Nick Smith

@mywoodstock, is this issue still relevant?

@rayraykay, is there a CI pipeline we can run to test that this doesn't break packaging? Alternatively we could put under a switch. We were thinking it could simplify packaging...

> I asked this privately to @tapspatel , however could you guys please explain the specific error you're seeing, and where you're building and placing the metal SOs? We have...

I think currently metal seems to be calling patchelf on the fly to add this rpath, for example a system without patchelf fails with: ``` >>> import ttnn /bin/sh: 1:...

> > This change would enable you to not have to run patchelf or need it installed. > > That's true if `libtt_metal.so` is in the same dir, which in...

> We're about to do some work on the wheel, and honestly that's not a bad idea. We may go with that. > > In any case, it doesn't sound...

@mywoodstock, let us know your thoughts. I think we can probably help make this change if you don't object.

FYI @LPanosTT

Met offline and came up with this pseudo code as a plan: ```c++ // accepts/returns host tensors Tensor prepareWeightsChannelsLastForTTNN(Tensor weightsChannelLast); Tensor prepareWeightsChannelsFirstForTTNN(Tensor weightsChannelFirst) { return prepareWeightsChannelsLastForTTNN(transpose(weightsChannelFirst)); } Tensor prepareBiasChannelsLastForTTNN(Tensor biasChannelLast);...

@LPanosTT, which format is not supported? Channels last weights?