Open melooooody opened 2 months ago
SelectiveScanMamba is a wrapper of selective_scan_fn. SelectiveScanCore delete some codes in mamba so that the code runs faster. SelectiveScanOflex support input with dtype float16 and output with float32, so that you don't have to convert the input into float32 in advance, which also raises the speed.
貌似SelectiveScanOflex效率要高些?