Skip to content

Fix Qwen attention mask crash with diffusers >=0.37 #748

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
jaretburkett merged 2 commits into from
Mar 23, 2026
+21 −11
Merged

Fix Qwen attention mask crash with diffusers >=0.37 #748

Show file tree
Hide file tree
Changes from all commits
Commits
Show all changes
2 commits
Select commit Hold shift + click to select a range
6cc5dde
Fix Qwen Image mask handling
Rasaboun Mar 23, 2026
c9a9955
Fix Qwen attention mask crash with diffusers >=0.37
Rasaboun Mar 23, 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
16 changes: 10 additions & 6 deletions extensions_built_in/diffusion_models/qwen_image/qwen_image.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -355,12 +355,16 @@ def get_prompt_embeds(self, prompt: str) -> PromptEmbeds:
if self.pipeline.text_encoder.device != self.device_torch:
self.pipeline.text_encoder.to(self.device_torch)

max_sequence_length = 1024

prompt_embeds, prompt_embeds_mask = self.pipeline._get_qwen_prompt_embeds(prompt, self.device_torch)
prompt_embeds = prompt_embeds[:, :max_sequence_length]
prompt_embeds_mask = prompt_embeds_mask[:, :max_sequence_length]

prompt_embeds, prompt_embeds_mask = self.pipeline.encode_prompt(
prompt,
device=self.device_torch,
num_images_per_prompt=1,
)
# diffusers >=0.37 returns None when all tokens are valid (no padding)
if prompt_embeds_mask is None:
prompt_embeds_mask = torch.ones(
prompt_embeds.shape[:2], device=prompt_embeds.device, dtype=torch.int64
)
pe = PromptEmbeds(prompt_embeds)
pe.attention_mask = prompt_embeds_mask
return pe
Expand Down
8 changes: 6 additions & 2 deletions extensions_built_in/diffusion_models/qwen_image/qwen_image_edit.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -197,6 +197,11 @@ def get_prompt_embeds(self, prompt: str, control_images=None) -> PromptEmbeds:
device=self.device_torch,
num_images_per_prompt=1,
)
# diffusers >=0.37 returns None when all tokens are valid (no padding)
if prompt_embeds_mask is None:
prompt_embeds_mask = torch.ones(
prompt_embeds.shape[:2], device=prompt_embeds.device, dtype=torch.int64
)
pe = PromptEmbeds(prompt_embeds)
pe.attention_mask = prompt_embeds_mask
return pe
Expand Down Expand Up @@ -251,14 +256,13 @@ def get_noise_prediction(
)
txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist()
enc_hs = text_embeddings.text_embeds.to(self.device_torch, self.torch_dtype)
prompt_embeds_mask = text_embeddings.attention_mask.to(self.device_torch, dtype=torch.int64)

noise_pred = self.transformer(
hidden_states=latent_model_input.to(self.device_torch, self.torch_dtype),
timestep=timestep / 1000,
guidance=None,
encoder_hidden_states=enc_hs,
encoder_hidden_states_mask=prompt_embeds_mask,
encoder_hidden_states_mask=prompt_embeds_mask.detach(),
img_shapes=img_shapes,
txt_seq_lens=txt_seq_lens,
return_dict=False,
Expand Down
8 changes: 5 additions & 3 deletions extensions_built_in/diffusion_models/qwen_image/qwen_image_edit_plus.py
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -192,6 +192,11 @@ def get_prompt_embeds(self, prompt: str, control_images=None) -> PromptEmbeds:
device=self.device_torch,
num_images_per_prompt=1,
)
# diffusers >=0.37 returns None when all tokens are valid (no padding)
if prompt_embeds_mask is None:
prompt_embeds_mask = torch.ones(
prompt_embeds.shape[:2], device=prompt_embeds.device, dtype=torch.int64
)
pe = PromptEmbeds(prompt_embeds)
pe.attention_mask = prompt_embeds_mask
return pe
Expand Down Expand Up @@ -323,9 +328,6 @@ def get_noise_prediction(
)
txt_seq_lens = prompt_embeds_mask.sum(dim=1).tolist()
enc_hs = text_embeddings.text_embeds.to(self.device_torch, self.torch_dtype)
prompt_embeds_mask = text_embeddings.attention_mask.to(
self.device_torch, dtype=torch.int64
)

noise_pred = self.transformer(
hidden_states=latent_model_input.to(
Expand Down
Loading