How would you optimize a large Transformer model for deployment on Tenstorrent hardware when SRAM is limited?
