while i'm a fan of wafer's work, 1024 input token length is not a great benchmark anymore, that number is useful only at 1-4x single node a100 / h100 at various

while i'm a fan of wafer's work, 1024 input token length is not a great benchmark anymore, that number is useful only at 1-4x single node a100 / h100 at various concurrency levels