Token-throughput accounting: tokens pushed through the layer stack by the
two-forward path vs the prefix-shared single forward, and the saving ratio.
(Each two-forward branch processes L-1 = P+R-1 input tokens; the shared path
processes T = P+Rc+Rr.)
Token-throughput accounting: tokens pushed through the layer stack by the two-forward path vs the prefix-shared single forward, and the saving ratio. (Each two-forward branch processes L-1 = P+R-1 input tokens; the shared path processes T = P+Rc+Rr.)