I a decentralized setting I am sending model weights from clients to the server using state_dict(), but as known state_dict() contains also buffers beside the learnable parameters (BatchNorm buffers). Is there a way to send only the learnable parameters without these buffers?