question about the inference path

[too-z]

When looking at Mamba's code, the inference path that processes the full sequence as input—rather than using a step-by-step method (where only the last step is used as input)—seems identical to the training-time forward pass, as discussed in the link below:

#187

Does this mean that constant-time inference is only achievable with the step-by-step method?

Thanks in advance!