can you give a compare with existing implementations?

[zsz00]

I appreciate your excellent work, especially the example https://juliamltools.github.io/shakespeare-gpt

there are some existing implementations of MultiHeadAttention and Transformer:
FluxML/Flux.jl#2146
https://github.com/chengchingwen/NeuralAttentionlib.jl
https://github.com/chengchingwen/Transformers.jl

can you give a compare with existing implementations ?
why you want to implementation this again ?

[JuliaMLTools]

Hello! The other version is slow and has an overly complicated interface in my opinion.

[JuliaMLTools]

TransformerBlocks.jl is intended to be minimalist and hyper-optimized for performance.