Skip to content

check_perlayer_mse: skip empty-input hook calls (Qwen3 MoE) #2

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
heroarmor merged 1 commit into from
May 17, 2026
Merged

check_perlayer_mse: skip empty-input hook calls (Qwen3 MoE) #2

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
7 changes: 7 additions & 0 deletions model_qwen4b/check_perlayer_mse.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -40,6 +40,13 @@ def pre(mod, args):

def post(mod, args, output):
x = captured.pop("x")
# Qwen3 MoE: experts that receive 0 routed tokens in this step get
# an empty input. F.mse_loss handles numel()==0 cleanly, but
# `.abs().max()` and `.abs().mean()` blow up on empty tensors
# ("Expected reduction dim to be specified for input.numel() == 0").
# Skip — there is no SC error to record for an unused expert call.
if x.numel() == 0:
return
with torch.no_grad():
ref = F.linear(x.to(weight.dtype), weight, bias).float()
out = output.float()
Expand Down