Open odjuricicTT opened 4 days ago
This seems to be preventing us from properly sharding op outputs. We end up with TT-metal asserting that shard shape is not tilized.
Hey @odjuricicTT, please take a look at this discussion: https://github.com/tenstorrent/tt-mlir/issues/272
Currently, we don't have a way to model layout properly, hence I am trying to understand what can be the default behavior for the compiler.
There seem to be inconsistencies in TTNN IR op layout inputs and outputs. Currently all op inputs are set to TILE layout, while all outputs are ROW_MAJOR. Example:
All input to_layout ops convert to TILE layout, while all op outputs are ROW_MAJOR.
@sdjordjevicTT Opening this issue for discussion.