Skip to content

[Fix] model_loader: Neutral UE8M0 block scales for dummy weights #47337

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
merrymercy merged 1 commit into from
Oct 31, 2026
Merged

[Fix] model_loader: Neutral UE8M0 block scales for dummy weights #47337

Show file tree
Hide file tree
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
5 changes: 5 additions & 0 deletions python/sglang/srt/model_loader/weight_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -1756,6 +1756,11 @@ def initialize_dummy_weights(
the parameter's number of elements and its data type.
"""
for name, param in model.state_dict().items():
# UE8M0 MXFP8 block scales (uint8): 127 is 2**0. Left at zero, every
# MXFP8 weight scales to 2**-127.
if name.endswith("weight_scale_inv") and param.dtype == torch.uint8:
param.fill_(127)
continue
if torch.is_floating_point(param):
# Neutral FP4 global scales keep fused gate/up projections compatible.
if name.endswith(("weight_scale_inv", "weight_scale_2")):
Expand Down
14 changes: 14 additions & 0 deletions test/registered/unit/model_loader/test_weight_utils.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -267,6 +267,20 @@ def test_nvfp4_dummy_scales_are_neutral(self):
self.assertTrue(torch.equal(model.w2_weight_scale_2, torch.ones(4)))
self.assertTrue(torch.all((model.weight >= 0.25) & (model.weight <= 0.75)))

def test_ue8m0_dummy_block_scales_are_neutral(self):
model = torch.nn.Linear(8, 4, bias=False)
model.register_buffer(
"w13_weight_scale_inv", torch.zeros(4, 2, dtype=torch.uint8)
)

initialize_dummy_weights(model, low=0.25, high=0.75)

self.assertTrue(
torch.equal(
model.w13_weight_scale_inv, torch.full((4, 2), 127, dtype=torch.uint8)
)
)


if __name__ == "__main__":
unittest.main()
Loading