Jit hiccup with a 1.8.0 nightly build

[KFrank]

Hello Forum!

I noticed that jit “quit working” on a particular use case with the
1.8.0.dev20201203 nightly build.

Specifically, a function that showed an about 2x jit speed-up with
version 1.6.0 shows no speed-up with version 1.8.0.

Details can be found in this post:

How to implement efficient element-wise 'ax+b' function

Hi Zhi and Alex! The short story is that I can reproduce your results. I’ve also done some further experimentation. The best I could come up with was to apply Alex’s torchscript suggestion to your “w-b” version. I also tried a Conv2d version – even though it doesn’t have BatchNorm2d’s normalization step, it was also slower. Note, torchscript gave a nice speedup with pytorch version 1.6.0, but not with a version 1.8.0 nightly build. I used your original problem – same-shaped tensors, a…

Best.

K. Frank

[Michael_Suo]

Thanks for the report! Do you mind making an issue on Github with the same repro? We changed optimizers in that time period so I suspect it may be a regression with a new optimizer but we need to check.

[KFrank]

Hi Michael!

@Michael_Suo

Do you mind making an issue on Github with the same repro?

Let me apologize for not creating a github issue. It turns out I would
need a github account to do so, which I don’t have. It might be fastest
if someone with a github account could link to my post, or something.

Best.

K. Frank

[Michael_Suo]

Sure! I’ll make one in the morning and update you