gchanan
gchanan
is it actually rejected? I seem to remember it just gives you the wrong answer.
Isn't the size "overload" called sizes? I'm not sure why.
actually `size` and `sizes` are dispatched differently; `sizes` gives you the fake-ATen sizes, `size` dispatches to TH/THC sizes.
Over in https://github.com/pytorch/pytorch/pull/3570, I made size/stride with dim arguments dispatch to the sizes/strides view, so at least they are consistent.
Did we decide about `size()` and `stride()`? Should we add these as aliases of `sizes()` / `strides()`? Rename `sizes()` and `strides()` to theses? Neither?
What is the requirement of the grad_output normally? Isn't it roughly that its the same type of tensor (and memory format?) as the output tensor? Would it be valid to...
``` def sin_backward(grad_output, input): if input.is_sparse(): return x.sin() return x.cos() ``` What is `x`? What I'm trying to get at is if the _current_ rules allow you to write something...
Even more than stability, Declarations.yaml is trying to solve a fundamentally different problem, which is to generate VariableType (i.e.: add autograd support). It's not sensible to guarantee stability on an...
Great writeup! A couple of questions/comments: > dtype=bf16 Why does this do AMP bf16 instead of bf16? > Digging through the AdamW documentation, the extra memory seems to be related...
pack and unpack seem worth doing. The other parts (i.e. compress/expand) could be useful, but I'm not sure it's worth doing -- it seems like at that point you'd be...