Skip to content

Fix Z-Image FP16 overflow via downscaling #11187

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
vanDuven wants to merge 2 commits into
base: master
Choose a base branch
from
+10 −2
Open

Fix Z-Image FP16 overflow via downscaling #11187

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
a8ea695
Fix Z Image FP16 overflow via downscaling
vanDuven Dec 8, 2025
5905513
revert fp16 support for Lumina2
vanDuven Dec 11, 2025
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
12 changes: 10 additions & 2 deletions comfy/ldm/lumina/model.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -119,6 +119,9 @@ def forward(
xv = xv.unsqueeze(3).repeat(1, 1, 1, n_rep, 1).flatten(2, 3)
output = optimized_attention_masked(xq.movedim(1, 2), xk.movedim(1, 2), xv.movedim(1, 2), self.n_local_heads, x_mask, skip_reshape=True, transformer_options=transformer_options)

if output.dtype == torch.float16:
output.div_(4)

return self.out(output)


Expand Down Expand Up @@ -175,8 +178,12 @@ def __init__(
def _forward_silu_gating(self, x1, x3):
return clamp_fp16(F.silu(x1) * x3)

def forward(self, x):
return self.w2(self._forward_silu_gating(self.w1(x), self.w3(x)))
def forward(self, x, apply_fp16_downscale=False):
x3 = self.w3(x)
if x.dtype == torch.float16 and apply_fp16_downscale:
x3.div_(32)

return self.w2(self._forward_silu_gating(self.w1(x), x3))


class JointTransformerBlock(nn.Module):
Expand Down Expand Up @@ -287,6 +294,7 @@ def forward(
x = x + gate_mlp.unsqueeze(1).tanh() * self.ffn_norm2(
clamp_fp16(self.feed_forward(
modulate(self.ffn_norm1(x), scale_mlp),
apply_fp16_downscale=True,
))
)
else:
Expand Down