Open lukasheinrich opened 3 years ago
initial experiments show that modulo some smalllish fixes pytorch XLA could work
This is great to see @lukasheinrich. Can you comment on how this fits in with Issue #1244?
Ah, I now see after looking at the docs and the pytorch/xla
repo that it doesn't seem trivial to install like torch
is and is probably best used as a Docker image.
I think it might be fairly related. you just pass a different device
to the tensors. The CLs computation might be more tricky sincce XLA has a static graph so nont sure how it dealls with conditionals
here's the diff https://github.com/scikit-hep/pyhf/pull/1249/files
I think it might be fairly related. you just pass a different
device
to the tensors. The CLs computation might be more tricky sincce XLA has a static graph so nont sure how it dealls with conditionals
Nice. Yeah looking at the diff it seems pretty reasonable so far.
initial experiments show that modulo some smalllish fixes pytorch XLA could work
Though it's unclear whether having multiple high-level tensor libs (jax, pytorch) that can target XLA iss benefiical (though it's the same w/ GPU)
This is the XLA graph