The code of this function is basically the same as the code of the function in PIRScaleDotProductPattern above, and rotate_half, mlp, etc. also have the same problem. complex networks should be a combination of basic networks, rather than a re implementation.

done

There are many variations of MLP, such as having 2 or 3 linear layers, with bias or no bias, and the activation function may also be different, so the name of MLPPattern needs to be carefully considered.

also, is it necessary to use the prefix "PIR"?

given that pir as default feature for paddlepaddle, remove all prefix "PIR" here

it is better to move all these configurations which required for each subclass to the parent class?

shall we set the shape smaller without losing generality and reducing memory usage.

done

Should this break be removed, such as pattern=(matmul, add, matmul, add,) and prog=(matmul, add, matmul, add, matmul, add, Relu, ...) will it cause a matching interruption?

It seems like there's no difference between these two set, just keep one?

using sequence parallel but not requiring transpose in MLP, is it better to call reshard_mlp_layer_input?

done

using sequence parallel but not requiring transpose in MLP, is it better to call reshard_mlp_layer_output?

done

logging can be added here and some key locations for easy debugging.

I will add logging in next PR. Thanks