Skip to content

Fix: Remove hardcoded CUDA autocast in Kandinsky 5 to fix import warning #12814

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
adi776borate wants to merge 2 commits into
base: main
Choose a base branch
from
+5 −3
Open

Fix: Remove hardcoded CUDA autocast in Kandinsky 5 to fix import warning #12814

Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
77ffd6a
Fix: Remove hardcoded CUDA autocast in Kandinsky 5 to fix import warning
adi776borate Dec 9, 2025
9afba5f
Apply style fixes
github-actions[bot] 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
8 changes: 5 additions & 3 deletions src/diffusers/models/transformers/transformer_kandinsky.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -165,10 +165,12 @@ def __init__(self, model_dim, time_dim, max_period=10000.0):
self.activation = nn.SiLU()
self.out_layer = nn.Linear(time_dim, time_dim, bias=True)

@torch.autocast(device_type="cuda", dtype=torch.float32)
def forward(self, time):
args = torch.outer(time, self.freqs.to(device=time.device))
time = time.to(dtype=torch.float32)
Copy link
Collaborator

@yiyixuxu yiyixuxu Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
time = time.to(dtype=torch.float32)
origintal_dtype = time.dtype
time = time.to(dtype=torch.float32)

freqs = self.freqs.to(device=time.device, dtype=torch.float32)
args = torch.outer(time, freqs)
time_embed = torch.cat([torch.cos(args), torch.sin(args)], dim=-1)
time_embed = time_embed.to(dtype=self.in_layer.weight.dtype)
Copy link
Collaborator

@yiyixuxu yiyixuxu Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Suggested change
time_embed = time_embed.to(dtype=self.in_layer.weight.dtype)
time_embed = time_embed.to(dtype=original_dtype)

Copy link
Contributor Author

@adi776borate adi776borate Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The reason I cast to self.in_layer.weight.dtype instead of original_dtype is to prevent runtime crashes on backends like XPU as mentioned by @vladmandic here.
If users load the pipeline in float16, and we pass time_embed as float32, that will raise an error, won't it?
I might be wrong, correct me if so.

time_embed = self.out_layer(self.activation(self.in_layer(time_embed)))
return time_embed

Expand Down Expand Up @@ -269,8 +271,8 @@ def __init__(self, time_dim, model_dim, num_params):
self.out_layer.weight.data.zero_()
self.out_layer.bias.data.zero_()

@torch.autocast(device_type="cuda", dtype=torch.float32)
def forward(self, x):
x = x.to(dtype=self.out_layer.weight.dtype)
Copy link
Collaborator

@yiyixuxu yiyixuxu Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

umm actually this did not look correct to me - we want to upcast it to float32, no?

Copy link
Contributor Author

@adi776borate adi776borate Dec 11, 2025

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Similarly, if we force x to float32 here, we might hit the same mismatch crash if the out_layer weights are float16/bfloat16.

return self.out_layer(self.activation(x))


Expand Down