Dependencies between kernels

[tobiasgrantner]

The mamba-ssm kernel currently still requires the causal-conv1d library to be installed for some operations, e.g. in the file selective_scan_interface.py functions are imported from the package:

try:
    from causal_conv1d import causal_conv1d_fn
    from causal_conv1d.causal_conv1d_interface import causal_conv1d_cuda
except ImportError:
    causal_conv1d_fn = None
    causal_conv1d_cuda = None

However, a kernel for causal-conv1d already exists. Is it possible/supported to instead load the functions from the causal-conv1d kernel to get rid of the dependency on the library?

[danieldk]

It would be possible to use get_kernel inside the mamba-ssm kernel. We discussed kernel-kernel dependencies a bit a while ago, but haven't decided yet to do it. Thanks for bringing this up! We'll discuss a bit what the best approach is to tackle this.

[tobiasgrantner]

Thank you for your work on these kernels and the library itself, I think it's already a huge improvement over how kernels were handled before. However, right now, the mamba-ssm kernel only includes a warning about its dependence on the einops and transformers libraries. Additionally, the kernel is incompatible with the newest releases of the causal-conv1d package (e.g. line 3 of the code snippet above fails with the newest version). It would be great if more information on the required dependencies could be provided until the causal-conv1d kernel can be used inside the mamba-ssm kernel directly.

And in general, do you plan to maintain these kernels independently from the upstream?