attention.py has at the moment two concurrent attention implementations which essentially do the exact same thing:

• diffusers/src/diffusers/models/attention.py

  Line 256 in 62608a9

  class AttentionBlock(nn.Module):

  
  and

• diffusers/src/diffusers/models/cross_attention.py

  Line 30 in 62608a9

  class CrossAttention(nn.Module):

Both

We should start deprecating

Deprecating this class won't be easy as it essentially means we have to force people to re-upload their weights. Essentially every model checkpoint that made use of

I would propose to do this in the following way:

• 1. To begin with when

     diffusers/src/diffusers/models/attention.py

     Line 256 in 62608a9

     class AttentionBlock(nn.Module):

     is called the code will convert the weights on the fly to

     diffusers/src/diffusers/models/cross_attention.py

     Line 30 in 62608a9

     class CrossAttention(nn.Module):

     with a very clear deprecation message that explains in detail how to one can save & re-upload the weights to remove the deprecation message
• 2. Open a mass PR on all checkpoints that make use of

     diffusers/src/diffusers/models/attention.py

     Line 256 in 62608a9

     class AttentionBlock(nn.Module):

     (can be retrieved via the config) to convert the weights to the new format.

Happy to help with this PR. @williamberman are you maybe interested in taking this one?