Fix Qwen attention mask crash with diffusers >=0.37

[Rasaboun]

Problem

diffusers v0.37 (PR #12987) optimizes all-ones attention masks to None in encode_prompt() when there is no padding. This causes an AttributeError: 'NoneType' object has no attribute 'to' crash in all three Qwen image extensions during training.

The recent fix in 295094b addresses qwen_image.py by switching to the private _get_qwen_prompt_embeds() API, but qwen_image_edit.py and qwen_image_edit_plus.py still crash.

Fix

Reconstruct the all-ones mask right after encode_prompt() returns None, in get_prompt_embeds() of all three Qwen variants:

prompt_embeds, prompt_embeds_mask = self.pipeline.encode_prompt(...)
if prompt_embeds_mask is None:
    prompt_embeds_mask = torch.ones(
        prompt_embeds.shape[:2], device=prompt_embeds.device, dtype=torch.int64
    )

This approach:

• Uses the public encode_prompt() API (not _get_qwen_prompt_embeds)
• Works with both old and new diffusers versions
• Requires zero downstream changes — the mask is always a tensor
• Fixes all three variants: base, edit, and edit_plus

Also removes redundant duplicate mask assignments in qwen_image_edit.py and qwen_image_edit_plus.py.

Fixes #740

[Copilot]

Pull request overview

Fixes a training-time crash in the Qwen image model extensions when used with diffusers>=0.37, where encode_prompt() may return None for an all-ones attention mask (no padding).

Changes:

• Reconstruct an all-ones prompt_embeds_mask tensor when encode_prompt() returns None in all Qwen variants (base/edit/edit_plus).
• Remove redundant re-assignment of prompt_embeds_mask from text_embeddings.attention_mask in the edit/edit_plus noise prediction paths.
• Ensure the mask passed into the transformer is a tensor (and in qwen_image_edit.py, passed as a detached tensor).

Reviewed changes

Copilot reviewed 3 out of 3 changed files in this pull request and generated no comments.

File	Description
extensions_built_in/diffusion_models/qwen_image/qwen_image.py	Switches prompt embedding extraction back to public encode_prompt() and guards against None masks by recreating an all-ones mask.
extensions_built_in/diffusion_models/qwen_image/qwen_image_edit.py	Adds the same None-mask reconstruction and removes redundant mask reassignment; passes a detached mask into the transformer.
extensions_built_in/diffusion_models/qwen_image/qwen_image_edit_plus.py	Adds the same None-mask reconstruction and removes redundant mask reassignment in noise prediction.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[jaretburkett]

Thank you! @Rasaboun