Fix #12116: preserve boolean dtype for attention masks in ChromaPipeline #12263
Conversation
- Convert attention masks to bool and prevent dtype corruption - Fix both positive and negative mask handling in _get_t5_prompt_embeds - Remove float conversion in _prepare_attention_mask method Fixes huggingface#12116
akshay-babbar
changed the title
akshay-babbar
changed the title
akshay-babbar
changed the title
|
hello @DN6, just checking in to see if you’ve had a chance to look at the above PR. If you’re not the right person or are keeping busy, would you mind pointing me to someone who could review it? Thanks! |
|
thanks @akshay-babbar |
|
The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update. |
|
Thanks for the response! I'm new to diffusers, so still learning best practices through the docs and codebase. Do let me know if there are any issues with my changes. I used these these 3 prompts - [ negative prompt used - Below are the results! Please review and let me know your feedback and any next steps. Thanks! BeforeMan
Doctor
King
After Code ChangesMan
Doctor
King
|
Problem
Fixes #12116
Short prompts generate corrupted images due to attention mask dtype conversion bug.
Root Cause
Attention masks converted from bool → float16/bfloat16, but PyTorch's scaled_dot_product_attention requires boolean masks.
Solution
_get_t5_prompt_embeds_prepare_attention_maskTesting
✅ Added @slow unit tests for dtype preservation
✅ Verified fix with prompts: "man", "cat"
✅ All tests pass locally
Please review when you have a chance. Thank you for your time and consideration!