Transformer tricks: Precomputing the first layer

MoE
Re-assign community