[Feature Request] Groups added to Conv2d about mlx

[awni]

Is anyone working on groups for convolution? E.g. @Stealeristaken ?

I am closing #637 as it's inactive but it might be useful as a starting point.

[awni]

I don't think anyone is actively working on it, would be great to have you work on it @Rifur13 !

[Stealeristaken]

I'm eager to get this working as well. I can pick this up if no one is working on it.

feel free to take. gl!

[dc-dc-dc]

Should the assert be removed and create a separate issue to improve Convolution speed on GPU when dilation is not 1? Creates more visibility to the core issue and opens the possibility of a contributor able to help out

[a1eaiactaest]

agree with @dc-dc-dc, this would be really helpful for implementations that require dilation

[Stealeristaken]

Is anyone working on groups for convolution? E.g. @Stealeristaken ?

I am closing #637 as it's inactive but it might be useful as a starting point.

Been a bit busy these days. I tried adding groups to convs, but it didn't work very well. So, I closed that pull request. I'll try again in the next few days.

[Rifur13]

I'm eager to get this working as well. I can pick this up if no one is working on it.

[dc-dc-dc]

@jagrit06 in conv2d op call, there is an assert to check if dilations is set to 1, but from a quick look through the implementation it seems to support dilations. I removed this check and messed with it a little comparing to pytorch and am seeing the same results with different values for dilations. Is there a specific case where dilations isn't supported?

[jagrit06]

Well at the moment the convolutions with dilations revert to a fairly naive and slow kernel on the GPU (though I think we do fine on the CPU)

That restriction was added to basically side step that performance hit

[Jimexist]

FWIW this PR now also depends on a groups param present.

[a1eaiactaest]

Added dilation in #766