question about RotaryPEMultiHeadAttention: rotary_percentage

[YOONSEOKHEO]

I confirmed that there is code in the RotaryPEMultiHeadAttention class that reduces the dimension using a parameter called rope_percentage.
(URL:

annotated_deep_learning_paper_implementations/labml_nn/transformers/rope/__init__.py

Line 205 in 285cb37

d_rope = int(self.d_k * rope_percentage)

)

I am curious in what cases you would set rope_percentage to a value less than 1.

(Of course, in experiment.py, we confirmed that rope_percentage is set to 1.0.)

[vpj]

I'm also not sure. I usually set it to 1. I have seen implementations where it's set to 0.5. I guess they do it so that some dimensions never get rotated and it makes it easier for the model to use attention only using content with no interference from the positional information.