Where can I find more details about location attributes and the algorithm that it implements to decide on the location category?
I am observing a behaviour where in the IR (which contains attention op), MLIR dumps the fused location as just the array of source location and not array of named location (even if the layer is named) after fusing the ops. Wherease a minor addition to the IR (dropout+attention) dumps the fused Location as an array of NamedLocations
This sounds like it’s a frontend generating MLIR making different choices. MLIR just dumps what’s created. Either front ends or passes creating ops decide what kind of locations they want to create. Common case for fused patterns is a fused loc.
1 Like