[core] reuse AttentionMixin for compatible classes
#12463
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. |
| for name, module in self.named_children(): | ||
| fn_recursive_attn_processor(name, module, processor) | ||
|
|
||
| # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor |
There was a problem hiding this comment.
Perhaps it's out of the scope for this PR, but I see that a lot of models additionally have a set_default_attn_processor method, usually # Copied from diffusers.models.unets.unet_2d_condition.UNet2DConditionModel.set_default_attn_processor. Do you think it makes sense to add this method to AttentionMixin?
There was a problem hiding this comment.
IMO, not yet since AttentionMixin is fairly agnostic to the model-type but set_default_attn_processor relies on some custom attention processor types. For UNet2DConditionModel, we have:
diffusers/src/diffusers/models/unets/unet_2d_condition.py
Lines 762 to 769 in fa468c5
However, for AutoencoderKL Temporal Decoder:
I'd be down to the refactoring, though. Cc: @DN6
dg845
left a comment
dg845
left a comment
There was a problem hiding this comment.
Looks good to me! I think AuraFlowTransformer2DModel and AudioLDM2UNet2DConditionModel have their attn_processor/set_attn_processor methods deleted but are missing the corresponding change to inherit from AttentionMixin.
|
Thanks for those catches, @dg845. Should have been fixed by now. |
|
LGTM :) |
|
@DN6 okay to go? |
|
@DN6 a gentle ping. |
What does this PR do?
Many models use
"# Copied from ..."implementations ofattn_processorsandset_attn_processor. They are basically the same as what we have implemented indiffusers/src/diffusers/models/attention.py
Line 39 in 693d8a3
This PR makes those models inherit from
AttentionMixinand removes the copied-over implementations.I decided to leave
fuse_qkv_projectionsandunfuse_qkv_projectionsout of this PR because some models don't have attention processors implemented in a way that would make this seamless. But the methods removed in this PR should be very harmless.