Used selective_scan_cuda and causal_conv1d_cuda, but still very slow to train

[nice66]

I used selective_scan_cuda and causal_conv1d_cuda. At first I was training at 30 seconds a round, but when I downloaded a couple of libraries, there was a problem ——ImportError: libcudart.so.11.0:. I know it's cuda so I re-downloaded mamba once but this time my training speed changed to 15 minutes a round. Does anyone know the reason for this please?

[AnaRhisT94]

I created batch decoding code, this can speedup your training. Don't use the cuda version
#477