Training

[XyFighting]

Hi! Thank you for your excellent work!
When I run the code on the local device, the code runs fine, but when I run it on the server, it's very slow, the problem is shown as follows. Could you suggest the way to solve it?

[MzeroMiko]

Did you find the warning 'Triton not installed, fall back to pytorch implements.' or 'Can not import selective_scan_cuda_oflex. This affects speed.' in the log file?

[XyFighting]

Thank you for your reply. I find the warning 'Can not import selective_scan_cuda_oflex', but when I run the code on my local device, the speed remains normal with the above warning. Could you give further suggestions?

[MzeroMiko]

In the code, we would try import oflex (kernels/selective_scan) fisrt, if not found, selective_scan_cuda_core (but actually, this is deprecated), if not found again, selective_scan_cuda (standard mamba implementation). If all above are not found, then we use torch version to make it at least usable.

So I guess that in your environment, all the accelerated patches are not installed correctly.

[XyFighting]

Thank you for your reply, how to install selective_scan_cuda_oflex?

[MzeroMiko]

you can cd into kernels/selective_scan and pip install .

[XyFighting]

Thank you for your reply, I have solved above issue.