Skip to content

Fix non-scalar input amax in preprocess_linear_fusion for MoE export #1264

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

Open
AEON-7 wants to merge 2 commits into
base: main
Choose a base branch
from
Open

Fix non-scalar input amax in preprocess_linear_fusion for MoE export #1264

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6929ecb
Fix non-scalar input amax in preprocess_linear_fusion for MoE export
AEON-7 Apr 15, 2026
aebe752
Merge branch 'main' into aeon7/fix-nonscalar-input-amax-moe-export
AEON-7 Apr 24, 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
20 changes: 15 additions & 5 deletions modelopt/torch/export/quant_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1375,11 +1375,21 @@ def preprocess_linear_fusion(modules: list[torch.nn.Module], resmooth_only=False
return

if modules[0].input_quantizer.is_enabled and modules[0].input_quantizer.amax is not None:
assert modules[0].input_quantizer.amax.numel() == 1, (
"Only support scalar input quant amax"
)

input_amax = torch.max(torch.stack([module.input_quantizer.amax for module in modules]))
if modules[0].input_quantizer.amax.numel() == 1:
# Scalar amax (e.g. dense layers with per-tensor activation quant):
# unify via scalar max across the modules being fused.
input_amax = torch.max(
torch.stack([module.input_quantizer.amax for module in modules])
)
else:
# Non-scalar amax (e.g. NVFP4 per-channel input quantizer on
# per-expert-decomposed MoE). Modules being fused here share the
# same input tensor, so their per-channel amax vectors are
# identical by construction. Elementwise max is a no-op in that
# case and is the correct unification rule if they ever differ.
input_amax = torch.stack(
[module.input_quantizer.amax for module in modules]
).amax(dim=0)
for module in modules:
module.input_quantizer.amax = input_amax

Expand Down