Comment by janalsncm

One thing I would like in PyTorch is for a tensor’s shape to be a fundamental part of its type. That is, disable implicit broadcasting and if an operation would require adding dimensions to inputs, require those inputs to be type cast to the correct shape first.

I can’t tell you how much time I have wasted on broadcasting bugs, where operations “work” but they aren’t doing what I want them to.

Jax can do this but no one uses Jax because of other reasons.