[Bug Fix] Add partial rotary factor support for Phi-4 and upgrade to transformers v4.50.0 #3984
Conversation
|
@zhaochenyang20 ready to be reviewed. Some inconsistencies in the CI with accuracy but should be good |
|
@adarshxs thanks. yi and me can help to rerun the CI. @yizhang2077 could you help to review this? |
|
@zhaochenyang20 @yizhang2077 any update on this? |
There was a problem hiding this comment.
@adarshxs Sorry I am late. Thanks for your work, I leave some comments here~
|
@adarshxs great work!!! do not rebase with main, let me rerun for you |
|
@adarshxs @zhaochenyang20 @yizhang2077 @mickqian You are great!! |
| # fix: for Qwen2-VL model, inject default 'size' if not provided. | ||
| if config.model_type in {"qwen2_vl"}: | ||
| if "size" not in kwargs: | ||
| kwargs["size"] = {"shortest_edge": 3136, "longest_edge": 1003520} |
There was a problem hiding this comment.
I would like to ask about the intention of injecting the default ‘size’ here for the Qwen2-VL model. I noticed that after Transformers version 4.54.0, this injection no longer works. I’m not sure whether I need to adjust it in order to make it work again.
There was a problem hiding this comment.
as far as i remember, prior to transformers v4.50.0, Qwen2-VL model's preprocessor_config.json only contained min_pixels/max_pixels and no explicit shortest_edge or longest_edge. as a result, loading those models under 4.50.0 would immediately throw ValueError: size must contain 'shortest_edge' and 'longest_edge' keys.
Motivation
Fixes: #3935
Modifications
Add partial rotary embedding support and upgrade to
transformers==4.50.0Also fix Qwen2.5VL which breaks when upgraded to
transformers==4.50.0fromtransformers==4.48.3Also minor fixes to reference_hf.py script
Checklist