-pick: Fix mtp_num_layers and clip_qk issues (#2581, #2776) #3075

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
3 commits
Select commit Hold shift + click to select a range
1490789
Fix: don't enter branch if mtp_num_layers == 0 (#2581)
rj42 Dec 9, 2025
62a1cb4
Fix clip_qk for virtual pipeline size > 1 (#2776)
juntaowww Jan 15, 2026
e66f596
Merge branch 'main' into lit/pick_pr2581_to_main
BestJuly Feb 5, 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
2 changes: 1 addition & 1 deletion megatron/core/models/gpt/gpt_model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -608,7 +608,7 @@ def _postprocess(
if not self.post_process:
return hidden_states

if self.config.mtp_num_layers is not None:
if self.config.mtp_num_layers:
hidden_states = process_mtp_loss(
hidden_states=hidden_states,
labels=labels,
Expand Down
5 changes: 5 additions & 0 deletions megatron/core/optimizer/qk_clip.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -22,6 +22,11 @@ def clip_qk(model, log_max_only=False) -> float:
for model_chunk in model:
for transformer_layer in model_chunk.module.module.decoder.layers:
if hasattr(transformer_layer.self_attention, 'clip_qk'):
if (
transformer_layer.self_attention.core_attention.current_max_attn_logits
is None
):
continue
torch.distributed.all_reduce(
transformer_layer.self_attention.core_attention.current_max_attn_logits,
op=torch.distributed.ReduceOp.MAX,
Expand Down
Loading