Default conv weights on host

[jnie-TT]

Defaults conv weights layout to host since they need to be prepared on host anyway. Patches uplift CI failures on tt-torch and tt-forge-fe.

FYI: @kmabeeTT @AleksKnezevic @pilkicTT

[mtopalovicTT]

Side question. What was motivation for TTNNLayoutFuncInputOutputTypeRewriter pass? I might have missed your explanation somewhere. To my knowledge we don't have cases where inputs/outputs of func is on device

[jnie-TT]

Side question. What was motivation for TTNNLayoutFuncInputOutputTypeRewriter pass? I might have missed your explanation somewhere. To my knowledge we don't have cases where inputs/outputs of func is on device

Hey @mtopalovicTT yeah we didn't previously but with a recent runtime stitching PR that went in all inputs/outputs now default on dram and are tiled, so that we can pre-push weights of models to device

[mtopalovicTT]

Side question. What was motivation for TTNNLayoutFuncInputOutputTypeRewriter pass? I might have missed your explanation somewhere. To my knowledge we don't have cases where inputs/outputs of func is on device

Hey @mtopalovicTT yeah we didn't previously but with a recent runtime stitching PR that went in all inputs/outputs now default on dram and are tiled, so that we can pre-push weights of models to device

@jnie-TT Ahhh I somehow missed that part when I reviewed that change. Makes sense now.

[kmabeeTT]

Thanks Jackson. Since FEs are all busted right now on models let's merge when CI passes before 24 hours since this was already tested against forge and torch CI.