Fix path generation in IP Adapter #6564
Conversation
| image_encoder = CLIPVisionModelWithProjection.from_pretrained( | ||
| pretrained_model_name_or_path_or_dict, | ||
| subfolder=os.path.join(subfolder, "image_encoder"), | ||
| subfolder=Path(subfolder, "image_encoder").as_posix(), |
There was a problem hiding this comment.
It seems to be a good idea to follow this throughout in the library to better support Windows.
What say @patrickvonplaten @yiyixuxu?
| if all(proc.__class__ in ADDED_KV_ATTENTION_PROCESSORS for proc in self.attn_processors.values()): | ||
| processor = AttnAddedKVProcessor() | ||
| processor = ( | ||
| AttnAddedKVProcessor2_0() if hasattr(F, "scaled_dot_product_attention") else AttnAddedKVProcessor() |
There was a problem hiding this comment.
Hmm I still understand AttnAddedKVProcessor and AttnProcessor really as the default attn processor. This is a bit of a breaking change. Could we maybe move this change out of the PR for now? cc @yiyixuxu @sayakpaul and @DN6 as well
There was a problem hiding this comment.
That should be the case indeed. No optimized things as the default.
patrickvonplaten
left a comment
patrickvonplaten
left a comment
There was a problem hiding this comment.
Hey @fabiorigano,
Super nice fix by using posix()! Could we for now maybe try to focus only on this change and not also mix a change of set_default_attn_processor in this PR ?
fabiorigano
changed the title
|
Removed changes on |
|
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. |
|
@sayakpaul just to be sure about the purpose of this fix: os.path.join seem to be working in all the other instances in the library, while it doesn't work when used in .from_pretrained(os.path.join()). I don't know if it worth it to update all instances in the library |
|
That is weird. But it's good to follow a convention across the board I think. |
* Fix path generation on Windows * Update set_default_attn_processors * Use pathlib * Fix quality * Fix copy * Revert changes in set_default_attn_processors --------- Co-authored-by: Sayak Paul <[email protected]>
5 participants
What does this PR do?
Fixes #6561
and #6192 (comment)@patrickvonplaten @sayakpaul