Unable to get speedup same as from BetterTransformers deprecation

[horeqade]

Hey guys, I'm having an issue with deprecated BetterTransformers, my models are slowed up 2-3x times on CPU.

System Info

I am using Python 3.11
Increased lib versions

torch 1.13 -> 2.7
transformers 4.24 -> 4.48

Information
I was using code like this:
BetterTransformer.transform(self.model.bert, keep_original_model=False)
with absolutely basic fine-tuned bert on CPU

Now I'm trying to do (Added attn_implementation='sdpa'):

config = AutoConfig.from_pretrained(bert_config_path, local_files_only=True, attn_implementation='sdpa')
self.bert = BertModel(config=config)

And I got slowing down from 2-3sec to around 4-6sec.
There are 3 bert-based models in my service and each of them is contributing to slowing down.
One of the models (NER) is initiated like this:

cls._instance = pipeline(
  'token-classification',
  model=model_path,
  aggregation_strategy='average',
  device=device,
  accelerator='bettertransformer',
)

I got it, that deprecation is inevitable, but maybe you can give me some advices on what I'm missing with increasing speed of model?
Now I'm using only attn_implementation='sdpa', which is anyway set by default according to docs, and it didn't gave any speed :(
Maybe I should do smth else? Flash attention backends are not cool for CPU as I know

[horeqade]

Found a solution:

attn_implementation='eager'
BetterTransformer.transform(self.model.bert, keep_original_model=False)

Without eager implementation the BetterTransformers was throwing a error, that waas saying that model is already BetterTransformed.
Feels like here is a space for a fix - not throw a error when sdpa is activated, coz there are other optimizations in BetterTransformers.

Problem can be closed