Windows support?

[d8ahazard]

It builds (CU124), inference on windows (Zonos) works.

Required install of triton to compile. https://github.com/woct0rdho/triton-windows/releases/download/v3.2.0-windows.post9/triton-3.2.0-cp310-cp310-win_amd64.whl

Pre-built wheels for anybody who wants to play at home:

https://github.com/d8ahazard/AudioLab/releases/download/1.0.0/causal_conv1d-1.5.0.post8-cp310-cp310-win_amd64.whl https://github.com/d8ahazard/AudioLab/releases/download/1.0.0/mamba_ssm-2.2.4-cp310-cp310-win_amd64.whl

[BBC-Esq]

Do you have any wheels regarding python 3.11 and 3.12?

[PinkysBrain]

This seems to combine a lot of things. The build/dependence improvements, the changes needed for MSVC to compile the BOOL_SWITCH sections, a whole lot of nitpicking, using the reference implementation rather than CUDA for selective_scan_fn and a bugfix in selective_scan_bwd_kernel.cuh.

Does the use of the reference implementation for selective_scan_fn impact speed?

[loscrossos]

hey @d8ahazard

thanks for your windows fork!

would you mind updating your code to the latest ssm?

and also adding blackwell support to it?

here is the code you can use:

https://github.com/state-spaces/mamba/pull/735

[TimexPeachtree]

This works just tested with IBM Granite - 4.0 Hybrid Nano on Win 11 with MSVC 2022 but please see to minimize the changes and let the existing comments remain in source.