Fix QwenImage txt_seq_lens handling #12702
Conversation
|
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. |
|
just a few comments, not a full review:
|
| raise ValueError(f"`txt_seq_lens` must have length {batch_size}, but got {len(txt_seq_lens)} instead.") | ||
| text_seq_len = max(text_seq_len, max(txt_seq_lens)) | ||
| elif encoder_hidden_states_mask is not None: | ||
| text_seq_len = max(text_seq_len, int(encoder_hidden_states_mask.sum(dim=1).max().item())) |
There was a problem hiding this comment.
This only works if the attention mask is in the form of [True, True, True, ..., False, False, False]. While this is the case in the most common use case of text attention masks, it doesn't have to be the case.
If the mask is [True, False, True, False, True, False], self.pos_embed receives an incorrect sequence length
There was a problem hiding this comment.
I think this is resolved, but I let you close it if you agree
|
thanks @dxqb the idea was to remove |
- Remove seq_lens parameter from dispatch_attention_fn - Update varlen backends to extract seqlens from masks - Update QwenImage to pass 2D joint_attention_mask - Fix native backend to handle 2D boolean masks - Fix sage_varlen seqlens_q to match seqlens_k for self-attention Note: sage_varlen still producing black images, needs further investigation
…to txt_seq_lens
|
some benchmarks with various backends: code: 
|
dxqb
left a comment
dxqb
left a comment
There was a problem hiding this comment.
I can confirm that this PR has the same output in a simple sampling test as this workaround: #12294 (comment)
|
thanks @dxqb can you test now? |
|
@dxqb you can also send a PR to this branch or fix this PR directly |
I'll send the small changes in your PR that I needed (like support for text sequence length per sample), but keep the PR overall separate - it's a separate feature |
… attention update merge with attention selection, add FLASH_SPLIT
yiyixuxu
left a comment
yiyixuxu
left a comment
There was a problem hiding this comment.
thanks so much, I went over again and left one feedback
I will also ask Qwen team to do a final review now:)
| encoder_hidden_states_mask=encoder_hidden_states_mask, | ||
| timestep=timestep, | ||
| img_shapes=img_shapes, | ||
| txt_seq_lens=txt_seq_lens, |
There was a problem hiding this comment.
should we also build joint mask for controlnet?
not merge-blocking though, I don't think it breaks anything
There was a problem hiding this comment.
good catch i think we should add it
Co-authored-by: YiYi Xu <yixu310@gmail.com>
7 participants
What does this PR do?
txt_seq_lensplumbing from all QwenImage pipelines and modular steps; the transformer now infers text length from encoder inputs/masks and validates optional overrides.encoder_hidden_states_maskinside the double-stream attention, avoiding fullseq_len²masks while keeping padding tokens masked.txt_seq_lensvalues and encoder masks are handled safely.Fixes #12344
Before submitting
documentation guidelines, and
here are tips on formatting docstrings.
Who can review?
Anyone in the community is free to review the PR once the tests have passed. Feel free to tag
members/contributors who may be interested in your PR.