Reduce row does not work with 16x16 tiny tile (single face) #35356
Description
For Moe Fused Gate, we have 1x9 data that we want to reduce to a single value using reduce row.
We are currently using a 16x16 tiny tile, but reduce row hangs. If we change it to 16x32 or 32x32 tile, the reduction no longer hangs.
We can use the workaround above for our functionality testing, but we require the single face support for perf.
If it's possible to use a smaller tile for better perf, ex 1x16 or 8x16 then we should update reduce to support that as well.