As you may have seen, we announced that the next major version of PyTorch will be PyTorch 2.0. You might be interested in reading PT2 Manifesto - Google Docs and PT2 Architecture - Google Docs , a pair of documents which we wrote a half a year ago when we were trying to form alignment on what exactly PT2 was going to be. I went through the docs before posting and interestingly very little needed to be updated.
I’ll also be podcasting about lots of aspects of PT2, the first episode is up at https://pytorch-dev-podcast.simplecast.com/episodes/pytorch-20
Thanks for sharing these docs @ezyang. The PT2 Manifesto document recommends reading PyTorch 2.0 Performance Bets for context on performance but the bets document is not public. Can you guys make it public?
Came here to leave the same comment :))
Would be really useful to have already established trained models/baselines to experiment with 2.0!
Thank you guys!
I’ll check if we can pub it.
For preestablished trained models/baselines, most of us are using the scripts in
benchmarks/dynamo to run performance experiments. If you have torchbench/huggingface/timm installed, you can also run them; e.g.,
python benchmarks/dynamo/huggingface.py --performance --float32 -dcuda --training --inductor --no-skip --dashboard --only DebertaForQuestionAnswering --cold_start_latency
Thank you very much, will check it out!