More memory efficient implementation / improvements on julius for > 100K sequence length?

[penguinshin]

Julius works great, but I notice that the memory usage grows super-linearly with filter length. Are there any settings to change or improvements to be made that could reduce gpu memory consumption?

[adefossez]

You can try deactivating the fft based implementation (not sure anymore what is the name of the parameter), especially if you have an input that requires a gradient. It's tuned for speed but not sure I do everything the best way for memory. But then things will be a lot slower. With the regular implem it will be slow but memory usage should be linear.