AutoencoderRAE loading error with older transformers #13279

Activity

1. arijit-hub

   added

   bugSomething isn't workingSomething isn't working

   on Mar 17, 2026
2. kashif commented on Mar 17, 2026

   

   kashif

   on Mar 17, 2026

   Contributor

   More actions

   thanks @arijit-hub checking

   👍React with 👍1Reacted by Arijit Ghosh
3. kashif

   self-assigned this

   on Mar 17, 2026
4. kashif commented on Mar 18, 2026

   

   kashif

   on Mar 18, 2026

   Contributor

   More actions

   I have fixed ImageReward to work with transformers v5, can you try with:

     pip install "ImageReward @ git+https://github.com/kashif/ImageReward.git@transofrmers-v5"
   

   👍React with 👍1Reacted by Arijit Ghosh
5. arijit-hub commented on Mar 18, 2026

   

   arijit-hub

   on Mar 18, 2026

   Author

   More actions

   Hye,

   Thanks for looking into it so quickly. Really appreciate it.

   Unfortunately, this leads to import error in ImageReward due to transformers :(

   ImportError: cannot import name 'apply_chunking_to_forward' from 'transformers.modeling_utils'
   

   Maybe doing something in diffusers makes more sense than changing ImageReward?

   Again, thanks a lot for looking into it!
6. lavinal712 commented on Oct 4, 2026

   

   lavinal712

   on Oct 4, 2026

   Last edited by lavinal712

   Contributor

   More actions

   small patch to src/diffusers/models/modeling_utils.py

   @contextmanager
   def no_init_weights():
       """
       Context manager to globally disable weight initialization to speed up loading large models. To do that, all the
       torch.nn.init function are all replaced with skip.
       """
   
       def _skip_init(*args, **kwargs):
           pass
   
       hf_modeling_utils = None
       hf_init_weights = None
       if is_transformers_available():
           from transformers import modeling_utils as hf_modeling_utils
   
           hf_init_weights = hf_modeling_utils._init_weights
           hf_modeling_utils._init_weights = False
       for name, init_func in TORCH_INIT_FUNCTIONS.items():
           setattr(torch.nn.init, name, _skip_init)
       try:
           yield
       finally:
           # Restore the original initialization functions
           for name, init_func in TORCH_INIT_FUNCTIONS.items():
               setattr(torch.nn.init, name, init_func)
           if hf_modeling_utils is not None:
               hf_modeling_utils._init_weights = hf_init_weights