Skip to content

Fix float8_e4m3fnuz input scaling range #1528

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
maskyuanzh wants to merge 1 commit into
base: main
Choose a base branch
from
+2 −7
Open

Fix float8_e4m3fnuz input scaling range #1528

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
9 changes: 2 additions & 7 deletions diffsynth/core/vram/layers.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -330,13 +330,8 @@ def fp8_linear(
input = input.reshape(-1, origin_shape[-1])

x_max = torch.max(torch.abs(input), dim=-1, keepdim=True).values
fp8_max = 448.0
# For float8_e4m3fnuz, the maximum representable value is half of that of e4m3fn.
# To avoid overflow and ensure numerical compatibility during FP8 computation,
# we scale down the input by 2.0 in advance.
# This scaling will be compensated later during the final result scaling.
if self.computation_dtype == torch.float8_e4m3fnuz:
fp8_max = fp8_max / 2.0
# Use the maximum finite value defined by the selected FP8 dtype.
fp8_max = torch.finfo(self.computation_dtype).max
scale_a = torch.clamp(x_max / fp8_max, min=1.0).float().to(device=device)
scale_b = torch.ones((weight.shape[0], 1)).to(device=device)
input = input / (scale_a + 1e-8)
Expand Down