Michael Misiewicz
Michael Misiewicz
I'd also add that I saw the same effects when running `timeit` with the toy examples in `docs/source/notes/mps.rst`. (e.g.: `timeit.timeit(lambda: x * 2, number=100000)` on both mps/GPU.
Example of the toy example run: ```python In [17]: # toy example mps ...: import timeit ...: import torch ...: import random ...: ...: x = torch.ones(5000, device="mps") ...: timeit.timeit(lambda:...
> The neural engine can't be used for training anyway. It only supports Float16, Int8, and UInt8, and is only accessible through CoreML and MLCompute. PyTorch uses neither of these...
Fascinating. That hypothesis might also explain why the delta is so much worse with the toy example compared to the full size BERT.
@kulinseth @albanD I noticed this ticket's been subject to triage, and a few other folks have filed issues regarding similar observations. Do you know how the [figure in the press...
@philipturner Thanks for the great read. @kulinseth thanks for sharing, I'll try running that on my system, and I'm curious to dig in, because when I ran my own benchmarks...
Agreed for XLSX export! And CSVs too!
Interesting, thanks for the context. Updating the output might be helpful for confusion reduction.
Amazing news! Confirmed it looks good with the latest commit!
Seconded!