Implicit broadcasting

I'm facing errors when using implicit broadcasting like x[...,-1] = x[...,-1] + 5

Strangely, it sometimes works when run independently but fails if I then feed x into another op like y = gather(x,-1,indices)

I also face this issue with simpler cases like x[...,0] = 5.0

I was wondering if this is explicit behavior and if there's any standard examples or solutions. Do I just need to use torch.repeat explicitly? The issue is that my code already has many of these ops so replacing each manually would take sometime. And the torch trace works just fine.

Start by reproducing the implicit-broadcasting assignments x[...,-1] = x[...,-1] + 5 and x[...,0] = 5.0, then feed the result to gather(x,-1,indices) and compare that with an independently run operation and torch trace. Done means establishing whether these cases are supported and identifying the relevant conversion behavior or documenting the required handling.

Written by the indexing model from the issue text.