Skip to content

fix(moe): align EP expert weight dtype with activation dtype #1913

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
jQizhang wants to merge 2 commits into
base: main
Choose a base branch
from
+9 −0
Open

fix(moe): align EP expert weight dtype with activation dtype #1913

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a48ef45
fix(moe): align EP expert weight dtype with activation dtype (#1863)
jQizhang Apr 19, 2026
31f4157
Merge branch 'main' into fix/moe-expert-dtype-1863
jQizhang Apr 20, 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
9 changes: 9 additions & 0 deletions nemo_automodel/components/moe/experts.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -312,6 +312,11 @@ def forward(
else None
)

# Match activation dtype for grouped_mm; covers the case where FSDP2's
# MixedPrecisionPolicy does not reach EP DTensors.
gate_and_up_projs = gate_and_up_projs.to(x.dtype)
down_projs = down_projs.to(x.dtype)

# EP variable-length all-gather
if ep_size > 1:
ep_group = ep_mesh.get_group()
Expand Down Expand Up @@ -688,6 +693,10 @@ def forward(
gate_and_up_projs = self.gate_and_up_projs.to_local()
down_projs = self.down_projs.to_local()

# Match activation dtype for grouped_mm; see GroupedExperts.forward.
gate_and_up_projs = gate_and_up_projs.to(permuted_local_hidden_states.dtype)
down_projs = down_projs.to(permuted_local_hidden_states.dtype)

if torch.count_nonzero(tokens_per_expert) > 0:
if self.use_torch_mm:
tokens_per_expert_gpu = tokens_per_expert.to(
Expand Down
Loading