Using Mamba's recurrent mode during evaluation

[JVP15]

So, coincidentally, I've also done some experiments using Mamba instead of a Transformer for Decision Mamba (most recent repo can be found here: https://github.com/lmco/DecisionMamba). Have you tried using Mamba's inference_params during evals (essentially, the 'recurrent' mode for Mamba) instead of the parallel mode (with the restricted context length)? It's something I was testing, but I didn't get very far with it.