3 ms·

I made a version using PyTorch 2.0, if anyone cares: https://colab.research.google.com/drive/1iar9NEmOjWf-RPCxVklwL6EIa4FPZRee?usp=sharing https://colab.researc

by voz_ 4y ago

I made a version using PyTorch 2.0, if anyone cares: https://colab.research.google.com/drive/1iar9NEmOjWf-RPCxVklwL6EIa4FPZRee?usp=sharing https://colab.research.google.com/drive/1iar9NEmOjWf-RPCxVkl... The Good: replacing gpt(x) w/ opt_gpt(x) speeds it up ~1.5-2x at first blush (really crude measurement) The Bad: The very first run took 43s on the initial compile! Gross! We had a mode for progress bar somewhere, but it would be cool if it was on by default. It would also be neat if there was a better set of utils around preheating. The Ugly: Nothing ugly?

swyx 4y ago
thank you for sharing! any expected benefits you see from switching this to pytorch 2.0?
- voz_ 4y ago
  Speed, but, honestly, if I am being critical, this example is both beautiful and extremely spartan. I think on an example so thoroughly bare, it is hard to see any benefits here.