Memory consumption on k2.rnnt_loss_smoothed?

[jesus-villalba]

Hi, In the k2's pruned loss, it seems that only the rnnt_loss_pruned function is benchmarked here: https://github.com/csukuangfj/transducer-loss-benchmarking/blob/8373897fd86a64425baa198247cde439f10b1bd2/benchmark_k2_pruned.py#L141 But not the functions used to select the ranges. Have you considered the memory required by rnnt_loss_smoothed? Thanks

[csukuangfj]

Everything between https://github.com/csukuangfj/transducer-loss-benchmarking/blob/8373897fd86a64425baa198247cde439f10b1bd2/benchmark_k2_pruned.py#L88 and https://github.com/csukuangfj/transducer-loss-benchmarking/blob/8373897fd86a64425baa198247cde439f10b1bd2/benchmark_k2_pruned.py#L158 is benchmarked