Open dmakoviichuk-tt opened 1 week ago
Are you able to use ttnn.bcast here instead as a workaround?
@dmakoviichuk-tt @eyonland https://github.com/tenstorrent/tt-metal/pull/13673
The above PR holds functionality support for broadcasting the second dimension using repeat. This will affect the performance as we are using the repeat op but we can use it now for functionality and update this later once broadcast support is updated
Describe the bug Ttnn::multiply do not broadcast in second dimension. Fails silently by throwing near NaN/Inf valuesTo Reproduce Steps to reproduce the behavior:
It will print a garbage. Expected behavior it should work.
Additional context Currently it requires us to generate broadcasted vectors on cpu. It is a huge problem for a training performance. @eyonland if you don't think you can fix it please let me know I'll find another owner.